From 10a2406232ef782790de01c24844adb0d573a6e4 Mon Sep 17 00:00:00 2001
From: Alyssa Rosenzweig <alyssa@collabora.com>
Date: Mon, 6 Jun 2022 13:56:50 -0400
Subject: [PATCH] pan/perf: Fix performance counters on G57

The performance counter layout depends on the number of L2 blocks and the number
of shader cores. It doesn't make a ton of sense to hardcode these into the XML
files. Instead, let's make the coutner offsets in the XML files relative to the
categories (blocks), so we can calculate the offsets of the categories
themselves at runtime based on the computed layout. This fixes performance
counters on Mali-G57 as implemented on MT8192.

There is little code change here, mainly churn from changing the XML definition.
Postprocessing for the XML to make it suitable for Mesa uses Antonio Caggiano's
https://gitlab.freedesktop.org/panfrost/hwc-helper tool.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Antonio Caggiano <antonio.caggiano@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16803>
---
 src/panfrost/perf/G31.xml         | 292 ++++++++++++++---------------
 src/panfrost/perf/G51.xml         | 292 ++++++++++++++---------------
 src/panfrost/perf/G52.xml         | 296 ++++++++++++++---------------
 src/panfrost/perf/G57.xml         | 300 +++++++++++++++---------------
 src/panfrost/perf/G68.xml         | 300 +++++++++++++++---------------
 src/panfrost/perf/G71.xml         | 292 ++++++++++++++---------------
 src/panfrost/perf/G72.xml         | 292 ++++++++++++++---------------
 src/panfrost/perf/G76.xml         | 296 ++++++++++++++---------------
 src/panfrost/perf/G77.xml         | 300 +++++++++++++++---------------
 src/panfrost/perf/G78.xml         | 300 +++++++++++++++---------------
 src/panfrost/perf/T72x.xml        | 128 ++++++-------
 src/panfrost/perf/T76x.xml        | 158 ++++++++--------
 src/panfrost/perf/T82x.xml        | 142 +++++++-------
 src/panfrost/perf/T83x.xml        | 142 +++++++-------
 src/panfrost/perf/T86x.xml        | 158 ++++++++--------
 src/panfrost/perf/T88x.xml        | 158 ++++++++--------
 src/panfrost/perf/pan_gen_perf.py |   2 +-
 src/panfrost/perf/pan_perf.c      |  24 ++-
 src/panfrost/perf/pan_perf.h      |  11 +-
 19 files changed, 1949 insertions(+), 1934 deletions(-)

diff --git a/src/panfrost/perf/G31.xml b/src/panfrost/perf/G31.xml
index 4fef36640b5..7400b1ddaf6 100644
--- a/src/panfrost/perf/G31.xml
+++ b/src/panfrost/perf/G31.xml
@@ -1,6 +1,6 @@
 <!--
 Copyright © 2017-2020 ARM Limited.
-Copyright © 2021 Collabora, Ltd.
+Copyright © 2021-2022 Collabora, Ltd.
 Author: Antonio Caggiano <antonio.caggiano@collabora.com>
 
 Permission is hereby granted, free of charge, to any person obtaining a
@@ -24,154 +24,154 @@ SOFTWARE.
 -->
 <metrics id="TDVx">
     <category name="Job Manager" per_cpu="no">
-        <event offset="6" counter="GPU_ACTIVE" title="GPU Cycles" name="GPU active" description="The number of cycles where the GPU has a workload of any type queued for processing." units="cycles" />
-        <event offset="7" advanced="yes" counter="IRQ_ACTIVE" title="GPU Cycles" name="Interrupt active" description="The number of cycles where the GPU has a pending interrupt." units="cycles" />
-        <event offset="8" advanced="yes" counter="JS0_JOBS" title="GPU Jobs" name="Fragment jobs" description="The number of jobs processed by the GPU fragment queue." units="jobs" />
-        <event offset="9" counter="JS0_TASKS" title="GPU Tasks" name="Fragment tasks" description="The number of 32x32 pixel tasks processed by the GPU fragment queue." units="tasks" />
-        <event offset="10" counter="JS0_ACTIVE" title="GPU Cycles" name="Fragment queue active" description="The number of cycles where work is queued for processing in the GPU fragment queue." units="cycles" />
-        <event offset="12" advanced="yes" counter="JS0_WAIT_READ" title="GPU Wait Cycles" name="Fragment descriptor reads cycles" description="The number of cycles where queued fragment work is waiting for a descriptor load." units="cycles" />
-        <event offset="13" advanced="yes" counter="JS0_WAIT_ISSUE" title="GPU Wait Cycles" name="Fragment job issue cycles" description="The number of cycles where queued fragment work is waiting for an available processor." units="cycles" />
-        <event offset="14" advanced="yes" counter="JS0_WAIT_DEPEND" title="GPU Wait Cycles" name="Fragment job dependency cycles" description="The number of cycles where queued fragment work is waiting for dependent work to complete." units="cycles" />
-        <event offset="15" advanced="yes" counter="JS0_WAIT_FINISH" title="GPU Wait Cycles" name="Fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued fragment work to complete." units="cycles" />
-        <event offset="16" advanced="yes" counter="JS1_JOBS" title="GPU Jobs" name="Non-fragment jobs" description="The number of jobs processed by the GPU non-fragment queue." units="jobs" />
-        <event offset="17" advanced="yes" counter="JS1_TASKS" title="GPU Tasks" name="Non-fragment tasks" description="The number of tasks processed by the GPU non-fragment queue." units="tasks" />
-        <event offset="18" counter="JS1_ACTIVE" title="GPU Cycles" name="Non-fragment queue active" description="The number of cycles where work is queued in the GPU non-fragment queue." units="cycles" />
-        <event offset="20" advanced="yes" counter="JS1_WAIT_READ" title="GPU Wait Cycles" name="Non-fragment descriptor read cycles" description="The number number of cycles where queued non-fragment work is waiting for a descriptor load." units="cycles" />
-        <event offset="21" advanced="yes" counter="JS1_WAIT_ISSUE" title="GPU Wait Cycles" name="Non-fragment job issue cycles" description="The number of cycles where queued non-fragment work is waiting for an available processor." units="cycles" />
-        <event offset="22" advanced="yes" counter="JS1_WAIT_DEPEND" title="GPU Wait Cycles" name="Non-fragment job dependency cycles" description="The number of cycles where queued non-fragment work is waiting for dependent work to complete." units="cycles" />
-        <event offset="23" advanced="yes" counter="JS1_WAIT_FINISH" title="GPU Wait Cycles" name="Non-fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued non-fragment work to complete." units="cycles" />
-        <event offset="24" advanced="yes" counter="JS2_JOBS" title="GPU Jobs" name="Reserved jobs" description="The number of jobs processed by the GPU reserved queue." units="jobs" />
-        <event offset="25" advanced="yes" counter="JS2_TASKS" title="GPU Tasks" name="Reserved tasks" description="The number of tasks processed by the GPU reserved queue." units="tasks" />
-        <event offset="26" advanced="yes" counter="JS2_ACTIVE" title="GPU Cycles" name="Reserved queue active" description="The number of cycles where work is queued in the GPU reserved queue." units="cycles" />
-        <event offset="28" advanced="yes" counter="JS2_WAIT_READ" title="GPU Wait Cycles" name="Reserved descriptor read cycles" description="The number of cycles where queued reserved work is waiting for a descriptor load." units="cycles" />
-        <event offset="29" advanced="yes" counter="JS2_WAIT_ISSUE" title="GPU Wait Cycles" name="Reserved job issue cycles" description="The number of cycles where queued reserved work is waiting for an available processor." units="cycles" />
-        <event offset="30" advanced="yes" counter="JS2_WAIT_DEPEND" title="GPU Wait Cycles" name="Reserved job dependency cycles" description="The number of cycles where queued reserved work is waiting for dependent work to complete." units="cycles" />
-        <event offset="31" advanced="yes" counter="JS2_WAIT_FINISH" title="GPU Wait Cycles" name="Reserved job finish cycles" description="The number of cycles where the GPU is waiting for issued reserved work to complete." units="cycles" />
+        <event offset="6" counter="GPU_ACTIVE" title="GPU Cycles" name="GPU active" description="The number of cycles where the GPU has a workload of any type queued for processing." units="cycles"/>
+        <event offset="7" advanced="yes" counter="IRQ_ACTIVE" title="GPU Cycles" name="Interrupt active" description="The number of cycles where the GPU has a pending interrupt." units="cycles"/>
+        <event offset="8" advanced="yes" counter="JS0_JOBS" title="GPU Jobs" name="Fragment jobs" description="The number of jobs processed by the GPU fragment queue." units="jobs"/>
+        <event offset="9" counter="JS0_TASKS" title="GPU Tasks" name="Fragment tasks" description="The number of 32x32 pixel tasks processed by the GPU fragment queue." units="tasks"/>
+        <event offset="10" counter="JS0_ACTIVE" title="GPU Cycles" name="Fragment queue active" description="The number of cycles where work is queued for processing in the GPU fragment queue." units="cycles"/>
+        <event offset="12" advanced="yes" counter="JS0_WAIT_READ" title="GPU Wait Cycles" name="Fragment descriptor reads cycles" description="The number of cycles where queued fragment work is waiting for a descriptor load." units="cycles"/>
+        <event offset="13" advanced="yes" counter="JS0_WAIT_ISSUE" title="GPU Wait Cycles" name="Fragment job issue cycles" description="The number of cycles where queued fragment work is waiting for an available processor." units="cycles"/>
+        <event offset="14" advanced="yes" counter="JS0_WAIT_DEPEND" title="GPU Wait Cycles" name="Fragment job dependency cycles" description="The number of cycles where queued fragment work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="15" advanced="yes" counter="JS0_WAIT_FINISH" title="GPU Wait Cycles" name="Fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued fragment work to complete." units="cycles"/>
+        <event offset="16" advanced="yes" counter="JS1_JOBS" title="GPU Jobs" name="Non-fragment jobs" description="The number of jobs processed by the GPU non-fragment queue." units="jobs"/>
+        <event offset="17" advanced="yes" counter="JS1_TASKS" title="GPU Tasks" name="Non-fragment tasks" description="The number of tasks processed by the GPU non-fragment queue." units="tasks"/>
+        <event offset="18" counter="JS1_ACTIVE" title="GPU Cycles" name="Non-fragment queue active" description="The number of cycles where work is queued in the GPU non-fragment queue." units="cycles"/>
+        <event offset="20" advanced="yes" counter="JS1_WAIT_READ" title="GPU Wait Cycles" name="Non-fragment descriptor read cycles" description="The number number of cycles where queued non-fragment work is waiting for a descriptor load." units="cycles"/>
+        <event offset="21" advanced="yes" counter="JS1_WAIT_ISSUE" title="GPU Wait Cycles" name="Non-fragment job issue cycles" description="The number of cycles where queued non-fragment work is waiting for an available processor." units="cycles"/>
+        <event offset="22" advanced="yes" counter="JS1_WAIT_DEPEND" title="GPU Wait Cycles" name="Non-fragment job dependency cycles" description="The number of cycles where queued non-fragment work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="23" advanced="yes" counter="JS1_WAIT_FINISH" title="GPU Wait Cycles" name="Non-fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued non-fragment work to complete." units="cycles"/>
+        <event offset="24" advanced="yes" counter="JS2_JOBS" title="GPU Jobs" name="Reserved jobs" description="The number of jobs processed by the GPU reserved queue." units="jobs"/>
+        <event offset="25" advanced="yes" counter="JS2_TASKS" title="GPU Tasks" name="Reserved tasks" description="The number of tasks processed by the GPU reserved queue." units="tasks"/>
+        <event offset="26" advanced="yes" counter="JS2_ACTIVE" title="GPU Cycles" name="Reserved queue active" description="The number of cycles where work is queued in the GPU reserved queue." units="cycles"/>
+        <event offset="28" advanced="yes" counter="JS2_WAIT_READ" title="GPU Wait Cycles" name="Reserved descriptor read cycles" description="The number of cycles where queued reserved work is waiting for a descriptor load." units="cycles"/>
+        <event offset="29" advanced="yes" counter="JS2_WAIT_ISSUE" title="GPU Wait Cycles" name="Reserved job issue cycles" description="The number of cycles where queued reserved work is waiting for an available processor." units="cycles"/>
+        <event offset="30" advanced="yes" counter="JS2_WAIT_DEPEND" title="GPU Wait Cycles" name="Reserved job dependency cycles" description="The number of cycles where queued reserved work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="31" advanced="yes" counter="JS2_WAIT_FINISH" title="GPU Wait Cycles" name="Reserved job finish cycles" description="The number of cycles where the GPU is waiting for issued reserved work to complete." units="cycles"/>
     </category>
     <category name="Tiler" per_cpu="no">
-        <event offset="68" counter="TILER_ACTIVE" title="GPU Cycles" name="Tiler active" description="The number of cycles where the tiler has a workload queued for processing." units="cycles" />
-        <event offset="70" counter="TRIANGLES" title="Input Primitives" name="Triangle primitives" description="The number of input triangle primitives." units="primitives" />
-        <event offset="71" counter="LINES" title="Input Primitives" name="Line primitives" description="The number of input line primitives." units="primitives" />
-        <event offset="72" counter="POINTS" title="Input Primitives" name="Point primitives" description="The number of input point primitives." units="primitives" />
-        <event offset="73" counter="FRONT_FACING" title="Visible Primitives" name="Front-facing primitives" description="The number of front-facing triangles that are visible after culling." units="primitives" />
-        <event offset="74" counter="BACK_FACING" title="Visible Primitives" name="Back-facing primitives" description="The number of back-facing triangles that are visible after culling." units="primitives" />
-        <event offset="75" counter="PRIM_VISIBLE" title="Primitive Culling" name="Visible primitives" description="The number of primitives that are visible after culling." units="primitives" />
-        <event offset="76" counter="PRIM_CULLED" title="Primitive Culling" name="Facing and XY plane test culled primitives" description="The number of primitives that are culled by facing or frustum XY plane tests." units="primitives" />
-        <event offset="77" counter="PRIM_CLIPPED" title="Primitive Culling" name="Z plane test culled primitives" description="The number of primitives that are culled by frustum Z plane tests." units="primitives" />
-        <event offset="78" counter="PRIM_SAT_CULLED" title="Primitive Culling" name="Sample test culled primitives" description="The number of primitives culled by the sample coverage test." units="primitives" />
-        <event offset="81" advanced="yes" counter="BUS_READ" title="Tiler L2 Accesses" name="Read beats" description="The number of internal bus data read cycles made by the tiler." units="beats" />
-        <event offset="83" advanced="yes" counter="BUS_WRITE" title="Tiler L2 Accesses" name="Write beats" description="The number of internal bus data write cycles made by the tiler." units="beats" />
-        <event offset="85" counter="IDVS_POS_SHAD_REQ" title="Tiler Shading Requests" name="Position shading requests" description="The number of position shading requests in the IDVS flow." units="requests" />
-        <event offset="87" advanced="yes" counter="IDVS_POS_SHAD_STALL" title="Tiler Cycles" name="Position shading stall cycles" description="The number of cycles where the tiler has a stalled position shading request." units="cycles" />
-        <event offset="88" advanced="yes" counter="IDVS_POS_FIFO_FULL" title="Tiler Cycles" name="Position FIFO full cycles" description="The number of cycles where the tiler has a stalled position shading buffer." units="cycles" />
-        <event offset="90" advanced="yes" counter="VCACHE_HIT" title="Tiler Vertex Cache" name="Position cache hits" description="The number of position lookups that result in a hit in the vertex cache." units="requests" />
-        <event offset="91" advanced="yes" counter="VCACHE_MISS" title="Tiler Vertex Cache" name="Position cache misses" description="The number of position lookups that miss in the vertex cache." units="requests" />
-        <event offset="95" advanced="yes" counter="VFETCH_STALL" title="Tiler Cycles" name="Primitive assembly busy stall cycles" description="The number of cycles where the tiler is stalled waiting for primitive assembly." units="cycles" />
-        <event offset="98" advanced="yes" counter="IDVS_VBU_HIT" title="Tiler Vertex Cache" name="Varying cache hits" description="The number of varying lookups that result in a hit in the vertex cache." units="requests" />
-        <event offset="99" advanced="yes" counter="IDVS_VBU_MISS" title="Tiler Vertex Cache" name="Varying cache misses" description="The number of varying lookups that miss in the vertex cache." units="requests" />
-        <event offset="101" counter="IDVS_VAR_SHAD_REQ" title="Tiler Shading Requests" name="Varying shading requests" description="The number of varying shading requests in the IDVS flow." units="requests" />
-        <event offset="102" advanced="yes" counter="IDVS_VAR_SHAD_STALL" title="Tiler Cycles" name="Varying shading stall cycles" description="The number of cycles where the tiler has a stalled varying shading request." units="cycles" />
-        <event offset="118" advanced="yes" counter="WRBUF_NO_AXI_ID_STALL" title="Tiler Cycles" name="Write buffer transaction stall cycles" description="The number of cycles where the tiler write buffer can not send data because it has no available write IDs." units="cycles" />
-        <event offset="119" advanced="yes" counter="WRBUF_AXI_STALL" title="Tiler Cycles" name="Write buffer write stall cycles" description="The number of cycles where the tiler write buffer can not send data because the bus is not ready." units="cycles" />
-    </category>
-    <category name="Shader Core" per_cpu="no">
-        <event offset="196" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment active" description="The number of cycles where the shader core is processing a fragment workload." units="cycles" />
-        <event offset="197" advanced="yes" counter="FRAG_PRIMITIVES" title="Core Primitives" name="Read primitives" description="The number of primitives read from the tile list by the fragment front-end." units="primitives" />
-        <event offset="198" counter="FRAG_PRIM_RAST" title="Core Primitives" name="Rasterized primitives" description="The number of primitives being rasterized." units="primitives" />
-        <event offset="199" counter="FRAG_FPK_ACTIVE" title="Core Cycles" name="Fragment FPKB active" description="The number of cycles where at least one quad is present in the pre-pipe quad queue." units="cycles" />
-        <event offset="201" counter="FRAG_WARPS" title="Core Warps" name="Fragment warps" description="The number of fragment warps created." units="warps" />
-        <event offset="202" counter="FRAG_PARTIAL_WARPS" title="Core Warps" name="Partial fragment warps" description="The number of fragment warps containing helper threads that do not correspond to a hit sample point." units="warps" />
-        <event offset="203" counter="FRAG_QUADS_RAST" title="Core Quads" name="Rasterized quads" description="The number of quads generated by the rasterization phase." units="quads" />
-        <event offset="204" counter="FRAG_QUADS_EZS_TEST" title="Core Quads" name="Early ZS tested quads" description="The number of quads that are undergoing early depth and stencil testing." units="quads" />
-        <event offset="205" counter="FRAG_QUADS_EZS_UPDATE" title="Core Quads" name="Early ZS updated quads" description="The number of quads undergoing early depth and stencil testing, that are capable of updating the framebuffer." units="quads" />
-        <event offset="206" counter="FRAG_QUADS_EZS_KILL" title="Core Quads" name="Early ZS killed quads" description="The number of quads killed by early depth and stencil testing." units="quads" />
-        <event offset="207" counter="FRAG_LZS_TEST" title="Core Quads" name="Late ZS tested quads" description="The number of quads undergoing late depth and stencil testing." units="quads" />
-        <event offset="208" counter="FRAG_LZS_KILL" title="Core Quads" name="Late ZS killed quads" description="The number of quads killed by late depth and stencil testing." units="quads" />
-        <event offset="210" counter="FRAG_PTILES" title="Core Tiles" name="Tiles" description="The number of tiles processed by the shader core." units="tiles" />
-        <event offset="211" counter="FRAG_TRANS_ELIM" title="Core Tiles" name="Constant tiles killed" description="The number of tiles killed by transaction elimination." units="tiles" />
-        <event offset="212" counter="QUAD_FPK_KILLER" title="Core Quads" name="FPK occluder quads" description="The number of quads that are valid occluders for hidden surface removal." units="quads" />
-        <event offset="214" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Non-fragment active" description="The number of cycles where the shader core is processing some non-fragment workload." units="cycles" />
-        <event offset="215" advanced="yes" counter="COMPUTE_TASKS" title="Core Tasks" name="Non-fragment tasks" description="The number of non-fragment tasks issued to the shader core." units="tasks" />
-        <event offset="216" counter="COMPUTE_WARPS" title="Core Warps" name="Non-fragment warps" description="The number of non-fragment warps created." units="warps" />
-        <event offset="217" advanced="yes" counter="COMPUTE_STARVING" title="Core Starvation Cycles" name="Non-fragment starvation cycles" description="The number of cycles where the shader core is processing a non-fragment workload and there are no new threads available for execution." units="cycles" />
-        <event offset="218" counter="EXEC_CORE_ACTIVE" title="Core Cycles" name="Execution core active" description="The number of cycles where the shader core is processing at least one warp." units="cycles" />
-        <event offset="219" advanced="yes" counter="EXEC_ACTIVE" title="Core Cycles" name="Execution engine active" description="The number of cycles where the execution engine unit is processing at least one thread." units="cycles" />
-        <event offset="220" counter="EXEC_INSTR_COUNT" title="Core EE Instructions" name="Executed instructions" description="The number of instructions executed per warp." units="instructions" />
-        <event offset="221" counter="EXEC_INSTR_DIVERGED" title="Core EE Instructions" name="Diverged instructions" description="The number of instructions executed per warp, that have control flow divergence." units="instructions" />
-        <event offset="222" advanced="yes" counter="EXEC_INSTR_STARVING" title="Core Starvation Cycles" name="Execution engine starvation cycles" description="The number of cycles where no new threads are available for execution." units="cycles" />
-        <event offset="223" advanced="yes" counter="ARITH_INSTR_SINGLE_FMA" title="Core EE Instructions" name="Arithmetic instructions" description="The number of instructions where the workload is a single FMA pipe arithmetic operation." units="instructions" />
-        <event offset="224" advanced="yes" counter="ARITH_INSTR_DOUBLE" title="Core EE Instructions" name="Dual Arithmetic instructions" description="The number of instructions where the workload is one FMA pipe arithmetic operation and one ADD pipe arithmetic operation." units="instructions" />
-        <event offset="225" advanced="yes" counter="ARITH_INSTR_MSG" title="Core EE Instructions" name="Arithmetic + Message instructions" description="The number of instructions where the workload is one FMA pipe arithmetic operation and one ADD pipe message operation" units="instructions" />
-        <event offset="226" advanced="yes" counter="ARITH_INSTR_MSG_ONLY" title="Core EE Instructions" name="Message instructions" description="The number of instructions where the workload is a single ADD pipe message operation, with no FMA pipe operation" units="instructions" />
-        <event offset="227" counter="TEX_MSGI_NUM_QUADS" title="Core Texture Quads" name="Texture requests" description="The number of quad-width texture operations processed by the texture unit." units="quads" />
-        <event offset="228" counter="TEX_DFCH_NUM_PASSES" title="Core Texture Quads" name="Texture issues" description="The number of quad-width filtering passes." units="issues" />
-        <event offset="229" counter="TEX_DFCH_NUM_PASSES_MISS" title="Core Texture Quads" name="Descriptor misses" description="The number of quad-width filtering passes that miss in the resource or sampler descriptor cache." units="requests" />
-        <event offset="230" counter="TEX_DFCH_NUM_PASSES_MIP_MAP" title="Core Texture Quads" name="Mipmapped texture issues" description="The number of quad-width filtering passes that use a mipmapped texture." units="issues" />
-        <event offset="231" counter="TEX_TIDX_NUM_SPLIT_MIP_MAP" title="Core Texture Quads" name="Trilinear filtered issues" description="The number of quad-width filtering passes that use a trilinear filter." units="issues" />
-        <event offset="232" counter="TEX_TFCH_NUM_LINES_FETCHED" title="Core Texture Line Fetches" name="Line fetches" description="The number of texture line fetches from the L2 cache." units="issues" />
-        <event offset="233" counter="TEX_TFCH_NUM_LINES_FETCHED_BLOCK_COMPRESSED" title="Core Texture Line Fetches" name="Compressed line fetches" description="The number of texture line fetches from the L2 cache that are block compressed textures." units="issues" />
-        <event offset="234" counter="TEX_TFCH_NUM_OPERATIONS" title="Core Texture Cycles" name="Cache lookups" description="The number of texture cache lookup cycles." units="requests" />
-        <event offset="235" counter="TEX_FILT_NUM_OPERATIONS" title="Core Texture Cycles" name="Texturing active" description="The number of texture filtering issue cycles." units="cycles" />
-        <event offset="236" counter="LS_MEM_READ_FULL" title="Core Load/Store Cycles" name="Full read cycles" description="The number of full-width load/store cache reads." units="cycles" />
-        <event offset="237" counter="LS_MEM_READ_SHORT" title="Core Load/Store Cycles" name="Partial read cycles" description="The number of partial-width load/store cache reads." units="cycles" />
-        <event offset="238" counter="LS_MEM_WRITE_FULL" title="Core Load/Store Cycles" name="Full write cycles" description="The number of full-width load/store cache writes." units="cycles" />
-        <event offset="239" counter="LS_MEM_WRITE_SHORT" title="Core Load/Store Cycles" name="Partial write cycles" description="The number of partial-width load/store cache writes." units="cycles" />
-        <event offset="240" counter="LS_MEM_ATOMIC" title="Core Load/Store Cycles" name="Atomic access cycles" description="The number of load/store atomic accesses." units="cycles" />
-        <event offset="241" counter="VARY_INSTR" title="Core Varying Requests" name="Interpolation requests" description="The number of warp-width interpolation operations processed by the varying unit." units="instructions" />
-        <event offset="242" counter="VARY_SLOT_32" title="Core Varying Cycles" name="32-bit interpolation active" description="The number of 32-bit interpolation cycles processed by the varying unit." units="cycles" />
-        <event offset="243" counter="VARY_SLOT_16" title="Core Varying Cycles" name="16-bit interpolation active" description="The number of 16-bit interpolation cycles processed by the varying unit." units="cycles" />
-        <event offset="244" advanced="yes" counter="ATTR_INSTR" title="Core Attribute Requests" name="Attribute requests" description="The number of instructions executed by the attribute unit." units="instructions" />
-        <event offset="245" advanced="yes" counter="ARITH_INSTR_FP_MUL" title="Core EE Instructions" name="Multiplier instructions" description="The number of instructions where the workload uses floating-point multiplier hardware." units="instructions" />
-        <event offset="246" counter="BEATS_RD_FTC" title="Core L2 Reads" name="Fragment L2 read beats" description="The number of read beats received by the fixed-function fragment front-end." units="beats" />
-        <event offset="247" counter="BEATS_RD_FTC_EXT" title="Core L2 Reads" name="Fragment external read beats" description="The number of read beats received by the fixed-function fragment front-end that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="248" counter="BEATS_RD_LSC" title="Core L2 Reads" name="Load/store L2 read beats" description="The number of read beats received by the load/store unit." units="beats" />
-        <event offset="249" counter="BEATS_RD_LSC_EXT" title="Core L2 Reads" name="Load/store external read beats" description="The number of read beats received by the load/store unit that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="250" counter="BEATS_RD_TEX" title="Core L2 Reads" name="Texture L2 read beats" description="The number of read beats received by the texture unit." units="beats" />
-        <event offset="251" counter="BEATS_RD_TEX_EXT" title="Core L2 Reads" name="Texture external read beats" description="The number of read beats received by the texture unit that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="252" advanced="yes" counter="BEATS_RD_OTHER" title="Core L2 Reads" name="Other L2 read beats" description="The number of read beats received by a unit that is not specifically identified." units="beats" />
-        <event offset="255" counter="BEATS_WR_LSC_WB" title="Core Writes" name="Load/store writeback write beats" description="The number of write beats by the load/store unit that are due to writeback." units="beats" />
-        <event offset="254" counter="BEATS_WR_TIB" title="Core Writes" name="Tile buffer write beats" description="The number of write beats sent by the tile buffer writeback unit." units="beats" />
-        <event offset="253" counter="BEATS_WR_LSC_OTHER" title="Core Writes" name="Load/store other write beats" description="The number of write beats by the load/store unit that are due to any reason other than writeback." units="beats" />
+        <event offset="4" counter="TILER_ACTIVE" title="GPU Cycles" name="Tiler active" description="The number of cycles where the tiler has a workload queued for processing." units="cycles"/>
+        <event offset="6" counter="TRIANGLES" title="Input Primitives" name="Triangle primitives" description="The number of input triangle primitives." units="primitives"/>
+        <event offset="7" counter="LINES" title="Input Primitives" name="Line primitives" description="The number of input line primitives." units="primitives"/>
+        <event offset="8" counter="POINTS" title="Input Primitives" name="Point primitives" description="The number of input point primitives." units="primitives"/>
+        <event offset="9" counter="FRONT_FACING" title="Visible Primitives" name="Front-facing primitives" description="The number of front-facing triangles that are visible after culling." units="primitives"/>
+        <event offset="10" counter="BACK_FACING" title="Visible Primitives" name="Back-facing primitives" description="The number of back-facing triangles that are visible after culling." units="primitives"/>
+        <event offset="11" counter="PRIM_VISIBLE" title="Primitive Culling" name="Visible primitives" description="The number of primitives that are visible after culling." units="primitives"/>
+        <event offset="12" counter="PRIM_CULLED" title="Primitive Culling" name="Facing and XY plane test culled primitives" description="The number of primitives that are culled by facing or frustum XY plane tests." units="primitives"/>
+        <event offset="13" counter="PRIM_CLIPPED" title="Primitive Culling" name="Z plane test culled primitives" description="The number of primitives that are culled by frustum Z plane tests." units="primitives"/>
+        <event offset="14" counter="PRIM_SAT_CULLED" title="Primitive Culling" name="Sample test culled primitives" description="The number of primitives culled by the sample coverage test." units="primitives"/>
+        <event offset="17" advanced="yes" counter="BUS_READ" title="Tiler L2 Accesses" name="Read beats" description="The number of internal bus data read cycles made by the tiler." units="beats"/>
+        <event offset="19" advanced="yes" counter="BUS_WRITE" title="Tiler L2 Accesses" name="Write beats" description="The number of internal bus data write cycles made by the tiler." units="beats"/>
+        <event offset="21" counter="IDVS_POS_SHAD_REQ" title="Tiler Shading Requests" name="Position shading requests" description="The number of position shading requests in the IDVS flow." units="requests"/>
+        <event offset="23" advanced="yes" counter="IDVS_POS_SHAD_STALL" title="Tiler Cycles" name="Position shading stall cycles" description="The number of cycles where the tiler has a stalled position shading request." units="cycles"/>
+        <event offset="24" advanced="yes" counter="IDVS_POS_FIFO_FULL" title="Tiler Cycles" name="Position FIFO full cycles" description="The number of cycles where the tiler has a stalled position shading buffer." units="cycles"/>
+        <event offset="26" advanced="yes" counter="VCACHE_HIT" title="Tiler Vertex Cache" name="Position cache hits" description="The number of position lookups that result in a hit in the vertex cache." units="requests"/>
+        <event offset="27" advanced="yes" counter="VCACHE_MISS" title="Tiler Vertex Cache" name="Position cache misses" description="The number of position lookups that miss in the vertex cache." units="requests"/>
+        <event offset="31" advanced="yes" counter="VFETCH_STALL" title="Tiler Cycles" name="Primitive assembly busy stall cycles" description="The number of cycles where the tiler is stalled waiting for primitive assembly." units="cycles"/>
+        <event offset="34" advanced="yes" counter="IDVS_VBU_HIT" title="Tiler Vertex Cache" name="Varying cache hits" description="The number of varying lookups that result in a hit in the vertex cache." units="requests"/>
+        <event offset="35" advanced="yes" counter="IDVS_VBU_MISS" title="Tiler Vertex Cache" name="Varying cache misses" description="The number of varying lookups that miss in the vertex cache." units="requests"/>
+        <event offset="37" counter="IDVS_VAR_SHAD_REQ" title="Tiler Shading Requests" name="Varying shading requests" description="The number of varying shading requests in the IDVS flow." units="requests"/>
+        <event offset="38" advanced="yes" counter="IDVS_VAR_SHAD_STALL" title="Tiler Cycles" name="Varying shading stall cycles" description="The number of cycles where the tiler has a stalled varying shading request." units="cycles"/>
+        <event offset="54" advanced="yes" counter="WRBUF_NO_AXI_ID_STALL" title="Tiler Cycles" name="Write buffer transaction stall cycles" description="The number of cycles where the tiler write buffer can not send data because it has no available write IDs." units="cycles"/>
+        <event offset="55" advanced="yes" counter="WRBUF_AXI_STALL" title="Tiler Cycles" name="Write buffer write stall cycles" description="The number of cycles where the tiler write buffer can not send data because the bus is not ready." units="cycles"/>
     </category>
     <category name="Memory System" per_cpu="no">
-        <event offset="132" advanced="yes" counter="MMU_REQUESTS" title="MMU Stage 1 Translations" name="MMU lookups" description="The number of main MMU address translations performed." units="requests" />
-        <event offset="144" advanced="yes" counter="L2_RD_MSG_IN" title="L2 Cache Requests" name="Read requests" description="The number of L2 cache read requests from internal masters." units="requests" />
-        <event offset="145" advanced="yes" counter="L2_RD_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Read stall cycles" description="The number of cycles L2 cache read requests from internal masters are stalled." units="cycles" />
-        <event offset="146" advanced="yes" counter="L2_WR_MSG_IN" title="L2 Cache Requests" name="Write requests" description="The number of L2 cache write requests from internal masters." units="requests" />
-        <event offset="147" advanced="yes" counter="L2_WR_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Write stall cycles" description="The number of cycles where L2 cache write requests from internal masters are stalled." units="cycles" />
-        <event offset="148" advanced="yes" counter="L2_SNP_MSG_IN" title="L2 Cache Requests" name="Snoop requests" description="The number of L2 snoop requests from internal masters." units="requests" />
-        <event offset="149" advanced="yes" counter="L2_SNP_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Snoop stall cycles" description="The number of cycles where L2 cache snoop requests from internal masters are stalled." units="cycles" />
-        <event offset="150" advanced="yes" counter="L2_RD_MSG_OUT" title="L2 Cache Requests" name="L1 read requests" description="The number of L1 cache read requests sent by the L2 cache to an internal master." units="requests" />
-        <event offset="151" advanced="yes" counter="L2_RD_MSG_OUT_STALL" title="L2 Cache Stall Cycles" name="L1 read stall cycles" description="The number of cycles where L1 cache read requests sent by the L2 cache to an internal master are stalled." units="cycles" />
-        <event offset="152" advanced="yes" counter="L2_WR_MSG_OUT" title="L2 Cache Requests" name="L1 write requests" description="The number of L1 cache write responses sent by the L2 cache to an internal master." units="requests" />
-        <event offset="153" counter="L2_ANY_LOOKUP" title="L2 Cache Lookups" name="Any lookup" description="The number of L2 cache lookups performed." units="requests" />
-        <event offset="154" counter="L2_READ_LOOKUP" title="L2 Cache Lookups" name="Read lookup" description="The number of L2 cache read lookups performed." units="requests" />
-        <event offset="155" counter="L2_WRITE_LOOKUP" title="L2 Cache Lookups" name="Write lookup" description="The number of L2 cache write lookups performed." units="requests" />
-        <event offset="156" advanced="yes" counter="L2_EXT_SNOOP_LOOKUP" title="L2 Cache Lookups" name="External snoop lookups" description="The number of coherency snoop lookups performed that were triggered by an external master." units="requests" />
-        <event offset="157" counter="L2_EXT_READ" title="External Bus Accesses" name="Read transaction" description="The number of external read transactions." units="transactions" />
-        <event offset="158" advanced="yes" counter="L2_EXT_READ_NOSNP" title="External Bus Accesses" name="ReadNoSnoop transactions" description="The number of external non-coherent read transactions." units="transactions" />
-        <event offset="159" advanced="yes" counter="L2_EXT_READ_UNIQUE" title="External Bus Accesses" name="ReadUnique transactions" description="The number of external coherent read unique transactions." units="transactions" />
-        <event offset="160" counter="L2_EXT_READ_BEATS" title="External Bus Beats" name="Read beat" description="The number of external bus data read cycles." units="beats" />
-        <event offset="161" counter="L2_EXT_AR_STALL" title="External Bus Stalls" name="Read stall cycles" description="The number of cycles where a read is stalled waiting for the external bus." units="cycles" />
-        <event offset="162" counter="L2_EXT_AR_CNT_Q1" title="External Bus Outstanding Reads" name="0-25% outstanding" description="The number of read transactions initiated when 0-25% of the maximum are in use." units="transactions" />
-        <event offset="163" counter="L2_EXT_AR_CNT_Q2" title="External Bus Outstanding Reads" name="25-50% outstanding" description="The number of read transactions initiated when 25-50% of the maximum are in use." units="transactions" />
-        <event offset="164" counter="L2_EXT_AR_CNT_Q3" title="External Bus Outstanding Reads" name="50-75% outstanding" description="The number of read transactions initiated when 50-75% of the maximum are in use." units="transactions" />
-        <event offset="165" counter="L2_EXT_RRESP_0_127" title="External Bus Read Latency" name="0-127 cycles" description="The number of data beats returned 0-127 cycles after the read request." units="beats" />
-        <event offset="166" counter="L2_EXT_RRESP_128_191" title="External Bus Read Latency" name="128-191 cycles" description="The number of data beats returned 128-191 cycles after the read request." units="beats" />
-        <event offset="167" counter="L2_EXT_RRESP_192_255" title="External Bus Read Latency" name="192-255 cycles" description="The number of data beats returned 192-255 cycles after the read request." units="beats" />
-        <event offset="168" counter="L2_EXT_RRESP_256_319" title="External Bus Read Latency" name="256-319 cycles" description="The number of data beats returned 256-319 cycles after the read request." units="beats" />
-        <event offset="169" counter="L2_EXT_RRESP_320_383" title="External Bus Read Latency" name="320-383 cycles" description="The number of data beats returned 320-383 cycles after the read request." units="beats" />
-        <event offset="170" counter="L2_EXT_WRITE" title="External Bus Accesses" name="Write transaction" description="The number of external write transactions." units="transactions" />
-        <event offset="171" advanced="yes" counter="L2_EXT_WRITE_NOSNP_FULL" title="External Bus Accesses" name="WriteNoSnoopFull transactions" description="The number of external non-coherent full write transactions." units="transactions" />
-        <event offset="172" advanced="yes" counter="L2_EXT_WRITE_NOSNP_PTL" title="External Bus Accesses" name="WriteNoSnoopPartial transactions" description="The number of external non-coherent partial write transactions." units="transactions" />
-        <event offset="173" advanced="yes" counter="L2_EXT_WRITE_SNP_FULL" title="External Bus Accesses" name="WriteSnoopFull transactions" description="The number of external coherent full write transactions." units="transactions" />
-        <event offset="174" advanced="yes" counter="L2_EXT_WRITE_SNP_PTL" title="External Bus Accesses" name="WriteSnoopPartial transactions" description="The number of external coherent partial write transactions." units="transactions" />
-        <event offset="175" counter="L2_EXT_WRITE_BEATS" title="External Bus Beats" name="Write beat" description="The number of external bus data write cycles." units="beats" />
-        <event offset="176" counter="L2_EXT_W_STALL" title="External Bus Stalls" name="Write stall cycles" description="The number of cycles where a write is stalled waiting for the external bus." units="cycles" />
-        <event offset="177" counter="L2_EXT_AW_CNT_Q1" title="External Bus Outstanding Writes" name="0-25% outstanding" description="The number of write transactions initiated when 0-25% of the maximum are in use." units="transactions" />
-        <event offset="178" counter="L2_EXT_AW_CNT_Q2" title="External Bus Outstanding Writes" name="25-50% outstanding" description="The number of write transactions initiated when 25-50% of the maximum are in use." units="transactions" />
-        <event offset="179" counter="L2_EXT_AW_CNT_Q3" title="External Bus Outstanding Writes" name="50-75% outstanding" description="The number of write transactions initiated when 50-75% of the maximum are in use." units="transactions" />
-        <event offset="180" advanced="yes" counter="L2_EXT_SNOOP" title="External Bus Accesses" name="Snoop transactions" description="The number of coherency snoops triggered by external masters." units="transactions" />
-        <event offset="181" advanced="yes" counter="L2_EXT_SNOOP_STALL" title="External Bus Stalls" name="Snoop stall cycles" description="The number of cycles where a coherency snoop triggered by external master is stalled." units="cycles" />
+        <event offset="4" advanced="yes" counter="MMU_REQUESTS" title="MMU Stage 1 Translations" name="MMU lookups" description="The number of main MMU address translations performed." units="requests"/>
+        <event offset="16" advanced="yes" counter="L2_RD_MSG_IN" title="L2 Cache Requests" name="Read requests" description="The number of L2 cache read requests from internal masters." units="requests"/>
+        <event offset="17" advanced="yes" counter="L2_RD_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Read stall cycles" description="The number of cycles L2 cache read requests from internal masters are stalled." units="cycles"/>
+        <event offset="18" advanced="yes" counter="L2_WR_MSG_IN" title="L2 Cache Requests" name="Write requests" description="The number of L2 cache write requests from internal masters." units="requests"/>
+        <event offset="19" advanced="yes" counter="L2_WR_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Write stall cycles" description="The number of cycles where L2 cache write requests from internal masters are stalled." units="cycles"/>
+        <event offset="20" advanced="yes" counter="L2_SNP_MSG_IN" title="L2 Cache Requests" name="Snoop requests" description="The number of L2 snoop requests from internal masters." units="requests"/>
+        <event offset="21" advanced="yes" counter="L2_SNP_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Snoop stall cycles" description="The number of cycles where L2 cache snoop requests from internal masters are stalled." units="cycles"/>
+        <event offset="22" advanced="yes" counter="L2_RD_MSG_OUT" title="L2 Cache Requests" name="L1 read requests" description="The number of L1 cache read requests sent by the L2 cache to an internal master." units="requests"/>
+        <event offset="23" advanced="yes" counter="L2_RD_MSG_OUT_STALL" title="L2 Cache Stall Cycles" name="L1 read stall cycles" description="The number of cycles where L1 cache read requests sent by the L2 cache to an internal master are stalled." units="cycles"/>
+        <event offset="24" advanced="yes" counter="L2_WR_MSG_OUT" title="L2 Cache Requests" name="L1 write requests" description="The number of L1 cache write responses sent by the L2 cache to an internal master." units="requests"/>
+        <event offset="25" counter="L2_ANY_LOOKUP" title="L2 Cache Lookups" name="Any lookup" description="The number of L2 cache lookups performed." units="requests"/>
+        <event offset="26" counter="L2_READ_LOOKUP" title="L2 Cache Lookups" name="Read lookup" description="The number of L2 cache read lookups performed." units="requests"/>
+        <event offset="27" counter="L2_WRITE_LOOKUP" title="L2 Cache Lookups" name="Write lookup" description="The number of L2 cache write lookups performed." units="requests"/>
+        <event offset="28" advanced="yes" counter="L2_EXT_SNOOP_LOOKUP" title="L2 Cache Lookups" name="External snoop lookups" description="The number of coherency snoop lookups performed that were triggered by an external master." units="requests"/>
+        <event offset="29" counter="L2_EXT_READ" title="External Bus Accesses" name="Read transaction" description="The number of external read transactions." units="transactions"/>
+        <event offset="30" advanced="yes" counter="L2_EXT_READ_NOSNP" title="External Bus Accesses" name="ReadNoSnoop transactions" description="The number of external non-coherent read transactions." units="transactions"/>
+        <event offset="31" advanced="yes" counter="L2_EXT_READ_UNIQUE" title="External Bus Accesses" name="ReadUnique transactions" description="The number of external coherent read unique transactions." units="transactions"/>
+        <event offset="32" counter="L2_EXT_READ_BEATS" title="External Bus Beats" name="Read beat" description="The number of external bus data read cycles." units="beats"/>
+        <event offset="33" counter="L2_EXT_AR_STALL" title="External Bus Stalls" name="Read stall cycles" description="The number of cycles where a read is stalled waiting for the external bus." units="cycles"/>
+        <event offset="34" counter="L2_EXT_AR_CNT_Q1" title="External Bus Outstanding Reads" name="0-25% outstanding" description="The number of read transactions initiated when 0-25% of the maximum are in use." units="transactions"/>
+        <event offset="35" counter="L2_EXT_AR_CNT_Q2" title="External Bus Outstanding Reads" name="25-50% outstanding" description="The number of read transactions initiated when 25-50% of the maximum are in use." units="transactions"/>
+        <event offset="36" counter="L2_EXT_AR_CNT_Q3" title="External Bus Outstanding Reads" name="50-75% outstanding" description="The number of read transactions initiated when 50-75% of the maximum are in use." units="transactions"/>
+        <event offset="37" counter="L2_EXT_RRESP_0_127" title="External Bus Read Latency" name="0-127 cycles" description="The number of data beats returned 0-127 cycles after the read request." units="beats"/>
+        <event offset="38" counter="L2_EXT_RRESP_128_191" title="External Bus Read Latency" name="128-191 cycles" description="The number of data beats returned 128-191 cycles after the read request." units="beats"/>
+        <event offset="39" counter="L2_EXT_RRESP_192_255" title="External Bus Read Latency" name="192-255 cycles" description="The number of data beats returned 192-255 cycles after the read request." units="beats"/>
+        <event offset="40" counter="L2_EXT_RRESP_256_319" title="External Bus Read Latency" name="256-319 cycles" description="The number of data beats returned 256-319 cycles after the read request." units="beats"/>
+        <event offset="41" counter="L2_EXT_RRESP_320_383" title="External Bus Read Latency" name="320-383 cycles" description="The number of data beats returned 320-383 cycles after the read request." units="beats"/>
+        <event offset="42" counter="L2_EXT_WRITE" title="External Bus Accesses" name="Write transaction" description="The number of external write transactions." units="transactions"/>
+        <event offset="43" advanced="yes" counter="L2_EXT_WRITE_NOSNP_FULL" title="External Bus Accesses" name="WriteNoSnoopFull transactions" description="The number of external non-coherent full write transactions." units="transactions"/>
+        <event offset="44" advanced="yes" counter="L2_EXT_WRITE_NOSNP_PTL" title="External Bus Accesses" name="WriteNoSnoopPartial transactions" description="The number of external non-coherent partial write transactions." units="transactions"/>
+        <event offset="45" advanced="yes" counter="L2_EXT_WRITE_SNP_FULL" title="External Bus Accesses" name="WriteSnoopFull transactions" description="The number of external coherent full write transactions." units="transactions"/>
+        <event offset="46" advanced="yes" counter="L2_EXT_WRITE_SNP_PTL" title="External Bus Accesses" name="WriteSnoopPartial transactions" description="The number of external coherent partial write transactions." units="transactions"/>
+        <event offset="47" counter="L2_EXT_WRITE_BEATS" title="External Bus Beats" name="Write beat" description="The number of external bus data write cycles." units="beats"/>
+        <event offset="48" counter="L2_EXT_W_STALL" title="External Bus Stalls" name="Write stall cycles" description="The number of cycles where a write is stalled waiting for the external bus." units="cycles"/>
+        <event offset="49" counter="L2_EXT_AW_CNT_Q1" title="External Bus Outstanding Writes" name="0-25% outstanding" description="The number of write transactions initiated when 0-25% of the maximum are in use." units="transactions"/>
+        <event offset="50" counter="L2_EXT_AW_CNT_Q2" title="External Bus Outstanding Writes" name="25-50% outstanding" description="The number of write transactions initiated when 25-50% of the maximum are in use." units="transactions"/>
+        <event offset="51" counter="L2_EXT_AW_CNT_Q3" title="External Bus Outstanding Writes" name="50-75% outstanding" description="The number of write transactions initiated when 50-75% of the maximum are in use." units="transactions"/>
+        <event offset="52" advanced="yes" counter="L2_EXT_SNOOP" title="External Bus Accesses" name="Snoop transactions" description="The number of coherency snoops triggered by external masters." units="transactions"/>
+        <event offset="53" advanced="yes" counter="L2_EXT_SNOOP_STALL" title="External Bus Stalls" name="Snoop stall cycles" description="The number of cycles where a coherency snoop triggered by external master is stalled." units="cycles"/>
+    </category>
+    <category name="Shader Core" per_cpu="no">
+        <event offset="4" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment active" description="The number of cycles where the shader core is processing a fragment workload." units="cycles"/>
+        <event offset="5" advanced="yes" counter="FRAG_PRIMITIVES" title="Core Primitives" name="Read primitives" description="The number of primitives read from the tile list by the fragment front-end." units="primitives"/>
+        <event offset="6" counter="FRAG_PRIM_RAST" title="Core Primitives" name="Rasterized primitives" description="The number of primitives being rasterized." units="primitives"/>
+        <event offset="7" counter="FRAG_FPK_ACTIVE" title="Core Cycles" name="Fragment FPKB active" description="The number of cycles where at least one quad is present in the pre-pipe quad queue." units="cycles"/>
+        <event offset="9" counter="FRAG_WARPS" title="Core Warps" name="Fragment warps" description="The number of fragment warps created." units="warps"/>
+        <event offset="10" counter="FRAG_PARTIAL_WARPS" title="Core Warps" name="Partial fragment warps" description="The number of fragment warps containing helper threads that do not correspond to a hit sample point." units="warps"/>
+        <event offset="11" counter="FRAG_QUADS_RAST" title="Core Quads" name="Rasterized quads" description="The number of quads generated by the rasterization phase." units="quads"/>
+        <event offset="12" counter="FRAG_QUADS_EZS_TEST" title="Core Quads" name="Early ZS tested quads" description="The number of quads that are undergoing early depth and stencil testing." units="quads"/>
+        <event offset="13" counter="FRAG_QUADS_EZS_UPDATE" title="Core Quads" name="Early ZS updated quads" description="The number of quads undergoing early depth and stencil testing, that are capable of updating the framebuffer." units="quads"/>
+        <event offset="14" counter="FRAG_QUADS_EZS_KILL" title="Core Quads" name="Early ZS killed quads" description="The number of quads killed by early depth and stencil testing." units="quads"/>
+        <event offset="15" counter="FRAG_LZS_TEST" title="Core Quads" name="Late ZS tested quads" description="The number of quads undergoing late depth and stencil testing." units="quads"/>
+        <event offset="16" counter="FRAG_LZS_KILL" title="Core Quads" name="Late ZS killed quads" description="The number of quads killed by late depth and stencil testing." units="quads"/>
+        <event offset="18" counter="FRAG_PTILES" title="Core Tiles" name="Tiles" description="The number of tiles processed by the shader core." units="tiles"/>
+        <event offset="19" counter="FRAG_TRANS_ELIM" title="Core Tiles" name="Constant tiles killed" description="The number of tiles killed by transaction elimination." units="tiles"/>
+        <event offset="20" counter="QUAD_FPK_KILLER" title="Core Quads" name="FPK occluder quads" description="The number of quads that are valid occluders for hidden surface removal." units="quads"/>
+        <event offset="22" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Non-fragment active" description="The number of cycles where the shader core is processing some non-fragment workload." units="cycles"/>
+        <event offset="23" advanced="yes" counter="COMPUTE_TASKS" title="Core Tasks" name="Non-fragment tasks" description="The number of non-fragment tasks issued to the shader core." units="tasks"/>
+        <event offset="24" counter="COMPUTE_WARPS" title="Core Warps" name="Non-fragment warps" description="The number of non-fragment warps created." units="warps"/>
+        <event offset="25" advanced="yes" counter="COMPUTE_STARVING" title="Core Starvation Cycles" name="Non-fragment starvation cycles" description="The number of cycles where the shader core is processing a non-fragment workload and there are no new threads available for execution." units="cycles"/>
+        <event offset="26" counter="EXEC_CORE_ACTIVE" title="Core Cycles" name="Execution core active" description="The number of cycles where the shader core is processing at least one warp." units="cycles"/>
+        <event offset="27" advanced="yes" counter="EXEC_ACTIVE" title="Core Cycles" name="Execution engine active" description="The number of cycles where the execution engine unit is processing at least one thread." units="cycles"/>
+        <event offset="28" counter="EXEC_INSTR_COUNT" title="Core EE Instructions" name="Executed instructions" description="The number of instructions executed per warp." units="instructions"/>
+        <event offset="29" counter="EXEC_INSTR_DIVERGED" title="Core EE Instructions" name="Diverged instructions" description="The number of instructions executed per warp, that have control flow divergence." units="instructions"/>
+        <event offset="30" advanced="yes" counter="EXEC_INSTR_STARVING" title="Core Starvation Cycles" name="Execution engine starvation cycles" description="The number of cycles where no new threads are available for execution." units="cycles"/>
+        <event offset="31" advanced="yes" counter="ARITH_INSTR_SINGLE_FMA" title="Core EE Instructions" name="Arithmetic instructions" description="The number of instructions where the workload is a single FMA pipe arithmetic operation." units="instructions"/>
+        <event offset="32" advanced="yes" counter="ARITH_INSTR_DOUBLE" title="Core EE Instructions" name="Dual Arithmetic instructions" description="The number of instructions where the workload is one FMA pipe arithmetic operation and one ADD pipe arithmetic operation." units="instructions"/>
+        <event offset="33" advanced="yes" counter="ARITH_INSTR_MSG" title="Core EE Instructions" name="Arithmetic + Message instructions" description="The number of instructions where the workload is one FMA pipe arithmetic operation and one ADD pipe message operation" units="instructions"/>
+        <event offset="34" advanced="yes" counter="ARITH_INSTR_MSG_ONLY" title="Core EE Instructions" name="Message instructions" description="The number of instructions where the workload is a single ADD pipe message operation, with no FMA pipe operation" units="instructions"/>
+        <event offset="35" counter="TEX_MSGI_NUM_QUADS" title="Core Texture Quads" name="Texture requests" description="The number of quad-width texture operations processed by the texture unit." units="quads"/>
+        <event offset="36" counter="TEX_DFCH_NUM_PASSES" title="Core Texture Quads" name="Texture issues" description="The number of quad-width filtering passes." units="issues"/>
+        <event offset="37" counter="TEX_DFCH_NUM_PASSES_MISS" title="Core Texture Quads" name="Descriptor misses" description="The number of quad-width filtering passes that miss in the resource or sampler descriptor cache." units="requests"/>
+        <event offset="38" counter="TEX_DFCH_NUM_PASSES_MIP_MAP" title="Core Texture Quads" name="Mipmapped texture issues" description="The number of quad-width filtering passes that use a mipmapped texture." units="issues"/>
+        <event offset="39" counter="TEX_TIDX_NUM_SPLIT_MIP_MAP" title="Core Texture Quads" name="Trilinear filtered issues" description="The number of quad-width filtering passes that use a trilinear filter." units="issues"/>
+        <event offset="40" counter="TEX_TFCH_NUM_LINES_FETCHED" title="Core Texture Line Fetches" name="Line fetches" description="The number of texture line fetches from the L2 cache." units="issues"/>
+        <event offset="41" counter="TEX_TFCH_NUM_LINES_FETCHED_BLOCK_COMPRESSED" title="Core Texture Line Fetches" name="Compressed line fetches" description="The number of texture line fetches from the L2 cache that are block compressed textures." units="issues"/>
+        <event offset="42" counter="TEX_TFCH_NUM_OPERATIONS" title="Core Texture Cycles" name="Cache lookups" description="The number of texture cache lookup cycles." units="requests"/>
+        <event offset="43" counter="TEX_FILT_NUM_OPERATIONS" title="Core Texture Cycles" name="Texturing active" description="The number of texture filtering issue cycles." units="cycles"/>
+        <event offset="44" counter="LS_MEM_READ_FULL" title="Core Load/Store Cycles" name="Full read cycles" description="The number of full-width load/store cache reads." units="cycles"/>
+        <event offset="45" counter="LS_MEM_READ_SHORT" title="Core Load/Store Cycles" name="Partial read cycles" description="The number of partial-width load/store cache reads." units="cycles"/>
+        <event offset="46" counter="LS_MEM_WRITE_FULL" title="Core Load/Store Cycles" name="Full write cycles" description="The number of full-width load/store cache writes." units="cycles"/>
+        <event offset="47" counter="LS_MEM_WRITE_SHORT" title="Core Load/Store Cycles" name="Partial write cycles" description="The number of partial-width load/store cache writes." units="cycles"/>
+        <event offset="48" counter="LS_MEM_ATOMIC" title="Core Load/Store Cycles" name="Atomic access cycles" description="The number of load/store atomic accesses." units="cycles"/>
+        <event offset="49" counter="VARY_INSTR" title="Core Varying Requests" name="Interpolation requests" description="The number of warp-width interpolation operations processed by the varying unit." units="instructions"/>
+        <event offset="50" counter="VARY_SLOT_32" title="Core Varying Cycles" name="32-bit interpolation active" description="The number of 32-bit interpolation cycles processed by the varying unit." units="cycles"/>
+        <event offset="51" counter="VARY_SLOT_16" title="Core Varying Cycles" name="16-bit interpolation active" description="The number of 16-bit interpolation cycles processed by the varying unit." units="cycles"/>
+        <event offset="52" advanced="yes" counter="ATTR_INSTR" title="Core Attribute Requests" name="Attribute requests" description="The number of instructions executed by the attribute unit." units="instructions"/>
+        <event offset="53" advanced="yes" counter="ARITH_INSTR_FP_MUL" title="Core EE Instructions" name="Multiplier instructions" description="The number of instructions where the workload uses floating-point multiplier hardware." units="instructions"/>
+        <event offset="54" counter="BEATS_RD_FTC" title="Core L2 Reads" name="Fragment L2 read beats" description="The number of read beats received by the fixed-function fragment front-end." units="beats"/>
+        <event offset="55" counter="BEATS_RD_FTC_EXT" title="Core L2 Reads" name="Fragment external read beats" description="The number of read beats received by the fixed-function fragment front-end that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="56" counter="BEATS_RD_LSC" title="Core L2 Reads" name="Load/store L2 read beats" description="The number of read beats received by the load/store unit." units="beats"/>
+        <event offset="57" counter="BEATS_RD_LSC_EXT" title="Core L2 Reads" name="Load/store external read beats" description="The number of read beats received by the load/store unit that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="58" counter="BEATS_RD_TEX" title="Core L2 Reads" name="Texture L2 read beats" description="The number of read beats received by the texture unit." units="beats"/>
+        <event offset="59" counter="BEATS_RD_TEX_EXT" title="Core L2 Reads" name="Texture external read beats" description="The number of read beats received by the texture unit that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="60" advanced="yes" counter="BEATS_RD_OTHER" title="Core L2 Reads" name="Other L2 read beats" description="The number of read beats received by a unit that is not specifically identified." units="beats"/>
+        <event offset="63" counter="BEATS_WR_LSC_WB" title="Core Writes" name="Load/store writeback write beats" description="The number of write beats by the load/store unit that are due to writeback." units="beats"/>
+        <event offset="62" counter="BEATS_WR_TIB" title="Core Writes" name="Tile buffer write beats" description="The number of write beats sent by the tile buffer writeback unit." units="beats"/>
+        <event offset="61" counter="BEATS_WR_LSC_OTHER" title="Core Writes" name="Load/store other write beats" description="The number of write beats by the load/store unit that are due to any reason other than writeback." units="beats"/>
     </category>
 </metrics>
\ No newline at end of file
diff --git a/src/panfrost/perf/G51.xml b/src/panfrost/perf/G51.xml
index 3bb25c724c2..cad669d3c3b 100644
--- a/src/panfrost/perf/G51.xml
+++ b/src/panfrost/perf/G51.xml
@@ -1,6 +1,6 @@
 <!--
 Copyright © 2017-2020 ARM Limited.
-Copyright © 2021 Collabora, Ltd.
+Copyright © 2021-2022 Collabora, Ltd.
 Author: Antonio Caggiano <antonio.caggiano@collabora.com>
 
 Permission is hereby granted, free of charge, to any person obtaining a
@@ -24,154 +24,154 @@ SOFTWARE.
 -->
 <metrics id="TSIx">
     <category name="Job Manager" per_cpu="no">
-        <event offset="6" counter="GPU_ACTIVE" title="GPU Cycles" name="GPU active" description="The number of cycles where the GPU has a workload of any type queued for processing." units="cycles" />
-        <event offset="7" advanced="yes" counter="IRQ_ACTIVE" title="GPU Cycles" name="Interrupt active" description="The number of cycles where the GPU has a pending interrupt." units="cycles" />
-        <event offset="8" advanced="yes" counter="JS0_JOBS" title="GPU Jobs" name="Fragment jobs" description="The number of jobs processed by the GPU fragment queue." units="jobs" />
-        <event offset="9" counter="JS0_TASKS" title="GPU Tasks" name="Fragment tasks" description="The number of 32x32 pixel tasks processed by the GPU fragment queue." units="tasks" />
-        <event offset="10" counter="JS0_ACTIVE" title="GPU Cycles" name="Fragment queue active" description="The number of cycles where work is queued for processing in the GPU fragment queue." units="cycles" />
-        <event offset="12" advanced="yes" counter="JS0_WAIT_READ" title="GPU Wait Cycles" name="Fragment descriptor reads cycles" description="The number of cycles where queued fragment work is waiting for a descriptor load." units="cycles" />
-        <event offset="13" advanced="yes" counter="JS0_WAIT_ISSUE" title="GPU Wait Cycles" name="Fragment job issue cycles" description="The number of cycles where queued fragment work is waiting for an available processor." units="cycles" />
-        <event offset="14" advanced="yes" counter="JS0_WAIT_DEPEND" title="GPU Wait Cycles" name="Fragment job dependency cycles" description="The number of cycles where queued fragment work is waiting for dependent work to complete." units="cycles" />
-        <event offset="15" advanced="yes" counter="JS0_WAIT_FINISH" title="GPU Wait Cycles" name="Fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued fragment work to complete." units="cycles" />
-        <event offset="16" advanced="yes" counter="JS1_JOBS" title="GPU Jobs" name="Non-fragment jobs" description="The number of jobs processed by the GPU non-fragment queue." units="jobs" />
-        <event offset="17" advanced="yes" counter="JS1_TASKS" title="GPU Tasks" name="Non-fragment tasks" description="The number of tasks processed by the GPU non-fragment queue." units="tasks" />
-        <event offset="18" counter="JS1_ACTIVE" title="GPU Cycles" name="Non-fragment queue active" description="The number of cycles where work is queued in the GPU non-fragment queue." units="cycles" />
-        <event offset="20" advanced="yes" counter="JS1_WAIT_READ" title="GPU Wait Cycles" name="Non-fragment descriptor read cycles" description="The number number of cycles where queued non-fragment work is waiting for a descriptor load." units="cycles" />
-        <event offset="21" advanced="yes" counter="JS1_WAIT_ISSUE" title="GPU Wait Cycles" name="Non-fragment job issue cycles" description="The number of cycles where queued non-fragment work is waiting for an available processor." units="cycles" />
-        <event offset="22" advanced="yes" counter="JS1_WAIT_DEPEND" title="GPU Wait Cycles" name="Non-fragment job dependency cycles" description="The number of cycles where queued non-fragment work is waiting for dependent work to complete." units="cycles" />
-        <event offset="23" advanced="yes" counter="JS1_WAIT_FINISH" title="GPU Wait Cycles" name="Non-fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued non-fragment work to complete." units="cycles" />
-        <event offset="24" advanced="yes" counter="JS2_JOBS" title="GPU Jobs" name="Reserved jobs" description="The number of jobs processed by the GPU reserved queue." units="jobs" />
-        <event offset="25" advanced="yes" counter="JS2_TASKS" title="GPU Tasks" name="Reserved tasks" description="The number of tasks processed by the GPU reserved queue." units="tasks" />
-        <event offset="26" advanced="yes" counter="JS2_ACTIVE" title="GPU Cycles" name="Reserved queue active" description="The number of cycles where work is queued in the GPU reserved queue." units="cycles" />
-        <event offset="28" advanced="yes" counter="JS2_WAIT_READ" title="GPU Wait Cycles" name="Reserved descriptor read cycles" description="The number of cycles where queued reserved work is waiting for a descriptor load." units="cycles" />
-        <event offset="29" advanced="yes" counter="JS2_WAIT_ISSUE" title="GPU Wait Cycles" name="Reserved job issue cycles" description="The number of cycles where queued reserved work is waiting for an available processor." units="cycles" />
-        <event offset="30" advanced="yes" counter="JS2_WAIT_DEPEND" title="GPU Wait Cycles" name="Reserved job dependency cycles" description="The number of cycles where queued reserved work is waiting for dependent work to complete." units="cycles" />
-        <event offset="31" advanced="yes" counter="JS2_WAIT_FINISH" title="GPU Wait Cycles" name="Reserved job finish cycles" description="The number of cycles where the GPU is waiting for issued reserved work to complete." units="cycles" />
+        <event offset="6" counter="GPU_ACTIVE" title="GPU Cycles" name="GPU active" description="The number of cycles where the GPU has a workload of any type queued for processing." units="cycles"/>
+        <event offset="7" advanced="yes" counter="IRQ_ACTIVE" title="GPU Cycles" name="Interrupt active" description="The number of cycles where the GPU has a pending interrupt." units="cycles"/>
+        <event offset="8" advanced="yes" counter="JS0_JOBS" title="GPU Jobs" name="Fragment jobs" description="The number of jobs processed by the GPU fragment queue." units="jobs"/>
+        <event offset="9" counter="JS0_TASKS" title="GPU Tasks" name="Fragment tasks" description="The number of 32x32 pixel tasks processed by the GPU fragment queue." units="tasks"/>
+        <event offset="10" counter="JS0_ACTIVE" title="GPU Cycles" name="Fragment queue active" description="The number of cycles where work is queued for processing in the GPU fragment queue." units="cycles"/>
+        <event offset="12" advanced="yes" counter="JS0_WAIT_READ" title="GPU Wait Cycles" name="Fragment descriptor reads cycles" description="The number of cycles where queued fragment work is waiting for a descriptor load." units="cycles"/>
+        <event offset="13" advanced="yes" counter="JS0_WAIT_ISSUE" title="GPU Wait Cycles" name="Fragment job issue cycles" description="The number of cycles where queued fragment work is waiting for an available processor." units="cycles"/>
+        <event offset="14" advanced="yes" counter="JS0_WAIT_DEPEND" title="GPU Wait Cycles" name="Fragment job dependency cycles" description="The number of cycles where queued fragment work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="15" advanced="yes" counter="JS0_WAIT_FINISH" title="GPU Wait Cycles" name="Fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued fragment work to complete." units="cycles"/>
+        <event offset="16" advanced="yes" counter="JS1_JOBS" title="GPU Jobs" name="Non-fragment jobs" description="The number of jobs processed by the GPU non-fragment queue." units="jobs"/>
+        <event offset="17" advanced="yes" counter="JS1_TASKS" title="GPU Tasks" name="Non-fragment tasks" description="The number of tasks processed by the GPU non-fragment queue." units="tasks"/>
+        <event offset="18" counter="JS1_ACTIVE" title="GPU Cycles" name="Non-fragment queue active" description="The number of cycles where work is queued in the GPU non-fragment queue." units="cycles"/>
+        <event offset="20" advanced="yes" counter="JS1_WAIT_READ" title="GPU Wait Cycles" name="Non-fragment descriptor read cycles" description="The number number of cycles where queued non-fragment work is waiting for a descriptor load." units="cycles"/>
+        <event offset="21" advanced="yes" counter="JS1_WAIT_ISSUE" title="GPU Wait Cycles" name="Non-fragment job issue cycles" description="The number of cycles where queued non-fragment work is waiting for an available processor." units="cycles"/>
+        <event offset="22" advanced="yes" counter="JS1_WAIT_DEPEND" title="GPU Wait Cycles" name="Non-fragment job dependency cycles" description="The number of cycles where queued non-fragment work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="23" advanced="yes" counter="JS1_WAIT_FINISH" title="GPU Wait Cycles" name="Non-fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued non-fragment work to complete." units="cycles"/>
+        <event offset="24" advanced="yes" counter="JS2_JOBS" title="GPU Jobs" name="Reserved jobs" description="The number of jobs processed by the GPU reserved queue." units="jobs"/>
+        <event offset="25" advanced="yes" counter="JS2_TASKS" title="GPU Tasks" name="Reserved tasks" description="The number of tasks processed by the GPU reserved queue." units="tasks"/>
+        <event offset="26" advanced="yes" counter="JS2_ACTIVE" title="GPU Cycles" name="Reserved queue active" description="The number of cycles where work is queued in the GPU reserved queue." units="cycles"/>
+        <event offset="28" advanced="yes" counter="JS2_WAIT_READ" title="GPU Wait Cycles" name="Reserved descriptor read cycles" description="The number of cycles where queued reserved work is waiting for a descriptor load." units="cycles"/>
+        <event offset="29" advanced="yes" counter="JS2_WAIT_ISSUE" title="GPU Wait Cycles" name="Reserved job issue cycles" description="The number of cycles where queued reserved work is waiting for an available processor." units="cycles"/>
+        <event offset="30" advanced="yes" counter="JS2_WAIT_DEPEND" title="GPU Wait Cycles" name="Reserved job dependency cycles" description="The number of cycles where queued reserved work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="31" advanced="yes" counter="JS2_WAIT_FINISH" title="GPU Wait Cycles" name="Reserved job finish cycles" description="The number of cycles where the GPU is waiting for issued reserved work to complete." units="cycles"/>
     </category>
     <category name="Tiler" per_cpu="no">
-        <event offset="68" counter="TILER_ACTIVE" title="GPU Cycles" name="Tiler active" description="The number of cycles where the tiler has a workload queued for processing." units="cycles" />
-        <event offset="70" counter="TRIANGLES" title="Input Primitives" name="Triangle primitives" description="The number of input triangle primitives." units="primitives" />
-        <event offset="71" counter="LINES" title="Input Primitives" name="Line primitives" description="The number of input line primitives." units="primitives" />
-        <event offset="72" counter="POINTS" title="Input Primitives" name="Point primitives" description="The number of input point primitives." units="primitives" />
-        <event offset="73" counter="FRONT_FACING" title="Visible Primitives" name="Front-facing primitives" description="The number of front-facing triangles that are visible after culling." units="primitives" />
-        <event offset="74" counter="BACK_FACING" title="Visible Primitives" name="Back-facing primitives" description="The number of back-facing triangles that are visible after culling." units="primitives" />
-        <event offset="75" counter="PRIM_VISIBLE" title="Primitive Culling" name="Visible primitives" description="The number of primitives that are visible after culling." units="primitives" />
-        <event offset="76" counter="PRIM_CULLED" title="Primitive Culling" name="Facing and XY plane test culled primitives" description="The number of primitives that are culled by facing or frustum XY plane tests." units="primitives" />
-        <event offset="77" counter="PRIM_CLIPPED" title="Primitive Culling" name="Z plane test culled primitives" description="The number of primitives that are culled by frustum Z plane tests." units="primitives" />
-        <event offset="78" counter="PRIM_SAT_CULLED" title="Primitive Culling" name="Sample test culled primitives" description="The number of primitives culled by the sample coverage test." units="primitives" />
-        <event offset="81" advanced="yes" counter="BUS_READ" title="Tiler L2 Accesses" name="Read beats" description="The number of internal bus data read cycles made by the tiler." units="beats" />
-        <event offset="83" advanced="yes" counter="BUS_WRITE" title="Tiler L2 Accesses" name="Write beats" description="The number of internal bus data write cycles made by the tiler." units="beats" />
-        <event offset="85" counter="IDVS_POS_SHAD_REQ" title="Tiler Shading Requests" name="Position shading requests" description="The number of position shading requests in the IDVS flow." units="requests" />
-        <event offset="87" advanced="yes" counter="IDVS_POS_SHAD_STALL" title="Tiler Cycles" name="Position shading stall cycles" description="The number of cycles where the tiler has a stalled position shading request." units="cycles" />
-        <event offset="88" advanced="yes" counter="IDVS_POS_FIFO_FULL" title="Tiler Cycles" name="Position FIFO full cycles" description="The number of cycles where the tiler has a stalled position shading buffer." units="cycles" />
-        <event offset="90" advanced="yes" counter="VCACHE_HIT" title="Tiler Vertex Cache" name="Position cache hits" description="The number of position lookups that result in a hit in the vertex cache." units="requests" />
-        <event offset="91" advanced="yes" counter="VCACHE_MISS" title="Tiler Vertex Cache" name="Position cache misses" description="The number of position lookups that miss in the vertex cache." units="requests" />
-        <event offset="95" advanced="yes" counter="VFETCH_STALL" title="Tiler Cycles" name="Primitive assembly busy stall cycles" description="The number of cycles where the tiler is stalled waiting for primitive assembly." units="cycles" />
-        <event offset="98" advanced="yes" counter="IDVS_VBU_HIT" title="Tiler Vertex Cache" name="Varying cache hits" description="The number of varying lookups that result in a hit in the vertex cache." units="requests" />
-        <event offset="99" advanced="yes" counter="IDVS_VBU_MISS" title="Tiler Vertex Cache" name="Varying cache misses" description="The number of varying lookups that miss in the vertex cache." units="requests" />
-        <event offset="101" counter="IDVS_VAR_SHAD_REQ" title="Tiler Shading Requests" name="Varying shading requests" description="The number of varying shading requests in the IDVS flow." units="requests" />
-        <event offset="102" advanced="yes" counter="IDVS_VAR_SHAD_STALL" title="Tiler Cycles" name="Varying shading stall cycles" description="The number of cycles where the tiler has a stalled varying shading request." units="cycles" />
-        <event offset="118" advanced="yes" counter="WRBUF_NO_AXI_ID_STALL" title="Tiler Cycles" name="Write buffer transaction stall cycles" description="The number of cycles where the tiler write buffer can not send data because it has no available write IDs." units="cycles" />
-        <event offset="119" advanced="yes" counter="WRBUF_AXI_STALL" title="Tiler Cycles" name="Write buffer write stall cycles" description="The number of cycles where the tiler write buffer can not send data because the bus is not ready." units="cycles" />
-    </category>
-    <category name="Shader Core" per_cpu="no">
-        <event offset="196" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment active" description="The number of cycles where the shader core is processing a fragment workload." units="cycles" />
-        <event offset="197" advanced="yes" counter="FRAG_PRIMITIVES" title="Core Primitives" name="Read primitives" description="The number of primitives read from the tile list by the fragment front-end." units="primitives" />
-        <event offset="198" counter="FRAG_PRIM_RAST" title="Core Primitives" name="Rasterized primitives" description="The number of primitives being rasterized." units="primitives" />
-        <event offset="199" counter="FRAG_FPK_ACTIVE" title="Core Cycles" name="Fragment FPKB active" description="The number of cycles where at least one quad is present in the pre-pipe quad queue." units="cycles" />
-        <event offset="201" counter="FRAG_WARPS" title="Core Warps" name="Fragment warps" description="The number of fragment warps created." units="warps" />
-        <event offset="202" counter="FRAG_PARTIAL_WARPS" title="Core Warps" name="Partial fragment warps" description="The number of fragment warps containing helper threads that do not correspond to a hit sample point." units="warps" />
-        <event offset="203" counter="FRAG_QUADS_RAST" title="Core Quads" name="Rasterized quads" description="The number of quads generated by the rasterization phase." units="quads" />
-        <event offset="204" counter="FRAG_QUADS_EZS_TEST" title="Core Quads" name="Early ZS tested quads" description="The number of quads that are undergoing early depth and stencil testing." units="quads" />
-        <event offset="205" counter="FRAG_QUADS_EZS_UPDATE" title="Core Quads" name="Early ZS updated quads" description="The number of quads undergoing early depth and stencil testing, that are capable of updating the framebuffer." units="quads" />
-        <event offset="206" counter="FRAG_QUADS_EZS_KILL" title="Core Quads" name="Early ZS killed quads" description="The number of quads killed by early depth and stencil testing." units="quads" />
-        <event offset="207" counter="FRAG_LZS_TEST" title="Core Quads" name="Late ZS tested quads" description="The number of quads undergoing late depth and stencil testing." units="quads" />
-        <event offset="208" counter="FRAG_LZS_KILL" title="Core Quads" name="Late ZS killed quads" description="The number of quads killed by late depth and stencil testing." units="quads" />
-        <event offset="210" counter="FRAG_PTILES" title="Core Tiles" name="Tiles" description="The number of tiles processed by the shader core." units="tiles" />
-        <event offset="211" counter="FRAG_TRANS_ELIM" title="Core Tiles" name="Constant tiles killed" description="The number of tiles killed by transaction elimination." units="tiles" />
-        <event offset="212" counter="QUAD_FPK_KILLER" title="Core Quads" name="FPK occluder quads" description="The number of quads that are valid occluders for hidden surface removal." units="quads" />
-        <event offset="214" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Non-fragment active" description="The number of cycles where the shader core is processing some non-fragment workload." units="cycles" />
-        <event offset="215" advanced="yes" counter="COMPUTE_TASKS" title="Core Tasks" name="Non-fragment tasks" description="The number of non-fragment tasks issued to the shader core." units="tasks" />
-        <event offset="216" counter="COMPUTE_WARPS" title="Core Warps" name="Non-fragment warps" description="The number of non-fragment warps created." units="warps" />
-        <event offset="217" advanced="yes" counter="COMPUTE_STARVING" title="Core Starvation Cycles" name="Non-fragment starvation cycles" description="The number of cycles where the shader core is processing a non-fragment workload and there are no new threads available for execution." units="cycles" />
-        <event offset="218" counter="EXEC_CORE_ACTIVE" title="Core Cycles" name="Execution core active" description="The number of cycles where the shader core is processing at least one warp." units="cycles" />
-        <event offset="219" advanced="yes" counter="EXEC_ACTIVE" title="Core Cycles" name="Execution engine active" description="The number of cycles where the execution engine unit is processing at least one thread." units="cycles" />
-        <event offset="220" counter="EXEC_INSTR_COUNT" title="Core EE Instructions" name="Executed instructions" description="The number of instructions executed per warp." units="instructions" />
-        <event offset="221" counter="EXEC_INSTR_DIVERGED" title="Core EE Instructions" name="Diverged instructions" description="The number of instructions executed per warp, that have control flow divergence." units="instructions" />
-        <event offset="222" advanced="yes" counter="EXEC_INSTR_STARVING" title="Core Starvation Cycles" name="Execution engine starvation cycles" description="The number of cycles where no new threads are available for execution." units="cycles" />
-        <event offset="223" advanced="yes" counter="ARITH_INSTR_SINGLE_FMA" title="Core EE Instructions" name="Arithmetic instructions" description="The number of instructions where the workload is a single FMA pipe arithmetic operation." units="instructions" />
-        <event offset="224" advanced="yes" counter="ARITH_INSTR_DOUBLE" title="Core EE Instructions" name="Dual Arithmetic instructions" description="The number of instructions where the workload is one FMA pipe arithmetic operation and one ADD pipe arithmetic operation." units="instructions" />
-        <event offset="225" advanced="yes" counter="ARITH_INSTR_MSG" title="Core EE Instructions" name="Arithmetic + Message instructions" description="The number of instructions where the workload is one FMA pipe arithmetic operation and one ADD pipe message operation" units="instructions" />
-        <event offset="226" advanced="yes" counter="ARITH_INSTR_MSG_ONLY" title="Core EE Instructions" name="Message instructions" description="The number of instructions where the workload is a single ADD pipe message operation, with no FMA pipe operation" units="instructions" />
-        <event offset="227" counter="TEX_MSGI_NUM_QUADS" title="Core Texture Quads" name="Texture requests" description="The number of quad-width texture operations processed by the texture unit." units="quads" />
-        <event offset="228" counter="TEX_DFCH_NUM_PASSES" title="Core Texture Quads" name="Texture issues" description="The number of quad-width filtering passes." units="issues" />
-        <event offset="229" counter="TEX_DFCH_NUM_PASSES_MISS" title="Core Texture Quads" name="Descriptor misses" description="The number of quad-width filtering passes that miss in the resource or sampler descriptor cache." units="requests" />
-        <event offset="230" counter="TEX_DFCH_NUM_PASSES_MIP_MAP" title="Core Texture Quads" name="Mipmapped texture issues" description="The number of quad-width filtering passes that use a mipmapped texture." units="issues" />
-        <event offset="231" counter="TEX_TIDX_NUM_SPLIT_MIP_MAP" title="Core Texture Quads" name="Trilinear filtered issues" description="The number of quad-width filtering passes that use a trilinear filter." units="issues" />
-        <event offset="232" counter="TEX_TFCH_NUM_LINES_FETCHED" title="Core Texture Line Fetches" name="Line fetches" description="The number of texture line fetches from the L2 cache." units="issues" />
-        <event offset="233" counter="TEX_TFCH_NUM_LINES_FETCHED_BLOCK_COMPRESSED" title="Core Texture Line Fetches" name="Compressed line fetches" description="The number of texture line fetches from the L2 cache that are block compressed textures." units="issues" />
-        <event offset="234" counter="TEX_TFCH_NUM_OPERATIONS" title="Core Texture Cycles" name="Cache lookups" description="The number of texture cache lookup cycles." units="requests" />
-        <event offset="235" counter="TEX_FILT_NUM_OPERATIONS" title="Core Texture Cycles" name="Texturing active" description="The number of texture filtering issue cycles." units="cycles" />
-        <event offset="236" counter="LS_MEM_READ_FULL" title="Core Load/Store Cycles" name="Full read cycles" description="The number of full-width load/store cache reads." units="cycles" />
-        <event offset="237" counter="LS_MEM_READ_SHORT" title="Core Load/Store Cycles" name="Partial read cycles" description="The number of partial-width load/store cache reads." units="cycles" />
-        <event offset="238" counter="LS_MEM_WRITE_FULL" title="Core Load/Store Cycles" name="Full write cycles" description="The number of full-width load/store cache writes." units="cycles" />
-        <event offset="239" counter="LS_MEM_WRITE_SHORT" title="Core Load/Store Cycles" name="Partial write cycles" description="The number of partial-width load/store cache writes." units="cycles" />
-        <event offset="240" counter="LS_MEM_ATOMIC" title="Core Load/Store Cycles" name="Atomic access cycles" description="The number of load/store atomic accesses." units="cycles" />
-        <event offset="241" counter="VARY_INSTR" title="Core Varying Requests" name="Interpolation requests" description="The number of warp-width interpolation operations processed by the varying unit." units="instructions" />
-        <event offset="242" counter="VARY_SLOT_32" title="Core Varying Cycles" name="32-bit interpolation active" description="The number of 32-bit interpolation cycles processed by the varying unit." units="cycles" />
-        <event offset="243" counter="VARY_SLOT_16" title="Core Varying Cycles" name="16-bit interpolation active" description="The number of 16-bit interpolation cycles processed by the varying unit." units="cycles" />
-        <event offset="244" advanced="yes" counter="ATTR_INSTR" title="Core Attribute Requests" name="Attribute requests" description="The number of instructions executed by the attribute unit." units="instructions" />
-        <event offset="245" advanced="yes" counter="ARITH_INSTR_FP_MUL" title="Core EE Instructions" name="Multiplier instructions" description="The number of instructions where the workload uses floating-point multiplier hardware." units="instructions" />
-        <event offset="246" counter="BEATS_RD_FTC" title="Core L2 Reads" name="Fragment L2 read beats" description="The number of read beats received by the fixed-function fragment front-end." units="beats" />
-        <event offset="247" counter="BEATS_RD_FTC_EXT" title="Core L2 Reads" name="Fragment external read beats" description="The number of read beats received by the fixed-function fragment front-end that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="248" counter="BEATS_RD_LSC" title="Core L2 Reads" name="Load/store L2 read beats" description="The number of read beats received by the load/store unit." units="beats" />
-        <event offset="249" counter="BEATS_RD_LSC_EXT" title="Core L2 Reads" name="Load/store external read beats" description="The number of read beats received by the load/store unit that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="250" counter="BEATS_RD_TEX" title="Core L2 Reads" name="Texture L2 read beats" description="The number of read beats received by the texture unit." units="beats" />
-        <event offset="251" counter="BEATS_RD_TEX_EXT" title="Core L2 Reads" name="Texture external read beats" description="The number of read beats received by the texture unit that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="252" advanced="yes" counter="BEATS_RD_OTHER" title="Core L2 Reads" name="Other L2 read beats" description="The number of read beats received by a unit that is not specifically identified." units="beats" />
-        <event offset="253" counter="BEATS_WR_LSC_OTHER" title="Core Writes" name="Load/store other write beats" description="The number of write beats by the load/store unit that are due to any reason other than writeback." units="beats" />
-        <event offset="254" counter="BEATS_WR_TIB" title="Core Writes" name="Tile buffer write beats" description="The number of write beats sent by the tile buffer writeback unit." units="beats" />
-        <event offset="255" counter="BEATS_WR_LSC_WB" title="Core Writes" name="Load/store writeback write beats" description="The number of write beats by the load/store unit that are due to writeback." units="beats" />
+        <event offset="4" counter="TILER_ACTIVE" title="GPU Cycles" name="Tiler active" description="The number of cycles where the tiler has a workload queued for processing." units="cycles"/>
+        <event offset="6" counter="TRIANGLES" title="Input Primitives" name="Triangle primitives" description="The number of input triangle primitives." units="primitives"/>
+        <event offset="7" counter="LINES" title="Input Primitives" name="Line primitives" description="The number of input line primitives." units="primitives"/>
+        <event offset="8" counter="POINTS" title="Input Primitives" name="Point primitives" description="The number of input point primitives." units="primitives"/>
+        <event offset="9" counter="FRONT_FACING" title="Visible Primitives" name="Front-facing primitives" description="The number of front-facing triangles that are visible after culling." units="primitives"/>
+        <event offset="10" counter="BACK_FACING" title="Visible Primitives" name="Back-facing primitives" description="The number of back-facing triangles that are visible after culling." units="primitives"/>
+        <event offset="11" counter="PRIM_VISIBLE" title="Primitive Culling" name="Visible primitives" description="The number of primitives that are visible after culling." units="primitives"/>
+        <event offset="12" counter="PRIM_CULLED" title="Primitive Culling" name="Facing and XY plane test culled primitives" description="The number of primitives that are culled by facing or frustum XY plane tests." units="primitives"/>
+        <event offset="13" counter="PRIM_CLIPPED" title="Primitive Culling" name="Z plane test culled primitives" description="The number of primitives that are culled by frustum Z plane tests." units="primitives"/>
+        <event offset="14" counter="PRIM_SAT_CULLED" title="Primitive Culling" name="Sample test culled primitives" description="The number of primitives culled by the sample coverage test." units="primitives"/>
+        <event offset="17" advanced="yes" counter="BUS_READ" title="Tiler L2 Accesses" name="Read beats" description="The number of internal bus data read cycles made by the tiler." units="beats"/>
+        <event offset="19" advanced="yes" counter="BUS_WRITE" title="Tiler L2 Accesses" name="Write beats" description="The number of internal bus data write cycles made by the tiler." units="beats"/>
+        <event offset="21" counter="IDVS_POS_SHAD_REQ" title="Tiler Shading Requests" name="Position shading requests" description="The number of position shading requests in the IDVS flow." units="requests"/>
+        <event offset="23" advanced="yes" counter="IDVS_POS_SHAD_STALL" title="Tiler Cycles" name="Position shading stall cycles" description="The number of cycles where the tiler has a stalled position shading request." units="cycles"/>
+        <event offset="24" advanced="yes" counter="IDVS_POS_FIFO_FULL" title="Tiler Cycles" name="Position FIFO full cycles" description="The number of cycles where the tiler has a stalled position shading buffer." units="cycles"/>
+        <event offset="26" advanced="yes" counter="VCACHE_HIT" title="Tiler Vertex Cache" name="Position cache hits" description="The number of position lookups that result in a hit in the vertex cache." units="requests"/>
+        <event offset="27" advanced="yes" counter="VCACHE_MISS" title="Tiler Vertex Cache" name="Position cache misses" description="The number of position lookups that miss in the vertex cache." units="requests"/>
+        <event offset="31" advanced="yes" counter="VFETCH_STALL" title="Tiler Cycles" name="Primitive assembly busy stall cycles" description="The number of cycles where the tiler is stalled waiting for primitive assembly." units="cycles"/>
+        <event offset="34" advanced="yes" counter="IDVS_VBU_HIT" title="Tiler Vertex Cache" name="Varying cache hits" description="The number of varying lookups that result in a hit in the vertex cache." units="requests"/>
+        <event offset="35" advanced="yes" counter="IDVS_VBU_MISS" title="Tiler Vertex Cache" name="Varying cache misses" description="The number of varying lookups that miss in the vertex cache." units="requests"/>
+        <event offset="37" counter="IDVS_VAR_SHAD_REQ" title="Tiler Shading Requests" name="Varying shading requests" description="The number of varying shading requests in the IDVS flow." units="requests"/>
+        <event offset="38" advanced="yes" counter="IDVS_VAR_SHAD_STALL" title="Tiler Cycles" name="Varying shading stall cycles" description="The number of cycles where the tiler has a stalled varying shading request." units="cycles"/>
+        <event offset="54" advanced="yes" counter="WRBUF_NO_AXI_ID_STALL" title="Tiler Cycles" name="Write buffer transaction stall cycles" description="The number of cycles where the tiler write buffer can not send data because it has no available write IDs." units="cycles"/>
+        <event offset="55" advanced="yes" counter="WRBUF_AXI_STALL" title="Tiler Cycles" name="Write buffer write stall cycles" description="The number of cycles where the tiler write buffer can not send data because the bus is not ready." units="cycles"/>
     </category>
     <category name="Memory System" per_cpu="no">
-        <event offset="132" advanced="yes" counter="MMU_REQUESTS" title="MMU Stage 1 Translations" name="MMU lookups" description="The number of main MMU address translations performed." units="requests" />
-        <event offset="144" advanced="yes" counter="L2_RD_MSG_IN" title="L2 Cache Requests" name="Read requests" description="The number of L2 cache read requests from internal masters." units="requests" />
-        <event offset="145" advanced="yes" counter="L2_RD_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Read stall cycles" description="The number of cycles L2 cache read requests from internal masters are stalled." units="cycles" />
-        <event offset="146" advanced="yes" counter="L2_WR_MSG_IN" title="L2 Cache Requests" name="Write requests" description="The number of L2 cache write requests from internal masters." units="requests" />
-        <event offset="147" advanced="yes" counter="L2_WR_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Write stall cycles" description="The number of cycles where L2 cache write requests from internal masters are stalled." units="cycles" />
-        <event offset="148" advanced="yes" counter="L2_SNP_MSG_IN" title="L2 Cache Requests" name="Snoop requests" description="The number of L2 snoop requests from internal masters." units="requests" />
-        <event offset="149" advanced="yes" counter="L2_SNP_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Snoop stall cycles" description="The number of cycles where L2 cache snoop requests from internal masters are stalled." units="cycles" />
-        <event offset="150" advanced="yes" counter="L2_RD_MSG_OUT" title="L2 Cache Requests" name="L1 read requests" description="The number of L1 cache read requests sent by the L2 cache to an internal master." units="requests" />
-        <event offset="151" advanced="yes" counter="L2_RD_MSG_OUT_STALL" title="L2 Cache Stall Cycles" name="L1 read stall cycles" description="The number of cycles where L1 cache read requests sent by the L2 cache to an internal master are stalled." units="cycles" />
-        <event offset="152" advanced="yes" counter="L2_WR_MSG_OUT" title="L2 Cache Requests" name="L1 write requests" description="The number of L1 cache write responses sent by the L2 cache to an internal master." units="requests" />
-        <event offset="153" counter="L2_ANY_LOOKUP" title="L2 Cache Lookups" name="Any lookup" description="The number of L2 cache lookups performed." units="requests" />
-        <event offset="154" counter="L2_READ_LOOKUP" title="L2 Cache Lookups" name="Read lookup" description="The number of L2 cache read lookups performed." units="requests" />
-        <event offset="155" counter="L2_WRITE_LOOKUP" title="L2 Cache Lookups" name="Write lookup" description="The number of L2 cache write lookups performed." units="requests" />
-        <event offset="156" advanced="yes" counter="L2_EXT_SNOOP_LOOKUP" title="L2 Cache Lookups" name="External snoop lookups" description="The number of coherency snoop lookups performed that were triggered by an external master." units="requests" />
-        <event offset="157" counter="L2_EXT_READ" title="External Bus Accesses" name="Read transaction" description="The number of external read transactions." units="transactions" />
-        <event offset="158" advanced="yes" counter="L2_EXT_READ_NOSNP" title="External Bus Accesses" name="ReadNoSnoop transactions" description="The number of external non-coherent read transactions." units="transactions" />
-        <event offset="159" advanced="yes" counter="L2_EXT_READ_UNIQUE" title="External Bus Accesses" name="ReadUnique transactions" description="The number of external coherent read unique transactions." units="transactions" />
-        <event offset="160" counter="L2_EXT_READ_BEATS" title="External Bus Beats" name="Read beat" description="The number of external bus data read cycles." units="beats" />
-        <event offset="161" counter="L2_EXT_AR_STALL" title="External Bus Stalls" name="Read stall cycles" description="The number of cycles where a read is stalled waiting for the external bus." units="cycles" />
-        <event offset="162" counter="L2_EXT_AR_CNT_Q1" title="External Bus Outstanding Reads" name="0-25% outstanding" description="The number of read transactions initiated when 0-25% of the maximum are in use." units="transactions" />
-        <event offset="163" counter="L2_EXT_AR_CNT_Q2" title="External Bus Outstanding Reads" name="25-50% outstanding" description="The number of read transactions initiated when 25-50% of the maximum are in use." units="transactions" />
-        <event offset="164" counter="L2_EXT_AR_CNT_Q3" title="External Bus Outstanding Reads" name="50-75% outstanding" description="The number of read transactions initiated when 50-75% of the maximum are in use." units="transactions" />
-        <event offset="165" counter="L2_EXT_RRESP_0_127" title="External Bus Read Latency" name="0-127 cycles" description="The number of data beats returned 0-127 cycles after the read request." units="beats" />
-        <event offset="166" counter="L2_EXT_RRESP_128_191" title="External Bus Read Latency" name="128-191 cycles" description="The number of data beats returned 128-191 cycles after the read request." units="beats" />
-        <event offset="167" counter="L2_EXT_RRESP_192_255" title="External Bus Read Latency" name="192-255 cycles" description="The number of data beats returned 192-255 cycles after the read request." units="beats" />
-        <event offset="168" counter="L2_EXT_RRESP_256_319" title="External Bus Read Latency" name="256-319 cycles" description="The number of data beats returned 256-319 cycles after the read request." units="beats" />
-        <event offset="169" counter="L2_EXT_RRESP_320_383" title="External Bus Read Latency" name="320-383 cycles" description="The number of data beats returned 320-383 cycles after the read request." units="beats" />
-        <event offset="170" counter="L2_EXT_WRITE" title="External Bus Accesses" name="Write transaction" description="The number of external write transactions." units="transactions" />
-        <event offset="171" advanced="yes" counter="L2_EXT_WRITE_NOSNP_FULL" title="External Bus Accesses" name="WriteNoSnoopFull transactions" description="The number of external non-coherent full write transactions." units="transactions" />
-        <event offset="172" advanced="yes" counter="L2_EXT_WRITE_NOSNP_PTL" title="External Bus Accesses" name="WriteNoSnoopPartial transactions" description="The number of external non-coherent partial write transactions." units="transactions" />
-        <event offset="173" advanced="yes" counter="L2_EXT_WRITE_SNP_FULL" title="External Bus Accesses" name="WriteSnoopFull transactions" description="The number of external coherent full write transactions." units="transactions" />
-        <event offset="174" advanced="yes" counter="L2_EXT_WRITE_SNP_PTL" title="External Bus Accesses" name="WriteSnoopPartial transactions" description="The number of external coherent partial write transactions." units="transactions" />
-        <event offset="175" counter="L2_EXT_WRITE_BEATS" title="External Bus Beats" name="Write beat" description="The number of external bus data write cycles." units="beats" />
-        <event offset="176" counter="L2_EXT_W_STALL" title="External Bus Stalls" name="Write stall cycles" description="The number of cycles where a write is stalled waiting for the external bus." units="cycles" />
-        <event offset="177" counter="L2_EXT_AW_CNT_Q1" title="External Bus Outstanding Writes" name="0-25% outstanding" description="The number of write transactions initiated when 0-25% of the maximum are in use." units="transactions" />
-        <event offset="178" counter="L2_EXT_AW_CNT_Q2" title="External Bus Outstanding Writes" name="25-50% outstanding" description="The number of write transactions initiated when 25-50% of the maximum are in use." units="transactions" />
-        <event offset="179" counter="L2_EXT_AW_CNT_Q3" title="External Bus Outstanding Writes" name="50-75% outstanding" description="The number of write transactions initiated when 50-75% of the maximum are in use." units="transactions" />
-        <event offset="180" advanced="yes" counter="L2_EXT_SNOOP" title="External Bus Accesses" name="Snoop transactions" description="The number of coherency snoops triggered by external masters." units="transactions" />
-        <event offset="181" advanced="yes" counter="L2_EXT_SNOOP_STALL" title="External Bus Stalls" name="Snoop stall cycles" description="The number of cycles where a coherency snoop triggered by external master is stalled." units="cycles" />
+        <event offset="4" advanced="yes" counter="MMU_REQUESTS" title="MMU Stage 1 Translations" name="MMU lookups" description="The number of main MMU address translations performed." units="requests"/>
+        <event offset="16" advanced="yes" counter="L2_RD_MSG_IN" title="L2 Cache Requests" name="Read requests" description="The number of L2 cache read requests from internal masters." units="requests"/>
+        <event offset="17" advanced="yes" counter="L2_RD_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Read stall cycles" description="The number of cycles L2 cache read requests from internal masters are stalled." units="cycles"/>
+        <event offset="18" advanced="yes" counter="L2_WR_MSG_IN" title="L2 Cache Requests" name="Write requests" description="The number of L2 cache write requests from internal masters." units="requests"/>
+        <event offset="19" advanced="yes" counter="L2_WR_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Write stall cycles" description="The number of cycles where L2 cache write requests from internal masters are stalled." units="cycles"/>
+        <event offset="20" advanced="yes" counter="L2_SNP_MSG_IN" title="L2 Cache Requests" name="Snoop requests" description="The number of L2 snoop requests from internal masters." units="requests"/>
+        <event offset="21" advanced="yes" counter="L2_SNP_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Snoop stall cycles" description="The number of cycles where L2 cache snoop requests from internal masters are stalled." units="cycles"/>
+        <event offset="22" advanced="yes" counter="L2_RD_MSG_OUT" title="L2 Cache Requests" name="L1 read requests" description="The number of L1 cache read requests sent by the L2 cache to an internal master." units="requests"/>
+        <event offset="23" advanced="yes" counter="L2_RD_MSG_OUT_STALL" title="L2 Cache Stall Cycles" name="L1 read stall cycles" description="The number of cycles where L1 cache read requests sent by the L2 cache to an internal master are stalled." units="cycles"/>
+        <event offset="24" advanced="yes" counter="L2_WR_MSG_OUT" title="L2 Cache Requests" name="L1 write requests" description="The number of L1 cache write responses sent by the L2 cache to an internal master." units="requests"/>
+        <event offset="25" counter="L2_ANY_LOOKUP" title="L2 Cache Lookups" name="Any lookup" description="The number of L2 cache lookups performed." units="requests"/>
+        <event offset="26" counter="L2_READ_LOOKUP" title="L2 Cache Lookups" name="Read lookup" description="The number of L2 cache read lookups performed." units="requests"/>
+        <event offset="27" counter="L2_WRITE_LOOKUP" title="L2 Cache Lookups" name="Write lookup" description="The number of L2 cache write lookups performed." units="requests"/>
+        <event offset="28" advanced="yes" counter="L2_EXT_SNOOP_LOOKUP" title="L2 Cache Lookups" name="External snoop lookups" description="The number of coherency snoop lookups performed that were triggered by an external master." units="requests"/>
+        <event offset="29" counter="L2_EXT_READ" title="External Bus Accesses" name="Read transaction" description="The number of external read transactions." units="transactions"/>
+        <event offset="30" advanced="yes" counter="L2_EXT_READ_NOSNP" title="External Bus Accesses" name="ReadNoSnoop transactions" description="The number of external non-coherent read transactions." units="transactions"/>
+        <event offset="31" advanced="yes" counter="L2_EXT_READ_UNIQUE" title="External Bus Accesses" name="ReadUnique transactions" description="The number of external coherent read unique transactions." units="transactions"/>
+        <event offset="32" counter="L2_EXT_READ_BEATS" title="External Bus Beats" name="Read beat" description="The number of external bus data read cycles." units="beats"/>
+        <event offset="33" counter="L2_EXT_AR_STALL" title="External Bus Stalls" name="Read stall cycles" description="The number of cycles where a read is stalled waiting for the external bus." units="cycles"/>
+        <event offset="34" counter="L2_EXT_AR_CNT_Q1" title="External Bus Outstanding Reads" name="0-25% outstanding" description="The number of read transactions initiated when 0-25% of the maximum are in use." units="transactions"/>
+        <event offset="35" counter="L2_EXT_AR_CNT_Q2" title="External Bus Outstanding Reads" name="25-50% outstanding" description="The number of read transactions initiated when 25-50% of the maximum are in use." units="transactions"/>
+        <event offset="36" counter="L2_EXT_AR_CNT_Q3" title="External Bus Outstanding Reads" name="50-75% outstanding" description="The number of read transactions initiated when 50-75% of the maximum are in use." units="transactions"/>
+        <event offset="37" counter="L2_EXT_RRESP_0_127" title="External Bus Read Latency" name="0-127 cycles" description="The number of data beats returned 0-127 cycles after the read request." units="beats"/>
+        <event offset="38" counter="L2_EXT_RRESP_128_191" title="External Bus Read Latency" name="128-191 cycles" description="The number of data beats returned 128-191 cycles after the read request." units="beats"/>
+        <event offset="39" counter="L2_EXT_RRESP_192_255" title="External Bus Read Latency" name="192-255 cycles" description="The number of data beats returned 192-255 cycles after the read request." units="beats"/>
+        <event offset="40" counter="L2_EXT_RRESP_256_319" title="External Bus Read Latency" name="256-319 cycles" description="The number of data beats returned 256-319 cycles after the read request." units="beats"/>
+        <event offset="41" counter="L2_EXT_RRESP_320_383" title="External Bus Read Latency" name="320-383 cycles" description="The number of data beats returned 320-383 cycles after the read request." units="beats"/>
+        <event offset="42" counter="L2_EXT_WRITE" title="External Bus Accesses" name="Write transaction" description="The number of external write transactions." units="transactions"/>
+        <event offset="43" advanced="yes" counter="L2_EXT_WRITE_NOSNP_FULL" title="External Bus Accesses" name="WriteNoSnoopFull transactions" description="The number of external non-coherent full write transactions." units="transactions"/>
+        <event offset="44" advanced="yes" counter="L2_EXT_WRITE_NOSNP_PTL" title="External Bus Accesses" name="WriteNoSnoopPartial transactions" description="The number of external non-coherent partial write transactions." units="transactions"/>
+        <event offset="45" advanced="yes" counter="L2_EXT_WRITE_SNP_FULL" title="External Bus Accesses" name="WriteSnoopFull transactions" description="The number of external coherent full write transactions." units="transactions"/>
+        <event offset="46" advanced="yes" counter="L2_EXT_WRITE_SNP_PTL" title="External Bus Accesses" name="WriteSnoopPartial transactions" description="The number of external coherent partial write transactions." units="transactions"/>
+        <event offset="47" counter="L2_EXT_WRITE_BEATS" title="External Bus Beats" name="Write beat" description="The number of external bus data write cycles." units="beats"/>
+        <event offset="48" counter="L2_EXT_W_STALL" title="External Bus Stalls" name="Write stall cycles" description="The number of cycles where a write is stalled waiting for the external bus." units="cycles"/>
+        <event offset="49" counter="L2_EXT_AW_CNT_Q1" title="External Bus Outstanding Writes" name="0-25% outstanding" description="The number of write transactions initiated when 0-25% of the maximum are in use." units="transactions"/>
+        <event offset="50" counter="L2_EXT_AW_CNT_Q2" title="External Bus Outstanding Writes" name="25-50% outstanding" description="The number of write transactions initiated when 25-50% of the maximum are in use." units="transactions"/>
+        <event offset="51" counter="L2_EXT_AW_CNT_Q3" title="External Bus Outstanding Writes" name="50-75% outstanding" description="The number of write transactions initiated when 50-75% of the maximum are in use." units="transactions"/>
+        <event offset="52" advanced="yes" counter="L2_EXT_SNOOP" title="External Bus Accesses" name="Snoop transactions" description="The number of coherency snoops triggered by external masters." units="transactions"/>
+        <event offset="53" advanced="yes" counter="L2_EXT_SNOOP_STALL" title="External Bus Stalls" name="Snoop stall cycles" description="The number of cycles where a coherency snoop triggered by external master is stalled." units="cycles"/>
+    </category>
+    <category name="Shader Core" per_cpu="no">
+        <event offset="4" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment active" description="The number of cycles where the shader core is processing a fragment workload." units="cycles"/>
+        <event offset="5" advanced="yes" counter="FRAG_PRIMITIVES" title="Core Primitives" name="Read primitives" description="The number of primitives read from the tile list by the fragment front-end." units="primitives"/>
+        <event offset="6" counter="FRAG_PRIM_RAST" title="Core Primitives" name="Rasterized primitives" description="The number of primitives being rasterized." units="primitives"/>
+        <event offset="7" counter="FRAG_FPK_ACTIVE" title="Core Cycles" name="Fragment FPKB active" description="The number of cycles where at least one quad is present in the pre-pipe quad queue." units="cycles"/>
+        <event offset="9" counter="FRAG_WARPS" title="Core Warps" name="Fragment warps" description="The number of fragment warps created." units="warps"/>
+        <event offset="10" counter="FRAG_PARTIAL_WARPS" title="Core Warps" name="Partial fragment warps" description="The number of fragment warps containing helper threads that do not correspond to a hit sample point." units="warps"/>
+        <event offset="11" counter="FRAG_QUADS_RAST" title="Core Quads" name="Rasterized quads" description="The number of quads generated by the rasterization phase." units="quads"/>
+        <event offset="12" counter="FRAG_QUADS_EZS_TEST" title="Core Quads" name="Early ZS tested quads" description="The number of quads that are undergoing early depth and stencil testing." units="quads"/>
+        <event offset="13" counter="FRAG_QUADS_EZS_UPDATE" title="Core Quads" name="Early ZS updated quads" description="The number of quads undergoing early depth and stencil testing, that are capable of updating the framebuffer." units="quads"/>
+        <event offset="14" counter="FRAG_QUADS_EZS_KILL" title="Core Quads" name="Early ZS killed quads" description="The number of quads killed by early depth and stencil testing." units="quads"/>
+        <event offset="15" counter="FRAG_LZS_TEST" title="Core Quads" name="Late ZS tested quads" description="The number of quads undergoing late depth and stencil testing." units="quads"/>
+        <event offset="16" counter="FRAG_LZS_KILL" title="Core Quads" name="Late ZS killed quads" description="The number of quads killed by late depth and stencil testing." units="quads"/>
+        <event offset="18" counter="FRAG_PTILES" title="Core Tiles" name="Tiles" description="The number of tiles processed by the shader core." units="tiles"/>
+        <event offset="19" counter="FRAG_TRANS_ELIM" title="Core Tiles" name="Constant tiles killed" description="The number of tiles killed by transaction elimination." units="tiles"/>
+        <event offset="20" counter="QUAD_FPK_KILLER" title="Core Quads" name="FPK occluder quads" description="The number of quads that are valid occluders for hidden surface removal." units="quads"/>
+        <event offset="22" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Non-fragment active" description="The number of cycles where the shader core is processing some non-fragment workload." units="cycles"/>
+        <event offset="23" advanced="yes" counter="COMPUTE_TASKS" title="Core Tasks" name="Non-fragment tasks" description="The number of non-fragment tasks issued to the shader core." units="tasks"/>
+        <event offset="24" counter="COMPUTE_WARPS" title="Core Warps" name="Non-fragment warps" description="The number of non-fragment warps created." units="warps"/>
+        <event offset="25" advanced="yes" counter="COMPUTE_STARVING" title="Core Starvation Cycles" name="Non-fragment starvation cycles" description="The number of cycles where the shader core is processing a non-fragment workload and there are no new threads available for execution." units="cycles"/>
+        <event offset="26" counter="EXEC_CORE_ACTIVE" title="Core Cycles" name="Execution core active" description="The number of cycles where the shader core is processing at least one warp." units="cycles"/>
+        <event offset="27" advanced="yes" counter="EXEC_ACTIVE" title="Core Cycles" name="Execution engine active" description="The number of cycles where the execution engine unit is processing at least one thread." units="cycles"/>
+        <event offset="28" counter="EXEC_INSTR_COUNT" title="Core EE Instructions" name="Executed instructions" description="The number of instructions executed per warp." units="instructions"/>
+        <event offset="29" counter="EXEC_INSTR_DIVERGED" title="Core EE Instructions" name="Diverged instructions" description="The number of instructions executed per warp, that have control flow divergence." units="instructions"/>
+        <event offset="30" advanced="yes" counter="EXEC_INSTR_STARVING" title="Core Starvation Cycles" name="Execution engine starvation cycles" description="The number of cycles where no new threads are available for execution." units="cycles"/>
+        <event offset="31" advanced="yes" counter="ARITH_INSTR_SINGLE_FMA" title="Core EE Instructions" name="Arithmetic instructions" description="The number of instructions where the workload is a single FMA pipe arithmetic operation." units="instructions"/>
+        <event offset="32" advanced="yes" counter="ARITH_INSTR_DOUBLE" title="Core EE Instructions" name="Dual Arithmetic instructions" description="The number of instructions where the workload is one FMA pipe arithmetic operation and one ADD pipe arithmetic operation." units="instructions"/>
+        <event offset="33" advanced="yes" counter="ARITH_INSTR_MSG" title="Core EE Instructions" name="Arithmetic + Message instructions" description="The number of instructions where the workload is one FMA pipe arithmetic operation and one ADD pipe message operation" units="instructions"/>
+        <event offset="34" advanced="yes" counter="ARITH_INSTR_MSG_ONLY" title="Core EE Instructions" name="Message instructions" description="The number of instructions where the workload is a single ADD pipe message operation, with no FMA pipe operation" units="instructions"/>
+        <event offset="35" counter="TEX_MSGI_NUM_QUADS" title="Core Texture Quads" name="Texture requests" description="The number of quad-width texture operations processed by the texture unit." units="quads"/>
+        <event offset="36" counter="TEX_DFCH_NUM_PASSES" title="Core Texture Quads" name="Texture issues" description="The number of quad-width filtering passes." units="issues"/>
+        <event offset="37" counter="TEX_DFCH_NUM_PASSES_MISS" title="Core Texture Quads" name="Descriptor misses" description="The number of quad-width filtering passes that miss in the resource or sampler descriptor cache." units="requests"/>
+        <event offset="38" counter="TEX_DFCH_NUM_PASSES_MIP_MAP" title="Core Texture Quads" name="Mipmapped texture issues" description="The number of quad-width filtering passes that use a mipmapped texture." units="issues"/>
+        <event offset="39" counter="TEX_TIDX_NUM_SPLIT_MIP_MAP" title="Core Texture Quads" name="Trilinear filtered issues" description="The number of quad-width filtering passes that use a trilinear filter." units="issues"/>
+        <event offset="40" counter="TEX_TFCH_NUM_LINES_FETCHED" title="Core Texture Line Fetches" name="Line fetches" description="The number of texture line fetches from the L2 cache." units="issues"/>
+        <event offset="41" counter="TEX_TFCH_NUM_LINES_FETCHED_BLOCK_COMPRESSED" title="Core Texture Line Fetches" name="Compressed line fetches" description="The number of texture line fetches from the L2 cache that are block compressed textures." units="issues"/>
+        <event offset="42" counter="TEX_TFCH_NUM_OPERATIONS" title="Core Texture Cycles" name="Cache lookups" description="The number of texture cache lookup cycles." units="requests"/>
+        <event offset="43" counter="TEX_FILT_NUM_OPERATIONS" title="Core Texture Cycles" name="Texturing active" description="The number of texture filtering issue cycles." units="cycles"/>
+        <event offset="44" counter="LS_MEM_READ_FULL" title="Core Load/Store Cycles" name="Full read cycles" description="The number of full-width load/store cache reads." units="cycles"/>
+        <event offset="45" counter="LS_MEM_READ_SHORT" title="Core Load/Store Cycles" name="Partial read cycles" description="The number of partial-width load/store cache reads." units="cycles"/>
+        <event offset="46" counter="LS_MEM_WRITE_FULL" title="Core Load/Store Cycles" name="Full write cycles" description="The number of full-width load/store cache writes." units="cycles"/>
+        <event offset="47" counter="LS_MEM_WRITE_SHORT" title="Core Load/Store Cycles" name="Partial write cycles" description="The number of partial-width load/store cache writes." units="cycles"/>
+        <event offset="48" counter="LS_MEM_ATOMIC" title="Core Load/Store Cycles" name="Atomic access cycles" description="The number of load/store atomic accesses." units="cycles"/>
+        <event offset="49" counter="VARY_INSTR" title="Core Varying Requests" name="Interpolation requests" description="The number of warp-width interpolation operations processed by the varying unit." units="instructions"/>
+        <event offset="50" counter="VARY_SLOT_32" title="Core Varying Cycles" name="32-bit interpolation active" description="The number of 32-bit interpolation cycles processed by the varying unit." units="cycles"/>
+        <event offset="51" counter="VARY_SLOT_16" title="Core Varying Cycles" name="16-bit interpolation active" description="The number of 16-bit interpolation cycles processed by the varying unit." units="cycles"/>
+        <event offset="52" advanced="yes" counter="ATTR_INSTR" title="Core Attribute Requests" name="Attribute requests" description="The number of instructions executed by the attribute unit." units="instructions"/>
+        <event offset="53" advanced="yes" counter="ARITH_INSTR_FP_MUL" title="Core EE Instructions" name="Multiplier instructions" description="The number of instructions where the workload uses floating-point multiplier hardware." units="instructions"/>
+        <event offset="54" counter="BEATS_RD_FTC" title="Core L2 Reads" name="Fragment L2 read beats" description="The number of read beats received by the fixed-function fragment front-end." units="beats"/>
+        <event offset="55" counter="BEATS_RD_FTC_EXT" title="Core L2 Reads" name="Fragment external read beats" description="The number of read beats received by the fixed-function fragment front-end that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="56" counter="BEATS_RD_LSC" title="Core L2 Reads" name="Load/store L2 read beats" description="The number of read beats received by the load/store unit." units="beats"/>
+        <event offset="57" counter="BEATS_RD_LSC_EXT" title="Core L2 Reads" name="Load/store external read beats" description="The number of read beats received by the load/store unit that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="58" counter="BEATS_RD_TEX" title="Core L2 Reads" name="Texture L2 read beats" description="The number of read beats received by the texture unit." units="beats"/>
+        <event offset="59" counter="BEATS_RD_TEX_EXT" title="Core L2 Reads" name="Texture external read beats" description="The number of read beats received by the texture unit that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="60" advanced="yes" counter="BEATS_RD_OTHER" title="Core L2 Reads" name="Other L2 read beats" description="The number of read beats received by a unit that is not specifically identified." units="beats"/>
+        <event offset="61" counter="BEATS_WR_LSC_OTHER" title="Core Writes" name="Load/store other write beats" description="The number of write beats by the load/store unit that are due to any reason other than writeback." units="beats"/>
+        <event offset="62" counter="BEATS_WR_TIB" title="Core Writes" name="Tile buffer write beats" description="The number of write beats sent by the tile buffer writeback unit." units="beats"/>
+        <event offset="63" counter="BEATS_WR_LSC_WB" title="Core Writes" name="Load/store writeback write beats" description="The number of write beats by the load/store unit that are due to writeback." units="beats"/>
     </category>
 </metrics>
\ No newline at end of file
diff --git a/src/panfrost/perf/G52.xml b/src/panfrost/perf/G52.xml
index 02b21da397f..35a9eaa7ba3 100644
--- a/src/panfrost/perf/G52.xml
+++ b/src/panfrost/perf/G52.xml
@@ -1,6 +1,6 @@
 <!--
 Copyright © 2017-2020 ARM Limited.
-Copyright © 2021 Collabora, Ltd.
+Copyright © 2021-2022 Collabora, Ltd.
 Author: Antonio Caggiano <antonio.caggiano@collabora.com>
 
 Permission is hereby granted, free of charge, to any person obtaining a
@@ -24,156 +24,156 @@ SOFTWARE.
 -->
 <metrics id="TGOx">
     <category name="Job Manager" per_cpu="no">
-        <event offset="6" counter="GPU_ACTIVE" title="GPU Cycles" name="GPU active" description="The number of cycles where the GPU has a workload of any type queued for processing." units="cycles" />
-        <event offset="7" advanced="yes" counter="IRQ_ACTIVE" title="GPU Cycles" name="Interrupt active" description="The number of cycles where the GPU has a pending interrupt." units="cycles" />
-        <event offset="8" advanced="yes" counter="JS0_JOBS" title="GPU Jobs" name="Fragment jobs" description="The number of jobs processed by the GPU fragment queue." units="jobs" />
-        <event offset="9" counter="JS0_TASKS" title="GPU Tasks" name="Fragment tasks" description="The number of 32x32 pixel tasks processed by the GPU fragment queue." units="tasks" />
-        <event offset="10" counter="JS0_ACTIVE" title="GPU Cycles" name="Fragment queue active" description="The number of cycles where work is queued for processing in the GPU fragment queue." units="cycles" />
-        <event offset="12" advanced="yes" counter="JS0_WAIT_READ" title="GPU Wait Cycles" name="Fragment descriptor reads cycles" description="The number of cycles where queued fragment work is waiting for a descriptor load." units="cycles" />
-        <event offset="13" advanced="yes" counter="JS0_WAIT_ISSUE" title="GPU Wait Cycles" name="Fragment job issue cycles" description="The number of cycles where queued fragment work is waiting for an available processor." units="cycles" />
-        <event offset="14" advanced="yes" counter="JS0_WAIT_DEPEND" title="GPU Wait Cycles" name="Fragment job dependency cycles" description="The number of cycles where queued fragment work is waiting for dependent work to complete." units="cycles" />
-        <event offset="15" advanced="yes" counter="JS0_WAIT_FINISH" title="GPU Wait Cycles" name="Fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued fragment work to complete." units="cycles" />
-        <event offset="16" advanced="yes" counter="JS1_JOBS" title="GPU Jobs" name="Non-fragment jobs" description="The number of jobs processed by the GPU non-fragment queue." units="jobs" />
-        <event offset="17" advanced="yes" counter="JS1_TASKS" title="GPU Tasks" name="Non-fragment tasks" description="The number of tasks processed by the GPU non-fragment queue." units="tasks" />
-        <event offset="18" counter="JS1_ACTIVE" title="GPU Cycles" name="Non-fragment queue active" description="The number of cycles where work is queued in the GPU non-fragment queue." units="cycles" />
-        <event offset="20" advanced="yes" counter="JS1_WAIT_READ" title="GPU Wait Cycles" name="Non-fragment descriptor read cycles" description="The number number of cycles where queued non-fragment work is waiting for a descriptor load." units="cycles" />
-        <event offset="21" advanced="yes" counter="JS1_WAIT_ISSUE" title="GPU Wait Cycles" name="Non-fragment job issue cycles" description="The number of cycles where queued non-fragment work is waiting for an available processor." units="cycles" />
-        <event offset="22" advanced="yes" counter="JS1_WAIT_DEPEND" title="GPU Wait Cycles" name="Non-fragment job dependency cycles" description="The number of cycles where queued non-fragment work is waiting for dependent work to complete." units="cycles" />
-        <event offset="23" advanced="yes" counter="JS1_WAIT_FINISH" title="GPU Wait Cycles" name="Non-fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued non-fragment work to complete." units="cycles" />
-        <event offset="24" advanced="yes" counter="JS2_JOBS" title="GPU Jobs" name="Reserved jobs" description="The number of jobs processed by the GPU reserved queue." units="jobs" />
-        <event offset="25" advanced="yes" counter="JS2_TASKS" title="GPU Tasks" name="Reserved tasks" description="The number of tasks processed by the GPU reserved queue." units="tasks" />
-        <event offset="26" advanced="yes" counter="JS2_ACTIVE" title="GPU Cycles" name="Reserved queue active" description="The number of cycles where work is queued in the GPU reserved queue." units="cycles" />
-        <event offset="28" advanced="yes" counter="JS2_WAIT_READ" title="GPU Wait Cycles" name="Reserved descriptor read cycles" description="The number of cycles where queued reserved work is waiting for a descriptor load." units="cycles" />
-        <event offset="29" advanced="yes" counter="JS2_WAIT_ISSUE" title="GPU Wait Cycles" name="Reserved job issue cycles" description="The number of cycles where queued reserved work is waiting for an available processor." units="cycles" />
-        <event offset="30" advanced="yes" counter="JS2_WAIT_DEPEND" title="GPU Wait Cycles" name="Reserved job dependency cycles" description="The number of cycles where queued reserved work is waiting for dependent work to complete." units="cycles" />
-        <event offset="31" advanced="yes" counter="JS2_WAIT_FINISH" title="GPU Wait Cycles" name="Reserved job finish cycles" description="The number of cycles where the GPU is waiting for issued reserved work to complete." units="cycles" />
+        <event offset="6" counter="GPU_ACTIVE" title="GPU Cycles" name="GPU active" description="The number of cycles where the GPU has a workload of any type queued for processing." units="cycles"/>
+        <event offset="7" advanced="yes" counter="IRQ_ACTIVE" title="GPU Cycles" name="Interrupt active" description="The number of cycles where the GPU has a pending interrupt." units="cycles"/>
+        <event offset="8" advanced="yes" counter="JS0_JOBS" title="GPU Jobs" name="Fragment jobs" description="The number of jobs processed by the GPU fragment queue." units="jobs"/>
+        <event offset="9" counter="JS0_TASKS" title="GPU Tasks" name="Fragment tasks" description="The number of 32x32 pixel tasks processed by the GPU fragment queue." units="tasks"/>
+        <event offset="10" counter="JS0_ACTIVE" title="GPU Cycles" name="Fragment queue active" description="The number of cycles where work is queued for processing in the GPU fragment queue." units="cycles"/>
+        <event offset="12" advanced="yes" counter="JS0_WAIT_READ" title="GPU Wait Cycles" name="Fragment descriptor reads cycles" description="The number of cycles where queued fragment work is waiting for a descriptor load." units="cycles"/>
+        <event offset="13" advanced="yes" counter="JS0_WAIT_ISSUE" title="GPU Wait Cycles" name="Fragment job issue cycles" description="The number of cycles where queued fragment work is waiting for an available processor." units="cycles"/>
+        <event offset="14" advanced="yes" counter="JS0_WAIT_DEPEND" title="GPU Wait Cycles" name="Fragment job dependency cycles" description="The number of cycles where queued fragment work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="15" advanced="yes" counter="JS0_WAIT_FINISH" title="GPU Wait Cycles" name="Fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued fragment work to complete." units="cycles"/>
+        <event offset="16" advanced="yes" counter="JS1_JOBS" title="GPU Jobs" name="Non-fragment jobs" description="The number of jobs processed by the GPU non-fragment queue." units="jobs"/>
+        <event offset="17" advanced="yes" counter="JS1_TASKS" title="GPU Tasks" name="Non-fragment tasks" description="The number of tasks processed by the GPU non-fragment queue." units="tasks"/>
+        <event offset="18" counter="JS1_ACTIVE" title="GPU Cycles" name="Non-fragment queue active" description="The number of cycles where work is queued in the GPU non-fragment queue." units="cycles"/>
+        <event offset="20" advanced="yes" counter="JS1_WAIT_READ" title="GPU Wait Cycles" name="Non-fragment descriptor read cycles" description="The number number of cycles where queued non-fragment work is waiting for a descriptor load." units="cycles"/>
+        <event offset="21" advanced="yes" counter="JS1_WAIT_ISSUE" title="GPU Wait Cycles" name="Non-fragment job issue cycles" description="The number of cycles where queued non-fragment work is waiting for an available processor." units="cycles"/>
+        <event offset="22" advanced="yes" counter="JS1_WAIT_DEPEND" title="GPU Wait Cycles" name="Non-fragment job dependency cycles" description="The number of cycles where queued non-fragment work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="23" advanced="yes" counter="JS1_WAIT_FINISH" title="GPU Wait Cycles" name="Non-fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued non-fragment work to complete." units="cycles"/>
+        <event offset="24" advanced="yes" counter="JS2_JOBS" title="GPU Jobs" name="Reserved jobs" description="The number of jobs processed by the GPU reserved queue." units="jobs"/>
+        <event offset="25" advanced="yes" counter="JS2_TASKS" title="GPU Tasks" name="Reserved tasks" description="The number of tasks processed by the GPU reserved queue." units="tasks"/>
+        <event offset="26" advanced="yes" counter="JS2_ACTIVE" title="GPU Cycles" name="Reserved queue active" description="The number of cycles where work is queued in the GPU reserved queue." units="cycles"/>
+        <event offset="28" advanced="yes" counter="JS2_WAIT_READ" title="GPU Wait Cycles" name="Reserved descriptor read cycles" description="The number of cycles where queued reserved work is waiting for a descriptor load." units="cycles"/>
+        <event offset="29" advanced="yes" counter="JS2_WAIT_ISSUE" title="GPU Wait Cycles" name="Reserved job issue cycles" description="The number of cycles where queued reserved work is waiting for an available processor." units="cycles"/>
+        <event offset="30" advanced="yes" counter="JS2_WAIT_DEPEND" title="GPU Wait Cycles" name="Reserved job dependency cycles" description="The number of cycles where queued reserved work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="31" advanced="yes" counter="JS2_WAIT_FINISH" title="GPU Wait Cycles" name="Reserved job finish cycles" description="The number of cycles where the GPU is waiting for issued reserved work to complete." units="cycles"/>
     </category>
     <category name="Tiler" per_cpu="no">
-        <event offset="68" counter="TILER_ACTIVE" title="GPU Cycles" name="Tiler active" description="The number of cycles where the tiler has a workload queued for processing." units="cycles" />
-        <event offset="70" counter="TRIANGLES" title="Input Primitives" name="Triangle primitives" description="The number of input triangle primitives." units="primitives" />
-        <event offset="71" counter="LINES" title="Input Primitives" name="Line primitives" description="The number of input line primitives." units="primitives" />
-        <event offset="72" counter="POINTS" title="Input Primitives" name="Point primitives" description="The number of input point primitives." units="primitives" />
-        <event offset="73" counter="FRONT_FACING" title="Visible Primitives" name="Front-facing primitives" description="The number of front-facing triangles that are visible after culling." units="primitives" />
-        <event offset="74" counter="BACK_FACING" title="Visible Primitives" name="Back-facing primitives" description="The number of back-facing triangles that are visible after culling." units="primitives" />
-        <event offset="75" counter="PRIM_VISIBLE" title="Primitive Culling" name="Visible primitives" description="The number of primitives that are visible after culling." units="primitives" />
-        <event offset="76" counter="PRIM_CULLED" title="Primitive Culling" name="Facing and XY plane test culled primitives" description="The number of primitives that are culled by facing or frustum XY plane tests." units="primitives" />
-        <event offset="77" counter="PRIM_CLIPPED" title="Primitive Culling" name="Z plane test culled primitives" description="The number of primitives that are culled by frustum Z plane tests." units="primitives" />
-        <event offset="78" counter="PRIM_SAT_CULLED" title="Primitive Culling" name="Sample test culled primitives" description="The number of primitives culled by the sample coverage test." units="primitives" />
-        <event offset="81" advanced="yes" counter="BUS_READ" title="Tiler L2 Accesses" name="Read beats" description="The number of internal bus data read cycles made by the tiler." units="beats" />
-        <event offset="83" advanced="yes" counter="BUS_WRITE" title="Tiler L2 Accesses" name="Write beats" description="The number of internal bus data write cycles made by the tiler." units="beats" />
-        <event offset="85" counter="IDVS_POS_SHAD_REQ" title="Tiler Shading Requests" name="Position shading requests" description="The number of position shading requests in the IDVS flow." units="requests" />
-        <event offset="87" advanced="yes" counter="IDVS_POS_SHAD_STALL" title="Tiler Cycles" name="Position shading stall cycles" description="The number of cycles where the tiler has a stalled position shading request." units="cycles" />
-        <event offset="88" advanced="yes" counter="IDVS_POS_FIFO_FULL" title="Tiler Cycles" name="Position FIFO full cycles" description="The number of cycles where the tiler has a stalled position shading buffer." units="cycles" />
-        <event offset="90" advanced="yes" counter="VCACHE_HIT" title="Tiler Vertex Cache" name="Position cache hits" description="The number of position lookups that result in a hit in the vertex cache." units="requests" />
-        <event offset="91" advanced="yes" counter="VCACHE_MISS" title="Tiler Vertex Cache" name="Position cache misses" description="The number of position lookups that miss in the vertex cache." units="requests" />
-        <event offset="95" advanced="yes" counter="VFETCH_STALL" title="Tiler Cycles" name="Primitive assembly busy stall cycles" description="The number of cycles where the tiler is stalled waiting for primitive assembly." units="cycles" />
-        <event offset="98" advanced="yes" counter="IDVS_VBU_HIT" title="Tiler Vertex Cache" name="Varying cache hits" description="The number of varying lookups that result in a hit in the vertex cache." units="requests" />
-        <event offset="99" advanced="yes" counter="IDVS_VBU_MISS" title="Tiler Vertex Cache" name="Varying cache misses" description="The number of varying lookups that miss in the vertex cache." units="requests" />
-        <event offset="101" counter="IDVS_VAR_SHAD_REQ" title="Tiler Shading Requests" name="Varying shading requests" description="The number of varying shading requests in the IDVS flow." units="requests" />
-        <event offset="102" advanced="yes" counter="IDVS_VAR_SHAD_STALL" title="Tiler Cycles" name="Varying shading stall cycles" description="The number of cycles where the tiler has a stalled varying shading request." units="cycles" />
-        <event offset="118" advanced="yes" counter="WRBUF_NO_AXI_ID_STALL" title="Tiler Cycles" name="Write buffer transaction stall cycles" description="The number of cycles where the tiler write buffer can not send data because it has no available write IDs." units="cycles" />
-        <event offset="119" advanced="yes" counter="WRBUF_AXI_STALL" title="Tiler Cycles" name="Write buffer write stall cycles" description="The number of cycles where the tiler write buffer can not send data because the bus is not ready." units="cycles" />
-    </category>
-    <category name="Shader Core" per_cpu="no">
-        <event offset="196" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment active" description="The number of cycles where the shader core is processing a fragment workload." units="cycles" />
-        <event offset="197" advanced="yes" counter="FRAG_PRIMITIVES" title="Core Primitives" name="Read primitives" description="The number of primitives read from the tile list by the fragment front-end." units="primitives" />
-        <event offset="198" counter="FRAG_PRIM_RAST" title="Core Primitives" name="Rasterized primitives" description="The number of primitives being rasterized." units="primitives" />
-        <event offset="199" counter="FRAG_FPK_ACTIVE" title="Core Cycles" name="Fragment FPKB active" description="The number of cycles where at least one quad is present in the pre-pipe quad queue." units="cycles" />
-        <event offset="201" counter="FRAG_WARPS" title="Core Warps" name="Fragment warps" description="The number of fragment warps created." units="warps" />
-        <event offset="202" counter="FRAG_PARTIAL_WARPS" title="Core Warps" name="Partial fragment warps" description="The number of fragment warps containing helper threads that do not correspond to a hit sample point." units="warps" />
-        <event offset="203" counter="FRAG_QUADS_RAST" title="Core Quads" name="Rasterized quads" description="The number of quads generated by the rasterization phase." units="quads" />
-        <event offset="204" counter="FRAG_QUADS_EZS_TEST" title="Core Quads" name="Early ZS tested quads" description="The number of quads that are undergoing early depth and stencil testing." units="quads" />
-        <event offset="205" counter="FRAG_QUADS_EZS_UPDATE" title="Core Quads" name="Early ZS updated quads" description="The number of quads undergoing early depth and stencil testing, that are capable of updating the framebuffer." units="quads" />
-        <event offset="206" counter="FRAG_QUADS_EZS_KILL" title="Core Quads" name="Early ZS killed quads" description="The number of quads killed by early depth and stencil testing." units="quads" />
-        <event offset="207" counter="FRAG_LZS_TEST" title="Core Quads" name="Late ZS tested quads" description="The number of quads undergoing late depth and stencil testing." units="quads" />
-        <event offset="208" counter="FRAG_LZS_KILL" title="Core Quads" name="Late ZS killed quads" description="The number of quads killed by late depth and stencil testing." units="quads" />
-        <event offset="209" counter="WARP_REG_SIZE_64" title="Core Warps" name="All register warps" description="The number of warps that require more than 32 registers." units="warps" />
-        <event offset="210" counter="FRAG_PTILES" title="Core Tiles" name="Tiles" description="The number of tiles processed by the shader core." units="tiles" />
-        <event offset="211" counter="FRAG_TRANS_ELIM" title="Core Tiles" name="Constant tiles killed" description="The number of tiles killed by transaction elimination." units="tiles" />
-        <event offset="212" counter="QUAD_FPK_KILLER" title="Core Quads" name="FPK occluder quads" description="The number of quads that are valid occluders for hidden surface removal." units="quads" />
-        <event offset="213" counter="FULL_QUAD_WARPS" title="Core Warps" name="Full quad warps" description="The number of warps that are fully populated with quads." units="warps" />
-        <event offset="214" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Non-fragment active" description="The number of cycles where the shader core is processing some non-fragment workload." units="cycles" />
-        <event offset="215" advanced="yes" counter="COMPUTE_TASKS" title="Core Tasks" name="Non-fragment tasks" description="The number of non-fragment tasks issued to the shader core." units="tasks" />
-        <event offset="216" counter="COMPUTE_WARPS" title="Core Warps" name="Non-fragment warps" description="The number of non-fragment warps created." units="warps" />
-        <event offset="217" advanced="yes" counter="COMPUTE_STARVING" title="Core Starvation Cycles" name="Non-fragment starvation cycles" description="The number of cycles where the shader core is processing a non-fragment workload and there are no new threads available for execution." units="cycles" />
-        <event offset="218" counter="EXEC_CORE_ACTIVE" title="Core Cycles" name="Execution core active" description="The number of cycles where the shader core is processing at least one warp." units="cycles" />
-        <event offset="219" advanced="yes" counter="EXEC_ACTIVE" title="Core Cycles" name="Execution engine active" description="The number of cycles where the execution engine unit is processing at least one thread." units="cycles" />
-        <event offset="220" counter="EXEC_INSTR_COUNT" title="Core EE Instructions" name="Executed instructions" description="The number of instructions executed per warp." units="instructions" />
-        <event offset="221" counter="EXEC_INSTR_DIVERGED" title="Core EE Instructions" name="Diverged instructions" description="The number of instructions executed per warp, that have control flow divergence." units="instructions" />
-        <event offset="222" advanced="yes" counter="EXEC_INSTR_STARVING" title="Core Starvation Cycles" name="Execution engine starvation cycles" description="The number of cycles where no new threads are available for execution." units="cycles" />
-        <event offset="223" advanced="yes" counter="ARITH_INSTR_SINGLE_FMA" title="Core EE Instructions" name="Arithmetic instructions" description="The number of instructions where the workload is a single FMA pipe arithmetic operation." units="instructions" />
-        <event offset="224" advanced="yes" counter="ARITH_INSTR_DOUBLE" title="Core EE Instructions" name="Dual Arithmetic instructions" description="The number of instructions where the workload is one FMA pipe arithmetic operation and one ADD pipe arithmetic operation." units="instructions" />
-        <event offset="225" advanced="yes" counter="ARITH_INSTR_MSG" title="Core EE Instructions" name="Arithmetic + Message instructions" description="The number of instructions where the workload is one FMA pipe arithmetic operation and one ADD pipe message operation" units="instructions" />
-        <event offset="226" advanced="yes" counter="ARITH_INSTR_MSG_ONLY" title="Core EE Instructions" name="Message instructions" description="The number of instructions where the workload is a single ADD pipe message operation, with no FMA pipe operation" units="instructions" />
-        <event offset="227" counter="TEX_MSGI_NUM_QUADS" title="Core Texture Quads" name="Texture requests" description="The number of quad-width texture operations processed by the texture unit." units="quads" />
-        <event offset="228" counter="TEX_DFCH_NUM_PASSES" title="Core Texture Quads" name="Texture issues" description="The number of quad-width filtering passes." units="issues" />
-        <event offset="229" counter="TEX_DFCH_NUM_PASSES_MISS" title="Core Texture Quads" name="Descriptor misses" description="The number of quad-width filtering passes that miss in the resource or sampler descriptor cache." units="requests" />
-        <event offset="230" counter="TEX_DFCH_NUM_PASSES_MIP_MAP" title="Core Texture Quads" name="Mipmapped texture issues" description="The number of quad-width filtering passes that use a mipmapped texture." units="issues" />
-        <event offset="231" counter="TEX_TIDX_NUM_SPLIT_MIP_MAP" title="Core Texture Quads" name="Trilinear filtered issues" description="The number of quad-width filtering passes that use a trilinear filter." units="issues" />
-        <event offset="232" counter="TEX_TFCH_NUM_LINES_FETCHED" title="Core Texture Line Fetches" name="Line fetches" description="The number of texture line fetches from the L2 cache." units="issues" />
-        <event offset="233" counter="TEX_TFCH_NUM_LINES_FETCHED_BLOCK_COMPRESSED" title="Core Texture Line Fetches" name="Compressed line fetches" description="The number of texture line fetches from the L2 cache that are block compressed textures." units="issues" />
-        <event offset="234" counter="TEX_TFCH_NUM_OPERATIONS" title="Core Texture Cycles" name="Cache lookups" description="The number of texture cache lookup cycles." units="requests" />
-        <event offset="235" counter="TEX_FILT_NUM_OPERATIONS" title="Core Texture Cycles" name="Texturing active" description="The number of texture filtering issue cycles." units="cycles" />
-        <event offset="236" counter="LS_MEM_READ_FULL" title="Core Load/Store Cycles" name="Full read cycles" description="The number of full-width load/store cache reads." units="cycles" />
-        <event offset="237" counter="LS_MEM_READ_SHORT" title="Core Load/Store Cycles" name="Partial read cycles" description="The number of partial-width load/store cache reads." units="cycles" />
-        <event offset="238" counter="LS_MEM_WRITE_FULL" title="Core Load/Store Cycles" name="Full write cycles" description="The number of full-width load/store cache writes." units="cycles" />
-        <event offset="239" counter="LS_MEM_WRITE_SHORT" title="Core Load/Store Cycles" name="Partial write cycles" description="The number of partial-width load/store cache writes." units="cycles" />
-        <event offset="240" counter="LS_MEM_ATOMIC" title="Core Load/Store Cycles" name="Atomic access cycles" description="The number of load/store atomic accesses." units="cycles" />
-        <event offset="241" counter="VARY_INSTR" title="Core Varying Requests" name="Interpolation requests" description="The number of warp-width interpolation operations processed by the varying unit." units="instructions" />
-        <event offset="242" counter="VARY_SLOT_32" title="Core Varying Cycles" name="32-bit interpolation active" description="The number of 32-bit interpolation cycles processed by the varying unit." units="cycles" />
-        <event offset="243" counter="VARY_SLOT_16" title="Core Varying Cycles" name="16-bit interpolation active" description="The number of 16-bit interpolation cycles processed by the varying unit." units="cycles" />
-        <event offset="244" advanced="yes" counter="ATTR_INSTR" title="Core Attribute Requests" name="Attribute requests" description="The number of instructions executed by the attribute unit." units="instructions" />
-        <event offset="245" advanced="yes" counter="ARITH_INSTR_FP_MUL" title="Core EE Instructions" name="Multiplier instructions" description="The number of instructions where the workload uses floating-point multiplier hardware." units="instructions" />
-        <event offset="246" counter="BEATS_RD_FTC" title="Core L2 Reads" name="Fragment L2 read beats" description="The number of read beats received by the fixed-function fragment front-end." units="beats" />
-        <event offset="247" counter="BEATS_RD_FTC_EXT" title="Core L2 Reads" name="Fragment external read beats" description="The number of read beats received by the fixed-function fragment front-end that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="248" counter="BEATS_RD_LSC" title="Core L2 Reads" name="Load/store L2 read beats" description="The number of read beats received by the load/store unit." units="beats" />
-        <event offset="249" counter="BEATS_RD_LSC_EXT" title="Core L2 Reads" name="Load/store external read beats" description="The number of read beats received by the load/store unit that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="250" counter="BEATS_RD_TEX" title="Core L2 Reads" name="Texture L2 read beats" description="The number of read beats received by the texture unit." units="beats" />
-        <event offset="251" counter="BEATS_RD_TEX_EXT" title="Core L2 Reads" name="Texture external read beats" description="The number of read beats received by the texture unit that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="252" advanced="yes" counter="BEATS_RD_OTHER" title="Core L2 Reads" name="Other L2 read beats" description="The number of read beats received by a unit that is not specifically identified." units="beats" />
-        <event offset="253" counter="BEATS_WR_LSC_WB" title="Core Writes" name="Load/store writeback write beats" description="The number of write beats by the load/store unit that are due to writeback." units="beats" />
-        <event offset="254" counter="BEATS_WR_TIB" title="Core Writes" name="Tile buffer write beats" description="The number of write beats sent by the tile buffer writeback unit." units="beats" />
-        <event offset="255" counter="BEATS_WR_LSC_OTHER" title="Core Writes" name="Load/store other write beats" description="The number of write beats by the load/store unit that are due to any reason other than writeback." units="beats" />
+        <event offset="4" counter="TILER_ACTIVE" title="GPU Cycles" name="Tiler active" description="The number of cycles where the tiler has a workload queued for processing." units="cycles"/>
+        <event offset="6" counter="TRIANGLES" title="Input Primitives" name="Triangle primitives" description="The number of input triangle primitives." units="primitives"/>
+        <event offset="7" counter="LINES" title="Input Primitives" name="Line primitives" description="The number of input line primitives." units="primitives"/>
+        <event offset="8" counter="POINTS" title="Input Primitives" name="Point primitives" description="The number of input point primitives." units="primitives"/>
+        <event offset="9" counter="FRONT_FACING" title="Visible Primitives" name="Front-facing primitives" description="The number of front-facing triangles that are visible after culling." units="primitives"/>
+        <event offset="10" counter="BACK_FACING" title="Visible Primitives" name="Back-facing primitives" description="The number of back-facing triangles that are visible after culling." units="primitives"/>
+        <event offset="11" counter="PRIM_VISIBLE" title="Primitive Culling" name="Visible primitives" description="The number of primitives that are visible after culling." units="primitives"/>
+        <event offset="12" counter="PRIM_CULLED" title="Primitive Culling" name="Facing and XY plane test culled primitives" description="The number of primitives that are culled by facing or frustum XY plane tests." units="primitives"/>
+        <event offset="13" counter="PRIM_CLIPPED" title="Primitive Culling" name="Z plane test culled primitives" description="The number of primitives that are culled by frustum Z plane tests." units="primitives"/>
+        <event offset="14" counter="PRIM_SAT_CULLED" title="Primitive Culling" name="Sample test culled primitives" description="The number of primitives culled by the sample coverage test." units="primitives"/>
+        <event offset="17" advanced="yes" counter="BUS_READ" title="Tiler L2 Accesses" name="Read beats" description="The number of internal bus data read cycles made by the tiler." units="beats"/>
+        <event offset="19" advanced="yes" counter="BUS_WRITE" title="Tiler L2 Accesses" name="Write beats" description="The number of internal bus data write cycles made by the tiler." units="beats"/>
+        <event offset="21" counter="IDVS_POS_SHAD_REQ" title="Tiler Shading Requests" name="Position shading requests" description="The number of position shading requests in the IDVS flow." units="requests"/>
+        <event offset="23" advanced="yes" counter="IDVS_POS_SHAD_STALL" title="Tiler Cycles" name="Position shading stall cycles" description="The number of cycles where the tiler has a stalled position shading request." units="cycles"/>
+        <event offset="24" advanced="yes" counter="IDVS_POS_FIFO_FULL" title="Tiler Cycles" name="Position FIFO full cycles" description="The number of cycles where the tiler has a stalled position shading buffer." units="cycles"/>
+        <event offset="26" advanced="yes" counter="VCACHE_HIT" title="Tiler Vertex Cache" name="Position cache hits" description="The number of position lookups that result in a hit in the vertex cache." units="requests"/>
+        <event offset="27" advanced="yes" counter="VCACHE_MISS" title="Tiler Vertex Cache" name="Position cache misses" description="The number of position lookups that miss in the vertex cache." units="requests"/>
+        <event offset="31" advanced="yes" counter="VFETCH_STALL" title="Tiler Cycles" name="Primitive assembly busy stall cycles" description="The number of cycles where the tiler is stalled waiting for primitive assembly." units="cycles"/>
+        <event offset="34" advanced="yes" counter="IDVS_VBU_HIT" title="Tiler Vertex Cache" name="Varying cache hits" description="The number of varying lookups that result in a hit in the vertex cache." units="requests"/>
+        <event offset="35" advanced="yes" counter="IDVS_VBU_MISS" title="Tiler Vertex Cache" name="Varying cache misses" description="The number of varying lookups that miss in the vertex cache." units="requests"/>
+        <event offset="37" counter="IDVS_VAR_SHAD_REQ" title="Tiler Shading Requests" name="Varying shading requests" description="The number of varying shading requests in the IDVS flow." units="requests"/>
+        <event offset="38" advanced="yes" counter="IDVS_VAR_SHAD_STALL" title="Tiler Cycles" name="Varying shading stall cycles" description="The number of cycles where the tiler has a stalled varying shading request." units="cycles"/>
+        <event offset="54" advanced="yes" counter="WRBUF_NO_AXI_ID_STALL" title="Tiler Cycles" name="Write buffer transaction stall cycles" description="The number of cycles where the tiler write buffer can not send data because it has no available write IDs." units="cycles"/>
+        <event offset="55" advanced="yes" counter="WRBUF_AXI_STALL" title="Tiler Cycles" name="Write buffer write stall cycles" description="The number of cycles where the tiler write buffer can not send data because the bus is not ready." units="cycles"/>
     </category>
     <category name="Memory System" per_cpu="no">
-        <event offset="132" advanced="yes" counter="MMU_REQUESTS" title="MMU Stage 1 Translations" name="MMU lookups" description="The number of main MMU address translations performed." units="requests" />
-        <event offset="144" advanced="yes" counter="L2_RD_MSG_IN" title="L2 Cache Requests" name="Read requests" description="The number of L2 cache read requests from internal masters." units="requests" />
-        <event offset="145" advanced="yes" counter="L2_RD_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Read stall cycles" description="The number of cycles L2 cache read requests from internal masters are stalled." units="cycles" />
-        <event offset="146" advanced="yes" counter="L2_WR_MSG_IN" title="L2 Cache Requests" name="Write requests" description="The number of L2 cache write requests from internal masters." units="requests" />
-        <event offset="147" advanced="yes" counter="L2_WR_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Write stall cycles" description="The number of cycles where L2 cache write requests from internal masters are stalled." units="cycles" />
-        <event offset="148" advanced="yes" counter="L2_SNP_MSG_IN" title="L2 Cache Requests" name="Snoop requests" description="The number of L2 snoop requests from internal masters." units="requests" />
-        <event offset="149" advanced="yes" counter="L2_SNP_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Snoop stall cycles" description="The number of cycles where L2 cache snoop requests from internal masters are stalled." units="cycles" />
-        <event offset="150" advanced="yes" counter="L2_RD_MSG_OUT" title="L2 Cache Requests" name="L1 read requests" description="The number of L1 cache read requests sent by the L2 cache to an internal master." units="requests" />
-        <event offset="151" advanced="yes" counter="L2_RD_MSG_OUT_STALL" title="L2 Cache Stall Cycles" name="L1 read stall cycles" description="The number of cycles where L1 cache read requests sent by the L2 cache to an internal master are stalled." units="cycles" />
-        <event offset="152" advanced="yes" counter="L2_WR_MSG_OUT" title="L2 Cache Requests" name="L1 write requests" description="The number of L1 cache write responses sent by the L2 cache to an internal master." units="requests" />
-        <event offset="153" counter="L2_ANY_LOOKUP" title="L2 Cache Lookups" name="Any lookup" description="The number of L2 cache lookups performed." units="requests" />
-        <event offset="154" counter="L2_READ_LOOKUP" title="L2 Cache Lookups" name="Read lookup" description="The number of L2 cache read lookups performed." units="requests" />
-        <event offset="155" counter="L2_WRITE_LOOKUP" title="L2 Cache Lookups" name="Write lookup" description="The number of L2 cache write lookups performed." units="requests" />
-        <event offset="156" advanced="yes" counter="L2_EXT_SNOOP_LOOKUP" title="L2 Cache Lookups" name="External snoop lookups" description="The number of coherency snoop lookups performed that were triggered by an external master." units="requests" />
-        <event offset="157" counter="L2_EXT_READ" title="External Bus Accesses" name="Read transaction" description="The number of external read transactions." units="transactions" />
-        <event offset="158" advanced="yes" counter="L2_EXT_READ_NOSNP" title="External Bus Accesses" name="ReadNoSnoop transactions" description="The number of external non-coherent read transactions." units="transactions" />
-        <event offset="159" advanced="yes" counter="L2_EXT_READ_UNIQUE" title="External Bus Accesses" name="ReadUnique transactions" description="The number of external coherent read unique transactions." units="transactions" />
-        <event offset="160" counter="L2_EXT_READ_BEATS" title="External Bus Beats" name="Read beat" description="The number of external bus data read cycles." units="beats" />
-        <event offset="161" counter="L2_EXT_AR_STALL" title="External Bus Stalls" name="Read stall cycles" description="The number of cycles where a read is stalled waiting for the external bus." units="cycles" />
-        <event offset="162" counter="L2_EXT_AR_CNT_Q1" title="External Bus Outstanding Reads" name="0-25% outstanding" description="The number of read transactions initiated when 0-25% of the maximum are in use." units="transactions" />
-        <event offset="163" counter="L2_EXT_AR_CNT_Q2" title="External Bus Outstanding Reads" name="25-50% outstanding" description="The number of read transactions initiated when 25-50% of the maximum are in use." units="transactions" />
-        <event offset="164" counter="L2_EXT_AR_CNT_Q3" title="External Bus Outstanding Reads" name="50-75% outstanding" description="The number of read transactions initiated when 50-75% of the maximum are in use." units="transactions" />
-        <event offset="165" counter="L2_EXT_RRESP_0_127" title="External Bus Read Latency" name="0-127 cycles" description="The number of data beats returned 0-127 cycles after the read request." units="beats" />
-        <event offset="166" counter="L2_EXT_RRESP_128_191" title="External Bus Read Latency" name="128-191 cycles" description="The number of data beats returned 128-191 cycles after the read request." units="beats" />
-        <event offset="167" counter="L2_EXT_RRESP_192_255" title="External Bus Read Latency" name="192-255 cycles" description="The number of data beats returned 192-255 cycles after the read request." units="beats" />
-        <event offset="168" counter="L2_EXT_RRESP_256_319" title="External Bus Read Latency" name="256-319 cycles" description="The number of data beats returned 256-319 cycles after the read request." units="beats" />
-        <event offset="169" counter="L2_EXT_RRESP_320_383" title="External Bus Read Latency" name="320-383 cycles" description="The number of data beats returned 320-383 cycles after the read request." units="beats" />
-        <event offset="170" counter="L2_EXT_WRITE" title="External Bus Accesses" name="Write transaction" description="The number of external write transactions." units="transactions" />
-        <event offset="171" advanced="yes" counter="L2_EXT_WRITE_NOSNP_FULL" title="External Bus Accesses" name="WriteNoSnoopFull transactions" description="The number of external non-coherent full write transactions." units="transactions" />
-        <event offset="172" advanced="yes" counter="L2_EXT_WRITE_NOSNP_PTL" title="External Bus Accesses" name="WriteNoSnoopPartial transactions" description="The number of external non-coherent partial write transactions." units="transactions" />
-        <event offset="173" advanced="yes" counter="L2_EXT_WRITE_SNP_FULL" title="External Bus Accesses" name="WriteSnoopFull transactions" description="The number of external coherent full write transactions." units="transactions" />
-        <event offset="174" advanced="yes" counter="L2_EXT_WRITE_SNP_PTL" title="External Bus Accesses" name="WriteSnoopPartial transactions" description="The number of external coherent partial write transactions." units="transactions" />
-        <event offset="175" counter="L2_EXT_WRITE_BEATS" title="External Bus Beats" name="Write beat" description="The number of external bus data write cycles." units="beats" />
-        <event offset="176" counter="L2_EXT_W_STALL" title="External Bus Stalls" name="Write stall cycles" description="The number of cycles where a write is stalled waiting for the external bus." units="cycles" />
-        <event offset="177" counter="L2_EXT_AW_CNT_Q1" title="External Bus Outstanding Writes" name="0-25% outstanding" description="The number of write transactions initiated when 0-25% of the maximum are in use." units="transactions" />
-        <event offset="178" counter="L2_EXT_AW_CNT_Q2" title="External Bus Outstanding Writes" name="25-50% outstanding" description="The number of write transactions initiated when 25-50% of the maximum are in use." units="transactions" />
-        <event offset="179" counter="L2_EXT_AW_CNT_Q3" title="External Bus Outstanding Writes" name="50-75% outstanding" description="The number of write transactions initiated when 50-75% of the maximum are in use." units="transactions" />
-        <event offset="180" advanced="yes" counter="L2_EXT_SNOOP" title="External Bus Accesses" name="Snoop transactions" description="The number of coherency snoops triggered by external masters." units="transactions" />
-        <event offset="181" advanced="yes" counter="L2_EXT_SNOOP_STALL" title="External Bus Stalls" name="Snoop stall cycles" description="The number of cycles where a coherency snoop triggered by external master is stalled." units="cycles" />
+        <event offset="4" advanced="yes" counter="MMU_REQUESTS" title="MMU Stage 1 Translations" name="MMU lookups" description="The number of main MMU address translations performed." units="requests"/>
+        <event offset="16" advanced="yes" counter="L2_RD_MSG_IN" title="L2 Cache Requests" name="Read requests" description="The number of L2 cache read requests from internal masters." units="requests"/>
+        <event offset="17" advanced="yes" counter="L2_RD_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Read stall cycles" description="The number of cycles L2 cache read requests from internal masters are stalled." units="cycles"/>
+        <event offset="18" advanced="yes" counter="L2_WR_MSG_IN" title="L2 Cache Requests" name="Write requests" description="The number of L2 cache write requests from internal masters." units="requests"/>
+        <event offset="19" advanced="yes" counter="L2_WR_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Write stall cycles" description="The number of cycles where L2 cache write requests from internal masters are stalled." units="cycles"/>
+        <event offset="20" advanced="yes" counter="L2_SNP_MSG_IN" title="L2 Cache Requests" name="Snoop requests" description="The number of L2 snoop requests from internal masters." units="requests"/>
+        <event offset="21" advanced="yes" counter="L2_SNP_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Snoop stall cycles" description="The number of cycles where L2 cache snoop requests from internal masters are stalled." units="cycles"/>
+        <event offset="22" advanced="yes" counter="L2_RD_MSG_OUT" title="L2 Cache Requests" name="L1 read requests" description="The number of L1 cache read requests sent by the L2 cache to an internal master." units="requests"/>
+        <event offset="23" advanced="yes" counter="L2_RD_MSG_OUT_STALL" title="L2 Cache Stall Cycles" name="L1 read stall cycles" description="The number of cycles where L1 cache read requests sent by the L2 cache to an internal master are stalled." units="cycles"/>
+        <event offset="24" advanced="yes" counter="L2_WR_MSG_OUT" title="L2 Cache Requests" name="L1 write requests" description="The number of L1 cache write responses sent by the L2 cache to an internal master." units="requests"/>
+        <event offset="25" counter="L2_ANY_LOOKUP" title="L2 Cache Lookups" name="Any lookup" description="The number of L2 cache lookups performed." units="requests"/>
+        <event offset="26" counter="L2_READ_LOOKUP" title="L2 Cache Lookups" name="Read lookup" description="The number of L2 cache read lookups performed." units="requests"/>
+        <event offset="27" counter="L2_WRITE_LOOKUP" title="L2 Cache Lookups" name="Write lookup" description="The number of L2 cache write lookups performed." units="requests"/>
+        <event offset="28" advanced="yes" counter="L2_EXT_SNOOP_LOOKUP" title="L2 Cache Lookups" name="External snoop lookups" description="The number of coherency snoop lookups performed that were triggered by an external master." units="requests"/>
+        <event offset="29" counter="L2_EXT_READ" title="External Bus Accesses" name="Read transaction" description="The number of external read transactions." units="transactions"/>
+        <event offset="30" advanced="yes" counter="L2_EXT_READ_NOSNP" title="External Bus Accesses" name="ReadNoSnoop transactions" description="The number of external non-coherent read transactions." units="transactions"/>
+        <event offset="31" advanced="yes" counter="L2_EXT_READ_UNIQUE" title="External Bus Accesses" name="ReadUnique transactions" description="The number of external coherent read unique transactions." units="transactions"/>
+        <event offset="32" counter="L2_EXT_READ_BEATS" title="External Bus Beats" name="Read beat" description="The number of external bus data read cycles." units="beats"/>
+        <event offset="33" counter="L2_EXT_AR_STALL" title="External Bus Stalls" name="Read stall cycles" description="The number of cycles where a read is stalled waiting for the external bus." units="cycles"/>
+        <event offset="34" counter="L2_EXT_AR_CNT_Q1" title="External Bus Outstanding Reads" name="0-25% outstanding" description="The number of read transactions initiated when 0-25% of the maximum are in use." units="transactions"/>
+        <event offset="35" counter="L2_EXT_AR_CNT_Q2" title="External Bus Outstanding Reads" name="25-50% outstanding" description="The number of read transactions initiated when 25-50% of the maximum are in use." units="transactions"/>
+        <event offset="36" counter="L2_EXT_AR_CNT_Q3" title="External Bus Outstanding Reads" name="50-75% outstanding" description="The number of read transactions initiated when 50-75% of the maximum are in use." units="transactions"/>
+        <event offset="37" counter="L2_EXT_RRESP_0_127" title="External Bus Read Latency" name="0-127 cycles" description="The number of data beats returned 0-127 cycles after the read request." units="beats"/>
+        <event offset="38" counter="L2_EXT_RRESP_128_191" title="External Bus Read Latency" name="128-191 cycles" description="The number of data beats returned 128-191 cycles after the read request." units="beats"/>
+        <event offset="39" counter="L2_EXT_RRESP_192_255" title="External Bus Read Latency" name="192-255 cycles" description="The number of data beats returned 192-255 cycles after the read request." units="beats"/>
+        <event offset="40" counter="L2_EXT_RRESP_256_319" title="External Bus Read Latency" name="256-319 cycles" description="The number of data beats returned 256-319 cycles after the read request." units="beats"/>
+        <event offset="41" counter="L2_EXT_RRESP_320_383" title="External Bus Read Latency" name="320-383 cycles" description="The number of data beats returned 320-383 cycles after the read request." units="beats"/>
+        <event offset="42" counter="L2_EXT_WRITE" title="External Bus Accesses" name="Write transaction" description="The number of external write transactions." units="transactions"/>
+        <event offset="43" advanced="yes" counter="L2_EXT_WRITE_NOSNP_FULL" title="External Bus Accesses" name="WriteNoSnoopFull transactions" description="The number of external non-coherent full write transactions." units="transactions"/>
+        <event offset="44" advanced="yes" counter="L2_EXT_WRITE_NOSNP_PTL" title="External Bus Accesses" name="WriteNoSnoopPartial transactions" description="The number of external non-coherent partial write transactions." units="transactions"/>
+        <event offset="45" advanced="yes" counter="L2_EXT_WRITE_SNP_FULL" title="External Bus Accesses" name="WriteSnoopFull transactions" description="The number of external coherent full write transactions." units="transactions"/>
+        <event offset="46" advanced="yes" counter="L2_EXT_WRITE_SNP_PTL" title="External Bus Accesses" name="WriteSnoopPartial transactions" description="The number of external coherent partial write transactions." units="transactions"/>
+        <event offset="47" counter="L2_EXT_WRITE_BEATS" title="External Bus Beats" name="Write beat" description="The number of external bus data write cycles." units="beats"/>
+        <event offset="48" counter="L2_EXT_W_STALL" title="External Bus Stalls" name="Write stall cycles" description="The number of cycles where a write is stalled waiting for the external bus." units="cycles"/>
+        <event offset="49" counter="L2_EXT_AW_CNT_Q1" title="External Bus Outstanding Writes" name="0-25% outstanding" description="The number of write transactions initiated when 0-25% of the maximum are in use." units="transactions"/>
+        <event offset="50" counter="L2_EXT_AW_CNT_Q2" title="External Bus Outstanding Writes" name="25-50% outstanding" description="The number of write transactions initiated when 25-50% of the maximum are in use." units="transactions"/>
+        <event offset="51" counter="L2_EXT_AW_CNT_Q3" title="External Bus Outstanding Writes" name="50-75% outstanding" description="The number of write transactions initiated when 50-75% of the maximum are in use." units="transactions"/>
+        <event offset="52" advanced="yes" counter="L2_EXT_SNOOP" title="External Bus Accesses" name="Snoop transactions" description="The number of coherency snoops triggered by external masters." units="transactions"/>
+        <event offset="53" advanced="yes" counter="L2_EXT_SNOOP_STALL" title="External Bus Stalls" name="Snoop stall cycles" description="The number of cycles where a coherency snoop triggered by external master is stalled." units="cycles"/>
+    </category>
+    <category name="Shader Core" per_cpu="no">
+        <event offset="4" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment active" description="The number of cycles where the shader core is processing a fragment workload." units="cycles"/>
+        <event offset="5" advanced="yes" counter="FRAG_PRIMITIVES" title="Core Primitives" name="Read primitives" description="The number of primitives read from the tile list by the fragment front-end." units="primitives"/>
+        <event offset="6" counter="FRAG_PRIM_RAST" title="Core Primitives" name="Rasterized primitives" description="The number of primitives being rasterized." units="primitives"/>
+        <event offset="7" counter="FRAG_FPK_ACTIVE" title="Core Cycles" name="Fragment FPKB active" description="The number of cycles where at least one quad is present in the pre-pipe quad queue." units="cycles"/>
+        <event offset="9" counter="FRAG_WARPS" title="Core Warps" name="Fragment warps" description="The number of fragment warps created." units="warps"/>
+        <event offset="10" counter="FRAG_PARTIAL_WARPS" title="Core Warps" name="Partial fragment warps" description="The number of fragment warps containing helper threads that do not correspond to a hit sample point." units="warps"/>
+        <event offset="11" counter="FRAG_QUADS_RAST" title="Core Quads" name="Rasterized quads" description="The number of quads generated by the rasterization phase." units="quads"/>
+        <event offset="12" counter="FRAG_QUADS_EZS_TEST" title="Core Quads" name="Early ZS tested quads" description="The number of quads that are undergoing early depth and stencil testing." units="quads"/>
+        <event offset="13" counter="FRAG_QUADS_EZS_UPDATE" title="Core Quads" name="Early ZS updated quads" description="The number of quads undergoing early depth and stencil testing, that are capable of updating the framebuffer." units="quads"/>
+        <event offset="14" counter="FRAG_QUADS_EZS_KILL" title="Core Quads" name="Early ZS killed quads" description="The number of quads killed by early depth and stencil testing." units="quads"/>
+        <event offset="15" counter="FRAG_LZS_TEST" title="Core Quads" name="Late ZS tested quads" description="The number of quads undergoing late depth and stencil testing." units="quads"/>
+        <event offset="16" counter="FRAG_LZS_KILL" title="Core Quads" name="Late ZS killed quads" description="The number of quads killed by late depth and stencil testing." units="quads"/>
+        <event offset="17" counter="WARP_REG_SIZE_64" title="Core Warps" name="All register warps" description="The number of warps that require more than 32 registers." units="warps"/>
+        <event offset="18" counter="FRAG_PTILES" title="Core Tiles" name="Tiles" description="The number of tiles processed by the shader core." units="tiles"/>
+        <event offset="19" counter="FRAG_TRANS_ELIM" title="Core Tiles" name="Constant tiles killed" description="The number of tiles killed by transaction elimination." units="tiles"/>
+        <event offset="20" counter="QUAD_FPK_KILLER" title="Core Quads" name="FPK occluder quads" description="The number of quads that are valid occluders for hidden surface removal." units="quads"/>
+        <event offset="21" counter="FULL_QUAD_WARPS" title="Core Warps" name="Full quad warps" description="The number of warps that are fully populated with quads." units="warps"/>
+        <event offset="22" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Non-fragment active" description="The number of cycles where the shader core is processing some non-fragment workload." units="cycles"/>
+        <event offset="23" advanced="yes" counter="COMPUTE_TASKS" title="Core Tasks" name="Non-fragment tasks" description="The number of non-fragment tasks issued to the shader core." units="tasks"/>
+        <event offset="24" counter="COMPUTE_WARPS" title="Core Warps" name="Non-fragment warps" description="The number of non-fragment warps created." units="warps"/>
+        <event offset="25" advanced="yes" counter="COMPUTE_STARVING" title="Core Starvation Cycles" name="Non-fragment starvation cycles" description="The number of cycles where the shader core is processing a non-fragment workload and there are no new threads available for execution." units="cycles"/>
+        <event offset="26" counter="EXEC_CORE_ACTIVE" title="Core Cycles" name="Execution core active" description="The number of cycles where the shader core is processing at least one warp." units="cycles"/>
+        <event offset="27" advanced="yes" counter="EXEC_ACTIVE" title="Core Cycles" name="Execution engine active" description="The number of cycles where the execution engine unit is processing at least one thread." units="cycles"/>
+        <event offset="28" counter="EXEC_INSTR_COUNT" title="Core EE Instructions" name="Executed instructions" description="The number of instructions executed per warp." units="instructions"/>
+        <event offset="29" counter="EXEC_INSTR_DIVERGED" title="Core EE Instructions" name="Diverged instructions" description="The number of instructions executed per warp, that have control flow divergence." units="instructions"/>
+        <event offset="30" advanced="yes" counter="EXEC_INSTR_STARVING" title="Core Starvation Cycles" name="Execution engine starvation cycles" description="The number of cycles where no new threads are available for execution." units="cycles"/>
+        <event offset="31" advanced="yes" counter="ARITH_INSTR_SINGLE_FMA" title="Core EE Instructions" name="Arithmetic instructions" description="The number of instructions where the workload is a single FMA pipe arithmetic operation." units="instructions"/>
+        <event offset="32" advanced="yes" counter="ARITH_INSTR_DOUBLE" title="Core EE Instructions" name="Dual Arithmetic instructions" description="The number of instructions where the workload is one FMA pipe arithmetic operation and one ADD pipe arithmetic operation." units="instructions"/>
+        <event offset="33" advanced="yes" counter="ARITH_INSTR_MSG" title="Core EE Instructions" name="Arithmetic + Message instructions" description="The number of instructions where the workload is one FMA pipe arithmetic operation and one ADD pipe message operation" units="instructions"/>
+        <event offset="34" advanced="yes" counter="ARITH_INSTR_MSG_ONLY" title="Core EE Instructions" name="Message instructions" description="The number of instructions where the workload is a single ADD pipe message operation, with no FMA pipe operation" units="instructions"/>
+        <event offset="35" counter="TEX_MSGI_NUM_QUADS" title="Core Texture Quads" name="Texture requests" description="The number of quad-width texture operations processed by the texture unit." units="quads"/>
+        <event offset="36" counter="TEX_DFCH_NUM_PASSES" title="Core Texture Quads" name="Texture issues" description="The number of quad-width filtering passes." units="issues"/>
+        <event offset="37" counter="TEX_DFCH_NUM_PASSES_MISS" title="Core Texture Quads" name="Descriptor misses" description="The number of quad-width filtering passes that miss in the resource or sampler descriptor cache." units="requests"/>
+        <event offset="38" counter="TEX_DFCH_NUM_PASSES_MIP_MAP" title="Core Texture Quads" name="Mipmapped texture issues" description="The number of quad-width filtering passes that use a mipmapped texture." units="issues"/>
+        <event offset="39" counter="TEX_TIDX_NUM_SPLIT_MIP_MAP" title="Core Texture Quads" name="Trilinear filtered issues" description="The number of quad-width filtering passes that use a trilinear filter." units="issues"/>
+        <event offset="40" counter="TEX_TFCH_NUM_LINES_FETCHED" title="Core Texture Line Fetches" name="Line fetches" description="The number of texture line fetches from the L2 cache." units="issues"/>
+        <event offset="41" counter="TEX_TFCH_NUM_LINES_FETCHED_BLOCK_COMPRESSED" title="Core Texture Line Fetches" name="Compressed line fetches" description="The number of texture line fetches from the L2 cache that are block compressed textures." units="issues"/>
+        <event offset="42" counter="TEX_TFCH_NUM_OPERATIONS" title="Core Texture Cycles" name="Cache lookups" description="The number of texture cache lookup cycles." units="requests"/>
+        <event offset="43" counter="TEX_FILT_NUM_OPERATIONS" title="Core Texture Cycles" name="Texturing active" description="The number of texture filtering issue cycles." units="cycles"/>
+        <event offset="44" counter="LS_MEM_READ_FULL" title="Core Load/Store Cycles" name="Full read cycles" description="The number of full-width load/store cache reads." units="cycles"/>
+        <event offset="45" counter="LS_MEM_READ_SHORT" title="Core Load/Store Cycles" name="Partial read cycles" description="The number of partial-width load/store cache reads." units="cycles"/>
+        <event offset="46" counter="LS_MEM_WRITE_FULL" title="Core Load/Store Cycles" name="Full write cycles" description="The number of full-width load/store cache writes." units="cycles"/>
+        <event offset="47" counter="LS_MEM_WRITE_SHORT" title="Core Load/Store Cycles" name="Partial write cycles" description="The number of partial-width load/store cache writes." units="cycles"/>
+        <event offset="48" counter="LS_MEM_ATOMIC" title="Core Load/Store Cycles" name="Atomic access cycles" description="The number of load/store atomic accesses." units="cycles"/>
+        <event offset="49" counter="VARY_INSTR" title="Core Varying Requests" name="Interpolation requests" description="The number of warp-width interpolation operations processed by the varying unit." units="instructions"/>
+        <event offset="50" counter="VARY_SLOT_32" title="Core Varying Cycles" name="32-bit interpolation active" description="The number of 32-bit interpolation cycles processed by the varying unit." units="cycles"/>
+        <event offset="51" counter="VARY_SLOT_16" title="Core Varying Cycles" name="16-bit interpolation active" description="The number of 16-bit interpolation cycles processed by the varying unit." units="cycles"/>
+        <event offset="52" advanced="yes" counter="ATTR_INSTR" title="Core Attribute Requests" name="Attribute requests" description="The number of instructions executed by the attribute unit." units="instructions"/>
+        <event offset="53" advanced="yes" counter="ARITH_INSTR_FP_MUL" title="Core EE Instructions" name="Multiplier instructions" description="The number of instructions where the workload uses floating-point multiplier hardware." units="instructions"/>
+        <event offset="54" counter="BEATS_RD_FTC" title="Core L2 Reads" name="Fragment L2 read beats" description="The number of read beats received by the fixed-function fragment front-end." units="beats"/>
+        <event offset="55" counter="BEATS_RD_FTC_EXT" title="Core L2 Reads" name="Fragment external read beats" description="The number of read beats received by the fixed-function fragment front-end that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="56" counter="BEATS_RD_LSC" title="Core L2 Reads" name="Load/store L2 read beats" description="The number of read beats received by the load/store unit." units="beats"/>
+        <event offset="57" counter="BEATS_RD_LSC_EXT" title="Core L2 Reads" name="Load/store external read beats" description="The number of read beats received by the load/store unit that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="58" counter="BEATS_RD_TEX" title="Core L2 Reads" name="Texture L2 read beats" description="The number of read beats received by the texture unit." units="beats"/>
+        <event offset="59" counter="BEATS_RD_TEX_EXT" title="Core L2 Reads" name="Texture external read beats" description="The number of read beats received by the texture unit that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="60" advanced="yes" counter="BEATS_RD_OTHER" title="Core L2 Reads" name="Other L2 read beats" description="The number of read beats received by a unit that is not specifically identified." units="beats"/>
+        <event offset="61" counter="BEATS_WR_LSC_WB" title="Core Writes" name="Load/store writeback write beats" description="The number of write beats by the load/store unit that are due to writeback." units="beats"/>
+        <event offset="62" counter="BEATS_WR_TIB" title="Core Writes" name="Tile buffer write beats" description="The number of write beats sent by the tile buffer writeback unit." units="beats"/>
+        <event offset="63" counter="BEATS_WR_LSC_OTHER" title="Core Writes" name="Load/store other write beats" description="The number of write beats by the load/store unit that are due to any reason other than writeback." units="beats"/>
     </category>
 </metrics>
\ No newline at end of file
diff --git a/src/panfrost/perf/G57.xml b/src/panfrost/perf/G57.xml
index 20f9869cd93..e696d1702a0 100644
--- a/src/panfrost/perf/G57.xml
+++ b/src/panfrost/perf/G57.xml
@@ -1,6 +1,6 @@
 <!--
 Copyright © 2017-2020 ARM Limited.
-Copyright © 2021 Collabora, Ltd.
+Copyright © 2021-2022 Collabora, Ltd.
 Author: Antonio Caggiano <antonio.caggiano@collabora.com>
 
 Permission is hereby granted, free of charge, to any person obtaining a
@@ -24,156 +24,156 @@ SOFTWARE.
 -->
 <metrics id="TNAx">
     <category name="Job Manager" per_cpu="no">
-        <event offset="6" counter="GPU_ACTIVE" title="GPU Cycles" name="GPU active" description="The number of cycles where the GPU has a workload of any type queued for processing." units="cycles" />
-        <event offset="7" advanced="yes" counter="IRQ_ACTIVE" title="GPU Cycles" name="Interrupt active" description="The number of cycles where the GPU has a pending interrupt." units="cycles" />
-        <event offset="8" advanced="yes" counter="JS0_JOBS" title="GPU Jobs" name="Fragment jobs" description="The number of jobs processed by the GPU fragment queue." units="jobs" />
-        <event offset="9" counter="JS0_TASKS" title="GPU Tasks" name="Fragment tasks" description="The number of 32x32 pixel tasks processed by the GPU fragment queue." units="tasks" />
-        <event offset="10" counter="JS0_ACTIVE" title="GPU Cycles" name="Fragment queue active" description="The number of cycles where work is queued for processing in the GPU fragment queue." units="cycles" />
-        <event offset="12" advanced="yes" counter="JS0_WAIT_READ" title="GPU Wait Cycles" name="Fragment descriptor reads cycles" description="The number of cycles where queued fragment work is waiting for a descriptor load." units="cycles" />
-        <event offset="13" advanced="yes" counter="JS0_WAIT_ISSUE" title="GPU Wait Cycles" name="Fragment job issue cycles" description="The number of cycles where queued fragment work is waiting for an available processor." units="cycles" />
-        <event offset="14" advanced="yes" counter="JS0_WAIT_DEPEND" title="GPU Wait Cycles" name="Fragment job dependency cycles" description="The number of cycles where queued fragment work is waiting for dependent work to complete." units="cycles" />
-        <event offset="15" advanced="yes" counter="JS0_WAIT_FINISH" title="GPU Wait Cycles" name="Fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued fragment work to complete." units="cycles" />
-        <event offset="16" advanced="yes" counter="JS1_JOBS" title="GPU Jobs" name="Non-fragment jobs" description="The number of jobs processed by the GPU non-fragment queue." units="jobs" />
-        <event offset="17" advanced="yes" counter="JS1_TASKS" title="GPU Tasks" name="Non-fragment tasks" description="The number of tasks processed by the GPU non-fragment queue." units="tasks" />
-        <event offset="18" counter="JS1_ACTIVE" title="GPU Cycles" name="Non-fragment queue active" description="The number of cycles where work is queued in the GPU non-fragment queue." units="cycles" />
-        <event offset="20" advanced="yes" counter="JS1_WAIT_READ" title="GPU Wait Cycles" name="Non-fragment descriptor read cycles" description="The number number of cycles where queued non-fragment work is waiting for a descriptor load." units="cycles" />
-        <event offset="21" advanced="yes" counter="JS1_WAIT_ISSUE" title="GPU Wait Cycles" name="Non-fragment job issue cycles" description="The number of cycles where queued non-fragment work is waiting for an available processor." units="cycles" />
-        <event offset="22" advanced="yes" counter="JS1_WAIT_DEPEND" title="GPU Wait Cycles" name="Non-fragment job dependency cycles" description="The number of cycles where queued non-fragment work is waiting for dependent work to complete." units="cycles" />
-        <event offset="23" advanced="yes" counter="JS1_WAIT_FINISH" title="GPU Wait Cycles" name="Non-fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued non-fragment work to complete." units="cycles" />
-        <event offset="24" advanced="yes" counter="JS2_JOBS" title="GPU Jobs" name="Reserved jobs" description="The number of jobs processed by the GPU reserved queue." units="jobs" />
-        <event offset="25" advanced="yes" counter="JS2_TASKS" title="GPU Tasks" name="Reserved tasks" description="The number of tasks processed by the GPU reserved queue." units="tasks" />
-        <event offset="26" advanced="yes" counter="JS2_ACTIVE" title="GPU Cycles" name="Reserved queue active" description="The number of cycles where work is queued in the GPU reserved queue." units="cycles" />
-        <event offset="28" advanced="yes" counter="JS2_WAIT_READ" title="GPU Wait Cycles" name="Reserved descriptor read cycles" description="The number of cycles where queued reserved work is waiting for a descriptor load." units="cycles" />
-        <event offset="29" advanced="yes" counter="JS2_WAIT_ISSUE" title="GPU Wait Cycles" name="Reserved job issue cycles" description="The number of cycles where queued reserved work is waiting for an available processor." units="cycles" />
-        <event offset="30" advanced="yes" counter="JS2_WAIT_DEPEND" title="GPU Wait Cycles" name="Reserved job dependency cycles" description="The number of cycles where queued reserved work is waiting for dependent work to complete." units="cycles" />
-        <event offset="31" advanced="yes" counter="JS2_WAIT_FINISH" title="GPU Wait Cycles" name="Reserved job finish cycles" description="The number of cycles where the GPU is waiting for issued reserved work to complete." units="cycles" />
-    </category>
-    <category name="Memory System" per_cpu="no">
-        <event offset="196" advanced="yes" counter="MMU_REQUESTS" title="MMU Stage 1 Translations" name="MMU lookups" description="The number of main MMU address translations performed." units="requests" />
-        <event offset="208" advanced="yes" counter="L2_RD_MSG_IN" title="L2 Cache Requests" name="Read requests" description="The number of L2 cache read requests from internal masters." units="requests" />
-        <event offset="209" advanced="yes" counter="L2_RD_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Read stall cycles" description="The number of cycles L2 cache read requests from internal masters are stalled." units="cycles" />
-        <event offset="210" advanced="yes" counter="L2_WR_MSG_IN" title="L2 Cache Requests" name="Write requests" description="The number of L2 cache write requests from internal masters." units="requests" />
-        <event offset="211" advanced="yes" counter="L2_WR_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Write stall cycles" description="The number of cycles where L2 cache write requests from internal masters are stalled." units="cycles" />
-        <event offset="212" advanced="yes" counter="L2_SNP_MSG_IN" title="L2 Cache Requests" name="Snoop requests" description="The number of L2 snoop requests from internal masters." units="requests" />
-        <event offset="213" advanced="yes" counter="L2_SNP_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Snoop stall cycles" description="The number of cycles where L2 cache snoop requests from internal masters are stalled." units="cycles" />
-        <event offset="214" advanced="yes" counter="L2_RD_MSG_OUT" title="L2 Cache Requests" name="L1 read requests" description="The number of L1 cache read requests sent by the L2 cache to an internal master." units="requests" />
-        <event offset="215" advanced="yes" counter="L2_RD_MSG_OUT_STALL" title="L2 Cache Stall Cycles" name="L1 read stall cycles" description="The number of cycles where L1 cache read requests sent by the L2 cache to an internal master are stalled." units="cycles" />
-        <event offset="216" advanced="yes" counter="L2_WR_MSG_OUT" title="L2 Cache Requests" name="L1 write requests" description="The number of L1 cache write responses sent by the L2 cache to an internal master." units="requests" />
-        <event offset="217" counter="L2_ANY_LOOKUP" title="L2 Cache Lookups" name="Any lookup" description="The number of L2 cache lookups performed." units="requests" />
-        <event offset="218" counter="L2_READ_LOOKUP" title="L2 Cache Lookups" name="Read lookup" description="The number of L2 cache read lookups performed." units="requests" />
-        <event offset="219" counter="L2_WRITE_LOOKUP" title="L2 Cache Lookups" name="Write lookup" description="The number of L2 cache write lookups performed." units="requests" />
-        <event offset="220" advanced="yes" counter="L2_EXT_SNOOP_LOOKUP" title="L2 Cache Lookups" name="External snoop lookups" description="The number of coherency snoop lookups performed that were triggered by an external master." units="requests" />
-        <event offset="221" counter="L2_EXT_READ" title="External Bus Accesses" name="Read transaction" description="The number of external read transactions." units="transactions" />
-        <event offset="222" advanced="yes" counter="L2_EXT_READ_NOSNP" title="External Bus Accesses" name="ReadNoSnoop transactions" description="The number of external non-coherent read transactions." units="transactions" />
-        <event offset="223" advanced="yes" counter="L2_EXT_READ_UNIQUE" title="External Bus Accesses" name="ReadUnique transactions" description="The number of external coherent read unique transactions." units="transactions" />
-        <event offset="224" counter="L2_EXT_READ_BEATS" title="External Bus Beats" name="Read beat" description="The number of external bus data read cycles." units="beats" />
-        <event offset="225" counter="L2_EXT_AR_STALL" title="External Bus Stalls" name="Read stall cycles" description="The number of cycles where a read is stalled waiting for the external bus." units="cycles" />
-        <event offset="226" counter="L2_EXT_AR_CNT_Q1" title="External Bus Outstanding Reads" name="0-25% outstanding" description="The number of read transactions initiated when 0-25% of the maximum are in use." units="transactions" />
-        <event offset="227" counter="L2_EXT_AR_CNT_Q2" title="External Bus Outstanding Reads" name="25-50% outstanding" description="The number of read transactions initiated when 25-50% of the maximum are in use." units="transactions" />
-        <event offset="228" counter="L2_EXT_AR_CNT_Q3" title="External Bus Outstanding Reads" name="50-75% outstanding" description="The number of read transactions initiated when 50-75% of the maximum are in use." units="transactions" />
-        <event offset="229" counter="L2_EXT_RRESP_0_127" title="External Bus Read Latency" name="0-127 cycles" description="The number of data beats returned 0-127 cycles after the read request." units="beats" />
-        <event offset="230" counter="L2_EXT_RRESP_128_191" title="External Bus Read Latency" name="128-191 cycles" description="The number of data beats returned 128-191 cycles after the read request." units="beats" />
-        <event offset="231" counter="L2_EXT_RRESP_192_255" title="External Bus Read Latency" name="192-255 cycles" description="The number of data beats returned 192-255 cycles after the read request." units="beats" />
-        <event offset="232" counter="L2_EXT_RRESP_256_319" title="External Bus Read Latency" name="256-319 cycles" description="The number of data beats returned 256-319 cycles after the read request." units="beats" />
-        <event offset="233" counter="L2_EXT_RRESP_320_383" title="External Bus Read Latency" name="320-383 cycles" description="The number of data beats returned 320-383 cycles after the read request." units="beats" />
-        <event offset="234" counter="L2_EXT_WRITE" title="External Bus Accesses" name="Write transaction" description="The number of external write transactions." units="transactions" />
-        <event offset="235" advanced="yes" counter="L2_EXT_WRITE_NOSNP_FULL" title="External Bus Accesses" name="WriteNoSnoopFull transactions" description="The number of external non-coherent full write transactions." units="transactions" />
-        <event offset="236" advanced="yes" counter="L2_EXT_WRITE_NOSNP_PTL" title="External Bus Accesses" name="WriteNoSnoopPartial transactions" description="The number of external non-coherent partial write transactions." units="transactions" />
-        <event offset="237" advanced="yes" counter="L2_EXT_WRITE_SNP_FULL" title="External Bus Accesses" name="WriteSnoopFull transactions" description="The number of external coherent full write transactions." units="transactions" />
-        <event offset="238" advanced="yes" counter="L2_EXT_WRITE_SNP_PTL" title="External Bus Accesses" name="WriteSnoopPartial transactions" description="The number of external coherent partial write transactions." units="transactions" />
-        <event offset="239" counter="L2_EXT_WRITE_BEATS" title="External Bus Beats" name="Write beat" description="The number of external bus data write cycles." units="beats" />
-        <event offset="240" counter="L2_EXT_W_STALL" title="External Bus Stalls" name="Write stall cycles" description="The number of cycles where a write is stalled waiting for the external bus." units="cycles" />
-        <event offset="241" counter="L2_EXT_AW_CNT_Q1" title="External Bus Outstanding Writes" name="0-25% outstanding" description="The number of write transactions initiated when 0-25% of the maximum are in use." units="transactions" />
-        <event offset="242" counter="L2_EXT_AW_CNT_Q2" title="External Bus Outstanding Writes" name="25-50% outstanding" description="The number of write transactions initiated when 25-50% of the maximum are in use." units="transactions" />
-        <event offset="243" counter="L2_EXT_AW_CNT_Q3" title="External Bus Outstanding Writes" name="50-75% outstanding" description="The number of write transactions initiated when 50-75% of the maximum are in use." units="transactions" />
-        <event offset="244" advanced="yes" counter="L2_EXT_SNOOP" title="External Bus Accesses" name="Snoop transactions" description="The number of coherency snoops triggered by external masters." units="transactions" />
-        <event offset="245" advanced="yes" counter="L2_EXT_SNOOP_STALL" title="External Bus Stalls" name="Snoop stall cycles" description="The number of cycles where a coherency snoop triggered by external master is stalled." units="cycles" />
-    </category>
-    <category name="Shader Core" per_cpu="no">
-        <event offset="196" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment active" description="The number of cycles where the shader core is processing a fragment workload." units="cycles" />
-        <event offset="197" advanced="yes" counter="FRAG_PRIMITIVES_OUT" title="Core Primitives" name="Read primitives" description="The number of primitives read from the tile list by the fragment front-end." units="primitives" />
-        <event offset="198" counter="FRAG_PRIM_RAST" title="Core Primitives" name="Rasterized primitives" description="The number of primitives being rasterized." units="primitives" />
-        <event offset="199" counter="FRAG_FPK_ACTIVE" title="Core Cycles" name="Fragment FPKB active" description="The number of cycles where at least one quad is present in the pre-pipe quad queue." units="cycles" />
-        <event offset="201" counter="FRAG_WARPS" title="Core Warps" name="Fragment warps" description="The number of fragment warps created." units="warps" />
-        <event offset="202" counter="FRAG_PARTIAL_QUADS_RAST" title="Core Quads" name="Partial rasterized quads" description="The number of partially-rasterized fragment quads created." units="quads" />
-        <event offset="203" counter="FRAG_QUADS_RAST" title="Core Quads" name="Rasterized quads" description="The number of quads generated by the rasterization phase." units="quads" />
-        <event offset="204" counter="FRAG_QUADS_EZS_TEST" title="Core Quads" name="Early ZS tested quads" description="The number of quads that are undergoing early depth and stencil testing." units="quads" />
-        <event offset="205" counter="FRAG_QUADS_EZS_UPDATE" title="Core Quads" name="Early ZS updated quads" description="The number of quads undergoing early depth and stencil testing, that are capable of updating the framebuffer." units="quads" />
-        <event offset="206" counter="FRAG_QUADS_EZS_KILL" title="Core Quads" name="Early ZS killed quads" description="The number of quads killed by early depth and stencil testing." units="quads" />
-        <event offset="207" counter="FRAG_LZS_TEST" title="Core Quads" name="Late ZS tested quads" description="The number of quads undergoing late depth and stencil testing." units="quads" />
-        <event offset="208" counter="FRAG_LZS_KILL" title="Core Quads" name="Late ZS killed quads" description="The number of quads killed by late depth and stencil testing." units="quads" />
-        <event offset="209" counter="WARP_REG_SIZE_64" title="Core Warps" name="All register warps" description="The number of warps that require more than 32 registers." units="warps" />
-        <event offset="210" counter="FRAG_PTILES" title="Core Tiles" name="Tiles" description="The number of tiles processed by the shader core." units="tiles" />
-        <event offset="211" counter="FRAG_TRANS_ELIM" title="Core Tiles" name="Constant tiles killed" description="The number of tiles killed by transaction elimination." units="tiles" />
-        <event offset="212" counter="QUAD_FPK_KILLER" title="Core Quads" name="FPK occluder quads" description="The number of quads that are valid occluders for hidden surface removal." units="quads" />
-        <event offset="213" counter="FULL_QUAD_WARPS" title="Core Warps" name="Full quad warps" description="The number of warps that are fully populated with quads." units="warps" />
-        <event offset="214" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Non-fragment active" description="The number of cycles where the shader core is processing some non-fragment workload." units="cycles" />
-        <event offset="215" advanced="yes" counter="COMPUTE_TASKS" title="Core Tasks" name="Non-fragment tasks" description="The number of non-fragment tasks issued to the shader core." units="tasks" />
-        <event offset="216" counter="COMPUTE_WARPS" title="Core Warps" name="Non-fragment warps" description="The number of non-fragment warps created." units="warps" />
-        <event offset="217" advanced="yes" counter="COMPUTE_STARVING" title="Core Starvation Cycles" name="Non-fragment starvation cycles" description="The number of cycles where the shader core is processing a non-fragment workload and there are no new threads available for execution." units="cycles" />
-        <event offset="218" counter="EXEC_CORE_ACTIVE" title="Core Cycles" name="Execution core active" description="The number of cycles where the shader core is processing at least one warp." units="cycles" />
-        <event offset="219" counter="EXEC_INSTR_FMA" title="Core PU Instructions" name="FMA instructions" description="The number of instructions issued to the FMA pipe." units="instructions" />
-        <event offset="220" counter="EXEC_INSTR_CVT" title="Core PU Instructions" name="CVT instructions" description="The number of instructions issued to the CVT pipe." units="instructions" />
-        <event offset="221" counter="EXEC_INSTR_SFU" title="Core PU Instructions" name="SFU instructions" description="The number of instructions issued to the SFU pipe." units="instructions" />
-        <event offset="222" counter="EXEC_INSTR_MSG" title="Core PU Instructions" name="Message instructions" description="The number of instructions issued to the MSG pipe." units="instructions" />
-        <event offset="223" counter="EXEC_INSTR_DIVERGED" title="Core EE Instructions" name="Diverged instructions" description="The number of instructions executed per warp, that have control flow divergence." units="instructions" />
-        <event offset="224" advanced="yes" counter="EXEC_ICACHE_MISS" title="Core PU Instructions" name="Instruction cache misses" description="The number of instruction cache misses." units="requests" />
-        <event offset="225" advanced="yes" counter="EXEC_STARVE_ARITH" title="Core Starvation Cycles" name="Execution engine starvation cycles" description="The number of cycles where the processing unit is starved of work." units="cycles" />
-        <event offset="226" counter="CALL_BLEND_SHADER" title="Core PU Instructions" name="Blend shader calls" description="The number of blend shader invocations executed." units="instructions" />
-        <event offset="227" counter="TEX_MSGI_NUM_FLITS" title="Texture Bus" name="Input beats" description="The number of texture request message data beats." units="beats" />
-        <event offset="228" counter="TEX_DFCH_CLK_STALLED" title="Core Texture Stalls" name="Descriptor stall cycles" description="The number of cycles where a quad is stalled on texture descriptor fetch." units="cycles" />
-        <event offset="229" counter="TEX_TFCH_CLK_STALLED" title="Core Texture Stalls" name="Fetch queue stall cycles" description="The number of cycles where a quad is stalled on entering texture fetch because the fetch queue is full." units="cycles" />
-        <event offset="230" counter="TEX_TFCH_STARVED_PENDING_DATA_FETCH" title="Core Texture Stalls" name="Filtering unit stall cycles" description="The number of cycles where the filtering unit is idle and there is at least one quad present in the texture data fetch queue." units="cycles" />
-        <event offset="231" counter="TEX_FILT_NUM_OPERATIONS" title="Core Texture Cycles" name="Texturing active" description="The number of texture filtering issue cycles." units="cycles" />
-        <event offset="232" counter="TEX_FILT_NUM_FXR_OPERATIONS" title="Core Texture Cycles" name="4x bilinear filtering active" description="The number of cycles where the filtering unit uses the 4x path to implement nearest or bilinear filtering." units="cycles" />
-        <event offset="233" counter="TEX_FILT_NUM_FST_OPERATIONS" title="Core Texture Cycles" name="2x trilinear filtering active" description="The number of cycles where the filtering unit uses the 4x path to implement trilinear filtering." units="cycles" />
-        <event offset="234" counter="TEX_MSGO_NUM_MSG" title="Core Texture Quads" name="Texture requests" description="The number of quad-width texture operations processed by the texture unit." units="quads" />
-        <event offset="235" counter="TEX_MSGO_NUM_FLITS" title="Texture Bus" name="Output beats" description="The number of texture response message data beats." units="beats" />
-        <event offset="236" counter="LS_MEM_READ_FULL" title="Core Load/Store Cycles" name="Full read cycles" description="The number of full-width load/store cache reads." units="cycles" />
-        <event offset="237" counter="LS_MEM_READ_SHORT" title="Core Load/Store Cycles" name="Partial read cycles" description="The number of partial-width load/store cache reads." units="cycles" />
-        <event offset="238" counter="LS_MEM_WRITE_FULL" title="Core Load/Store Cycles" name="Full write cycles" description="The number of full-width load/store cache writes." units="cycles" />
-        <event offset="239" counter="LS_MEM_WRITE_SHORT" title="Core Load/Store Cycles" name="Partial write cycles" description="The number of partial-width load/store cache writes." units="cycles" />
-        <event offset="240" counter="LS_MEM_ATOMIC" title="Core Load/Store Cycles" name="Atomic access cycles" description="The number of load/store atomic accesses." units="cycles" />
-        <event offset="241" counter="VARY_INSTR" title="Core Varying Requests" name="Interpolation requests" description="The number of warp-width interpolation operations processed by the varying unit." units="instructions" />
-        <event offset="242" counter="VARY_SLOT_32" title="Core Varying Cycles" name="32-bit interpolation active" description="The number of 32-bit interpolation cycles processed by the varying unit." units="cycles" />
-        <event offset="243" counter="VARY_SLOT_16" title="Core Varying Cycles" name="16-bit interpolation active" description="The number of 16-bit interpolation cycles processed by the varying unit." units="cycles" />
-        <event offset="244" advanced="yes" counter="ATTR_INSTR" title="Core Attribute Requests" name="Attribute requests" description="The number of instructions executed by the attribute unit." units="instructions" />
-        <event offset="245" advanced="yes" counter="ARITH_INSTR_FP_MUL" title="Core EE Instructions" name="Multiplier instructions" description="The number of instructions where the workload uses floating-point multiplier hardware." units="instructions" />
-        <event offset="246" counter="BEATS_RD_FTC" title="Core L2 Reads" name="Fragment L2 read beats" description="The number of read beats received by the fixed-function fragment front-end." units="beats" />
-        <event offset="247" counter="BEATS_RD_FTC_EXT" title="Core L2 Reads" name="Fragment external read beats" description="The number of read beats received by the fixed-function fragment front-end that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="248" counter="BEATS_RD_LSC" title="Core L2 Reads" name="Load/store L2 read beats" description="The number of read beats received by the load/store unit." units="beats" />
-        <event offset="249" counter="BEATS_RD_LSC_EXT" title="Core L2 Reads" name="Load/store external read beats" description="The number of read beats received by the load/store unit that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="250" counter="BEATS_RD_TEX" title="Core L2 Reads" name="Texture L2 read beats" description="The number of read beats received by the texture unit." units="beats" />
-        <event offset="251" counter="BEATS_RD_TEX_EXT" title="Core L2 Reads" name="Texture external read beats" description="The number of read beats received by the texture unit that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="252" advanced="yes" counter="BEATS_RD_OTHER" title="Core L2 Reads" name="Other L2 read beats" description="The number of read beats received by a unit that is not specifically identified." units="beats" />
-        <event offset="253" counter="BEATS_WR_LSC_OTHER" title="Core Writes" name="Load/store other write beats" description="The number of write beats by the load/store unit that are due to any reason other than writeback." units="beats" />
-        <event offset="254" counter="BEATS_WR_TIB" title="Core Writes" name="Tile buffer write beats" description="The number of write beats sent by the tile buffer writeback unit." units="beats" />
-        <event offset="255" counter="BEATS_WR_LSC_WB" title="Core Writes" name="Load/store writeback write beats" description="The number of write beats by the load/store unit that are due to writeback." units="beats" />
+        <event offset="6" counter="GPU_ACTIVE" title="GPU Cycles" name="GPU active" description="The number of cycles where the GPU has a workload of any type queued for processing." units="cycles"/>
+        <event offset="7" advanced="yes" counter="IRQ_ACTIVE" title="GPU Cycles" name="Interrupt active" description="The number of cycles where the GPU has a pending interrupt." units="cycles"/>
+        <event offset="8" advanced="yes" counter="JS0_JOBS" title="GPU Jobs" name="Fragment jobs" description="The number of jobs processed by the GPU fragment queue." units="jobs"/>
+        <event offset="9" counter="JS0_TASKS" title="GPU Tasks" name="Fragment tasks" description="The number of 32x32 pixel tasks processed by the GPU fragment queue." units="tasks"/>
+        <event offset="10" counter="JS0_ACTIVE" title="GPU Cycles" name="Fragment queue active" description="The number of cycles where work is queued for processing in the GPU fragment queue." units="cycles"/>
+        <event offset="12" advanced="yes" counter="JS0_WAIT_READ" title="GPU Wait Cycles" name="Fragment descriptor reads cycles" description="The number of cycles where queued fragment work is waiting for a descriptor load." units="cycles"/>
+        <event offset="13" advanced="yes" counter="JS0_WAIT_ISSUE" title="GPU Wait Cycles" name="Fragment job issue cycles" description="The number of cycles where queued fragment work is waiting for an available processor." units="cycles"/>
+        <event offset="14" advanced="yes" counter="JS0_WAIT_DEPEND" title="GPU Wait Cycles" name="Fragment job dependency cycles" description="The number of cycles where queued fragment work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="15" advanced="yes" counter="JS0_WAIT_FINISH" title="GPU Wait Cycles" name="Fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued fragment work to complete." units="cycles"/>
+        <event offset="16" advanced="yes" counter="JS1_JOBS" title="GPU Jobs" name="Non-fragment jobs" description="The number of jobs processed by the GPU non-fragment queue." units="jobs"/>
+        <event offset="17" advanced="yes" counter="JS1_TASKS" title="GPU Tasks" name="Non-fragment tasks" description="The number of tasks processed by the GPU non-fragment queue." units="tasks"/>
+        <event offset="18" counter="JS1_ACTIVE" title="GPU Cycles" name="Non-fragment queue active" description="The number of cycles where work is queued in the GPU non-fragment queue." units="cycles"/>
+        <event offset="20" advanced="yes" counter="JS1_WAIT_READ" title="GPU Wait Cycles" name="Non-fragment descriptor read cycles" description="The number number of cycles where queued non-fragment work is waiting for a descriptor load." units="cycles"/>
+        <event offset="21" advanced="yes" counter="JS1_WAIT_ISSUE" title="GPU Wait Cycles" name="Non-fragment job issue cycles" description="The number of cycles where queued non-fragment work is waiting for an available processor." units="cycles"/>
+        <event offset="22" advanced="yes" counter="JS1_WAIT_DEPEND" title="GPU Wait Cycles" name="Non-fragment job dependency cycles" description="The number of cycles where queued non-fragment work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="23" advanced="yes" counter="JS1_WAIT_FINISH" title="GPU Wait Cycles" name="Non-fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued non-fragment work to complete." units="cycles"/>
+        <event offset="24" advanced="yes" counter="JS2_JOBS" title="GPU Jobs" name="Reserved jobs" description="The number of jobs processed by the GPU reserved queue." units="jobs"/>
+        <event offset="25" advanced="yes" counter="JS2_TASKS" title="GPU Tasks" name="Reserved tasks" description="The number of tasks processed by the GPU reserved queue." units="tasks"/>
+        <event offset="26" advanced="yes" counter="JS2_ACTIVE" title="GPU Cycles" name="Reserved queue active" description="The number of cycles where work is queued in the GPU reserved queue." units="cycles"/>
+        <event offset="28" advanced="yes" counter="JS2_WAIT_READ" title="GPU Wait Cycles" name="Reserved descriptor read cycles" description="The number of cycles where queued reserved work is waiting for a descriptor load." units="cycles"/>
+        <event offset="29" advanced="yes" counter="JS2_WAIT_ISSUE" title="GPU Wait Cycles" name="Reserved job issue cycles" description="The number of cycles where queued reserved work is waiting for an available processor." units="cycles"/>
+        <event offset="30" advanced="yes" counter="JS2_WAIT_DEPEND" title="GPU Wait Cycles" name="Reserved job dependency cycles" description="The number of cycles where queued reserved work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="31" advanced="yes" counter="JS2_WAIT_FINISH" title="GPU Wait Cycles" name="Reserved job finish cycles" description="The number of cycles where the GPU is waiting for issued reserved work to complete." units="cycles"/>
     </category>
     <category name="Tiler" per_cpu="no">
-        <event offset="4" counter="TILER_ACTIVE" title="GPU Cycles" name="Tiler active" description="The number of cycles where the tiler has a workload queued for processing." units="cycles" />
-        <event offset="6" counter="TRIANGLES" title="Input Primitives" name="Triangle primitives" description="The number of input triangle primitives." units="primitives" />
-        <event offset="7" counter="LINES" title="Input Primitives" name="Line primitives" description="The number of input line primitives." units="primitives" />
-        <event offset="8" counter="POINTS" title="Input Primitives" name="Point primitives" description="The number of input point primitives." units="primitives" />
-        <event offset="9" counter="FRONT_FACING" title="Visible Primitives" name="Front-facing primitives" description="The number of front-facing triangles that are visible after culling." units="primitives" />
-        <event offset="10" counter="BACK_FACING" title="Visible Primitives" name="Back-facing primitives" description="The number of back-facing triangles that are visible after culling." units="primitives" />
-        <event offset="11" counter="PRIM_VISIBLE" title="Primitive Culling" name="Visible primitives" description="The number of primitives that are visible after culling." units="primitives" />
-        <event offset="12" counter="PRIM_CULLED" title="Primitive Culling" name="Facing and XY plane test culled primitives" description="The number of primitives that are culled by facing or frustum XY plane tests." units="primitives" />
-        <event offset="13" counter="PRIM_CLIPPED" title="Primitive Culling" name="Z plane test culled primitives" description="The number of primitives that are culled by frustum Z plane tests." units="primitives" />
-        <event offset="14" counter="PRIM_SAT_CULLED" title="Primitive Culling" name="Sample test culled primitives" description="The number of primitives culled by the sample coverage test." units="primitives" />
-        <event offset="17" advanced="yes" counter="BUS_READ" title="Tiler L2 Accesses" name="Read beats" description="The number of internal bus data read cycles made by the tiler." units="beats" />
-        <event offset="19" advanced="yes" counter="BUS_WRITE" title="Tiler L2 Accesses" name="Write beats" description="The number of internal bus data write cycles made by the tiler." units="beats" />
-        <event offset="21" counter="IDVS_POS_SHAD_REQ" title="Tiler Shading Requests" name="Position shading requests" description="The number of position shading requests in the IDVS flow." units="requests" />
-        <event offset="23" advanced="yes" counter="IDVS_POS_SHAD_STALL" title="Tiler Cycles" name="Position shading stall cycles" description="The number of cycles where the tiler has a stalled position shading request." units="cycles" />
-        <event offset="24" advanced="yes" counter="IDVS_POS_FIFO_FULL" title="Tiler Cycles" name="Position FIFO full cycles" description="The number of cycles where the tiler has a stalled position shading buffer." units="cycles" />
-        <event offset="26" advanced="yes" counter="VCACHE_HIT" title="Tiler Vertex Cache" name="Position cache hits" description="The number of position lookups that result in a hit in the vertex cache." units="requests" />
-        <event offset="27" advanced="yes" counter="VCACHE_MISS" title="Tiler Vertex Cache" name="Position cache misses" description="The number of position lookups that miss in the vertex cache." units="requests" />
-        <event offset="31" advanced="yes" counter="VFETCH_STALL" title="Tiler Cycles" name="Primitive assembly busy stall cycles" description="The number of cycles where the tiler is stalled waiting for primitive assembly." units="cycles" />
-        <event offset="34" advanced="yes" counter="IDVS_VBU_HIT" title="Tiler Vertex Cache" name="Varying cache hits" description="The number of varying lookups that result in a hit in the vertex cache." units="requests" />
-        <event offset="35" advanced="yes" counter="IDVS_VBU_MISS" title="Tiler Vertex Cache" name="Varying cache misses" description="The number of varying lookups that miss in the vertex cache." units="requests" />
-        <event offset="37" counter="IDVS_VAR_SHAD_REQ" title="Tiler Shading Requests" name="Varying shading requests" description="The number of varying shading requests in the IDVS flow." units="requests" />
-        <event offset="38" advanced="yes" counter="IDVS_VAR_SHAD_STALL" title="Tiler Cycles" name="Varying shading stall cycles" description="The number of cycles where the tiler has a stalled varying shading request." units="cycles" />
-        <event offset="54" advanced="yes" counter="WRBUF_NO_AXI_ID_STALL" title="Tiler Cycles" name="Write buffer transaction stall cycles" description="The number of cycles where the tiler write buffer can not send data because it has no available write IDs." units="cycles" />
-        <event offset="55" advanced="yes" counter="WRBUF_AXI_STALL" title="Tiler Cycles" name="Write buffer write stall cycles" description="The number of cycles where the tiler write buffer can not send data because the bus is not ready." units="cycles" />
+        <event offset="4" counter="TILER_ACTIVE" title="GPU Cycles" name="Tiler active" description="The number of cycles where the tiler has a workload queued for processing." units="cycles"/>
+        <event offset="6" counter="TRIANGLES" title="Input Primitives" name="Triangle primitives" description="The number of input triangle primitives." units="primitives"/>
+        <event offset="7" counter="LINES" title="Input Primitives" name="Line primitives" description="The number of input line primitives." units="primitives"/>
+        <event offset="8" counter="POINTS" title="Input Primitives" name="Point primitives" description="The number of input point primitives." units="primitives"/>
+        <event offset="9" counter="FRONT_FACING" title="Visible Primitives" name="Front-facing primitives" description="The number of front-facing triangles that are visible after culling." units="primitives"/>
+        <event offset="10" counter="BACK_FACING" title="Visible Primitives" name="Back-facing primitives" description="The number of back-facing triangles that are visible after culling." units="primitives"/>
+        <event offset="11" counter="PRIM_VISIBLE" title="Primitive Culling" name="Visible primitives" description="The number of primitives that are visible after culling." units="primitives"/>
+        <event offset="12" counter="PRIM_CULLED" title="Primitive Culling" name="Facing and XY plane test culled primitives" description="The number of primitives that are culled by facing or frustum XY plane tests." units="primitives"/>
+        <event offset="13" counter="PRIM_CLIPPED" title="Primitive Culling" name="Z plane test culled primitives" description="The number of primitives that are culled by frustum Z plane tests." units="primitives"/>
+        <event offset="14" counter="PRIM_SAT_CULLED" title="Primitive Culling" name="Sample test culled primitives" description="The number of primitives culled by the sample coverage test." units="primitives"/>
+        <event offset="17" advanced="yes" counter="BUS_READ" title="Tiler L2 Accesses" name="Read beats" description="The number of internal bus data read cycles made by the tiler." units="beats"/>
+        <event offset="19" advanced="yes" counter="BUS_WRITE" title="Tiler L2 Accesses" name="Write beats" description="The number of internal bus data write cycles made by the tiler." units="beats"/>
+        <event offset="21" counter="IDVS_POS_SHAD_REQ" title="Tiler Shading Requests" name="Position shading requests" description="The number of position shading requests in the IDVS flow." units="requests"/>
+        <event offset="23" advanced="yes" counter="IDVS_POS_SHAD_STALL" title="Tiler Cycles" name="Position shading stall cycles" description="The number of cycles where the tiler has a stalled position shading request." units="cycles"/>
+        <event offset="24" advanced="yes" counter="IDVS_POS_FIFO_FULL" title="Tiler Cycles" name="Position FIFO full cycles" description="The number of cycles where the tiler has a stalled position shading buffer." units="cycles"/>
+        <event offset="26" advanced="yes" counter="VCACHE_HIT" title="Tiler Vertex Cache" name="Position cache hits" description="The number of position lookups that result in a hit in the vertex cache." units="requests"/>
+        <event offset="27" advanced="yes" counter="VCACHE_MISS" title="Tiler Vertex Cache" name="Position cache misses" description="The number of position lookups that miss in the vertex cache." units="requests"/>
+        <event offset="31" advanced="yes" counter="VFETCH_STALL" title="Tiler Cycles" name="Primitive assembly busy stall cycles" description="The number of cycles where the tiler is stalled waiting for primitive assembly." units="cycles"/>
+        <event offset="34" advanced="yes" counter="IDVS_VBU_HIT" title="Tiler Vertex Cache" name="Varying cache hits" description="The number of varying lookups that result in a hit in the vertex cache." units="requests"/>
+        <event offset="35" advanced="yes" counter="IDVS_VBU_MISS" title="Tiler Vertex Cache" name="Varying cache misses" description="The number of varying lookups that miss in the vertex cache." units="requests"/>
+        <event offset="37" counter="IDVS_VAR_SHAD_REQ" title="Tiler Shading Requests" name="Varying shading requests" description="The number of varying shading requests in the IDVS flow." units="requests"/>
+        <event offset="38" advanced="yes" counter="IDVS_VAR_SHAD_STALL" title="Tiler Cycles" name="Varying shading stall cycles" description="The number of cycles where the tiler has a stalled varying shading request." units="cycles"/>
+        <event offset="54" advanced="yes" counter="WRBUF_NO_AXI_ID_STALL" title="Tiler Cycles" name="Write buffer transaction stall cycles" description="The number of cycles where the tiler write buffer can not send data because it has no available write IDs." units="cycles"/>
+        <event offset="55" advanced="yes" counter="WRBUF_AXI_STALL" title="Tiler Cycles" name="Write buffer write stall cycles" description="The number of cycles where the tiler write buffer can not send data because the bus is not ready." units="cycles"/>
+    </category>
+    <category name="Memory System" per_cpu="no">
+        <event offset="4" advanced="yes" counter="MMU_REQUESTS" title="MMU Stage 1 Translations" name="MMU lookups" description="The number of main MMU address translations performed." units="requests"/>
+        <event offset="16" advanced="yes" counter="L2_RD_MSG_IN" title="L2 Cache Requests" name="Read requests" description="The number of L2 cache read requests from internal masters." units="requests"/>
+        <event offset="17" advanced="yes" counter="L2_RD_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Read stall cycles" description="The number of cycles L2 cache read requests from internal masters are stalled." units="cycles"/>
+        <event offset="18" advanced="yes" counter="L2_WR_MSG_IN" title="L2 Cache Requests" name="Write requests" description="The number of L2 cache write requests from internal masters." units="requests"/>
+        <event offset="19" advanced="yes" counter="L2_WR_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Write stall cycles" description="The number of cycles where L2 cache write requests from internal masters are stalled." units="cycles"/>
+        <event offset="20" advanced="yes" counter="L2_SNP_MSG_IN" title="L2 Cache Requests" name="Snoop requests" description="The number of L2 snoop requests from internal masters." units="requests"/>
+        <event offset="21" advanced="yes" counter="L2_SNP_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Snoop stall cycles" description="The number of cycles where L2 cache snoop requests from internal masters are stalled." units="cycles"/>
+        <event offset="22" advanced="yes" counter="L2_RD_MSG_OUT" title="L2 Cache Requests" name="L1 read requests" description="The number of L1 cache read requests sent by the L2 cache to an internal master." units="requests"/>
+        <event offset="23" advanced="yes" counter="L2_RD_MSG_OUT_STALL" title="L2 Cache Stall Cycles" name="L1 read stall cycles" description="The number of cycles where L1 cache read requests sent by the L2 cache to an internal master are stalled." units="cycles"/>
+        <event offset="24" advanced="yes" counter="L2_WR_MSG_OUT" title="L2 Cache Requests" name="L1 write requests" description="The number of L1 cache write responses sent by the L2 cache to an internal master." units="requests"/>
+        <event offset="25" counter="L2_ANY_LOOKUP" title="L2 Cache Lookups" name="Any lookup" description="The number of L2 cache lookups performed." units="requests"/>
+        <event offset="26" counter="L2_READ_LOOKUP" title="L2 Cache Lookups" name="Read lookup" description="The number of L2 cache read lookups performed." units="requests"/>
+        <event offset="27" counter="L2_WRITE_LOOKUP" title="L2 Cache Lookups" name="Write lookup" description="The number of L2 cache write lookups performed." units="requests"/>
+        <event offset="28" advanced="yes" counter="L2_EXT_SNOOP_LOOKUP" title="L2 Cache Lookups" name="External snoop lookups" description="The number of coherency snoop lookups performed that were triggered by an external master." units="requests"/>
+        <event offset="29" counter="L2_EXT_READ" title="External Bus Accesses" name="Read transaction" description="The number of external read transactions." units="transactions"/>
+        <event offset="30" advanced="yes" counter="L2_EXT_READ_NOSNP" title="External Bus Accesses" name="ReadNoSnoop transactions" description="The number of external non-coherent read transactions." units="transactions"/>
+        <event offset="31" advanced="yes" counter="L2_EXT_READ_UNIQUE" title="External Bus Accesses" name="ReadUnique transactions" description="The number of external coherent read unique transactions." units="transactions"/>
+        <event offset="32" counter="L2_EXT_READ_BEATS" title="External Bus Beats" name="Read beat" description="The number of external bus data read cycles." units="beats"/>
+        <event offset="33" counter="L2_EXT_AR_STALL" title="External Bus Stalls" name="Read stall cycles" description="The number of cycles where a read is stalled waiting for the external bus." units="cycles"/>
+        <event offset="34" counter="L2_EXT_AR_CNT_Q1" title="External Bus Outstanding Reads" name="0-25% outstanding" description="The number of read transactions initiated when 0-25% of the maximum are in use." units="transactions"/>
+        <event offset="35" counter="L2_EXT_AR_CNT_Q2" title="External Bus Outstanding Reads" name="25-50% outstanding" description="The number of read transactions initiated when 25-50% of the maximum are in use." units="transactions"/>
+        <event offset="36" counter="L2_EXT_AR_CNT_Q3" title="External Bus Outstanding Reads" name="50-75% outstanding" description="The number of read transactions initiated when 50-75% of the maximum are in use." units="transactions"/>
+        <event offset="37" counter="L2_EXT_RRESP_0_127" title="External Bus Read Latency" name="0-127 cycles" description="The number of data beats returned 0-127 cycles after the read request." units="beats"/>
+        <event offset="38" counter="L2_EXT_RRESP_128_191" title="External Bus Read Latency" name="128-191 cycles" description="The number of data beats returned 128-191 cycles after the read request." units="beats"/>
+        <event offset="39" counter="L2_EXT_RRESP_192_255" title="External Bus Read Latency" name="192-255 cycles" description="The number of data beats returned 192-255 cycles after the read request." units="beats"/>
+        <event offset="40" counter="L2_EXT_RRESP_256_319" title="External Bus Read Latency" name="256-319 cycles" description="The number of data beats returned 256-319 cycles after the read request." units="beats"/>
+        <event offset="41" counter="L2_EXT_RRESP_320_383" title="External Bus Read Latency" name="320-383 cycles" description="The number of data beats returned 320-383 cycles after the read request." units="beats"/>
+        <event offset="42" counter="L2_EXT_WRITE" title="External Bus Accesses" name="Write transaction" description="The number of external write transactions." units="transactions"/>
+        <event offset="43" advanced="yes" counter="L2_EXT_WRITE_NOSNP_FULL" title="External Bus Accesses" name="WriteNoSnoopFull transactions" description="The number of external non-coherent full write transactions." units="transactions"/>
+        <event offset="44" advanced="yes" counter="L2_EXT_WRITE_NOSNP_PTL" title="External Bus Accesses" name="WriteNoSnoopPartial transactions" description="The number of external non-coherent partial write transactions." units="transactions"/>
+        <event offset="45" advanced="yes" counter="L2_EXT_WRITE_SNP_FULL" title="External Bus Accesses" name="WriteSnoopFull transactions" description="The number of external coherent full write transactions." units="transactions"/>
+        <event offset="46" advanced="yes" counter="L2_EXT_WRITE_SNP_PTL" title="External Bus Accesses" name="WriteSnoopPartial transactions" description="The number of external coherent partial write transactions." units="transactions"/>
+        <event offset="47" counter="L2_EXT_WRITE_BEATS" title="External Bus Beats" name="Write beat" description="The number of external bus data write cycles." units="beats"/>
+        <event offset="48" counter="L2_EXT_W_STALL" title="External Bus Stalls" name="Write stall cycles" description="The number of cycles where a write is stalled waiting for the external bus." units="cycles"/>
+        <event offset="49" counter="L2_EXT_AW_CNT_Q1" title="External Bus Outstanding Writes" name="0-25% outstanding" description="The number of write transactions initiated when 0-25% of the maximum are in use." units="transactions"/>
+        <event offset="50" counter="L2_EXT_AW_CNT_Q2" title="External Bus Outstanding Writes" name="25-50% outstanding" description="The number of write transactions initiated when 25-50% of the maximum are in use." units="transactions"/>
+        <event offset="51" counter="L2_EXT_AW_CNT_Q3" title="External Bus Outstanding Writes" name="50-75% outstanding" description="The number of write transactions initiated when 50-75% of the maximum are in use." units="transactions"/>
+        <event offset="52" advanced="yes" counter="L2_EXT_SNOOP" title="External Bus Accesses" name="Snoop transactions" description="The number of coherency snoops triggered by external masters." units="transactions"/>
+        <event offset="53" advanced="yes" counter="L2_EXT_SNOOP_STALL" title="External Bus Stalls" name="Snoop stall cycles" description="The number of cycles where a coherency snoop triggered by external master is stalled." units="cycles"/>
+    </category>
+    <category name="Shader Core" per_cpu="no">
+        <event offset="4" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment active" description="The number of cycles where the shader core is processing a fragment workload." units="cycles"/>
+        <event offset="5" advanced="yes" counter="FRAG_PRIMITIVES_OUT" title="Core Primitives" name="Read primitives" description="The number of primitives read from the tile list by the fragment front-end." units="primitives"/>
+        <event offset="6" counter="FRAG_PRIM_RAST" title="Core Primitives" name="Rasterized primitives" description="The number of primitives being rasterized." units="primitives"/>
+        <event offset="7" counter="FRAG_FPK_ACTIVE" title="Core Cycles" name="Fragment FPKB active" description="The number of cycles where at least one quad is present in the pre-pipe quad queue." units="cycles"/>
+        <event offset="9" counter="FRAG_WARPS" title="Core Warps" name="Fragment warps" description="The number of fragment warps created." units="warps"/>
+        <event offset="10" counter="FRAG_PARTIAL_QUADS_RAST" title="Core Quads" name="Partial rasterized quads" description="The number of partially-rasterized fragment quads created." units="quads"/>
+        <event offset="11" counter="FRAG_QUADS_RAST" title="Core Quads" name="Rasterized quads" description="The number of quads generated by the rasterization phase." units="quads"/>
+        <event offset="12" counter="FRAG_QUADS_EZS_TEST" title="Core Quads" name="Early ZS tested quads" description="The number of quads that are undergoing early depth and stencil testing." units="quads"/>
+        <event offset="13" counter="FRAG_QUADS_EZS_UPDATE" title="Core Quads" name="Early ZS updated quads" description="The number of quads undergoing early depth and stencil testing, that are capable of updating the framebuffer." units="quads"/>
+        <event offset="14" counter="FRAG_QUADS_EZS_KILL" title="Core Quads" name="Early ZS killed quads" description="The number of quads killed by early depth and stencil testing." units="quads"/>
+        <event offset="15" counter="FRAG_LZS_TEST" title="Core Quads" name="Late ZS tested quads" description="The number of quads undergoing late depth and stencil testing." units="quads"/>
+        <event offset="16" counter="FRAG_LZS_KILL" title="Core Quads" name="Late ZS killed quads" description="The number of quads killed by late depth and stencil testing." units="quads"/>
+        <event offset="17" counter="WARP_REG_SIZE_64" title="Core Warps" name="All register warps" description="The number of warps that require more than 32 registers." units="warps"/>
+        <event offset="18" counter="FRAG_PTILES" title="Core Tiles" name="Tiles" description="The number of tiles processed by the shader core." units="tiles"/>
+        <event offset="19" counter="FRAG_TRANS_ELIM" title="Core Tiles" name="Constant tiles killed" description="The number of tiles killed by transaction elimination." units="tiles"/>
+        <event offset="20" counter="QUAD_FPK_KILLER" title="Core Quads" name="FPK occluder quads" description="The number of quads that are valid occluders for hidden surface removal." units="quads"/>
+        <event offset="21" counter="FULL_QUAD_WARPS" title="Core Warps" name="Full quad warps" description="The number of warps that are fully populated with quads." units="warps"/>
+        <event offset="22" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Non-fragment active" description="The number of cycles where the shader core is processing some non-fragment workload." units="cycles"/>
+        <event offset="23" advanced="yes" counter="COMPUTE_TASKS" title="Core Tasks" name="Non-fragment tasks" description="The number of non-fragment tasks issued to the shader core." units="tasks"/>
+        <event offset="24" counter="COMPUTE_WARPS" title="Core Warps" name="Non-fragment warps" description="The number of non-fragment warps created." units="warps"/>
+        <event offset="25" advanced="yes" counter="COMPUTE_STARVING" title="Core Starvation Cycles" name="Non-fragment starvation cycles" description="The number of cycles where the shader core is processing a non-fragment workload and there are no new threads available for execution." units="cycles"/>
+        <event offset="26" counter="EXEC_CORE_ACTIVE" title="Core Cycles" name="Execution core active" description="The number of cycles where the shader core is processing at least one warp." units="cycles"/>
+        <event offset="27" counter="EXEC_INSTR_FMA" title="Core PU Instructions" name="FMA instructions" description="The number of instructions issued to the FMA pipe." units="instructions"/>
+        <event offset="28" counter="EXEC_INSTR_CVT" title="Core PU Instructions" name="CVT instructions" description="The number of instructions issued to the CVT pipe." units="instructions"/>
+        <event offset="29" counter="EXEC_INSTR_SFU" title="Core PU Instructions" name="SFU instructions" description="The number of instructions issued to the SFU pipe." units="instructions"/>
+        <event offset="30" counter="EXEC_INSTR_MSG" title="Core PU Instructions" name="Message instructions" description="The number of instructions issued to the MSG pipe." units="instructions"/>
+        <event offset="31" counter="EXEC_INSTR_DIVERGED" title="Core EE Instructions" name="Diverged instructions" description="The number of instructions executed per warp, that have control flow divergence." units="instructions"/>
+        <event offset="32" advanced="yes" counter="EXEC_ICACHE_MISS" title="Core PU Instructions" name="Instruction cache misses" description="The number of instruction cache misses." units="requests"/>
+        <event offset="33" advanced="yes" counter="EXEC_STARVE_ARITH" title="Core Starvation Cycles" name="Execution engine starvation cycles" description="The number of cycles where the processing unit is starved of work." units="cycles"/>
+        <event offset="34" counter="CALL_BLEND_SHADER" title="Core PU Instructions" name="Blend shader calls" description="The number of blend shader invocations executed." units="instructions"/>
+        <event offset="35" counter="TEX_MSGI_NUM_FLITS" title="Texture Bus" name="Input beats" description="The number of texture request message data beats." units="beats"/>
+        <event offset="36" counter="TEX_DFCH_CLK_STALLED" title="Core Texture Stalls" name="Descriptor stall cycles" description="The number of cycles where a quad is stalled on texture descriptor fetch." units="cycles"/>
+        <event offset="37" counter="TEX_TFCH_CLK_STALLED" title="Core Texture Stalls" name="Fetch queue stall cycles" description="The number of cycles where a quad is stalled on entering texture fetch because the fetch queue is full." units="cycles"/>
+        <event offset="38" counter="TEX_TFCH_STARVED_PENDING_DATA_FETCH" title="Core Texture Stalls" name="Filtering unit stall cycles" description="The number of cycles where the filtering unit is idle and there is at least one quad present in the texture data fetch queue." units="cycles"/>
+        <event offset="39" counter="TEX_FILT_NUM_OPERATIONS" title="Core Texture Cycles" name="Texturing active" description="The number of texture filtering issue cycles." units="cycles"/>
+        <event offset="40" counter="TEX_FILT_NUM_FXR_OPERATIONS" title="Core Texture Cycles" name="4x bilinear filtering active" description="The number of cycles where the filtering unit uses the 4x path to implement nearest or bilinear filtering." units="cycles"/>
+        <event offset="41" counter="TEX_FILT_NUM_FST_OPERATIONS" title="Core Texture Cycles" name="2x trilinear filtering active" description="The number of cycles where the filtering unit uses the 4x path to implement trilinear filtering." units="cycles"/>
+        <event offset="42" counter="TEX_MSGO_NUM_MSG" title="Core Texture Quads" name="Texture requests" description="The number of quad-width texture operations processed by the texture unit." units="quads"/>
+        <event offset="43" counter="TEX_MSGO_NUM_FLITS" title="Texture Bus" name="Output beats" description="The number of texture response message data beats." units="beats"/>
+        <event offset="44" counter="LS_MEM_READ_FULL" title="Core Load/Store Cycles" name="Full read cycles" description="The number of full-width load/store cache reads." units="cycles"/>
+        <event offset="45" counter="LS_MEM_READ_SHORT" title="Core Load/Store Cycles" name="Partial read cycles" description="The number of partial-width load/store cache reads." units="cycles"/>
+        <event offset="46" counter="LS_MEM_WRITE_FULL" title="Core Load/Store Cycles" name="Full write cycles" description="The number of full-width load/store cache writes." units="cycles"/>
+        <event offset="47" counter="LS_MEM_WRITE_SHORT" title="Core Load/Store Cycles" name="Partial write cycles" description="The number of partial-width load/store cache writes." units="cycles"/>
+        <event offset="48" counter="LS_MEM_ATOMIC" title="Core Load/Store Cycles" name="Atomic access cycles" description="The number of load/store atomic accesses." units="cycles"/>
+        <event offset="49" counter="VARY_INSTR" title="Core Varying Requests" name="Interpolation requests" description="The number of warp-width interpolation operations processed by the varying unit." units="instructions"/>
+        <event offset="50" counter="VARY_SLOT_32" title="Core Varying Cycles" name="32-bit interpolation active" description="The number of 32-bit interpolation cycles processed by the varying unit." units="cycles"/>
+        <event offset="51" counter="VARY_SLOT_16" title="Core Varying Cycles" name="16-bit interpolation active" description="The number of 16-bit interpolation cycles processed by the varying unit." units="cycles"/>
+        <event offset="52" advanced="yes" counter="ATTR_INSTR" title="Core Attribute Requests" name="Attribute requests" description="The number of instructions executed by the attribute unit." units="instructions"/>
+        <event offset="53" advanced="yes" counter="ARITH_INSTR_FP_MUL" title="Core EE Instructions" name="Multiplier instructions" description="The number of instructions where the workload uses floating-point multiplier hardware." units="instructions"/>
+        <event offset="54" counter="BEATS_RD_FTC" title="Core L2 Reads" name="Fragment L2 read beats" description="The number of read beats received by the fixed-function fragment front-end." units="beats"/>
+        <event offset="55" counter="BEATS_RD_FTC_EXT" title="Core L2 Reads" name="Fragment external read beats" description="The number of read beats received by the fixed-function fragment front-end that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="56" counter="BEATS_RD_LSC" title="Core L2 Reads" name="Load/store L2 read beats" description="The number of read beats received by the load/store unit." units="beats"/>
+        <event offset="57" counter="BEATS_RD_LSC_EXT" title="Core L2 Reads" name="Load/store external read beats" description="The number of read beats received by the load/store unit that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="58" counter="BEATS_RD_TEX" title="Core L2 Reads" name="Texture L2 read beats" description="The number of read beats received by the texture unit." units="beats"/>
+        <event offset="59" counter="BEATS_RD_TEX_EXT" title="Core L2 Reads" name="Texture external read beats" description="The number of read beats received by the texture unit that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="60" advanced="yes" counter="BEATS_RD_OTHER" title="Core L2 Reads" name="Other L2 read beats" description="The number of read beats received by a unit that is not specifically identified." units="beats"/>
+        <event offset="61" counter="BEATS_WR_LSC_OTHER" title="Core Writes" name="Load/store other write beats" description="The number of write beats by the load/store unit that are due to any reason other than writeback." units="beats"/>
+        <event offset="62" counter="BEATS_WR_TIB" title="Core Writes" name="Tile buffer write beats" description="The number of write beats sent by the tile buffer writeback unit." units="beats"/>
+        <event offset="63" counter="BEATS_WR_LSC_WB" title="Core Writes" name="Load/store writeback write beats" description="The number of write beats by the load/store unit that are due to writeback." units="beats"/>
     </category>
 </metrics>
\ No newline at end of file
diff --git a/src/panfrost/perf/G68.xml b/src/panfrost/perf/G68.xml
index 9095d9a8c53..de21afafca3 100644
--- a/src/panfrost/perf/G68.xml
+++ b/src/panfrost/perf/G68.xml
@@ -1,6 +1,6 @@
 <!--
 Copyright © 2017-2020 ARM Limited.
-Copyright © 2021 Collabora, Ltd.
+Copyright © 2021-2022 Collabora, Ltd.
 Author: Antonio Caggiano <antonio.caggiano@collabora.com>
 
 Permission is hereby granted, free of charge, to any person obtaining a
@@ -24,156 +24,156 @@ SOFTWARE.
 -->
 <metrics id="TOTx">
     <category name="Job Manager" per_cpu="no">
-        <event offset="6" counter="GPU_ACTIVE" title="GPU Cycles" name="GPU active" description="The number of cycles where the GPU has a workload of any type queued for processing." units="cycles" />
-        <event offset="7" advanced="yes" counter="IRQ_ACTIVE" title="GPU Cycles" name="Interrupt active" description="The number of cycles where the GPU has a pending interrupt." units="cycles" />
-        <event offset="8" advanced="yes" counter="JS0_JOBS" title="GPU Jobs" name="Fragment jobs" description="The number of jobs processed by the GPU fragment queue." units="jobs" />
-        <event offset="9" counter="JS0_TASKS" title="GPU Tasks" name="Fragment tasks" description="The number of 32x32 pixel tasks processed by the GPU fragment queue." units="tasks" />
-        <event offset="10" counter="JS0_ACTIVE" title="GPU Cycles" name="Fragment queue active" description="The number of cycles where work is queued for processing in the GPU fragment queue." units="cycles" />
-        <event offset="12" advanced="yes" counter="JS0_WAIT_READ" title="GPU Wait Cycles" name="Fragment descriptor reads cycles" description="The number of cycles where queued fragment work is waiting for a descriptor load." units="cycles" />
-        <event offset="13" advanced="yes" counter="JS0_WAIT_ISSUE" title="GPU Wait Cycles" name="Fragment job issue cycles" description="The number of cycles where queued fragment work is waiting for an available processor." units="cycles" />
-        <event offset="14" advanced="yes" counter="JS0_WAIT_DEPEND" title="GPU Wait Cycles" name="Fragment job dependency cycles" description="The number of cycles where queued fragment work is waiting for dependent work to complete." units="cycles" />
-        <event offset="15" advanced="yes" counter="JS0_WAIT_FINISH" title="GPU Wait Cycles" name="Fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued fragment work to complete." units="cycles" />
-        <event offset="16" advanced="yes" counter="JS1_JOBS" title="GPU Jobs" name="Non-fragment jobs" description="The number of jobs processed by the GPU non-fragment queue." units="jobs" />
-        <event offset="17" advanced="yes" counter="JS1_TASKS" title="GPU Tasks" name="Non-fragment tasks" description="The number of tasks processed by the GPU non-fragment queue." units="tasks" />
-        <event offset="18" counter="JS1_ACTIVE" title="GPU Cycles" name="Non-fragment queue active" description="The number of cycles where work is queued in the GPU non-fragment queue." units="cycles" />
-        <event offset="20" advanced="yes" counter="JS1_WAIT_READ" title="GPU Wait Cycles" name="Non-fragment descriptor read cycles" description="The number number of cycles where queued non-fragment work is waiting for a descriptor load." units="cycles" />
-        <event offset="21" advanced="yes" counter="JS1_WAIT_ISSUE" title="GPU Wait Cycles" name="Non-fragment job issue cycles" description="The number of cycles where queued non-fragment work is waiting for an available processor." units="cycles" />
-        <event offset="22" advanced="yes" counter="JS1_WAIT_DEPEND" title="GPU Wait Cycles" name="Non-fragment job dependency cycles" description="The number of cycles where queued non-fragment work is waiting for dependent work to complete." units="cycles" />
-        <event offset="23" advanced="yes" counter="JS1_WAIT_FINISH" title="GPU Wait Cycles" name="Non-fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued non-fragment work to complete." units="cycles" />
-        <event offset="24" advanced="yes" counter="JS2_JOBS" title="GPU Jobs" name="Reserved jobs" description="The number of jobs processed by the GPU reserved queue." units="jobs" />
-        <event offset="25" advanced="yes" counter="JS2_TASKS" title="GPU Tasks" name="Reserved tasks" description="The number of tasks processed by the GPU reserved queue." units="tasks" />
-        <event offset="26" advanced="yes" counter="JS2_ACTIVE" title="GPU Cycles" name="Reserved queue active" description="The number of cycles where work is queued in the GPU reserved queue." units="cycles" />
-        <event offset="28" advanced="yes" counter="JS2_WAIT_READ" title="GPU Wait Cycles" name="Reserved descriptor read cycles" description="The number of cycles where queued reserved work is waiting for a descriptor load." units="cycles" />
-        <event offset="29" advanced="yes" counter="JS2_WAIT_ISSUE" title="GPU Wait Cycles" name="Reserved job issue cycles" description="The number of cycles where queued reserved work is waiting for an available processor." units="cycles" />
-        <event offset="30" advanced="yes" counter="JS2_WAIT_DEPEND" title="GPU Wait Cycles" name="Reserved job dependency cycles" description="The number of cycles where queued reserved work is waiting for dependent work to complete." units="cycles" />
-        <event offset="31" advanced="yes" counter="JS2_WAIT_FINISH" title="GPU Wait Cycles" name="Reserved job finish cycles" description="The number of cycles where the GPU is waiting for issued reserved work to complete." units="cycles" />
-    </category>
-    <category name="Memory System" per_cpu="no">
-        <event offset="196" advanced="yes" counter="MMU_REQUESTS" title="MMU Stage 1 Translations" name="MMU lookups" description="The number of main MMU address translations performed." units="requests" />
-        <event offset="208" advanced="yes" counter="L2_RD_MSG_IN" title="L2 Cache Requests" name="Read requests" description="The number of L2 cache read requests from internal masters." units="requests" />
-        <event offset="209" advanced="yes" counter="L2_RD_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Read stall cycles" description="The number of cycles L2 cache read requests from internal masters are stalled." units="cycles" />
-        <event offset="210" advanced="yes" counter="L2_WR_MSG_IN" title="L2 Cache Requests" name="Write requests" description="The number of L2 cache write requests from internal masters." units="requests" />
-        <event offset="211" advanced="yes" counter="L2_WR_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Write stall cycles" description="The number of cycles where L2 cache write requests from internal masters are stalled." units="cycles" />
-        <event offset="212" advanced="yes" counter="L2_SNP_MSG_IN" title="L2 Cache Requests" name="Snoop requests" description="The number of L2 snoop requests from internal masters." units="requests" />
-        <event offset="213" advanced="yes" counter="L2_SNP_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Snoop stall cycles" description="The number of cycles where L2 cache snoop requests from internal masters are stalled." units="cycles" />
-        <event offset="214" advanced="yes" counter="L2_RD_MSG_OUT" title="L2 Cache Requests" name="L1 read requests" description="The number of L1 cache read requests sent by the L2 cache to an internal master." units="requests" />
-        <event offset="215" advanced="yes" counter="L2_RD_MSG_OUT_STALL" title="L2 Cache Stall Cycles" name="L1 read stall cycles" description="The number of cycles where L1 cache read requests sent by the L2 cache to an internal master are stalled." units="cycles" />
-        <event offset="216" advanced="yes" counter="L2_WR_MSG_OUT" title="L2 Cache Requests" name="L1 write requests" description="The number of L1 cache write responses sent by the L2 cache to an internal master." units="requests" />
-        <event offset="217" counter="L2_ANY_LOOKUP" title="L2 Cache Lookups" name="Any lookup" description="The number of L2 cache lookups performed." units="requests" />
-        <event offset="218" counter="L2_READ_LOOKUP" title="L2 Cache Lookups" name="Read lookup" description="The number of L2 cache read lookups performed." units="requests" />
-        <event offset="219" counter="L2_WRITE_LOOKUP" title="L2 Cache Lookups" name="Write lookup" description="The number of L2 cache write lookups performed." units="requests" />
-        <event offset="220" advanced="yes" counter="L2_EXT_SNOOP_LOOKUP" title="L2 Cache Lookups" name="External snoop lookups" description="The number of coherency snoop lookups performed that were triggered by an external master." units="requests" />
-        <event offset="221" counter="L2_EXT_READ" title="External Bus Accesses" name="Read transaction" description="The number of external read transactions." units="transactions" />
-        <event offset="222" advanced="yes" counter="L2_EXT_READ_NOSNP" title="External Bus Accesses" name="ReadNoSnoop transactions" description="The number of external non-coherent read transactions." units="transactions" />
-        <event offset="223" advanced="yes" counter="L2_EXT_READ_UNIQUE" title="External Bus Accesses" name="ReadUnique transactions" description="The number of external coherent read unique transactions." units="transactions" />
-        <event offset="224" counter="L2_EXT_READ_BEATS" title="External Bus Beats" name="Read beat" description="The number of external bus data read cycles." units="beats" />
-        <event offset="225" counter="L2_EXT_AR_STALL" title="External Bus Stalls" name="Read stall cycles" description="The number of cycles where a read is stalled waiting for the external bus." units="cycles" />
-        <event offset="226" counter="L2_EXT_AR_CNT_Q1" title="External Bus Outstanding Reads" name="0-25% outstanding" description="The number of read transactions initiated when 0-25% of the maximum are in use." units="transactions" />
-        <event offset="227" counter="L2_EXT_AR_CNT_Q2" title="External Bus Outstanding Reads" name="25-50% outstanding" description="The number of read transactions initiated when 25-50% of the maximum are in use." units="transactions" />
-        <event offset="228" counter="L2_EXT_AR_CNT_Q3" title="External Bus Outstanding Reads" name="50-75% outstanding" description="The number of read transactions initiated when 50-75% of the maximum are in use." units="transactions" />
-        <event offset="229" counter="L2_EXT_RRESP_0_127" title="External Bus Read Latency" name="0-127 cycles" description="The number of data beats returned 0-127 cycles after the read request." units="beats" />
-        <event offset="230" counter="L2_EXT_RRESP_128_191" title="External Bus Read Latency" name="128-191 cycles" description="The number of data beats returned 128-191 cycles after the read request." units="beats" />
-        <event offset="231" counter="L2_EXT_RRESP_192_255" title="External Bus Read Latency" name="192-255 cycles" description="The number of data beats returned 192-255 cycles after the read request." units="beats" />
-        <event offset="232" counter="L2_EXT_RRESP_256_319" title="External Bus Read Latency" name="256-319 cycles" description="The number of data beats returned 256-319 cycles after the read request." units="beats" />
-        <event offset="233" counter="L2_EXT_RRESP_320_383" title="External Bus Read Latency" name="320-383 cycles" description="The number of data beats returned 320-383 cycles after the read request." units="beats" />
-        <event offset="234" counter="L2_EXT_WRITE" title="External Bus Accesses" name="Write transaction" description="The number of external write transactions." units="transactions" />
-        <event offset="235" advanced="yes" counter="L2_EXT_WRITE_NOSNP_FULL" title="External Bus Accesses" name="WriteNoSnoopFull transactions" description="The number of external non-coherent full write transactions." units="transactions" />
-        <event offset="236" advanced="yes" counter="L2_EXT_WRITE_NOSNP_PTL" title="External Bus Accesses" name="WriteNoSnoopPartial transactions" description="The number of external non-coherent partial write transactions." units="transactions" />
-        <event offset="237" advanced="yes" counter="L2_EXT_WRITE_SNP_FULL" title="External Bus Accesses" name="WriteSnoopFull transactions" description="The number of external coherent full write transactions." units="transactions" />
-        <event offset="238" advanced="yes" counter="L2_EXT_WRITE_SNP_PTL" title="External Bus Accesses" name="WriteSnoopPartial transactions" description="The number of external coherent partial write transactions." units="transactions" />
-        <event offset="239" counter="L2_EXT_WRITE_BEATS" title="External Bus Beats" name="Write beat" description="The number of external bus data write cycles." units="beats" />
-        <event offset="240" counter="L2_EXT_W_STALL" title="External Bus Stalls" name="Write stall cycles" description="The number of cycles where a write is stalled waiting for the external bus." units="cycles" />
-        <event offset="241" counter="L2_EXT_AW_CNT_Q1" title="External Bus Outstanding Writes" name="0-25% outstanding" description="The number of write transactions initiated when 0-25% of the maximum are in use." units="transactions" />
-        <event offset="242" counter="L2_EXT_AW_CNT_Q2" title="External Bus Outstanding Writes" name="25-50% outstanding" description="The number of write transactions initiated when 25-50% of the maximum are in use." units="transactions" />
-        <event offset="243" counter="L2_EXT_AW_CNT_Q3" title="External Bus Outstanding Writes" name="50-75% outstanding" description="The number of write transactions initiated when 50-75% of the maximum are in use." units="transactions" />
-        <event offset="244" advanced="yes" counter="L2_EXT_SNOOP" title="External Bus Accesses" name="Snoop transactions" description="The number of coherency snoops triggered by external masters." units="transactions" />
-        <event offset="245" advanced="yes" counter="L2_EXT_SNOOP_STALL" title="External Bus Stalls" name="Snoop stall cycles" description="The number of cycles where a coherency snoop triggered by external master is stalled." units="cycles" />
-    </category>
-    <category name="Shader Core" per_cpu="no">
-        <event offset="196" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment active" description="The number of cycles where the shader core is processing a fragment workload." units="cycles" />
-        <event offset="197" advanced="yes" counter="FRAG_PRIMITIVES_OUT" title="Core Primitives" name="Read primitives" description="The number of primitives read from the tile list by the fragment front-end." units="primitives" />
-        <event offset="198" counter="FRAG_PRIM_RAST" title="Core Primitives" name="Rasterized primitives" description="The number of primitives being rasterized." units="primitives" />
-        <event offset="199" counter="FRAG_FPK_ACTIVE" title="Core Cycles" name="Fragment FPKB active" description="The number of cycles where at least one quad is present in the pre-pipe quad queue." units="cycles" />
-        <event offset="201" counter="FRAG_WARPS" title="Core Warps" name="Fragment warps" description="The number of fragment warps created." units="warps" />
-        <event offset="202" counter="FRAG_PARTIAL_QUADS_RAST" title="Core Quads" name="Partial rasterized quads" description="The number of partially-rasterized fragment quads created." units="quads" />
-        <event offset="203" counter="FRAG_QUADS_RAST" title="Core Quads" name="Rasterized quads" description="The number of quads generated by the rasterization phase." units="quads" />
-        <event offset="204" counter="FRAG_QUADS_EZS_TEST" title="Core Quads" name="Early ZS tested quads" description="The number of quads that are undergoing early depth and stencil testing." units="quads" />
-        <event offset="205" counter="FRAG_QUADS_EZS_UPDATE" title="Core Quads" name="Early ZS updated quads" description="The number of quads undergoing early depth and stencil testing, that are capable of updating the framebuffer." units="quads" />
-        <event offset="206" counter="FRAG_QUADS_EZS_KILL" title="Core Quads" name="Early ZS killed quads" description="The number of quads killed by early depth and stencil testing." units="quads" />
-        <event offset="207" counter="FRAG_LZS_TEST" title="Core Quads" name="Late ZS tested quads" description="The number of quads undergoing late depth and stencil testing." units="quads" />
-        <event offset="208" counter="FRAG_LZS_KILL" title="Core Quads" name="Late ZS killed quads" description="The number of quads killed by late depth and stencil testing." units="quads" />
-        <event offset="209" counter="WARP_REG_SIZE_64" title="Core Warps" name="All register warps" description="The number of warps that require more than 32 registers." units="warps" />
-        <event offset="210" counter="FRAG_PTILES" title="Core Tiles" name="Tiles" description="The number of tiles processed by the shader core." units="tiles" />
-        <event offset="211" counter="FRAG_TRANS_ELIM" title="Core Tiles" name="Constant tiles killed" description="The number of tiles killed by transaction elimination." units="tiles" />
-        <event offset="212" counter="QUAD_FPK_KILLER" title="Core Quads" name="FPK occluder quads" description="The number of quads that are valid occluders for hidden surface removal." units="quads" />
-        <event offset="213" counter="FULL_QUAD_WARPS" title="Core Warps" name="Full quad warps" description="The number of warps that are fully populated with quads." units="warps" />
-        <event offset="214" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Non-fragment active" description="The number of cycles where the shader core is processing some non-fragment workload." units="cycles" />
-        <event offset="215" advanced="yes" counter="COMPUTE_TASKS" title="Core Tasks" name="Non-fragment tasks" description="The number of non-fragment tasks issued to the shader core." units="tasks" />
-        <event offset="216" counter="COMPUTE_WARPS" title="Core Warps" name="Non-fragment warps" description="The number of non-fragment warps created." units="warps" />
-        <event offset="217" advanced="yes" counter="COMPUTE_STARVING" title="Core Starvation Cycles" name="Non-fragment starvation cycles" description="The number of cycles where the shader core is processing a non-fragment workload and there are no new threads available for execution." units="cycles" />
-        <event offset="218" counter="EXEC_CORE_ACTIVE" title="Core Cycles" name="Execution core active" description="The number of cycles where the shader core is processing at least one warp." units="cycles" />
-        <event offset="219" counter="EXEC_INSTR_FMA" title="Core PU Instructions" name="FMA instructions" description="The number of instructions issued to the FMA pipe." units="instructions" />
-        <event offset="220" counter="EXEC_INSTR_CVT" title="Core PU Instructions" name="CVT instructions" description="The number of instructions issued to the CVT pipe." units="instructions" />
-        <event offset="221" counter="EXEC_INSTR_SFU" title="Core PU Instructions" name="SFU instructions" description="The number of instructions issued to the SFU pipe." units="instructions" />
-        <event offset="222" counter="EXEC_INSTR_MSG" title="Core PU Instructions" name="Message instructions" description="The number of instructions issued to the MSG pipe." units="instructions" />
-        <event offset="223" counter="EXEC_INSTR_DIVERGED" title="Core EE Instructions" name="Diverged instructions" description="The number of instructions executed per warp, that have control flow divergence." units="instructions" />
-        <event offset="224" advanced="yes" counter="EXEC_ICACHE_MISS" title="Core PU Instructions" name="Instruction cache misses" description="The number of instruction cache misses." units="requests" />
-        <event offset="225" advanced="yes" counter="EXEC_STARVE_ARITH" title="Core Starvation Cycles" name="Execution engine starvation cycles" description="The number of cycles where the processing unit is starved of work." units="cycles" />
-        <event offset="226" counter="CALL_BLEND_SHADER" title="Core PU Instructions" name="Blend shader calls" description="The number of blend shader invocations executed." units="instructions" />
-        <event offset="227" counter="TEX_MSGI_NUM_FLITS" title="Texture Bus" name="Input beats" description="The number of texture request message data beats." units="beats" />
-        <event offset="228" counter="TEX_DFCH_CLK_STALLED" title="Core Texture Stalls" name="Descriptor stall cycles" description="The number of cycles where a quad is stalled on texture descriptor fetch." units="cycles" />
-        <event offset="229" counter="TEX_TFCH_CLK_STALLED" title="Core Texture Stalls" name="Fetch queue stall cycles" description="The number of cycles where a quad is stalled on entering texture fetch because the fetch queue is full." units="cycles" />
-        <event offset="230" counter="TEX_TFCH_STARVED_PENDING_DATA_FETCH" title="Core Texture Stalls" name="Filtering unit stall cycles" description="The number of cycles where the filtering unit is idle and there is at least one quad present in the texture data fetch queue." units="cycles" />
-        <event offset="231" counter="TEX_FILT_NUM_OPERATIONS" title="Core Texture Cycles" name="Texturing active" description="The number of texture filtering issue cycles." units="cycles" />
-        <event offset="232" counter="TEX_FILT_NUM_FXR_OPERATIONS" title="Core Texture Cycles" name="4x bilinear filtering active" description="The number of cycles where the filtering unit uses the 4x path to implement nearest or bilinear filtering." units="cycles" />
-        <event offset="233" counter="TEX_FILT_NUM_FST_OPERATIONS" title="Core Texture Cycles" name="2x trilinear filtering active" description="The number of cycles where the filtering unit uses the 4x path to implement trilinear filtering." units="cycles" />
-        <event offset="234" counter="TEX_MSGO_NUM_MSG" title="Core Texture Quads" name="Texture requests" description="The number of quad-width texture operations processed by the texture unit." units="quads" />
-        <event offset="235" counter="TEX_MSGO_NUM_FLITS" title="Texture Bus" name="Output beats" description="The number of texture response message data beats." units="beats" />
-        <event offset="236" counter="LS_MEM_READ_FULL" title="Core Load/Store Cycles" name="Full read cycles" description="The number of full-width load/store cache reads." units="cycles" />
-        <event offset="237" counter="LS_MEM_READ_SHORT" title="Core Load/Store Cycles" name="Partial read cycles" description="The number of partial-width load/store cache reads." units="cycles" />
-        <event offset="238" counter="LS_MEM_WRITE_FULL" title="Core Load/Store Cycles" name="Full write cycles" description="The number of full-width load/store cache writes." units="cycles" />
-        <event offset="239" counter="LS_MEM_WRITE_SHORT" title="Core Load/Store Cycles" name="Partial write cycles" description="The number of partial-width load/store cache writes." units="cycles" />
-        <event offset="240" counter="LS_MEM_ATOMIC" title="Core Load/Store Cycles" name="Atomic access cycles" description="The number of load/store atomic accesses." units="cycles" />
-        <event offset="241" counter="VARY_INSTR" title="Core Varying Requests" name="Interpolation requests" description="The number of warp-width interpolation operations processed by the varying unit." units="instructions" />
-        <event offset="242" counter="VARY_SLOT_32" title="Core Varying Cycles" name="32-bit interpolation active" description="The number of 32-bit interpolation cycles processed by the varying unit." units="cycles" />
-        <event offset="243" counter="VARY_SLOT_16" title="Core Varying Cycles" name="16-bit interpolation active" description="The number of 16-bit interpolation cycles processed by the varying unit." units="cycles" />
-        <event offset="244" advanced="yes" counter="ATTR_INSTR" title="Core Attribute Requests" name="Attribute requests" description="The number of instructions executed by the attribute unit." units="instructions" />
-        <event offset="245" counter="SHADER_CORE_ACTIVE" title="Core Cycles" name="Any active" description="The number of cycles where the shader core is processing either a non-fragment workload or a fragment workload." units="cycles" />
-        <event offset="246" counter="BEATS_RD_FTC" title="Core L2 Reads" name="Fragment L2 read beats" description="The number of read beats received by the fixed-function fragment front-end." units="beats" />
-        <event offset="247" counter="BEATS_RD_FTC_EXT" title="Core L2 Reads" name="Fragment external read beats" description="The number of read beats received by the fixed-function fragment front-end that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="248" counter="BEATS_RD_LSC" title="Core L2 Reads" name="Load/store L2 read beats" description="The number of read beats received by the load/store unit." units="beats" />
-        <event offset="249" counter="BEATS_RD_LSC_EXT" title="Core L2 Reads" name="Load/store external read beats" description="The number of read beats received by the load/store unit that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="250" counter="BEATS_RD_TEX" title="Core L2 Reads" name="Texture L2 read beats" description="The number of read beats received by the texture unit." units="beats" />
-        <event offset="251" counter="BEATS_RD_TEX_EXT" title="Core L2 Reads" name="Texture external read beats" description="The number of read beats received by the texture unit that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="252" advanced="yes" counter="BEATS_RD_OTHER" title="Core L2 Reads" name="Other L2 read beats" description="The number of read beats received by a unit that is not specifically identified." units="beats" />
-        <event offset="253" counter="BEATS_WR_LSC_OTHER" title="Core Writes" name="Load/store other write beats" description="The number of write beats by the load/store unit that are due to any reason other than writeback." units="beats" />
-        <event offset="254" counter="BEATS_WR_TIB" title="Core Writes" name="Tile buffer write beats" description="The number of write beats sent by the tile buffer writeback unit." units="beats" />
-        <event offset="255" counter="BEATS_WR_LSC_WB" title="Core Writes" name="Load/store writeback write beats" description="The number of write beats by the load/store unit that are due to writeback." units="beats" />
+        <event offset="6" counter="GPU_ACTIVE" title="GPU Cycles" name="GPU active" description="The number of cycles where the GPU has a workload of any type queued for processing." units="cycles"/>
+        <event offset="7" advanced="yes" counter="IRQ_ACTIVE" title="GPU Cycles" name="Interrupt active" description="The number of cycles where the GPU has a pending interrupt." units="cycles"/>
+        <event offset="8" advanced="yes" counter="JS0_JOBS" title="GPU Jobs" name="Fragment jobs" description="The number of jobs processed by the GPU fragment queue." units="jobs"/>
+        <event offset="9" counter="JS0_TASKS" title="GPU Tasks" name="Fragment tasks" description="The number of 32x32 pixel tasks processed by the GPU fragment queue." units="tasks"/>
+        <event offset="10" counter="JS0_ACTIVE" title="GPU Cycles" name="Fragment queue active" description="The number of cycles where work is queued for processing in the GPU fragment queue." units="cycles"/>
+        <event offset="12" advanced="yes" counter="JS0_WAIT_READ" title="GPU Wait Cycles" name="Fragment descriptor reads cycles" description="The number of cycles where queued fragment work is waiting for a descriptor load." units="cycles"/>
+        <event offset="13" advanced="yes" counter="JS0_WAIT_ISSUE" title="GPU Wait Cycles" name="Fragment job issue cycles" description="The number of cycles where queued fragment work is waiting for an available processor." units="cycles"/>
+        <event offset="14" advanced="yes" counter="JS0_WAIT_DEPEND" title="GPU Wait Cycles" name="Fragment job dependency cycles" description="The number of cycles where queued fragment work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="15" advanced="yes" counter="JS0_WAIT_FINISH" title="GPU Wait Cycles" name="Fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued fragment work to complete." units="cycles"/>
+        <event offset="16" advanced="yes" counter="JS1_JOBS" title="GPU Jobs" name="Non-fragment jobs" description="The number of jobs processed by the GPU non-fragment queue." units="jobs"/>
+        <event offset="17" advanced="yes" counter="JS1_TASKS" title="GPU Tasks" name="Non-fragment tasks" description="The number of tasks processed by the GPU non-fragment queue." units="tasks"/>
+        <event offset="18" counter="JS1_ACTIVE" title="GPU Cycles" name="Non-fragment queue active" description="The number of cycles where work is queued in the GPU non-fragment queue." units="cycles"/>
+        <event offset="20" advanced="yes" counter="JS1_WAIT_READ" title="GPU Wait Cycles" name="Non-fragment descriptor read cycles" description="The number number of cycles where queued non-fragment work is waiting for a descriptor load." units="cycles"/>
+        <event offset="21" advanced="yes" counter="JS1_WAIT_ISSUE" title="GPU Wait Cycles" name="Non-fragment job issue cycles" description="The number of cycles where queued non-fragment work is waiting for an available processor." units="cycles"/>
+        <event offset="22" advanced="yes" counter="JS1_WAIT_DEPEND" title="GPU Wait Cycles" name="Non-fragment job dependency cycles" description="The number of cycles where queued non-fragment work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="23" advanced="yes" counter="JS1_WAIT_FINISH" title="GPU Wait Cycles" name="Non-fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued non-fragment work to complete." units="cycles"/>
+        <event offset="24" advanced="yes" counter="JS2_JOBS" title="GPU Jobs" name="Reserved jobs" description="The number of jobs processed by the GPU reserved queue." units="jobs"/>
+        <event offset="25" advanced="yes" counter="JS2_TASKS" title="GPU Tasks" name="Reserved tasks" description="The number of tasks processed by the GPU reserved queue." units="tasks"/>
+        <event offset="26" advanced="yes" counter="JS2_ACTIVE" title="GPU Cycles" name="Reserved queue active" description="The number of cycles where work is queued in the GPU reserved queue." units="cycles"/>
+        <event offset="28" advanced="yes" counter="JS2_WAIT_READ" title="GPU Wait Cycles" name="Reserved descriptor read cycles" description="The number of cycles where queued reserved work is waiting for a descriptor load." units="cycles"/>
+        <event offset="29" advanced="yes" counter="JS2_WAIT_ISSUE" title="GPU Wait Cycles" name="Reserved job issue cycles" description="The number of cycles where queued reserved work is waiting for an available processor." units="cycles"/>
+        <event offset="30" advanced="yes" counter="JS2_WAIT_DEPEND" title="GPU Wait Cycles" name="Reserved job dependency cycles" description="The number of cycles where queued reserved work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="31" advanced="yes" counter="JS2_WAIT_FINISH" title="GPU Wait Cycles" name="Reserved job finish cycles" description="The number of cycles where the GPU is waiting for issued reserved work to complete." units="cycles"/>
     </category>
     <category name="Tiler" per_cpu="no">
-        <event offset="4" counter="TILER_ACTIVE" title="GPU Cycles" name="Tiler active" description="The number of cycles where the tiler has a workload queued for processing." units="cycles" />
-        <event offset="6" counter="TRIANGLES" title="Input Primitives" name="Triangle primitives" description="The number of input triangle primitives." units="primitives" />
-        <event offset="7" counter="LINES" title="Input Primitives" name="Line primitives" description="The number of input line primitives." units="primitives" />
-        <event offset="8" counter="POINTS" title="Input Primitives" name="Point primitives" description="The number of input point primitives." units="primitives" />
-        <event offset="9" counter="FRONT_FACING" title="Visible Primitives" name="Front-facing primitives" description="The number of front-facing triangles that are visible after culling." units="primitives" />
-        <event offset="10" counter="BACK_FACING" title="Visible Primitives" name="Back-facing primitives" description="The number of back-facing triangles that are visible after culling." units="primitives" />
-        <event offset="11" counter="PRIM_VISIBLE" title="Primitive Culling" name="Visible primitives" description="The number of primitives that are visible after culling." units="primitives" />
-        <event offset="12" counter="PRIM_CULLED" title="Primitive Culling" name="Facing and XY plane test culled primitives" description="The number of primitives that are culled by facing or frustum XY plane tests." units="primitives" />
-        <event offset="13" counter="PRIM_CLIPPED" title="Primitive Culling" name="Z plane test culled primitives" description="The number of primitives that are culled by frustum Z plane tests." units="primitives" />
-        <event offset="14" counter="PRIM_SAT_CULLED" title="Primitive Culling" name="Sample test culled primitives" description="The number of primitives culled by the sample coverage test." units="primitives" />
-        <event offset="17" advanced="yes" counter="BUS_READ" title="Tiler L2 Accesses" name="Read beats" description="The number of internal bus data read cycles made by the tiler." units="beats" />
-        <event offset="19" advanced="yes" counter="BUS_WRITE" title="Tiler L2 Accesses" name="Write beats" description="The number of internal bus data write cycles made by the tiler." units="beats" />
-        <event offset="21" counter="IDVS_POS_SHAD_REQ" title="Tiler Shading Requests" name="Position shading requests" description="The number of position shading requests in the IDVS flow." units="requests" />
-        <event offset="23" advanced="yes" counter="IDVS_POS_SHAD_STALL" title="Tiler Cycles" name="Position shading stall cycles" description="The number of cycles where the tiler has a stalled position shading request." units="cycles" />
-        <event offset="24" advanced="yes" counter="IDVS_POS_FIFO_FULL" title="Tiler Cycles" name="Position FIFO full cycles" description="The number of cycles where the tiler has a stalled position shading buffer." units="cycles" />
-        <event offset="26" advanced="yes" counter="VCACHE_HIT" title="Tiler Vertex Cache" name="Position cache hits" description="The number of position lookups that result in a hit in the vertex cache." units="requests" />
-        <event offset="27" advanced="yes" counter="VCACHE_MISS" title="Tiler Vertex Cache" name="Position cache misses" description="The number of position lookups that miss in the vertex cache." units="requests" />
-        <event offset="31" advanced="yes" counter="VFETCH_STALL" title="Tiler Cycles" name="Primitive assembly busy stall cycles" description="The number of cycles where the tiler is stalled waiting for primitive assembly." units="cycles" />
-        <event offset="34" advanced="yes" counter="IDVS_VBU_HIT" title="Tiler Vertex Cache" name="Varying cache hits" description="The number of varying lookups that result in a hit in the vertex cache." units="requests" />
-        <event offset="35" advanced="yes" counter="IDVS_VBU_MISS" title="Tiler Vertex Cache" name="Varying cache misses" description="The number of varying lookups that miss in the vertex cache." units="requests" />
-        <event offset="37" counter="IDVS_VAR_SHAD_REQ" title="Tiler Shading Requests" name="Varying shading requests" description="The number of varying shading requests in the IDVS flow." units="requests" />
-        <event offset="38" advanced="yes" counter="IDVS_VAR_SHAD_STALL" title="Tiler Cycles" name="Varying shading stall cycles" description="The number of cycles where the tiler has a stalled varying shading request." units="cycles" />
-        <event offset="54" advanced="yes" counter="WRBUF_NO_AXI_ID_STALL" title="Tiler Cycles" name="Write buffer transaction stall cycles" description="The number of cycles where the tiler write buffer can not send data because it has no available write IDs." units="cycles" />
-        <event offset="55" advanced="yes" counter="WRBUF_AXI_STALL" title="Tiler Cycles" name="Write buffer write stall cycles" description="The number of cycles where the tiler write buffer can not send data because the bus is not ready." units="cycles" />
+        <event offset="4" counter="TILER_ACTIVE" title="GPU Cycles" name="Tiler active" description="The number of cycles where the tiler has a workload queued for processing." units="cycles"/>
+        <event offset="6" counter="TRIANGLES" title="Input Primitives" name="Triangle primitives" description="The number of input triangle primitives." units="primitives"/>
+        <event offset="7" counter="LINES" title="Input Primitives" name="Line primitives" description="The number of input line primitives." units="primitives"/>
+        <event offset="8" counter="POINTS" title="Input Primitives" name="Point primitives" description="The number of input point primitives." units="primitives"/>
+        <event offset="9" counter="FRONT_FACING" title="Visible Primitives" name="Front-facing primitives" description="The number of front-facing triangles that are visible after culling." units="primitives"/>
+        <event offset="10" counter="BACK_FACING" title="Visible Primitives" name="Back-facing primitives" description="The number of back-facing triangles that are visible after culling." units="primitives"/>
+        <event offset="11" counter="PRIM_VISIBLE" title="Primitive Culling" name="Visible primitives" description="The number of primitives that are visible after culling." units="primitives"/>
+        <event offset="12" counter="PRIM_CULLED" title="Primitive Culling" name="Facing and XY plane test culled primitives" description="The number of primitives that are culled by facing or frustum XY plane tests." units="primitives"/>
+        <event offset="13" counter="PRIM_CLIPPED" title="Primitive Culling" name="Z plane test culled primitives" description="The number of primitives that are culled by frustum Z plane tests." units="primitives"/>
+        <event offset="14" counter="PRIM_SAT_CULLED" title="Primitive Culling" name="Sample test culled primitives" description="The number of primitives culled by the sample coverage test." units="primitives"/>
+        <event offset="17" advanced="yes" counter="BUS_READ" title="Tiler L2 Accesses" name="Read beats" description="The number of internal bus data read cycles made by the tiler." units="beats"/>
+        <event offset="19" advanced="yes" counter="BUS_WRITE" title="Tiler L2 Accesses" name="Write beats" description="The number of internal bus data write cycles made by the tiler." units="beats"/>
+        <event offset="21" counter="IDVS_POS_SHAD_REQ" title="Tiler Shading Requests" name="Position shading requests" description="The number of position shading requests in the IDVS flow." units="requests"/>
+        <event offset="23" advanced="yes" counter="IDVS_POS_SHAD_STALL" title="Tiler Cycles" name="Position shading stall cycles" description="The number of cycles where the tiler has a stalled position shading request." units="cycles"/>
+        <event offset="24" advanced="yes" counter="IDVS_POS_FIFO_FULL" title="Tiler Cycles" name="Position FIFO full cycles" description="The number of cycles where the tiler has a stalled position shading buffer." units="cycles"/>
+        <event offset="26" advanced="yes" counter="VCACHE_HIT" title="Tiler Vertex Cache" name="Position cache hits" description="The number of position lookups that result in a hit in the vertex cache." units="requests"/>
+        <event offset="27" advanced="yes" counter="VCACHE_MISS" title="Tiler Vertex Cache" name="Position cache misses" description="The number of position lookups that miss in the vertex cache." units="requests"/>
+        <event offset="31" advanced="yes" counter="VFETCH_STALL" title="Tiler Cycles" name="Primitive assembly busy stall cycles" description="The number of cycles where the tiler is stalled waiting for primitive assembly." units="cycles"/>
+        <event offset="34" advanced="yes" counter="IDVS_VBU_HIT" title="Tiler Vertex Cache" name="Varying cache hits" description="The number of varying lookups that result in a hit in the vertex cache." units="requests"/>
+        <event offset="35" advanced="yes" counter="IDVS_VBU_MISS" title="Tiler Vertex Cache" name="Varying cache misses" description="The number of varying lookups that miss in the vertex cache." units="requests"/>
+        <event offset="37" counter="IDVS_VAR_SHAD_REQ" title="Tiler Shading Requests" name="Varying shading requests" description="The number of varying shading requests in the IDVS flow." units="requests"/>
+        <event offset="38" advanced="yes" counter="IDVS_VAR_SHAD_STALL" title="Tiler Cycles" name="Varying shading stall cycles" description="The number of cycles where the tiler has a stalled varying shading request." units="cycles"/>
+        <event offset="54" advanced="yes" counter="WRBUF_NO_AXI_ID_STALL" title="Tiler Cycles" name="Write buffer transaction stall cycles" description="The number of cycles where the tiler write buffer can not send data because it has no available write IDs." units="cycles"/>
+        <event offset="55" advanced="yes" counter="WRBUF_AXI_STALL" title="Tiler Cycles" name="Write buffer write stall cycles" description="The number of cycles where the tiler write buffer can not send data because the bus is not ready." units="cycles"/>
+    </category>
+    <category name="Memory System" per_cpu="no">
+        <event offset="4" advanced="yes" counter="MMU_REQUESTS" title="MMU Stage 1 Translations" name="MMU lookups" description="The number of main MMU address translations performed." units="requests"/>
+        <event offset="16" advanced="yes" counter="L2_RD_MSG_IN" title="L2 Cache Requests" name="Read requests" description="The number of L2 cache read requests from internal masters." units="requests"/>
+        <event offset="17" advanced="yes" counter="L2_RD_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Read stall cycles" description="The number of cycles L2 cache read requests from internal masters are stalled." units="cycles"/>
+        <event offset="18" advanced="yes" counter="L2_WR_MSG_IN" title="L2 Cache Requests" name="Write requests" description="The number of L2 cache write requests from internal masters." units="requests"/>
+        <event offset="19" advanced="yes" counter="L2_WR_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Write stall cycles" description="The number of cycles where L2 cache write requests from internal masters are stalled." units="cycles"/>
+        <event offset="20" advanced="yes" counter="L2_SNP_MSG_IN" title="L2 Cache Requests" name="Snoop requests" description="The number of L2 snoop requests from internal masters." units="requests"/>
+        <event offset="21" advanced="yes" counter="L2_SNP_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Snoop stall cycles" description="The number of cycles where L2 cache snoop requests from internal masters are stalled." units="cycles"/>
+        <event offset="22" advanced="yes" counter="L2_RD_MSG_OUT" title="L2 Cache Requests" name="L1 read requests" description="The number of L1 cache read requests sent by the L2 cache to an internal master." units="requests"/>
+        <event offset="23" advanced="yes" counter="L2_RD_MSG_OUT_STALL" title="L2 Cache Stall Cycles" name="L1 read stall cycles" description="The number of cycles where L1 cache read requests sent by the L2 cache to an internal master are stalled." units="cycles"/>
+        <event offset="24" advanced="yes" counter="L2_WR_MSG_OUT" title="L2 Cache Requests" name="L1 write requests" description="The number of L1 cache write responses sent by the L2 cache to an internal master." units="requests"/>
+        <event offset="25" counter="L2_ANY_LOOKUP" title="L2 Cache Lookups" name="Any lookup" description="The number of L2 cache lookups performed." units="requests"/>
+        <event offset="26" counter="L2_READ_LOOKUP" title="L2 Cache Lookups" name="Read lookup" description="The number of L2 cache read lookups performed." units="requests"/>
+        <event offset="27" counter="L2_WRITE_LOOKUP" title="L2 Cache Lookups" name="Write lookup" description="The number of L2 cache write lookups performed." units="requests"/>
+        <event offset="28" advanced="yes" counter="L2_EXT_SNOOP_LOOKUP" title="L2 Cache Lookups" name="External snoop lookups" description="The number of coherency snoop lookups performed that were triggered by an external master." units="requests"/>
+        <event offset="29" counter="L2_EXT_READ" title="External Bus Accesses" name="Read transaction" description="The number of external read transactions." units="transactions"/>
+        <event offset="30" advanced="yes" counter="L2_EXT_READ_NOSNP" title="External Bus Accesses" name="ReadNoSnoop transactions" description="The number of external non-coherent read transactions." units="transactions"/>
+        <event offset="31" advanced="yes" counter="L2_EXT_READ_UNIQUE" title="External Bus Accesses" name="ReadUnique transactions" description="The number of external coherent read unique transactions." units="transactions"/>
+        <event offset="32" counter="L2_EXT_READ_BEATS" title="External Bus Beats" name="Read beat" description="The number of external bus data read cycles." units="beats"/>
+        <event offset="33" counter="L2_EXT_AR_STALL" title="External Bus Stalls" name="Read stall cycles" description="The number of cycles where a read is stalled waiting for the external bus." units="cycles"/>
+        <event offset="34" counter="L2_EXT_AR_CNT_Q1" title="External Bus Outstanding Reads" name="0-25% outstanding" description="The number of read transactions initiated when 0-25% of the maximum are in use." units="transactions"/>
+        <event offset="35" counter="L2_EXT_AR_CNT_Q2" title="External Bus Outstanding Reads" name="25-50% outstanding" description="The number of read transactions initiated when 25-50% of the maximum are in use." units="transactions"/>
+        <event offset="36" counter="L2_EXT_AR_CNT_Q3" title="External Bus Outstanding Reads" name="50-75% outstanding" description="The number of read transactions initiated when 50-75% of the maximum are in use." units="transactions"/>
+        <event offset="37" counter="L2_EXT_RRESP_0_127" title="External Bus Read Latency" name="0-127 cycles" description="The number of data beats returned 0-127 cycles after the read request." units="beats"/>
+        <event offset="38" counter="L2_EXT_RRESP_128_191" title="External Bus Read Latency" name="128-191 cycles" description="The number of data beats returned 128-191 cycles after the read request." units="beats"/>
+        <event offset="39" counter="L2_EXT_RRESP_192_255" title="External Bus Read Latency" name="192-255 cycles" description="The number of data beats returned 192-255 cycles after the read request." units="beats"/>
+        <event offset="40" counter="L2_EXT_RRESP_256_319" title="External Bus Read Latency" name="256-319 cycles" description="The number of data beats returned 256-319 cycles after the read request." units="beats"/>
+        <event offset="41" counter="L2_EXT_RRESP_320_383" title="External Bus Read Latency" name="320-383 cycles" description="The number of data beats returned 320-383 cycles after the read request." units="beats"/>
+        <event offset="42" counter="L2_EXT_WRITE" title="External Bus Accesses" name="Write transaction" description="The number of external write transactions." units="transactions"/>
+        <event offset="43" advanced="yes" counter="L2_EXT_WRITE_NOSNP_FULL" title="External Bus Accesses" name="WriteNoSnoopFull transactions" description="The number of external non-coherent full write transactions." units="transactions"/>
+        <event offset="44" advanced="yes" counter="L2_EXT_WRITE_NOSNP_PTL" title="External Bus Accesses" name="WriteNoSnoopPartial transactions" description="The number of external non-coherent partial write transactions." units="transactions"/>
+        <event offset="45" advanced="yes" counter="L2_EXT_WRITE_SNP_FULL" title="External Bus Accesses" name="WriteSnoopFull transactions" description="The number of external coherent full write transactions." units="transactions"/>
+        <event offset="46" advanced="yes" counter="L2_EXT_WRITE_SNP_PTL" title="External Bus Accesses" name="WriteSnoopPartial transactions" description="The number of external coherent partial write transactions." units="transactions"/>
+        <event offset="47" counter="L2_EXT_WRITE_BEATS" title="External Bus Beats" name="Write beat" description="The number of external bus data write cycles." units="beats"/>
+        <event offset="48" counter="L2_EXT_W_STALL" title="External Bus Stalls" name="Write stall cycles" description="The number of cycles where a write is stalled waiting for the external bus." units="cycles"/>
+        <event offset="49" counter="L2_EXT_AW_CNT_Q1" title="External Bus Outstanding Writes" name="0-25% outstanding" description="The number of write transactions initiated when 0-25% of the maximum are in use." units="transactions"/>
+        <event offset="50" counter="L2_EXT_AW_CNT_Q2" title="External Bus Outstanding Writes" name="25-50% outstanding" description="The number of write transactions initiated when 25-50% of the maximum are in use." units="transactions"/>
+        <event offset="51" counter="L2_EXT_AW_CNT_Q3" title="External Bus Outstanding Writes" name="50-75% outstanding" description="The number of write transactions initiated when 50-75% of the maximum are in use." units="transactions"/>
+        <event offset="52" advanced="yes" counter="L2_EXT_SNOOP" title="External Bus Accesses" name="Snoop transactions" description="The number of coherency snoops triggered by external masters." units="transactions"/>
+        <event offset="53" advanced="yes" counter="L2_EXT_SNOOP_STALL" title="External Bus Stalls" name="Snoop stall cycles" description="The number of cycles where a coherency snoop triggered by external master is stalled." units="cycles"/>
+    </category>
+    <category name="Shader Core" per_cpu="no">
+        <event offset="4" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment active" description="The number of cycles where the shader core is processing a fragment workload." units="cycles"/>
+        <event offset="5" advanced="yes" counter="FRAG_PRIMITIVES_OUT" title="Core Primitives" name="Read primitives" description="The number of primitives read from the tile list by the fragment front-end." units="primitives"/>
+        <event offset="6" counter="FRAG_PRIM_RAST" title="Core Primitives" name="Rasterized primitives" description="The number of primitives being rasterized." units="primitives"/>
+        <event offset="7" counter="FRAG_FPK_ACTIVE" title="Core Cycles" name="Fragment FPKB active" description="The number of cycles where at least one quad is present in the pre-pipe quad queue." units="cycles"/>
+        <event offset="9" counter="FRAG_WARPS" title="Core Warps" name="Fragment warps" description="The number of fragment warps created." units="warps"/>
+        <event offset="10" counter="FRAG_PARTIAL_QUADS_RAST" title="Core Quads" name="Partial rasterized quads" description="The number of partially-rasterized fragment quads created." units="quads"/>
+        <event offset="11" counter="FRAG_QUADS_RAST" title="Core Quads" name="Rasterized quads" description="The number of quads generated by the rasterization phase." units="quads"/>
+        <event offset="12" counter="FRAG_QUADS_EZS_TEST" title="Core Quads" name="Early ZS tested quads" description="The number of quads that are undergoing early depth and stencil testing." units="quads"/>
+        <event offset="13" counter="FRAG_QUADS_EZS_UPDATE" title="Core Quads" name="Early ZS updated quads" description="The number of quads undergoing early depth and stencil testing, that are capable of updating the framebuffer." units="quads"/>
+        <event offset="14" counter="FRAG_QUADS_EZS_KILL" title="Core Quads" name="Early ZS killed quads" description="The number of quads killed by early depth and stencil testing." units="quads"/>
+        <event offset="15" counter="FRAG_LZS_TEST" title="Core Quads" name="Late ZS tested quads" description="The number of quads undergoing late depth and stencil testing." units="quads"/>
+        <event offset="16" counter="FRAG_LZS_KILL" title="Core Quads" name="Late ZS killed quads" description="The number of quads killed by late depth and stencil testing." units="quads"/>
+        <event offset="17" counter="WARP_REG_SIZE_64" title="Core Warps" name="All register warps" description="The number of warps that require more than 32 registers." units="warps"/>
+        <event offset="18" counter="FRAG_PTILES" title="Core Tiles" name="Tiles" description="The number of tiles processed by the shader core." units="tiles"/>
+        <event offset="19" counter="FRAG_TRANS_ELIM" title="Core Tiles" name="Constant tiles killed" description="The number of tiles killed by transaction elimination." units="tiles"/>
+        <event offset="20" counter="QUAD_FPK_KILLER" title="Core Quads" name="FPK occluder quads" description="The number of quads that are valid occluders for hidden surface removal." units="quads"/>
+        <event offset="21" counter="FULL_QUAD_WARPS" title="Core Warps" name="Full quad warps" description="The number of warps that are fully populated with quads." units="warps"/>
+        <event offset="22" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Non-fragment active" description="The number of cycles where the shader core is processing some non-fragment workload." units="cycles"/>
+        <event offset="23" advanced="yes" counter="COMPUTE_TASKS" title="Core Tasks" name="Non-fragment tasks" description="The number of non-fragment tasks issued to the shader core." units="tasks"/>
+        <event offset="24" counter="COMPUTE_WARPS" title="Core Warps" name="Non-fragment warps" description="The number of non-fragment warps created." units="warps"/>
+        <event offset="25" advanced="yes" counter="COMPUTE_STARVING" title="Core Starvation Cycles" name="Non-fragment starvation cycles" description="The number of cycles where the shader core is processing a non-fragment workload and there are no new threads available for execution." units="cycles"/>
+        <event offset="26" counter="EXEC_CORE_ACTIVE" title="Core Cycles" name="Execution core active" description="The number of cycles where the shader core is processing at least one warp." units="cycles"/>
+        <event offset="27" counter="EXEC_INSTR_FMA" title="Core PU Instructions" name="FMA instructions" description="The number of instructions issued to the FMA pipe." units="instructions"/>
+        <event offset="28" counter="EXEC_INSTR_CVT" title="Core PU Instructions" name="CVT instructions" description="The number of instructions issued to the CVT pipe." units="instructions"/>
+        <event offset="29" counter="EXEC_INSTR_SFU" title="Core PU Instructions" name="SFU instructions" description="The number of instructions issued to the SFU pipe." units="instructions"/>
+        <event offset="30" counter="EXEC_INSTR_MSG" title="Core PU Instructions" name="Message instructions" description="The number of instructions issued to the MSG pipe." units="instructions"/>
+        <event offset="31" counter="EXEC_INSTR_DIVERGED" title="Core EE Instructions" name="Diverged instructions" description="The number of instructions executed per warp, that have control flow divergence." units="instructions"/>
+        <event offset="32" advanced="yes" counter="EXEC_ICACHE_MISS" title="Core PU Instructions" name="Instruction cache misses" description="The number of instruction cache misses." units="requests"/>
+        <event offset="33" advanced="yes" counter="EXEC_STARVE_ARITH" title="Core Starvation Cycles" name="Execution engine starvation cycles" description="The number of cycles where the processing unit is starved of work." units="cycles"/>
+        <event offset="34" counter="CALL_BLEND_SHADER" title="Core PU Instructions" name="Blend shader calls" description="The number of blend shader invocations executed." units="instructions"/>
+        <event offset="35" counter="TEX_MSGI_NUM_FLITS" title="Texture Bus" name="Input beats" description="The number of texture request message data beats." units="beats"/>
+        <event offset="36" counter="TEX_DFCH_CLK_STALLED" title="Core Texture Stalls" name="Descriptor stall cycles" description="The number of cycles where a quad is stalled on texture descriptor fetch." units="cycles"/>
+        <event offset="37" counter="TEX_TFCH_CLK_STALLED" title="Core Texture Stalls" name="Fetch queue stall cycles" description="The number of cycles where a quad is stalled on entering texture fetch because the fetch queue is full." units="cycles"/>
+        <event offset="38" counter="TEX_TFCH_STARVED_PENDING_DATA_FETCH" title="Core Texture Stalls" name="Filtering unit stall cycles" description="The number of cycles where the filtering unit is idle and there is at least one quad present in the texture data fetch queue." units="cycles"/>
+        <event offset="39" counter="TEX_FILT_NUM_OPERATIONS" title="Core Texture Cycles" name="Texturing active" description="The number of texture filtering issue cycles." units="cycles"/>
+        <event offset="40" counter="TEX_FILT_NUM_FXR_OPERATIONS" title="Core Texture Cycles" name="4x bilinear filtering active" description="The number of cycles where the filtering unit uses the 4x path to implement nearest or bilinear filtering." units="cycles"/>
+        <event offset="41" counter="TEX_FILT_NUM_FST_OPERATIONS" title="Core Texture Cycles" name="2x trilinear filtering active" description="The number of cycles where the filtering unit uses the 4x path to implement trilinear filtering." units="cycles"/>
+        <event offset="42" counter="TEX_MSGO_NUM_MSG" title="Core Texture Quads" name="Texture requests" description="The number of quad-width texture operations processed by the texture unit." units="quads"/>
+        <event offset="43" counter="TEX_MSGO_NUM_FLITS" title="Texture Bus" name="Output beats" description="The number of texture response message data beats." units="beats"/>
+        <event offset="44" counter="LS_MEM_READ_FULL" title="Core Load/Store Cycles" name="Full read cycles" description="The number of full-width load/store cache reads." units="cycles"/>
+        <event offset="45" counter="LS_MEM_READ_SHORT" title="Core Load/Store Cycles" name="Partial read cycles" description="The number of partial-width load/store cache reads." units="cycles"/>
+        <event offset="46" counter="LS_MEM_WRITE_FULL" title="Core Load/Store Cycles" name="Full write cycles" description="The number of full-width load/store cache writes." units="cycles"/>
+        <event offset="47" counter="LS_MEM_WRITE_SHORT" title="Core Load/Store Cycles" name="Partial write cycles" description="The number of partial-width load/store cache writes." units="cycles"/>
+        <event offset="48" counter="LS_MEM_ATOMIC" title="Core Load/Store Cycles" name="Atomic access cycles" description="The number of load/store atomic accesses." units="cycles"/>
+        <event offset="49" counter="VARY_INSTR" title="Core Varying Requests" name="Interpolation requests" description="The number of warp-width interpolation operations processed by the varying unit." units="instructions"/>
+        <event offset="50" counter="VARY_SLOT_32" title="Core Varying Cycles" name="32-bit interpolation active" description="The number of 32-bit interpolation cycles processed by the varying unit." units="cycles"/>
+        <event offset="51" counter="VARY_SLOT_16" title="Core Varying Cycles" name="16-bit interpolation active" description="The number of 16-bit interpolation cycles processed by the varying unit." units="cycles"/>
+        <event offset="52" advanced="yes" counter="ATTR_INSTR" title="Core Attribute Requests" name="Attribute requests" description="The number of instructions executed by the attribute unit." units="instructions"/>
+        <event offset="53" counter="SHADER_CORE_ACTIVE" title="Core Cycles" name="Any active" description="The number of cycles where the shader core is processing either a non-fragment workload or a fragment workload." units="cycles"/>
+        <event offset="54" counter="BEATS_RD_FTC" title="Core L2 Reads" name="Fragment L2 read beats" description="The number of read beats received by the fixed-function fragment front-end." units="beats"/>
+        <event offset="55" counter="BEATS_RD_FTC_EXT" title="Core L2 Reads" name="Fragment external read beats" description="The number of read beats received by the fixed-function fragment front-end that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="56" counter="BEATS_RD_LSC" title="Core L2 Reads" name="Load/store L2 read beats" description="The number of read beats received by the load/store unit." units="beats"/>
+        <event offset="57" counter="BEATS_RD_LSC_EXT" title="Core L2 Reads" name="Load/store external read beats" description="The number of read beats received by the load/store unit that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="58" counter="BEATS_RD_TEX" title="Core L2 Reads" name="Texture L2 read beats" description="The number of read beats received by the texture unit." units="beats"/>
+        <event offset="59" counter="BEATS_RD_TEX_EXT" title="Core L2 Reads" name="Texture external read beats" description="The number of read beats received by the texture unit that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="60" advanced="yes" counter="BEATS_RD_OTHER" title="Core L2 Reads" name="Other L2 read beats" description="The number of read beats received by a unit that is not specifically identified." units="beats"/>
+        <event offset="61" counter="BEATS_WR_LSC_OTHER" title="Core Writes" name="Load/store other write beats" description="The number of write beats by the load/store unit that are due to any reason other than writeback." units="beats"/>
+        <event offset="62" counter="BEATS_WR_TIB" title="Core Writes" name="Tile buffer write beats" description="The number of write beats sent by the tile buffer writeback unit." units="beats"/>
+        <event offset="63" counter="BEATS_WR_LSC_WB" title="Core Writes" name="Load/store writeback write beats" description="The number of write beats by the load/store unit that are due to writeback." units="beats"/>
     </category>
 </metrics>
\ No newline at end of file
diff --git a/src/panfrost/perf/G71.xml b/src/panfrost/perf/G71.xml
index 080fd619f93..760e045b588 100644
--- a/src/panfrost/perf/G71.xml
+++ b/src/panfrost/perf/G71.xml
@@ -1,6 +1,6 @@
 <!--
 Copyright © 2017-2020 ARM Limited.
-Copyright © 2021 Collabora, Ltd.
+Copyright © 2021-2022 Collabora, Ltd.
 Author: Antonio Caggiano <antonio.caggiano@collabora.com>
 
 Permission is hereby granted, free of charge, to any person obtaining a
@@ -24,154 +24,154 @@ SOFTWARE.
 -->
 <metrics id="TMIx">
     <category name="Job Manager" per_cpu="no">
-        <event offset="6" counter="GPU_ACTIVE" title="GPU Cycles" name="GPU active" description="The number of cycles where the GPU has a workload of any type queued for processing." units="cycles" />
-        <event offset="7" advanced="yes" counter="IRQ_ACTIVE" title="GPU Cycles" name="Interrupt active" description="The number of cycles where the GPU has a pending interrupt." units="cycles" />
-        <event offset="8" advanced="yes" counter="JS0_JOBS" title="GPU Jobs" name="Fragment jobs" description="The number of jobs processed by the GPU fragment queue." units="jobs" />
-        <event offset="9" counter="JS0_TASKS" title="GPU Tasks" name="Fragment tasks" description="The number of 32x32 pixel tasks processed by the GPU fragment queue." units="tasks" />
-        <event offset="10" counter="JS0_ACTIVE" title="GPU Cycles" name="Fragment queue active" description="The number of cycles where work is queued for processing in the GPU fragment queue." units="cycles" />
-        <event offset="12" advanced="yes" counter="JS0_WAIT_READ" title="GPU Wait Cycles" name="Fragment descriptor reads cycles" description="The number of cycles where queued fragment work is waiting for a descriptor load." units="cycles" />
-        <event offset="13" advanced="yes" counter="JS0_WAIT_ISSUE" title="GPU Wait Cycles" name="Fragment job issue cycles" description="The number of cycles where queued fragment work is waiting for an available processor." units="cycles" />
-        <event offset="14" advanced="yes" counter="JS0_WAIT_DEPEND" title="GPU Wait Cycles" name="Fragment job dependency cycles" description="The number of cycles where queued fragment work is waiting for dependent work to complete." units="cycles" />
-        <event offset="15" advanced="yes" counter="JS0_WAIT_FINISH" title="GPU Wait Cycles" name="Fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued fragment work to complete." units="cycles" />
-        <event offset="16" advanced="yes" counter="JS1_JOBS" title="GPU Jobs" name="Non-fragment jobs" description="The number of jobs processed by the GPU non-fragment queue." units="jobs" />
-        <event offset="17" advanced="yes" counter="JS1_TASKS" title="GPU Tasks" name="Non-fragment tasks" description="The number of tasks processed by the GPU non-fragment queue." units="tasks" />
-        <event offset="18" counter="JS1_ACTIVE" title="GPU Cycles" name="Non-fragment queue active" description="The number of cycles where work is queued in the GPU non-fragment queue." units="cycles" />
-        <event offset="20" advanced="yes" counter="JS1_WAIT_READ" title="GPU Wait Cycles" name="Non-fragment descriptor read cycles" description="The number number of cycles where queued non-fragment work is waiting for a descriptor load." units="cycles" />
-        <event offset="21" advanced="yes" counter="JS1_WAIT_ISSUE" title="GPU Wait Cycles" name="Non-fragment job issue cycles" description="The number of cycles where queued non-fragment work is waiting for an available processor." units="cycles" />
-        <event offset="22" advanced="yes" counter="JS1_WAIT_DEPEND" title="GPU Wait Cycles" name="Non-fragment job dependency cycles" description="The number of cycles where queued non-fragment work is waiting for dependent work to complete." units="cycles" />
-        <event offset="23" advanced="yes" counter="JS1_WAIT_FINISH" title="GPU Wait Cycles" name="Non-fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued non-fragment work to complete." units="cycles" />
-        <event offset="24" advanced="yes" counter="JS2_JOBS" title="GPU Jobs" name="Reserved jobs" description="The number of jobs processed by the GPU reserved queue." units="jobs" />
-        <event offset="25" advanced="yes" counter="JS2_TASKS" title="GPU Tasks" name="Reserved tasks" description="The number of tasks processed by the GPU reserved queue." units="tasks" />
-        <event offset="26" advanced="yes" counter="JS2_ACTIVE" title="GPU Cycles" name="Reserved queue active" description="The number of cycles where work is queued in the GPU reserved queue." units="cycles" />
-        <event offset="28" advanced="yes" counter="JS2_WAIT_READ" title="GPU Wait Cycles" name="Reserved descriptor read cycles" description="The number of cycles where queued reserved work is waiting for a descriptor load." units="cycles" />
-        <event offset="29" advanced="yes" counter="JS2_WAIT_ISSUE" title="GPU Wait Cycles" name="Reserved job issue cycles" description="The number of cycles where queued reserved work is waiting for an available processor." units="cycles" />
-        <event offset="30" advanced="yes" counter="JS2_WAIT_DEPEND" title="GPU Wait Cycles" name="Reserved job dependency cycles" description="The number of cycles where queued reserved work is waiting for dependent work to complete." units="cycles" />
-        <event offset="31" advanced="yes" counter="JS2_WAIT_FINISH" title="GPU Wait Cycles" name="Reserved job finish cycles" description="The number of cycles where the GPU is waiting for issued reserved work to complete." units="cycles" />
+        <event offset="6" counter="GPU_ACTIVE" title="GPU Cycles" name="GPU active" description="The number of cycles where the GPU has a workload of any type queued for processing." units="cycles"/>
+        <event offset="7" advanced="yes" counter="IRQ_ACTIVE" title="GPU Cycles" name="Interrupt active" description="The number of cycles where the GPU has a pending interrupt." units="cycles"/>
+        <event offset="8" advanced="yes" counter="JS0_JOBS" title="GPU Jobs" name="Fragment jobs" description="The number of jobs processed by the GPU fragment queue." units="jobs"/>
+        <event offset="9" counter="JS0_TASKS" title="GPU Tasks" name="Fragment tasks" description="The number of 32x32 pixel tasks processed by the GPU fragment queue." units="tasks"/>
+        <event offset="10" counter="JS0_ACTIVE" title="GPU Cycles" name="Fragment queue active" description="The number of cycles where work is queued for processing in the GPU fragment queue." units="cycles"/>
+        <event offset="12" advanced="yes" counter="JS0_WAIT_READ" title="GPU Wait Cycles" name="Fragment descriptor reads cycles" description="The number of cycles where queued fragment work is waiting for a descriptor load." units="cycles"/>
+        <event offset="13" advanced="yes" counter="JS0_WAIT_ISSUE" title="GPU Wait Cycles" name="Fragment job issue cycles" description="The number of cycles where queued fragment work is waiting for an available processor." units="cycles"/>
+        <event offset="14" advanced="yes" counter="JS0_WAIT_DEPEND" title="GPU Wait Cycles" name="Fragment job dependency cycles" description="The number of cycles where queued fragment work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="15" advanced="yes" counter="JS0_WAIT_FINISH" title="GPU Wait Cycles" name="Fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued fragment work to complete." units="cycles"/>
+        <event offset="16" advanced="yes" counter="JS1_JOBS" title="GPU Jobs" name="Non-fragment jobs" description="The number of jobs processed by the GPU non-fragment queue." units="jobs"/>
+        <event offset="17" advanced="yes" counter="JS1_TASKS" title="GPU Tasks" name="Non-fragment tasks" description="The number of tasks processed by the GPU non-fragment queue." units="tasks"/>
+        <event offset="18" counter="JS1_ACTIVE" title="GPU Cycles" name="Non-fragment queue active" description="The number of cycles where work is queued in the GPU non-fragment queue." units="cycles"/>
+        <event offset="20" advanced="yes" counter="JS1_WAIT_READ" title="GPU Wait Cycles" name="Non-fragment descriptor read cycles" description="The number number of cycles where queued non-fragment work is waiting for a descriptor load." units="cycles"/>
+        <event offset="21" advanced="yes" counter="JS1_WAIT_ISSUE" title="GPU Wait Cycles" name="Non-fragment job issue cycles" description="The number of cycles where queued non-fragment work is waiting for an available processor." units="cycles"/>
+        <event offset="22" advanced="yes" counter="JS1_WAIT_DEPEND" title="GPU Wait Cycles" name="Non-fragment job dependency cycles" description="The number of cycles where queued non-fragment work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="23" advanced="yes" counter="JS1_WAIT_FINISH" title="GPU Wait Cycles" name="Non-fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued non-fragment work to complete." units="cycles"/>
+        <event offset="24" advanced="yes" counter="JS2_JOBS" title="GPU Jobs" name="Reserved jobs" description="The number of jobs processed by the GPU reserved queue." units="jobs"/>
+        <event offset="25" advanced="yes" counter="JS2_TASKS" title="GPU Tasks" name="Reserved tasks" description="The number of tasks processed by the GPU reserved queue." units="tasks"/>
+        <event offset="26" advanced="yes" counter="JS2_ACTIVE" title="GPU Cycles" name="Reserved queue active" description="The number of cycles where work is queued in the GPU reserved queue." units="cycles"/>
+        <event offset="28" advanced="yes" counter="JS2_WAIT_READ" title="GPU Wait Cycles" name="Reserved descriptor read cycles" description="The number of cycles where queued reserved work is waiting for a descriptor load." units="cycles"/>
+        <event offset="29" advanced="yes" counter="JS2_WAIT_ISSUE" title="GPU Wait Cycles" name="Reserved job issue cycles" description="The number of cycles where queued reserved work is waiting for an available processor." units="cycles"/>
+        <event offset="30" advanced="yes" counter="JS2_WAIT_DEPEND" title="GPU Wait Cycles" name="Reserved job dependency cycles" description="The number of cycles where queued reserved work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="31" advanced="yes" counter="JS2_WAIT_FINISH" title="GPU Wait Cycles" name="Reserved job finish cycles" description="The number of cycles where the GPU is waiting for issued reserved work to complete." units="cycles"/>
     </category>
     <category name="Tiler" per_cpu="no">
-        <event offset="68" counter="TILER_ACTIVE" title="GPU Cycles" name="Tiler active" description="The number of cycles where the tiler has a workload queued for processing." units="cycles" />
-        <event offset="70" counter="TRIANGLES" title="Input Primitives" name="Triangle primitives" description="The number of input triangle primitives." units="primitives" />
-        <event offset="71" counter="LINES" title="Input Primitives" name="Line primitives" description="The number of input line primitives." units="primitives" />
-        <event offset="72" counter="POINTS" title="Input Primitives" name="Point primitives" description="The number of input point primitives." units="primitives" />
-        <event offset="73" counter="FRONT_FACING" title="Visible Primitives" name="Front-facing primitives" description="The number of front-facing triangles that are visible after culling." units="primitives" />
-        <event offset="74" counter="BACK_FACING" title="Visible Primitives" name="Back-facing primitives" description="The number of back-facing triangles that are visible after culling." units="primitives" />
-        <event offset="75" counter="PRIM_VISIBLE" title="Primitive Culling" name="Visible primitives" description="The number of primitives that are visible after culling." units="primitives" />
-        <event offset="76" counter="PRIM_CULLED" title="Primitive Culling" name="Facing and XY plane test culled primitives" description="The number of primitives that are culled by facing or frustum XY plane tests." units="primitives" />
-        <event offset="77" counter="PRIM_CLIPPED" title="Primitive Culling" name="Z plane test culled primitives" description="The number of primitives that are culled by frustum Z plane tests." units="primitives" />
-        <event offset="78" counter="PRIM_SAT_CULLED" title="Primitive Culling" name="Sample test culled primitives" description="The number of primitives culled by the sample coverage test." units="primitives" />
-        <event offset="81" advanced="yes" counter="BUS_READ" title="Tiler L2 Accesses" name="Read beats" description="The number of internal bus data read cycles made by the tiler." units="beats" />
-        <event offset="83" advanced="yes" counter="BUS_WRITE" title="Tiler L2 Accesses" name="Write beats" description="The number of internal bus data write cycles made by the tiler." units="beats" />
-        <event offset="85" counter="IDVS_POS_SHAD_REQ" title="Tiler Shading Requests" name="Position shading requests" description="The number of position shading requests in the IDVS flow." units="requests" />
-        <event offset="87" advanced="yes" counter="IDVS_POS_SHAD_STALL" title="Tiler Cycles" name="Position shading stall cycles" description="The number of cycles where the tiler has a stalled position shading request." units="cycles" />
-        <event offset="88" advanced="yes" counter="IDVS_POS_FIFO_FULL" title="Tiler Cycles" name="Position FIFO full cycles" description="The number of cycles where the tiler has a stalled position shading buffer." units="cycles" />
-        <event offset="90" advanced="yes" counter="VCACHE_HIT" title="Tiler Vertex Cache" name="Position cache hits" description="The number of position lookups that result in a hit in the vertex cache." units="requests" />
-        <event offset="91" advanced="yes" counter="VCACHE_MISS" title="Tiler Vertex Cache" name="Position cache misses" description="The number of position lookups that miss in the vertex cache." units="requests" />
-        <event offset="95" advanced="yes" counter="VFETCH_STALL" title="Tiler Cycles" name="Primitive assembly busy stall cycles" description="The number of cycles where the tiler is stalled waiting for primitive assembly." units="cycles" />
-        <event offset="98" advanced="yes" counter="IDVS_VBU_HIT" title="Tiler Vertex Cache" name="Varying cache hits" description="The number of varying lookups that result in a hit in the vertex cache." units="requests" />
-        <event offset="99" advanced="yes" counter="IDVS_VBU_MISS" title="Tiler Vertex Cache" name="Varying cache misses" description="The number of varying lookups that miss in the vertex cache." units="requests" />
-        <event offset="101" counter="IDVS_VAR_SHAD_REQ" title="Tiler Shading Requests" name="Varying shading requests" description="The number of varying shading requests in the IDVS flow." units="requests" />
-        <event offset="102" advanced="yes" counter="IDVS_VAR_SHAD_STALL" title="Tiler Cycles" name="Varying shading stall cycles" description="The number of cycles where the tiler has a stalled varying shading request." units="cycles" />
-        <event offset="118" advanced="yes" counter="WRBUF_NO_AXI_ID_STALL" title="Tiler Cycles" name="Write buffer transaction stall cycles" description="The number of cycles where the tiler write buffer can not send data because it has no available write IDs." units="cycles" />
-        <event offset="119" advanced="yes" counter="WRBUF_AXI_STALL" title="Tiler Cycles" name="Write buffer write stall cycles" description="The number of cycles where the tiler write buffer can not send data because the bus is not ready." units="cycles" />
-    </category>
-    <category name="Shader Core" per_cpu="no">
-        <event offset="196" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment active" description="The number of cycles where the shader core is processing a fragment workload." units="cycles" />
-        <event offset="197" advanced="yes" counter="FRAG_PRIMITIVES" title="Core Primitives" name="Read primitives" description="The number of primitives read from the tile list by the fragment front-end." units="primitives" />
-        <event offset="198" counter="FRAG_PRIM_RAST" title="Core Primitives" name="Rasterized primitives" description="The number of primitives being rasterized." units="primitives" />
-        <event offset="199" counter="FRAG_FPK_ACTIVE" title="Core Cycles" name="Fragment FPKB active" description="The number of cycles where at least one quad is present in the pre-pipe quad queue." units="cycles" />
-        <event offset="201" counter="FRAG_WARPS" title="Core Warps" name="Fragment warps" description="The number of fragment warps created." units="warps" />
-        <event offset="202" counter="FRAG_PARTIAL_WARPS" title="Core Warps" name="Partial fragment warps" description="The number of fragment warps containing helper threads that do not correspond to a hit sample point." units="warps" />
-        <event offset="203" counter="FRAG_QUADS_RAST" title="Core Quads" name="Rasterized quads" description="The number of quads generated by the rasterization phase." units="quads" />
-        <event offset="204" counter="FRAG_QUADS_EZS_TEST" title="Core Quads" name="Early ZS tested quads" description="The number of quads that are undergoing early depth and stencil testing." units="quads" />
-        <event offset="205" counter="FRAG_QUADS_EZS_UPDATE" title="Core Quads" name="Early ZS updated quads" description="The number of quads undergoing early depth and stencil testing, that are capable of updating the framebuffer." units="quads" />
-        <event offset="206" counter="FRAG_QUADS_EZS_KILL" title="Core Quads" name="Early ZS killed quads" description="The number of quads killed by early depth and stencil testing." units="quads" />
-        <event offset="207" counter="FRAG_LZS_TEST" title="Core Quads" name="Late ZS tested quads" description="The number of quads undergoing late depth and stencil testing." units="quads" />
-        <event offset="208" counter="FRAG_LZS_KILL" title="Core Quads" name="Late ZS killed quads" description="The number of quads killed by late depth and stencil testing." units="quads" />
-        <event offset="210" counter="FRAG_PTILES" title="Core Tiles" name="Tiles" description="The number of tiles processed by the shader core." units="tiles" />
-        <event offset="211" counter="FRAG_TRANS_ELIM" title="Core Tiles" name="Constant tiles killed" description="The number of tiles killed by transaction elimination." units="tiles" />
-        <event offset="212" counter="QUAD_FPK_KILLER" title="Core Quads" name="FPK occluder quads" description="The number of quads that are valid occluders for hidden surface removal." units="quads" />
-        <event offset="214" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Non-fragment active" description="The number of cycles where the shader core is processing some non-fragment workload." units="cycles" />
-        <event offset="215" advanced="yes" counter="COMPUTE_TASKS" title="Core Tasks" name="Non-fragment tasks" description="The number of non-fragment tasks issued to the shader core." units="tasks" />
-        <event offset="216" counter="COMPUTE_WARPS" title="Core Warps" name="Non-fragment warps" description="The number of non-fragment warps created." units="warps" />
-        <event offset="217" advanced="yes" counter="COMPUTE_STARVING" title="Core Starvation Cycles" name="Non-fragment starvation cycles" description="The number of cycles where the shader core is processing a non-fragment workload and there are no new threads available for execution." units="cycles" />
-        <event offset="218" counter="EXEC_CORE_ACTIVE" title="Core Cycles" name="Execution core active" description="The number of cycles where the shader core is processing at least one warp." units="cycles" />
-        <event offset="219" advanced="yes" counter="EXEC_ACTIVE" title="Core Cycles" name="Execution engine active" description="The number of cycles where the execution engine unit is processing at least one thread." units="cycles" />
-        <event offset="220" counter="EXEC_INSTR_COUNT" title="Core EE Instructions" name="Executed instructions" description="The number of instructions executed per warp." units="instructions" />
-        <event offset="221" counter="EXEC_INSTR_DIVERGED" title="Core EE Instructions" name="Diverged instructions" description="The number of instructions executed per warp, that have control flow divergence." units="instructions" />
-        <event offset="222" advanced="yes" counter="EXEC_INSTR_STARVING" title="Core Starvation Cycles" name="Execution engine starvation cycles" description="The number of cycles where no new threads are available for execution." units="cycles" />
-        <event offset="223" advanced="yes" counter="ARITH_INSTR_SINGLE_FMA" title="Core EE Instructions" name="Arithmetic instructions" description="The number of instructions where the workload is a single FMA pipe arithmetic operation." units="instructions" />
-        <event offset="224" advanced="yes" counter="ARITH_INSTR_DOUBLE" title="Core EE Instructions" name="Dual Arithmetic instructions" description="The number of instructions where the workload is one FMA pipe arithmetic operation and one ADD pipe arithmetic operation." units="instructions" />
-        <event offset="225" advanced="yes" counter="ARITH_INSTR_MSG" title="Core EE Instructions" name="Arithmetic + Message instructions" description="The number of instructions where the workload is one FMA pipe arithmetic operation and one ADD pipe message operation" units="instructions" />
-        <event offset="226" advanced="yes" counter="ARITH_INSTR_MSG_ONLY" title="Core EE Instructions" name="Message instructions" description="The number of instructions where the workload is a single ADD pipe message operation, with no FMA pipe operation" units="instructions" />
-        <event offset="227" counter="TEX_INSTR" title="Core Texture Requests" name="Texture requests" description="The number of thread-width texture operations processed." units="instructions" />
-        <event offset="228" counter="TEX_INSTR_MIPMAP" title="Core Texture Requests" name="Mipmapped texture request" description="The number of texture operations that act on a mipmapped texture." units="instructions" />
-        <event offset="229" counter="TEX_INSTR_COMPRESSED" title="Core Texture Requests" name="Compressed texture requests" description="The number of texture operations acting on a compressed texture." units="instructions" />
-        <event offset="230" counter="TEX_INSTR_3D" title="Core Texture Requests" name="3D texture requests" description="The number of texture operations acting on a 3D texture." units="instructions" />
-        <event offset="231" counter="TEX_INSTR_TRILINEAR" title="Core Texture Requests" name="Trilinear filtered requests" description="The number of texture operations using a trilinear texture filter." units="instructions" />
-        <event offset="232" counter="TEX_COORD_ISSUE" title="Core Texture Cycles" name="Texturing active" description="The number of texture filtering issue cycles." units="cycles" />
-        <event offset="233" advanced="yes" counter="TEX_COORD_STALL" title="Core Texture Cycles" name="Coordinate stall cycles" description="The number of clock cycles where threads are stalled at the texel coordinate calculation stage." units="cycles" />
-        <event offset="234" advanced="yes" counter="TEX_STARVE_CACHE" title="Core Texture Cycles" name="Line fill stall cycles" description="The number of clock cycles where at least one thread is waiting for data from the texture cache, but no lookup is completed." units="cycles" />
-        <event offset="235" advanced="yes" counter="TEX_STARVE_FILTER" title="Core Texture Cycles" name="Partial data stall cycles" description="The number of clock cycles where at least one thread fetched some data from the texture cache, but no filtering operation is started." units="cycles" />
-        <event offset="236" counter="LS_MEM_READ_FULL" title="Core Load/Store Cycles" name="Full read cycles" description="The number of full-width load/store cache reads." units="cycles" />
-        <event offset="237" counter="LS_MEM_READ_SHORT" title="Core Load/Store Cycles" name="Partial read cycles" description="The number of partial-width load/store cache reads." units="cycles" />
-        <event offset="238" counter="LS_MEM_WRITE_FULL" title="Core Load/Store Cycles" name="Full write cycles" description="The number of full-width load/store cache writes." units="cycles" />
-        <event offset="239" counter="LS_MEM_WRITE_SHORT" title="Core Load/Store Cycles" name="Partial write cycles" description="The number of partial-width load/store cache writes." units="cycles" />
-        <event offset="240" counter="LS_MEM_ATOMIC" title="Core Load/Store Cycles" name="Atomic access cycles" description="The number of load/store atomic accesses." units="cycles" />
-        <event offset="241" counter="VARY_INSTR" title="Core Varying Requests" name="Interpolation requests" description="The number of warp-width interpolation operations processed by the varying unit." units="instructions" />
-        <event offset="242" counter="VARY_SLOT_32" title="Core Varying Cycles" name="32-bit interpolation active" description="The number of 32-bit interpolation cycles processed by the varying unit." units="cycles" />
-        <event offset="243" counter="VARY_SLOT_16" title="Core Varying Cycles" name="16-bit interpolation active" description="The number of 16-bit interpolation cycles processed by the varying unit." units="cycles" />
-        <event offset="244" advanced="yes" counter="ATTR_INSTR" title="Core Attribute Requests" name="Attribute requests" description="The number of instructions executed by the attribute unit." units="instructions" />
-        <event offset="245" advanced="yes" counter="ARITH_INSTR_FP_MUL" title="Core EE Instructions" name="Multiplier instructions" description="The number of instructions where the workload uses floating-point multiplier hardware." units="instructions" />
-        <event offset="246" counter="BEATS_RD_FTC" title="Core L2 Reads" name="Fragment L2 read beats" description="The number of read beats received by the fixed-function fragment front-end." units="beats" />
-        <event offset="247" counter="BEATS_RD_FTC_EXT" title="Core L2 Reads" name="Fragment external read beats" description="The number of read beats received by the fixed-function fragment front-end that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="248" counter="BEATS_RD_LSC" title="Core L2 Reads" name="Load/store L2 read beats" description="The number of read beats received by the load/store unit." units="beats" />
-        <event offset="249" counter="BEATS_RD_LSC_EXT" title="Core L2 Reads" name="Load/store external read beats" description="The number of read beats received by the load/store unit that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="250" counter="BEATS_RD_TEX" title="Core L2 Reads" name="Texture L2 read beats" description="The number of read beats received by the texture unit." units="beats" />
-        <event offset="251" counter="BEATS_RD_TEX_EXT" title="Core L2 Reads" name="Texture external read beats" description="The number of read beats received by the texture unit that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="252" advanced="yes" counter="BEATS_RD_OTHER" title="Core L2 Reads" name="Other L2 read beats" description="The number of read beats received by a unit that is not specifically identified." units="beats" />
-        <event offset="253" counter="BEATS_WR_LSC" title="Core Writes" name="Load/store write beats" description="The number of write beats sent by the load/store unit." units="beats" />
-        <event offset="254" counter="BEATS_WR_TIB" title="Core Writes" name="Tile buffer write beats" description="The number of write beats sent by the tile buffer writeback unit." units="beats" />
-        <event offset="255" advanced="yes" counter="BEATS_WR_OTHER" title="Core Writes" name="Other write beats" description="The number of write beats sent by any unit that is not specifically identified." units="beats" />
+        <event offset="4" counter="TILER_ACTIVE" title="GPU Cycles" name="Tiler active" description="The number of cycles where the tiler has a workload queued for processing." units="cycles"/>
+        <event offset="6" counter="TRIANGLES" title="Input Primitives" name="Triangle primitives" description="The number of input triangle primitives." units="primitives"/>
+        <event offset="7" counter="LINES" title="Input Primitives" name="Line primitives" description="The number of input line primitives." units="primitives"/>
+        <event offset="8" counter="POINTS" title="Input Primitives" name="Point primitives" description="The number of input point primitives." units="primitives"/>
+        <event offset="9" counter="FRONT_FACING" title="Visible Primitives" name="Front-facing primitives" description="The number of front-facing triangles that are visible after culling." units="primitives"/>
+        <event offset="10" counter="BACK_FACING" title="Visible Primitives" name="Back-facing primitives" description="The number of back-facing triangles that are visible after culling." units="primitives"/>
+        <event offset="11" counter="PRIM_VISIBLE" title="Primitive Culling" name="Visible primitives" description="The number of primitives that are visible after culling." units="primitives"/>
+        <event offset="12" counter="PRIM_CULLED" title="Primitive Culling" name="Facing and XY plane test culled primitives" description="The number of primitives that are culled by facing or frustum XY plane tests." units="primitives"/>
+        <event offset="13" counter="PRIM_CLIPPED" title="Primitive Culling" name="Z plane test culled primitives" description="The number of primitives that are culled by frustum Z plane tests." units="primitives"/>
+        <event offset="14" counter="PRIM_SAT_CULLED" title="Primitive Culling" name="Sample test culled primitives" description="The number of primitives culled by the sample coverage test." units="primitives"/>
+        <event offset="17" advanced="yes" counter="BUS_READ" title="Tiler L2 Accesses" name="Read beats" description="The number of internal bus data read cycles made by the tiler." units="beats"/>
+        <event offset="19" advanced="yes" counter="BUS_WRITE" title="Tiler L2 Accesses" name="Write beats" description="The number of internal bus data write cycles made by the tiler." units="beats"/>
+        <event offset="21" counter="IDVS_POS_SHAD_REQ" title="Tiler Shading Requests" name="Position shading requests" description="The number of position shading requests in the IDVS flow." units="requests"/>
+        <event offset="23" advanced="yes" counter="IDVS_POS_SHAD_STALL" title="Tiler Cycles" name="Position shading stall cycles" description="The number of cycles where the tiler has a stalled position shading request." units="cycles"/>
+        <event offset="24" advanced="yes" counter="IDVS_POS_FIFO_FULL" title="Tiler Cycles" name="Position FIFO full cycles" description="The number of cycles where the tiler has a stalled position shading buffer." units="cycles"/>
+        <event offset="26" advanced="yes" counter="VCACHE_HIT" title="Tiler Vertex Cache" name="Position cache hits" description="The number of position lookups that result in a hit in the vertex cache." units="requests"/>
+        <event offset="27" advanced="yes" counter="VCACHE_MISS" title="Tiler Vertex Cache" name="Position cache misses" description="The number of position lookups that miss in the vertex cache." units="requests"/>
+        <event offset="31" advanced="yes" counter="VFETCH_STALL" title="Tiler Cycles" name="Primitive assembly busy stall cycles" description="The number of cycles where the tiler is stalled waiting for primitive assembly." units="cycles"/>
+        <event offset="34" advanced="yes" counter="IDVS_VBU_HIT" title="Tiler Vertex Cache" name="Varying cache hits" description="The number of varying lookups that result in a hit in the vertex cache." units="requests"/>
+        <event offset="35" advanced="yes" counter="IDVS_VBU_MISS" title="Tiler Vertex Cache" name="Varying cache misses" description="The number of varying lookups that miss in the vertex cache." units="requests"/>
+        <event offset="37" counter="IDVS_VAR_SHAD_REQ" title="Tiler Shading Requests" name="Varying shading requests" description="The number of varying shading requests in the IDVS flow." units="requests"/>
+        <event offset="38" advanced="yes" counter="IDVS_VAR_SHAD_STALL" title="Tiler Cycles" name="Varying shading stall cycles" description="The number of cycles where the tiler has a stalled varying shading request." units="cycles"/>
+        <event offset="54" advanced="yes" counter="WRBUF_NO_AXI_ID_STALL" title="Tiler Cycles" name="Write buffer transaction stall cycles" description="The number of cycles where the tiler write buffer can not send data because it has no available write IDs." units="cycles"/>
+        <event offset="55" advanced="yes" counter="WRBUF_AXI_STALL" title="Tiler Cycles" name="Write buffer write stall cycles" description="The number of cycles where the tiler write buffer can not send data because the bus is not ready." units="cycles"/>
     </category>
     <category name="Memory System" per_cpu="no">
-        <event offset="132" advanced="yes" counter="MMU_REQUESTS" title="MMU Stage 1 Translations" name="MMU lookups" description="The number of main MMU address translations performed." units="requests" />
-        <event offset="144" advanced="yes" counter="L2_RD_MSG_IN" title="L2 Cache Requests" name="Read requests" description="The number of L2 cache read requests from internal masters." units="requests" />
-        <event offset="145" advanced="yes" counter="L2_RD_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Read stall cycles" description="The number of cycles L2 cache read requests from internal masters are stalled." units="cycles" />
-        <event offset="146" advanced="yes" counter="L2_WR_MSG_IN" title="L2 Cache Requests" name="Write requests" description="The number of L2 cache write requests from internal masters." units="requests" />
-        <event offset="147" advanced="yes" counter="L2_WR_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Write stall cycles" description="The number of cycles where L2 cache write requests from internal masters are stalled." units="cycles" />
-        <event offset="148" advanced="yes" counter="L2_SNP_MSG_IN" title="L2 Cache Requests" name="Snoop requests" description="The number of L2 snoop requests from internal masters." units="requests" />
-        <event offset="149" advanced="yes" counter="L2_SNP_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Snoop stall cycles" description="The number of cycles where L2 cache snoop requests from internal masters are stalled." units="cycles" />
-        <event offset="150" advanced="yes" counter="L2_RD_MSG_OUT" title="L2 Cache Requests" name="L1 read requests" description="The number of L1 cache read requests sent by the L2 cache to an internal master." units="requests" />
-        <event offset="151" advanced="yes" counter="L2_RD_MSG_OUT_STALL" title="L2 Cache Stall Cycles" name="L1 read stall cycles" description="The number of cycles where L1 cache read requests sent by the L2 cache to an internal master are stalled." units="cycles" />
-        <event offset="152" advanced="yes" counter="L2_WR_MSG_OUT" title="L2 Cache Requests" name="L1 write requests" description="The number of L1 cache write responses sent by the L2 cache to an internal master." units="requests" />
-        <event offset="153" counter="L2_ANY_LOOKUP" title="L2 Cache Lookups" name="Any lookup" description="The number of L2 cache lookups performed." units="requests" />
-        <event offset="154" counter="L2_READ_LOOKUP" title="L2 Cache Lookups" name="Read lookup" description="The number of L2 cache read lookups performed." units="requests" />
-        <event offset="155" counter="L2_WRITE_LOOKUP" title="L2 Cache Lookups" name="Write lookup" description="The number of L2 cache write lookups performed." units="requests" />
-        <event offset="156" advanced="yes" counter="L2_EXT_SNOOP_LOOKUP" title="L2 Cache Lookups" name="External snoop lookups" description="The number of coherency snoop lookups performed that were triggered by an external master." units="requests" />
-        <event offset="157" counter="L2_EXT_READ" title="External Bus Accesses" name="Read transaction" description="The number of external read transactions." units="transactions" />
-        <event offset="158" advanced="yes" counter="L2_EXT_READ_NOSNP" title="External Bus Accesses" name="ReadNoSnoop transactions" description="The number of external non-coherent read transactions." units="transactions" />
-        <event offset="159" advanced="yes" counter="L2_EXT_READ_UNIQUE" title="External Bus Accesses" name="ReadUnique transactions" description="The number of external coherent read unique transactions." units="transactions" />
-        <event offset="160" counter="L2_EXT_READ_BEATS" title="External Bus Beats" name="Read beat" description="The number of external bus data read cycles." units="beats" />
-        <event offset="161" counter="L2_EXT_AR_STALL" title="External Bus Stalls" name="Read stall cycles" description="The number of cycles where a read is stalled waiting for the external bus." units="cycles" />
-        <event offset="162" counter="L2_EXT_AR_CNT_Q1" title="External Bus Outstanding Reads" name="0-25% outstanding" description="The number of read transactions initiated when 0-25% of the maximum are in use." units="transactions" />
-        <event offset="163" counter="L2_EXT_AR_CNT_Q2" title="External Bus Outstanding Reads" name="25-50% outstanding" description="The number of read transactions initiated when 25-50% of the maximum are in use." units="transactions" />
-        <event offset="164" counter="L2_EXT_AR_CNT_Q3" title="External Bus Outstanding Reads" name="50-75% outstanding" description="The number of read transactions initiated when 50-75% of the maximum are in use." units="transactions" />
-        <event offset="165" counter="L2_EXT_RRESP_0_127" title="External Bus Read Latency" name="0-127 cycles" description="The number of data beats returned 0-127 cycles after the read request." units="beats" />
-        <event offset="166" counter="L2_EXT_RRESP_128_191" title="External Bus Read Latency" name="128-191 cycles" description="The number of data beats returned 128-191 cycles after the read request." units="beats" />
-        <event offset="167" counter="L2_EXT_RRESP_192_255" title="External Bus Read Latency" name="192-255 cycles" description="The number of data beats returned 192-255 cycles after the read request." units="beats" />
-        <event offset="168" counter="L2_EXT_RRESP_256_319" title="External Bus Read Latency" name="256-319 cycles" description="The number of data beats returned 256-319 cycles after the read request." units="beats" />
-        <event offset="169" counter="L2_EXT_RRESP_320_383" title="External Bus Read Latency" name="320-383 cycles" description="The number of data beats returned 320-383 cycles after the read request." units="beats" />
-        <event offset="170" counter="L2_EXT_WRITE" title="External Bus Accesses" name="Write transaction" description="The number of external write transactions." units="transactions" />
-        <event offset="171" advanced="yes" counter="L2_EXT_WRITE_NOSNP_FULL" title="External Bus Accesses" name="WriteNoSnoopFull transactions" description="The number of external non-coherent full write transactions." units="transactions" />
-        <event offset="172" advanced="yes" counter="L2_EXT_WRITE_NOSNP_PTL" title="External Bus Accesses" name="WriteNoSnoopPartial transactions" description="The number of external non-coherent partial write transactions." units="transactions" />
-        <event offset="173" advanced="yes" counter="L2_EXT_WRITE_SNP_FULL" title="External Bus Accesses" name="WriteSnoopFull transactions" description="The number of external coherent full write transactions." units="transactions" />
-        <event offset="174" advanced="yes" counter="L2_EXT_WRITE_SNP_PTL" title="External Bus Accesses" name="WriteSnoopPartial transactions" description="The number of external coherent partial write transactions." units="transactions" />
-        <event offset="175" counter="L2_EXT_WRITE_BEATS" title="External Bus Beats" name="Write beat" description="The number of external bus data write cycles." units="beats" />
-        <event offset="176" counter="L2_EXT_W_STALL" title="External Bus Stalls" name="Write stall cycles" description="The number of cycles where a write is stalled waiting for the external bus." units="cycles" />
-        <event offset="177" counter="L2_EXT_AW_CNT_Q1" title="External Bus Outstanding Writes" name="0-25% outstanding" description="The number of write transactions initiated when 0-25% of the maximum are in use." units="transactions" />
-        <event offset="178" counter="L2_EXT_AW_CNT_Q2" title="External Bus Outstanding Writes" name="25-50% outstanding" description="The number of write transactions initiated when 25-50% of the maximum are in use." units="transactions" />
-        <event offset="179" counter="L2_EXT_AW_CNT_Q3" title="External Bus Outstanding Writes" name="50-75% outstanding" description="The number of write transactions initiated when 50-75% of the maximum are in use." units="transactions" />
-        <event offset="180" advanced="yes" counter="L2_EXT_SNOOP" title="External Bus Accesses" name="Snoop transactions" description="The number of coherency snoops triggered by external masters." units="transactions" />
-        <event offset="181" advanced="yes" counter="L2_EXT_SNOOP_STALL" title="External Bus Stalls" name="Snoop stall cycles" description="The number of cycles where a coherency snoop triggered by external master is stalled." units="cycles" />
+        <event offset="4" advanced="yes" counter="MMU_REQUESTS" title="MMU Stage 1 Translations" name="MMU lookups" description="The number of main MMU address translations performed." units="requests"/>
+        <event offset="16" advanced="yes" counter="L2_RD_MSG_IN" title="L2 Cache Requests" name="Read requests" description="The number of L2 cache read requests from internal masters." units="requests"/>
+        <event offset="17" advanced="yes" counter="L2_RD_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Read stall cycles" description="The number of cycles L2 cache read requests from internal masters are stalled." units="cycles"/>
+        <event offset="18" advanced="yes" counter="L2_WR_MSG_IN" title="L2 Cache Requests" name="Write requests" description="The number of L2 cache write requests from internal masters." units="requests"/>
+        <event offset="19" advanced="yes" counter="L2_WR_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Write stall cycles" description="The number of cycles where L2 cache write requests from internal masters are stalled." units="cycles"/>
+        <event offset="20" advanced="yes" counter="L2_SNP_MSG_IN" title="L2 Cache Requests" name="Snoop requests" description="The number of L2 snoop requests from internal masters." units="requests"/>
+        <event offset="21" advanced="yes" counter="L2_SNP_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Snoop stall cycles" description="The number of cycles where L2 cache snoop requests from internal masters are stalled." units="cycles"/>
+        <event offset="22" advanced="yes" counter="L2_RD_MSG_OUT" title="L2 Cache Requests" name="L1 read requests" description="The number of L1 cache read requests sent by the L2 cache to an internal master." units="requests"/>
+        <event offset="23" advanced="yes" counter="L2_RD_MSG_OUT_STALL" title="L2 Cache Stall Cycles" name="L1 read stall cycles" description="The number of cycles where L1 cache read requests sent by the L2 cache to an internal master are stalled." units="cycles"/>
+        <event offset="24" advanced="yes" counter="L2_WR_MSG_OUT" title="L2 Cache Requests" name="L1 write requests" description="The number of L1 cache write responses sent by the L2 cache to an internal master." units="requests"/>
+        <event offset="25" counter="L2_ANY_LOOKUP" title="L2 Cache Lookups" name="Any lookup" description="The number of L2 cache lookups performed." units="requests"/>
+        <event offset="26" counter="L2_READ_LOOKUP" title="L2 Cache Lookups" name="Read lookup" description="The number of L2 cache read lookups performed." units="requests"/>
+        <event offset="27" counter="L2_WRITE_LOOKUP" title="L2 Cache Lookups" name="Write lookup" description="The number of L2 cache write lookups performed." units="requests"/>
+        <event offset="28" advanced="yes" counter="L2_EXT_SNOOP_LOOKUP" title="L2 Cache Lookups" name="External snoop lookups" description="The number of coherency snoop lookups performed that were triggered by an external master." units="requests"/>
+        <event offset="29" counter="L2_EXT_READ" title="External Bus Accesses" name="Read transaction" description="The number of external read transactions." units="transactions"/>
+        <event offset="30" advanced="yes" counter="L2_EXT_READ_NOSNP" title="External Bus Accesses" name="ReadNoSnoop transactions" description="The number of external non-coherent read transactions." units="transactions"/>
+        <event offset="31" advanced="yes" counter="L2_EXT_READ_UNIQUE" title="External Bus Accesses" name="ReadUnique transactions" description="The number of external coherent read unique transactions." units="transactions"/>
+        <event offset="32" counter="L2_EXT_READ_BEATS" title="External Bus Beats" name="Read beat" description="The number of external bus data read cycles." units="beats"/>
+        <event offset="33" counter="L2_EXT_AR_STALL" title="External Bus Stalls" name="Read stall cycles" description="The number of cycles where a read is stalled waiting for the external bus." units="cycles"/>
+        <event offset="34" counter="L2_EXT_AR_CNT_Q1" title="External Bus Outstanding Reads" name="0-25% outstanding" description="The number of read transactions initiated when 0-25% of the maximum are in use." units="transactions"/>
+        <event offset="35" counter="L2_EXT_AR_CNT_Q2" title="External Bus Outstanding Reads" name="25-50% outstanding" description="The number of read transactions initiated when 25-50% of the maximum are in use." units="transactions"/>
+        <event offset="36" counter="L2_EXT_AR_CNT_Q3" title="External Bus Outstanding Reads" name="50-75% outstanding" description="The number of read transactions initiated when 50-75% of the maximum are in use." units="transactions"/>
+        <event offset="37" counter="L2_EXT_RRESP_0_127" title="External Bus Read Latency" name="0-127 cycles" description="The number of data beats returned 0-127 cycles after the read request." units="beats"/>
+        <event offset="38" counter="L2_EXT_RRESP_128_191" title="External Bus Read Latency" name="128-191 cycles" description="The number of data beats returned 128-191 cycles after the read request." units="beats"/>
+        <event offset="39" counter="L2_EXT_RRESP_192_255" title="External Bus Read Latency" name="192-255 cycles" description="The number of data beats returned 192-255 cycles after the read request." units="beats"/>
+        <event offset="40" counter="L2_EXT_RRESP_256_319" title="External Bus Read Latency" name="256-319 cycles" description="The number of data beats returned 256-319 cycles after the read request." units="beats"/>
+        <event offset="41" counter="L2_EXT_RRESP_320_383" title="External Bus Read Latency" name="320-383 cycles" description="The number of data beats returned 320-383 cycles after the read request." units="beats"/>
+        <event offset="42" counter="L2_EXT_WRITE" title="External Bus Accesses" name="Write transaction" description="The number of external write transactions." units="transactions"/>
+        <event offset="43" advanced="yes" counter="L2_EXT_WRITE_NOSNP_FULL" title="External Bus Accesses" name="WriteNoSnoopFull transactions" description="The number of external non-coherent full write transactions." units="transactions"/>
+        <event offset="44" advanced="yes" counter="L2_EXT_WRITE_NOSNP_PTL" title="External Bus Accesses" name="WriteNoSnoopPartial transactions" description="The number of external non-coherent partial write transactions." units="transactions"/>
+        <event offset="45" advanced="yes" counter="L2_EXT_WRITE_SNP_FULL" title="External Bus Accesses" name="WriteSnoopFull transactions" description="The number of external coherent full write transactions." units="transactions"/>
+        <event offset="46" advanced="yes" counter="L2_EXT_WRITE_SNP_PTL" title="External Bus Accesses" name="WriteSnoopPartial transactions" description="The number of external coherent partial write transactions." units="transactions"/>
+        <event offset="47" counter="L2_EXT_WRITE_BEATS" title="External Bus Beats" name="Write beat" description="The number of external bus data write cycles." units="beats"/>
+        <event offset="48" counter="L2_EXT_W_STALL" title="External Bus Stalls" name="Write stall cycles" description="The number of cycles where a write is stalled waiting for the external bus." units="cycles"/>
+        <event offset="49" counter="L2_EXT_AW_CNT_Q1" title="External Bus Outstanding Writes" name="0-25% outstanding" description="The number of write transactions initiated when 0-25% of the maximum are in use." units="transactions"/>
+        <event offset="50" counter="L2_EXT_AW_CNT_Q2" title="External Bus Outstanding Writes" name="25-50% outstanding" description="The number of write transactions initiated when 25-50% of the maximum are in use." units="transactions"/>
+        <event offset="51" counter="L2_EXT_AW_CNT_Q3" title="External Bus Outstanding Writes" name="50-75% outstanding" description="The number of write transactions initiated when 50-75% of the maximum are in use." units="transactions"/>
+        <event offset="52" advanced="yes" counter="L2_EXT_SNOOP" title="External Bus Accesses" name="Snoop transactions" description="The number of coherency snoops triggered by external masters." units="transactions"/>
+        <event offset="53" advanced="yes" counter="L2_EXT_SNOOP_STALL" title="External Bus Stalls" name="Snoop stall cycles" description="The number of cycles where a coherency snoop triggered by external master is stalled." units="cycles"/>
+    </category>
+    <category name="Shader Core" per_cpu="no">
+        <event offset="4" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment active" description="The number of cycles where the shader core is processing a fragment workload." units="cycles"/>
+        <event offset="5" advanced="yes" counter="FRAG_PRIMITIVES" title="Core Primitives" name="Read primitives" description="The number of primitives read from the tile list by the fragment front-end." units="primitives"/>
+        <event offset="6" counter="FRAG_PRIM_RAST" title="Core Primitives" name="Rasterized primitives" description="The number of primitives being rasterized." units="primitives"/>
+        <event offset="7" counter="FRAG_FPK_ACTIVE" title="Core Cycles" name="Fragment FPKB active" description="The number of cycles where at least one quad is present in the pre-pipe quad queue." units="cycles"/>
+        <event offset="9" counter="FRAG_WARPS" title="Core Warps" name="Fragment warps" description="The number of fragment warps created." units="warps"/>
+        <event offset="10" counter="FRAG_PARTIAL_WARPS" title="Core Warps" name="Partial fragment warps" description="The number of fragment warps containing helper threads that do not correspond to a hit sample point." units="warps"/>
+        <event offset="11" counter="FRAG_QUADS_RAST" title="Core Quads" name="Rasterized quads" description="The number of quads generated by the rasterization phase." units="quads"/>
+        <event offset="12" counter="FRAG_QUADS_EZS_TEST" title="Core Quads" name="Early ZS tested quads" description="The number of quads that are undergoing early depth and stencil testing." units="quads"/>
+        <event offset="13" counter="FRAG_QUADS_EZS_UPDATE" title="Core Quads" name="Early ZS updated quads" description="The number of quads undergoing early depth and stencil testing, that are capable of updating the framebuffer." units="quads"/>
+        <event offset="14" counter="FRAG_QUADS_EZS_KILL" title="Core Quads" name="Early ZS killed quads" description="The number of quads killed by early depth and stencil testing." units="quads"/>
+        <event offset="15" counter="FRAG_LZS_TEST" title="Core Quads" name="Late ZS tested quads" description="The number of quads undergoing late depth and stencil testing." units="quads"/>
+        <event offset="16" counter="FRAG_LZS_KILL" title="Core Quads" name="Late ZS killed quads" description="The number of quads killed by late depth and stencil testing." units="quads"/>
+        <event offset="18" counter="FRAG_PTILES" title="Core Tiles" name="Tiles" description="The number of tiles processed by the shader core." units="tiles"/>
+        <event offset="19" counter="FRAG_TRANS_ELIM" title="Core Tiles" name="Constant tiles killed" description="The number of tiles killed by transaction elimination." units="tiles"/>
+        <event offset="20" counter="QUAD_FPK_KILLER" title="Core Quads" name="FPK occluder quads" description="The number of quads that are valid occluders for hidden surface removal." units="quads"/>
+        <event offset="22" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Non-fragment active" description="The number of cycles where the shader core is processing some non-fragment workload." units="cycles"/>
+        <event offset="23" advanced="yes" counter="COMPUTE_TASKS" title="Core Tasks" name="Non-fragment tasks" description="The number of non-fragment tasks issued to the shader core." units="tasks"/>
+        <event offset="24" counter="COMPUTE_WARPS" title="Core Warps" name="Non-fragment warps" description="The number of non-fragment warps created." units="warps"/>
+        <event offset="25" advanced="yes" counter="COMPUTE_STARVING" title="Core Starvation Cycles" name="Non-fragment starvation cycles" description="The number of cycles where the shader core is processing a non-fragment workload and there are no new threads available for execution." units="cycles"/>
+        <event offset="26" counter="EXEC_CORE_ACTIVE" title="Core Cycles" name="Execution core active" description="The number of cycles where the shader core is processing at least one warp." units="cycles"/>
+        <event offset="27" advanced="yes" counter="EXEC_ACTIVE" title="Core Cycles" name="Execution engine active" description="The number of cycles where the execution engine unit is processing at least one thread." units="cycles"/>
+        <event offset="28" counter="EXEC_INSTR_COUNT" title="Core EE Instructions" name="Executed instructions" description="The number of instructions executed per warp." units="instructions"/>
+        <event offset="29" counter="EXEC_INSTR_DIVERGED" title="Core EE Instructions" name="Diverged instructions" description="The number of instructions executed per warp, that have control flow divergence." units="instructions"/>
+        <event offset="30" advanced="yes" counter="EXEC_INSTR_STARVING" title="Core Starvation Cycles" name="Execution engine starvation cycles" description="The number of cycles where no new threads are available for execution." units="cycles"/>
+        <event offset="31" advanced="yes" counter="ARITH_INSTR_SINGLE_FMA" title="Core EE Instructions" name="Arithmetic instructions" description="The number of instructions where the workload is a single FMA pipe arithmetic operation." units="instructions"/>
+        <event offset="32" advanced="yes" counter="ARITH_INSTR_DOUBLE" title="Core EE Instructions" name="Dual Arithmetic instructions" description="The number of instructions where the workload is one FMA pipe arithmetic operation and one ADD pipe arithmetic operation." units="instructions"/>
+        <event offset="33" advanced="yes" counter="ARITH_INSTR_MSG" title="Core EE Instructions" name="Arithmetic + Message instructions" description="The number of instructions where the workload is one FMA pipe arithmetic operation and one ADD pipe message operation" units="instructions"/>
+        <event offset="34" advanced="yes" counter="ARITH_INSTR_MSG_ONLY" title="Core EE Instructions" name="Message instructions" description="The number of instructions where the workload is a single ADD pipe message operation, with no FMA pipe operation" units="instructions"/>
+        <event offset="35" counter="TEX_INSTR" title="Core Texture Requests" name="Texture requests" description="The number of thread-width texture operations processed." units="instructions"/>
+        <event offset="36" counter="TEX_INSTR_MIPMAP" title="Core Texture Requests" name="Mipmapped texture request" description="The number of texture operations that act on a mipmapped texture." units="instructions"/>
+        <event offset="37" counter="TEX_INSTR_COMPRESSED" title="Core Texture Requests" name="Compressed texture requests" description="The number of texture operations acting on a compressed texture." units="instructions"/>
+        <event offset="38" counter="TEX_INSTR_3D" title="Core Texture Requests" name="3D texture requests" description="The number of texture operations acting on a 3D texture." units="instructions"/>
+        <event offset="39" counter="TEX_INSTR_TRILINEAR" title="Core Texture Requests" name="Trilinear filtered requests" description="The number of texture operations using a trilinear texture filter." units="instructions"/>
+        <event offset="40" counter="TEX_COORD_ISSUE" title="Core Texture Cycles" name="Texturing active" description="The number of texture filtering issue cycles." units="cycles"/>
+        <event offset="41" advanced="yes" counter="TEX_COORD_STALL" title="Core Texture Cycles" name="Coordinate stall cycles" description="The number of clock cycles where threads are stalled at the texel coordinate calculation stage." units="cycles"/>
+        <event offset="42" advanced="yes" counter="TEX_STARVE_CACHE" title="Core Texture Cycles" name="Line fill stall cycles" description="The number of clock cycles where at least one thread is waiting for data from the texture cache, but no lookup is completed." units="cycles"/>
+        <event offset="43" advanced="yes" counter="TEX_STARVE_FILTER" title="Core Texture Cycles" name="Partial data stall cycles" description="The number of clock cycles where at least one thread fetched some data from the texture cache, but no filtering operation is started." units="cycles"/>
+        <event offset="44" counter="LS_MEM_READ_FULL" title="Core Load/Store Cycles" name="Full read cycles" description="The number of full-width load/store cache reads." units="cycles"/>
+        <event offset="45" counter="LS_MEM_READ_SHORT" title="Core Load/Store Cycles" name="Partial read cycles" description="The number of partial-width load/store cache reads." units="cycles"/>
+        <event offset="46" counter="LS_MEM_WRITE_FULL" title="Core Load/Store Cycles" name="Full write cycles" description="The number of full-width load/store cache writes." units="cycles"/>
+        <event offset="47" counter="LS_MEM_WRITE_SHORT" title="Core Load/Store Cycles" name="Partial write cycles" description="The number of partial-width load/store cache writes." units="cycles"/>
+        <event offset="48" counter="LS_MEM_ATOMIC" title="Core Load/Store Cycles" name="Atomic access cycles" description="The number of load/store atomic accesses." units="cycles"/>
+        <event offset="49" counter="VARY_INSTR" title="Core Varying Requests" name="Interpolation requests" description="The number of warp-width interpolation operations processed by the varying unit." units="instructions"/>
+        <event offset="50" counter="VARY_SLOT_32" title="Core Varying Cycles" name="32-bit interpolation active" description="The number of 32-bit interpolation cycles processed by the varying unit." units="cycles"/>
+        <event offset="51" counter="VARY_SLOT_16" title="Core Varying Cycles" name="16-bit interpolation active" description="The number of 16-bit interpolation cycles processed by the varying unit." units="cycles"/>
+        <event offset="52" advanced="yes" counter="ATTR_INSTR" title="Core Attribute Requests" name="Attribute requests" description="The number of instructions executed by the attribute unit." units="instructions"/>
+        <event offset="53" advanced="yes" counter="ARITH_INSTR_FP_MUL" title="Core EE Instructions" name="Multiplier instructions" description="The number of instructions where the workload uses floating-point multiplier hardware." units="instructions"/>
+        <event offset="54" counter="BEATS_RD_FTC" title="Core L2 Reads" name="Fragment L2 read beats" description="The number of read beats received by the fixed-function fragment front-end." units="beats"/>
+        <event offset="55" counter="BEATS_RD_FTC_EXT" title="Core L2 Reads" name="Fragment external read beats" description="The number of read beats received by the fixed-function fragment front-end that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="56" counter="BEATS_RD_LSC" title="Core L2 Reads" name="Load/store L2 read beats" description="The number of read beats received by the load/store unit." units="beats"/>
+        <event offset="57" counter="BEATS_RD_LSC_EXT" title="Core L2 Reads" name="Load/store external read beats" description="The number of read beats received by the load/store unit that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="58" counter="BEATS_RD_TEX" title="Core L2 Reads" name="Texture L2 read beats" description="The number of read beats received by the texture unit." units="beats"/>
+        <event offset="59" counter="BEATS_RD_TEX_EXT" title="Core L2 Reads" name="Texture external read beats" description="The number of read beats received by the texture unit that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="60" advanced="yes" counter="BEATS_RD_OTHER" title="Core L2 Reads" name="Other L2 read beats" description="The number of read beats received by a unit that is not specifically identified." units="beats"/>
+        <event offset="61" counter="BEATS_WR_LSC" title="Core Writes" name="Load/store write beats" description="The number of write beats sent by the load/store unit." units="beats"/>
+        <event offset="62" counter="BEATS_WR_TIB" title="Core Writes" name="Tile buffer write beats" description="The number of write beats sent by the tile buffer writeback unit." units="beats"/>
+        <event offset="63" advanced="yes" counter="BEATS_WR_OTHER" title="Core Writes" name="Other write beats" description="The number of write beats sent by any unit that is not specifically identified." units="beats"/>
     </category>
 </metrics>
\ No newline at end of file
diff --git a/src/panfrost/perf/G72.xml b/src/panfrost/perf/G72.xml
index 80b420131f6..3f642ff9c2b 100644
--- a/src/panfrost/perf/G72.xml
+++ b/src/panfrost/perf/G72.xml
@@ -1,6 +1,6 @@
 <!--
 Copyright © 2017-2020 ARM Limited.
-Copyright © 2021 Collabora, Ltd.
+Copyright © 2021-2022 Collabora, Ltd.
 Author: Antonio Caggiano <antonio.caggiano@collabora.com>
 
 Permission is hereby granted, free of charge, to any person obtaining a
@@ -24,154 +24,154 @@ SOFTWARE.
 -->
 <metrics id="THEx">
     <category name="Job Manager" per_cpu="no">
-        <event offset="6" counter="GPU_ACTIVE" title="GPU Cycles" name="GPU active" description="The number of cycles where the GPU has a workload of any type queued for processing." units="cycles" />
-        <event offset="7" advanced="yes" counter="IRQ_ACTIVE" title="GPU Cycles" name="Interrupt active" description="The number of cycles where the GPU has a pending interrupt." units="cycles" />
-        <event offset="8" advanced="yes" counter="JS0_JOBS" title="GPU Jobs" name="Fragment jobs" description="The number of jobs processed by the GPU fragment queue." units="jobs" />
-        <event offset="9" counter="JS0_TASKS" title="GPU Tasks" name="Fragment tasks" description="The number of 32x32 pixel tasks processed by the GPU fragment queue." units="tasks" />
-        <event offset="10" counter="JS0_ACTIVE" title="GPU Cycles" name="Fragment queue active" description="The number of cycles where work is queued for processing in the GPU fragment queue." units="cycles" />
-        <event offset="12" advanced="yes" counter="JS0_WAIT_READ" title="GPU Wait Cycles" name="Fragment descriptor reads cycles" description="The number of cycles where queued fragment work is waiting for a descriptor load." units="cycles" />
-        <event offset="13" advanced="yes" counter="JS0_WAIT_ISSUE" title="GPU Wait Cycles" name="Fragment job issue cycles" description="The number of cycles where queued fragment work is waiting for an available processor." units="cycles" />
-        <event offset="14" advanced="yes" counter="JS0_WAIT_DEPEND" title="GPU Wait Cycles" name="Fragment job dependency cycles" description="The number of cycles where queued fragment work is waiting for dependent work to complete." units="cycles" />
-        <event offset="15" advanced="yes" counter="JS0_WAIT_FINISH" title="GPU Wait Cycles" name="Fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued fragment work to complete." units="cycles" />
-        <event offset="16" advanced="yes" counter="JS1_JOBS" title="GPU Jobs" name="Non-fragment jobs" description="The number of jobs processed by the GPU non-fragment queue." units="jobs" />
-        <event offset="17" advanced="yes" counter="JS1_TASKS" title="GPU Tasks" name="Non-fragment tasks" description="The number of tasks processed by the GPU non-fragment queue." units="tasks" />
-        <event offset="18" counter="JS1_ACTIVE" title="GPU Cycles" name="Non-fragment queue active" description="The number of cycles where work is queued in the GPU non-fragment queue." units="cycles" />
-        <event offset="20" advanced="yes" counter="JS1_WAIT_READ" title="GPU Wait Cycles" name="Non-fragment descriptor read cycles" description="The number number of cycles where queued non-fragment work is waiting for a descriptor load." units="cycles" />
-        <event offset="21" advanced="yes" counter="JS1_WAIT_ISSUE" title="GPU Wait Cycles" name="Non-fragment job issue cycles" description="The number of cycles where queued non-fragment work is waiting for an available processor." units="cycles" />
-        <event offset="22" advanced="yes" counter="JS1_WAIT_DEPEND" title="GPU Wait Cycles" name="Non-fragment job dependency cycles" description="The number of cycles where queued non-fragment work is waiting for dependent work to complete." units="cycles" />
-        <event offset="23" advanced="yes" counter="JS1_WAIT_FINISH" title="GPU Wait Cycles" name="Non-fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued non-fragment work to complete." units="cycles" />
-        <event offset="24" advanced="yes" counter="JS2_JOBS" title="GPU Jobs" name="Reserved jobs" description="The number of jobs processed by the GPU reserved queue." units="jobs" />
-        <event offset="25" advanced="yes" counter="JS2_TASKS" title="GPU Tasks" name="Reserved tasks" description="The number of tasks processed by the GPU reserved queue." units="tasks" />
-        <event offset="26" advanced="yes" counter="JS2_ACTIVE" title="GPU Cycles" name="Reserved queue active" description="The number of cycles where work is queued in the GPU reserved queue." units="cycles" />
-        <event offset="28" advanced="yes" counter="JS2_WAIT_READ" title="GPU Wait Cycles" name="Reserved descriptor read cycles" description="The number of cycles where queued reserved work is waiting for a descriptor load." units="cycles" />
-        <event offset="29" advanced="yes" counter="JS2_WAIT_ISSUE" title="GPU Wait Cycles" name="Reserved job issue cycles" description="The number of cycles where queued reserved work is waiting for an available processor." units="cycles" />
-        <event offset="30" advanced="yes" counter="JS2_WAIT_DEPEND" title="GPU Wait Cycles" name="Reserved job dependency cycles" description="The number of cycles where queued reserved work is waiting for dependent work to complete." units="cycles" />
-        <event offset="31" advanced="yes" counter="JS2_WAIT_FINISH" title="GPU Wait Cycles" name="Reserved job finish cycles" description="The number of cycles where the GPU is waiting for issued reserved work to complete." units="cycles" />
+        <event offset="6" counter="GPU_ACTIVE" title="GPU Cycles" name="GPU active" description="The number of cycles where the GPU has a workload of any type queued for processing." units="cycles"/>
+        <event offset="7" advanced="yes" counter="IRQ_ACTIVE" title="GPU Cycles" name="Interrupt active" description="The number of cycles where the GPU has a pending interrupt." units="cycles"/>
+        <event offset="8" advanced="yes" counter="JS0_JOBS" title="GPU Jobs" name="Fragment jobs" description="The number of jobs processed by the GPU fragment queue." units="jobs"/>
+        <event offset="9" counter="JS0_TASKS" title="GPU Tasks" name="Fragment tasks" description="The number of 32x32 pixel tasks processed by the GPU fragment queue." units="tasks"/>
+        <event offset="10" counter="JS0_ACTIVE" title="GPU Cycles" name="Fragment queue active" description="The number of cycles where work is queued for processing in the GPU fragment queue." units="cycles"/>
+        <event offset="12" advanced="yes" counter="JS0_WAIT_READ" title="GPU Wait Cycles" name="Fragment descriptor reads cycles" description="The number of cycles where queued fragment work is waiting for a descriptor load." units="cycles"/>
+        <event offset="13" advanced="yes" counter="JS0_WAIT_ISSUE" title="GPU Wait Cycles" name="Fragment job issue cycles" description="The number of cycles where queued fragment work is waiting for an available processor." units="cycles"/>
+        <event offset="14" advanced="yes" counter="JS0_WAIT_DEPEND" title="GPU Wait Cycles" name="Fragment job dependency cycles" description="The number of cycles where queued fragment work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="15" advanced="yes" counter="JS0_WAIT_FINISH" title="GPU Wait Cycles" name="Fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued fragment work to complete." units="cycles"/>
+        <event offset="16" advanced="yes" counter="JS1_JOBS" title="GPU Jobs" name="Non-fragment jobs" description="The number of jobs processed by the GPU non-fragment queue." units="jobs"/>
+        <event offset="17" advanced="yes" counter="JS1_TASKS" title="GPU Tasks" name="Non-fragment tasks" description="The number of tasks processed by the GPU non-fragment queue." units="tasks"/>
+        <event offset="18" counter="JS1_ACTIVE" title="GPU Cycles" name="Non-fragment queue active" description="The number of cycles where work is queued in the GPU non-fragment queue." units="cycles"/>
+        <event offset="20" advanced="yes" counter="JS1_WAIT_READ" title="GPU Wait Cycles" name="Non-fragment descriptor read cycles" description="The number number of cycles where queued non-fragment work is waiting for a descriptor load." units="cycles"/>
+        <event offset="21" advanced="yes" counter="JS1_WAIT_ISSUE" title="GPU Wait Cycles" name="Non-fragment job issue cycles" description="The number of cycles where queued non-fragment work is waiting for an available processor." units="cycles"/>
+        <event offset="22" advanced="yes" counter="JS1_WAIT_DEPEND" title="GPU Wait Cycles" name="Non-fragment job dependency cycles" description="The number of cycles where queued non-fragment work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="23" advanced="yes" counter="JS1_WAIT_FINISH" title="GPU Wait Cycles" name="Non-fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued non-fragment work to complete." units="cycles"/>
+        <event offset="24" advanced="yes" counter="JS2_JOBS" title="GPU Jobs" name="Reserved jobs" description="The number of jobs processed by the GPU reserved queue." units="jobs"/>
+        <event offset="25" advanced="yes" counter="JS2_TASKS" title="GPU Tasks" name="Reserved tasks" description="The number of tasks processed by the GPU reserved queue." units="tasks"/>
+        <event offset="26" advanced="yes" counter="JS2_ACTIVE" title="GPU Cycles" name="Reserved queue active" description="The number of cycles where work is queued in the GPU reserved queue." units="cycles"/>
+        <event offset="28" advanced="yes" counter="JS2_WAIT_READ" title="GPU Wait Cycles" name="Reserved descriptor read cycles" description="The number of cycles where queued reserved work is waiting for a descriptor load." units="cycles"/>
+        <event offset="29" advanced="yes" counter="JS2_WAIT_ISSUE" title="GPU Wait Cycles" name="Reserved job issue cycles" description="The number of cycles where queued reserved work is waiting for an available processor." units="cycles"/>
+        <event offset="30" advanced="yes" counter="JS2_WAIT_DEPEND" title="GPU Wait Cycles" name="Reserved job dependency cycles" description="The number of cycles where queued reserved work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="31" advanced="yes" counter="JS2_WAIT_FINISH" title="GPU Wait Cycles" name="Reserved job finish cycles" description="The number of cycles where the GPU is waiting for issued reserved work to complete." units="cycles"/>
     </category>
     <category name="Tiler" per_cpu="no">
-        <event offset="68" counter="TILER_ACTIVE" title="GPU Cycles" name="Tiler active" description="The number of cycles where the tiler has a workload queued for processing." units="cycles" />
-        <event offset="70" counter="TRIANGLES" title="Input Primitives" name="Triangle primitives" description="The number of input triangle primitives." units="primitives" />
-        <event offset="71" counter="LINES" title="Input Primitives" name="Line primitives" description="The number of input line primitives." units="primitives" />
-        <event offset="72" counter="POINTS" title="Input Primitives" name="Point primitives" description="The number of input point primitives." units="primitives" />
-        <event offset="73" counter="FRONT_FACING" title="Visible Primitives" name="Front-facing primitives" description="The number of front-facing triangles that are visible after culling." units="primitives" />
-        <event offset="74" counter="BACK_FACING" title="Visible Primitives" name="Back-facing primitives" description="The number of back-facing triangles that are visible after culling." units="primitives" />
-        <event offset="75" counter="PRIM_VISIBLE" title="Primitive Culling" name="Visible primitives" description="The number of primitives that are visible after culling." units="primitives" />
-        <event offset="76" counter="PRIM_CULLED" title="Primitive Culling" name="Facing and XY plane test culled primitives" description="The number of primitives that are culled by facing or frustum XY plane tests." units="primitives" />
-        <event offset="77" counter="PRIM_CLIPPED" title="Primitive Culling" name="Z plane test culled primitives" description="The number of primitives that are culled by frustum Z plane tests." units="primitives" />
-        <event offset="78" counter="PRIM_SAT_CULLED" title="Primitive Culling" name="Sample test culled primitives" description="The number of primitives culled by the sample coverage test." units="primitives" />
-        <event offset="81" advanced="yes" counter="BUS_READ" title="Tiler L2 Accesses" name="Read beats" description="The number of internal bus data read cycles made by the tiler." units="beats" />
-        <event offset="83" advanced="yes" counter="BUS_WRITE" title="Tiler L2 Accesses" name="Write beats" description="The number of internal bus data write cycles made by the tiler." units="beats" />
-        <event offset="85" counter="IDVS_POS_SHAD_REQ" title="Tiler Shading Requests" name="Position shading requests" description="The number of position shading requests in the IDVS flow." units="requests" />
-        <event offset="87" advanced="yes" counter="IDVS_POS_SHAD_STALL" title="Tiler Cycles" name="Position shading stall cycles" description="The number of cycles where the tiler has a stalled position shading request." units="cycles" />
-        <event offset="88" advanced="yes" counter="IDVS_POS_FIFO_FULL" title="Tiler Cycles" name="Position FIFO full cycles" description="The number of cycles where the tiler has a stalled position shading buffer." units="cycles" />
-        <event offset="90" advanced="yes" counter="VCACHE_HIT" title="Tiler Vertex Cache" name="Position cache hits" description="The number of position lookups that result in a hit in the vertex cache." units="requests" />
-        <event offset="91" advanced="yes" counter="VCACHE_MISS" title="Tiler Vertex Cache" name="Position cache misses" description="The number of position lookups that miss in the vertex cache." units="requests" />
-        <event offset="95" advanced="yes" counter="VFETCH_STALL" title="Tiler Cycles" name="Primitive assembly busy stall cycles" description="The number of cycles where the tiler is stalled waiting for primitive assembly." units="cycles" />
-        <event offset="98" advanced="yes" counter="IDVS_VBU_HIT" title="Tiler Vertex Cache" name="Varying cache hits" description="The number of varying lookups that result in a hit in the vertex cache." units="requests" />
-        <event offset="99" advanced="yes" counter="IDVS_VBU_MISS" title="Tiler Vertex Cache" name="Varying cache misses" description="The number of varying lookups that miss in the vertex cache." units="requests" />
-        <event offset="101" counter="IDVS_VAR_SHAD_REQ" title="Tiler Shading Requests" name="Varying shading requests" description="The number of varying shading requests in the IDVS flow." units="requests" />
-        <event offset="102" advanced="yes" counter="IDVS_VAR_SHAD_STALL" title="Tiler Cycles" name="Varying shading stall cycles" description="The number of cycles where the tiler has a stalled varying shading request." units="cycles" />
-        <event offset="118" advanced="yes" counter="WRBUF_NO_AXI_ID_STALL" title="Tiler Cycles" name="Write buffer transaction stall cycles" description="The number of cycles where the tiler write buffer can not send data because it has no available write IDs." units="cycles" />
-        <event offset="119" advanced="yes" counter="WRBUF_AXI_STALL" title="Tiler Cycles" name="Write buffer write stall cycles" description="The number of cycles where the tiler write buffer can not send data because the bus is not ready." units="cycles" />
-    </category>
-    <category name="Shader Core" per_cpu="no">
-        <event offset="196" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment active" description="The number of cycles where the shader core is processing a fragment workload." units="cycles" />
-        <event offset="197" advanced="yes" counter="FRAG_PRIMITIVES" title="Core Primitives" name="Read primitives" description="The number of primitives read from the tile list by the fragment front-end." units="primitives" />
-        <event offset="198" counter="FRAG_PRIM_RAST" title="Core Primitives" name="Rasterized primitives" description="The number of primitives being rasterized." units="primitives" />
-        <event offset="199" counter="FRAG_FPK_ACTIVE" title="Core Cycles" name="Fragment FPKB active" description="The number of cycles where at least one quad is present in the pre-pipe quad queue." units="cycles" />
-        <event offset="201" counter="FRAG_WARPS" title="Core Warps" name="Fragment warps" description="The number of fragment warps created." units="warps" />
-        <event offset="202" counter="FRAG_PARTIAL_WARPS" title="Core Warps" name="Partial fragment warps" description="The number of fragment warps containing helper threads that do not correspond to a hit sample point." units="warps" />
-        <event offset="203" counter="FRAG_QUADS_RAST" title="Core Quads" name="Rasterized quads" description="The number of quads generated by the rasterization phase." units="quads" />
-        <event offset="204" counter="FRAG_QUADS_EZS_TEST" title="Core Quads" name="Early ZS tested quads" description="The number of quads that are undergoing early depth and stencil testing." units="quads" />
-        <event offset="205" counter="FRAG_QUADS_EZS_UPDATE" title="Core Quads" name="Early ZS updated quads" description="The number of quads undergoing early depth and stencil testing, that are capable of updating the framebuffer." units="quads" />
-        <event offset="206" counter="FRAG_QUADS_EZS_KILL" title="Core Quads" name="Early ZS killed quads" description="The number of quads killed by early depth and stencil testing." units="quads" />
-        <event offset="207" counter="FRAG_LZS_TEST" title="Core Quads" name="Late ZS tested quads" description="The number of quads undergoing late depth and stencil testing." units="quads" />
-        <event offset="208" counter="FRAG_LZS_KILL" title="Core Quads" name="Late ZS killed quads" description="The number of quads killed by late depth and stencil testing." units="quads" />
-        <event offset="210" counter="FRAG_PTILES" title="Core Tiles" name="Tiles" description="The number of tiles processed by the shader core." units="tiles" />
-        <event offset="211" counter="FRAG_TRANS_ELIM" title="Core Tiles" name="Constant tiles killed" description="The number of tiles killed by transaction elimination." units="tiles" />
-        <event offset="212" counter="QUAD_FPK_KILLER" title="Core Quads" name="FPK occluder quads" description="The number of quads that are valid occluders for hidden surface removal." units="quads" />
-        <event offset="214" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Non-fragment active" description="The number of cycles where the shader core is processing some non-fragment workload." units="cycles" />
-        <event offset="215" advanced="yes" counter="COMPUTE_TASKS" title="Core Tasks" name="Non-fragment tasks" description="The number of non-fragment tasks issued to the shader core." units="tasks" />
-        <event offset="216" counter="COMPUTE_WARPS" title="Core Warps" name="Non-fragment warps" description="The number of non-fragment warps created." units="warps" />
-        <event offset="217" advanced="yes" counter="COMPUTE_STARVING" title="Core Starvation Cycles" name="Non-fragment starvation cycles" description="The number of cycles where the shader core is processing a non-fragment workload and there are no new threads available for execution." units="cycles" />
-        <event offset="218" counter="EXEC_CORE_ACTIVE" title="Core Cycles" name="Execution core active" description="The number of cycles where the shader core is processing at least one warp." units="cycles" />
-        <event offset="219" advanced="yes" counter="EXEC_ACTIVE" title="Core Cycles" name="Execution engine active" description="The number of cycles where the execution engine unit is processing at least one thread." units="cycles" />
-        <event offset="220" counter="EXEC_INSTR_COUNT" title="Core EE Instructions" name="Executed instructions" description="The number of instructions executed per warp." units="instructions" />
-        <event offset="221" counter="EXEC_INSTR_DIVERGED" title="Core EE Instructions" name="Diverged instructions" description="The number of instructions executed per warp, that have control flow divergence." units="instructions" />
-        <event offset="222" advanced="yes" counter="EXEC_INSTR_STARVING" title="Core Starvation Cycles" name="Execution engine starvation cycles" description="The number of cycles where no new threads are available for execution." units="cycles" />
-        <event offset="223" advanced="yes" counter="ARITH_INSTR_SINGLE_FMA" title="Core EE Instructions" name="Arithmetic instructions" description="The number of instructions where the workload is a single FMA pipe arithmetic operation." units="instructions" />
-        <event offset="224" advanced="yes" counter="ARITH_INSTR_DOUBLE" title="Core EE Instructions" name="Dual Arithmetic instructions" description="The number of instructions where the workload is one FMA pipe arithmetic operation and one ADD pipe arithmetic operation." units="instructions" />
-        <event offset="225" advanced="yes" counter="ARITH_INSTR_MSG" title="Core EE Instructions" name="Arithmetic + Message instructions" description="The number of instructions where the workload is one FMA pipe arithmetic operation and one ADD pipe message operation" units="instructions" />
-        <event offset="226" advanced="yes" counter="ARITH_INSTR_MSG_ONLY" title="Core EE Instructions" name="Message instructions" description="The number of instructions where the workload is a single ADD pipe message operation, with no FMA pipe operation" units="instructions" />
-        <event offset="227" counter="TEX_INSTR" title="Core Texture Requests" name="Texture requests" description="The number of thread-width texture operations processed." units="instructions" />
-        <event offset="228" counter="TEX_INSTR_MIPMAP" title="Core Texture Requests" name="Mipmapped texture request" description="The number of texture operations that act on a mipmapped texture." units="instructions" />
-        <event offset="229" counter="TEX_INSTR_COMPRESSED" title="Core Texture Requests" name="Compressed texture requests" description="The number of texture operations acting on a compressed texture." units="instructions" />
-        <event offset="230" counter="TEX_INSTR_3D" title="Core Texture Requests" name="3D texture requests" description="The number of texture operations acting on a 3D texture." units="instructions" />
-        <event offset="231" counter="TEX_INSTR_TRILINEAR" title="Core Texture Requests" name="Trilinear filtered requests" description="The number of texture operations using a trilinear texture filter." units="instructions" />
-        <event offset="232" counter="TEX_COORD_ISSUE" title="Core Texture Cycles" name="Texturing active" description="The number of texture filtering issue cycles." units="cycles" />
-        <event offset="233" advanced="yes" counter="TEX_COORD_STALL" title="Core Texture Cycles" name="Coordinate stall cycles" description="The number of clock cycles where threads are stalled at the texel coordinate calculation stage." units="cycles" />
-        <event offset="234" advanced="yes" counter="TEX_STARVE_CACHE" title="Core Texture Cycles" name="Line fill stall cycles" description="The number of clock cycles where at least one thread is waiting for data from the texture cache, but no lookup is completed." units="cycles" />
-        <event offset="235" advanced="yes" counter="TEX_STARVE_FILTER" title="Core Texture Cycles" name="Partial data stall cycles" description="The number of clock cycles where at least one thread fetched some data from the texture cache, but no filtering operation is started." units="cycles" />
-        <event offset="236" counter="LS_MEM_READ_FULL" title="Core Load/Store Cycles" name="Full read cycles" description="The number of full-width load/store cache reads." units="cycles" />
-        <event offset="237" counter="LS_MEM_READ_SHORT" title="Core Load/Store Cycles" name="Partial read cycles" description="The number of partial-width load/store cache reads." units="cycles" />
-        <event offset="238" counter="LS_MEM_WRITE_FULL" title="Core Load/Store Cycles" name="Full write cycles" description="The number of full-width load/store cache writes." units="cycles" />
-        <event offset="239" counter="LS_MEM_WRITE_SHORT" title="Core Load/Store Cycles" name="Partial write cycles" description="The number of partial-width load/store cache writes." units="cycles" />
-        <event offset="240" counter="LS_MEM_ATOMIC" title="Core Load/Store Cycles" name="Atomic access cycles" description="The number of load/store atomic accesses." units="cycles" />
-        <event offset="241" counter="VARY_INSTR" title="Core Varying Requests" name="Interpolation requests" description="The number of warp-width interpolation operations processed by the varying unit." units="instructions" />
-        <event offset="242" counter="VARY_SLOT_32" title="Core Varying Cycles" name="32-bit interpolation active" description="The number of 32-bit interpolation cycles processed by the varying unit." units="cycles" />
-        <event offset="243" counter="VARY_SLOT_16" title="Core Varying Cycles" name="16-bit interpolation active" description="The number of 16-bit interpolation cycles processed by the varying unit." units="cycles" />
-        <event offset="244" advanced="yes" counter="ATTR_INSTR" title="Core Attribute Requests" name="Attribute requests" description="The number of instructions executed by the attribute unit." units="instructions" />
-        <event offset="245" advanced="yes" counter="ARITH_INSTR_FP_MUL" title="Core EE Instructions" name="Multiplier instructions" description="The number of instructions where the workload uses floating-point multiplier hardware." units="instructions" />
-        <event offset="246" counter="BEATS_RD_FTC" title="Core L2 Reads" name="Fragment L2 read beats" description="The number of read beats received by the fixed-function fragment front-end." units="beats" />
-        <event offset="247" counter="BEATS_RD_FTC_EXT" title="Core L2 Reads" name="Fragment external read beats" description="The number of read beats received by the fixed-function fragment front-end that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="248" counter="BEATS_RD_LSC" title="Core L2 Reads" name="Load/store L2 read beats" description="The number of read beats received by the load/store unit." units="beats" />
-        <event offset="249" counter="BEATS_RD_LSC_EXT" title="Core L2 Reads" name="Load/store external read beats" description="The number of read beats received by the load/store unit that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="250" counter="BEATS_RD_TEX" title="Core L2 Reads" name="Texture L2 read beats" description="The number of read beats received by the texture unit." units="beats" />
-        <event offset="251" counter="BEATS_RD_TEX_EXT" title="Core L2 Reads" name="Texture external read beats" description="The number of read beats received by the texture unit that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="252" advanced="yes" counter="BEATS_RD_OTHER" title="Core L2 Reads" name="Other L2 read beats" description="The number of read beats received by a unit that is not specifically identified." units="beats" />
-        <event offset="253" counter="BEATS_WR_LSC" title="Core Writes" name="Load/store write beats" description="The number of write beats sent by the load/store unit." units="beats" />
-        <event offset="254" counter="BEATS_WR_TIB" title="Core Writes" name="Tile buffer write beats" description="The number of write beats sent by the tile buffer writeback unit." units="beats" />
-        <event offset="255" advanced="yes" counter="BEATS_WR_OTHER" title="Core Writes" name="Other write beats" description="The number of write beats sent by any unit that is not specifically identified." units="beats" />
+        <event offset="4" counter="TILER_ACTIVE" title="GPU Cycles" name="Tiler active" description="The number of cycles where the tiler has a workload queued for processing." units="cycles"/>
+        <event offset="6" counter="TRIANGLES" title="Input Primitives" name="Triangle primitives" description="The number of input triangle primitives." units="primitives"/>
+        <event offset="7" counter="LINES" title="Input Primitives" name="Line primitives" description="The number of input line primitives." units="primitives"/>
+        <event offset="8" counter="POINTS" title="Input Primitives" name="Point primitives" description="The number of input point primitives." units="primitives"/>
+        <event offset="9" counter="FRONT_FACING" title="Visible Primitives" name="Front-facing primitives" description="The number of front-facing triangles that are visible after culling." units="primitives"/>
+        <event offset="10" counter="BACK_FACING" title="Visible Primitives" name="Back-facing primitives" description="The number of back-facing triangles that are visible after culling." units="primitives"/>
+        <event offset="11" counter="PRIM_VISIBLE" title="Primitive Culling" name="Visible primitives" description="The number of primitives that are visible after culling." units="primitives"/>
+        <event offset="12" counter="PRIM_CULLED" title="Primitive Culling" name="Facing and XY plane test culled primitives" description="The number of primitives that are culled by facing or frustum XY plane tests." units="primitives"/>
+        <event offset="13" counter="PRIM_CLIPPED" title="Primitive Culling" name="Z plane test culled primitives" description="The number of primitives that are culled by frustum Z plane tests." units="primitives"/>
+        <event offset="14" counter="PRIM_SAT_CULLED" title="Primitive Culling" name="Sample test culled primitives" description="The number of primitives culled by the sample coverage test." units="primitives"/>
+        <event offset="17" advanced="yes" counter="BUS_READ" title="Tiler L2 Accesses" name="Read beats" description="The number of internal bus data read cycles made by the tiler." units="beats"/>
+        <event offset="19" advanced="yes" counter="BUS_WRITE" title="Tiler L2 Accesses" name="Write beats" description="The number of internal bus data write cycles made by the tiler." units="beats"/>
+        <event offset="21" counter="IDVS_POS_SHAD_REQ" title="Tiler Shading Requests" name="Position shading requests" description="The number of position shading requests in the IDVS flow." units="requests"/>
+        <event offset="23" advanced="yes" counter="IDVS_POS_SHAD_STALL" title="Tiler Cycles" name="Position shading stall cycles" description="The number of cycles where the tiler has a stalled position shading request." units="cycles"/>
+        <event offset="24" advanced="yes" counter="IDVS_POS_FIFO_FULL" title="Tiler Cycles" name="Position FIFO full cycles" description="The number of cycles where the tiler has a stalled position shading buffer." units="cycles"/>
+        <event offset="26" advanced="yes" counter="VCACHE_HIT" title="Tiler Vertex Cache" name="Position cache hits" description="The number of position lookups that result in a hit in the vertex cache." units="requests"/>
+        <event offset="27" advanced="yes" counter="VCACHE_MISS" title="Tiler Vertex Cache" name="Position cache misses" description="The number of position lookups that miss in the vertex cache." units="requests"/>
+        <event offset="31" advanced="yes" counter="VFETCH_STALL" title="Tiler Cycles" name="Primitive assembly busy stall cycles" description="The number of cycles where the tiler is stalled waiting for primitive assembly." units="cycles"/>
+        <event offset="34" advanced="yes" counter="IDVS_VBU_HIT" title="Tiler Vertex Cache" name="Varying cache hits" description="The number of varying lookups that result in a hit in the vertex cache." units="requests"/>
+        <event offset="35" advanced="yes" counter="IDVS_VBU_MISS" title="Tiler Vertex Cache" name="Varying cache misses" description="The number of varying lookups that miss in the vertex cache." units="requests"/>
+        <event offset="37" counter="IDVS_VAR_SHAD_REQ" title="Tiler Shading Requests" name="Varying shading requests" description="The number of varying shading requests in the IDVS flow." units="requests"/>
+        <event offset="38" advanced="yes" counter="IDVS_VAR_SHAD_STALL" title="Tiler Cycles" name="Varying shading stall cycles" description="The number of cycles where the tiler has a stalled varying shading request." units="cycles"/>
+        <event offset="54" advanced="yes" counter="WRBUF_NO_AXI_ID_STALL" title="Tiler Cycles" name="Write buffer transaction stall cycles" description="The number of cycles where the tiler write buffer can not send data because it has no available write IDs." units="cycles"/>
+        <event offset="55" advanced="yes" counter="WRBUF_AXI_STALL" title="Tiler Cycles" name="Write buffer write stall cycles" description="The number of cycles where the tiler write buffer can not send data because the bus is not ready." units="cycles"/>
     </category>
     <category name="Memory System" per_cpu="no">
-        <event offset="132" advanced="yes" counter="MMU_REQUESTS" title="MMU Stage 1 Translations" name="MMU lookups" description="The number of main MMU address translations performed." units="requests" />
-        <event offset="144" advanced="yes" counter="L2_RD_MSG_IN" title="L2 Cache Requests" name="Read requests" description="The number of L2 cache read requests from internal masters." units="requests" />
-        <event offset="145" advanced="yes" counter="L2_RD_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Read stall cycles" description="The number of cycles L2 cache read requests from internal masters are stalled." units="cycles" />
-        <event offset="146" advanced="yes" counter="L2_WR_MSG_IN" title="L2 Cache Requests" name="Write requests" description="The number of L2 cache write requests from internal masters." units="requests" />
-        <event offset="147" advanced="yes" counter="L2_WR_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Write stall cycles" description="The number of cycles where L2 cache write requests from internal masters are stalled." units="cycles" />
-        <event offset="148" advanced="yes" counter="L2_SNP_MSG_IN" title="L2 Cache Requests" name="Snoop requests" description="The number of L2 snoop requests from internal masters." units="requests" />
-        <event offset="149" advanced="yes" counter="L2_SNP_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Snoop stall cycles" description="The number of cycles where L2 cache snoop requests from internal masters are stalled." units="cycles" />
-        <event offset="150" advanced="yes" counter="L2_RD_MSG_OUT" title="L2 Cache Requests" name="L1 read requests" description="The number of L1 cache read requests sent by the L2 cache to an internal master." units="requests" />
-        <event offset="151" advanced="yes" counter="L2_RD_MSG_OUT_STALL" title="L2 Cache Stall Cycles" name="L1 read stall cycles" description="The number of cycles where L1 cache read requests sent by the L2 cache to an internal master are stalled." units="cycles" />
-        <event offset="152" advanced="yes" counter="L2_WR_MSG_OUT" title="L2 Cache Requests" name="L1 write requests" description="The number of L1 cache write responses sent by the L2 cache to an internal master." units="requests" />
-        <event offset="153" counter="L2_ANY_LOOKUP" title="L2 Cache Lookups" name="Any lookup" description="The number of L2 cache lookups performed." units="requests" />
-        <event offset="154" counter="L2_READ_LOOKUP" title="L2 Cache Lookups" name="Read lookup" description="The number of L2 cache read lookups performed." units="requests" />
-        <event offset="155" counter="L2_WRITE_LOOKUP" title="L2 Cache Lookups" name="Write lookup" description="The number of L2 cache write lookups performed." units="requests" />
-        <event offset="156" advanced="yes" counter="L2_EXT_SNOOP_LOOKUP" title="L2 Cache Lookups" name="External snoop lookups" description="The number of coherency snoop lookups performed that were triggered by an external master." units="requests" />
-        <event offset="157" counter="L2_EXT_READ" title="External Bus Accesses" name="Read transaction" description="The number of external read transactions." units="transactions" />
-        <event offset="158" advanced="yes" counter="L2_EXT_READ_NOSNP" title="External Bus Accesses" name="ReadNoSnoop transactions" description="The number of external non-coherent read transactions." units="transactions" />
-        <event offset="159" advanced="yes" counter="L2_EXT_READ_UNIQUE" title="External Bus Accesses" name="ReadUnique transactions" description="The number of external coherent read unique transactions." units="transactions" />
-        <event offset="160" counter="L2_EXT_READ_BEATS" title="External Bus Beats" name="Read beat" description="The number of external bus data read cycles." units="beats" />
-        <event offset="161" counter="L2_EXT_AR_STALL" title="External Bus Stalls" name="Read stall cycles" description="The number of cycles where a read is stalled waiting for the external bus." units="cycles" />
-        <event offset="162" counter="L2_EXT_AR_CNT_Q1" title="External Bus Outstanding Reads" name="0-25% outstanding" description="The number of read transactions initiated when 0-25% of the maximum are in use." units="transactions" />
-        <event offset="163" counter="L2_EXT_AR_CNT_Q2" title="External Bus Outstanding Reads" name="25-50% outstanding" description="The number of read transactions initiated when 25-50% of the maximum are in use." units="transactions" />
-        <event offset="164" counter="L2_EXT_AR_CNT_Q3" title="External Bus Outstanding Reads" name="50-75% outstanding" description="The number of read transactions initiated when 50-75% of the maximum are in use." units="transactions" />
-        <event offset="165" counter="L2_EXT_RRESP_0_127" title="External Bus Read Latency" name="0-127 cycles" description="The number of data beats returned 0-127 cycles after the read request." units="beats" />
-        <event offset="166" counter="L2_EXT_RRESP_128_191" title="External Bus Read Latency" name="128-191 cycles" description="The number of data beats returned 128-191 cycles after the read request." units="beats" />
-        <event offset="167" counter="L2_EXT_RRESP_192_255" title="External Bus Read Latency" name="192-255 cycles" description="The number of data beats returned 192-255 cycles after the read request." units="beats" />
-        <event offset="168" counter="L2_EXT_RRESP_256_319" title="External Bus Read Latency" name="256-319 cycles" description="The number of data beats returned 256-319 cycles after the read request." units="beats" />
-        <event offset="169" counter="L2_EXT_RRESP_320_383" title="External Bus Read Latency" name="320-383 cycles" description="The number of data beats returned 320-383 cycles after the read request." units="beats" />
-        <event offset="170" counter="L2_EXT_WRITE" title="External Bus Accesses" name="Write transaction" description="The number of external write transactions." units="transactions" />
-        <event offset="171" advanced="yes" counter="L2_EXT_WRITE_NOSNP_FULL" title="External Bus Accesses" name="WriteNoSnoopFull transactions" description="The number of external non-coherent full write transactions." units="transactions" />
-        <event offset="172" advanced="yes" counter="L2_EXT_WRITE_NOSNP_PTL" title="External Bus Accesses" name="WriteNoSnoopPartial transactions" description="The number of external non-coherent partial write transactions." units="transactions" />
-        <event offset="173" advanced="yes" counter="L2_EXT_WRITE_SNP_FULL" title="External Bus Accesses" name="WriteSnoopFull transactions" description="The number of external coherent full write transactions." units="transactions" />
-        <event offset="174" advanced="yes" counter="L2_EXT_WRITE_SNP_PTL" title="External Bus Accesses" name="WriteSnoopPartial transactions" description="The number of external coherent partial write transactions." units="transactions" />
-        <event offset="175" counter="L2_EXT_WRITE_BEATS" title="External Bus Beats" name="Write beat" description="The number of external bus data write cycles." units="beats" />
-        <event offset="176" counter="L2_EXT_W_STALL" title="External Bus Stalls" name="Write stall cycles" description="The number of cycles where a write is stalled waiting for the external bus." units="cycles" />
-        <event offset="177" counter="L2_EXT_AW_CNT_Q1" title="External Bus Outstanding Writes" name="0-25% outstanding" description="The number of write transactions initiated when 0-25% of the maximum are in use." units="transactions" />
-        <event offset="178" counter="L2_EXT_AW_CNT_Q2" title="External Bus Outstanding Writes" name="25-50% outstanding" description="The number of write transactions initiated when 25-50% of the maximum are in use." units="transactions" />
-        <event offset="179" counter="L2_EXT_AW_CNT_Q3" title="External Bus Outstanding Writes" name="50-75% outstanding" description="The number of write transactions initiated when 50-75% of the maximum are in use." units="transactions" />
-        <event offset="180" advanced="yes" counter="L2_EXT_SNOOP" title="External Bus Accesses" name="Snoop transactions" description="The number of coherency snoops triggered by external masters." units="transactions" />
-        <event offset="181" advanced="yes" counter="L2_EXT_SNOOP_STALL" title="External Bus Stalls" name="Snoop stall cycles" description="The number of cycles where a coherency snoop triggered by external master is stalled." units="cycles" />
+        <event offset="4" advanced="yes" counter="MMU_REQUESTS" title="MMU Stage 1 Translations" name="MMU lookups" description="The number of main MMU address translations performed." units="requests"/>
+        <event offset="16" advanced="yes" counter="L2_RD_MSG_IN" title="L2 Cache Requests" name="Read requests" description="The number of L2 cache read requests from internal masters." units="requests"/>
+        <event offset="17" advanced="yes" counter="L2_RD_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Read stall cycles" description="The number of cycles L2 cache read requests from internal masters are stalled." units="cycles"/>
+        <event offset="18" advanced="yes" counter="L2_WR_MSG_IN" title="L2 Cache Requests" name="Write requests" description="The number of L2 cache write requests from internal masters." units="requests"/>
+        <event offset="19" advanced="yes" counter="L2_WR_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Write stall cycles" description="The number of cycles where L2 cache write requests from internal masters are stalled." units="cycles"/>
+        <event offset="20" advanced="yes" counter="L2_SNP_MSG_IN" title="L2 Cache Requests" name="Snoop requests" description="The number of L2 snoop requests from internal masters." units="requests"/>
+        <event offset="21" advanced="yes" counter="L2_SNP_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Snoop stall cycles" description="The number of cycles where L2 cache snoop requests from internal masters are stalled." units="cycles"/>
+        <event offset="22" advanced="yes" counter="L2_RD_MSG_OUT" title="L2 Cache Requests" name="L1 read requests" description="The number of L1 cache read requests sent by the L2 cache to an internal master." units="requests"/>
+        <event offset="23" advanced="yes" counter="L2_RD_MSG_OUT_STALL" title="L2 Cache Stall Cycles" name="L1 read stall cycles" description="The number of cycles where L1 cache read requests sent by the L2 cache to an internal master are stalled." units="cycles"/>
+        <event offset="24" advanced="yes" counter="L2_WR_MSG_OUT" title="L2 Cache Requests" name="L1 write requests" description="The number of L1 cache write responses sent by the L2 cache to an internal master." units="requests"/>
+        <event offset="25" counter="L2_ANY_LOOKUP" title="L2 Cache Lookups" name="Any lookup" description="The number of L2 cache lookups performed." units="requests"/>
+        <event offset="26" counter="L2_READ_LOOKUP" title="L2 Cache Lookups" name="Read lookup" description="The number of L2 cache read lookups performed." units="requests"/>
+        <event offset="27" counter="L2_WRITE_LOOKUP" title="L2 Cache Lookups" name="Write lookup" description="The number of L2 cache write lookups performed." units="requests"/>
+        <event offset="28" advanced="yes" counter="L2_EXT_SNOOP_LOOKUP" title="L2 Cache Lookups" name="External snoop lookups" description="The number of coherency snoop lookups performed that were triggered by an external master." units="requests"/>
+        <event offset="29" counter="L2_EXT_READ" title="External Bus Accesses" name="Read transaction" description="The number of external read transactions." units="transactions"/>
+        <event offset="30" advanced="yes" counter="L2_EXT_READ_NOSNP" title="External Bus Accesses" name="ReadNoSnoop transactions" description="The number of external non-coherent read transactions." units="transactions"/>
+        <event offset="31" advanced="yes" counter="L2_EXT_READ_UNIQUE" title="External Bus Accesses" name="ReadUnique transactions" description="The number of external coherent read unique transactions." units="transactions"/>
+        <event offset="32" counter="L2_EXT_READ_BEATS" title="External Bus Beats" name="Read beat" description="The number of external bus data read cycles." units="beats"/>
+        <event offset="33" counter="L2_EXT_AR_STALL" title="External Bus Stalls" name="Read stall cycles" description="The number of cycles where a read is stalled waiting for the external bus." units="cycles"/>
+        <event offset="34" counter="L2_EXT_AR_CNT_Q1" title="External Bus Outstanding Reads" name="0-25% outstanding" description="The number of read transactions initiated when 0-25% of the maximum are in use." units="transactions"/>
+        <event offset="35" counter="L2_EXT_AR_CNT_Q2" title="External Bus Outstanding Reads" name="25-50% outstanding" description="The number of read transactions initiated when 25-50% of the maximum are in use." units="transactions"/>
+        <event offset="36" counter="L2_EXT_AR_CNT_Q3" title="External Bus Outstanding Reads" name="50-75% outstanding" description="The number of read transactions initiated when 50-75% of the maximum are in use." units="transactions"/>
+        <event offset="37" counter="L2_EXT_RRESP_0_127" title="External Bus Read Latency" name="0-127 cycles" description="The number of data beats returned 0-127 cycles after the read request." units="beats"/>
+        <event offset="38" counter="L2_EXT_RRESP_128_191" title="External Bus Read Latency" name="128-191 cycles" description="The number of data beats returned 128-191 cycles after the read request." units="beats"/>
+        <event offset="39" counter="L2_EXT_RRESP_192_255" title="External Bus Read Latency" name="192-255 cycles" description="The number of data beats returned 192-255 cycles after the read request." units="beats"/>
+        <event offset="40" counter="L2_EXT_RRESP_256_319" title="External Bus Read Latency" name="256-319 cycles" description="The number of data beats returned 256-319 cycles after the read request." units="beats"/>
+        <event offset="41" counter="L2_EXT_RRESP_320_383" title="External Bus Read Latency" name="320-383 cycles" description="The number of data beats returned 320-383 cycles after the read request." units="beats"/>
+        <event offset="42" counter="L2_EXT_WRITE" title="External Bus Accesses" name="Write transaction" description="The number of external write transactions." units="transactions"/>
+        <event offset="43" advanced="yes" counter="L2_EXT_WRITE_NOSNP_FULL" title="External Bus Accesses" name="WriteNoSnoopFull transactions" description="The number of external non-coherent full write transactions." units="transactions"/>
+        <event offset="44" advanced="yes" counter="L2_EXT_WRITE_NOSNP_PTL" title="External Bus Accesses" name="WriteNoSnoopPartial transactions" description="The number of external non-coherent partial write transactions." units="transactions"/>
+        <event offset="45" advanced="yes" counter="L2_EXT_WRITE_SNP_FULL" title="External Bus Accesses" name="WriteSnoopFull transactions" description="The number of external coherent full write transactions." units="transactions"/>
+        <event offset="46" advanced="yes" counter="L2_EXT_WRITE_SNP_PTL" title="External Bus Accesses" name="WriteSnoopPartial transactions" description="The number of external coherent partial write transactions." units="transactions"/>
+        <event offset="47" counter="L2_EXT_WRITE_BEATS" title="External Bus Beats" name="Write beat" description="The number of external bus data write cycles." units="beats"/>
+        <event offset="48" counter="L2_EXT_W_STALL" title="External Bus Stalls" name="Write stall cycles" description="The number of cycles where a write is stalled waiting for the external bus." units="cycles"/>
+        <event offset="49" counter="L2_EXT_AW_CNT_Q1" title="External Bus Outstanding Writes" name="0-25% outstanding" description="The number of write transactions initiated when 0-25% of the maximum are in use." units="transactions"/>
+        <event offset="50" counter="L2_EXT_AW_CNT_Q2" title="External Bus Outstanding Writes" name="25-50% outstanding" description="The number of write transactions initiated when 25-50% of the maximum are in use." units="transactions"/>
+        <event offset="51" counter="L2_EXT_AW_CNT_Q3" title="External Bus Outstanding Writes" name="50-75% outstanding" description="The number of write transactions initiated when 50-75% of the maximum are in use." units="transactions"/>
+        <event offset="52" advanced="yes" counter="L2_EXT_SNOOP" title="External Bus Accesses" name="Snoop transactions" description="The number of coherency snoops triggered by external masters." units="transactions"/>
+        <event offset="53" advanced="yes" counter="L2_EXT_SNOOP_STALL" title="External Bus Stalls" name="Snoop stall cycles" description="The number of cycles where a coherency snoop triggered by external master is stalled." units="cycles"/>
+    </category>
+    <category name="Shader Core" per_cpu="no">
+        <event offset="4" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment active" description="The number of cycles where the shader core is processing a fragment workload." units="cycles"/>
+        <event offset="5" advanced="yes" counter="FRAG_PRIMITIVES" title="Core Primitives" name="Read primitives" description="The number of primitives read from the tile list by the fragment front-end." units="primitives"/>
+        <event offset="6" counter="FRAG_PRIM_RAST" title="Core Primitives" name="Rasterized primitives" description="The number of primitives being rasterized." units="primitives"/>
+        <event offset="7" counter="FRAG_FPK_ACTIVE" title="Core Cycles" name="Fragment FPKB active" description="The number of cycles where at least one quad is present in the pre-pipe quad queue." units="cycles"/>
+        <event offset="9" counter="FRAG_WARPS" title="Core Warps" name="Fragment warps" description="The number of fragment warps created." units="warps"/>
+        <event offset="10" counter="FRAG_PARTIAL_WARPS" title="Core Warps" name="Partial fragment warps" description="The number of fragment warps containing helper threads that do not correspond to a hit sample point." units="warps"/>
+        <event offset="11" counter="FRAG_QUADS_RAST" title="Core Quads" name="Rasterized quads" description="The number of quads generated by the rasterization phase." units="quads"/>
+        <event offset="12" counter="FRAG_QUADS_EZS_TEST" title="Core Quads" name="Early ZS tested quads" description="The number of quads that are undergoing early depth and stencil testing." units="quads"/>
+        <event offset="13" counter="FRAG_QUADS_EZS_UPDATE" title="Core Quads" name="Early ZS updated quads" description="The number of quads undergoing early depth and stencil testing, that are capable of updating the framebuffer." units="quads"/>
+        <event offset="14" counter="FRAG_QUADS_EZS_KILL" title="Core Quads" name="Early ZS killed quads" description="The number of quads killed by early depth and stencil testing." units="quads"/>
+        <event offset="15" counter="FRAG_LZS_TEST" title="Core Quads" name="Late ZS tested quads" description="The number of quads undergoing late depth and stencil testing." units="quads"/>
+        <event offset="16" counter="FRAG_LZS_KILL" title="Core Quads" name="Late ZS killed quads" description="The number of quads killed by late depth and stencil testing." units="quads"/>
+        <event offset="18" counter="FRAG_PTILES" title="Core Tiles" name="Tiles" description="The number of tiles processed by the shader core." units="tiles"/>
+        <event offset="19" counter="FRAG_TRANS_ELIM" title="Core Tiles" name="Constant tiles killed" description="The number of tiles killed by transaction elimination." units="tiles"/>
+        <event offset="20" counter="QUAD_FPK_KILLER" title="Core Quads" name="FPK occluder quads" description="The number of quads that are valid occluders for hidden surface removal." units="quads"/>
+        <event offset="22" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Non-fragment active" description="The number of cycles where the shader core is processing some non-fragment workload." units="cycles"/>
+        <event offset="23" advanced="yes" counter="COMPUTE_TASKS" title="Core Tasks" name="Non-fragment tasks" description="The number of non-fragment tasks issued to the shader core." units="tasks"/>
+        <event offset="24" counter="COMPUTE_WARPS" title="Core Warps" name="Non-fragment warps" description="The number of non-fragment warps created." units="warps"/>
+        <event offset="25" advanced="yes" counter="COMPUTE_STARVING" title="Core Starvation Cycles" name="Non-fragment starvation cycles" description="The number of cycles where the shader core is processing a non-fragment workload and there are no new threads available for execution." units="cycles"/>
+        <event offset="26" counter="EXEC_CORE_ACTIVE" title="Core Cycles" name="Execution core active" description="The number of cycles where the shader core is processing at least one warp." units="cycles"/>
+        <event offset="27" advanced="yes" counter="EXEC_ACTIVE" title="Core Cycles" name="Execution engine active" description="The number of cycles where the execution engine unit is processing at least one thread." units="cycles"/>
+        <event offset="28" counter="EXEC_INSTR_COUNT" title="Core EE Instructions" name="Executed instructions" description="The number of instructions executed per warp." units="instructions"/>
+        <event offset="29" counter="EXEC_INSTR_DIVERGED" title="Core EE Instructions" name="Diverged instructions" description="The number of instructions executed per warp, that have control flow divergence." units="instructions"/>
+        <event offset="30" advanced="yes" counter="EXEC_INSTR_STARVING" title="Core Starvation Cycles" name="Execution engine starvation cycles" description="The number of cycles where no new threads are available for execution." units="cycles"/>
+        <event offset="31" advanced="yes" counter="ARITH_INSTR_SINGLE_FMA" title="Core EE Instructions" name="Arithmetic instructions" description="The number of instructions where the workload is a single FMA pipe arithmetic operation." units="instructions"/>
+        <event offset="32" advanced="yes" counter="ARITH_INSTR_DOUBLE" title="Core EE Instructions" name="Dual Arithmetic instructions" description="The number of instructions where the workload is one FMA pipe arithmetic operation and one ADD pipe arithmetic operation." units="instructions"/>
+        <event offset="33" advanced="yes" counter="ARITH_INSTR_MSG" title="Core EE Instructions" name="Arithmetic + Message instructions" description="The number of instructions where the workload is one FMA pipe arithmetic operation and one ADD pipe message operation" units="instructions"/>
+        <event offset="34" advanced="yes" counter="ARITH_INSTR_MSG_ONLY" title="Core EE Instructions" name="Message instructions" description="The number of instructions where the workload is a single ADD pipe message operation, with no FMA pipe operation" units="instructions"/>
+        <event offset="35" counter="TEX_INSTR" title="Core Texture Requests" name="Texture requests" description="The number of thread-width texture operations processed." units="instructions"/>
+        <event offset="36" counter="TEX_INSTR_MIPMAP" title="Core Texture Requests" name="Mipmapped texture request" description="The number of texture operations that act on a mipmapped texture." units="instructions"/>
+        <event offset="37" counter="TEX_INSTR_COMPRESSED" title="Core Texture Requests" name="Compressed texture requests" description="The number of texture operations acting on a compressed texture." units="instructions"/>
+        <event offset="38" counter="TEX_INSTR_3D" title="Core Texture Requests" name="3D texture requests" description="The number of texture operations acting on a 3D texture." units="instructions"/>
+        <event offset="39" counter="TEX_INSTR_TRILINEAR" title="Core Texture Requests" name="Trilinear filtered requests" description="The number of texture operations using a trilinear texture filter." units="instructions"/>
+        <event offset="40" counter="TEX_COORD_ISSUE" title="Core Texture Cycles" name="Texturing active" description="The number of texture filtering issue cycles." units="cycles"/>
+        <event offset="41" advanced="yes" counter="TEX_COORD_STALL" title="Core Texture Cycles" name="Coordinate stall cycles" description="The number of clock cycles where threads are stalled at the texel coordinate calculation stage." units="cycles"/>
+        <event offset="42" advanced="yes" counter="TEX_STARVE_CACHE" title="Core Texture Cycles" name="Line fill stall cycles" description="The number of clock cycles where at least one thread is waiting for data from the texture cache, but no lookup is completed." units="cycles"/>
+        <event offset="43" advanced="yes" counter="TEX_STARVE_FILTER" title="Core Texture Cycles" name="Partial data stall cycles" description="The number of clock cycles where at least one thread fetched some data from the texture cache, but no filtering operation is started." units="cycles"/>
+        <event offset="44" counter="LS_MEM_READ_FULL" title="Core Load/Store Cycles" name="Full read cycles" description="The number of full-width load/store cache reads." units="cycles"/>
+        <event offset="45" counter="LS_MEM_READ_SHORT" title="Core Load/Store Cycles" name="Partial read cycles" description="The number of partial-width load/store cache reads." units="cycles"/>
+        <event offset="46" counter="LS_MEM_WRITE_FULL" title="Core Load/Store Cycles" name="Full write cycles" description="The number of full-width load/store cache writes." units="cycles"/>
+        <event offset="47" counter="LS_MEM_WRITE_SHORT" title="Core Load/Store Cycles" name="Partial write cycles" description="The number of partial-width load/store cache writes." units="cycles"/>
+        <event offset="48" counter="LS_MEM_ATOMIC" title="Core Load/Store Cycles" name="Atomic access cycles" description="The number of load/store atomic accesses." units="cycles"/>
+        <event offset="49" counter="VARY_INSTR" title="Core Varying Requests" name="Interpolation requests" description="The number of warp-width interpolation operations processed by the varying unit." units="instructions"/>
+        <event offset="50" counter="VARY_SLOT_32" title="Core Varying Cycles" name="32-bit interpolation active" description="The number of 32-bit interpolation cycles processed by the varying unit." units="cycles"/>
+        <event offset="51" counter="VARY_SLOT_16" title="Core Varying Cycles" name="16-bit interpolation active" description="The number of 16-bit interpolation cycles processed by the varying unit." units="cycles"/>
+        <event offset="52" advanced="yes" counter="ATTR_INSTR" title="Core Attribute Requests" name="Attribute requests" description="The number of instructions executed by the attribute unit." units="instructions"/>
+        <event offset="53" advanced="yes" counter="ARITH_INSTR_FP_MUL" title="Core EE Instructions" name="Multiplier instructions" description="The number of instructions where the workload uses floating-point multiplier hardware." units="instructions"/>
+        <event offset="54" counter="BEATS_RD_FTC" title="Core L2 Reads" name="Fragment L2 read beats" description="The number of read beats received by the fixed-function fragment front-end." units="beats"/>
+        <event offset="55" counter="BEATS_RD_FTC_EXT" title="Core L2 Reads" name="Fragment external read beats" description="The number of read beats received by the fixed-function fragment front-end that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="56" counter="BEATS_RD_LSC" title="Core L2 Reads" name="Load/store L2 read beats" description="The number of read beats received by the load/store unit." units="beats"/>
+        <event offset="57" counter="BEATS_RD_LSC_EXT" title="Core L2 Reads" name="Load/store external read beats" description="The number of read beats received by the load/store unit that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="58" counter="BEATS_RD_TEX" title="Core L2 Reads" name="Texture L2 read beats" description="The number of read beats received by the texture unit." units="beats"/>
+        <event offset="59" counter="BEATS_RD_TEX_EXT" title="Core L2 Reads" name="Texture external read beats" description="The number of read beats received by the texture unit that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="60" advanced="yes" counter="BEATS_RD_OTHER" title="Core L2 Reads" name="Other L2 read beats" description="The number of read beats received by a unit that is not specifically identified." units="beats"/>
+        <event offset="61" counter="BEATS_WR_LSC" title="Core Writes" name="Load/store write beats" description="The number of write beats sent by the load/store unit." units="beats"/>
+        <event offset="62" counter="BEATS_WR_TIB" title="Core Writes" name="Tile buffer write beats" description="The number of write beats sent by the tile buffer writeback unit." units="beats"/>
+        <event offset="63" advanced="yes" counter="BEATS_WR_OTHER" title="Core Writes" name="Other write beats" description="The number of write beats sent by any unit that is not specifically identified." units="beats"/>
     </category>
 </metrics>
\ No newline at end of file
diff --git a/src/panfrost/perf/G76.xml b/src/panfrost/perf/G76.xml
index d0b44366893..43bc89a3afa 100644
--- a/src/panfrost/perf/G76.xml
+++ b/src/panfrost/perf/G76.xml
@@ -1,6 +1,6 @@
 <!--
 Copyright © 2017-2020 ARM Limited.
-Copyright © 2021 Collabora, Ltd.
+Copyright © 2021-2022 Collabora, Ltd.
 Author: Antonio Caggiano <antonio.caggiano@collabora.com>
 
 Permission is hereby granted, free of charge, to any person obtaining a
@@ -24,156 +24,156 @@ SOFTWARE.
 -->
 <metrics id="TNOx">
     <category name="Job Manager" per_cpu="no">
-        <event offset="6" counter="GPU_ACTIVE" title="GPU Cycles" name="GPU active" description="The number of cycles where the GPU has a workload of any type queued for processing." units="cycles" />
-        <event offset="7" advanced="yes" counter="IRQ_ACTIVE" title="GPU Cycles" name="Interrupt active" description="The number of cycles where the GPU has a pending interrupt." units="cycles" />
-        <event offset="8" advanced="yes" counter="JS0_JOBS" title="GPU Jobs" name="Fragment jobs" description="The number of jobs processed by the GPU fragment queue." units="jobs" />
-        <event offset="9" counter="JS0_TASKS" title="GPU Tasks" name="Fragment tasks" description="The number of 32x32 pixel tasks processed by the GPU fragment queue." units="tasks" />
-        <event offset="10" counter="JS0_ACTIVE" title="GPU Cycles" name="Fragment queue active" description="The number of cycles where work is queued for processing in the GPU fragment queue." units="cycles" />
-        <event offset="12" advanced="yes" counter="JS0_WAIT_READ" title="GPU Wait Cycles" name="Fragment descriptor reads cycles" description="The number of cycles where queued fragment work is waiting for a descriptor load." units="cycles" />
-        <event offset="13" advanced="yes" counter="JS0_WAIT_ISSUE" title="GPU Wait Cycles" name="Fragment job issue cycles" description="The number of cycles where queued fragment work is waiting for an available processor." units="cycles" />
-        <event offset="14" advanced="yes" counter="JS0_WAIT_DEPEND" title="GPU Wait Cycles" name="Fragment job dependency cycles" description="The number of cycles where queued fragment work is waiting for dependent work to complete." units="cycles" />
-        <event offset="15" advanced="yes" counter="JS0_WAIT_FINISH" title="GPU Wait Cycles" name="Fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued fragment work to complete." units="cycles" />
-        <event offset="16" advanced="yes" counter="JS1_JOBS" title="GPU Jobs" name="Non-fragment jobs" description="The number of jobs processed by the GPU non-fragment queue." units="jobs" />
-        <event offset="17" advanced="yes" counter="JS1_TASKS" title="GPU Tasks" name="Non-fragment tasks" description="The number of tasks processed by the GPU non-fragment queue." units="tasks" />
-        <event offset="18" counter="JS1_ACTIVE" title="GPU Cycles" name="Non-fragment queue active" description="The number of cycles where work is queued in the GPU non-fragment queue." units="cycles" />
-        <event offset="20" advanced="yes" counter="JS1_WAIT_READ" title="GPU Wait Cycles" name="Non-fragment descriptor read cycles" description="The number number of cycles where queued non-fragment work is waiting for a descriptor load." units="cycles" />
-        <event offset="21" advanced="yes" counter="JS1_WAIT_ISSUE" title="GPU Wait Cycles" name="Non-fragment job issue cycles" description="The number of cycles where queued non-fragment work is waiting for an available processor." units="cycles" />
-        <event offset="22" advanced="yes" counter="JS1_WAIT_DEPEND" title="GPU Wait Cycles" name="Non-fragment job dependency cycles" description="The number of cycles where queued non-fragment work is waiting for dependent work to complete." units="cycles" />
-        <event offset="23" advanced="yes" counter="JS1_WAIT_FINISH" title="GPU Wait Cycles" name="Non-fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued non-fragment work to complete." units="cycles" />
-        <event offset="24" advanced="yes" counter="JS2_JOBS" title="GPU Jobs" name="Reserved jobs" description="The number of jobs processed by the GPU reserved queue." units="jobs" />
-        <event offset="25" advanced="yes" counter="JS2_TASKS" title="GPU Tasks" name="Reserved tasks" description="The number of tasks processed by the GPU reserved queue." units="tasks" />
-        <event offset="26" advanced="yes" counter="JS2_ACTIVE" title="GPU Cycles" name="Reserved queue active" description="The number of cycles where work is queued in the GPU reserved queue." units="cycles" />
-        <event offset="28" advanced="yes" counter="JS2_WAIT_READ" title="GPU Wait Cycles" name="Reserved descriptor read cycles" description="The number of cycles where queued reserved work is waiting for a descriptor load." units="cycles" />
-        <event offset="29" advanced="yes" counter="JS2_WAIT_ISSUE" title="GPU Wait Cycles" name="Reserved job issue cycles" description="The number of cycles where queued reserved work is waiting for an available processor." units="cycles" />
-        <event offset="30" advanced="yes" counter="JS2_WAIT_DEPEND" title="GPU Wait Cycles" name="Reserved job dependency cycles" description="The number of cycles where queued reserved work is waiting for dependent work to complete." units="cycles" />
-        <event offset="31" advanced="yes" counter="JS2_WAIT_FINISH" title="GPU Wait Cycles" name="Reserved job finish cycles" description="The number of cycles where the GPU is waiting for issued reserved work to complete." units="cycles" />
+        <event offset="6" counter="GPU_ACTIVE" title="GPU Cycles" name="GPU active" description="The number of cycles where the GPU has a workload of any type queued for processing." units="cycles"/>
+        <event offset="7" advanced="yes" counter="IRQ_ACTIVE" title="GPU Cycles" name="Interrupt active" description="The number of cycles where the GPU has a pending interrupt." units="cycles"/>
+        <event offset="8" advanced="yes" counter="JS0_JOBS" title="GPU Jobs" name="Fragment jobs" description="The number of jobs processed by the GPU fragment queue." units="jobs"/>
+        <event offset="9" counter="JS0_TASKS" title="GPU Tasks" name="Fragment tasks" description="The number of 32x32 pixel tasks processed by the GPU fragment queue." units="tasks"/>
+        <event offset="10" counter="JS0_ACTIVE" title="GPU Cycles" name="Fragment queue active" description="The number of cycles where work is queued for processing in the GPU fragment queue." units="cycles"/>
+        <event offset="12" advanced="yes" counter="JS0_WAIT_READ" title="GPU Wait Cycles" name="Fragment descriptor reads cycles" description="The number of cycles where queued fragment work is waiting for a descriptor load." units="cycles"/>
+        <event offset="13" advanced="yes" counter="JS0_WAIT_ISSUE" title="GPU Wait Cycles" name="Fragment job issue cycles" description="The number of cycles where queued fragment work is waiting for an available processor." units="cycles"/>
+        <event offset="14" advanced="yes" counter="JS0_WAIT_DEPEND" title="GPU Wait Cycles" name="Fragment job dependency cycles" description="The number of cycles where queued fragment work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="15" advanced="yes" counter="JS0_WAIT_FINISH" title="GPU Wait Cycles" name="Fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued fragment work to complete." units="cycles"/>
+        <event offset="16" advanced="yes" counter="JS1_JOBS" title="GPU Jobs" name="Non-fragment jobs" description="The number of jobs processed by the GPU non-fragment queue." units="jobs"/>
+        <event offset="17" advanced="yes" counter="JS1_TASKS" title="GPU Tasks" name="Non-fragment tasks" description="The number of tasks processed by the GPU non-fragment queue." units="tasks"/>
+        <event offset="18" counter="JS1_ACTIVE" title="GPU Cycles" name="Non-fragment queue active" description="The number of cycles where work is queued in the GPU non-fragment queue." units="cycles"/>
+        <event offset="20" advanced="yes" counter="JS1_WAIT_READ" title="GPU Wait Cycles" name="Non-fragment descriptor read cycles" description="The number number of cycles where queued non-fragment work is waiting for a descriptor load." units="cycles"/>
+        <event offset="21" advanced="yes" counter="JS1_WAIT_ISSUE" title="GPU Wait Cycles" name="Non-fragment job issue cycles" description="The number of cycles where queued non-fragment work is waiting for an available processor." units="cycles"/>
+        <event offset="22" advanced="yes" counter="JS1_WAIT_DEPEND" title="GPU Wait Cycles" name="Non-fragment job dependency cycles" description="The number of cycles where queued non-fragment work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="23" advanced="yes" counter="JS1_WAIT_FINISH" title="GPU Wait Cycles" name="Non-fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued non-fragment work to complete." units="cycles"/>
+        <event offset="24" advanced="yes" counter="JS2_JOBS" title="GPU Jobs" name="Reserved jobs" description="The number of jobs processed by the GPU reserved queue." units="jobs"/>
+        <event offset="25" advanced="yes" counter="JS2_TASKS" title="GPU Tasks" name="Reserved tasks" description="The number of tasks processed by the GPU reserved queue." units="tasks"/>
+        <event offset="26" advanced="yes" counter="JS2_ACTIVE" title="GPU Cycles" name="Reserved queue active" description="The number of cycles where work is queued in the GPU reserved queue." units="cycles"/>
+        <event offset="28" advanced="yes" counter="JS2_WAIT_READ" title="GPU Wait Cycles" name="Reserved descriptor read cycles" description="The number of cycles where queued reserved work is waiting for a descriptor load." units="cycles"/>
+        <event offset="29" advanced="yes" counter="JS2_WAIT_ISSUE" title="GPU Wait Cycles" name="Reserved job issue cycles" description="The number of cycles where queued reserved work is waiting for an available processor." units="cycles"/>
+        <event offset="30" advanced="yes" counter="JS2_WAIT_DEPEND" title="GPU Wait Cycles" name="Reserved job dependency cycles" description="The number of cycles where queued reserved work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="31" advanced="yes" counter="JS2_WAIT_FINISH" title="GPU Wait Cycles" name="Reserved job finish cycles" description="The number of cycles where the GPU is waiting for issued reserved work to complete." units="cycles"/>
     </category>
     <category name="Tiler" per_cpu="no">
-        <event offset="68" counter="TILER_ACTIVE" title="GPU Cycles" name="Tiler active" description="The number of cycles where the tiler has a workload queued for processing." units="cycles" />
-        <event offset="70" counter="TRIANGLES" title="Input Primitives" name="Triangle primitives" description="The number of input triangle primitives." units="primitives" />
-        <event offset="71" counter="LINES" title="Input Primitives" name="Line primitives" description="The number of input line primitives." units="primitives" />
-        <event offset="72" counter="POINTS" title="Input Primitives" name="Point primitives" description="The number of input point primitives." units="primitives" />
-        <event offset="73" counter="FRONT_FACING" title="Visible Primitives" name="Front-facing primitives" description="The number of front-facing triangles that are visible after culling." units="primitives" />
-        <event offset="74" counter="BACK_FACING" title="Visible Primitives" name="Back-facing primitives" description="The number of back-facing triangles that are visible after culling." units="primitives" />
-        <event offset="75" counter="PRIM_VISIBLE" title="Primitive Culling" name="Visible primitives" description="The number of primitives that are visible after culling." units="primitives" />
-        <event offset="76" counter="PRIM_CULLED" title="Primitive Culling" name="Facing and XY plane test culled primitives" description="The number of primitives that are culled by facing or frustum XY plane tests." units="primitives" />
-        <event offset="77" counter="PRIM_CLIPPED" title="Primitive Culling" name="Z plane test culled primitives" description="The number of primitives that are culled by frustum Z plane tests." units="primitives" />
-        <event offset="78" counter="PRIM_SAT_CULLED" title="Primitive Culling" name="Sample test culled primitives" description="The number of primitives culled by the sample coverage test." units="primitives" />
-        <event offset="81" advanced="yes" counter="BUS_READ" title="Tiler L2 Accesses" name="Read beats" description="The number of internal bus data read cycles made by the tiler." units="beats" />
-        <event offset="83" advanced="yes" counter="BUS_WRITE" title="Tiler L2 Accesses" name="Write beats" description="The number of internal bus data write cycles made by the tiler." units="beats" />
-        <event offset="85" counter="IDVS_POS_SHAD_REQ" title="Tiler Shading Requests" name="Position shading requests" description="The number of position shading requests in the IDVS flow." units="requests" />
-        <event offset="87" advanced="yes" counter="IDVS_POS_SHAD_STALL" title="Tiler Cycles" name="Position shading stall cycles" description="The number of cycles where the tiler has a stalled position shading request." units="cycles" />
-        <event offset="88" advanced="yes" counter="IDVS_POS_FIFO_FULL" title="Tiler Cycles" name="Position FIFO full cycles" description="The number of cycles where the tiler has a stalled position shading buffer." units="cycles" />
-        <event offset="90" advanced="yes" counter="VCACHE_HIT" title="Tiler Vertex Cache" name="Position cache hits" description="The number of position lookups that result in a hit in the vertex cache." units="requests" />
-        <event offset="91" advanced="yes" counter="VCACHE_MISS" title="Tiler Vertex Cache" name="Position cache misses" description="The number of position lookups that miss in the vertex cache." units="requests" />
-        <event offset="95" advanced="yes" counter="VFETCH_STALL" title="Tiler Cycles" name="Primitive assembly busy stall cycles" description="The number of cycles where the tiler is stalled waiting for primitive assembly." units="cycles" />
-        <event offset="98" advanced="yes" counter="IDVS_VBU_HIT" title="Tiler Vertex Cache" name="Varying cache hits" description="The number of varying lookups that result in a hit in the vertex cache." units="requests" />
-        <event offset="99" advanced="yes" counter="IDVS_VBU_MISS" title="Tiler Vertex Cache" name="Varying cache misses" description="The number of varying lookups that miss in the vertex cache." units="requests" />
-        <event offset="101" counter="IDVS_VAR_SHAD_REQ" title="Tiler Shading Requests" name="Varying shading requests" description="The number of varying shading requests in the IDVS flow." units="requests" />
-        <event offset="102" advanced="yes" counter="IDVS_VAR_SHAD_STALL" title="Tiler Cycles" name="Varying shading stall cycles" description="The number of cycles where the tiler has a stalled varying shading request." units="cycles" />
-        <event offset="118" advanced="yes" counter="WRBUF_NO_AXI_ID_STALL" title="Tiler Cycles" name="Write buffer transaction stall cycles" description="The number of cycles where the tiler write buffer can not send data because it has no available write IDs." units="cycles" />
-        <event offset="119" advanced="yes" counter="WRBUF_AXI_STALL" title="Tiler Cycles" name="Write buffer write stall cycles" description="The number of cycles where the tiler write buffer can not send data because the bus is not ready." units="cycles" />
-    </category>
-    <category name="Shader Core" per_cpu="no">
-        <event offset="196" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment active" description="The number of cycles where the shader core is processing a fragment workload." units="cycles" />
-        <event offset="197" advanced="yes" counter="FRAG_PRIMITIVES" title="Core Primitives" name="Read primitives" description="The number of primitives read from the tile list by the fragment front-end." units="primitives" />
-        <event offset="198" counter="FRAG_PRIM_RAST" title="Core Primitives" name="Rasterized primitives" description="The number of primitives being rasterized." units="primitives" />
-        <event offset="199" counter="FRAG_FPK_ACTIVE" title="Core Cycles" name="Fragment FPKB active" description="The number of cycles where at least one quad is present in the pre-pipe quad queue." units="cycles" />
-        <event offset="201" counter="FRAG_WARPS" title="Core Warps" name="Fragment warps" description="The number of fragment warps created." units="warps" />
-        <event offset="202" counter="FRAG_PARTIAL_WARPS" title="Core Warps" name="Partial fragment warps" description="The number of fragment warps containing helper threads that do not correspond to a hit sample point." units="warps" />
-        <event offset="203" counter="FRAG_QUADS_RAST" title="Core Quads" name="Rasterized quads" description="The number of quads generated by the rasterization phase." units="quads" />
-        <event offset="204" counter="FRAG_QUADS_EZS_TEST" title="Core Quads" name="Early ZS tested quads" description="The number of quads that are undergoing early depth and stencil testing." units="quads" />
-        <event offset="205" counter="FRAG_QUADS_EZS_UPDATE" title="Core Quads" name="Early ZS updated quads" description="The number of quads undergoing early depth and stencil testing, that are capable of updating the framebuffer." units="quads" />
-        <event offset="206" counter="FRAG_QUADS_EZS_KILL" title="Core Quads" name="Early ZS killed quads" description="The number of quads killed by early depth and stencil testing." units="quads" />
-        <event offset="207" counter="FRAG_LZS_TEST" title="Core Quads" name="Late ZS tested quads" description="The number of quads undergoing late depth and stencil testing." units="quads" />
-        <event offset="208" counter="FRAG_LZS_KILL" title="Core Quads" name="Late ZS killed quads" description="The number of quads killed by late depth and stencil testing." units="quads" />
-        <event offset="209" counter="WARP_REG_SIZE_64" title="Core Warps" name="All register warps" description="The number of warps that require more than 32 registers." units="warps" />
-        <event offset="210" counter="FRAG_PTILES" title="Core Tiles" name="Tiles" description="The number of tiles processed by the shader core." units="tiles" />
-        <event offset="211" counter="FRAG_TRANS_ELIM" title="Core Tiles" name="Constant tiles killed" description="The number of tiles killed by transaction elimination." units="tiles" />
-        <event offset="212" counter="QUAD_FPK_KILLER" title="Core Quads" name="FPK occluder quads" description="The number of quads that are valid occluders for hidden surface removal." units="quads" />
-        <event offset="213" counter="FULL_QUAD_WARPS" title="Core Warps" name="Full quad warps" description="The number of warps that are fully populated with quads." units="warps" />
-        <event offset="214" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Non-fragment active" description="The number of cycles where the shader core is processing some non-fragment workload." units="cycles" />
-        <event offset="215" advanced="yes" counter="COMPUTE_TASKS" title="Core Tasks" name="Non-fragment tasks" description="The number of non-fragment tasks issued to the shader core." units="tasks" />
-        <event offset="216" counter="COMPUTE_WARPS" title="Core Warps" name="Non-fragment warps" description="The number of non-fragment warps created." units="warps" />
-        <event offset="217" advanced="yes" counter="COMPUTE_STARVING" title="Core Starvation Cycles" name="Non-fragment starvation cycles" description="The number of cycles where the shader core is processing a non-fragment workload and there are no new threads available for execution." units="cycles" />
-        <event offset="218" counter="EXEC_CORE_ACTIVE" title="Core Cycles" name="Execution core active" description="The number of cycles where the shader core is processing at least one warp." units="cycles" />
-        <event offset="219" advanced="yes" counter="EXEC_ACTIVE" title="Core Cycles" name="Execution engine active" description="The number of cycles where the execution engine unit is processing at least one thread." units="cycles" />
-        <event offset="220" counter="EXEC_INSTR_COUNT" title="Core EE Instructions" name="Executed instructions" description="The number of instructions executed per warp." units="instructions" />
-        <event offset="221" counter="EXEC_INSTR_DIVERGED" title="Core EE Instructions" name="Diverged instructions" description="The number of instructions executed per warp, that have control flow divergence." units="instructions" />
-        <event offset="222" advanced="yes" counter="EXEC_INSTR_STARVING" title="Core Starvation Cycles" name="Execution engine starvation cycles" description="The number of cycles where no new threads are available for execution." units="cycles" />
-        <event offset="223" advanced="yes" counter="ARITH_INSTR_SINGLE_FMA" title="Core EE Instructions" name="Arithmetic instructions" description="The number of instructions where the workload is a single FMA pipe arithmetic operation." units="instructions" />
-        <event offset="224" advanced="yes" counter="ARITH_INSTR_DOUBLE" title="Core EE Instructions" name="Dual Arithmetic instructions" description="The number of instructions where the workload is one FMA pipe arithmetic operation and one ADD pipe arithmetic operation." units="instructions" />
-        <event offset="225" advanced="yes" counter="ARITH_INSTR_MSG" title="Core EE Instructions" name="Arithmetic + Message instructions" description="The number of instructions where the workload is one FMA pipe arithmetic operation and one ADD pipe message operation" units="instructions" />
-        <event offset="226" advanced="yes" counter="ARITH_INSTR_MSG_ONLY" title="Core EE Instructions" name="Message instructions" description="The number of instructions where the workload is a single ADD pipe message operation, with no FMA pipe operation" units="instructions" />
-        <event offset="227" counter="TEX_MSGI_NUM_QUADS" title="Core Texture Quads" name="Texture requests" description="The number of quad-width texture operations processed by the texture unit." units="quads" />
-        <event offset="228" counter="TEX_DFCH_NUM_PASSES" title="Core Texture Quads" name="Texture issues" description="The number of quad-width filtering passes." units="issues" />
-        <event offset="229" counter="TEX_DFCH_NUM_PASSES_MISS" title="Core Texture Quads" name="Descriptor misses" description="The number of quad-width filtering passes that miss in the resource or sampler descriptor cache." units="requests" />
-        <event offset="230" counter="TEX_DFCH_NUM_PASSES_MIP_MAP" title="Core Texture Quads" name="Mipmapped texture issues" description="The number of quad-width filtering passes that use a mipmapped texture." units="issues" />
-        <event offset="231" counter="TEX_TIDX_NUM_SPLIT_MIP_MAP" title="Core Texture Quads" name="Trilinear filtered issues" description="The number of quad-width filtering passes that use a trilinear filter." units="issues" />
-        <event offset="232" counter="TEX_TFCH_NUM_LINES_FETCHED" title="Core Texture Line Fetches" name="Line fetches" description="The number of texture line fetches from the L2 cache." units="issues" />
-        <event offset="233" counter="TEX_TFCH_NUM_LINES_FETCHED_BLOCK_COMPRESSED" title="Core Texture Line Fetches" name="Compressed line fetches" description="The number of texture line fetches from the L2 cache that are block compressed textures." units="issues" />
-        <event offset="234" counter="TEX_TFCH_NUM_OPERATIONS" title="Core Texture Cycles" name="Cache lookups" description="The number of texture cache lookup cycles." units="requests" />
-        <event offset="235" counter="TEX_FILT_NUM_OPERATIONS" title="Core Texture Cycles" name="Texturing active" description="The number of texture filtering issue cycles." units="cycles" />
-        <event offset="236" counter="LS_MEM_READ_FULL" title="Core Load/Store Cycles" name="Full read cycles" description="The number of full-width load/store cache reads." units="cycles" />
-        <event offset="237" counter="LS_MEM_READ_SHORT" title="Core Load/Store Cycles" name="Partial read cycles" description="The number of partial-width load/store cache reads." units="cycles" />
-        <event offset="238" counter="LS_MEM_WRITE_FULL" title="Core Load/Store Cycles" name="Full write cycles" description="The number of full-width load/store cache writes." units="cycles" />
-        <event offset="239" counter="LS_MEM_WRITE_SHORT" title="Core Load/Store Cycles" name="Partial write cycles" description="The number of partial-width load/store cache writes." units="cycles" />
-        <event offset="240" counter="LS_MEM_ATOMIC" title="Core Load/Store Cycles" name="Atomic access cycles" description="The number of load/store atomic accesses." units="cycles" />
-        <event offset="241" counter="VARY_INSTR" title="Core Varying Requests" name="Interpolation requests" description="The number of warp-width interpolation operations processed by the varying unit." units="instructions" />
-        <event offset="242" counter="VARY_SLOT_32" title="Core Varying Cycles" name="32-bit interpolation active" description="The number of 32-bit interpolation cycles processed by the varying unit." units="cycles" />
-        <event offset="243" counter="VARY_SLOT_16" title="Core Varying Cycles" name="16-bit interpolation active" description="The number of 16-bit interpolation cycles processed by the varying unit." units="cycles" />
-        <event offset="244" advanced="yes" counter="ATTR_INSTR" title="Core Attribute Requests" name="Attribute requests" description="The number of instructions executed by the attribute unit." units="instructions" />
-        <event offset="245" advanced="yes" counter="ARITH_INSTR_FP_MUL" title="Core EE Instructions" name="Multiplier instructions" description="The number of instructions where the workload uses floating-point multiplier hardware." units="instructions" />
-        <event offset="246" counter="BEATS_RD_FTC" title="Core L2 Reads" name="Fragment L2 read beats" description="The number of read beats received by the fixed-function fragment front-end." units="beats" />
-        <event offset="247" counter="BEATS_RD_FTC_EXT" title="Core L2 Reads" name="Fragment external read beats" description="The number of read beats received by the fixed-function fragment front-end that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="248" counter="BEATS_RD_LSC" title="Core L2 Reads" name="Load/store L2 read beats" description="The number of read beats received by the load/store unit." units="beats" />
-        <event offset="249" counter="BEATS_RD_LSC_EXT" title="Core L2 Reads" name="Load/store external read beats" description="The number of read beats received by the load/store unit that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="250" counter="BEATS_RD_TEX" title="Core L2 Reads" name="Texture L2 read beats" description="The number of read beats received by the texture unit." units="beats" />
-        <event offset="251" counter="BEATS_RD_TEX_EXT" title="Core L2 Reads" name="Texture external read beats" description="The number of read beats received by the texture unit that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="252" advanced="yes" counter="BEATS_RD_OTHER" title="Core L2 Reads" name="Other L2 read beats" description="The number of read beats received by a unit that is not specifically identified." units="beats" />
-        <event offset="253" counter="BEATS_WR_LSC_OTHER" title="Core Writes" name="Load/store other write beats" description="The number of write beats by the load/store unit that are due to any reason other than writeback." units="beats" />
-        <event offset="254" counter="BEATS_WR_TIB" title="Core Writes" name="Tile buffer write beats" description="The number of write beats sent by the tile buffer writeback unit." units="beats" />
-        <event offset="255" counter="BEATS_WR_LSC_WB" title="Core Writes" name="Load/store writeback write beats" description="The number of write beats by the load/store unit that are due to writeback." units="beats" />
+        <event offset="4" counter="TILER_ACTIVE" title="GPU Cycles" name="Tiler active" description="The number of cycles where the tiler has a workload queued for processing." units="cycles"/>
+        <event offset="6" counter="TRIANGLES" title="Input Primitives" name="Triangle primitives" description="The number of input triangle primitives." units="primitives"/>
+        <event offset="7" counter="LINES" title="Input Primitives" name="Line primitives" description="The number of input line primitives." units="primitives"/>
+        <event offset="8" counter="POINTS" title="Input Primitives" name="Point primitives" description="The number of input point primitives." units="primitives"/>
+        <event offset="9" counter="FRONT_FACING" title="Visible Primitives" name="Front-facing primitives" description="The number of front-facing triangles that are visible after culling." units="primitives"/>
+        <event offset="10" counter="BACK_FACING" title="Visible Primitives" name="Back-facing primitives" description="The number of back-facing triangles that are visible after culling." units="primitives"/>
+        <event offset="11" counter="PRIM_VISIBLE" title="Primitive Culling" name="Visible primitives" description="The number of primitives that are visible after culling." units="primitives"/>
+        <event offset="12" counter="PRIM_CULLED" title="Primitive Culling" name="Facing and XY plane test culled primitives" description="The number of primitives that are culled by facing or frustum XY plane tests." units="primitives"/>
+        <event offset="13" counter="PRIM_CLIPPED" title="Primitive Culling" name="Z plane test culled primitives" description="The number of primitives that are culled by frustum Z plane tests." units="primitives"/>
+        <event offset="14" counter="PRIM_SAT_CULLED" title="Primitive Culling" name="Sample test culled primitives" description="The number of primitives culled by the sample coverage test." units="primitives"/>
+        <event offset="17" advanced="yes" counter="BUS_READ" title="Tiler L2 Accesses" name="Read beats" description="The number of internal bus data read cycles made by the tiler." units="beats"/>
+        <event offset="19" advanced="yes" counter="BUS_WRITE" title="Tiler L2 Accesses" name="Write beats" description="The number of internal bus data write cycles made by the tiler." units="beats"/>
+        <event offset="21" counter="IDVS_POS_SHAD_REQ" title="Tiler Shading Requests" name="Position shading requests" description="The number of position shading requests in the IDVS flow." units="requests"/>
+        <event offset="23" advanced="yes" counter="IDVS_POS_SHAD_STALL" title="Tiler Cycles" name="Position shading stall cycles" description="The number of cycles where the tiler has a stalled position shading request." units="cycles"/>
+        <event offset="24" advanced="yes" counter="IDVS_POS_FIFO_FULL" title="Tiler Cycles" name="Position FIFO full cycles" description="The number of cycles where the tiler has a stalled position shading buffer." units="cycles"/>
+        <event offset="26" advanced="yes" counter="VCACHE_HIT" title="Tiler Vertex Cache" name="Position cache hits" description="The number of position lookups that result in a hit in the vertex cache." units="requests"/>
+        <event offset="27" advanced="yes" counter="VCACHE_MISS" title="Tiler Vertex Cache" name="Position cache misses" description="The number of position lookups that miss in the vertex cache." units="requests"/>
+        <event offset="31" advanced="yes" counter="VFETCH_STALL" title="Tiler Cycles" name="Primitive assembly busy stall cycles" description="The number of cycles where the tiler is stalled waiting for primitive assembly." units="cycles"/>
+        <event offset="34" advanced="yes" counter="IDVS_VBU_HIT" title="Tiler Vertex Cache" name="Varying cache hits" description="The number of varying lookups that result in a hit in the vertex cache." units="requests"/>
+        <event offset="35" advanced="yes" counter="IDVS_VBU_MISS" title="Tiler Vertex Cache" name="Varying cache misses" description="The number of varying lookups that miss in the vertex cache." units="requests"/>
+        <event offset="37" counter="IDVS_VAR_SHAD_REQ" title="Tiler Shading Requests" name="Varying shading requests" description="The number of varying shading requests in the IDVS flow." units="requests"/>
+        <event offset="38" advanced="yes" counter="IDVS_VAR_SHAD_STALL" title="Tiler Cycles" name="Varying shading stall cycles" description="The number of cycles where the tiler has a stalled varying shading request." units="cycles"/>
+        <event offset="54" advanced="yes" counter="WRBUF_NO_AXI_ID_STALL" title="Tiler Cycles" name="Write buffer transaction stall cycles" description="The number of cycles where the tiler write buffer can not send data because it has no available write IDs." units="cycles"/>
+        <event offset="55" advanced="yes" counter="WRBUF_AXI_STALL" title="Tiler Cycles" name="Write buffer write stall cycles" description="The number of cycles where the tiler write buffer can not send data because the bus is not ready." units="cycles"/>
     </category>
     <category name="Memory System" per_cpu="no">
-        <event offset="132" advanced="yes" counter="MMU_REQUESTS" title="MMU Stage 1 Translations" name="MMU lookups" description="The number of main MMU address translations performed." units="requests" />
-        <event offset="144" advanced="yes" counter="L2_RD_MSG_IN" title="L2 Cache Requests" name="Read requests" description="The number of L2 cache read requests from internal masters." units="requests" />
-        <event offset="145" advanced="yes" counter="L2_RD_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Read stall cycles" description="The number of cycles L2 cache read requests from internal masters are stalled." units="cycles" />
-        <event offset="146" advanced="yes" counter="L2_WR_MSG_IN" title="L2 Cache Requests" name="Write requests" description="The number of L2 cache write requests from internal masters." units="requests" />
-        <event offset="147" advanced="yes" counter="L2_WR_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Write stall cycles" description="The number of cycles where L2 cache write requests from internal masters are stalled." units="cycles" />
-        <event offset="148" advanced="yes" counter="L2_SNP_MSG_IN" title="L2 Cache Requests" name="Snoop requests" description="The number of L2 snoop requests from internal masters." units="requests" />
-        <event offset="149" advanced="yes" counter="L2_SNP_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Snoop stall cycles" description="The number of cycles where L2 cache snoop requests from internal masters are stalled." units="cycles" />
-        <event offset="150" advanced="yes" counter="L2_RD_MSG_OUT" title="L2 Cache Requests" name="L1 read requests" description="The number of L1 cache read requests sent by the L2 cache to an internal master." units="requests" />
-        <event offset="151" advanced="yes" counter="L2_RD_MSG_OUT_STALL" title="L2 Cache Stall Cycles" name="L1 read stall cycles" description="The number of cycles where L1 cache read requests sent by the L2 cache to an internal master are stalled." units="cycles" />
-        <event offset="152" advanced="yes" counter="L2_WR_MSG_OUT" title="L2 Cache Requests" name="L1 write requests" description="The number of L1 cache write responses sent by the L2 cache to an internal master." units="requests" />
-        <event offset="153" counter="L2_ANY_LOOKUP" title="L2 Cache Lookups" name="Any lookup" description="The number of L2 cache lookups performed." units="requests" />
-        <event offset="154" counter="L2_READ_LOOKUP" title="L2 Cache Lookups" name="Read lookup" description="The number of L2 cache read lookups performed." units="requests" />
-        <event offset="155" counter="L2_WRITE_LOOKUP" title="L2 Cache Lookups" name="Write lookup" description="The number of L2 cache write lookups performed." units="requests" />
-        <event offset="156" advanced="yes" counter="L2_EXT_SNOOP_LOOKUP" title="L2 Cache Lookups" name="External snoop lookups" description="The number of coherency snoop lookups performed that were triggered by an external master." units="requests" />
-        <event offset="157" counter="L2_EXT_READ" title="External Bus Accesses" name="Read transaction" description="The number of external read transactions." units="transactions" />
-        <event offset="158" advanced="yes" counter="L2_EXT_READ_NOSNP" title="External Bus Accesses" name="ReadNoSnoop transactions" description="The number of external non-coherent read transactions." units="transactions" />
-        <event offset="159" advanced="yes" counter="L2_EXT_READ_UNIQUE" title="External Bus Accesses" name="ReadUnique transactions" description="The number of external coherent read unique transactions." units="transactions" />
-        <event offset="160" counter="L2_EXT_READ_BEATS" title="External Bus Beats" name="Read beat" description="The number of external bus data read cycles." units="beats" />
-        <event offset="161" counter="L2_EXT_AR_STALL" title="External Bus Stalls" name="Read stall cycles" description="The number of cycles where a read is stalled waiting for the external bus." units="cycles" />
-        <event offset="162" counter="L2_EXT_AR_CNT_Q1" title="External Bus Outstanding Reads" name="0-25% outstanding" description="The number of read transactions initiated when 0-25% of the maximum are in use." units="transactions" />
-        <event offset="163" counter="L2_EXT_AR_CNT_Q2" title="External Bus Outstanding Reads" name="25-50% outstanding" description="The number of read transactions initiated when 25-50% of the maximum are in use." units="transactions" />
-        <event offset="164" counter="L2_EXT_AR_CNT_Q3" title="External Bus Outstanding Reads" name="50-75% outstanding" description="The number of read transactions initiated when 50-75% of the maximum are in use." units="transactions" />
-        <event offset="165" counter="L2_EXT_RRESP_0_127" title="External Bus Read Latency" name="0-127 cycles" description="The number of data beats returned 0-127 cycles after the read request." units="beats" />
-        <event offset="166" counter="L2_EXT_RRESP_128_191" title="External Bus Read Latency" name="128-191 cycles" description="The number of data beats returned 128-191 cycles after the read request." units="beats" />
-        <event offset="167" counter="L2_EXT_RRESP_192_255" title="External Bus Read Latency" name="192-255 cycles" description="The number of data beats returned 192-255 cycles after the read request." units="beats" />
-        <event offset="168" counter="L2_EXT_RRESP_256_319" title="External Bus Read Latency" name="256-319 cycles" description="The number of data beats returned 256-319 cycles after the read request." units="beats" />
-        <event offset="169" counter="L2_EXT_RRESP_320_383" title="External Bus Read Latency" name="320-383 cycles" description="The number of data beats returned 320-383 cycles after the read request." units="beats" />
-        <event offset="170" counter="L2_EXT_WRITE" title="External Bus Accesses" name="Write transaction" description="The number of external write transactions." units="transactions" />
-        <event offset="171" advanced="yes" counter="L2_EXT_WRITE_NOSNP_FULL" title="External Bus Accesses" name="WriteNoSnoopFull transactions" description="The number of external non-coherent full write transactions." units="transactions" />
-        <event offset="172" advanced="yes" counter="L2_EXT_WRITE_NOSNP_PTL" title="External Bus Accesses" name="WriteNoSnoopPartial transactions" description="The number of external non-coherent partial write transactions." units="transactions" />
-        <event offset="173" advanced="yes" counter="L2_EXT_WRITE_SNP_FULL" title="External Bus Accesses" name="WriteSnoopFull transactions" description="The number of external coherent full write transactions." units="transactions" />
-        <event offset="174" advanced="yes" counter="L2_EXT_WRITE_SNP_PTL" title="External Bus Accesses" name="WriteSnoopPartial transactions" description="The number of external coherent partial write transactions." units="transactions" />
-        <event offset="175" counter="L2_EXT_WRITE_BEATS" title="External Bus Beats" name="Write beat" description="The number of external bus data write cycles." units="beats" />
-        <event offset="176" counter="L2_EXT_W_STALL" title="External Bus Stalls" name="Write stall cycles" description="The number of cycles where a write is stalled waiting for the external bus." units="cycles" />
-        <event offset="177" counter="L2_EXT_AW_CNT_Q1" title="External Bus Outstanding Writes" name="0-25% outstanding" description="The number of write transactions initiated when 0-25% of the maximum are in use." units="transactions" />
-        <event offset="178" counter="L2_EXT_AW_CNT_Q2" title="External Bus Outstanding Writes" name="25-50% outstanding" description="The number of write transactions initiated when 25-50% of the maximum are in use." units="transactions" />
-        <event offset="179" counter="L2_EXT_AW_CNT_Q3" title="External Bus Outstanding Writes" name="50-75% outstanding" description="The number of write transactions initiated when 50-75% of the maximum are in use." units="transactions" />
-        <event offset="180" advanced="yes" counter="L2_EXT_SNOOP" title="External Bus Accesses" name="Snoop transactions" description="The number of coherency snoops triggered by external masters." units="transactions" />
-        <event offset="181" advanced="yes" counter="L2_EXT_SNOOP_STALL" title="External Bus Stalls" name="Snoop stall cycles" description="The number of cycles where a coherency snoop triggered by external master is stalled." units="cycles" />
+        <event offset="4" advanced="yes" counter="MMU_REQUESTS" title="MMU Stage 1 Translations" name="MMU lookups" description="The number of main MMU address translations performed." units="requests"/>
+        <event offset="16" advanced="yes" counter="L2_RD_MSG_IN" title="L2 Cache Requests" name="Read requests" description="The number of L2 cache read requests from internal masters." units="requests"/>
+        <event offset="17" advanced="yes" counter="L2_RD_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Read stall cycles" description="The number of cycles L2 cache read requests from internal masters are stalled." units="cycles"/>
+        <event offset="18" advanced="yes" counter="L2_WR_MSG_IN" title="L2 Cache Requests" name="Write requests" description="The number of L2 cache write requests from internal masters." units="requests"/>
+        <event offset="19" advanced="yes" counter="L2_WR_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Write stall cycles" description="The number of cycles where L2 cache write requests from internal masters are stalled." units="cycles"/>
+        <event offset="20" advanced="yes" counter="L2_SNP_MSG_IN" title="L2 Cache Requests" name="Snoop requests" description="The number of L2 snoop requests from internal masters." units="requests"/>
+        <event offset="21" advanced="yes" counter="L2_SNP_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Snoop stall cycles" description="The number of cycles where L2 cache snoop requests from internal masters are stalled." units="cycles"/>
+        <event offset="22" advanced="yes" counter="L2_RD_MSG_OUT" title="L2 Cache Requests" name="L1 read requests" description="The number of L1 cache read requests sent by the L2 cache to an internal master." units="requests"/>
+        <event offset="23" advanced="yes" counter="L2_RD_MSG_OUT_STALL" title="L2 Cache Stall Cycles" name="L1 read stall cycles" description="The number of cycles where L1 cache read requests sent by the L2 cache to an internal master are stalled." units="cycles"/>
+        <event offset="24" advanced="yes" counter="L2_WR_MSG_OUT" title="L2 Cache Requests" name="L1 write requests" description="The number of L1 cache write responses sent by the L2 cache to an internal master." units="requests"/>
+        <event offset="25" counter="L2_ANY_LOOKUP" title="L2 Cache Lookups" name="Any lookup" description="The number of L2 cache lookups performed." units="requests"/>
+        <event offset="26" counter="L2_READ_LOOKUP" title="L2 Cache Lookups" name="Read lookup" description="The number of L2 cache read lookups performed." units="requests"/>
+        <event offset="27" counter="L2_WRITE_LOOKUP" title="L2 Cache Lookups" name="Write lookup" description="The number of L2 cache write lookups performed." units="requests"/>
+        <event offset="28" advanced="yes" counter="L2_EXT_SNOOP_LOOKUP" title="L2 Cache Lookups" name="External snoop lookups" description="The number of coherency snoop lookups performed that were triggered by an external master." units="requests"/>
+        <event offset="29" counter="L2_EXT_READ" title="External Bus Accesses" name="Read transaction" description="The number of external read transactions." units="transactions"/>
+        <event offset="30" advanced="yes" counter="L2_EXT_READ_NOSNP" title="External Bus Accesses" name="ReadNoSnoop transactions" description="The number of external non-coherent read transactions." units="transactions"/>
+        <event offset="31" advanced="yes" counter="L2_EXT_READ_UNIQUE" title="External Bus Accesses" name="ReadUnique transactions" description="The number of external coherent read unique transactions." units="transactions"/>
+        <event offset="32" counter="L2_EXT_READ_BEATS" title="External Bus Beats" name="Read beat" description="The number of external bus data read cycles." units="beats"/>
+        <event offset="33" counter="L2_EXT_AR_STALL" title="External Bus Stalls" name="Read stall cycles" description="The number of cycles where a read is stalled waiting for the external bus." units="cycles"/>
+        <event offset="34" counter="L2_EXT_AR_CNT_Q1" title="External Bus Outstanding Reads" name="0-25% outstanding" description="The number of read transactions initiated when 0-25% of the maximum are in use." units="transactions"/>
+        <event offset="35" counter="L2_EXT_AR_CNT_Q2" title="External Bus Outstanding Reads" name="25-50% outstanding" description="The number of read transactions initiated when 25-50% of the maximum are in use." units="transactions"/>
+        <event offset="36" counter="L2_EXT_AR_CNT_Q3" title="External Bus Outstanding Reads" name="50-75% outstanding" description="The number of read transactions initiated when 50-75% of the maximum are in use." units="transactions"/>
+        <event offset="37" counter="L2_EXT_RRESP_0_127" title="External Bus Read Latency" name="0-127 cycles" description="The number of data beats returned 0-127 cycles after the read request." units="beats"/>
+        <event offset="38" counter="L2_EXT_RRESP_128_191" title="External Bus Read Latency" name="128-191 cycles" description="The number of data beats returned 128-191 cycles after the read request." units="beats"/>
+        <event offset="39" counter="L2_EXT_RRESP_192_255" title="External Bus Read Latency" name="192-255 cycles" description="The number of data beats returned 192-255 cycles after the read request." units="beats"/>
+        <event offset="40" counter="L2_EXT_RRESP_256_319" title="External Bus Read Latency" name="256-319 cycles" description="The number of data beats returned 256-319 cycles after the read request." units="beats"/>
+        <event offset="41" counter="L2_EXT_RRESP_320_383" title="External Bus Read Latency" name="320-383 cycles" description="The number of data beats returned 320-383 cycles after the read request." units="beats"/>
+        <event offset="42" counter="L2_EXT_WRITE" title="External Bus Accesses" name="Write transaction" description="The number of external write transactions." units="transactions"/>
+        <event offset="43" advanced="yes" counter="L2_EXT_WRITE_NOSNP_FULL" title="External Bus Accesses" name="WriteNoSnoopFull transactions" description="The number of external non-coherent full write transactions." units="transactions"/>
+        <event offset="44" advanced="yes" counter="L2_EXT_WRITE_NOSNP_PTL" title="External Bus Accesses" name="WriteNoSnoopPartial transactions" description="The number of external non-coherent partial write transactions." units="transactions"/>
+        <event offset="45" advanced="yes" counter="L2_EXT_WRITE_SNP_FULL" title="External Bus Accesses" name="WriteSnoopFull transactions" description="The number of external coherent full write transactions." units="transactions"/>
+        <event offset="46" advanced="yes" counter="L2_EXT_WRITE_SNP_PTL" title="External Bus Accesses" name="WriteSnoopPartial transactions" description="The number of external coherent partial write transactions." units="transactions"/>
+        <event offset="47" counter="L2_EXT_WRITE_BEATS" title="External Bus Beats" name="Write beat" description="The number of external bus data write cycles." units="beats"/>
+        <event offset="48" counter="L2_EXT_W_STALL" title="External Bus Stalls" name="Write stall cycles" description="The number of cycles where a write is stalled waiting for the external bus." units="cycles"/>
+        <event offset="49" counter="L2_EXT_AW_CNT_Q1" title="External Bus Outstanding Writes" name="0-25% outstanding" description="The number of write transactions initiated when 0-25% of the maximum are in use." units="transactions"/>
+        <event offset="50" counter="L2_EXT_AW_CNT_Q2" title="External Bus Outstanding Writes" name="25-50% outstanding" description="The number of write transactions initiated when 25-50% of the maximum are in use." units="transactions"/>
+        <event offset="51" counter="L2_EXT_AW_CNT_Q3" title="External Bus Outstanding Writes" name="50-75% outstanding" description="The number of write transactions initiated when 50-75% of the maximum are in use." units="transactions"/>
+        <event offset="52" advanced="yes" counter="L2_EXT_SNOOP" title="External Bus Accesses" name="Snoop transactions" description="The number of coherency snoops triggered by external masters." units="transactions"/>
+        <event offset="53" advanced="yes" counter="L2_EXT_SNOOP_STALL" title="External Bus Stalls" name="Snoop stall cycles" description="The number of cycles where a coherency snoop triggered by external master is stalled." units="cycles"/>
+    </category>
+    <category name="Shader Core" per_cpu="no">
+        <event offset="4" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment active" description="The number of cycles where the shader core is processing a fragment workload." units="cycles"/>
+        <event offset="5" advanced="yes" counter="FRAG_PRIMITIVES" title="Core Primitives" name="Read primitives" description="The number of primitives read from the tile list by the fragment front-end." units="primitives"/>
+        <event offset="6" counter="FRAG_PRIM_RAST" title="Core Primitives" name="Rasterized primitives" description="The number of primitives being rasterized." units="primitives"/>
+        <event offset="7" counter="FRAG_FPK_ACTIVE" title="Core Cycles" name="Fragment FPKB active" description="The number of cycles where at least one quad is present in the pre-pipe quad queue." units="cycles"/>
+        <event offset="9" counter="FRAG_WARPS" title="Core Warps" name="Fragment warps" description="The number of fragment warps created." units="warps"/>
+        <event offset="10" counter="FRAG_PARTIAL_WARPS" title="Core Warps" name="Partial fragment warps" description="The number of fragment warps containing helper threads that do not correspond to a hit sample point." units="warps"/>
+        <event offset="11" counter="FRAG_QUADS_RAST" title="Core Quads" name="Rasterized quads" description="The number of quads generated by the rasterization phase." units="quads"/>
+        <event offset="12" counter="FRAG_QUADS_EZS_TEST" title="Core Quads" name="Early ZS tested quads" description="The number of quads that are undergoing early depth and stencil testing." units="quads"/>
+        <event offset="13" counter="FRAG_QUADS_EZS_UPDATE" title="Core Quads" name="Early ZS updated quads" description="The number of quads undergoing early depth and stencil testing, that are capable of updating the framebuffer." units="quads"/>
+        <event offset="14" counter="FRAG_QUADS_EZS_KILL" title="Core Quads" name="Early ZS killed quads" description="The number of quads killed by early depth and stencil testing." units="quads"/>
+        <event offset="15" counter="FRAG_LZS_TEST" title="Core Quads" name="Late ZS tested quads" description="The number of quads undergoing late depth and stencil testing." units="quads"/>
+        <event offset="16" counter="FRAG_LZS_KILL" title="Core Quads" name="Late ZS killed quads" description="The number of quads killed by late depth and stencil testing." units="quads"/>
+        <event offset="17" counter="WARP_REG_SIZE_64" title="Core Warps" name="All register warps" description="The number of warps that require more than 32 registers." units="warps"/>
+        <event offset="18" counter="FRAG_PTILES" title="Core Tiles" name="Tiles" description="The number of tiles processed by the shader core." units="tiles"/>
+        <event offset="19" counter="FRAG_TRANS_ELIM" title="Core Tiles" name="Constant tiles killed" description="The number of tiles killed by transaction elimination." units="tiles"/>
+        <event offset="20" counter="QUAD_FPK_KILLER" title="Core Quads" name="FPK occluder quads" description="The number of quads that are valid occluders for hidden surface removal." units="quads"/>
+        <event offset="21" counter="FULL_QUAD_WARPS" title="Core Warps" name="Full quad warps" description="The number of warps that are fully populated with quads." units="warps"/>
+        <event offset="22" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Non-fragment active" description="The number of cycles where the shader core is processing some non-fragment workload." units="cycles"/>
+        <event offset="23" advanced="yes" counter="COMPUTE_TASKS" title="Core Tasks" name="Non-fragment tasks" description="The number of non-fragment tasks issued to the shader core." units="tasks"/>
+        <event offset="24" counter="COMPUTE_WARPS" title="Core Warps" name="Non-fragment warps" description="The number of non-fragment warps created." units="warps"/>
+        <event offset="25" advanced="yes" counter="COMPUTE_STARVING" title="Core Starvation Cycles" name="Non-fragment starvation cycles" description="The number of cycles where the shader core is processing a non-fragment workload and there are no new threads available for execution." units="cycles"/>
+        <event offset="26" counter="EXEC_CORE_ACTIVE" title="Core Cycles" name="Execution core active" description="The number of cycles where the shader core is processing at least one warp." units="cycles"/>
+        <event offset="27" advanced="yes" counter="EXEC_ACTIVE" title="Core Cycles" name="Execution engine active" description="The number of cycles where the execution engine unit is processing at least one thread." units="cycles"/>
+        <event offset="28" counter="EXEC_INSTR_COUNT" title="Core EE Instructions" name="Executed instructions" description="The number of instructions executed per warp." units="instructions"/>
+        <event offset="29" counter="EXEC_INSTR_DIVERGED" title="Core EE Instructions" name="Diverged instructions" description="The number of instructions executed per warp, that have control flow divergence." units="instructions"/>
+        <event offset="30" advanced="yes" counter="EXEC_INSTR_STARVING" title="Core Starvation Cycles" name="Execution engine starvation cycles" description="The number of cycles where no new threads are available for execution." units="cycles"/>
+        <event offset="31" advanced="yes" counter="ARITH_INSTR_SINGLE_FMA" title="Core EE Instructions" name="Arithmetic instructions" description="The number of instructions where the workload is a single FMA pipe arithmetic operation." units="instructions"/>
+        <event offset="32" advanced="yes" counter="ARITH_INSTR_DOUBLE" title="Core EE Instructions" name="Dual Arithmetic instructions" description="The number of instructions where the workload is one FMA pipe arithmetic operation and one ADD pipe arithmetic operation." units="instructions"/>
+        <event offset="33" advanced="yes" counter="ARITH_INSTR_MSG" title="Core EE Instructions" name="Arithmetic + Message instructions" description="The number of instructions where the workload is one FMA pipe arithmetic operation and one ADD pipe message operation" units="instructions"/>
+        <event offset="34" advanced="yes" counter="ARITH_INSTR_MSG_ONLY" title="Core EE Instructions" name="Message instructions" description="The number of instructions where the workload is a single ADD pipe message operation, with no FMA pipe operation" units="instructions"/>
+        <event offset="35" counter="TEX_MSGI_NUM_QUADS" title="Core Texture Quads" name="Texture requests" description="The number of quad-width texture operations processed by the texture unit." units="quads"/>
+        <event offset="36" counter="TEX_DFCH_NUM_PASSES" title="Core Texture Quads" name="Texture issues" description="The number of quad-width filtering passes." units="issues"/>
+        <event offset="37" counter="TEX_DFCH_NUM_PASSES_MISS" title="Core Texture Quads" name="Descriptor misses" description="The number of quad-width filtering passes that miss in the resource or sampler descriptor cache." units="requests"/>
+        <event offset="38" counter="TEX_DFCH_NUM_PASSES_MIP_MAP" title="Core Texture Quads" name="Mipmapped texture issues" description="The number of quad-width filtering passes that use a mipmapped texture." units="issues"/>
+        <event offset="39" counter="TEX_TIDX_NUM_SPLIT_MIP_MAP" title="Core Texture Quads" name="Trilinear filtered issues" description="The number of quad-width filtering passes that use a trilinear filter." units="issues"/>
+        <event offset="40" counter="TEX_TFCH_NUM_LINES_FETCHED" title="Core Texture Line Fetches" name="Line fetches" description="The number of texture line fetches from the L2 cache." units="issues"/>
+        <event offset="41" counter="TEX_TFCH_NUM_LINES_FETCHED_BLOCK_COMPRESSED" title="Core Texture Line Fetches" name="Compressed line fetches" description="The number of texture line fetches from the L2 cache that are block compressed textures." units="issues"/>
+        <event offset="42" counter="TEX_TFCH_NUM_OPERATIONS" title="Core Texture Cycles" name="Cache lookups" description="The number of texture cache lookup cycles." units="requests"/>
+        <event offset="43" counter="TEX_FILT_NUM_OPERATIONS" title="Core Texture Cycles" name="Texturing active" description="The number of texture filtering issue cycles." units="cycles"/>
+        <event offset="44" counter="LS_MEM_READ_FULL" title="Core Load/Store Cycles" name="Full read cycles" description="The number of full-width load/store cache reads." units="cycles"/>
+        <event offset="45" counter="LS_MEM_READ_SHORT" title="Core Load/Store Cycles" name="Partial read cycles" description="The number of partial-width load/store cache reads." units="cycles"/>
+        <event offset="46" counter="LS_MEM_WRITE_FULL" title="Core Load/Store Cycles" name="Full write cycles" description="The number of full-width load/store cache writes." units="cycles"/>
+        <event offset="47" counter="LS_MEM_WRITE_SHORT" title="Core Load/Store Cycles" name="Partial write cycles" description="The number of partial-width load/store cache writes." units="cycles"/>
+        <event offset="48" counter="LS_MEM_ATOMIC" title="Core Load/Store Cycles" name="Atomic access cycles" description="The number of load/store atomic accesses." units="cycles"/>
+        <event offset="49" counter="VARY_INSTR" title="Core Varying Requests" name="Interpolation requests" description="The number of warp-width interpolation operations processed by the varying unit." units="instructions"/>
+        <event offset="50" counter="VARY_SLOT_32" title="Core Varying Cycles" name="32-bit interpolation active" description="The number of 32-bit interpolation cycles processed by the varying unit." units="cycles"/>
+        <event offset="51" counter="VARY_SLOT_16" title="Core Varying Cycles" name="16-bit interpolation active" description="The number of 16-bit interpolation cycles processed by the varying unit." units="cycles"/>
+        <event offset="52" advanced="yes" counter="ATTR_INSTR" title="Core Attribute Requests" name="Attribute requests" description="The number of instructions executed by the attribute unit." units="instructions"/>
+        <event offset="53" advanced="yes" counter="ARITH_INSTR_FP_MUL" title="Core EE Instructions" name="Multiplier instructions" description="The number of instructions where the workload uses floating-point multiplier hardware." units="instructions"/>
+        <event offset="54" counter="BEATS_RD_FTC" title="Core L2 Reads" name="Fragment L2 read beats" description="The number of read beats received by the fixed-function fragment front-end." units="beats"/>
+        <event offset="55" counter="BEATS_RD_FTC_EXT" title="Core L2 Reads" name="Fragment external read beats" description="The number of read beats received by the fixed-function fragment front-end that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="56" counter="BEATS_RD_LSC" title="Core L2 Reads" name="Load/store L2 read beats" description="The number of read beats received by the load/store unit." units="beats"/>
+        <event offset="57" counter="BEATS_RD_LSC_EXT" title="Core L2 Reads" name="Load/store external read beats" description="The number of read beats received by the load/store unit that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="58" counter="BEATS_RD_TEX" title="Core L2 Reads" name="Texture L2 read beats" description="The number of read beats received by the texture unit." units="beats"/>
+        <event offset="59" counter="BEATS_RD_TEX_EXT" title="Core L2 Reads" name="Texture external read beats" description="The number of read beats received by the texture unit that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="60" advanced="yes" counter="BEATS_RD_OTHER" title="Core L2 Reads" name="Other L2 read beats" description="The number of read beats received by a unit that is not specifically identified." units="beats"/>
+        <event offset="61" counter="BEATS_WR_LSC_OTHER" title="Core Writes" name="Load/store other write beats" description="The number of write beats by the load/store unit that are due to any reason other than writeback." units="beats"/>
+        <event offset="62" counter="BEATS_WR_TIB" title="Core Writes" name="Tile buffer write beats" description="The number of write beats sent by the tile buffer writeback unit." units="beats"/>
+        <event offset="63" counter="BEATS_WR_LSC_WB" title="Core Writes" name="Load/store writeback write beats" description="The number of write beats by the load/store unit that are due to writeback." units="beats"/>
     </category>
 </metrics>
\ No newline at end of file
diff --git a/src/panfrost/perf/G77.xml b/src/panfrost/perf/G77.xml
index 60d99193b87..26ddfc07bb1 100644
--- a/src/panfrost/perf/G77.xml
+++ b/src/panfrost/perf/G77.xml
@@ -1,6 +1,6 @@
 <!--
 Copyright © 2017-2020 ARM Limited.
-Copyright © 2021 Collabora, Ltd.
+Copyright © 2021-2022 Collabora, Ltd.
 Author: Antonio Caggiano <antonio.caggiano@collabora.com>
 
 Permission is hereby granted, free of charge, to any person obtaining a
@@ -24,156 +24,156 @@ SOFTWARE.
 -->
 <metrics id="TTRx">
     <category name="Job Manager" per_cpu="no">
-        <event offset="6" counter="GPU_ACTIVE" title="GPU Cycles" name="GPU active" description="The number of cycles where the GPU has a workload of any type queued for processing." units="cycles" />
-        <event offset="7" advanced="yes" counter="IRQ_ACTIVE" title="GPU Cycles" name="Interrupt active" description="The number of cycles where the GPU has a pending interrupt." units="cycles" />
-        <event offset="8" advanced="yes" counter="JS0_JOBS" title="GPU Jobs" name="Fragment jobs" description="The number of jobs processed by the GPU fragment queue." units="jobs" />
-        <event offset="9" counter="JS0_TASKS" title="GPU Tasks" name="Fragment tasks" description="The number of 32x32 pixel tasks processed by the GPU fragment queue." units="tasks" />
-        <event offset="10" counter="JS0_ACTIVE" title="GPU Cycles" name="Fragment queue active" description="The number of cycles where work is queued for processing in the GPU fragment queue." units="cycles" />
-        <event offset="12" advanced="yes" counter="JS0_WAIT_READ" title="GPU Wait Cycles" name="Fragment descriptor reads cycles" description="The number of cycles where queued fragment work is waiting for a descriptor load." units="cycles" />
-        <event offset="13" advanced="yes" counter="JS0_WAIT_ISSUE" title="GPU Wait Cycles" name="Fragment job issue cycles" description="The number of cycles where queued fragment work is waiting for an available processor." units="cycles" />
-        <event offset="14" advanced="yes" counter="JS0_WAIT_DEPEND" title="GPU Wait Cycles" name="Fragment job dependency cycles" description="The number of cycles where queued fragment work is waiting for dependent work to complete." units="cycles" />
-        <event offset="15" advanced="yes" counter="JS0_WAIT_FINISH" title="GPU Wait Cycles" name="Fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued fragment work to complete." units="cycles" />
-        <event offset="16" advanced="yes" counter="JS1_JOBS" title="GPU Jobs" name="Non-fragment jobs" description="The number of jobs processed by the GPU non-fragment queue." units="jobs" />
-        <event offset="17" advanced="yes" counter="JS1_TASKS" title="GPU Tasks" name="Non-fragment tasks" description="The number of tasks processed by the GPU non-fragment queue." units="tasks" />
-        <event offset="18" counter="JS1_ACTIVE" title="GPU Cycles" name="Non-fragment queue active" description="The number of cycles where work is queued in the GPU non-fragment queue." units="cycles" />
-        <event offset="20" advanced="yes" counter="JS1_WAIT_READ" title="GPU Wait Cycles" name="Non-fragment descriptor read cycles" description="The number number of cycles where queued non-fragment work is waiting for a descriptor load." units="cycles" />
-        <event offset="21" advanced="yes" counter="JS1_WAIT_ISSUE" title="GPU Wait Cycles" name="Non-fragment job issue cycles" description="The number of cycles where queued non-fragment work is waiting for an available processor." units="cycles" />
-        <event offset="22" advanced="yes" counter="JS1_WAIT_DEPEND" title="GPU Wait Cycles" name="Non-fragment job dependency cycles" description="The number of cycles where queued non-fragment work is waiting for dependent work to complete." units="cycles" />
-        <event offset="23" advanced="yes" counter="JS1_WAIT_FINISH" title="GPU Wait Cycles" name="Non-fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued non-fragment work to complete." units="cycles" />
-        <event offset="24" advanced="yes" counter="JS2_JOBS" title="GPU Jobs" name="Reserved jobs" description="The number of jobs processed by the GPU reserved queue." units="jobs" />
-        <event offset="25" advanced="yes" counter="JS2_TASKS" title="GPU Tasks" name="Reserved tasks" description="The number of tasks processed by the GPU reserved queue." units="tasks" />
-        <event offset="26" advanced="yes" counter="JS2_ACTIVE" title="GPU Cycles" name="Reserved queue active" description="The number of cycles where work is queued in the GPU reserved queue." units="cycles" />
-        <event offset="28" advanced="yes" counter="JS2_WAIT_READ" title="GPU Wait Cycles" name="Reserved descriptor read cycles" description="The number of cycles where queued reserved work is waiting for a descriptor load." units="cycles" />
-        <event offset="29" advanced="yes" counter="JS2_WAIT_ISSUE" title="GPU Wait Cycles" name="Reserved job issue cycles" description="The number of cycles where queued reserved work is waiting for an available processor." units="cycles" />
-        <event offset="30" advanced="yes" counter="JS2_WAIT_DEPEND" title="GPU Wait Cycles" name="Reserved job dependency cycles" description="The number of cycles where queued reserved work is waiting for dependent work to complete." units="cycles" />
-        <event offset="31" advanced="yes" counter="JS2_WAIT_FINISH" title="GPU Wait Cycles" name="Reserved job finish cycles" description="The number of cycles where the GPU is waiting for issued reserved work to complete." units="cycles" />
-    </category>
-    <category name="Memory System" per_cpu="no">
-        <event offset="196" advanced="yes" counter="MMU_REQUESTS" title="MMU Stage 1 Translations" name="MMU lookups" description="The number of main MMU address translations performed." units="requests" />
-        <event offset="208" advanced="yes" counter="L2_RD_MSG_IN" title="L2 Cache Requests" name="Read requests" description="The number of L2 cache read requests from internal masters." units="requests" />
-        <event offset="209" advanced="yes" counter="L2_RD_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Read stall cycles" description="The number of cycles L2 cache read requests from internal masters are stalled." units="cycles" />
-        <event offset="210" advanced="yes" counter="L2_WR_MSG_IN" title="L2 Cache Requests" name="Write requests" description="The number of L2 cache write requests from internal masters." units="requests" />
-        <event offset="211" advanced="yes" counter="L2_WR_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Write stall cycles" description="The number of cycles where L2 cache write requests from internal masters are stalled." units="cycles" />
-        <event offset="212" advanced="yes" counter="L2_SNP_MSG_IN" title="L2 Cache Requests" name="Snoop requests" description="The number of L2 snoop requests from internal masters." units="requests" />
-        <event offset="213" advanced="yes" counter="L2_SNP_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Snoop stall cycles" description="The number of cycles where L2 cache snoop requests from internal masters are stalled." units="cycles" />
-        <event offset="214" advanced="yes" counter="L2_RD_MSG_OUT" title="L2 Cache Requests" name="L1 read requests" description="The number of L1 cache read requests sent by the L2 cache to an internal master." units="requests" />
-        <event offset="215" advanced="yes" counter="L2_RD_MSG_OUT_STALL" title="L2 Cache Stall Cycles" name="L1 read stall cycles" description="The number of cycles where L1 cache read requests sent by the L2 cache to an internal master are stalled." units="cycles" />
-        <event offset="216" advanced="yes" counter="L2_WR_MSG_OUT" title="L2 Cache Requests" name="L1 write requests" description="The number of L1 cache write responses sent by the L2 cache to an internal master." units="requests" />
-        <event offset="217" counter="L2_ANY_LOOKUP" title="L2 Cache Lookups" name="Any lookup" description="The number of L2 cache lookups performed." units="requests" />
-        <event offset="218" counter="L2_READ_LOOKUP" title="L2 Cache Lookups" name="Read lookup" description="The number of L2 cache read lookups performed." units="requests" />
-        <event offset="219" counter="L2_WRITE_LOOKUP" title="L2 Cache Lookups" name="Write lookup" description="The number of L2 cache write lookups performed." units="requests" />
-        <event offset="220" advanced="yes" counter="L2_EXT_SNOOP_LOOKUP" title="L2 Cache Lookups" name="External snoop lookups" description="The number of coherency snoop lookups performed that were triggered by an external master." units="requests" />
-        <event offset="221" counter="L2_EXT_READ" title="External Bus Accesses" name="Read transaction" description="The number of external read transactions." units="transactions" />
-        <event offset="222" advanced="yes" counter="L2_EXT_READ_NOSNP" title="External Bus Accesses" name="ReadNoSnoop transactions" description="The number of external non-coherent read transactions." units="transactions" />
-        <event offset="223" advanced="yes" counter="L2_EXT_READ_UNIQUE" title="External Bus Accesses" name="ReadUnique transactions" description="The number of external coherent read unique transactions." units="transactions" />
-        <event offset="224" counter="L2_EXT_READ_BEATS" title="External Bus Beats" name="Read beat" description="The number of external bus data read cycles." units="beats" />
-        <event offset="225" counter="L2_EXT_AR_STALL" title="External Bus Stalls" name="Read stall cycles" description="The number of cycles where a read is stalled waiting for the external bus." units="cycles" />
-        <event offset="226" counter="L2_EXT_AR_CNT_Q1" title="External Bus Outstanding Reads" name="0-25% outstanding" description="The number of read transactions initiated when 0-25% of the maximum are in use." units="transactions" />
-        <event offset="227" counter="L2_EXT_AR_CNT_Q2" title="External Bus Outstanding Reads" name="25-50% outstanding" description="The number of read transactions initiated when 25-50% of the maximum are in use." units="transactions" />
-        <event offset="228" counter="L2_EXT_AR_CNT_Q3" title="External Bus Outstanding Reads" name="50-75% outstanding" description="The number of read transactions initiated when 50-75% of the maximum are in use." units="transactions" />
-        <event offset="229" counter="L2_EXT_RRESP_0_127" title="External Bus Read Latency" name="0-127 cycles" description="The number of data beats returned 0-127 cycles after the read request." units="beats" />
-        <event offset="230" counter="L2_EXT_RRESP_128_191" title="External Bus Read Latency" name="128-191 cycles" description="The number of data beats returned 128-191 cycles after the read request." units="beats" />
-        <event offset="231" counter="L2_EXT_RRESP_192_255" title="External Bus Read Latency" name="192-255 cycles" description="The number of data beats returned 192-255 cycles after the read request." units="beats" />
-        <event offset="232" counter="L2_EXT_RRESP_256_319" title="External Bus Read Latency" name="256-319 cycles" description="The number of data beats returned 256-319 cycles after the read request." units="beats" />
-        <event offset="233" counter="L2_EXT_RRESP_320_383" title="External Bus Read Latency" name="320-383 cycles" description="The number of data beats returned 320-383 cycles after the read request." units="beats" />
-        <event offset="234" counter="L2_EXT_WRITE" title="External Bus Accesses" name="Write transaction" description="The number of external write transactions." units="transactions" />
-        <event offset="235" advanced="yes" counter="L2_EXT_WRITE_NOSNP_FULL" title="External Bus Accesses" name="WriteNoSnoopFull transactions" description="The number of external non-coherent full write transactions." units="transactions" />
-        <event offset="236" advanced="yes" counter="L2_EXT_WRITE_NOSNP_PTL" title="External Bus Accesses" name="WriteNoSnoopPartial transactions" description="The number of external non-coherent partial write transactions." units="transactions" />
-        <event offset="237" advanced="yes" counter="L2_EXT_WRITE_SNP_FULL" title="External Bus Accesses" name="WriteSnoopFull transactions" description="The number of external coherent full write transactions." units="transactions" />
-        <event offset="238" advanced="yes" counter="L2_EXT_WRITE_SNP_PTL" title="External Bus Accesses" name="WriteSnoopPartial transactions" description="The number of external coherent partial write transactions." units="transactions" />
-        <event offset="239" counter="L2_EXT_WRITE_BEATS" title="External Bus Beats" name="Write beat" description="The number of external bus data write cycles." units="beats" />
-        <event offset="240" counter="L2_EXT_W_STALL" title="External Bus Stalls" name="Write stall cycles" description="The number of cycles where a write is stalled waiting for the external bus." units="cycles" />
-        <event offset="241" counter="L2_EXT_AW_CNT_Q1" title="External Bus Outstanding Writes" name="0-25% outstanding" description="The number of write transactions initiated when 0-25% of the maximum are in use." units="transactions" />
-        <event offset="242" counter="L2_EXT_AW_CNT_Q2" title="External Bus Outstanding Writes" name="25-50% outstanding" description="The number of write transactions initiated when 25-50% of the maximum are in use." units="transactions" />
-        <event offset="243" counter="L2_EXT_AW_CNT_Q3" title="External Bus Outstanding Writes" name="50-75% outstanding" description="The number of write transactions initiated when 50-75% of the maximum are in use." units="transactions" />
-        <event offset="244" advanced="yes" counter="L2_EXT_SNOOP" title="External Bus Accesses" name="Snoop transactions" description="The number of coherency snoops triggered by external masters." units="transactions" />
-        <event offset="245" advanced="yes" counter="L2_EXT_SNOOP_STALL" title="External Bus Stalls" name="Snoop stall cycles" description="The number of cycles where a coherency snoop triggered by external master is stalled." units="cycles" />
-    </category>
-    <category name="Shader Core" per_cpu="no">
-        <event offset="196" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment active" description="The number of cycles where the shader core is processing a fragment workload." units="cycles" />
-        <event offset="197" advanced="yes" counter="FRAG_PRIMITIVES_OUT" title="Core Primitives" name="Read primitives" description="The number of primitives read from the tile list by the fragment front-end." units="primitives" />
-        <event offset="198" counter="FRAG_PRIM_RAST" title="Core Primitives" name="Rasterized primitives" description="The number of primitives being rasterized." units="primitives" />
-        <event offset="199" counter="FRAG_FPK_ACTIVE" title="Core Cycles" name="Fragment FPKB active" description="The number of cycles where at least one quad is present in the pre-pipe quad queue." units="cycles" />
-        <event offset="201" counter="FRAG_WARPS" title="Core Warps" name="Fragment warps" description="The number of fragment warps created." units="warps" />
-        <event offset="202" counter="FRAG_PARTIAL_QUADS_RAST" title="Core Quads" name="Partial rasterized quads" description="The number of partially-rasterized fragment quads created." units="quads" />
-        <event offset="203" counter="FRAG_QUADS_RAST" title="Core Quads" name="Rasterized quads" description="The number of quads generated by the rasterization phase." units="quads" />
-        <event offset="204" counter="FRAG_QUADS_EZS_TEST" title="Core Quads" name="Early ZS tested quads" description="The number of quads that are undergoing early depth and stencil testing." units="quads" />
-        <event offset="205" counter="FRAG_QUADS_EZS_UPDATE" title="Core Quads" name="Early ZS updated quads" description="The number of quads undergoing early depth and stencil testing, that are capable of updating the framebuffer." units="quads" />
-        <event offset="206" counter="FRAG_QUADS_EZS_KILL" title="Core Quads" name="Early ZS killed quads" description="The number of quads killed by early depth and stencil testing." units="quads" />
-        <event offset="207" counter="FRAG_LZS_TEST" title="Core Quads" name="Late ZS tested quads" description="The number of quads undergoing late depth and stencil testing." units="quads" />
-        <event offset="208" counter="FRAG_LZS_KILL" title="Core Quads" name="Late ZS killed quads" description="The number of quads killed by late depth and stencil testing." units="quads" />
-        <event offset="209" counter="WARP_REG_SIZE_64" title="Core Warps" name="All register warps" description="The number of warps that require more than 32 registers." units="warps" />
-        <event offset="210" counter="FRAG_PTILES" title="Core Tiles" name="Tiles" description="The number of tiles processed by the shader core." units="tiles" />
-        <event offset="211" counter="FRAG_TRANS_ELIM" title="Core Tiles" name="Constant tiles killed" description="The number of tiles killed by transaction elimination." units="tiles" />
-        <event offset="212" counter="QUAD_FPK_KILLER" title="Core Quads" name="FPK occluder quads" description="The number of quads that are valid occluders for hidden surface removal." units="quads" />
-        <event offset="213" counter="FULL_QUAD_WARPS" title="Core Warps" name="Full quad warps" description="The number of warps that are fully populated with quads." units="warps" />
-        <event offset="214" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Non-fragment active" description="The number of cycles where the shader core is processing some non-fragment workload." units="cycles" />
-        <event offset="215" advanced="yes" counter="COMPUTE_TASKS" title="Core Tasks" name="Non-fragment tasks" description="The number of non-fragment tasks issued to the shader core." units="tasks" />
-        <event offset="216" counter="COMPUTE_WARPS" title="Core Warps" name="Non-fragment warps" description="The number of non-fragment warps created." units="warps" />
-        <event offset="217" advanced="yes" counter="COMPUTE_STARVING" title="Core Starvation Cycles" name="Non-fragment starvation cycles" description="The number of cycles where the shader core is processing a non-fragment workload and there are no new threads available for execution." units="cycles" />
-        <event offset="218" counter="EXEC_CORE_ACTIVE" title="Core Cycles" name="Execution core active" description="The number of cycles where the shader core is processing at least one warp." units="cycles" />
-        <event offset="219" counter="EXEC_INSTR_FMA" title="Core PU Instructions" name="FMA instructions" description="The number of instructions issued to the FMA pipe." units="instructions" />
-        <event offset="220" counter="EXEC_INSTR_CVT" title="Core PU Instructions" name="CVT instructions" description="The number of instructions issued to the CVT pipe." units="instructions" />
-        <event offset="221" counter="EXEC_INSTR_SFU" title="Core PU Instructions" name="SFU instructions" description="The number of instructions issued to the SFU pipe." units="instructions" />
-        <event offset="222" counter="EXEC_INSTR_MSG" title="Core PU Instructions" name="Message instructions" description="The number of instructions issued to the MSG pipe." units="instructions" />
-        <event offset="223" counter="EXEC_INSTR_DIVERGED" title="Core EE Instructions" name="Diverged instructions" description="The number of instructions executed per warp, that have control flow divergence." units="instructions" />
-        <event offset="224" advanced="yes" counter="EXEC_ICACHE_MISS" title="Core PU Instructions" name="Instruction cache misses" description="The number of instruction cache misses." units="requests" />
-        <event offset="225" advanced="yes" counter="EXEC_STARVE_ARITH" title="Core Starvation Cycles" name="Execution engine starvation cycles" description="The number of cycles where the processing unit is starved of work." units="cycles" />
-        <event offset="226" counter="CALL_BLEND_SHADER" title="Core PU Instructions" name="Blend shader calls" description="The number of blend shader invocations executed." units="instructions" />
-        <event offset="227" counter="TEX_MSGI_NUM_FLITS" title="Texture Bus" name="Input beats" description="The number of texture request message data beats." units="beats" />
-        <event offset="228" counter="TEX_DFCH_CLK_STALLED" title="Core Texture Stalls" name="Descriptor stall cycles" description="The number of cycles where a quad is stalled on texture descriptor fetch." units="cycles" />
-        <event offset="229" counter="TEX_TFCH_CLK_STALLED" title="Core Texture Stalls" name="Fetch queue stall cycles" description="The number of cycles where a quad is stalled on entering texture fetch because the fetch queue is full." units="cycles" />
-        <event offset="230" counter="TEX_TFCH_STARVED_PENDING_DATA_FETCH" title="Core Texture Stalls" name="Filtering unit stall cycles" description="The number of cycles where the filtering unit is idle and there is at least one quad present in the texture data fetch queue." units="cycles" />
-        <event offset="231" counter="TEX_FILT_NUM_OPERATIONS" title="Core Texture Cycles" name="Texturing active" description="The number of texture filtering issue cycles." units="cycles" />
-        <event offset="232" counter="TEX_FILT_NUM_FXR_OPERATIONS" title="Core Texture Cycles" name="4x bilinear filtering active" description="The number of cycles where the filtering unit uses the 4x path to implement nearest or bilinear filtering." units="cycles" />
-        <event offset="233" counter="TEX_FILT_NUM_FST_OPERATIONS" title="Core Texture Cycles" name="2x trilinear filtering active" description="The number of cycles where the filtering unit uses the 4x path to implement trilinear filtering." units="cycles" />
-        <event offset="234" counter="TEX_MSGO_NUM_MSG" title="Core Texture Quads" name="Texture requests" description="The number of quad-width texture operations processed by the texture unit." units="quads" />
-        <event offset="235" counter="TEX_MSGO_NUM_FLITS" title="Texture Bus" name="Output beats" description="The number of texture response message data beats." units="beats" />
-        <event offset="236" counter="LS_MEM_READ_FULL" title="Core Load/Store Cycles" name="Full read cycles" description="The number of full-width load/store cache reads." units="cycles" />
-        <event offset="237" counter="LS_MEM_READ_SHORT" title="Core Load/Store Cycles" name="Partial read cycles" description="The number of partial-width load/store cache reads." units="cycles" />
-        <event offset="238" counter="LS_MEM_WRITE_FULL" title="Core Load/Store Cycles" name="Full write cycles" description="The number of full-width load/store cache writes." units="cycles" />
-        <event offset="239" counter="LS_MEM_WRITE_SHORT" title="Core Load/Store Cycles" name="Partial write cycles" description="The number of partial-width load/store cache writes." units="cycles" />
-        <event offset="240" counter="LS_MEM_ATOMIC" title="Core Load/Store Cycles" name="Atomic access cycles" description="The number of load/store atomic accesses." units="cycles" />
-        <event offset="241" counter="VARY_INSTR" title="Core Varying Requests" name="Interpolation requests" description="The number of warp-width interpolation operations processed by the varying unit." units="instructions" />
-        <event offset="242" counter="VARY_SLOT_32" title="Core Varying Cycles" name="32-bit interpolation active" description="The number of 32-bit interpolation cycles processed by the varying unit." units="cycles" />
-        <event offset="243" counter="VARY_SLOT_16" title="Core Varying Cycles" name="16-bit interpolation active" description="The number of 16-bit interpolation cycles processed by the varying unit." units="cycles" />
-        <event offset="244" advanced="yes" counter="ATTR_INSTR" title="Core Attribute Requests" name="Attribute requests" description="The number of instructions executed by the attribute unit." units="instructions" />
-        <event offset="245" advanced="yes" counter="ARITH_INSTR_FP_MUL" title="Core EE Instructions" name="Multiplier instructions" description="The number of instructions where the workload uses floating-point multiplier hardware." units="instructions" />
-        <event offset="246" counter="BEATS_RD_FTC" title="Core L2 Reads" name="Fragment L2 read beats" description="The number of read beats received by the fixed-function fragment front-end." units="beats" />
-        <event offset="247" counter="BEATS_RD_FTC_EXT" title="Core L2 Reads" name="Fragment external read beats" description="The number of read beats received by the fixed-function fragment front-end that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="248" counter="BEATS_RD_LSC" title="Core L2 Reads" name="Load/store L2 read beats" description="The number of read beats received by the load/store unit." units="beats" />
-        <event offset="249" counter="BEATS_RD_LSC_EXT" title="Core L2 Reads" name="Load/store external read beats" description="The number of read beats received by the load/store unit that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="250" counter="BEATS_RD_TEX" title="Core L2 Reads" name="Texture L2 read beats" description="The number of read beats received by the texture unit." units="beats" />
-        <event offset="251" counter="BEATS_RD_TEX_EXT" title="Core L2 Reads" name="Texture external read beats" description="The number of read beats received by the texture unit that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="252" advanced="yes" counter="BEATS_RD_OTHER" title="Core L2 Reads" name="Other L2 read beats" description="The number of read beats received by a unit that is not specifically identified." units="beats" />
-        <event offset="253" counter="BEATS_WR_LSC_OTHER" title="Core Writes" name="Load/store other write beats" description="The number of write beats by the load/store unit that are due to any reason other than writeback." units="beats" />
-        <event offset="254" counter="BEATS_WR_TIB" title="Core Writes" name="Tile buffer write beats" description="The number of write beats sent by the tile buffer writeback unit." units="beats" />
-        <event offset="255" counter="BEATS_WR_LSC_WB" title="Core Writes" name="Load/store writeback write beats" description="The number of write beats by the load/store unit that are due to writeback." units="beats" />
+        <event offset="6" counter="GPU_ACTIVE" title="GPU Cycles" name="GPU active" description="The number of cycles where the GPU has a workload of any type queued for processing." units="cycles"/>
+        <event offset="7" advanced="yes" counter="IRQ_ACTIVE" title="GPU Cycles" name="Interrupt active" description="The number of cycles where the GPU has a pending interrupt." units="cycles"/>
+        <event offset="8" advanced="yes" counter="JS0_JOBS" title="GPU Jobs" name="Fragment jobs" description="The number of jobs processed by the GPU fragment queue." units="jobs"/>
+        <event offset="9" counter="JS0_TASKS" title="GPU Tasks" name="Fragment tasks" description="The number of 32x32 pixel tasks processed by the GPU fragment queue." units="tasks"/>
+        <event offset="10" counter="JS0_ACTIVE" title="GPU Cycles" name="Fragment queue active" description="The number of cycles where work is queued for processing in the GPU fragment queue." units="cycles"/>
+        <event offset="12" advanced="yes" counter="JS0_WAIT_READ" title="GPU Wait Cycles" name="Fragment descriptor reads cycles" description="The number of cycles where queued fragment work is waiting for a descriptor load." units="cycles"/>
+        <event offset="13" advanced="yes" counter="JS0_WAIT_ISSUE" title="GPU Wait Cycles" name="Fragment job issue cycles" description="The number of cycles where queued fragment work is waiting for an available processor." units="cycles"/>
+        <event offset="14" advanced="yes" counter="JS0_WAIT_DEPEND" title="GPU Wait Cycles" name="Fragment job dependency cycles" description="The number of cycles where queued fragment work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="15" advanced="yes" counter="JS0_WAIT_FINISH" title="GPU Wait Cycles" name="Fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued fragment work to complete." units="cycles"/>
+        <event offset="16" advanced="yes" counter="JS1_JOBS" title="GPU Jobs" name="Non-fragment jobs" description="The number of jobs processed by the GPU non-fragment queue." units="jobs"/>
+        <event offset="17" advanced="yes" counter="JS1_TASKS" title="GPU Tasks" name="Non-fragment tasks" description="The number of tasks processed by the GPU non-fragment queue." units="tasks"/>
+        <event offset="18" counter="JS1_ACTIVE" title="GPU Cycles" name="Non-fragment queue active" description="The number of cycles where work is queued in the GPU non-fragment queue." units="cycles"/>
+        <event offset="20" advanced="yes" counter="JS1_WAIT_READ" title="GPU Wait Cycles" name="Non-fragment descriptor read cycles" description="The number number of cycles where queued non-fragment work is waiting for a descriptor load." units="cycles"/>
+        <event offset="21" advanced="yes" counter="JS1_WAIT_ISSUE" title="GPU Wait Cycles" name="Non-fragment job issue cycles" description="The number of cycles where queued non-fragment work is waiting for an available processor." units="cycles"/>
+        <event offset="22" advanced="yes" counter="JS1_WAIT_DEPEND" title="GPU Wait Cycles" name="Non-fragment job dependency cycles" description="The number of cycles where queued non-fragment work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="23" advanced="yes" counter="JS1_WAIT_FINISH" title="GPU Wait Cycles" name="Non-fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued non-fragment work to complete." units="cycles"/>
+        <event offset="24" advanced="yes" counter="JS2_JOBS" title="GPU Jobs" name="Reserved jobs" description="The number of jobs processed by the GPU reserved queue." units="jobs"/>
+        <event offset="25" advanced="yes" counter="JS2_TASKS" title="GPU Tasks" name="Reserved tasks" description="The number of tasks processed by the GPU reserved queue." units="tasks"/>
+        <event offset="26" advanced="yes" counter="JS2_ACTIVE" title="GPU Cycles" name="Reserved queue active" description="The number of cycles where work is queued in the GPU reserved queue." units="cycles"/>
+        <event offset="28" advanced="yes" counter="JS2_WAIT_READ" title="GPU Wait Cycles" name="Reserved descriptor read cycles" description="The number of cycles where queued reserved work is waiting for a descriptor load." units="cycles"/>
+        <event offset="29" advanced="yes" counter="JS2_WAIT_ISSUE" title="GPU Wait Cycles" name="Reserved job issue cycles" description="The number of cycles where queued reserved work is waiting for an available processor." units="cycles"/>
+        <event offset="30" advanced="yes" counter="JS2_WAIT_DEPEND" title="GPU Wait Cycles" name="Reserved job dependency cycles" description="The number of cycles where queued reserved work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="31" advanced="yes" counter="JS2_WAIT_FINISH" title="GPU Wait Cycles" name="Reserved job finish cycles" description="The number of cycles where the GPU is waiting for issued reserved work to complete." units="cycles"/>
     </category>
     <category name="Tiler" per_cpu="no">
-        <event offset="4" counter="TILER_ACTIVE" title="GPU Cycles" name="Tiler active" description="The number of cycles where the tiler has a workload queued for processing." units="cycles" />
-        <event offset="6" counter="TRIANGLES" title="Input Primitives" name="Triangle primitives" description="The number of input triangle primitives." units="primitives" />
-        <event offset="7" counter="LINES" title="Input Primitives" name="Line primitives" description="The number of input line primitives." units="primitives" />
-        <event offset="8" counter="POINTS" title="Input Primitives" name="Point primitives" description="The number of input point primitives." units="primitives" />
-        <event offset="9" counter="FRONT_FACING" title="Visible Primitives" name="Front-facing primitives" description="The number of front-facing triangles that are visible after culling." units="primitives" />
-        <event offset="10" counter="BACK_FACING" title="Visible Primitives" name="Back-facing primitives" description="The number of back-facing triangles that are visible after culling." units="primitives" />
-        <event offset="11" counter="PRIM_VISIBLE" title="Primitive Culling" name="Visible primitives" description="The number of primitives that are visible after culling." units="primitives" />
-        <event offset="12" counter="PRIM_CULLED" title="Primitive Culling" name="Facing and XY plane test culled primitives" description="The number of primitives that are culled by facing or frustum XY plane tests." units="primitives" />
-        <event offset="13" counter="PRIM_CLIPPED" title="Primitive Culling" name="Z plane test culled primitives" description="The number of primitives that are culled by frustum Z plane tests." units="primitives" />
-        <event offset="14" counter="PRIM_SAT_CULLED" title="Primitive Culling" name="Sample test culled primitives" description="The number of primitives culled by the sample coverage test." units="primitives" />
-        <event offset="17" advanced="yes" counter="BUS_READ" title="Tiler L2 Accesses" name="Read beats" description="The number of internal bus data read cycles made by the tiler." units="beats" />
-        <event offset="19" advanced="yes" counter="BUS_WRITE" title="Tiler L2 Accesses" name="Write beats" description="The number of internal bus data write cycles made by the tiler." units="beats" />
-        <event offset="21" counter="IDVS_POS_SHAD_REQ" title="Tiler Shading Requests" name="Position shading requests" description="The number of position shading requests in the IDVS flow." units="requests" />
-        <event offset="23" advanced="yes" counter="IDVS_POS_SHAD_STALL" title="Tiler Cycles" name="Position shading stall cycles" description="The number of cycles where the tiler has a stalled position shading request." units="cycles" />
-        <event offset="24" advanced="yes" counter="IDVS_POS_FIFO_FULL" title="Tiler Cycles" name="Position FIFO full cycles" description="The number of cycles where the tiler has a stalled position shading buffer." units="cycles" />
-        <event offset="26" advanced="yes" counter="VCACHE_HIT" title="Tiler Vertex Cache" name="Position cache hits" description="The number of position lookups that result in a hit in the vertex cache." units="requests" />
-        <event offset="27" advanced="yes" counter="VCACHE_MISS" title="Tiler Vertex Cache" name="Position cache misses" description="The number of position lookups that miss in the vertex cache." units="requests" />
-        <event offset="31" advanced="yes" counter="VFETCH_STALL" title="Tiler Cycles" name="Primitive assembly busy stall cycles" description="The number of cycles where the tiler is stalled waiting for primitive assembly." units="cycles" />
-        <event offset="34" advanced="yes" counter="IDVS_VBU_HIT" title="Tiler Vertex Cache" name="Varying cache hits" description="The number of varying lookups that result in a hit in the vertex cache." units="requests" />
-        <event offset="35" advanced="yes" counter="IDVS_VBU_MISS" title="Tiler Vertex Cache" name="Varying cache misses" description="The number of varying lookups that miss in the vertex cache." units="requests" />
-        <event offset="37" counter="IDVS_VAR_SHAD_REQ" title="Tiler Shading Requests" name="Varying shading requests" description="The number of varying shading requests in the IDVS flow." units="requests" />
-        <event offset="38" advanced="yes" counter="IDVS_VAR_SHAD_STALL" title="Tiler Cycles" name="Varying shading stall cycles" description="The number of cycles where the tiler has a stalled varying shading request." units="cycles" />
-        <event offset="54" advanced="yes" counter="WRBUF_NO_AXI_ID_STALL" title="Tiler Cycles" name="Write buffer transaction stall cycles" description="The number of cycles where the tiler write buffer can not send data because it has no available write IDs." units="cycles" />
-        <event offset="55" advanced="yes" counter="WRBUF_AXI_STALL" title="Tiler Cycles" name="Write buffer write stall cycles" description="The number of cycles where the tiler write buffer can not send data because the bus is not ready." units="cycles" />
+        <event offset="4" counter="TILER_ACTIVE" title="GPU Cycles" name="Tiler active" description="The number of cycles where the tiler has a workload queued for processing." units="cycles"/>
+        <event offset="6" counter="TRIANGLES" title="Input Primitives" name="Triangle primitives" description="The number of input triangle primitives." units="primitives"/>
+        <event offset="7" counter="LINES" title="Input Primitives" name="Line primitives" description="The number of input line primitives." units="primitives"/>
+        <event offset="8" counter="POINTS" title="Input Primitives" name="Point primitives" description="The number of input point primitives." units="primitives"/>
+        <event offset="9" counter="FRONT_FACING" title="Visible Primitives" name="Front-facing primitives" description="The number of front-facing triangles that are visible after culling." units="primitives"/>
+        <event offset="10" counter="BACK_FACING" title="Visible Primitives" name="Back-facing primitives" description="The number of back-facing triangles that are visible after culling." units="primitives"/>
+        <event offset="11" counter="PRIM_VISIBLE" title="Primitive Culling" name="Visible primitives" description="The number of primitives that are visible after culling." units="primitives"/>
+        <event offset="12" counter="PRIM_CULLED" title="Primitive Culling" name="Facing and XY plane test culled primitives" description="The number of primitives that are culled by facing or frustum XY plane tests." units="primitives"/>
+        <event offset="13" counter="PRIM_CLIPPED" title="Primitive Culling" name="Z plane test culled primitives" description="The number of primitives that are culled by frustum Z plane tests." units="primitives"/>
+        <event offset="14" counter="PRIM_SAT_CULLED" title="Primitive Culling" name="Sample test culled primitives" description="The number of primitives culled by the sample coverage test." units="primitives"/>
+        <event offset="17" advanced="yes" counter="BUS_READ" title="Tiler L2 Accesses" name="Read beats" description="The number of internal bus data read cycles made by the tiler." units="beats"/>
+        <event offset="19" advanced="yes" counter="BUS_WRITE" title="Tiler L2 Accesses" name="Write beats" description="The number of internal bus data write cycles made by the tiler." units="beats"/>
+        <event offset="21" counter="IDVS_POS_SHAD_REQ" title="Tiler Shading Requests" name="Position shading requests" description="The number of position shading requests in the IDVS flow." units="requests"/>
+        <event offset="23" advanced="yes" counter="IDVS_POS_SHAD_STALL" title="Tiler Cycles" name="Position shading stall cycles" description="The number of cycles where the tiler has a stalled position shading request." units="cycles"/>
+        <event offset="24" advanced="yes" counter="IDVS_POS_FIFO_FULL" title="Tiler Cycles" name="Position FIFO full cycles" description="The number of cycles where the tiler has a stalled position shading buffer." units="cycles"/>
+        <event offset="26" advanced="yes" counter="VCACHE_HIT" title="Tiler Vertex Cache" name="Position cache hits" description="The number of position lookups that result in a hit in the vertex cache." units="requests"/>
+        <event offset="27" advanced="yes" counter="VCACHE_MISS" title="Tiler Vertex Cache" name="Position cache misses" description="The number of position lookups that miss in the vertex cache." units="requests"/>
+        <event offset="31" advanced="yes" counter="VFETCH_STALL" title="Tiler Cycles" name="Primitive assembly busy stall cycles" description="The number of cycles where the tiler is stalled waiting for primitive assembly." units="cycles"/>
+        <event offset="34" advanced="yes" counter="IDVS_VBU_HIT" title="Tiler Vertex Cache" name="Varying cache hits" description="The number of varying lookups that result in a hit in the vertex cache." units="requests"/>
+        <event offset="35" advanced="yes" counter="IDVS_VBU_MISS" title="Tiler Vertex Cache" name="Varying cache misses" description="The number of varying lookups that miss in the vertex cache." units="requests"/>
+        <event offset="37" counter="IDVS_VAR_SHAD_REQ" title="Tiler Shading Requests" name="Varying shading requests" description="The number of varying shading requests in the IDVS flow." units="requests"/>
+        <event offset="38" advanced="yes" counter="IDVS_VAR_SHAD_STALL" title="Tiler Cycles" name="Varying shading stall cycles" description="The number of cycles where the tiler has a stalled varying shading request." units="cycles"/>
+        <event offset="54" advanced="yes" counter="WRBUF_NO_AXI_ID_STALL" title="Tiler Cycles" name="Write buffer transaction stall cycles" description="The number of cycles where the tiler write buffer can not send data because it has no available write IDs." units="cycles"/>
+        <event offset="55" advanced="yes" counter="WRBUF_AXI_STALL" title="Tiler Cycles" name="Write buffer write stall cycles" description="The number of cycles where the tiler write buffer can not send data because the bus is not ready." units="cycles"/>
+    </category>
+    <category name="Memory System" per_cpu="no">
+        <event offset="4" advanced="yes" counter="MMU_REQUESTS" title="MMU Stage 1 Translations" name="MMU lookups" description="The number of main MMU address translations performed." units="requests"/>
+        <event offset="16" advanced="yes" counter="L2_RD_MSG_IN" title="L2 Cache Requests" name="Read requests" description="The number of L2 cache read requests from internal masters." units="requests"/>
+        <event offset="17" advanced="yes" counter="L2_RD_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Read stall cycles" description="The number of cycles L2 cache read requests from internal masters are stalled." units="cycles"/>
+        <event offset="18" advanced="yes" counter="L2_WR_MSG_IN" title="L2 Cache Requests" name="Write requests" description="The number of L2 cache write requests from internal masters." units="requests"/>
+        <event offset="19" advanced="yes" counter="L2_WR_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Write stall cycles" description="The number of cycles where L2 cache write requests from internal masters are stalled." units="cycles"/>
+        <event offset="20" advanced="yes" counter="L2_SNP_MSG_IN" title="L2 Cache Requests" name="Snoop requests" description="The number of L2 snoop requests from internal masters." units="requests"/>
+        <event offset="21" advanced="yes" counter="L2_SNP_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Snoop stall cycles" description="The number of cycles where L2 cache snoop requests from internal masters are stalled." units="cycles"/>
+        <event offset="22" advanced="yes" counter="L2_RD_MSG_OUT" title="L2 Cache Requests" name="L1 read requests" description="The number of L1 cache read requests sent by the L2 cache to an internal master." units="requests"/>
+        <event offset="23" advanced="yes" counter="L2_RD_MSG_OUT_STALL" title="L2 Cache Stall Cycles" name="L1 read stall cycles" description="The number of cycles where L1 cache read requests sent by the L2 cache to an internal master are stalled." units="cycles"/>
+        <event offset="24" advanced="yes" counter="L2_WR_MSG_OUT" title="L2 Cache Requests" name="L1 write requests" description="The number of L1 cache write responses sent by the L2 cache to an internal master." units="requests"/>
+        <event offset="25" counter="L2_ANY_LOOKUP" title="L2 Cache Lookups" name="Any lookup" description="The number of L2 cache lookups performed." units="requests"/>
+        <event offset="26" counter="L2_READ_LOOKUP" title="L2 Cache Lookups" name="Read lookup" description="The number of L2 cache read lookups performed." units="requests"/>
+        <event offset="27" counter="L2_WRITE_LOOKUP" title="L2 Cache Lookups" name="Write lookup" description="The number of L2 cache write lookups performed." units="requests"/>
+        <event offset="28" advanced="yes" counter="L2_EXT_SNOOP_LOOKUP" title="L2 Cache Lookups" name="External snoop lookups" description="The number of coherency snoop lookups performed that were triggered by an external master." units="requests"/>
+        <event offset="29" counter="L2_EXT_READ" title="External Bus Accesses" name="Read transaction" description="The number of external read transactions." units="transactions"/>
+        <event offset="30" advanced="yes" counter="L2_EXT_READ_NOSNP" title="External Bus Accesses" name="ReadNoSnoop transactions" description="The number of external non-coherent read transactions." units="transactions"/>
+        <event offset="31" advanced="yes" counter="L2_EXT_READ_UNIQUE" title="External Bus Accesses" name="ReadUnique transactions" description="The number of external coherent read unique transactions." units="transactions"/>
+        <event offset="32" counter="L2_EXT_READ_BEATS" title="External Bus Beats" name="Read beat" description="The number of external bus data read cycles." units="beats"/>
+        <event offset="33" counter="L2_EXT_AR_STALL" title="External Bus Stalls" name="Read stall cycles" description="The number of cycles where a read is stalled waiting for the external bus." units="cycles"/>
+        <event offset="34" counter="L2_EXT_AR_CNT_Q1" title="External Bus Outstanding Reads" name="0-25% outstanding" description="The number of read transactions initiated when 0-25% of the maximum are in use." units="transactions"/>
+        <event offset="35" counter="L2_EXT_AR_CNT_Q2" title="External Bus Outstanding Reads" name="25-50% outstanding" description="The number of read transactions initiated when 25-50% of the maximum are in use." units="transactions"/>
+        <event offset="36" counter="L2_EXT_AR_CNT_Q3" title="External Bus Outstanding Reads" name="50-75% outstanding" description="The number of read transactions initiated when 50-75% of the maximum are in use." units="transactions"/>
+        <event offset="37" counter="L2_EXT_RRESP_0_127" title="External Bus Read Latency" name="0-127 cycles" description="The number of data beats returned 0-127 cycles after the read request." units="beats"/>
+        <event offset="38" counter="L2_EXT_RRESP_128_191" title="External Bus Read Latency" name="128-191 cycles" description="The number of data beats returned 128-191 cycles after the read request." units="beats"/>
+        <event offset="39" counter="L2_EXT_RRESP_192_255" title="External Bus Read Latency" name="192-255 cycles" description="The number of data beats returned 192-255 cycles after the read request." units="beats"/>
+        <event offset="40" counter="L2_EXT_RRESP_256_319" title="External Bus Read Latency" name="256-319 cycles" description="The number of data beats returned 256-319 cycles after the read request." units="beats"/>
+        <event offset="41" counter="L2_EXT_RRESP_320_383" title="External Bus Read Latency" name="320-383 cycles" description="The number of data beats returned 320-383 cycles after the read request." units="beats"/>
+        <event offset="42" counter="L2_EXT_WRITE" title="External Bus Accesses" name="Write transaction" description="The number of external write transactions." units="transactions"/>
+        <event offset="43" advanced="yes" counter="L2_EXT_WRITE_NOSNP_FULL" title="External Bus Accesses" name="WriteNoSnoopFull transactions" description="The number of external non-coherent full write transactions." units="transactions"/>
+        <event offset="44" advanced="yes" counter="L2_EXT_WRITE_NOSNP_PTL" title="External Bus Accesses" name="WriteNoSnoopPartial transactions" description="The number of external non-coherent partial write transactions." units="transactions"/>
+        <event offset="45" advanced="yes" counter="L2_EXT_WRITE_SNP_FULL" title="External Bus Accesses" name="WriteSnoopFull transactions" description="The number of external coherent full write transactions." units="transactions"/>
+        <event offset="46" advanced="yes" counter="L2_EXT_WRITE_SNP_PTL" title="External Bus Accesses" name="WriteSnoopPartial transactions" description="The number of external coherent partial write transactions." units="transactions"/>
+        <event offset="47" counter="L2_EXT_WRITE_BEATS" title="External Bus Beats" name="Write beat" description="The number of external bus data write cycles." units="beats"/>
+        <event offset="48" counter="L2_EXT_W_STALL" title="External Bus Stalls" name="Write stall cycles" description="The number of cycles where a write is stalled waiting for the external bus." units="cycles"/>
+        <event offset="49" counter="L2_EXT_AW_CNT_Q1" title="External Bus Outstanding Writes" name="0-25% outstanding" description="The number of write transactions initiated when 0-25% of the maximum are in use." units="transactions"/>
+        <event offset="50" counter="L2_EXT_AW_CNT_Q2" title="External Bus Outstanding Writes" name="25-50% outstanding" description="The number of write transactions initiated when 25-50% of the maximum are in use." units="transactions"/>
+        <event offset="51" counter="L2_EXT_AW_CNT_Q3" title="External Bus Outstanding Writes" name="50-75% outstanding" description="The number of write transactions initiated when 50-75% of the maximum are in use." units="transactions"/>
+        <event offset="52" advanced="yes" counter="L2_EXT_SNOOP" title="External Bus Accesses" name="Snoop transactions" description="The number of coherency snoops triggered by external masters." units="transactions"/>
+        <event offset="53" advanced="yes" counter="L2_EXT_SNOOP_STALL" title="External Bus Stalls" name="Snoop stall cycles" description="The number of cycles where a coherency snoop triggered by external master is stalled." units="cycles"/>
+    </category>
+    <category name="Shader Core" per_cpu="no">
+        <event offset="4" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment active" description="The number of cycles where the shader core is processing a fragment workload." units="cycles"/>
+        <event offset="5" advanced="yes" counter="FRAG_PRIMITIVES_OUT" title="Core Primitives" name="Read primitives" description="The number of primitives read from the tile list by the fragment front-end." units="primitives"/>
+        <event offset="6" counter="FRAG_PRIM_RAST" title="Core Primitives" name="Rasterized primitives" description="The number of primitives being rasterized." units="primitives"/>
+        <event offset="7" counter="FRAG_FPK_ACTIVE" title="Core Cycles" name="Fragment FPKB active" description="The number of cycles where at least one quad is present in the pre-pipe quad queue." units="cycles"/>
+        <event offset="9" counter="FRAG_WARPS" title="Core Warps" name="Fragment warps" description="The number of fragment warps created." units="warps"/>
+        <event offset="10" counter="FRAG_PARTIAL_QUADS_RAST" title="Core Quads" name="Partial rasterized quads" description="The number of partially-rasterized fragment quads created." units="quads"/>
+        <event offset="11" counter="FRAG_QUADS_RAST" title="Core Quads" name="Rasterized quads" description="The number of quads generated by the rasterization phase." units="quads"/>
+        <event offset="12" counter="FRAG_QUADS_EZS_TEST" title="Core Quads" name="Early ZS tested quads" description="The number of quads that are undergoing early depth and stencil testing." units="quads"/>
+        <event offset="13" counter="FRAG_QUADS_EZS_UPDATE" title="Core Quads" name="Early ZS updated quads" description="The number of quads undergoing early depth and stencil testing, that are capable of updating the framebuffer." units="quads"/>
+        <event offset="14" counter="FRAG_QUADS_EZS_KILL" title="Core Quads" name="Early ZS killed quads" description="The number of quads killed by early depth and stencil testing." units="quads"/>
+        <event offset="15" counter="FRAG_LZS_TEST" title="Core Quads" name="Late ZS tested quads" description="The number of quads undergoing late depth and stencil testing." units="quads"/>
+        <event offset="16" counter="FRAG_LZS_KILL" title="Core Quads" name="Late ZS killed quads" description="The number of quads killed by late depth and stencil testing." units="quads"/>
+        <event offset="17" counter="WARP_REG_SIZE_64" title="Core Warps" name="All register warps" description="The number of warps that require more than 32 registers." units="warps"/>
+        <event offset="18" counter="FRAG_PTILES" title="Core Tiles" name="Tiles" description="The number of tiles processed by the shader core." units="tiles"/>
+        <event offset="19" counter="FRAG_TRANS_ELIM" title="Core Tiles" name="Constant tiles killed" description="The number of tiles killed by transaction elimination." units="tiles"/>
+        <event offset="20" counter="QUAD_FPK_KILLER" title="Core Quads" name="FPK occluder quads" description="The number of quads that are valid occluders for hidden surface removal." units="quads"/>
+        <event offset="21" counter="FULL_QUAD_WARPS" title="Core Warps" name="Full quad warps" description="The number of warps that are fully populated with quads." units="warps"/>
+        <event offset="22" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Non-fragment active" description="The number of cycles where the shader core is processing some non-fragment workload." units="cycles"/>
+        <event offset="23" advanced="yes" counter="COMPUTE_TASKS" title="Core Tasks" name="Non-fragment tasks" description="The number of non-fragment tasks issued to the shader core." units="tasks"/>
+        <event offset="24" counter="COMPUTE_WARPS" title="Core Warps" name="Non-fragment warps" description="The number of non-fragment warps created." units="warps"/>
+        <event offset="25" advanced="yes" counter="COMPUTE_STARVING" title="Core Starvation Cycles" name="Non-fragment starvation cycles" description="The number of cycles where the shader core is processing a non-fragment workload and there are no new threads available for execution." units="cycles"/>
+        <event offset="26" counter="EXEC_CORE_ACTIVE" title="Core Cycles" name="Execution core active" description="The number of cycles where the shader core is processing at least one warp." units="cycles"/>
+        <event offset="27" counter="EXEC_INSTR_FMA" title="Core PU Instructions" name="FMA instructions" description="The number of instructions issued to the FMA pipe." units="instructions"/>
+        <event offset="28" counter="EXEC_INSTR_CVT" title="Core PU Instructions" name="CVT instructions" description="The number of instructions issued to the CVT pipe." units="instructions"/>
+        <event offset="29" counter="EXEC_INSTR_SFU" title="Core PU Instructions" name="SFU instructions" description="The number of instructions issued to the SFU pipe." units="instructions"/>
+        <event offset="30" counter="EXEC_INSTR_MSG" title="Core PU Instructions" name="Message instructions" description="The number of instructions issued to the MSG pipe." units="instructions"/>
+        <event offset="31" counter="EXEC_INSTR_DIVERGED" title="Core EE Instructions" name="Diverged instructions" description="The number of instructions executed per warp, that have control flow divergence." units="instructions"/>
+        <event offset="32" advanced="yes" counter="EXEC_ICACHE_MISS" title="Core PU Instructions" name="Instruction cache misses" description="The number of instruction cache misses." units="requests"/>
+        <event offset="33" advanced="yes" counter="EXEC_STARVE_ARITH" title="Core Starvation Cycles" name="Execution engine starvation cycles" description="The number of cycles where the processing unit is starved of work." units="cycles"/>
+        <event offset="34" counter="CALL_BLEND_SHADER" title="Core PU Instructions" name="Blend shader calls" description="The number of blend shader invocations executed." units="instructions"/>
+        <event offset="35" counter="TEX_MSGI_NUM_FLITS" title="Texture Bus" name="Input beats" description="The number of texture request message data beats." units="beats"/>
+        <event offset="36" counter="TEX_DFCH_CLK_STALLED" title="Core Texture Stalls" name="Descriptor stall cycles" description="The number of cycles where a quad is stalled on texture descriptor fetch." units="cycles"/>
+        <event offset="37" counter="TEX_TFCH_CLK_STALLED" title="Core Texture Stalls" name="Fetch queue stall cycles" description="The number of cycles where a quad is stalled on entering texture fetch because the fetch queue is full." units="cycles"/>
+        <event offset="38" counter="TEX_TFCH_STARVED_PENDING_DATA_FETCH" title="Core Texture Stalls" name="Filtering unit stall cycles" description="The number of cycles where the filtering unit is idle and there is at least one quad present in the texture data fetch queue." units="cycles"/>
+        <event offset="39" counter="TEX_FILT_NUM_OPERATIONS" title="Core Texture Cycles" name="Texturing active" description="The number of texture filtering issue cycles." units="cycles"/>
+        <event offset="40" counter="TEX_FILT_NUM_FXR_OPERATIONS" title="Core Texture Cycles" name="4x bilinear filtering active" description="The number of cycles where the filtering unit uses the 4x path to implement nearest or bilinear filtering." units="cycles"/>
+        <event offset="41" counter="TEX_FILT_NUM_FST_OPERATIONS" title="Core Texture Cycles" name="2x trilinear filtering active" description="The number of cycles where the filtering unit uses the 4x path to implement trilinear filtering." units="cycles"/>
+        <event offset="42" counter="TEX_MSGO_NUM_MSG" title="Core Texture Quads" name="Texture requests" description="The number of quad-width texture operations processed by the texture unit." units="quads"/>
+        <event offset="43" counter="TEX_MSGO_NUM_FLITS" title="Texture Bus" name="Output beats" description="The number of texture response message data beats." units="beats"/>
+        <event offset="44" counter="LS_MEM_READ_FULL" title="Core Load/Store Cycles" name="Full read cycles" description="The number of full-width load/store cache reads." units="cycles"/>
+        <event offset="45" counter="LS_MEM_READ_SHORT" title="Core Load/Store Cycles" name="Partial read cycles" description="The number of partial-width load/store cache reads." units="cycles"/>
+        <event offset="46" counter="LS_MEM_WRITE_FULL" title="Core Load/Store Cycles" name="Full write cycles" description="The number of full-width load/store cache writes." units="cycles"/>
+        <event offset="47" counter="LS_MEM_WRITE_SHORT" title="Core Load/Store Cycles" name="Partial write cycles" description="The number of partial-width load/store cache writes." units="cycles"/>
+        <event offset="48" counter="LS_MEM_ATOMIC" title="Core Load/Store Cycles" name="Atomic access cycles" description="The number of load/store atomic accesses." units="cycles"/>
+        <event offset="49" counter="VARY_INSTR" title="Core Varying Requests" name="Interpolation requests" description="The number of warp-width interpolation operations processed by the varying unit." units="instructions"/>
+        <event offset="50" counter="VARY_SLOT_32" title="Core Varying Cycles" name="32-bit interpolation active" description="The number of 32-bit interpolation cycles processed by the varying unit." units="cycles"/>
+        <event offset="51" counter="VARY_SLOT_16" title="Core Varying Cycles" name="16-bit interpolation active" description="The number of 16-bit interpolation cycles processed by the varying unit." units="cycles"/>
+        <event offset="52" advanced="yes" counter="ATTR_INSTR" title="Core Attribute Requests" name="Attribute requests" description="The number of instructions executed by the attribute unit." units="instructions"/>
+        <event offset="53" advanced="yes" counter="ARITH_INSTR_FP_MUL" title="Core EE Instructions" name="Multiplier instructions" description="The number of instructions where the workload uses floating-point multiplier hardware." units="instructions"/>
+        <event offset="54" counter="BEATS_RD_FTC" title="Core L2 Reads" name="Fragment L2 read beats" description="The number of read beats received by the fixed-function fragment front-end." units="beats"/>
+        <event offset="55" counter="BEATS_RD_FTC_EXT" title="Core L2 Reads" name="Fragment external read beats" description="The number of read beats received by the fixed-function fragment front-end that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="56" counter="BEATS_RD_LSC" title="Core L2 Reads" name="Load/store L2 read beats" description="The number of read beats received by the load/store unit." units="beats"/>
+        <event offset="57" counter="BEATS_RD_LSC_EXT" title="Core L2 Reads" name="Load/store external read beats" description="The number of read beats received by the load/store unit that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="58" counter="BEATS_RD_TEX" title="Core L2 Reads" name="Texture L2 read beats" description="The number of read beats received by the texture unit." units="beats"/>
+        <event offset="59" counter="BEATS_RD_TEX_EXT" title="Core L2 Reads" name="Texture external read beats" description="The number of read beats received by the texture unit that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="60" advanced="yes" counter="BEATS_RD_OTHER" title="Core L2 Reads" name="Other L2 read beats" description="The number of read beats received by a unit that is not specifically identified." units="beats"/>
+        <event offset="61" counter="BEATS_WR_LSC_OTHER" title="Core Writes" name="Load/store other write beats" description="The number of write beats by the load/store unit that are due to any reason other than writeback." units="beats"/>
+        <event offset="62" counter="BEATS_WR_TIB" title="Core Writes" name="Tile buffer write beats" description="The number of write beats sent by the tile buffer writeback unit." units="beats"/>
+        <event offset="63" counter="BEATS_WR_LSC_WB" title="Core Writes" name="Load/store writeback write beats" description="The number of write beats by the load/store unit that are due to writeback." units="beats"/>
     </category>
 </metrics>
\ No newline at end of file
diff --git a/src/panfrost/perf/G78.xml b/src/panfrost/perf/G78.xml
index fa7d9e95610..9288e316ab7 100644
--- a/src/panfrost/perf/G78.xml
+++ b/src/panfrost/perf/G78.xml
@@ -1,6 +1,6 @@
 <!--
 Copyright © 2017-2020 ARM Limited.
-Copyright © 2021 Collabora, Ltd.
+Copyright © 2021-2022 Collabora, Ltd.
 Author: Antonio Caggiano <antonio.caggiano@collabora.com>
 
 Permission is hereby granted, free of charge, to any person obtaining a
@@ -24,156 +24,156 @@ SOFTWARE.
 -->
 <metrics id="TBOx">
     <category name="Job Manager" per_cpu="no">
-        <event offset="6" counter="GPU_ACTIVE" title="GPU Cycles" name="GPU active" description="The number of cycles where the GPU has a workload of any type queued for processing." units="cycles" />
-        <event offset="7" advanced="yes" counter="IRQ_ACTIVE" title="GPU Cycles" name="Interrupt active" description="The number of cycles where the GPU has a pending interrupt." units="cycles" />
-        <event offset="8" advanced="yes" counter="JS0_JOBS" title="GPU Jobs" name="Fragment jobs" description="The number of jobs processed by the GPU fragment queue." units="jobs" />
-        <event offset="9" counter="JS0_TASKS" title="GPU Tasks" name="Fragment tasks" description="The number of 32x32 pixel tasks processed by the GPU fragment queue." units="tasks" />
-        <event offset="10" counter="JS0_ACTIVE" title="GPU Cycles" name="Fragment queue active" description="The number of cycles where work is queued for processing in the GPU fragment queue." units="cycles" />
-        <event offset="12" advanced="yes" counter="JS0_WAIT_READ" title="GPU Wait Cycles" name="Fragment descriptor reads cycles" description="The number of cycles where queued fragment work is waiting for a descriptor load." units="cycles" />
-        <event offset="13" advanced="yes" counter="JS0_WAIT_ISSUE" title="GPU Wait Cycles" name="Fragment job issue cycles" description="The number of cycles where queued fragment work is waiting for an available processor." units="cycles" />
-        <event offset="14" advanced="yes" counter="JS0_WAIT_DEPEND" title="GPU Wait Cycles" name="Fragment job dependency cycles" description="The number of cycles where queued fragment work is waiting for dependent work to complete." units="cycles" />
-        <event offset="15" advanced="yes" counter="JS0_WAIT_FINISH" title="GPU Wait Cycles" name="Fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued fragment work to complete." units="cycles" />
-        <event offset="16" advanced="yes" counter="JS1_JOBS" title="GPU Jobs" name="Non-fragment jobs" description="The number of jobs processed by the GPU non-fragment queue." units="jobs" />
-        <event offset="17" advanced="yes" counter="JS1_TASKS" title="GPU Tasks" name="Non-fragment tasks" description="The number of tasks processed by the GPU non-fragment queue." units="tasks" />
-        <event offset="18" counter="JS1_ACTIVE" title="GPU Cycles" name="Non-fragment queue active" description="The number of cycles where work is queued in the GPU non-fragment queue." units="cycles" />
-        <event offset="20" advanced="yes" counter="JS1_WAIT_READ" title="GPU Wait Cycles" name="Non-fragment descriptor read cycles" description="The number number of cycles where queued non-fragment work is waiting for a descriptor load." units="cycles" />
-        <event offset="21" advanced="yes" counter="JS1_WAIT_ISSUE" title="GPU Wait Cycles" name="Non-fragment job issue cycles" description="The number of cycles where queued non-fragment work is waiting for an available processor." units="cycles" />
-        <event offset="22" advanced="yes" counter="JS1_WAIT_DEPEND" title="GPU Wait Cycles" name="Non-fragment job dependency cycles" description="The number of cycles where queued non-fragment work is waiting for dependent work to complete." units="cycles" />
-        <event offset="23" advanced="yes" counter="JS1_WAIT_FINISH" title="GPU Wait Cycles" name="Non-fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued non-fragment work to complete." units="cycles" />
-        <event offset="24" advanced="yes" counter="JS2_JOBS" title="GPU Jobs" name="Reserved jobs" description="The number of jobs processed by the GPU reserved queue." units="jobs" />
-        <event offset="25" advanced="yes" counter="JS2_TASKS" title="GPU Tasks" name="Reserved tasks" description="The number of tasks processed by the GPU reserved queue." units="tasks" />
-        <event offset="26" advanced="yes" counter="JS2_ACTIVE" title="GPU Cycles" name="Reserved queue active" description="The number of cycles where work is queued in the GPU reserved queue." units="cycles" />
-        <event offset="28" advanced="yes" counter="JS2_WAIT_READ" title="GPU Wait Cycles" name="Reserved descriptor read cycles" description="The number of cycles where queued reserved work is waiting for a descriptor load." units="cycles" />
-        <event offset="29" advanced="yes" counter="JS2_WAIT_ISSUE" title="GPU Wait Cycles" name="Reserved job issue cycles" description="The number of cycles where queued reserved work is waiting for an available processor." units="cycles" />
-        <event offset="30" advanced="yes" counter="JS2_WAIT_DEPEND" title="GPU Wait Cycles" name="Reserved job dependency cycles" description="The number of cycles where queued reserved work is waiting for dependent work to complete." units="cycles" />
-        <event offset="31" advanced="yes" counter="JS2_WAIT_FINISH" title="GPU Wait Cycles" name="Reserved job finish cycles" description="The number of cycles where the GPU is waiting for issued reserved work to complete." units="cycles" />
-    </category>
-    <category name="Memory System" per_cpu="no">
-        <event offset="196" advanced="yes" counter="MMU_REQUESTS" title="MMU Stage 1 Translations" name="MMU lookups" description="The number of main MMU address translations performed." units="requests" />
-        <event offset="208" advanced="yes" counter="L2_RD_MSG_IN" title="L2 Cache Requests" name="Read requests" description="The number of L2 cache read requests from internal masters." units="requests" />
-        <event offset="209" advanced="yes" counter="L2_RD_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Read stall cycles" description="The number of cycles L2 cache read requests from internal masters are stalled." units="cycles" />
-        <event offset="210" advanced="yes" counter="L2_WR_MSG_IN" title="L2 Cache Requests" name="Write requests" description="The number of L2 cache write requests from internal masters." units="requests" />
-        <event offset="211" advanced="yes" counter="L2_WR_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Write stall cycles" description="The number of cycles where L2 cache write requests from internal masters are stalled." units="cycles" />
-        <event offset="212" advanced="yes" counter="L2_SNP_MSG_IN" title="L2 Cache Requests" name="Snoop requests" description="The number of L2 snoop requests from internal masters." units="requests" />
-        <event offset="213" advanced="yes" counter="L2_SNP_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Snoop stall cycles" description="The number of cycles where L2 cache snoop requests from internal masters are stalled." units="cycles" />
-        <event offset="214" advanced="yes" counter="L2_RD_MSG_OUT" title="L2 Cache Requests" name="L1 read requests" description="The number of L1 cache read requests sent by the L2 cache to an internal master." units="requests" />
-        <event offset="215" advanced="yes" counter="L2_RD_MSG_OUT_STALL" title="L2 Cache Stall Cycles" name="L1 read stall cycles" description="The number of cycles where L1 cache read requests sent by the L2 cache to an internal master are stalled." units="cycles" />
-        <event offset="216" advanced="yes" counter="L2_WR_MSG_OUT" title="L2 Cache Requests" name="L1 write requests" description="The number of L1 cache write responses sent by the L2 cache to an internal master." units="requests" />
-        <event offset="217" counter="L2_ANY_LOOKUP" title="L2 Cache Lookups" name="Any lookup" description="The number of L2 cache lookups performed." units="requests" />
-        <event offset="218" counter="L2_READ_LOOKUP" title="L2 Cache Lookups" name="Read lookup" description="The number of L2 cache read lookups performed." units="requests" />
-        <event offset="219" counter="L2_WRITE_LOOKUP" title="L2 Cache Lookups" name="Write lookup" description="The number of L2 cache write lookups performed." units="requests" />
-        <event offset="220" advanced="yes" counter="L2_EXT_SNOOP_LOOKUP" title="L2 Cache Lookups" name="External snoop lookups" description="The number of coherency snoop lookups performed that were triggered by an external master." units="requests" />
-        <event offset="221" counter="L2_EXT_READ" title="External Bus Accesses" name="Read transaction" description="The number of external read transactions." units="transactions" />
-        <event offset="222" advanced="yes" counter="L2_EXT_READ_NOSNP" title="External Bus Accesses" name="ReadNoSnoop transactions" description="The number of external non-coherent read transactions." units="transactions" />
-        <event offset="223" advanced="yes" counter="L2_EXT_READ_UNIQUE" title="External Bus Accesses" name="ReadUnique transactions" description="The number of external coherent read unique transactions." units="transactions" />
-        <event offset="224" counter="L2_EXT_READ_BEATS" title="External Bus Beats" name="Read beat" description="The number of external bus data read cycles." units="beats" />
-        <event offset="225" counter="L2_EXT_AR_STALL" title="External Bus Stalls" name="Read stall cycles" description="The number of cycles where a read is stalled waiting for the external bus." units="cycles" />
-        <event offset="226" counter="L2_EXT_AR_CNT_Q1" title="External Bus Outstanding Reads" name="0-25% outstanding" description="The number of read transactions initiated when 0-25% of the maximum are in use." units="transactions" />
-        <event offset="227" counter="L2_EXT_AR_CNT_Q2" title="External Bus Outstanding Reads" name="25-50% outstanding" description="The number of read transactions initiated when 25-50% of the maximum are in use." units="transactions" />
-        <event offset="228" counter="L2_EXT_AR_CNT_Q3" title="External Bus Outstanding Reads" name="50-75% outstanding" description="The number of read transactions initiated when 50-75% of the maximum are in use." units="transactions" />
-        <event offset="229" counter="L2_EXT_RRESP_0_127" title="External Bus Read Latency" name="0-127 cycles" description="The number of data beats returned 0-127 cycles after the read request." units="beats" />
-        <event offset="230" counter="L2_EXT_RRESP_128_191" title="External Bus Read Latency" name="128-191 cycles" description="The number of data beats returned 128-191 cycles after the read request." units="beats" />
-        <event offset="231" counter="L2_EXT_RRESP_192_255" title="External Bus Read Latency" name="192-255 cycles" description="The number of data beats returned 192-255 cycles after the read request." units="beats" />
-        <event offset="232" counter="L2_EXT_RRESP_256_319" title="External Bus Read Latency" name="256-319 cycles" description="The number of data beats returned 256-319 cycles after the read request." units="beats" />
-        <event offset="233" counter="L2_EXT_RRESP_320_383" title="External Bus Read Latency" name="320-383 cycles" description="The number of data beats returned 320-383 cycles after the read request." units="beats" />
-        <event offset="234" counter="L2_EXT_WRITE" title="External Bus Accesses" name="Write transaction" description="The number of external write transactions." units="transactions" />
-        <event offset="235" advanced="yes" counter="L2_EXT_WRITE_NOSNP_FULL" title="External Bus Accesses" name="WriteNoSnoopFull transactions" description="The number of external non-coherent full write transactions." units="transactions" />
-        <event offset="236" advanced="yes" counter="L2_EXT_WRITE_NOSNP_PTL" title="External Bus Accesses" name="WriteNoSnoopPartial transactions" description="The number of external non-coherent partial write transactions." units="transactions" />
-        <event offset="237" advanced="yes" counter="L2_EXT_WRITE_SNP_FULL" title="External Bus Accesses" name="WriteSnoopFull transactions" description="The number of external coherent full write transactions." units="transactions" />
-        <event offset="238" advanced="yes" counter="L2_EXT_WRITE_SNP_PTL" title="External Bus Accesses" name="WriteSnoopPartial transactions" description="The number of external coherent partial write transactions." units="transactions" />
-        <event offset="239" counter="L2_EXT_WRITE_BEATS" title="External Bus Beats" name="Write beat" description="The number of external bus data write cycles." units="beats" />
-        <event offset="240" counter="L2_EXT_W_STALL" title="External Bus Stalls" name="Write stall cycles" description="The number of cycles where a write is stalled waiting for the external bus." units="cycles" />
-        <event offset="241" counter="L2_EXT_AW_CNT_Q1" title="External Bus Outstanding Writes" name="0-25% outstanding" description="The number of write transactions initiated when 0-25% of the maximum are in use." units="transactions" />
-        <event offset="242" counter="L2_EXT_AW_CNT_Q2" title="External Bus Outstanding Writes" name="25-50% outstanding" description="The number of write transactions initiated when 25-50% of the maximum are in use." units="transactions" />
-        <event offset="243" counter="L2_EXT_AW_CNT_Q3" title="External Bus Outstanding Writes" name="50-75% outstanding" description="The number of write transactions initiated when 50-75% of the maximum are in use." units="transactions" />
-        <event offset="244" advanced="yes" counter="L2_EXT_SNOOP" title="External Bus Accesses" name="Snoop transactions" description="The number of coherency snoops triggered by external masters." units="transactions" />
-        <event offset="245" advanced="yes" counter="L2_EXT_SNOOP_STALL" title="External Bus Stalls" name="Snoop stall cycles" description="The number of cycles where a coherency snoop triggered by external master is stalled." units="cycles" />
-    </category>
-    <category name="Shader Core" per_cpu="no">
-        <event offset="196" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment active" description="The number of cycles where the shader core is processing a fragment workload." units="cycles" />
-        <event offset="197" advanced="yes" counter="FRAG_PRIMITIVES_OUT" title="Core Primitives" name="Read primitives" description="The number of primitives read from the tile list by the fragment front-end." units="primitives" />
-        <event offset="198" counter="FRAG_PRIM_RAST" title="Core Primitives" name="Rasterized primitives" description="The number of primitives being rasterized." units="primitives" />
-        <event offset="199" counter="FRAG_FPK_ACTIVE" title="Core Cycles" name="Fragment FPKB active" description="The number of cycles where at least one quad is present in the pre-pipe quad queue." units="cycles" />
-        <event offset="201" counter="FRAG_WARPS" title="Core Warps" name="Fragment warps" description="The number of fragment warps created." units="warps" />
-        <event offset="202" counter="FRAG_PARTIAL_QUADS_RAST" title="Core Quads" name="Partial rasterized quads" description="The number of partially-rasterized fragment quads created." units="quads" />
-        <event offset="203" counter="FRAG_QUADS_RAST" title="Core Quads" name="Rasterized quads" description="The number of quads generated by the rasterization phase." units="quads" />
-        <event offset="204" counter="FRAG_QUADS_EZS_TEST" title="Core Quads" name="Early ZS tested quads" description="The number of quads that are undergoing early depth and stencil testing." units="quads" />
-        <event offset="205" counter="FRAG_QUADS_EZS_UPDATE" title="Core Quads" name="Early ZS updated quads" description="The number of quads undergoing early depth and stencil testing, that are capable of updating the framebuffer." units="quads" />
-        <event offset="206" counter="FRAG_QUADS_EZS_KILL" title="Core Quads" name="Early ZS killed quads" description="The number of quads killed by early depth and stencil testing." units="quads" />
-        <event offset="207" counter="FRAG_LZS_TEST" title="Core Quads" name="Late ZS tested quads" description="The number of quads undergoing late depth and stencil testing." units="quads" />
-        <event offset="208" counter="FRAG_LZS_KILL" title="Core Quads" name="Late ZS killed quads" description="The number of quads killed by late depth and stencil testing." units="quads" />
-        <event offset="209" counter="WARP_REG_SIZE_64" title="Core Warps" name="All register warps" description="The number of warps that require more than 32 registers." units="warps" />
-        <event offset="210" counter="FRAG_PTILES" title="Core Tiles" name="Tiles" description="The number of tiles processed by the shader core." units="tiles" />
-        <event offset="211" counter="FRAG_TRANS_ELIM" title="Core Tiles" name="Constant tiles killed" description="The number of tiles killed by transaction elimination." units="tiles" />
-        <event offset="212" counter="QUAD_FPK_KILLER" title="Core Quads" name="FPK occluder quads" description="The number of quads that are valid occluders for hidden surface removal." units="quads" />
-        <event offset="213" counter="FULL_QUAD_WARPS" title="Core Warps" name="Full quad warps" description="The number of warps that are fully populated with quads." units="warps" />
-        <event offset="214" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Non-fragment active" description="The number of cycles where the shader core is processing some non-fragment workload." units="cycles" />
-        <event offset="215" advanced="yes" counter="COMPUTE_TASKS" title="Core Tasks" name="Non-fragment tasks" description="The number of non-fragment tasks issued to the shader core." units="tasks" />
-        <event offset="216" counter="COMPUTE_WARPS" title="Core Warps" name="Non-fragment warps" description="The number of non-fragment warps created." units="warps" />
-        <event offset="217" advanced="yes" counter="COMPUTE_STARVING" title="Core Starvation Cycles" name="Non-fragment starvation cycles" description="The number of cycles where the shader core is processing a non-fragment workload and there are no new threads available for execution." units="cycles" />
-        <event offset="218" counter="EXEC_CORE_ACTIVE" title="Core Cycles" name="Execution core active" description="The number of cycles where the shader core is processing at least one warp." units="cycles" />
-        <event offset="219" counter="EXEC_INSTR_FMA" title="Core PU Instructions" name="FMA instructions" description="The number of instructions issued to the FMA pipe." units="instructions" />
-        <event offset="220" counter="EXEC_INSTR_CVT" title="Core PU Instructions" name="CVT instructions" description="The number of instructions issued to the CVT pipe." units="instructions" />
-        <event offset="221" counter="EXEC_INSTR_SFU" title="Core PU Instructions" name="SFU instructions" description="The number of instructions issued to the SFU pipe." units="instructions" />
-        <event offset="222" counter="EXEC_INSTR_MSG" title="Core PU Instructions" name="Message instructions" description="The number of instructions issued to the MSG pipe." units="instructions" />
-        <event offset="223" counter="EXEC_INSTR_DIVERGED" title="Core EE Instructions" name="Diverged instructions" description="The number of instructions executed per warp, that have control flow divergence." units="instructions" />
-        <event offset="224" advanced="yes" counter="EXEC_ICACHE_MISS" title="Core PU Instructions" name="Instruction cache misses" description="The number of instruction cache misses." units="requests" />
-        <event offset="225" advanced="yes" counter="EXEC_STARVE_ARITH" title="Core Starvation Cycles" name="Execution engine starvation cycles" description="The number of cycles where the processing unit is starved of work." units="cycles" />
-        <event offset="226" counter="CALL_BLEND_SHADER" title="Core PU Instructions" name="Blend shader calls" description="The number of blend shader invocations executed." units="instructions" />
-        <event offset="227" counter="TEX_MSGI_NUM_FLITS" title="Texture Bus" name="Input beats" description="The number of texture request message data beats." units="beats" />
-        <event offset="228" counter="TEX_DFCH_CLK_STALLED" title="Core Texture Stalls" name="Descriptor stall cycles" description="The number of cycles where a quad is stalled on texture descriptor fetch." units="cycles" />
-        <event offset="229" counter="TEX_TFCH_CLK_STALLED" title="Core Texture Stalls" name="Fetch queue stall cycles" description="The number of cycles where a quad is stalled on entering texture fetch because the fetch queue is full." units="cycles" />
-        <event offset="230" counter="TEX_TFCH_STARVED_PENDING_DATA_FETCH" title="Core Texture Stalls" name="Filtering unit stall cycles" description="The number of cycles where the filtering unit is idle and there is at least one quad present in the texture data fetch queue." units="cycles" />
-        <event offset="231" counter="TEX_FILT_NUM_OPERATIONS" title="Core Texture Cycles" name="Texturing active" description="The number of texture filtering issue cycles." units="cycles" />
-        <event offset="232" counter="TEX_FILT_NUM_FXR_OPERATIONS" title="Core Texture Cycles" name="4x bilinear filtering active" description="The number of cycles where the filtering unit uses the 4x path to implement nearest or bilinear filtering." units="cycles" />
-        <event offset="233" counter="TEX_FILT_NUM_FST_OPERATIONS" title="Core Texture Cycles" name="2x trilinear filtering active" description="The number of cycles where the filtering unit uses the 4x path to implement trilinear filtering." units="cycles" />
-        <event offset="234" counter="TEX_MSGO_NUM_MSG" title="Core Texture Quads" name="Texture requests" description="The number of quad-width texture operations processed by the texture unit." units="quads" />
-        <event offset="235" counter="TEX_MSGO_NUM_FLITS" title="Texture Bus" name="Output beats" description="The number of texture response message data beats." units="beats" />
-        <event offset="236" counter="LS_MEM_READ_FULL" title="Core Load/Store Cycles" name="Full read cycles" description="The number of full-width load/store cache reads." units="cycles" />
-        <event offset="237" counter="LS_MEM_READ_SHORT" title="Core Load/Store Cycles" name="Partial read cycles" description="The number of partial-width load/store cache reads." units="cycles" />
-        <event offset="238" counter="LS_MEM_WRITE_FULL" title="Core Load/Store Cycles" name="Full write cycles" description="The number of full-width load/store cache writes." units="cycles" />
-        <event offset="239" counter="LS_MEM_WRITE_SHORT" title="Core Load/Store Cycles" name="Partial write cycles" description="The number of partial-width load/store cache writes." units="cycles" />
-        <event offset="240" counter="LS_MEM_ATOMIC" title="Core Load/Store Cycles" name="Atomic access cycles" description="The number of load/store atomic accesses." units="cycles" />
-        <event offset="241" counter="VARY_INSTR" title="Core Varying Requests" name="Interpolation requests" description="The number of warp-width interpolation operations processed by the varying unit." units="instructions" />
-        <event offset="242" counter="VARY_SLOT_32" title="Core Varying Cycles" name="32-bit interpolation active" description="The number of 32-bit interpolation cycles processed by the varying unit." units="cycles" />
-        <event offset="243" counter="VARY_SLOT_16" title="Core Varying Cycles" name="16-bit interpolation active" description="The number of 16-bit interpolation cycles processed by the varying unit." units="cycles" />
-        <event offset="244" advanced="yes" counter="ATTR_INSTR" title="Core Attribute Requests" name="Attribute requests" description="The number of instructions executed by the attribute unit." units="instructions" />
-        <event offset="245" counter="SHADER_CORE_ACTIVE" title="Core Cycles" name="Any active" description="The number of cycles where the shader core is processing either a non-fragment workload or a fragment workload." units="cycles" />
-        <event offset="246" counter="BEATS_RD_FTC" title="Core L2 Reads" name="Fragment L2 read beats" description="The number of read beats received by the fixed-function fragment front-end." units="beats" />
-        <event offset="247" counter="BEATS_RD_FTC_EXT" title="Core L2 Reads" name="Fragment external read beats" description="The number of read beats received by the fixed-function fragment front-end that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="248" counter="BEATS_RD_LSC" title="Core L2 Reads" name="Load/store L2 read beats" description="The number of read beats received by the load/store unit." units="beats" />
-        <event offset="249" counter="BEATS_RD_LSC_EXT" title="Core L2 Reads" name="Load/store external read beats" description="The number of read beats received by the load/store unit that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="250" counter="BEATS_RD_TEX" title="Core L2 Reads" name="Texture L2 read beats" description="The number of read beats received by the texture unit." units="beats" />
-        <event offset="251" counter="BEATS_RD_TEX_EXT" title="Core L2 Reads" name="Texture external read beats" description="The number of read beats received by the texture unit that required an external memory access due to an L2 cache miss." units="beats" />
-        <event offset="252" advanced="yes" counter="BEATS_RD_OTHER" title="Core L2 Reads" name="Other L2 read beats" description="The number of read beats received by a unit that is not specifically identified." units="beats" />
-        <event offset="253" counter="BEATS_WR_LSC_OTHER" title="Core Writes" name="Load/store other write beats" description="The number of write beats by the load/store unit that are due to any reason other than writeback." units="beats" />
-        <event offset="254" counter="BEATS_WR_TIB" title="Core Writes" name="Tile buffer write beats" description="The number of write beats sent by the tile buffer writeback unit." units="beats" />
-        <event offset="255" counter="BEATS_WR_LSC_WB" title="Core Writes" name="Load/store writeback write beats" description="The number of write beats by the load/store unit that are due to writeback." units="beats" />
+        <event offset="6" counter="GPU_ACTIVE" title="GPU Cycles" name="GPU active" description="The number of cycles where the GPU has a workload of any type queued for processing." units="cycles"/>
+        <event offset="7" advanced="yes" counter="IRQ_ACTIVE" title="GPU Cycles" name="Interrupt active" description="The number of cycles where the GPU has a pending interrupt." units="cycles"/>
+        <event offset="8" advanced="yes" counter="JS0_JOBS" title="GPU Jobs" name="Fragment jobs" description="The number of jobs processed by the GPU fragment queue." units="jobs"/>
+        <event offset="9" counter="JS0_TASKS" title="GPU Tasks" name="Fragment tasks" description="The number of 32x32 pixel tasks processed by the GPU fragment queue." units="tasks"/>
+        <event offset="10" counter="JS0_ACTIVE" title="GPU Cycles" name="Fragment queue active" description="The number of cycles where work is queued for processing in the GPU fragment queue." units="cycles"/>
+        <event offset="12" advanced="yes" counter="JS0_WAIT_READ" title="GPU Wait Cycles" name="Fragment descriptor reads cycles" description="The number of cycles where queued fragment work is waiting for a descriptor load." units="cycles"/>
+        <event offset="13" advanced="yes" counter="JS0_WAIT_ISSUE" title="GPU Wait Cycles" name="Fragment job issue cycles" description="The number of cycles where queued fragment work is waiting for an available processor." units="cycles"/>
+        <event offset="14" advanced="yes" counter="JS0_WAIT_DEPEND" title="GPU Wait Cycles" name="Fragment job dependency cycles" description="The number of cycles where queued fragment work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="15" advanced="yes" counter="JS0_WAIT_FINISH" title="GPU Wait Cycles" name="Fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued fragment work to complete." units="cycles"/>
+        <event offset="16" advanced="yes" counter="JS1_JOBS" title="GPU Jobs" name="Non-fragment jobs" description="The number of jobs processed by the GPU non-fragment queue." units="jobs"/>
+        <event offset="17" advanced="yes" counter="JS1_TASKS" title="GPU Tasks" name="Non-fragment tasks" description="The number of tasks processed by the GPU non-fragment queue." units="tasks"/>
+        <event offset="18" counter="JS1_ACTIVE" title="GPU Cycles" name="Non-fragment queue active" description="The number of cycles where work is queued in the GPU non-fragment queue." units="cycles"/>
+        <event offset="20" advanced="yes" counter="JS1_WAIT_READ" title="GPU Wait Cycles" name="Non-fragment descriptor read cycles" description="The number number of cycles where queued non-fragment work is waiting for a descriptor load." units="cycles"/>
+        <event offset="21" advanced="yes" counter="JS1_WAIT_ISSUE" title="GPU Wait Cycles" name="Non-fragment job issue cycles" description="The number of cycles where queued non-fragment work is waiting for an available processor." units="cycles"/>
+        <event offset="22" advanced="yes" counter="JS1_WAIT_DEPEND" title="GPU Wait Cycles" name="Non-fragment job dependency cycles" description="The number of cycles where queued non-fragment work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="23" advanced="yes" counter="JS1_WAIT_FINISH" title="GPU Wait Cycles" name="Non-fragment job finish cycles" description="The number of cycles where the GPU is waiting for issued non-fragment work to complete." units="cycles"/>
+        <event offset="24" advanced="yes" counter="JS2_JOBS" title="GPU Jobs" name="Reserved jobs" description="The number of jobs processed by the GPU reserved queue." units="jobs"/>
+        <event offset="25" advanced="yes" counter="JS2_TASKS" title="GPU Tasks" name="Reserved tasks" description="The number of tasks processed by the GPU reserved queue." units="tasks"/>
+        <event offset="26" advanced="yes" counter="JS2_ACTIVE" title="GPU Cycles" name="Reserved queue active" description="The number of cycles where work is queued in the GPU reserved queue." units="cycles"/>
+        <event offset="28" advanced="yes" counter="JS2_WAIT_READ" title="GPU Wait Cycles" name="Reserved descriptor read cycles" description="The number of cycles where queued reserved work is waiting for a descriptor load." units="cycles"/>
+        <event offset="29" advanced="yes" counter="JS2_WAIT_ISSUE" title="GPU Wait Cycles" name="Reserved job issue cycles" description="The number of cycles where queued reserved work is waiting for an available processor." units="cycles"/>
+        <event offset="30" advanced="yes" counter="JS2_WAIT_DEPEND" title="GPU Wait Cycles" name="Reserved job dependency cycles" description="The number of cycles where queued reserved work is waiting for dependent work to complete." units="cycles"/>
+        <event offset="31" advanced="yes" counter="JS2_WAIT_FINISH" title="GPU Wait Cycles" name="Reserved job finish cycles" description="The number of cycles where the GPU is waiting for issued reserved work to complete." units="cycles"/>
     </category>
     <category name="Tiler" per_cpu="no">
-        <event offset="4" counter="TILER_ACTIVE" title="GPU Cycles" name="Tiler active" description="The number of cycles where the tiler has a workload queued for processing." units="cycles" />
-        <event offset="6" counter="TRIANGLES" title="Input Primitives" name="Triangle primitives" description="The number of input triangle primitives." units="primitives" />
-        <event offset="7" counter="LINES" title="Input Primitives" name="Line primitives" description="The number of input line primitives." units="primitives" />
-        <event offset="8" counter="POINTS" title="Input Primitives" name="Point primitives" description="The number of input point primitives." units="primitives" />
-        <event offset="9" counter="FRONT_FACING" title="Visible Primitives" name="Front-facing primitives" description="The number of front-facing triangles that are visible after culling." units="primitives" />
-        <event offset="10" counter="BACK_FACING" title="Visible Primitives" name="Back-facing primitives" description="The number of back-facing triangles that are visible after culling." units="primitives" />
-        <event offset="11" counter="PRIM_VISIBLE" title="Primitive Culling" name="Visible primitives" description="The number of primitives that are visible after culling." units="primitives" />
-        <event offset="12" counter="PRIM_CULLED" title="Primitive Culling" name="Facing and XY plane test culled primitives" description="The number of primitives that are culled by facing or frustum XY plane tests." units="primitives" />
-        <event offset="13" counter="PRIM_CLIPPED" title="Primitive Culling" name="Z plane test culled primitives" description="The number of primitives that are culled by frustum Z plane tests." units="primitives" />
-        <event offset="14" counter="PRIM_SAT_CULLED" title="Primitive Culling" name="Sample test culled primitives" description="The number of primitives culled by the sample coverage test." units="primitives" />
-        <event offset="17" advanced="yes" counter="BUS_READ" title="Tiler L2 Accesses" name="Read beats" description="The number of internal bus data read cycles made by the tiler." units="beats" />
-        <event offset="19" advanced="yes" counter="BUS_WRITE" title="Tiler L2 Accesses" name="Write beats" description="The number of internal bus data write cycles made by the tiler." units="beats" />
-        <event offset="21" counter="IDVS_POS_SHAD_REQ" title="Tiler Shading Requests" name="Position shading requests" description="The number of position shading requests in the IDVS flow." units="requests" />
-        <event offset="23" advanced="yes" counter="IDVS_POS_SHAD_STALL" title="Tiler Cycles" name="Position shading stall cycles" description="The number of cycles where the tiler has a stalled position shading request." units="cycles" />
-        <event offset="24" advanced="yes" counter="IDVS_POS_FIFO_FULL" title="Tiler Cycles" name="Position FIFO full cycles" description="The number of cycles where the tiler has a stalled position shading buffer." units="cycles" />
-        <event offset="26" advanced="yes" counter="VCACHE_HIT" title="Tiler Vertex Cache" name="Position cache hits" description="The number of position lookups that result in a hit in the vertex cache." units="requests" />
-        <event offset="27" advanced="yes" counter="VCACHE_MISS" title="Tiler Vertex Cache" name="Position cache misses" description="The number of position lookups that miss in the vertex cache." units="requests" />
-        <event offset="31" advanced="yes" counter="VFETCH_STALL" title="Tiler Cycles" name="Primitive assembly busy stall cycles" description="The number of cycles where the tiler is stalled waiting for primitive assembly." units="cycles" />
-        <event offset="34" advanced="yes" counter="IDVS_VBU_HIT" title="Tiler Vertex Cache" name="Varying cache hits" description="The number of varying lookups that result in a hit in the vertex cache." units="requests" />
-        <event offset="35" advanced="yes" counter="IDVS_VBU_MISS" title="Tiler Vertex Cache" name="Varying cache misses" description="The number of varying lookups that miss in the vertex cache." units="requests" />
-        <event offset="37" counter="IDVS_VAR_SHAD_REQ" title="Tiler Shading Requests" name="Varying shading requests" description="The number of varying shading requests in the IDVS flow." units="requests" />
-        <event offset="38" advanced="yes" counter="IDVS_VAR_SHAD_STALL" title="Tiler Cycles" name="Varying shading stall cycles" description="The number of cycles where the tiler has a stalled varying shading request." units="cycles" />
-        <event offset="54" advanced="yes" counter="WRBUF_NO_AXI_ID_STALL" title="Tiler Cycles" name="Write buffer transaction stall cycles" description="The number of cycles where the tiler write buffer can not send data because it has no available write IDs." units="cycles" />
-        <event offset="55" advanced="yes" counter="WRBUF_AXI_STALL" title="Tiler Cycles" name="Write buffer write stall cycles" description="The number of cycles where the tiler write buffer can not send data because the bus is not ready." units="cycles" />
+        <event offset="4" counter="TILER_ACTIVE" title="GPU Cycles" name="Tiler active" description="The number of cycles where the tiler has a workload queued for processing." units="cycles"/>
+        <event offset="6" counter="TRIANGLES" title="Input Primitives" name="Triangle primitives" description="The number of input triangle primitives." units="primitives"/>
+        <event offset="7" counter="LINES" title="Input Primitives" name="Line primitives" description="The number of input line primitives." units="primitives"/>
+        <event offset="8" counter="POINTS" title="Input Primitives" name="Point primitives" description="The number of input point primitives." units="primitives"/>
+        <event offset="9" counter="FRONT_FACING" title="Visible Primitives" name="Front-facing primitives" description="The number of front-facing triangles that are visible after culling." units="primitives"/>
+        <event offset="10" counter="BACK_FACING" title="Visible Primitives" name="Back-facing primitives" description="The number of back-facing triangles that are visible after culling." units="primitives"/>
+        <event offset="11" counter="PRIM_VISIBLE" title="Primitive Culling" name="Visible primitives" description="The number of primitives that are visible after culling." units="primitives"/>
+        <event offset="12" counter="PRIM_CULLED" title="Primitive Culling" name="Facing and XY plane test culled primitives" description="The number of primitives that are culled by facing or frustum XY plane tests." units="primitives"/>
+        <event offset="13" counter="PRIM_CLIPPED" title="Primitive Culling" name="Z plane test culled primitives" description="The number of primitives that are culled by frustum Z plane tests." units="primitives"/>
+        <event offset="14" counter="PRIM_SAT_CULLED" title="Primitive Culling" name="Sample test culled primitives" description="The number of primitives culled by the sample coverage test." units="primitives"/>
+        <event offset="17" advanced="yes" counter="BUS_READ" title="Tiler L2 Accesses" name="Read beats" description="The number of internal bus data read cycles made by the tiler." units="beats"/>
+        <event offset="19" advanced="yes" counter="BUS_WRITE" title="Tiler L2 Accesses" name="Write beats" description="The number of internal bus data write cycles made by the tiler." units="beats"/>
+        <event offset="21" counter="IDVS_POS_SHAD_REQ" title="Tiler Shading Requests" name="Position shading requests" description="The number of position shading requests in the IDVS flow." units="requests"/>
+        <event offset="23" advanced="yes" counter="IDVS_POS_SHAD_STALL" title="Tiler Cycles" name="Position shading stall cycles" description="The number of cycles where the tiler has a stalled position shading request." units="cycles"/>
+        <event offset="24" advanced="yes" counter="IDVS_POS_FIFO_FULL" title="Tiler Cycles" name="Position FIFO full cycles" description="The number of cycles where the tiler has a stalled position shading buffer." units="cycles"/>
+        <event offset="26" advanced="yes" counter="VCACHE_HIT" title="Tiler Vertex Cache" name="Position cache hits" description="The number of position lookups that result in a hit in the vertex cache." units="requests"/>
+        <event offset="27" advanced="yes" counter="VCACHE_MISS" title="Tiler Vertex Cache" name="Position cache misses" description="The number of position lookups that miss in the vertex cache." units="requests"/>
+        <event offset="31" advanced="yes" counter="VFETCH_STALL" title="Tiler Cycles" name="Primitive assembly busy stall cycles" description="The number of cycles where the tiler is stalled waiting for primitive assembly." units="cycles"/>
+        <event offset="34" advanced="yes" counter="IDVS_VBU_HIT" title="Tiler Vertex Cache" name="Varying cache hits" description="The number of varying lookups that result in a hit in the vertex cache." units="requests"/>
+        <event offset="35" advanced="yes" counter="IDVS_VBU_MISS" title="Tiler Vertex Cache" name="Varying cache misses" description="The number of varying lookups that miss in the vertex cache." units="requests"/>
+        <event offset="37" counter="IDVS_VAR_SHAD_REQ" title="Tiler Shading Requests" name="Varying shading requests" description="The number of varying shading requests in the IDVS flow." units="requests"/>
+        <event offset="38" advanced="yes" counter="IDVS_VAR_SHAD_STALL" title="Tiler Cycles" name="Varying shading stall cycles" description="The number of cycles where the tiler has a stalled varying shading request." units="cycles"/>
+        <event offset="54" advanced="yes" counter="WRBUF_NO_AXI_ID_STALL" title="Tiler Cycles" name="Write buffer transaction stall cycles" description="The number of cycles where the tiler write buffer can not send data because it has no available write IDs." units="cycles"/>
+        <event offset="55" advanced="yes" counter="WRBUF_AXI_STALL" title="Tiler Cycles" name="Write buffer write stall cycles" description="The number of cycles where the tiler write buffer can not send data because the bus is not ready." units="cycles"/>
+    </category>
+    <category name="Memory System" per_cpu="no">
+        <event offset="4" advanced="yes" counter="MMU_REQUESTS" title="MMU Stage 1 Translations" name="MMU lookups" description="The number of main MMU address translations performed." units="requests"/>
+        <event offset="16" advanced="yes" counter="L2_RD_MSG_IN" title="L2 Cache Requests" name="Read requests" description="The number of L2 cache read requests from internal masters." units="requests"/>
+        <event offset="17" advanced="yes" counter="L2_RD_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Read stall cycles" description="The number of cycles L2 cache read requests from internal masters are stalled." units="cycles"/>
+        <event offset="18" advanced="yes" counter="L2_WR_MSG_IN" title="L2 Cache Requests" name="Write requests" description="The number of L2 cache write requests from internal masters." units="requests"/>
+        <event offset="19" advanced="yes" counter="L2_WR_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Write stall cycles" description="The number of cycles where L2 cache write requests from internal masters are stalled." units="cycles"/>
+        <event offset="20" advanced="yes" counter="L2_SNP_MSG_IN" title="L2 Cache Requests" name="Snoop requests" description="The number of L2 snoop requests from internal masters." units="requests"/>
+        <event offset="21" advanced="yes" counter="L2_SNP_MSG_IN_STALL" title="L2 Cache Stall Cycles" name="Snoop stall cycles" description="The number of cycles where L2 cache snoop requests from internal masters are stalled." units="cycles"/>
+        <event offset="22" advanced="yes" counter="L2_RD_MSG_OUT" title="L2 Cache Requests" name="L1 read requests" description="The number of L1 cache read requests sent by the L2 cache to an internal master." units="requests"/>
+        <event offset="23" advanced="yes" counter="L2_RD_MSG_OUT_STALL" title="L2 Cache Stall Cycles" name="L1 read stall cycles" description="The number of cycles where L1 cache read requests sent by the L2 cache to an internal master are stalled." units="cycles"/>
+        <event offset="24" advanced="yes" counter="L2_WR_MSG_OUT" title="L2 Cache Requests" name="L1 write requests" description="The number of L1 cache write responses sent by the L2 cache to an internal master." units="requests"/>
+        <event offset="25" counter="L2_ANY_LOOKUP" title="L2 Cache Lookups" name="Any lookup" description="The number of L2 cache lookups performed." units="requests"/>
+        <event offset="26" counter="L2_READ_LOOKUP" title="L2 Cache Lookups" name="Read lookup" description="The number of L2 cache read lookups performed." units="requests"/>
+        <event offset="27" counter="L2_WRITE_LOOKUP" title="L2 Cache Lookups" name="Write lookup" description="The number of L2 cache write lookups performed." units="requests"/>
+        <event offset="28" advanced="yes" counter="L2_EXT_SNOOP_LOOKUP" title="L2 Cache Lookups" name="External snoop lookups" description="The number of coherency snoop lookups performed that were triggered by an external master." units="requests"/>
+        <event offset="29" counter="L2_EXT_READ" title="External Bus Accesses" name="Read transaction" description="The number of external read transactions." units="transactions"/>
+        <event offset="30" advanced="yes" counter="L2_EXT_READ_NOSNP" title="External Bus Accesses" name="ReadNoSnoop transactions" description="The number of external non-coherent read transactions." units="transactions"/>
+        <event offset="31" advanced="yes" counter="L2_EXT_READ_UNIQUE" title="External Bus Accesses" name="ReadUnique transactions" description="The number of external coherent read unique transactions." units="transactions"/>
+        <event offset="32" counter="L2_EXT_READ_BEATS" title="External Bus Beats" name="Read beat" description="The number of external bus data read cycles." units="beats"/>
+        <event offset="33" counter="L2_EXT_AR_STALL" title="External Bus Stalls" name="Read stall cycles" description="The number of cycles where a read is stalled waiting for the external bus." units="cycles"/>
+        <event offset="34" counter="L2_EXT_AR_CNT_Q1" title="External Bus Outstanding Reads" name="0-25% outstanding" description="The number of read transactions initiated when 0-25% of the maximum are in use." units="transactions"/>
+        <event offset="35" counter="L2_EXT_AR_CNT_Q2" title="External Bus Outstanding Reads" name="25-50% outstanding" description="The number of read transactions initiated when 25-50% of the maximum are in use." units="transactions"/>
+        <event offset="36" counter="L2_EXT_AR_CNT_Q3" title="External Bus Outstanding Reads" name="50-75% outstanding" description="The number of read transactions initiated when 50-75% of the maximum are in use." units="transactions"/>
+        <event offset="37" counter="L2_EXT_RRESP_0_127" title="External Bus Read Latency" name="0-127 cycles" description="The number of data beats returned 0-127 cycles after the read request." units="beats"/>
+        <event offset="38" counter="L2_EXT_RRESP_128_191" title="External Bus Read Latency" name="128-191 cycles" description="The number of data beats returned 128-191 cycles after the read request." units="beats"/>
+        <event offset="39" counter="L2_EXT_RRESP_192_255" title="External Bus Read Latency" name="192-255 cycles" description="The number of data beats returned 192-255 cycles after the read request." units="beats"/>
+        <event offset="40" counter="L2_EXT_RRESP_256_319" title="External Bus Read Latency" name="256-319 cycles" description="The number of data beats returned 256-319 cycles after the read request." units="beats"/>
+        <event offset="41" counter="L2_EXT_RRESP_320_383" title="External Bus Read Latency" name="320-383 cycles" description="The number of data beats returned 320-383 cycles after the read request." units="beats"/>
+        <event offset="42" counter="L2_EXT_WRITE" title="External Bus Accesses" name="Write transaction" description="The number of external write transactions." units="transactions"/>
+        <event offset="43" advanced="yes" counter="L2_EXT_WRITE_NOSNP_FULL" title="External Bus Accesses" name="WriteNoSnoopFull transactions" description="The number of external non-coherent full write transactions." units="transactions"/>
+        <event offset="44" advanced="yes" counter="L2_EXT_WRITE_NOSNP_PTL" title="External Bus Accesses" name="WriteNoSnoopPartial transactions" description="The number of external non-coherent partial write transactions." units="transactions"/>
+        <event offset="45" advanced="yes" counter="L2_EXT_WRITE_SNP_FULL" title="External Bus Accesses" name="WriteSnoopFull transactions" description="The number of external coherent full write transactions." units="transactions"/>
+        <event offset="46" advanced="yes" counter="L2_EXT_WRITE_SNP_PTL" title="External Bus Accesses" name="WriteSnoopPartial transactions" description="The number of external coherent partial write transactions." units="transactions"/>
+        <event offset="47" counter="L2_EXT_WRITE_BEATS" title="External Bus Beats" name="Write beat" description="The number of external bus data write cycles." units="beats"/>
+        <event offset="48" counter="L2_EXT_W_STALL" title="External Bus Stalls" name="Write stall cycles" description="The number of cycles where a write is stalled waiting for the external bus." units="cycles"/>
+        <event offset="49" counter="L2_EXT_AW_CNT_Q1" title="External Bus Outstanding Writes" name="0-25% outstanding" description="The number of write transactions initiated when 0-25% of the maximum are in use." units="transactions"/>
+        <event offset="50" counter="L2_EXT_AW_CNT_Q2" title="External Bus Outstanding Writes" name="25-50% outstanding" description="The number of write transactions initiated when 25-50% of the maximum are in use." units="transactions"/>
+        <event offset="51" counter="L2_EXT_AW_CNT_Q3" title="External Bus Outstanding Writes" name="50-75% outstanding" description="The number of write transactions initiated when 50-75% of the maximum are in use." units="transactions"/>
+        <event offset="52" advanced="yes" counter="L2_EXT_SNOOP" title="External Bus Accesses" name="Snoop transactions" description="The number of coherency snoops triggered by external masters." units="transactions"/>
+        <event offset="53" advanced="yes" counter="L2_EXT_SNOOP_STALL" title="External Bus Stalls" name="Snoop stall cycles" description="The number of cycles where a coherency snoop triggered by external master is stalled." units="cycles"/>
+    </category>
+    <category name="Shader Core" per_cpu="no">
+        <event offset="4" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment active" description="The number of cycles where the shader core is processing a fragment workload." units="cycles"/>
+        <event offset="5" advanced="yes" counter="FRAG_PRIMITIVES_OUT" title="Core Primitives" name="Read primitives" description="The number of primitives read from the tile list by the fragment front-end." units="primitives"/>
+        <event offset="6" counter="FRAG_PRIM_RAST" title="Core Primitives" name="Rasterized primitives" description="The number of primitives being rasterized." units="primitives"/>
+        <event offset="7" counter="FRAG_FPK_ACTIVE" title="Core Cycles" name="Fragment FPKB active" description="The number of cycles where at least one quad is present in the pre-pipe quad queue." units="cycles"/>
+        <event offset="9" counter="FRAG_WARPS" title="Core Warps" name="Fragment warps" description="The number of fragment warps created." units="warps"/>
+        <event offset="10" counter="FRAG_PARTIAL_QUADS_RAST" title="Core Quads" name="Partial rasterized quads" description="The number of partially-rasterized fragment quads created." units="quads"/>
+        <event offset="11" counter="FRAG_QUADS_RAST" title="Core Quads" name="Rasterized quads" description="The number of quads generated by the rasterization phase." units="quads"/>
+        <event offset="12" counter="FRAG_QUADS_EZS_TEST" title="Core Quads" name="Early ZS tested quads" description="The number of quads that are undergoing early depth and stencil testing." units="quads"/>
+        <event offset="13" counter="FRAG_QUADS_EZS_UPDATE" title="Core Quads" name="Early ZS updated quads" description="The number of quads undergoing early depth and stencil testing, that are capable of updating the framebuffer." units="quads"/>
+        <event offset="14" counter="FRAG_QUADS_EZS_KILL" title="Core Quads" name="Early ZS killed quads" description="The number of quads killed by early depth and stencil testing." units="quads"/>
+        <event offset="15" counter="FRAG_LZS_TEST" title="Core Quads" name="Late ZS tested quads" description="The number of quads undergoing late depth and stencil testing." units="quads"/>
+        <event offset="16" counter="FRAG_LZS_KILL" title="Core Quads" name="Late ZS killed quads" description="The number of quads killed by late depth and stencil testing." units="quads"/>
+        <event offset="17" counter="WARP_REG_SIZE_64" title="Core Warps" name="All register warps" description="The number of warps that require more than 32 registers." units="warps"/>
+        <event offset="18" counter="FRAG_PTILES" title="Core Tiles" name="Tiles" description="The number of tiles processed by the shader core." units="tiles"/>
+        <event offset="19" counter="FRAG_TRANS_ELIM" title="Core Tiles" name="Constant tiles killed" description="The number of tiles killed by transaction elimination." units="tiles"/>
+        <event offset="20" counter="QUAD_FPK_KILLER" title="Core Quads" name="FPK occluder quads" description="The number of quads that are valid occluders for hidden surface removal." units="quads"/>
+        <event offset="21" counter="FULL_QUAD_WARPS" title="Core Warps" name="Full quad warps" description="The number of warps that are fully populated with quads." units="warps"/>
+        <event offset="22" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Non-fragment active" description="The number of cycles where the shader core is processing some non-fragment workload." units="cycles"/>
+        <event offset="23" advanced="yes" counter="COMPUTE_TASKS" title="Core Tasks" name="Non-fragment tasks" description="The number of non-fragment tasks issued to the shader core." units="tasks"/>
+        <event offset="24" counter="COMPUTE_WARPS" title="Core Warps" name="Non-fragment warps" description="The number of non-fragment warps created." units="warps"/>
+        <event offset="25" advanced="yes" counter="COMPUTE_STARVING" title="Core Starvation Cycles" name="Non-fragment starvation cycles" description="The number of cycles where the shader core is processing a non-fragment workload and there are no new threads available for execution." units="cycles"/>
+        <event offset="26" counter="EXEC_CORE_ACTIVE" title="Core Cycles" name="Execution core active" description="The number of cycles where the shader core is processing at least one warp." units="cycles"/>
+        <event offset="27" counter="EXEC_INSTR_FMA" title="Core PU Instructions" name="FMA instructions" description="The number of instructions issued to the FMA pipe." units="instructions"/>
+        <event offset="28" counter="EXEC_INSTR_CVT" title="Core PU Instructions" name="CVT instructions" description="The number of instructions issued to the CVT pipe." units="instructions"/>
+        <event offset="29" counter="EXEC_INSTR_SFU" title="Core PU Instructions" name="SFU instructions" description="The number of instructions issued to the SFU pipe." units="instructions"/>
+        <event offset="30" counter="EXEC_INSTR_MSG" title="Core PU Instructions" name="Message instructions" description="The number of instructions issued to the MSG pipe." units="instructions"/>
+        <event offset="31" counter="EXEC_INSTR_DIVERGED" title="Core EE Instructions" name="Diverged instructions" description="The number of instructions executed per warp, that have control flow divergence." units="instructions"/>
+        <event offset="32" advanced="yes" counter="EXEC_ICACHE_MISS" title="Core PU Instructions" name="Instruction cache misses" description="The number of instruction cache misses." units="requests"/>
+        <event offset="33" advanced="yes" counter="EXEC_STARVE_ARITH" title="Core Starvation Cycles" name="Execution engine starvation cycles" description="The number of cycles where the processing unit is starved of work." units="cycles"/>
+        <event offset="34" counter="CALL_BLEND_SHADER" title="Core PU Instructions" name="Blend shader calls" description="The number of blend shader invocations executed." units="instructions"/>
+        <event offset="35" counter="TEX_MSGI_NUM_FLITS" title="Texture Bus" name="Input beats" description="The number of texture request message data beats." units="beats"/>
+        <event offset="36" counter="TEX_DFCH_CLK_STALLED" title="Core Texture Stalls" name="Descriptor stall cycles" description="The number of cycles where a quad is stalled on texture descriptor fetch." units="cycles"/>
+        <event offset="37" counter="TEX_TFCH_CLK_STALLED" title="Core Texture Stalls" name="Fetch queue stall cycles" description="The number of cycles where a quad is stalled on entering texture fetch because the fetch queue is full." units="cycles"/>
+        <event offset="38" counter="TEX_TFCH_STARVED_PENDING_DATA_FETCH" title="Core Texture Stalls" name="Filtering unit stall cycles" description="The number of cycles where the filtering unit is idle and there is at least one quad present in the texture data fetch queue." units="cycles"/>
+        <event offset="39" counter="TEX_FILT_NUM_OPERATIONS" title="Core Texture Cycles" name="Texturing active" description="The number of texture filtering issue cycles." units="cycles"/>
+        <event offset="40" counter="TEX_FILT_NUM_FXR_OPERATIONS" title="Core Texture Cycles" name="4x bilinear filtering active" description="The number of cycles where the filtering unit uses the 4x path to implement nearest or bilinear filtering." units="cycles"/>
+        <event offset="41" counter="TEX_FILT_NUM_FST_OPERATIONS" title="Core Texture Cycles" name="2x trilinear filtering active" description="The number of cycles where the filtering unit uses the 4x path to implement trilinear filtering." units="cycles"/>
+        <event offset="42" counter="TEX_MSGO_NUM_MSG" title="Core Texture Quads" name="Texture requests" description="The number of quad-width texture operations processed by the texture unit." units="quads"/>
+        <event offset="43" counter="TEX_MSGO_NUM_FLITS" title="Texture Bus" name="Output beats" description="The number of texture response message data beats." units="beats"/>
+        <event offset="44" counter="LS_MEM_READ_FULL" title="Core Load/Store Cycles" name="Full read cycles" description="The number of full-width load/store cache reads." units="cycles"/>
+        <event offset="45" counter="LS_MEM_READ_SHORT" title="Core Load/Store Cycles" name="Partial read cycles" description="The number of partial-width load/store cache reads." units="cycles"/>
+        <event offset="46" counter="LS_MEM_WRITE_FULL" title="Core Load/Store Cycles" name="Full write cycles" description="The number of full-width load/store cache writes." units="cycles"/>
+        <event offset="47" counter="LS_MEM_WRITE_SHORT" title="Core Load/Store Cycles" name="Partial write cycles" description="The number of partial-width load/store cache writes." units="cycles"/>
+        <event offset="48" counter="LS_MEM_ATOMIC" title="Core Load/Store Cycles" name="Atomic access cycles" description="The number of load/store atomic accesses." units="cycles"/>
+        <event offset="49" counter="VARY_INSTR" title="Core Varying Requests" name="Interpolation requests" description="The number of warp-width interpolation operations processed by the varying unit." units="instructions"/>
+        <event offset="50" counter="VARY_SLOT_32" title="Core Varying Cycles" name="32-bit interpolation active" description="The number of 32-bit interpolation cycles processed by the varying unit." units="cycles"/>
+        <event offset="51" counter="VARY_SLOT_16" title="Core Varying Cycles" name="16-bit interpolation active" description="The number of 16-bit interpolation cycles processed by the varying unit." units="cycles"/>
+        <event offset="52" advanced="yes" counter="ATTR_INSTR" title="Core Attribute Requests" name="Attribute requests" description="The number of instructions executed by the attribute unit." units="instructions"/>
+        <event offset="53" counter="SHADER_CORE_ACTIVE" title="Core Cycles" name="Any active" description="The number of cycles where the shader core is processing either a non-fragment workload or a fragment workload." units="cycles"/>
+        <event offset="54" counter="BEATS_RD_FTC" title="Core L2 Reads" name="Fragment L2 read beats" description="The number of read beats received by the fixed-function fragment front-end." units="beats"/>
+        <event offset="55" counter="BEATS_RD_FTC_EXT" title="Core L2 Reads" name="Fragment external read beats" description="The number of read beats received by the fixed-function fragment front-end that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="56" counter="BEATS_RD_LSC" title="Core L2 Reads" name="Load/store L2 read beats" description="The number of read beats received by the load/store unit." units="beats"/>
+        <event offset="57" counter="BEATS_RD_LSC_EXT" title="Core L2 Reads" name="Load/store external read beats" description="The number of read beats received by the load/store unit that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="58" counter="BEATS_RD_TEX" title="Core L2 Reads" name="Texture L2 read beats" description="The number of read beats received by the texture unit." units="beats"/>
+        <event offset="59" counter="BEATS_RD_TEX_EXT" title="Core L2 Reads" name="Texture external read beats" description="The number of read beats received by the texture unit that required an external memory access due to an L2 cache miss." units="beats"/>
+        <event offset="60" advanced="yes" counter="BEATS_RD_OTHER" title="Core L2 Reads" name="Other L2 read beats" description="The number of read beats received by a unit that is not specifically identified." units="beats"/>
+        <event offset="61" counter="BEATS_WR_LSC_OTHER" title="Core Writes" name="Load/store other write beats" description="The number of write beats by the load/store unit that are due to any reason other than writeback." units="beats"/>
+        <event offset="62" counter="BEATS_WR_TIB" title="Core Writes" name="Tile buffer write beats" description="The number of write beats sent by the tile buffer writeback unit." units="beats"/>
+        <event offset="63" counter="BEATS_WR_LSC_WB" title="Core Writes" name="Load/store writeback write beats" description="The number of write beats by the load/store unit that are due to writeback." units="beats"/>
     </category>
 </metrics>
\ No newline at end of file
diff --git a/src/panfrost/perf/T72x.xml b/src/panfrost/perf/T72x.xml
index 819b43fec59..16a5f807382 100644
--- a/src/panfrost/perf/T72x.xml
+++ b/src/panfrost/perf/T72x.xml
@@ -1,6 +1,6 @@
 <!--
 Copyright © 2017-2020 ARM Limited.
-Copyright © 2021 Collabora, Ltd.
+Copyright © 2021-2022 Collabora, Ltd.
 Author: Antonio Caggiano <antonio.caggiano@collabora.com>
 
 Permission is hereby granted, free of charge, to any person obtaining a
@@ -24,72 +24,72 @@ SOFTWARE.
 -->
 <metrics id="T72x">
     <category name="Job Manager" per_cpu="no">
-        <event offset="4" counter="GPU_ACTIVE" title="Job Manager Cycles" name="GPU cycles" description="Number of cycles GPU active" units="cycles" />
-        <event offset="5" counter="IRQ_ACTIVE" title="Job Manager Cycles" name="IRQ cycles" description="Number of cycles GPU interrupt pending" units="cycles" />
-        <event offset="8" counter="JS0_ACTIVE" title="Job Manager Cycles" name="JS0 cycles" description="Number of cycles JS0 (fragment) active" units="cycles" />
-        <event offset="11" counter="JS1_ACTIVE" title="Job Manager Cycles" name="JS1 cycles" description="Number of cycles JS1 (vertex/tiler/compute) active" units="cycles" />
-        <event offset="14" counter="JS2_ACTIVE" title="Job Manager Cycles" name="JS2 cycles" description="Number of cycles JS2 (vertex/compute) active" units="cycles" />
-        <event offset="6" counter="JS0_JOBS" title="Job Manager Jobs" name="JS0 jobs" description="Number of Jobs (fragment) completed in JS0" units="jobs" />
-        <event offset="9" counter="JS1_JOBS" title="Job Manager Jobs" name="JS1 jobs" description="Number of Jobs (vertex/tiler/compute) completed in JS1" units="jobs" />
-        <event offset="12" counter="JS2_JOBS" title="Job Manager Jobs" name="JS2 jobs" description="Number of Jobs (vertex/compute) completed in JS2" units="jobs" />
-        <event offset="7" counter="JS0_TASKS" title="Job Manager" name="Pixels" description="Number of Pixels completed in JS0" multiplier="256" units="pixels" />
-        <event offset="10" counter="JS1_TASKS" title="Job Manager Tasks" name="JS1 tasks" description="Number of Tasks completed in JS1" units="tasks" />
-        <event offset="13" counter="JS2_TASKS" title="Job Manager Tasks" name="JS2 tasks" description="Number of Tasks completed in JS2" units="tasks" />
+        <event offset="4" counter="GPU_ACTIVE" title="Job Manager Cycles" name="GPU cycles" description="Number of cycles GPU active" units="cycles"/>
+        <event offset="5" counter="IRQ_ACTIVE" title="Job Manager Cycles" name="IRQ cycles" description="Number of cycles GPU interrupt pending" units="cycles"/>
+        <event offset="8" counter="JS0_ACTIVE" title="Job Manager Cycles" name="JS0 cycles" description="Number of cycles JS0 (fragment) active" units="cycles"/>
+        <event offset="11" counter="JS1_ACTIVE" title="Job Manager Cycles" name="JS1 cycles" description="Number of cycles JS1 (vertex/tiler/compute) active" units="cycles"/>
+        <event offset="14" counter="JS2_ACTIVE" title="Job Manager Cycles" name="JS2 cycles" description="Number of cycles JS2 (vertex/compute) active" units="cycles"/>
+        <event offset="6" counter="JS0_JOBS" title="Job Manager Jobs" name="JS0 jobs" description="Number of Jobs (fragment) completed in JS0" units="jobs"/>
+        <event offset="9" counter="JS1_JOBS" title="Job Manager Jobs" name="JS1 jobs" description="Number of Jobs (vertex/tiler/compute) completed in JS1" units="jobs"/>
+        <event offset="12" counter="JS2_JOBS" title="Job Manager Jobs" name="JS2 jobs" description="Number of Jobs (vertex/compute) completed in JS2" units="jobs"/>
+        <event offset="7" counter="JS0_TASKS" title="Job Manager" name="Pixels" description="Number of Pixels completed in JS0" multiplier="256" units="pixels"/>
+        <event offset="10" counter="JS1_TASKS" title="Job Manager Tasks" name="JS1 tasks" description="Number of Tasks completed in JS1" units="tasks"/>
+        <event offset="13" counter="JS2_TASKS" title="Job Manager Tasks" name="JS2 tasks" description="Number of Tasks completed in JS2" units="tasks"/>
     </category>
     <category name="Tiler" per_cpu="no">
-        <event offset="86" counter="TI_ACTIVE" title="Tiler Cycles" name="Tiler cycles" description="Number of cycles Tiler active" units="cycles" />
-        <event offset="70" counter="TI_POLYGONS" title="Tiler Primitives" name="Polygons" description="Number of polygons processed" units="primitives" />
-        <event offset="69" counter="TI_QUADS" title="Tiler Primitives" name="Quads" description="Number of quads processed" units="primitives" />
-        <event offset="68" counter="TI_TRIANGLES" title="Tiler Primitives" name="Triangles" description="Number of triangles processed" units="primitives" />
-        <event offset="72" counter="TI_LINES" title="Tiler Primitives" name="Lines" description="Number of lines processed" units="primitives" />
-        <event offset="71" counter="TI_POINTS" title="Tiler Primitives" name="Points" description="Number of points processed" units="primitives" />
-        <event offset="73" counter="TI_FRONT_FACING" title="Tiler Culling" name="Front facing prims" description="Number of front facing primitives" units="primitives" />
-        <event offset="74" counter="TI_BACK_FACING" title="Tiler Culling" name="Back facing prims" description="Number of back facing primitives" units="primitives" />
-        <event offset="75" counter="TI_PRIM_VISIBLE" title="Tiler Culling" name="Visible prims" description="Number of visible primitives" units="primitives" />
-        <event offset="76" counter="TI_PRIM_CULLED" title="Tiler Culling" name="Culled prims" description="Number of culled primitives" units="primitives" />
-        <event offset="77" counter="TI_PRIM_CLIPPED" title="Tiler Culling" name="Clipped prims" description="Number of clipped primitives" units="primitives" />
-    </category>
-    <category name="Shader Core" per_cpu="no">
-        <event offset="212" counter="TRIPIPE_ACTIVE" title="Core Cycles" name="Tripipe cycles" description="Number of cycles tripipe was active" units="cycles" />
-        <event offset="196" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment cycles" description="Number of cycles fragment processing was active" units="cycles" />
-        <event offset="209" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Compute cycles" description="Number of cycles vertex\compute processing was active" units="cycles" />
-        <event offset="206" counter="FRAG_CYCLES_NO_TILE" title="Core Cycles" name="Fragment cycles waiting for tile" description="Number of cycles spent waiting for a physical tile buffer" units="cycles" />
-        <event offset="199" counter="FRAG_THREADS" title="Fragment Threads" name="Fragment threads" description="Number of fragment threads started" units="threads" />
-        <event offset="200" counter="FRAG_DUMMY_THREADS" title="Fragment Threads" name="Dummy fragment threads" description="Number of dummy fragment threads started" units="threads" />
-        <event offset="204" counter="FRAG_THREADS_LZS_TEST" title="Fragment Threads" name="Fragment threads doing late ZS" description="Number of threads doing late ZS test" units="threads" />
-        <event offset="205" counter="FRAG_THREADS_LZS_KILLED" title="Fragment Threads" name="Fragment threads killed late ZS" description="Number of threads killed by late ZS test" units="threads" />
-        <event offset="210" counter="COMPUTE_TASKS" title="Compute Tasks" name="Compute tasks" description="Number of compute tasks" units="tasks" />
-        <event offset="211" counter="COMPUTE_THREADS" title="Compute Threads" name="Compute threads" description="Number of compute threads started" units="threads" />
-        <event offset="197" counter="FRAG_PRIMITIVES" title="Fragment Primitives" name="Primitives loaded" description="Number of primitives loaded from tiler" units="primitives" />
-        <event offset="198" counter="FRAG_PRIMITIVES_DROPPED" title="Fragment Primitives" name="Primitives dropped" description="Number of primitives dropped because out of tile" units="primitives" />
-        <event offset="201" counter="FRAG_QUADS_RAST" title="Fragment Quads" name="Quads rasterized" description="Number of quads rasterized" units="quads" />
-        <event offset="202" counter="FRAG_QUADS_EZS_TEST" title="Fragment Quads" name="Quads doing early ZS" description="Number of quads doing early ZS test" units="quads" />
-        <event offset="203" counter="FRAG_QUADS_EZS_KILLED" title="Fragment Quads" name="Quads killed early Z" description="Number of quads killed by early ZS test" units="quads" />
-        <event offset="207" counter="FRAG_NUM_TILES" title="Fragment" name="Pixels" description="Number of pixels rendered" multiplier="256" units="pixels" />
-        <event offset="208" counter="FRAG_TRANS_ELIM" title="Fragment Tasks" name="Tile writes killed by TE" description="Number of tile writes skipped by transaction elimination" units="tiles" />
-        <event offset="213" counter="ARITH_WORDS" title="Arithmetic Pipe" name="A instructions" description="Number of batched instructions executed by the A-pipe" units="instructions" />
-        <event offset="215" counter="LS_WORDS" title="Load/Store Pipe" name="LS instructions" description="Number of instructions completed by the LS-pipe" units="instructions" />
-        <event offset="216" counter="LS_ISSUES" title="Load/Store Pipe" name="LS instruction issues" description="Number of instructions issued to the LS-pipe, including restarts" units="instructions" />
-        <event offset="219" counter="TEX_WORDS" title="Texture Pipe" name="T instructions" description="Number of instructions completed by the T-pipe" units="instructions" />
-        <event offset="221" counter="TEX_ISSUES" title="Texture Pipe" name="T instruction issues" description="Number of threads through loop 2 address calculation" units="instructions" />
-        <event offset="222" counter="LSC_READ_HITS" title="Load/Store Cache Reads" name="Read hits" description="Number of read hits in the Load/Store cache" units="requests" />
-        <event offset="223" counter="LSC_READ_MISSES" title="Load/Store Cache Reads" name="Read misses" description="Number of read misses in the Load/Store cache" units="requests" />
-        <event offset="224" counter="LSC_WRITE_HITS" title="Load/Store Cache Writes" name="Write hits" description="Number of write hits in the Load/Store cache" units="requests" />
-        <event offset="225" counter="LSC_WRITE_MISSES" title="Load/Store Cache Writes" name="Write misses" description="Number of write misses in the Load/Store cache" units="requests" />
-        <event offset="226" counter="LSC_ATOMIC_HITS" title="Load/Store Cache Atomics" name="Atomic hits" description="Number of atomic hits in the Load/Store cache" units="requests" />
-        <event offset="227" counter="LSC_ATOMIC_MISSES" title="Load/Store Cache Atomics" name="Atomic misses" description="Number of atomic misses in the Load/Store cache" units="requests" />
-        <event offset="228" counter="LSC_LINE_FETCHES" title="Load/Store Cache Bus" name="Line fetches" description="Number of line fetches in the Load/Store cache" units="requests" />
-        <event offset="229" counter="LSC_DIRTY_LINE" title="Load/Store Cache Bus" name="Dirty line evictions" description="Number of dirty line evictions in the Load/Store cache" units="requests" />
-        <event offset="230" counter="LSC_SNOOPS" title="Load/Store Cache Bus" name="Snoops in to LSC" description="Number of coherent memory snoops in to the Load/Store cache" units="requests" />
+        <event offset="22" counter="TI_ACTIVE" title="Tiler Cycles" name="Tiler cycles" description="Number of cycles Tiler active" units="cycles"/>
+        <event offset="6" counter="TI_POLYGONS" title="Tiler Primitives" name="Polygons" description="Number of polygons processed" units="primitives"/>
+        <event offset="5" counter="TI_QUADS" title="Tiler Primitives" name="Quads" description="Number of quads processed" units="primitives"/>
+        <event offset="4" counter="TI_TRIANGLES" title="Tiler Primitives" name="Triangles" description="Number of triangles processed" units="primitives"/>
+        <event offset="8" counter="TI_LINES" title="Tiler Primitives" name="Lines" description="Number of lines processed" units="primitives"/>
+        <event offset="7" counter="TI_POINTS" title="Tiler Primitives" name="Points" description="Number of points processed" units="primitives"/>
+        <event offset="9" counter="TI_FRONT_FACING" title="Tiler Culling" name="Front facing prims" description="Number of front facing primitives" units="primitives"/>
+        <event offset="10" counter="TI_BACK_FACING" title="Tiler Culling" name="Back facing prims" description="Number of back facing primitives" units="primitives"/>
+        <event offset="11" counter="TI_PRIM_VISIBLE" title="Tiler Culling" name="Visible prims" description="Number of visible primitives" units="primitives"/>
+        <event offset="12" counter="TI_PRIM_CULLED" title="Tiler Culling" name="Culled prims" description="Number of culled primitives" units="primitives"/>
+        <event offset="13" counter="TI_PRIM_CLIPPED" title="Tiler Culling" name="Clipped prims" description="Number of clipped primitives" units="primitives"/>
     </category>
     <category name="L2 Cache" per_cpu="no">
-        <event offset="134" counter="L2_READ_SNOOP" title="L2 Cache Reads" name="Read snoops" description="Number of read transaction snoops" units="requests" />
-        <event offset="135" counter="L2_READ_HIT" title="L2 Cache Reads" name="L2 read hits" description="Number of reads hitting in the L2 cache" units="requests" />
-        <event offset="136" counter="L2_WRITE_SNOOP" title="L2 Cache Writes" name="Write snoops" description="Number of write transaction snoops" units="requests" />
-        <event offset="137" counter="L2_WRITE_HIT" title="L2 Cache Writes" name="L2 write hits" description="Number of writes hitting in the L2 cache" units="requests" />
-        <event offset="133" counter="L2_EXT_READ_BEAT" title="L2 Cache Ext Reads" name="External read bytes" description="Number of external bus read bytes" multiplier="16" units="bytes" />
-        <event offset="140" counter="L2_EXT_AR_STALL" title="L2 Cache Ext Reads" name="External bus stalls (AR)" description="Number of cycles a valid read address (AR) is stalled by the external interconnect" units="cycles" />
-        <event offset="132" counter="L2_EXT_WRITE_BEAT" title="L2 Cache Ext Writes" name="External write bytes" description="Number of external bus write bytes" multiplier="16" units="bytes" />
-        <event offset="141" counter="L2_EXT_W_STALL" title="L2 Cache Ext Writes" name="External bus stalls (W)" description="Number of cycles a valid write data (W channel) is stalled by the external interconnect" units="cycles" />
+        <event offset="6" counter="L2_READ_SNOOP" title="L2 Cache Reads" name="Read snoops" description="Number of read transaction snoops" units="requests"/>
+        <event offset="7" counter="L2_READ_HIT" title="L2 Cache Reads" name="L2 read hits" description="Number of reads hitting in the L2 cache" units="requests"/>
+        <event offset="8" counter="L2_WRITE_SNOOP" title="L2 Cache Writes" name="Write snoops" description="Number of write transaction snoops" units="requests"/>
+        <event offset="9" counter="L2_WRITE_HIT" title="L2 Cache Writes" name="L2 write hits" description="Number of writes hitting in the L2 cache" units="requests"/>
+        <event offset="5" counter="L2_EXT_READ_BEAT" title="L2 Cache Ext Reads" name="External read bytes" description="Number of external bus read bytes" multiplier="16" units="bytes"/>
+        <event offset="12" counter="L2_EXT_AR_STALL" title="L2 Cache Ext Reads" name="External bus stalls (AR)" description="Number of cycles a valid read address (AR) is stalled by the external interconnect" units="cycles"/>
+        <event offset="4" counter="L2_EXT_WRITE_BEAT" title="L2 Cache Ext Writes" name="External write bytes" description="Number of external bus write bytes" multiplier="16" units="bytes"/>
+        <event offset="13" counter="L2_EXT_W_STALL" title="L2 Cache Ext Writes" name="External bus stalls (W)" description="Number of cycles a valid write data (W channel) is stalled by the external interconnect" units="cycles"/>
+    </category>
+    <category name="Shader Core" per_cpu="no">
+        <event offset="20" counter="TRIPIPE_ACTIVE" title="Core Cycles" name="Tripipe cycles" description="Number of cycles tripipe was active" units="cycles"/>
+        <event offset="4" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment cycles" description="Number of cycles fragment processing was active" units="cycles"/>
+        <event offset="17" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Compute cycles" description="Number of cycles vertex\compute processing was active" units="cycles"/>
+        <event offset="14" counter="FRAG_CYCLES_NO_TILE" title="Core Cycles" name="Fragment cycles waiting for tile" description="Number of cycles spent waiting for a physical tile buffer" units="cycles"/>
+        <event offset="7" counter="FRAG_THREADS" title="Fragment Threads" name="Fragment threads" description="Number of fragment threads started" units="threads"/>
+        <event offset="8" counter="FRAG_DUMMY_THREADS" title="Fragment Threads" name="Dummy fragment threads" description="Number of dummy fragment threads started" units="threads"/>
+        <event offset="12" counter="FRAG_THREADS_LZS_TEST" title="Fragment Threads" name="Fragment threads doing late ZS" description="Number of threads doing late ZS test" units="threads"/>
+        <event offset="13" counter="FRAG_THREADS_LZS_KILLED" title="Fragment Threads" name="Fragment threads killed late ZS" description="Number of threads killed by late ZS test" units="threads"/>
+        <event offset="18" counter="COMPUTE_TASKS" title="Compute Tasks" name="Compute tasks" description="Number of compute tasks" units="tasks"/>
+        <event offset="19" counter="COMPUTE_THREADS" title="Compute Threads" name="Compute threads" description="Number of compute threads started" units="threads"/>
+        <event offset="5" counter="FRAG_PRIMITIVES" title="Fragment Primitives" name="Primitives loaded" description="Number of primitives loaded from tiler" units="primitives"/>
+        <event offset="6" counter="FRAG_PRIMITIVES_DROPPED" title="Fragment Primitives" name="Primitives dropped" description="Number of primitives dropped because out of tile" units="primitives"/>
+        <event offset="9" counter="FRAG_QUADS_RAST" title="Fragment Quads" name="Quads rasterized" description="Number of quads rasterized" units="quads"/>
+        <event offset="10" counter="FRAG_QUADS_EZS_TEST" title="Fragment Quads" name="Quads doing early ZS" description="Number of quads doing early ZS test" units="quads"/>
+        <event offset="11" counter="FRAG_QUADS_EZS_KILLED" title="Fragment Quads" name="Quads killed early Z" description="Number of quads killed by early ZS test" units="quads"/>
+        <event offset="15" counter="FRAG_NUM_TILES" title="Fragment" name="Pixels" description="Number of pixels rendered" multiplier="256" units="pixels"/>
+        <event offset="16" counter="FRAG_TRANS_ELIM" title="Fragment Tasks" name="Tile writes killed by TE" description="Number of tile writes skipped by transaction elimination" units="tiles"/>
+        <event offset="21" counter="ARITH_WORDS" title="Arithmetic Pipe" name="A instructions" description="Number of batched instructions executed by the A-pipe" units="instructions"/>
+        <event offset="23" counter="LS_WORDS" title="Load/Store Pipe" name="LS instructions" description="Number of instructions completed by the LS-pipe" units="instructions"/>
+        <event offset="24" counter="LS_ISSUES" title="Load/Store Pipe" name="LS instruction issues" description="Number of instructions issued to the LS-pipe, including restarts" units="instructions"/>
+        <event offset="27" counter="TEX_WORDS" title="Texture Pipe" name="T instructions" description="Number of instructions completed by the T-pipe" units="instructions"/>
+        <event offset="29" counter="TEX_ISSUES" title="Texture Pipe" name="T instruction issues" description="Number of threads through loop 2 address calculation" units="instructions"/>
+        <event offset="30" counter="LSC_READ_HITS" title="Load/Store Cache Reads" name="Read hits" description="Number of read hits in the Load/Store cache" units="requests"/>
+        <event offset="31" counter="LSC_READ_MISSES" title="Load/Store Cache Reads" name="Read misses" description="Number of read misses in the Load/Store cache" units="requests"/>
+        <event offset="32" counter="LSC_WRITE_HITS" title="Load/Store Cache Writes" name="Write hits" description="Number of write hits in the Load/Store cache" units="requests"/>
+        <event offset="33" counter="LSC_WRITE_MISSES" title="Load/Store Cache Writes" name="Write misses" description="Number of write misses in the Load/Store cache" units="requests"/>
+        <event offset="34" counter="LSC_ATOMIC_HITS" title="Load/Store Cache Atomics" name="Atomic hits" description="Number of atomic hits in the Load/Store cache" units="requests"/>
+        <event offset="35" counter="LSC_ATOMIC_MISSES" title="Load/Store Cache Atomics" name="Atomic misses" description="Number of atomic misses in the Load/Store cache" units="requests"/>
+        <event offset="36" counter="LSC_LINE_FETCHES" title="Load/Store Cache Bus" name="Line fetches" description="Number of line fetches in the Load/Store cache" units="requests"/>
+        <event offset="37" counter="LSC_DIRTY_LINE" title="Load/Store Cache Bus" name="Dirty line evictions" description="Number of dirty line evictions in the Load/Store cache" units="requests"/>
+        <event offset="38" counter="LSC_SNOOPS" title="Load/Store Cache Bus" name="Snoops in to LSC" description="Number of coherent memory snoops in to the Load/Store cache" units="requests"/>
     </category>
 </metrics>
\ No newline at end of file
diff --git a/src/panfrost/perf/T76x.xml b/src/panfrost/perf/T76x.xml
index 344d18f49ab..521ba49a49b 100644
--- a/src/panfrost/perf/T76x.xml
+++ b/src/panfrost/perf/T76x.xml
@@ -1,6 +1,6 @@
 <!--
 Copyright © 2017-2020 ARM Limited.
-Copyright © 2021 Collabora, Ltd.
+Copyright © 2021-2022 Collabora, Ltd.
 Author: Antonio Caggiano <antonio.caggiano@collabora.com>
 
 Permission is hereby granted, free of charge, to any person obtaining a
@@ -24,87 +24,87 @@ SOFTWARE.
 -->
 <metrics id="T76x">
     <category name="Job Manager" per_cpu="no">
-        <event offset="6" counter="GPU_ACTIVE" title="Job Manager Cycles" name="GPU cycles" description="Number of cycles GPU active" units="cycles" />
-        <event offset="7" counter="IRQ_ACTIVE" title="Job Manager Cycles" name="IRQ cycles" description="Number of cycles GPU interrupt pending" units="cycles" />
-        <event offset="10" counter="JS0_ACTIVE" title="Job Manager Cycles" name="JS0 cycles" description="Number of cycles JS0 (fragment) active" units="cycles" />
-        <event offset="18" counter="JS1_ACTIVE" title="Job Manager Cycles" name="JS1 cycles" description="Number of cycles JS1 (vertex/tiler/compute) active" units="cycles" />
-        <event offset="26" counter="JS2_ACTIVE" title="Job Manager Cycles" name="JS2 cycles" description="Number of cycles JS2 (vertex/compute) active" units="cycles" />
-        <event offset="8" counter="JS0_JOBS" title="Job Manager Jobs" name="JS0 jobs" description="Number of Jobs (fragment) completed in JS0" units="jobs" />
-        <event offset="16" counter="JS1_JOBS" title="Job Manager Jobs" name="JS1 jobs" description="Number of Jobs (vertex/tiler/compute) completed in JS1" units="jobs" />
-        <event offset="24" counter="JS2_JOBS" title="Job Manager Jobs" name="JS2 jobs" description="Number of Jobs (vertex/compute) completed in JS2" units="jobs" />
-        <event offset="9" counter="JS0_TASKS" title="Job Manager" name="Pixels" description="Number of Pixels completed in JS0" multiplier="1024" units="pixels" />
-        <event offset="17" counter="JS1_TASKS" title="Job Manager Tasks" name="JS1 tasks" description="Number of Tasks completed in JS1" units="tasks" />
-        <event offset="25" counter="JS2_TASKS" title="Job Manager Tasks" name="JS2 tasks" description="Number of Tasks completed in JS2" units="tasks" />
+        <event offset="6" counter="GPU_ACTIVE" title="Job Manager Cycles" name="GPU cycles" description="Number of cycles GPU active" units="cycles"/>
+        <event offset="7" counter="IRQ_ACTIVE" title="Job Manager Cycles" name="IRQ cycles" description="Number of cycles GPU interrupt pending" units="cycles"/>
+        <event offset="10" counter="JS0_ACTIVE" title="Job Manager Cycles" name="JS0 cycles" description="Number of cycles JS0 (fragment) active" units="cycles"/>
+        <event offset="18" counter="JS1_ACTIVE" title="Job Manager Cycles" name="JS1 cycles" description="Number of cycles JS1 (vertex/tiler/compute) active" units="cycles"/>
+        <event offset="26" counter="JS2_ACTIVE" title="Job Manager Cycles" name="JS2 cycles" description="Number of cycles JS2 (vertex/compute) active" units="cycles"/>
+        <event offset="8" counter="JS0_JOBS" title="Job Manager Jobs" name="JS0 jobs" description="Number of Jobs (fragment) completed in JS0" units="jobs"/>
+        <event offset="16" counter="JS1_JOBS" title="Job Manager Jobs" name="JS1 jobs" description="Number of Jobs (vertex/tiler/compute) completed in JS1" units="jobs"/>
+        <event offset="24" counter="JS2_JOBS" title="Job Manager Jobs" name="JS2 jobs" description="Number of Jobs (vertex/compute) completed in JS2" units="jobs"/>
+        <event offset="9" counter="JS0_TASKS" title="Job Manager" name="Pixels" description="Number of Pixels completed in JS0" multiplier="1024" units="pixels"/>
+        <event offset="17" counter="JS1_TASKS" title="Job Manager Tasks" name="JS1 tasks" description="Number of Tasks completed in JS1" units="tasks"/>
+        <event offset="25" counter="JS2_TASKS" title="Job Manager Tasks" name="JS2 tasks" description="Number of Tasks completed in JS2" units="tasks"/>
     </category>
     <category name="Tiler" per_cpu="no">
-        <event offset="109" counter="TI_ACTIVE" title="Tiler Cycles" name="Tiler cycles" description="Number of cycles Tiler active" units="cycles" />
-        <event offset="70" counter="TI_POLYGONS" title="Tiler Primitives" name="Polygons" description="Number of polygons processed" units="primitives" />
-        <event offset="69" counter="TI_QUADS" title="Tiler Primitives" name="Quads" description="Number of quads processed" units="primitives" />
-        <event offset="68" counter="TI_TRIANGLES" title="Tiler Primitives" name="Triangles" description="Number of triangles processed" units="primitives" />
-        <event offset="72" counter="TI_LINES" title="Tiler Primitives" name="Lines" description="Number of lines processed" units="primitives" />
-        <event offset="71" counter="TI_POINTS" title="Tiler Primitives" name="Points" description="Number of points processed" units="primitives" />
-        <event offset="75" counter="TI_FRONT_FACING" title="Tiler Culling" name="Front facing prims" description="Number of front facing primitives" units="primitives" />
-        <event offset="76" counter="TI_BACK_FACING" title="Tiler Culling" name="Back facing prims" description="Number of back facing primitives" units="primitives" />
-        <event offset="77" counter="TI_PRIM_VISIBLE" title="Tiler Culling" name="Visible prims" description="Number of visible primitives" units="primitives" />
-        <event offset="78" counter="TI_PRIM_CULLED" title="Tiler Culling" name="Culled prims" description="Number of culled primitives" units="primitives" />
-        <event offset="79" counter="TI_PRIM_CLIPPED" title="Tiler Culling" name="Clipped prims" description="Number of clipped primitives" units="primitives" />
-        <event offset="80" counter="TI_LEVEL0" title="Tiler Hierarchy" name="L0 prims" description="Number of primitives in hierarchy level 0" units="primitives" />
-        <event offset="81" counter="TI_LEVEL1" title="Tiler Hierarchy" name="L1 prims" description="Number of primitives in hierarchy level 1" units="primitives" />
-        <event offset="82" counter="TI_LEVEL2" title="Tiler Hierarchy" name="L2 prims" description="Number of primitives in hierarchy level 2" units="primitives" />
-        <event offset="83" counter="TI_LEVEL3" title="Tiler Hierarchy" name="L3 prims" description="Number of primitives in hierarchy level 3" units="primitives" />
-        <event offset="84" counter="TI_LEVEL4" title="Tiler Hierarchy" name="L4 prims" description="Number of primitives in hierarchy level 4" units="primitives" />
-        <event offset="85" counter="TI_LEVEL5" title="Tiler Hierarchy" name="L5 prims" description="Number of primitives in hierarchy level 5" units="primitives" />
-        <event offset="86" counter="TI_LEVEL6" title="Tiler Hierarchy" name="L6 prims" description="Number of primitives in hierarchy level 6" units="primitives" />
-        <event offset="87" counter="TI_LEVEL7" title="Tiler Hierarchy" name="L7 prims" description="Number of primitives in hierarchy level 7" units="primitives" />
-    </category>
-    <category name="Shader Core" per_cpu="no">
-        <event offset="218" counter="TRIPIPE_ACTIVE" title="Core Cycles" name="Tripipe cycles" description="Number of cycles tripipe was active" units="cycles" />
-        <event offset="196" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment cycles" description="Number of cycles fragment processing was active" units="cycles" />
-        <event offset="214" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Compute cycles" description="Number of cycles vertex\compute processing was active" units="cycles" />
-        <event offset="211" counter="FRAG_CYCLES_NO_TILE" title="Core Cycles" name="Fragment cycles waiting for tile" description="Number of cycles spent waiting for a physical tile buffer" units="cycles" />
-        <event offset="200" counter="FRAG_CYCLES_FPKQ_ACTIVE" title="Core Cycles" name="Fragment cycles pre-pipe buffer not empty" description="Number of cycles the pre-pipe queue contains quads" units="cycles" />
-        <event offset="204" counter="FRAG_THREADS" title="Fragment Threads" name="Fragment threads" description="Number of fragment threads started" units="threads" />
-        <event offset="205" counter="FRAG_DUMMY_THREADS" title="Fragment Threads" name="Dummy fragment threads" description="Number of dummy fragment threads started" units="threads" />
-        <event offset="209" counter="FRAG_THREADS_LZS_TEST" title="Fragment Threads" name="Fragment threads doing late ZS" description="Number of threads doing late ZS test" units="threads" />
-        <event offset="210" counter="FRAG_THREADS_LZS_KILLED" title="Fragment Threads" name="Fragment threads killed late ZS" description="Number of threads killed by late ZS test" units="threads" />
-        <event offset="215" counter="COMPUTE_TASKS" title="Compute Tasks" name="Compute tasks" description="Number of compute tasks" units="tasks" />
-        <event offset="216" counter="COMPUTE_THREADS" title="Compute Threads" name="Compute threads" description="Number of compute threads started" units="threads" />
-        <event offset="197" counter="FRAG_PRIMITIVES" title="Fragment Primitives" name="Primitives loaded" description="Number of primitives loaded from tiler" units="primitives" />
-        <event offset="198" counter="FRAG_PRIMITIVES_DROPPED" title="Fragment Primitives" name="Primitives dropped" description="Number of primitives dropped because out of tile" units="primitives" />
-        <event offset="206" counter="FRAG_QUADS_RAST" title="Fragment Quads" name="Quads rasterized" description="Number of quads rasterized" units="quads" />
-        <event offset="207" counter="FRAG_QUADS_EZS_TEST" title="Fragment Quads" name="Quads doing early ZS" description="Number of quads doing early ZS test" units="quads" />
-        <event offset="208" counter="FRAG_QUADS_EZS_KILLED" title="Fragment Quads" name="Quads killed early Z" description="Number of quads killed by early ZS test" units="quads" />
-        <event offset="212" counter="FRAG_NUM_TILES" title="Fragment" name="Pixels" description="Number of pixels rendered" multiplier="1024" units="pixels" />
-        <event offset="213" counter="FRAG_TRANS_ELIM" title="Fragment Tasks" name="Tile writes killed by TE" description="Number of tile writes skipped by transaction elimination" units="tiles" />
-        <event offset="219" counter="ARITH_WORDS" title="Arithmetic Pipe" name="A instructions" description="Number of instructions completed by the the A-pipe (normalized per pipeline)" units="instructions" />
-        <event offset="223" counter="LS_WORDS" title="Load/Store Pipe" name="LS instructions" description="Number of instructions completed by the LS-pipe" units="instructions" />
-        <event offset="224" counter="LS_ISSUES" title="Load/Store Pipe" name="LS instruction issues" description="Number of instructions issued to the LS-pipe, including restarts" units="instructions" />
-        <event offset="230" counter="TEX_WORDS" title="Texture Pipe" name="T instructions" description="Number of instructions completed by the T-pipe" units="instructions" />
-        <event offset="234" counter="TEX_ISSUES" title="Texture Pipe" name="T instruction issues" description="Number of threads through loop 2 address calculation" units="instructions" />
-        <event offset="235" counter="TEX_RECIRC_FMISS" title="Texture Pipe" name="Cache misses" description="Number of instructions in the T-pipe, recirculated due to cache miss" units="requests" />
-        <event offset="241" counter="LSC_READ_OP" title="Load/Store Cache Reads" name="Read operations" description="Number of read operations in the Load/Store cache" units="requests" />
-        <event offset="240" counter="LSC_READ_HITS" title="Load/Store Cache Reads" name="Read hits" description="Number of read hits in the Load/Store cache" units="requests" />
-        <event offset="243" counter="LSC_WRITE_OP" title="Load/Store Cache Writes" name="Write operations" description="Number of write operations in the Load/Store cache" units="requests" />
-        <event offset="242" counter="LSC_WRITE_HITS" title="Load/Store Cache Writes" name="Write hits" description="Number of write hits in the Load/Store cache" units="requests" />
-        <event offset="245" counter="LSC_ATOMIC_OP" title="Load/Store Cache Atomics" name="Atomic operations" description="Number of atomic operations in the Load/Store cache" units="requests" />
-        <event offset="244" counter="LSC_ATOMIC_HITS" title="Load/Store Cache Atomics" name="Atomic hits" description="Number of atomic hits in the Load/Store cache" units="requests" />
-        <event offset="246" counter="LSC_LINE_FETCHES" title="Load/Store Cache Bus" name="Line fetches" description="Number of line fetches in the Load/Store cache" units="requests" />
-        <event offset="247" counter="LSC_DIRTY_LINE" title="Load/Store Cache Bus" name="Dirty line evictions" description="Number of dirty line evictions in the Load/Store cache" units="requests" />
-        <event offset="248" counter="LSC_SNOOPS" title="Load/Store Cache Bus" name="Snoops in to LSC" description="Number of coherent memory snoops in to the Load/Store cache" units="requests" />
+        <event offset="45" counter="TI_ACTIVE" title="Tiler Cycles" name="Tiler cycles" description="Number of cycles Tiler active" units="cycles"/>
+        <event offset="6" counter="TI_POLYGONS" title="Tiler Primitives" name="Polygons" description="Number of polygons processed" units="primitives"/>
+        <event offset="5" counter="TI_QUADS" title="Tiler Primitives" name="Quads" description="Number of quads processed" units="primitives"/>
+        <event offset="4" counter="TI_TRIANGLES" title="Tiler Primitives" name="Triangles" description="Number of triangles processed" units="primitives"/>
+        <event offset="8" counter="TI_LINES" title="Tiler Primitives" name="Lines" description="Number of lines processed" units="primitives"/>
+        <event offset="7" counter="TI_POINTS" title="Tiler Primitives" name="Points" description="Number of points processed" units="primitives"/>
+        <event offset="11" counter="TI_FRONT_FACING" title="Tiler Culling" name="Front facing prims" description="Number of front facing primitives" units="primitives"/>
+        <event offset="12" counter="TI_BACK_FACING" title="Tiler Culling" name="Back facing prims" description="Number of back facing primitives" units="primitives"/>
+        <event offset="13" counter="TI_PRIM_VISIBLE" title="Tiler Culling" name="Visible prims" description="Number of visible primitives" units="primitives"/>
+        <event offset="14" counter="TI_PRIM_CULLED" title="Tiler Culling" name="Culled prims" description="Number of culled primitives" units="primitives"/>
+        <event offset="15" counter="TI_PRIM_CLIPPED" title="Tiler Culling" name="Clipped prims" description="Number of clipped primitives" units="primitives"/>
+        <event offset="16" counter="TI_LEVEL0" title="Tiler Hierarchy" name="L0 prims" description="Number of primitives in hierarchy level 0" units="primitives"/>
+        <event offset="17" counter="TI_LEVEL1" title="Tiler Hierarchy" name="L1 prims" description="Number of primitives in hierarchy level 1" units="primitives"/>
+        <event offset="18" counter="TI_LEVEL2" title="Tiler Hierarchy" name="L2 prims" description="Number of primitives in hierarchy level 2" units="primitives"/>
+        <event offset="19" counter="TI_LEVEL3" title="Tiler Hierarchy" name="L3 prims" description="Number of primitives in hierarchy level 3" units="primitives"/>
+        <event offset="20" counter="TI_LEVEL4" title="Tiler Hierarchy" name="L4 prims" description="Number of primitives in hierarchy level 4" units="primitives"/>
+        <event offset="21" counter="TI_LEVEL5" title="Tiler Hierarchy" name="L5 prims" description="Number of primitives in hierarchy level 5" units="primitives"/>
+        <event offset="22" counter="TI_LEVEL6" title="Tiler Hierarchy" name="L6 prims" description="Number of primitives in hierarchy level 6" units="primitives"/>
+        <event offset="23" counter="TI_LEVEL7" title="Tiler Hierarchy" name="L7 prims" description="Number of primitives in hierarchy level 7" units="primitives"/>
     </category>
     <category name="L2 Cache" per_cpu="no">
-        <event offset="161" counter="L2_READ_LOOKUP" title="L2 Cache Reads" name="L2 read lookups" description="Number of reads into the L2 cache" units="requests" />
-        <event offset="164" counter="L2_READ_SNOOP" title="L2 Cache Reads" name="Read snoops" description="Number of read transaction snoops" units="requests" />
-        <event offset="165" counter="L2_READ_HIT" title="L2 Cache Reads" name="L2 read hits" description="Number of reads hitting in the L2 cache" units="requests" />
-        <event offset="170" counter="L2_WRITE_SNOOP" title="L2 Cache Writes" name="Write snoops" description="Number of write transaction snoops" units="requests" />
-        <event offset="171" counter="L2_WRITE_HIT" title="L2 Cache Writes" name="L2 write hits" description="Number of writes hitting in the L2 cache" units="requests" />
-        <event offset="167" counter="L2_WRITE_LOOKUP" title="L2 Cache Writes" name="L2 write lookups" description="Number of writes into the L2 cache" units="requests" />
-        <event offset="159" counter="L2_EXT_READ_BEATS" title="L2 Cache Ext Reads" name="External read bytes" description="Number of external bus read bytes" multiplier="16" units="bytes" />
-        <event offset="182" counter="L2_EXT_AR_STALL" title="L2 Cache Ext Reads" name="External bus stalls (AR)" description="Number of cycles a valid read address (AR) is stalled by the external interconnect" units="cycles" />
-        <event offset="188" counter="L2_EXT_R_BUF_FULL" title="L2 Cache Ext Reads" name="External bus response buffer full" description="Number of cycles a valid request is blocked by a full response buffer" units="cycles" />
-        <event offset="184" counter="L2_EXT_RD_BUF_FULL" title="L2 Cache Ext Reads" name="External bus read data buffer full" description="Number of cycles a valid request is blocked by a full read data buffer" units="cycles" />
-        <event offset="158" counter="L2_EXT_WRITE_BEATS" title="L2 Cache Ext Writes" name="External write bytes" description="Number of external bus write bytes" multiplier="16" units="bytes" />
-        <event offset="186" counter="L2_EXT_W_STALL" title="L2 Cache Ext Writes" name="External bus stalls (W)" description="Number of cycles a valid write data (W channel) is stalled by the external interconnect" units="cycles" />
-        <event offset="187" counter="L2_EXT_W_BUF_FULL" title="L2 Cache Ext Writes" name="External bus write buffer full" description="Number of cycles a valid request is blocked by a full write buffer" units="cycles" />
+        <event offset="33" counter="L2_READ_LOOKUP" title="L2 Cache Reads" name="L2 read lookups" description="Number of reads into the L2 cache" units="requests"/>
+        <event offset="36" counter="L2_READ_SNOOP" title="L2 Cache Reads" name="Read snoops" description="Number of read transaction snoops" units="requests"/>
+        <event offset="37" counter="L2_READ_HIT" title="L2 Cache Reads" name="L2 read hits" description="Number of reads hitting in the L2 cache" units="requests"/>
+        <event offset="42" counter="L2_WRITE_SNOOP" title="L2 Cache Writes" name="Write snoops" description="Number of write transaction snoops" units="requests"/>
+        <event offset="43" counter="L2_WRITE_HIT" title="L2 Cache Writes" name="L2 write hits" description="Number of writes hitting in the L2 cache" units="requests"/>
+        <event offset="39" counter="L2_WRITE_LOOKUP" title="L2 Cache Writes" name="L2 write lookups" description="Number of writes into the L2 cache" units="requests"/>
+        <event offset="31" counter="L2_EXT_READ_BEATS" title="L2 Cache Ext Reads" name="External read bytes" description="Number of external bus read bytes" multiplier="16" units="bytes"/>
+        <event offset="54" counter="L2_EXT_AR_STALL" title="L2 Cache Ext Reads" name="External bus stalls (AR)" description="Number of cycles a valid read address (AR) is stalled by the external interconnect" units="cycles"/>
+        <event offset="60" counter="L2_EXT_R_BUF_FULL" title="L2 Cache Ext Reads" name="External bus response buffer full" description="Number of cycles a valid request is blocked by a full response buffer" units="cycles"/>
+        <event offset="56" counter="L2_EXT_RD_BUF_FULL" title="L2 Cache Ext Reads" name="External bus read data buffer full" description="Number of cycles a valid request is blocked by a full read data buffer" units="cycles"/>
+        <event offset="30" counter="L2_EXT_WRITE_BEATS" title="L2 Cache Ext Writes" name="External write bytes" description="Number of external bus write bytes" multiplier="16" units="bytes"/>
+        <event offset="58" counter="L2_EXT_W_STALL" title="L2 Cache Ext Writes" name="External bus stalls (W)" description="Number of cycles a valid write data (W channel) is stalled by the external interconnect" units="cycles"/>
+        <event offset="59" counter="L2_EXT_W_BUF_FULL" title="L2 Cache Ext Writes" name="External bus write buffer full" description="Number of cycles a valid request is blocked by a full write buffer" units="cycles"/>
+    </category>
+    <category name="Shader Core" per_cpu="no">
+        <event offset="26" counter="TRIPIPE_ACTIVE" title="Core Cycles" name="Tripipe cycles" description="Number of cycles tripipe was active" units="cycles"/>
+        <event offset="4" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment cycles" description="Number of cycles fragment processing was active" units="cycles"/>
+        <event offset="22" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Compute cycles" description="Number of cycles vertex\compute processing was active" units="cycles"/>
+        <event offset="19" counter="FRAG_CYCLES_NO_TILE" title="Core Cycles" name="Fragment cycles waiting for tile" description="Number of cycles spent waiting for a physical tile buffer" units="cycles"/>
+        <event offset="8" counter="FRAG_CYCLES_FPKQ_ACTIVE" title="Core Cycles" name="Fragment cycles pre-pipe buffer not empty" description="Number of cycles the pre-pipe queue contains quads" units="cycles"/>
+        <event offset="12" counter="FRAG_THREADS" title="Fragment Threads" name="Fragment threads" description="Number of fragment threads started" units="threads"/>
+        <event offset="13" counter="FRAG_DUMMY_THREADS" title="Fragment Threads" name="Dummy fragment threads" description="Number of dummy fragment threads started" units="threads"/>
+        <event offset="17" counter="FRAG_THREADS_LZS_TEST" title="Fragment Threads" name="Fragment threads doing late ZS" description="Number of threads doing late ZS test" units="threads"/>
+        <event offset="18" counter="FRAG_THREADS_LZS_KILLED" title="Fragment Threads" name="Fragment threads killed late ZS" description="Number of threads killed by late ZS test" units="threads"/>
+        <event offset="23" counter="COMPUTE_TASKS" title="Compute Tasks" name="Compute tasks" description="Number of compute tasks" units="tasks"/>
+        <event offset="24" counter="COMPUTE_THREADS" title="Compute Threads" name="Compute threads" description="Number of compute threads started" units="threads"/>
+        <event offset="5" counter="FRAG_PRIMITIVES" title="Fragment Primitives" name="Primitives loaded" description="Number of primitives loaded from tiler" units="primitives"/>
+        <event offset="6" counter="FRAG_PRIMITIVES_DROPPED" title="Fragment Primitives" name="Primitives dropped" description="Number of primitives dropped because out of tile" units="primitives"/>
+        <event offset="14" counter="FRAG_QUADS_RAST" title="Fragment Quads" name="Quads rasterized" description="Number of quads rasterized" units="quads"/>
+        <event offset="15" counter="FRAG_QUADS_EZS_TEST" title="Fragment Quads" name="Quads doing early ZS" description="Number of quads doing early ZS test" units="quads"/>
+        <event offset="16" counter="FRAG_QUADS_EZS_KILLED" title="Fragment Quads" name="Quads killed early Z" description="Number of quads killed by early ZS test" units="quads"/>
+        <event offset="20" counter="FRAG_NUM_TILES" title="Fragment" name="Pixels" description="Number of pixels rendered" multiplier="1024" units="pixels"/>
+        <event offset="21" counter="FRAG_TRANS_ELIM" title="Fragment Tasks" name="Tile writes killed by TE" description="Number of tile writes skipped by transaction elimination" units="tiles"/>
+        <event offset="27" counter="ARITH_WORDS" title="Arithmetic Pipe" name="A instructions" description="Number of instructions completed by the the A-pipe (normalized per pipeline)" units="instructions"/>
+        <event offset="31" counter="LS_WORDS" title="Load/Store Pipe" name="LS instructions" description="Number of instructions completed by the LS-pipe" units="instructions"/>
+        <event offset="32" counter="LS_ISSUES" title="Load/Store Pipe" name="LS instruction issues" description="Number of instructions issued to the LS-pipe, including restarts" units="instructions"/>
+        <event offset="38" counter="TEX_WORDS" title="Texture Pipe" name="T instructions" description="Number of instructions completed by the T-pipe" units="instructions"/>
+        <event offset="42" counter="TEX_ISSUES" title="Texture Pipe" name="T instruction issues" description="Number of threads through loop 2 address calculation" units="instructions"/>
+        <event offset="43" counter="TEX_RECIRC_FMISS" title="Texture Pipe" name="Cache misses" description="Number of instructions in the T-pipe, recirculated due to cache miss" units="requests"/>
+        <event offset="49" counter="LSC_READ_OP" title="Load/Store Cache Reads" name="Read operations" description="Number of read operations in the Load/Store cache" units="requests"/>
+        <event offset="48" counter="LSC_READ_HITS" title="Load/Store Cache Reads" name="Read hits" description="Number of read hits in the Load/Store cache" units="requests"/>
+        <event offset="51" counter="LSC_WRITE_OP" title="Load/Store Cache Writes" name="Write operations" description="Number of write operations in the Load/Store cache" units="requests"/>
+        <event offset="50" counter="LSC_WRITE_HITS" title="Load/Store Cache Writes" name="Write hits" description="Number of write hits in the Load/Store cache" units="requests"/>
+        <event offset="53" counter="LSC_ATOMIC_OP" title="Load/Store Cache Atomics" name="Atomic operations" description="Number of atomic operations in the Load/Store cache" units="requests"/>
+        <event offset="52" counter="LSC_ATOMIC_HITS" title="Load/Store Cache Atomics" name="Atomic hits" description="Number of atomic hits in the Load/Store cache" units="requests"/>
+        <event offset="54" counter="LSC_LINE_FETCHES" title="Load/Store Cache Bus" name="Line fetches" description="Number of line fetches in the Load/Store cache" units="requests"/>
+        <event offset="55" counter="LSC_DIRTY_LINE" title="Load/Store Cache Bus" name="Dirty line evictions" description="Number of dirty line evictions in the Load/Store cache" units="requests"/>
+        <event offset="56" counter="LSC_SNOOPS" title="Load/Store Cache Bus" name="Snoops in to LSC" description="Number of coherent memory snoops in to the Load/Store cache" units="requests"/>
     </category>
 </metrics>
\ No newline at end of file
diff --git a/src/panfrost/perf/T82x.xml b/src/panfrost/perf/T82x.xml
index abbe1ab5e13..beb86474f64 100644
--- a/src/panfrost/perf/T82x.xml
+++ b/src/panfrost/perf/T82x.xml
@@ -1,6 +1,6 @@
 <!--
 Copyright © 2017-2020 ARM Limited.
-Copyright © 2021 Collabora, Ltd.
+Copyright © 2021-2022 Collabora, Ltd.
 Author: Antonio Caggiano <antonio.caggiano@collabora.com>
 
 Permission is hereby granted, free of charge, to any person obtaining a
@@ -24,79 +24,79 @@ SOFTWARE.
 -->
 <metrics id="T82x">
     <category name="Job Manager" per_cpu="no">
-        <event offset="6" counter="GPU_ACTIVE" title="Job Manager Cycles" name="GPU cycles" description="Number of cycles GPU active" units="cycles" />
-        <event offset="7" counter="IRQ_ACTIVE" title="Job Manager Cycles" name="IRQ cycles" description="Number of cycles GPU interrupt pending" units="cycles" />
-        <event offset="10" counter="JS0_ACTIVE" title="Job Manager Cycles" name="JS0 cycles" description="Number of cycles JS0 (fragment) active" units="cycles" />
-        <event offset="18" counter="JS1_ACTIVE" title="Job Manager Cycles" name="JS1 cycles" description="Number of cycles JS1 (vertex/tiler/compute) active" units="cycles" />
-        <event offset="26" counter="JS2_ACTIVE" title="Job Manager Cycles" name="JS2 cycles" description="Number of cycles JS2 (vertex/compute) active" units="cycles" />
-        <event offset="8" counter="JS0_JOBS" title="Job Manager Jobs" name="JS0 jobs" description="Number of Jobs (fragment) completed in JS0" units="jobs" />
-        <event offset="16" counter="JS1_JOBS" title="Job Manager Jobs" name="JS1 jobs" description="Number of Jobs (vertex/tiler/compute) completed in JS1" units="jobs" />
-        <event offset="24" counter="JS2_JOBS" title="Job Manager Jobs" name="JS2 jobs" description="Number of Jobs (vertex/compute) completed in JS2" units="jobs" />
-        <event offset="9" counter="JS0_TASKS" title="Job Manager" name="Pixels" description="Number of Pixels completed in JS0" multiplier="1024" units="pixels" />
-        <event offset="17" counter="JS1_TASKS" title="Job Manager Tasks" name="JS1 tasks" description="Number of Tasks completed in JS1" units="tasks" />
-        <event offset="25" counter="JS2_TASKS" title="Job Manager Tasks" name="JS2 tasks" description="Number of Tasks completed in JS2" units="tasks" />
+        <event offset="6" counter="GPU_ACTIVE" title="Job Manager Cycles" name="GPU cycles" description="Number of cycles GPU active" units="cycles"/>
+        <event offset="7" counter="IRQ_ACTIVE" title="Job Manager Cycles" name="IRQ cycles" description="Number of cycles GPU interrupt pending" units="cycles"/>
+        <event offset="10" counter="JS0_ACTIVE" title="Job Manager Cycles" name="JS0 cycles" description="Number of cycles JS0 (fragment) active" units="cycles"/>
+        <event offset="18" counter="JS1_ACTIVE" title="Job Manager Cycles" name="JS1 cycles" description="Number of cycles JS1 (vertex/tiler/compute) active" units="cycles"/>
+        <event offset="26" counter="JS2_ACTIVE" title="Job Manager Cycles" name="JS2 cycles" description="Number of cycles JS2 (vertex/compute) active" units="cycles"/>
+        <event offset="8" counter="JS0_JOBS" title="Job Manager Jobs" name="JS0 jobs" description="Number of Jobs (fragment) completed in JS0" units="jobs"/>
+        <event offset="16" counter="JS1_JOBS" title="Job Manager Jobs" name="JS1 jobs" description="Number of Jobs (vertex/tiler/compute) completed in JS1" units="jobs"/>
+        <event offset="24" counter="JS2_JOBS" title="Job Manager Jobs" name="JS2 jobs" description="Number of Jobs (vertex/compute) completed in JS2" units="jobs"/>
+        <event offset="9" counter="JS0_TASKS" title="Job Manager" name="Pixels" description="Number of Pixels completed in JS0" multiplier="1024" units="pixels"/>
+        <event offset="17" counter="JS1_TASKS" title="Job Manager Tasks" name="JS1 tasks" description="Number of Tasks completed in JS1" units="tasks"/>
+        <event offset="25" counter="JS2_TASKS" title="Job Manager Tasks" name="JS2 tasks" description="Number of Tasks completed in JS2" units="tasks"/>
     </category>
     <category name="Tiler" per_cpu="no">
-        <event offset="86" counter="TI_ACTIVE" title="Tiler Cycles" name="Tiler cycles" description="Number of cycles Tiler active" units="cycles" />
-        <event offset="70" counter="TI_POLYGONS" title="Tiler Primitives" name="Polygons" description="Number of polygons processed" units="primitives" />
-        <event offset="69" counter="TI_QUADS" title="Tiler Primitives" name="Quads" description="Number of quads processed" units="primitives" />
-        <event offset="68" counter="TI_TRIANGLES" title="Tiler Primitives" name="Triangles" description="Number of triangles processed" units="primitives" />
-        <event offset="72" counter="TI_LINES" title="Tiler Primitives" name="Lines" description="Number of lines processed" units="primitives" />
-        <event offset="71" counter="TI_POINTS" title="Tiler Primitives" name="Points" description="Number of points processed" units="primitives" />
-        <event offset="73" counter="TI_FRONT_FACING" title="Tiler Culling" name="Front facing prims" description="Number of front facing primitives" units="primitives" />
-        <event offset="74" counter="TI_BACK_FACING" title="Tiler Culling" name="Back facing prims" description="Number of back facing primitives" units="primitives" />
-        <event offset="75" counter="TI_PRIM_VISIBLE" title="Tiler Culling" name="Visible prims" description="Number of visible primitives" units="primitives" />
-        <event offset="76" counter="TI_PRIM_CULLED" title="Tiler Culling" name="Culled prims" description="Number of culled primitives" units="primitives" />
-        <event offset="77" counter="TI_PRIM_CLIPPED" title="Tiler Culling" name="Clipped prims" description="Number of clipped primitives" units="primitives" />
-    </category>
-    <category name="Shader Core" per_cpu="no">
-        <event offset="218" counter="TRIPIPE_ACTIVE" title="Core Cycles" name="Tripipe cycles" description="Number of cycles tripipe was active" units="cycles" />
-        <event offset="196" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment cycles" description="Number of cycles fragment processing was active" units="cycles" />
-        <event offset="214" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Compute cycles" description="Number of cycles vertex\compute processing was active" units="cycles" />
-        <event offset="211" counter="FRAG_CYCLES_NO_TILE" title="Core Cycles" name="Fragment cycles waiting for tile" description="Number of cycles spent waiting for a physical tile buffer" units="cycles" />
-        <event offset="200" counter="FRAG_CYCLES_FPKQ_ACTIVE" title="Core Cycles" name="Fragment cycles pre-pipe buffer not empty" description="Number of cycles the pre-pipe queue contains quads" units="cycles" />
-        <event offset="204" counter="FRAG_THREADS" title="Fragment Threads" name="Fragment threads" description="Number of fragment threads started" units="threads" />
-        <event offset="205" counter="FRAG_DUMMY_THREADS" title="Fragment Threads" name="Dummy fragment threads" description="Number of dummy fragment threads started" units="threads" />
-        <event offset="209" counter="FRAG_THREADS_LZS_TEST" title="Fragment Threads" name="Fragment threads doing late ZS" description="Number of threads doing late ZS test" units="threads" />
-        <event offset="210" counter="FRAG_THREADS_LZS_KILLED" title="Fragment Threads" name="Fragment threads killed late ZS" description="Number of threads killed by late ZS test" units="threads" />
-        <event offset="215" counter="COMPUTE_TASKS" title="Compute Tasks" name="Compute tasks" description="Number of compute tasks" units="tasks" />
-        <event offset="216" counter="COMPUTE_THREADS" title="Compute Threads" name="Compute threads" description="Number of compute threads started" units="threads" />
-        <event offset="197" counter="FRAG_PRIMITIVES" title="Fragment Primitives" name="Primitives loaded" description="Number of primitives loaded from tiler" units="primitives" />
-        <event offset="198" counter="FRAG_PRIMITIVES_DROPPED" title="Fragment Primitives" name="Primitives dropped" description="Number of primitives dropped because out of tile" units="primitives" />
-        <event offset="206" counter="FRAG_QUADS_RAST" title="Fragment Quads" name="Quads rasterized" description="Number of quads rasterized" units="quads" />
-        <event offset="207" counter="FRAG_QUADS_EZS_TEST" title="Fragment Quads" name="Quads doing early ZS" description="Number of quads doing early ZS test" units="quads" />
-        <event offset="208" counter="FRAG_QUADS_EZS_KILLED" title="Fragment Quads" name="Quads killed early Z" description="Number of quads killed by early ZS test" units="quads" />
-        <event offset="212" counter="FRAG_NUM_TILES" title="Fragment" name="Pixels" description="Number of pixels rendered" multiplier="1024" units="pixels" />
-        <event offset="213" counter="FRAG_TRANS_ELIM" title="Fragment Tasks" name="Tile writes killed by TE" description="Number of tile writes skipped by transaction elimination" units="tiles" />
-        <event offset="219" counter="ARITH_WORDS" title="Arithmetic Pipe" name="A instructions" description="Number of batched instructions executed by the A-pipe (normalized per pipe)" units="instructions" />
-        <event offset="223" counter="LS_WORDS" title="Load/Store Pipe" name="LS instructions" description="Number of instructions completed by the LS-pipe" units="instructions" />
-        <event offset="224" counter="LS_ISSUES" title="Load/Store Pipe" name="LS instruction issues" description="Number of instructions issued to the LS-pipe, including restarts" units="instructions" />
-        <event offset="230" counter="TEX_WORDS" title="Texture Pipe" name="T instructions" description="Number of instructions completed by the T-pipe" units="instructions" />
-        <event offset="234" counter="TEX_ISSUES" title="Texture Pipe" name="T instruction issues" description="Number of threads through loop 2 address calculation" units="instructions" />
-        <event offset="235" counter="TEX_RECIRC_FMISS" title="Texture Pipe" name="Cache misses" description="Number of instructions in the T-pipe, recirculated due to cache miss" units="requests" />
-        <event offset="241" counter="LSC_READ_OP" title="Load/Store Cache Reads" name="Read operations" description="Number of read operations in the Load/Store cache" units="requests" />
-        <event offset="240" counter="LSC_READ_HITS" title="Load/Store Cache Reads" name="Read hits" description="Number of read hits in the Load/Store cache" units="requests" />
-        <event offset="243" counter="LSC_WRITE_OP" title="Load/Store Cache Writes" name="Write operations" description="Number of write operations in the Load/Store cache" units="requests" />
-        <event offset="242" counter="LSC_WRITE_HITS" title="Load/Store Cache Writes" name="Write hits" description="Number of write hits in the Load/Store cache" units="requests" />
-        <event offset="245" counter="LSC_ATOMIC_OP" title="Load/Store Cache Atomics" name="Atomic operations" description="Number of atomic operations in the Load/Store cache" units="requests" />
-        <event offset="244" counter="LSC_ATOMIC_HITS" title="Load/Store Cache Atomics" name="Atomic hits" description="Number of atomic hits in the Load/Store cache" units="requests" />
-        <event offset="246" counter="LSC_LINE_FETCHES" title="Load/Store Cache Bus" name="Line fetches" description="Number of line fetches in the Load/Store cache" units="requests" />
-        <event offset="247" counter="LSC_DIRTY_LINE" title="Load/Store Cache Bus" name="Dirty line evictions" description="Number of dirty line evictions in the Load/Store cache" units="requests" />
-        <event offset="248" counter="LSC_SNOOPS" title="Load/Store Cache Bus" name="Snoops in to LSC" description="Number of coherent memory snoops in to the Load/Store cache" units="requests" />
+        <event offset="22" counter="TI_ACTIVE" title="Tiler Cycles" name="Tiler cycles" description="Number of cycles Tiler active" units="cycles"/>
+        <event offset="6" counter="TI_POLYGONS" title="Tiler Primitives" name="Polygons" description="Number of polygons processed" units="primitives"/>
+        <event offset="5" counter="TI_QUADS" title="Tiler Primitives" name="Quads" description="Number of quads processed" units="primitives"/>
+        <event offset="4" counter="TI_TRIANGLES" title="Tiler Primitives" name="Triangles" description="Number of triangles processed" units="primitives"/>
+        <event offset="8" counter="TI_LINES" title="Tiler Primitives" name="Lines" description="Number of lines processed" units="primitives"/>
+        <event offset="7" counter="TI_POINTS" title="Tiler Primitives" name="Points" description="Number of points processed" units="primitives"/>
+        <event offset="9" counter="TI_FRONT_FACING" title="Tiler Culling" name="Front facing prims" description="Number of front facing primitives" units="primitives"/>
+        <event offset="10" counter="TI_BACK_FACING" title="Tiler Culling" name="Back facing prims" description="Number of back facing primitives" units="primitives"/>
+        <event offset="11" counter="TI_PRIM_VISIBLE" title="Tiler Culling" name="Visible prims" description="Number of visible primitives" units="primitives"/>
+        <event offset="12" counter="TI_PRIM_CULLED" title="Tiler Culling" name="Culled prims" description="Number of culled primitives" units="primitives"/>
+        <event offset="13" counter="TI_PRIM_CLIPPED" title="Tiler Culling" name="Clipped prims" description="Number of clipped primitives" units="primitives"/>
     </category>
     <category name="L2 Cache" per_cpu="no">
-        <event offset="161" counter="L2_READ_LOOKUP" title="L2 Cache Reads" name="L2 read lookups" description="Number of reads into the L2 cache" units="requests" />
-        <event offset="164" counter="L2_READ_SNOOP" title="L2 Cache Reads" name="Read snoops" description="Number of read transaction snoops" units="requests" />
-        <event offset="165" counter="L2_READ_HIT" title="L2 Cache Reads" name="L2 read hits" description="Number of reads hitting in the L2 cache" units="requests" />
-        <event offset="170" counter="L2_WRITE_SNOOP" title="L2 Cache Writes" name="Write snoops" description="Number of write transaction snoops" units="requests" />
-        <event offset="171" counter="L2_WRITE_HIT" title="L2 Cache Writes" name="L2 write hits" description="Number of writes hitting in the L2 cache" units="requests" />
-        <event offset="167" counter="L2_WRITE_LOOKUP" title="L2 Cache Writes" name="L2 write lookups" description="Number of writes into the L2 cache" units="requests" />
-        <event offset="159" counter="L2_EXT_READ_BEATS" title="L2 Cache Ext Reads" name="External read bytes" description="Number of external bus read bytes" multiplier="16" units="bytes" />
-        <event offset="182" counter="L2_EXT_AR_STALL" title="L2 Cache Ext Reads" name="External bus stalls (AR)" description="Number of cycles a valid read address (AR) is stalled by the external interconnect" units="cycles" />
-        <event offset="188" counter="L2_EXT_R_BUF_FULL" title="L2 Cache Ext Reads" name="External bus response buffer full" description="Number of cycles a valid request is blocked by a full response buffer" units="cycles" />
-        <event offset="184" counter="L2_EXT_RD_BUF_FULL" title="L2 Cache Ext Reads" name="External bus read data buffer full" description="Number of cycles a valid request is blocked by a full read data buffer" units="cycles" />
-        <event offset="158" counter="L2_EXT_WRITE_BEATS" title="L2 Cache Ext Writes" name="External write bytes" description="Number of external bus write bytes" multiplier="16" units="bytes" />
-        <event offset="186" counter="L2_EXT_W_STALL" title="L2 Cache Ext Writes" name="External bus stalls (W)" description="Number of cycles a valid write data (W channel) is stalled by the external interconnect" units="cycles" />
-        <event offset="187" counter="L2_EXT_W_BUF_FULL" title="L2 Cache Ext Writes" name="External bus write buffer full" description="Number of cycles a valid request is blocked by a full write buffer" units="cycles" />
+        <event offset="33" counter="L2_READ_LOOKUP" title="L2 Cache Reads" name="L2 read lookups" description="Number of reads into the L2 cache" units="requests"/>
+        <event offset="36" counter="L2_READ_SNOOP" title="L2 Cache Reads" name="Read snoops" description="Number of read transaction snoops" units="requests"/>
+        <event offset="37" counter="L2_READ_HIT" title="L2 Cache Reads" name="L2 read hits" description="Number of reads hitting in the L2 cache" units="requests"/>
+        <event offset="42" counter="L2_WRITE_SNOOP" title="L2 Cache Writes" name="Write snoops" description="Number of write transaction snoops" units="requests"/>
+        <event offset="43" counter="L2_WRITE_HIT" title="L2 Cache Writes" name="L2 write hits" description="Number of writes hitting in the L2 cache" units="requests"/>
+        <event offset="39" counter="L2_WRITE_LOOKUP" title="L2 Cache Writes" name="L2 write lookups" description="Number of writes into the L2 cache" units="requests"/>
+        <event offset="31" counter="L2_EXT_READ_BEATS" title="L2 Cache Ext Reads" name="External read bytes" description="Number of external bus read bytes" multiplier="16" units="bytes"/>
+        <event offset="54" counter="L2_EXT_AR_STALL" title="L2 Cache Ext Reads" name="External bus stalls (AR)" description="Number of cycles a valid read address (AR) is stalled by the external interconnect" units="cycles"/>
+        <event offset="60" counter="L2_EXT_R_BUF_FULL" title="L2 Cache Ext Reads" name="External bus response buffer full" description="Number of cycles a valid request is blocked by a full response buffer" units="cycles"/>
+        <event offset="56" counter="L2_EXT_RD_BUF_FULL" title="L2 Cache Ext Reads" name="External bus read data buffer full" description="Number of cycles a valid request is blocked by a full read data buffer" units="cycles"/>
+        <event offset="30" counter="L2_EXT_WRITE_BEATS" title="L2 Cache Ext Writes" name="External write bytes" description="Number of external bus write bytes" multiplier="16" units="bytes"/>
+        <event offset="58" counter="L2_EXT_W_STALL" title="L2 Cache Ext Writes" name="External bus stalls (W)" description="Number of cycles a valid write data (W channel) is stalled by the external interconnect" units="cycles"/>
+        <event offset="59" counter="L2_EXT_W_BUF_FULL" title="L2 Cache Ext Writes" name="External bus write buffer full" description="Number of cycles a valid request is blocked by a full write buffer" units="cycles"/>
+    </category>
+    <category name="Shader Core" per_cpu="no">
+        <event offset="26" counter="TRIPIPE_ACTIVE" title="Core Cycles" name="Tripipe cycles" description="Number of cycles tripipe was active" units="cycles"/>
+        <event offset="4" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment cycles" description="Number of cycles fragment processing was active" units="cycles"/>
+        <event offset="22" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Compute cycles" description="Number of cycles vertex\compute processing was active" units="cycles"/>
+        <event offset="19" counter="FRAG_CYCLES_NO_TILE" title="Core Cycles" name="Fragment cycles waiting for tile" description="Number of cycles spent waiting for a physical tile buffer" units="cycles"/>
+        <event offset="8" counter="FRAG_CYCLES_FPKQ_ACTIVE" title="Core Cycles" name="Fragment cycles pre-pipe buffer not empty" description="Number of cycles the pre-pipe queue contains quads" units="cycles"/>
+        <event offset="12" counter="FRAG_THREADS" title="Fragment Threads" name="Fragment threads" description="Number of fragment threads started" units="threads"/>
+        <event offset="13" counter="FRAG_DUMMY_THREADS" title="Fragment Threads" name="Dummy fragment threads" description="Number of dummy fragment threads started" units="threads"/>
+        <event offset="17" counter="FRAG_THREADS_LZS_TEST" title="Fragment Threads" name="Fragment threads doing late ZS" description="Number of threads doing late ZS test" units="threads"/>
+        <event offset="18" counter="FRAG_THREADS_LZS_KILLED" title="Fragment Threads" name="Fragment threads killed late ZS" description="Number of threads killed by late ZS test" units="threads"/>
+        <event offset="23" counter="COMPUTE_TASKS" title="Compute Tasks" name="Compute tasks" description="Number of compute tasks" units="tasks"/>
+        <event offset="24" counter="COMPUTE_THREADS" title="Compute Threads" name="Compute threads" description="Number of compute threads started" units="threads"/>
+        <event offset="5" counter="FRAG_PRIMITIVES" title="Fragment Primitives" name="Primitives loaded" description="Number of primitives loaded from tiler" units="primitives"/>
+        <event offset="6" counter="FRAG_PRIMITIVES_DROPPED" title="Fragment Primitives" name="Primitives dropped" description="Number of primitives dropped because out of tile" units="primitives"/>
+        <event offset="14" counter="FRAG_QUADS_RAST" title="Fragment Quads" name="Quads rasterized" description="Number of quads rasterized" units="quads"/>
+        <event offset="15" counter="FRAG_QUADS_EZS_TEST" title="Fragment Quads" name="Quads doing early ZS" description="Number of quads doing early ZS test" units="quads"/>
+        <event offset="16" counter="FRAG_QUADS_EZS_KILLED" title="Fragment Quads" name="Quads killed early Z" description="Number of quads killed by early ZS test" units="quads"/>
+        <event offset="20" counter="FRAG_NUM_TILES" title="Fragment" name="Pixels" description="Number of pixels rendered" multiplier="1024" units="pixels"/>
+        <event offset="21" counter="FRAG_TRANS_ELIM" title="Fragment Tasks" name="Tile writes killed by TE" description="Number of tile writes skipped by transaction elimination" units="tiles"/>
+        <event offset="27" counter="ARITH_WORDS" title="Arithmetic Pipe" name="A instructions" description="Number of batched instructions executed by the A-pipe (normalized per pipe)" units="instructions"/>
+        <event offset="31" counter="LS_WORDS" title="Load/Store Pipe" name="LS instructions" description="Number of instructions completed by the LS-pipe" units="instructions"/>
+        <event offset="32" counter="LS_ISSUES" title="Load/Store Pipe" name="LS instruction issues" description="Number of instructions issued to the LS-pipe, including restarts" units="instructions"/>
+        <event offset="38" counter="TEX_WORDS" title="Texture Pipe" name="T instructions" description="Number of instructions completed by the T-pipe" units="instructions"/>
+        <event offset="42" counter="TEX_ISSUES" title="Texture Pipe" name="T instruction issues" description="Number of threads through loop 2 address calculation" units="instructions"/>
+        <event offset="43" counter="TEX_RECIRC_FMISS" title="Texture Pipe" name="Cache misses" description="Number of instructions in the T-pipe, recirculated due to cache miss" units="requests"/>
+        <event offset="49" counter="LSC_READ_OP" title="Load/Store Cache Reads" name="Read operations" description="Number of read operations in the Load/Store cache" units="requests"/>
+        <event offset="48" counter="LSC_READ_HITS" title="Load/Store Cache Reads" name="Read hits" description="Number of read hits in the Load/Store cache" units="requests"/>
+        <event offset="51" counter="LSC_WRITE_OP" title="Load/Store Cache Writes" name="Write operations" description="Number of write operations in the Load/Store cache" units="requests"/>
+        <event offset="50" counter="LSC_WRITE_HITS" title="Load/Store Cache Writes" name="Write hits" description="Number of write hits in the Load/Store cache" units="requests"/>
+        <event offset="53" counter="LSC_ATOMIC_OP" title="Load/Store Cache Atomics" name="Atomic operations" description="Number of atomic operations in the Load/Store cache" units="requests"/>
+        <event offset="52" counter="LSC_ATOMIC_HITS" title="Load/Store Cache Atomics" name="Atomic hits" description="Number of atomic hits in the Load/Store cache" units="requests"/>
+        <event offset="54" counter="LSC_LINE_FETCHES" title="Load/Store Cache Bus" name="Line fetches" description="Number of line fetches in the Load/Store cache" units="requests"/>
+        <event offset="55" counter="LSC_DIRTY_LINE" title="Load/Store Cache Bus" name="Dirty line evictions" description="Number of dirty line evictions in the Load/Store cache" units="requests"/>
+        <event offset="56" counter="LSC_SNOOPS" title="Load/Store Cache Bus" name="Snoops in to LSC" description="Number of coherent memory snoops in to the Load/Store cache" units="requests"/>
     </category>
 </metrics>
\ No newline at end of file
diff --git a/src/panfrost/perf/T83x.xml b/src/panfrost/perf/T83x.xml
index 6834c9289b8..7f36dfd1e68 100644
--- a/src/panfrost/perf/T83x.xml
+++ b/src/panfrost/perf/T83x.xml
@@ -1,6 +1,6 @@
 <!--
 Copyright © 2017-2020 ARM Limited.
-Copyright © 2021 Collabora, Ltd.
+Copyright © 2021-2022 Collabora, Ltd.
 Author: Antonio Caggiano <antonio.caggiano@collabora.com>
 
 Permission is hereby granted, free of charge, to any person obtaining a
@@ -24,79 +24,79 @@ SOFTWARE.
 -->
 <metrics id="T83x">
     <category name="Job Manager" per_cpu="no">
-        <event offset="6" counter="GPU_ACTIVE" title="Job Manager Cycles" name="GPU cycles" description="Number of cycles GPU active" units="cycles" />
-        <event offset="7" counter="IRQ_ACTIVE" title="Job Manager Cycles" name="IRQ cycles" description="Number of cycles GPU interrupt pending" units="cycles" />
-        <event offset="10" counter="JS0_ACTIVE" title="Job Manager Cycles" name="JS0 cycles" description="Number of cycles JS0 (fragment) active" units="cycles" />
-        <event offset="18" counter="JS1_ACTIVE" title="Job Manager Cycles" name="JS1 cycles" description="Number of cycles JS1 (vertex/tiler/compute) active" units="cycles" />
-        <event offset="26" counter="JS2_ACTIVE" title="Job Manager Cycles" name="JS2 cycles" description="Number of cycles JS2 (vertex/compute) active" units="cycles" />
-        <event offset="8" counter="JS0_JOBS" title="Job Manager Jobs" name="JS0 jobs" description="Number of Jobs (fragment) completed in JS0" units="jobs" />
-        <event offset="16" counter="JS1_JOBS" title="Job Manager Jobs" name="JS1 jobs" description="Number of Jobs (vertex/tiler/compute) completed in JS1" units="jobs" />
-        <event offset="24" counter="JS2_JOBS" title="Job Manager Jobs" name="JS2 jobs" description="Number of Jobs (vertex/compute) completed in JS2" units="jobs" />
-        <event offset="9" counter="JS0_TASKS" title="Job Manager" name="Pixels" description="Number of Pixels completed in JS0" multiplier="1024" units="pixels" />
-        <event offset="17" counter="JS1_TASKS" title="Job Manager Tasks" name="JS1 tasks" description="Number of Tasks completed in JS1" units="tasks" />
-        <event offset="25" counter="JS2_TASKS" title="Job Manager Tasks" name="JS2 tasks" description="Number of Tasks completed in JS2" units="tasks" />
+        <event offset="6" counter="GPU_ACTIVE" title="Job Manager Cycles" name="GPU cycles" description="Number of cycles GPU active" units="cycles"/>
+        <event offset="7" counter="IRQ_ACTIVE" title="Job Manager Cycles" name="IRQ cycles" description="Number of cycles GPU interrupt pending" units="cycles"/>
+        <event offset="10" counter="JS0_ACTIVE" title="Job Manager Cycles" name="JS0 cycles" description="Number of cycles JS0 (fragment) active" units="cycles"/>
+        <event offset="18" counter="JS1_ACTIVE" title="Job Manager Cycles" name="JS1 cycles" description="Number of cycles JS1 (vertex/tiler/compute) active" units="cycles"/>
+        <event offset="26" counter="JS2_ACTIVE" title="Job Manager Cycles" name="JS2 cycles" description="Number of cycles JS2 (vertex/compute) active" units="cycles"/>
+        <event offset="8" counter="JS0_JOBS" title="Job Manager Jobs" name="JS0 jobs" description="Number of Jobs (fragment) completed in JS0" units="jobs"/>
+        <event offset="16" counter="JS1_JOBS" title="Job Manager Jobs" name="JS1 jobs" description="Number of Jobs (vertex/tiler/compute) completed in JS1" units="jobs"/>
+        <event offset="24" counter="JS2_JOBS" title="Job Manager Jobs" name="JS2 jobs" description="Number of Jobs (vertex/compute) completed in JS2" units="jobs"/>
+        <event offset="9" counter="JS0_TASKS" title="Job Manager" name="Pixels" description="Number of Pixels completed in JS0" multiplier="1024" units="pixels"/>
+        <event offset="17" counter="JS1_TASKS" title="Job Manager Tasks" name="JS1 tasks" description="Number of Tasks completed in JS1" units="tasks"/>
+        <event offset="25" counter="JS2_TASKS" title="Job Manager Tasks" name="JS2 tasks" description="Number of Tasks completed in JS2" units="tasks"/>
     </category>
     <category name="Tiler" per_cpu="no">
-        <event offset="86" counter="TI_ACTIVE" title="Tiler Cycles" name="Tiler cycles" description="Number of cycles Tiler active" units="cycles" />
-        <event offset="70" counter="TI_POLYGONS" title="Tiler Primitives" name="Polygons" description="Number of polygons processed" units="primitives" />
-        <event offset="69" counter="TI_QUADS" title="Tiler Primitives" name="Quads" description="Number of quads processed" units="primitives" />
-        <event offset="68" counter="TI_TRIANGLES" title="Tiler Primitives" name="Triangles" description="Number of triangles processed" units="primitives" />
-        <event offset="72" counter="TI_LINES" title="Tiler Primitives" name="Lines" description="Number of lines processed" units="primitives" />
-        <event offset="71" counter="TI_POINTS" title="Tiler Primitives" name="Points" description="Number of points processed" units="primitives" />
-        <event offset="73" counter="TI_FRONT_FACING" title="Tiler Culling" name="Front facing prims" description="Number of front facing primitives" units="primitives" />
-        <event offset="74" counter="TI_BACK_FACING" title="Tiler Culling" name="Back facing prims" description="Number of back facing primitives" units="primitives" />
-        <event offset="75" counter="TI_PRIM_VISIBLE" title="Tiler Culling" name="Visible prims" description="Number of visible primitives" units="primitives" />
-        <event offset="76" counter="TI_PRIM_CULLED" title="Tiler Culling" name="Culled prims" description="Number of culled primitives" units="primitives" />
-        <event offset="77" counter="TI_PRIM_CLIPPED" title="Tiler Culling" name="Clipped prims" description="Number of clipped primitives" units="primitives" />
-    </category>
-    <category name="Shader Core" per_cpu="no">
-        <event offset="218" counter="TRIPIPE_ACTIVE" title="Core Cycles" name="Tripipe cycles" description="Number of cycles tripipe was active" units="cycles" />
-        <event offset="196" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment cycles" description="Number of cycles fragment processing was active" units="cycles" />
-        <event offset="214" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Compute cycles" description="Number of cycles vertex\compute processing was active" units="cycles" />
-        <event offset="211" counter="FRAG_CYCLES_NO_TILE" title="Core Cycles" name="Fragment cycles waiting for tile" description="Number of cycles spent waiting for a physical tile buffer" units="cycles" />
-        <event offset="200" counter="FRAG_CYCLES_FPKQ_ACTIVE" title="Core Cycles" name="Fragment cycles pre-pipe buffer not empty" description="Number of cycles the pre-pipe queue contains quads" units="cycles" />
-        <event offset="204" counter="FRAG_THREADS" title="Fragment Threads" name="Fragment threads" description="Number of fragment threads started" units="threads" />
-        <event offset="205" counter="FRAG_DUMMY_THREADS" title="Fragment Threads" name="Dummy fragment threads" description="Number of dummy fragment threads started" units="threads" />
-        <event offset="209" counter="FRAG_THREADS_LZS_TEST" title="Fragment Threads" name="Fragment threads doing late ZS" description="Number of threads doing late ZS test" units="threads" />
-        <event offset="210" counter="FRAG_THREADS_LZS_KILLED" title="Fragment Threads" name="Fragment threads killed late ZS" description="Number of threads killed by late ZS test" units="threads" />
-        <event offset="215" counter="COMPUTE_TASKS" title="Compute Tasks" name="Compute tasks" description="Number of compute tasks" units="tasks" />
-        <event offset="216" counter="COMPUTE_THREADS" title="Compute Threads" name="Compute threads" description="Number of compute threads started" units="threads" />
-        <event offset="197" counter="FRAG_PRIMITIVES" title="Fragment Primitives" name="Primitives loaded" description="Number of primitives loaded from tiler" units="primitives" />
-        <event offset="198" counter="FRAG_PRIMITIVES_DROPPED" title="Fragment Primitives" name="Primitives dropped" description="Number of primitives dropped because out of tile" units="primitives" />
-        <event offset="206" counter="FRAG_QUADS_RAST" title="Fragment Quads" name="Quads rasterized" description="Number of quads rasterized" units="quads" />
-        <event offset="207" counter="FRAG_QUADS_EZS_TEST" title="Fragment Quads" name="Quads doing early ZS" description="Number of quads doing early ZS test" units="quads" />
-        <event offset="208" counter="FRAG_QUADS_EZS_KILLED" title="Fragment Quads" name="Quads killed early Z" description="Number of quads killed by early ZS test" units="quads" />
-        <event offset="212" counter="FRAG_NUM_TILES" title="Fragment" name="Pixels" description="Number of pixels rendered" multiplier="1024" units="pixels" />
-        <event offset="213" counter="FRAG_TRANS_ELIM" title="Fragment Tasks" name="Tile writes killed by TE" description="Number of tile writes skipped by transaction elimination" units="tiles" />
-        <event offset="219" counter="ARITH_WORDS" title="Arithmetic Pipe" name="A instructions" description="Number of batched instructions executed by the A-pipe (normalized per pipe)" units="instructions" />
-        <event offset="223" counter="LS_WORDS" title="Load/Store Pipe" name="LS instructions" description="Number of instructions completed by the LS-pipe" units="instructions" />
-        <event offset="224" counter="LS_ISSUES" title="Load/Store Pipe" name="LS instruction issues" description="Number of instructions issued to the LS-pipe, including restarts" units="instructions" />
-        <event offset="230" counter="TEX_WORDS" title="Texture Pipe" name="T instructions" description="Number of instructions completed by the T-pipe" units="instructions" />
-        <event offset="234" counter="TEX_ISSUES" title="Texture Pipe" name="T instruction issues" description="Number of threads through loop 2 address calculation" units="instructions" />
-        <event offset="235" counter="TEX_RECIRC_FMISS" title="Texture Pipe" name="Cache misses" description="Number of instructions in the T-pipe, recirculated due to cache miss" units="requests" />
-        <event offset="241" counter="LSC_READ_OP" title="Load/Store Cache Reads" name="Read operations" description="Number of read operations in the Load/Store cache" units="requests" />
-        <event offset="240" counter="LSC_READ_HITS" title="Load/Store Cache Reads" name="Read hits" description="Number of read hits in the Load/Store cache" units="requests" />
-        <event offset="243" counter="LSC_WRITE_OP" title="Load/Store Cache Writes" name="Write operations" description="Number of write operations in the Load/Store cache" units="requests" />
-        <event offset="242" counter="LSC_WRITE_HITS" title="Load/Store Cache Writes" name="Write hits" description="Number of write hits in the Load/Store cache" units="requests" />
-        <event offset="245" counter="LSC_ATOMIC_OP" title="Load/Store Cache Atomics" name="Atomic operations" description="Number of atomic operations in the Load/Store cache" units="requests" />
-        <event offset="244" counter="LSC_ATOMIC_HITS" title="Load/Store Cache Atomics" name="Atomic hits" description="Number of atomic hits in the Load/Store cache" units="requests" />
-        <event offset="246" counter="LSC_LINE_FETCHES" title="Load/Store Cache Bus" name="Line fetches" description="Number of line fetches in the Load/Store cache" units="requests" />
-        <event offset="247" counter="LSC_DIRTY_LINE" title="Load/Store Cache Bus" name="Dirty line evictions" description="Number of dirty line evictions in the Load/Store cache" units="requests" />
-        <event offset="248" counter="LSC_SNOOPS" title="Load/Store Cache Bus" name="Snoops in to LSC" description="Number of coherent memory snoops in to the Load/Store cache" units="requests" />
+        <event offset="22" counter="TI_ACTIVE" title="Tiler Cycles" name="Tiler cycles" description="Number of cycles Tiler active" units="cycles"/>
+        <event offset="6" counter="TI_POLYGONS" title="Tiler Primitives" name="Polygons" description="Number of polygons processed" units="primitives"/>
+        <event offset="5" counter="TI_QUADS" title="Tiler Primitives" name="Quads" description="Number of quads processed" units="primitives"/>
+        <event offset="4" counter="TI_TRIANGLES" title="Tiler Primitives" name="Triangles" description="Number of triangles processed" units="primitives"/>
+        <event offset="8" counter="TI_LINES" title="Tiler Primitives" name="Lines" description="Number of lines processed" units="primitives"/>
+        <event offset="7" counter="TI_POINTS" title="Tiler Primitives" name="Points" description="Number of points processed" units="primitives"/>
+        <event offset="9" counter="TI_FRONT_FACING" title="Tiler Culling" name="Front facing prims" description="Number of front facing primitives" units="primitives"/>
+        <event offset="10" counter="TI_BACK_FACING" title="Tiler Culling" name="Back facing prims" description="Number of back facing primitives" units="primitives"/>
+        <event offset="11" counter="TI_PRIM_VISIBLE" title="Tiler Culling" name="Visible prims" description="Number of visible primitives" units="primitives"/>
+        <event offset="12" counter="TI_PRIM_CULLED" title="Tiler Culling" name="Culled prims" description="Number of culled primitives" units="primitives"/>
+        <event offset="13" counter="TI_PRIM_CLIPPED" title="Tiler Culling" name="Clipped prims" description="Number of clipped primitives" units="primitives"/>
     </category>
     <category name="L2 Cache" per_cpu="no">
-        <event offset="161" counter="L2_READ_LOOKUP" title="L2 Cache Reads" name="L2 read lookups" description="Number of reads into the L2 cache" units="requests" />
-        <event offset="164" counter="L2_READ_SNOOP" title="L2 Cache Reads" name="Read snoops" description="Number of read transaction snoops" units="requests" />
-        <event offset="165" counter="L2_READ_HIT" title="L2 Cache Reads" name="L2 read hits" description="Number of reads hitting in the L2 cache" units="requests" />
-        <event offset="170" counter="L2_WRITE_SNOOP" title="L2 Cache Writes" name="Write snoops" description="Number of write transaction snoops" units="requests" />
-        <event offset="171" counter="L2_WRITE_HIT" title="L2 Cache Writes" name="L2 write hits" description="Number of writes hitting in the L2 cache" units="requests" />
-        <event offset="167" counter="L2_WRITE_LOOKUP" title="L2 Cache Writes" name="L2 write lookups" description="Number of writes into the L2 cache" units="requests" />
-        <event offset="159" counter="L2_EXT_READ_BEATS" title="L2 Cache Ext Reads" name="External read bytes" description="Number of external bus read bytes" multiplier="16" units="bytes" />
-        <event offset="182" counter="L2_EXT_AR_STALL" title="L2 Cache Ext Reads" name="External bus stalls (AR)" description="Number of cycles a valid read address (AR) is stalled by the external interconnect" units="cycles" />
-        <event offset="188" counter="L2_EXT_R_BUF_FULL" title="L2 Cache Ext Reads" name="External bus response buffer full" description="Number of cycles a valid request is blocked by a full response buffer" units="cycles" />
-        <event offset="184" counter="L2_EXT_RD_BUF_FULL" title="L2 Cache Ext Reads" name="External bus read data buffer full" description="Number of cycles a valid request is blocked by a full read data buffer" units="cycles" />
-        <event offset="158" counter="L2_EXT_WRITE_BEATS" title="L2 Cache Ext Writes" name="External write bytes" description="Number of external bus write bytes" multiplier="16" units="bytes" />
-        <event offset="186" counter="L2_EXT_W_STALL" title="L2 Cache Ext Writes" name="External bus stalls (W)" description="Number of cycles a valid write data (W channel) is stalled by the external interconnect" units="cycles" />
-        <event offset="187" counter="L2_EXT_W_BUF_FULL" title="L2 Cache Ext Writes" name="External bus write buffer full" description="Number of cycles a valid request is blocked by a full write buffer" units="cycles" />
+        <event offset="33" counter="L2_READ_LOOKUP" title="L2 Cache Reads" name="L2 read lookups" description="Number of reads into the L2 cache" units="requests"/>
+        <event offset="36" counter="L2_READ_SNOOP" title="L2 Cache Reads" name="Read snoops" description="Number of read transaction snoops" units="requests"/>
+        <event offset="37" counter="L2_READ_HIT" title="L2 Cache Reads" name="L2 read hits" description="Number of reads hitting in the L2 cache" units="requests"/>
+        <event offset="42" counter="L2_WRITE_SNOOP" title="L2 Cache Writes" name="Write snoops" description="Number of write transaction snoops" units="requests"/>
+        <event offset="43" counter="L2_WRITE_HIT" title="L2 Cache Writes" name="L2 write hits" description="Number of writes hitting in the L2 cache" units="requests"/>
+        <event offset="39" counter="L2_WRITE_LOOKUP" title="L2 Cache Writes" name="L2 write lookups" description="Number of writes into the L2 cache" units="requests"/>
+        <event offset="31" counter="L2_EXT_READ_BEATS" title="L2 Cache Ext Reads" name="External read bytes" description="Number of external bus read bytes" multiplier="16" units="bytes"/>
+        <event offset="54" counter="L2_EXT_AR_STALL" title="L2 Cache Ext Reads" name="External bus stalls (AR)" description="Number of cycles a valid read address (AR) is stalled by the external interconnect" units="cycles"/>
+        <event offset="60" counter="L2_EXT_R_BUF_FULL" title="L2 Cache Ext Reads" name="External bus response buffer full" description="Number of cycles a valid request is blocked by a full response buffer" units="cycles"/>
+        <event offset="56" counter="L2_EXT_RD_BUF_FULL" title="L2 Cache Ext Reads" name="External bus read data buffer full" description="Number of cycles a valid request is blocked by a full read data buffer" units="cycles"/>
+        <event offset="30" counter="L2_EXT_WRITE_BEATS" title="L2 Cache Ext Writes" name="External write bytes" description="Number of external bus write bytes" multiplier="16" units="bytes"/>
+        <event offset="58" counter="L2_EXT_W_STALL" title="L2 Cache Ext Writes" name="External bus stalls (W)" description="Number of cycles a valid write data (W channel) is stalled by the external interconnect" units="cycles"/>
+        <event offset="59" counter="L2_EXT_W_BUF_FULL" title="L2 Cache Ext Writes" name="External bus write buffer full" description="Number of cycles a valid request is blocked by a full write buffer" units="cycles"/>
+    </category>
+    <category name="Shader Core" per_cpu="no">
+        <event offset="26" counter="TRIPIPE_ACTIVE" title="Core Cycles" name="Tripipe cycles" description="Number of cycles tripipe was active" units="cycles"/>
+        <event offset="4" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment cycles" description="Number of cycles fragment processing was active" units="cycles"/>
+        <event offset="22" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Compute cycles" description="Number of cycles vertex\compute processing was active" units="cycles"/>
+        <event offset="19" counter="FRAG_CYCLES_NO_TILE" title="Core Cycles" name="Fragment cycles waiting for tile" description="Number of cycles spent waiting for a physical tile buffer" units="cycles"/>
+        <event offset="8" counter="FRAG_CYCLES_FPKQ_ACTIVE" title="Core Cycles" name="Fragment cycles pre-pipe buffer not empty" description="Number of cycles the pre-pipe queue contains quads" units="cycles"/>
+        <event offset="12" counter="FRAG_THREADS" title="Fragment Threads" name="Fragment threads" description="Number of fragment threads started" units="threads"/>
+        <event offset="13" counter="FRAG_DUMMY_THREADS" title="Fragment Threads" name="Dummy fragment threads" description="Number of dummy fragment threads started" units="threads"/>
+        <event offset="17" counter="FRAG_THREADS_LZS_TEST" title="Fragment Threads" name="Fragment threads doing late ZS" description="Number of threads doing late ZS test" units="threads"/>
+        <event offset="18" counter="FRAG_THREADS_LZS_KILLED" title="Fragment Threads" name="Fragment threads killed late ZS" description="Number of threads killed by late ZS test" units="threads"/>
+        <event offset="23" counter="COMPUTE_TASKS" title="Compute Tasks" name="Compute tasks" description="Number of compute tasks" units="tasks"/>
+        <event offset="24" counter="COMPUTE_THREADS" title="Compute Threads" name="Compute threads" description="Number of compute threads started" units="threads"/>
+        <event offset="5" counter="FRAG_PRIMITIVES" title="Fragment Primitives" name="Primitives loaded" description="Number of primitives loaded from tiler" units="primitives"/>
+        <event offset="6" counter="FRAG_PRIMITIVES_DROPPED" title="Fragment Primitives" name="Primitives dropped" description="Number of primitives dropped because out of tile" units="primitives"/>
+        <event offset="14" counter="FRAG_QUADS_RAST" title="Fragment Quads" name="Quads rasterized" description="Number of quads rasterized" units="quads"/>
+        <event offset="15" counter="FRAG_QUADS_EZS_TEST" title="Fragment Quads" name="Quads doing early ZS" description="Number of quads doing early ZS test" units="quads"/>
+        <event offset="16" counter="FRAG_QUADS_EZS_KILLED" title="Fragment Quads" name="Quads killed early Z" description="Number of quads killed by early ZS test" units="quads"/>
+        <event offset="20" counter="FRAG_NUM_TILES" title="Fragment" name="Pixels" description="Number of pixels rendered" multiplier="1024" units="pixels"/>
+        <event offset="21" counter="FRAG_TRANS_ELIM" title="Fragment Tasks" name="Tile writes killed by TE" description="Number of tile writes skipped by transaction elimination" units="tiles"/>
+        <event offset="27" counter="ARITH_WORDS" title="Arithmetic Pipe" name="A instructions" description="Number of batched instructions executed by the A-pipe (normalized per pipe)" units="instructions"/>
+        <event offset="31" counter="LS_WORDS" title="Load/Store Pipe" name="LS instructions" description="Number of instructions completed by the LS-pipe" units="instructions"/>
+        <event offset="32" counter="LS_ISSUES" title="Load/Store Pipe" name="LS instruction issues" description="Number of instructions issued to the LS-pipe, including restarts" units="instructions"/>
+        <event offset="38" counter="TEX_WORDS" title="Texture Pipe" name="T instructions" description="Number of instructions completed by the T-pipe" units="instructions"/>
+        <event offset="42" counter="TEX_ISSUES" title="Texture Pipe" name="T instruction issues" description="Number of threads through loop 2 address calculation" units="instructions"/>
+        <event offset="43" counter="TEX_RECIRC_FMISS" title="Texture Pipe" name="Cache misses" description="Number of instructions in the T-pipe, recirculated due to cache miss" units="requests"/>
+        <event offset="49" counter="LSC_READ_OP" title="Load/Store Cache Reads" name="Read operations" description="Number of read operations in the Load/Store cache" units="requests"/>
+        <event offset="48" counter="LSC_READ_HITS" title="Load/Store Cache Reads" name="Read hits" description="Number of read hits in the Load/Store cache" units="requests"/>
+        <event offset="51" counter="LSC_WRITE_OP" title="Load/Store Cache Writes" name="Write operations" description="Number of write operations in the Load/Store cache" units="requests"/>
+        <event offset="50" counter="LSC_WRITE_HITS" title="Load/Store Cache Writes" name="Write hits" description="Number of write hits in the Load/Store cache" units="requests"/>
+        <event offset="53" counter="LSC_ATOMIC_OP" title="Load/Store Cache Atomics" name="Atomic operations" description="Number of atomic operations in the Load/Store cache" units="requests"/>
+        <event offset="52" counter="LSC_ATOMIC_HITS" title="Load/Store Cache Atomics" name="Atomic hits" description="Number of atomic hits in the Load/Store cache" units="requests"/>
+        <event offset="54" counter="LSC_LINE_FETCHES" title="Load/Store Cache Bus" name="Line fetches" description="Number of line fetches in the Load/Store cache" units="requests"/>
+        <event offset="55" counter="LSC_DIRTY_LINE" title="Load/Store Cache Bus" name="Dirty line evictions" description="Number of dirty line evictions in the Load/Store cache" units="requests"/>
+        <event offset="56" counter="LSC_SNOOPS" title="Load/Store Cache Bus" name="Snoops in to LSC" description="Number of coherent memory snoops in to the Load/Store cache" units="requests"/>
     </category>
 </metrics>
\ No newline at end of file
diff --git a/src/panfrost/perf/T86x.xml b/src/panfrost/perf/T86x.xml
index d3a6d18a9c7..3dab882d43d 100644
--- a/src/panfrost/perf/T86x.xml
+++ b/src/panfrost/perf/T86x.xml
@@ -1,6 +1,6 @@
 <!--
 Copyright © 2017-2020 ARM Limited.
-Copyright © 2021 Collabora, Ltd.
+Copyright © 2021-2022 Collabora, Ltd.
 Author: Antonio Caggiano <antonio.caggiano@collabora.com>
 
 Permission is hereby granted, free of charge, to any person obtaining a
@@ -24,87 +24,87 @@ SOFTWARE.
 -->
 <metrics id="T86x">
     <category name="Job Manager" per_cpu="no">
-        <event offset="6" counter="GPU_ACTIVE" title="Job Manager Cycles" name="GPU cycles" description="Number of cycles GPU active" units="cycles" />
-        <event offset="7" counter="IRQ_ACTIVE" title="Job Manager Cycles" name="IRQ cycles" description="Number of cycles GPU interrupt pending" units="cycles" />
-        <event offset="10" counter="JS0_ACTIVE" title="Job Manager Cycles" name="JS0 cycles" description="Number of cycles JS0 (fragment) active" units="cycles" />
-        <event offset="18" counter="JS1_ACTIVE" title="Job Manager Cycles" name="JS1 cycles" description="Number of cycles JS1 (vertex/tiler/compute) active" units="cycles" />
-        <event offset="26" counter="JS2_ACTIVE" title="Job Manager Cycles" name="JS2 cycles" description="Number of cycles JS2 (vertex/compute) active" units="cycles" />
-        <event offset="8" counter="JS0_JOBS" title="Job Manager Jobs" name="JS0 jobs" description="Number of Jobs (fragment) completed in JS0" units="jobs" />
-        <event offset="16" counter="JS1_JOBS" title="Job Manager Jobs" name="JS1 jobs" description="Number of Jobs (vertex/tiler/compute) completed in JS1" units="jobs" />
-        <event offset="24" counter="JS2_JOBS" title="Job Manager Jobs" name="JS2 jobs" description="Number of Jobs (vertex/compute) completed in JS2" units="jobs" />
-        <event offset="9" counter="JS0_TASKS" title="Job Manager" name="Pixels" description="Number of Pixels completed in JS0" multiplier="1024" units="pixels" />
-        <event offset="17" counter="JS1_TASKS" title="Job Manager Tasks" name="JS1 tasks" description="Number of Tasks completed in JS1" units="tasks" />
-        <event offset="25" counter="JS2_TASKS" title="Job Manager Tasks" name="JS2 tasks" description="Number of Tasks completed in JS2" units="tasks" />
+        <event offset="6" counter="GPU_ACTIVE" title="Job Manager Cycles" name="GPU cycles" description="Number of cycles GPU active" units="cycles"/>
+        <event offset="7" counter="IRQ_ACTIVE" title="Job Manager Cycles" name="IRQ cycles" description="Number of cycles GPU interrupt pending" units="cycles"/>
+        <event offset="10" counter="JS0_ACTIVE" title="Job Manager Cycles" name="JS0 cycles" description="Number of cycles JS0 (fragment) active" units="cycles"/>
+        <event offset="18" counter="JS1_ACTIVE" title="Job Manager Cycles" name="JS1 cycles" description="Number of cycles JS1 (vertex/tiler/compute) active" units="cycles"/>
+        <event offset="26" counter="JS2_ACTIVE" title="Job Manager Cycles" name="JS2 cycles" description="Number of cycles JS2 (vertex/compute) active" units="cycles"/>
+        <event offset="8" counter="JS0_JOBS" title="Job Manager Jobs" name="JS0 jobs" description="Number of Jobs (fragment) completed in JS0" units="jobs"/>
+        <event offset="16" counter="JS1_JOBS" title="Job Manager Jobs" name="JS1 jobs" description="Number of Jobs (vertex/tiler/compute) completed in JS1" units="jobs"/>
+        <event offset="24" counter="JS2_JOBS" title="Job Manager Jobs" name="JS2 jobs" description="Number of Jobs (vertex/compute) completed in JS2" units="jobs"/>
+        <event offset="9" counter="JS0_TASKS" title="Job Manager" name="Pixels" description="Number of Pixels completed in JS0" multiplier="1024" units="pixels"/>
+        <event offset="17" counter="JS1_TASKS" title="Job Manager Tasks" name="JS1 tasks" description="Number of Tasks completed in JS1" units="tasks"/>
+        <event offset="25" counter="JS2_TASKS" title="Job Manager Tasks" name="JS2 tasks" description="Number of Tasks completed in JS2" units="tasks"/>
     </category>
     <category name="Tiler" per_cpu="no">
-        <event offset="109" counter="TI_ACTIVE" title="Tiler Cycles" name="Tiler cycles" description="Number of cycles Tiler active" units="cycles" />
-        <event offset="70" counter="TI_POLYGONS" title="Tiler Primitives" name="Polygons" description="Number of polygons processed" units="primitives" />
-        <event offset="69" counter="TI_QUADS" title="Tiler Primitives" name="Quads" description="Number of quads processed" units="primitives" />
-        <event offset="68" counter="TI_TRIANGLES" title="Tiler Primitives" name="Triangles" description="Number of triangles processed" units="primitives" />
-        <event offset="72" counter="TI_LINES" title="Tiler Primitives" name="Lines" description="Number of lines processed" units="primitives" />
-        <event offset="71" counter="TI_POINTS" title="Tiler Primitives" name="Points" description="Number of points processed" units="primitives" />
-        <event offset="75" counter="TI_FRONT_FACING" title="Tiler Culling" name="Front facing prims" description="Number of front facing primitives" units="primitives" />
-        <event offset="76" counter="TI_BACK_FACING" title="Tiler Culling" name="Back facing prims" description="Number of back facing primitives" units="primitives" />
-        <event offset="77" counter="TI_PRIM_VISIBLE" title="Tiler Culling" name="Visible prims" description="Number of visible primitives" units="primitives" />
-        <event offset="78" counter="TI_PRIM_CULLED" title="Tiler Culling" name="Culled prims" description="Number of culled primitives" units="primitives" />
-        <event offset="79" counter="TI_PRIM_CLIPPED" title="Tiler Culling" name="Clipped prims" description="Number of clipped primitives" units="primitives" />
-        <event offset="80" counter="TI_LEVEL0" title="Tiler Hierarchy" name="L0 prims" description="Number of primitives in hierarchy level 0" units="primitives" />
-        <event offset="81" counter="TI_LEVEL1" title="Tiler Hierarchy" name="L1 prims" description="Number of primitives in hierarchy level 1" units="primitives" />
-        <event offset="82" counter="TI_LEVEL2" title="Tiler Hierarchy" name="L2 prims" description="Number of primitives in hierarchy level 2" units="primitives" />
-        <event offset="83" counter="TI_LEVEL3" title="Tiler Hierarchy" name="L3 prims" description="Number of primitives in hierarchy level 3" units="primitives" />
-        <event offset="84" counter="TI_LEVEL4" title="Tiler Hierarchy" name="L4 prims" description="Number of primitives in hierarchy level 4" units="primitives" />
-        <event offset="85" counter="TI_LEVEL5" title="Tiler Hierarchy" name="L5 prims" description="Number of primitives in hierarchy level 5" units="primitives" />
-        <event offset="86" counter="TI_LEVEL6" title="Tiler Hierarchy" name="L6 prims" description="Number of primitives in hierarchy level 6" units="primitives" />
-        <event offset="87" counter="TI_LEVEL7" title="Tiler Hierarchy" name="L7 prims" description="Number of primitives in hierarchy level 7" units="primitives" />
-    </category>
-    <category name="Shader Core" per_cpu="no">
-        <event offset="218" counter="TRIPIPE_ACTIVE" title="Core Cycles" name="Tripipe cycles" description="Number of cycles tripipe was active" units="cycles" />
-        <event offset="196" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment cycles" description="Number of cycles fragment processing was active" units="cycles" />
-        <event offset="214" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Compute cycles" description="Number of cycles vertex\compute processing was active" units="cycles" />
-        <event offset="211" counter="FRAG_CYCLES_NO_TILE" title="Core Cycles" name="Fragment cycles waiting for tile" description="Number of cycles spent waiting for a physical tile buffer" units="cycles" />
-        <event offset="200" counter="FRAG_CYCLES_FPKQ_ACTIVE" title="Core Cycles" name="Fragment cycles pre-pipe buffer not empty" description="Number of cycles the pre-pipe queue contains quads" units="cycles" />
-        <event offset="204" counter="FRAG_THREADS" title="Fragment Threads" name="Fragment threads" description="Number of fragment threads started" units="threads" />
-        <event offset="205" counter="FRAG_DUMMY_THREADS" title="Fragment Threads" name="Dummy fragment threads" description="Number of dummy fragment threads started" units="threads" />
-        <event offset="209" counter="FRAG_THREADS_LZS_TEST" title="Fragment Threads" name="Fragment threads doing late ZS" description="Number of threads doing late ZS test" units="threads" />
-        <event offset="210" counter="FRAG_THREADS_LZS_KILLED" title="Fragment Threads" name="Fragment threads killed late ZS" description="Number of threads killed by late ZS test" units="threads" />
-        <event offset="215" counter="COMPUTE_TASKS" title="Compute Tasks" name="Compute tasks" description="Number of compute tasks" units="tasks" />
-        <event offset="216" counter="COMPUTE_THREADS" title="Compute Threads" name="Compute threads" description="Number of compute threads started" units="threads" />
-        <event offset="197" counter="FRAG_PRIMITIVES" title="Fragment Primitives" name="Primitives loaded" description="Number of primitives loaded from tiler" units="primitives" />
-        <event offset="198" counter="FRAG_PRIMITIVES_DROPPED" title="Fragment Primitives" name="Primitives dropped" description="Number of primitives dropped because out of tile" units="primitives" />
-        <event offset="206" counter="FRAG_QUADS_RAST" title="Fragment Quads" name="Quads rasterized" description="Number of quads rasterized" units="quads" />
-        <event offset="207" counter="FRAG_QUADS_EZS_TEST" title="Fragment Quads" name="Quads doing early ZS" description="Number of quads doing early ZS test" units="quads" />
-        <event offset="208" counter="FRAG_QUADS_EZS_KILLED" title="Fragment Quads" name="Quads killed early Z" description="Number of quads killed by early ZS test" units="quads" />
-        <event offset="212" counter="FRAG_NUM_TILES" title="Fragment" name="Pixels" description="Number of pixels rendered" multiplier="1024" units="pixels" />
-        <event offset="213" counter="FRAG_TRANS_ELIM" title="Fragment Tasks" name="Tile writes killed by TE" description="Number of tile writes skipped by transaction elimination" units="tiles" />
-        <event offset="219" counter="ARITH_WORDS" title="Arithmetic Pipe" name="A instructions" description="Number of instructions completed by the the A-pipe (normalized per pipeline)" units="instructions" />
-        <event offset="223" counter="LS_WORDS" title="Load/Store Pipe" name="LS instructions" description="Number of instructions completed by the LS-pipe" units="instructions" />
-        <event offset="224" counter="LS_ISSUES" title="Load/Store Pipe" name="LS instruction issues" description="Number of instructions issued to the LS-pipe, including restarts" units="instructions" />
-        <event offset="230" counter="TEX_WORDS" title="Texture Pipe" name="T instructions" description="Number of instructions completed by the T-pipe" units="instructions" />
-        <event offset="234" counter="TEX_ISSUES" title="Texture Pipe" name="T instruction issues" description="Number of threads through loop 2 address calculation" units="instructions" />
-        <event offset="235" counter="TEX_RECIRC_FMISS" title="Texture Pipe" name="Cache misses" description="Number of instructions in the T-pipe, recirculated due to cache miss" units="requests" />
-        <event offset="241" counter="LSC_READ_OP" title="Load/Store Cache Reads" name="Read operations" description="Number of read operations in the Load/Store cache" units="requests" />
-        <event offset="240" counter="LSC_READ_HITS" title="Load/Store Cache Reads" name="Read hits" description="Number of read hits in the Load/Store cache" units="requests" />
-        <event offset="243" counter="LSC_WRITE_OP" title="Load/Store Cache Writes" name="Write operations" description="Number of write operations in the Load/Store cache" units="requests" />
-        <event offset="242" counter="LSC_WRITE_HITS" title="Load/Store Cache Writes" name="Write hits" description="Number of write hits in the Load/Store cache" units="requests" />
-        <event offset="245" counter="LSC_ATOMIC_OP" title="Load/Store Cache Atomics" name="Atomic operations" description="Number of atomic operations in the Load/Store cache" units="requests" />
-        <event offset="244" counter="LSC_ATOMIC_HITS" title="Load/Store Cache Atomics" name="Atomic hits" description="Number of atomic hits in the Load/Store cache" units="requests" />
-        <event offset="246" counter="LSC_LINE_FETCHES" title="Load/Store Cache Bus" name="Line fetches" description="Number of line fetches in the Load/Store cache" units="requests" />
-        <event offset="247" counter="LSC_DIRTY_LINE" title="Load/Store Cache Bus" name="Dirty line evictions" description="Number of dirty line evictions in the Load/Store cache" units="requests" />
-        <event offset="248" counter="LSC_SNOOPS" title="Load/Store Cache Bus" name="Snoops in to LSC" description="Number of coherent memory snoops in to the Load/Store cache" units="requests" />
+        <event offset="45" counter="TI_ACTIVE" title="Tiler Cycles" name="Tiler cycles" description="Number of cycles Tiler active" units="cycles"/>
+        <event offset="6" counter="TI_POLYGONS" title="Tiler Primitives" name="Polygons" description="Number of polygons processed" units="primitives"/>
+        <event offset="5" counter="TI_QUADS" title="Tiler Primitives" name="Quads" description="Number of quads processed" units="primitives"/>
+        <event offset="4" counter="TI_TRIANGLES" title="Tiler Primitives" name="Triangles" description="Number of triangles processed" units="primitives"/>
+        <event offset="8" counter="TI_LINES" title="Tiler Primitives" name="Lines" description="Number of lines processed" units="primitives"/>
+        <event offset="7" counter="TI_POINTS" title="Tiler Primitives" name="Points" description="Number of points processed" units="primitives"/>
+        <event offset="11" counter="TI_FRONT_FACING" title="Tiler Culling" name="Front facing prims" description="Number of front facing primitives" units="primitives"/>
+        <event offset="12" counter="TI_BACK_FACING" title="Tiler Culling" name="Back facing prims" description="Number of back facing primitives" units="primitives"/>
+        <event offset="13" counter="TI_PRIM_VISIBLE" title="Tiler Culling" name="Visible prims" description="Number of visible primitives" units="primitives"/>
+        <event offset="14" counter="TI_PRIM_CULLED" title="Tiler Culling" name="Culled prims" description="Number of culled primitives" units="primitives"/>
+        <event offset="15" counter="TI_PRIM_CLIPPED" title="Tiler Culling" name="Clipped prims" description="Number of clipped primitives" units="primitives"/>
+        <event offset="16" counter="TI_LEVEL0" title="Tiler Hierarchy" name="L0 prims" description="Number of primitives in hierarchy level 0" units="primitives"/>
+        <event offset="17" counter="TI_LEVEL1" title="Tiler Hierarchy" name="L1 prims" description="Number of primitives in hierarchy level 1" units="primitives"/>
+        <event offset="18" counter="TI_LEVEL2" title="Tiler Hierarchy" name="L2 prims" description="Number of primitives in hierarchy level 2" units="primitives"/>
+        <event offset="19" counter="TI_LEVEL3" title="Tiler Hierarchy" name="L3 prims" description="Number of primitives in hierarchy level 3" units="primitives"/>
+        <event offset="20" counter="TI_LEVEL4" title="Tiler Hierarchy" name="L4 prims" description="Number of primitives in hierarchy level 4" units="primitives"/>
+        <event offset="21" counter="TI_LEVEL5" title="Tiler Hierarchy" name="L5 prims" description="Number of primitives in hierarchy level 5" units="primitives"/>
+        <event offset="22" counter="TI_LEVEL6" title="Tiler Hierarchy" name="L6 prims" description="Number of primitives in hierarchy level 6" units="primitives"/>
+        <event offset="23" counter="TI_LEVEL7" title="Tiler Hierarchy" name="L7 prims" description="Number of primitives in hierarchy level 7" units="primitives"/>
     </category>
     <category name="L2 Cache" per_cpu="no">
-        <event offset="161" counter="L2_READ_LOOKUP" title="L2 Cache Reads" name="L2 read lookups" description="Number of reads into the L2 cache" units="requests" />
-        <event offset="164" counter="L2_READ_SNOOP" title="L2 Cache Reads" name="Read snoops" description="Number of read transaction snoops" units="requests" />
-        <event offset="165" counter="L2_READ_HIT" title="L2 Cache Reads" name="L2 read hits" description="Number of reads hitting in the L2 cache" units="requests" />
-        <event offset="170" counter="L2_WRITE_SNOOP" title="L2 Cache Writes" name="Write snoops" description="Number of write transaction snoops" units="requests" />
-        <event offset="171" counter="L2_WRITE_HIT" title="L2 Cache Writes" name="L2 write hits" description="Number of writes hitting in the L2 cache" units="requests" />
-        <event offset="167" counter="L2_WRITE_LOOKUP" title="L2 Cache Writes" name="L2 write lookups" description="Number of writes into the L2 cache" units="requests" />
-        <event offset="159" counter="L2_EXT_READ_BEATS" title="L2 Cache Ext Reads" name="External read bytes" description="Number of external bus read bytes" multiplier="16" units="bytes" />
-        <event offset="182" counter="L2_EXT_AR_STALL" title="L2 Cache Ext Reads" name="External bus stalls (AR)" description="Number of cycles a valid read address (AR) is stalled by the external interconnect" units="cycles" />
-        <event offset="188" counter="L2_EXT_R_BUF_FULL" title="L2 Cache Ext Reads" name="External bus response buffer full" description="Number of cycles a valid request is blocked by a full response buffer" units="cycles" />
-        <event offset="184" counter="L2_EXT_RD_BUF_FULL" title="L2 Cache Ext Reads" name="External bus read data buffer full" description="Number of cycles a valid request is blocked by a full read data buffer" units="cycles" />
-        <event offset="158" counter="L2_EXT_WRITE_BEATS" title="L2 Cache Ext Writes" name="External write bytes" description="Number of external bus write bytes" multiplier="16" units="bytes" />
-        <event offset="186" counter="L2_EXT_W_STALL" title="L2 Cache Ext Writes" name="External bus stalls (W)" description="Number of cycles a valid write data (W channel) is stalled by the external interconnect" units="cycles" />
-        <event offset="187" counter="L2_EXT_W_BUF_FULL" title="L2 Cache Ext Writes" name="External bus write buffer full" description="Number of cycles a valid request is blocked by a full write buffer" units="cycles" />
+        <event offset="33" counter="L2_READ_LOOKUP" title="L2 Cache Reads" name="L2 read lookups" description="Number of reads into the L2 cache" units="requests"/>
+        <event offset="36" counter="L2_READ_SNOOP" title="L2 Cache Reads" name="Read snoops" description="Number of read transaction snoops" units="requests"/>
+        <event offset="37" counter="L2_READ_HIT" title="L2 Cache Reads" name="L2 read hits" description="Number of reads hitting in the L2 cache" units="requests"/>
+        <event offset="42" counter="L2_WRITE_SNOOP" title="L2 Cache Writes" name="Write snoops" description="Number of write transaction snoops" units="requests"/>
+        <event offset="43" counter="L2_WRITE_HIT" title="L2 Cache Writes" name="L2 write hits" description="Number of writes hitting in the L2 cache" units="requests"/>
+        <event offset="39" counter="L2_WRITE_LOOKUP" title="L2 Cache Writes" name="L2 write lookups" description="Number of writes into the L2 cache" units="requests"/>
+        <event offset="31" counter="L2_EXT_READ_BEATS" title="L2 Cache Ext Reads" name="External read bytes" description="Number of external bus read bytes" multiplier="16" units="bytes"/>
+        <event offset="54" counter="L2_EXT_AR_STALL" title="L2 Cache Ext Reads" name="External bus stalls (AR)" description="Number of cycles a valid read address (AR) is stalled by the external interconnect" units="cycles"/>
+        <event offset="60" counter="L2_EXT_R_BUF_FULL" title="L2 Cache Ext Reads" name="External bus response buffer full" description="Number of cycles a valid request is blocked by a full response buffer" units="cycles"/>
+        <event offset="56" counter="L2_EXT_RD_BUF_FULL" title="L2 Cache Ext Reads" name="External bus read data buffer full" description="Number of cycles a valid request is blocked by a full read data buffer" units="cycles"/>
+        <event offset="30" counter="L2_EXT_WRITE_BEATS" title="L2 Cache Ext Writes" name="External write bytes" description="Number of external bus write bytes" multiplier="16" units="bytes"/>
+        <event offset="58" counter="L2_EXT_W_STALL" title="L2 Cache Ext Writes" name="External bus stalls (W)" description="Number of cycles a valid write data (W channel) is stalled by the external interconnect" units="cycles"/>
+        <event offset="59" counter="L2_EXT_W_BUF_FULL" title="L2 Cache Ext Writes" name="External bus write buffer full" description="Number of cycles a valid request is blocked by a full write buffer" units="cycles"/>
+    </category>
+    <category name="Shader Core" per_cpu="no">
+        <event offset="26" counter="TRIPIPE_ACTIVE" title="Core Cycles" name="Tripipe cycles" description="Number of cycles tripipe was active" units="cycles"/>
+        <event offset="4" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment cycles" description="Number of cycles fragment processing was active" units="cycles"/>
+        <event offset="22" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Compute cycles" description="Number of cycles vertex\compute processing was active" units="cycles"/>
+        <event offset="19" counter="FRAG_CYCLES_NO_TILE" title="Core Cycles" name="Fragment cycles waiting for tile" description="Number of cycles spent waiting for a physical tile buffer" units="cycles"/>
+        <event offset="8" counter="FRAG_CYCLES_FPKQ_ACTIVE" title="Core Cycles" name="Fragment cycles pre-pipe buffer not empty" description="Number of cycles the pre-pipe queue contains quads" units="cycles"/>
+        <event offset="12" counter="FRAG_THREADS" title="Fragment Threads" name="Fragment threads" description="Number of fragment threads started" units="threads"/>
+        <event offset="13" counter="FRAG_DUMMY_THREADS" title="Fragment Threads" name="Dummy fragment threads" description="Number of dummy fragment threads started" units="threads"/>
+        <event offset="17" counter="FRAG_THREADS_LZS_TEST" title="Fragment Threads" name="Fragment threads doing late ZS" description="Number of threads doing late ZS test" units="threads"/>
+        <event offset="18" counter="FRAG_THREADS_LZS_KILLED" title="Fragment Threads" name="Fragment threads killed late ZS" description="Number of threads killed by late ZS test" units="threads"/>
+        <event offset="23" counter="COMPUTE_TASKS" title="Compute Tasks" name="Compute tasks" description="Number of compute tasks" units="tasks"/>
+        <event offset="24" counter="COMPUTE_THREADS" title="Compute Threads" name="Compute threads" description="Number of compute threads started" units="threads"/>
+        <event offset="5" counter="FRAG_PRIMITIVES" title="Fragment Primitives" name="Primitives loaded" description="Number of primitives loaded from tiler" units="primitives"/>
+        <event offset="6" counter="FRAG_PRIMITIVES_DROPPED" title="Fragment Primitives" name="Primitives dropped" description="Number of primitives dropped because out of tile" units="primitives"/>
+        <event offset="14" counter="FRAG_QUADS_RAST" title="Fragment Quads" name="Quads rasterized" description="Number of quads rasterized" units="quads"/>
+        <event offset="15" counter="FRAG_QUADS_EZS_TEST" title="Fragment Quads" name="Quads doing early ZS" description="Number of quads doing early ZS test" units="quads"/>
+        <event offset="16" counter="FRAG_QUADS_EZS_KILLED" title="Fragment Quads" name="Quads killed early Z" description="Number of quads killed by early ZS test" units="quads"/>
+        <event offset="20" counter="FRAG_NUM_TILES" title="Fragment" name="Pixels" description="Number of pixels rendered" multiplier="1024" units="pixels"/>
+        <event offset="21" counter="FRAG_TRANS_ELIM" title="Fragment Tasks" name="Tile writes killed by TE" description="Number of tile writes skipped by transaction elimination" units="tiles"/>
+        <event offset="27" counter="ARITH_WORDS" title="Arithmetic Pipe" name="A instructions" description="Number of instructions completed by the the A-pipe (normalized per pipeline)" units="instructions"/>
+        <event offset="31" counter="LS_WORDS" title="Load/Store Pipe" name="LS instructions" description="Number of instructions completed by the LS-pipe" units="instructions"/>
+        <event offset="32" counter="LS_ISSUES" title="Load/Store Pipe" name="LS instruction issues" description="Number of instructions issued to the LS-pipe, including restarts" units="instructions"/>
+        <event offset="38" counter="TEX_WORDS" title="Texture Pipe" name="T instructions" description="Number of instructions completed by the T-pipe" units="instructions"/>
+        <event offset="42" counter="TEX_ISSUES" title="Texture Pipe" name="T instruction issues" description="Number of threads through loop 2 address calculation" units="instructions"/>
+        <event offset="43" counter="TEX_RECIRC_FMISS" title="Texture Pipe" name="Cache misses" description="Number of instructions in the T-pipe, recirculated due to cache miss" units="requests"/>
+        <event offset="49" counter="LSC_READ_OP" title="Load/Store Cache Reads" name="Read operations" description="Number of read operations in the Load/Store cache" units="requests"/>
+        <event offset="48" counter="LSC_READ_HITS" title="Load/Store Cache Reads" name="Read hits" description="Number of read hits in the Load/Store cache" units="requests"/>
+        <event offset="51" counter="LSC_WRITE_OP" title="Load/Store Cache Writes" name="Write operations" description="Number of write operations in the Load/Store cache" units="requests"/>
+        <event offset="50" counter="LSC_WRITE_HITS" title="Load/Store Cache Writes" name="Write hits" description="Number of write hits in the Load/Store cache" units="requests"/>
+        <event offset="53" counter="LSC_ATOMIC_OP" title="Load/Store Cache Atomics" name="Atomic operations" description="Number of atomic operations in the Load/Store cache" units="requests"/>
+        <event offset="52" counter="LSC_ATOMIC_HITS" title="Load/Store Cache Atomics" name="Atomic hits" description="Number of atomic hits in the Load/Store cache" units="requests"/>
+        <event offset="54" counter="LSC_LINE_FETCHES" title="Load/Store Cache Bus" name="Line fetches" description="Number of line fetches in the Load/Store cache" units="requests"/>
+        <event offset="55" counter="LSC_DIRTY_LINE" title="Load/Store Cache Bus" name="Dirty line evictions" description="Number of dirty line evictions in the Load/Store cache" units="requests"/>
+        <event offset="56" counter="LSC_SNOOPS" title="Load/Store Cache Bus" name="Snoops in to LSC" description="Number of coherent memory snoops in to the Load/Store cache" units="requests"/>
     </category>
 </metrics>
\ No newline at end of file
diff --git a/src/panfrost/perf/T88x.xml b/src/panfrost/perf/T88x.xml
index ebbe0402be4..4a362d76984 100644
--- a/src/panfrost/perf/T88x.xml
+++ b/src/panfrost/perf/T88x.xml
@@ -1,6 +1,6 @@
 <!--
 Copyright © 2017-2020 ARM Limited.
-Copyright © 2021 Collabora, Ltd.
+Copyright © 2021-2022 Collabora, Ltd.
 Author: Antonio Caggiano <antonio.caggiano@collabora.com>
 
 Permission is hereby granted, free of charge, to any person obtaining a
@@ -24,87 +24,87 @@ SOFTWARE.
 -->
 <metrics id="T88x">
     <category name="Job Manager" per_cpu="no">
-        <event offset="6" counter="GPU_ACTIVE" title="Job Manager Cycles" name="GPU cycles" description="Number of cycles GPU active" units="cycles" />
-        <event offset="7" counter="IRQ_ACTIVE" title="Job Manager Cycles" name="IRQ cycles" description="Number of cycles GPU interrupt pending" units="cycles" />
-        <event offset="10" counter="JS0_ACTIVE" title="Job Manager Cycles" name="JS0 cycles" description="Number of cycles JS0 (fragment) active" units="cycles" />
-        <event offset="18" counter="JS1_ACTIVE" title="Job Manager Cycles" name="JS1 cycles" description="Number of cycles JS1 (vertex/tiler/compute) active" units="cycles" />
-        <event offset="26" counter="JS2_ACTIVE" title="Job Manager Cycles" name="JS2 cycles" description="Number of cycles JS2 (vertex/compute) active" units="cycles" />
-        <event offset="8" counter="JS0_JOBS" title="Job Manager Jobs" name="JS0 jobs" description="Number of Jobs (fragment) completed in JS0" units="jobs" />
-        <event offset="16" counter="JS1_JOBS" title="Job Manager Jobs" name="JS1 jobs" description="Number of Jobs (vertex/tiler/compute) completed in JS1" units="jobs" />
-        <event offset="24" counter="JS2_JOBS" title="Job Manager Jobs" name="JS2 jobs" description="Number of Jobs (vertex/compute) completed in JS2" units="jobs" />
-        <event offset="9" counter="JS0_TASKS" title="Job Manager" name="Pixels" description="Number of Pixels completed in JS0" multiplier="1024" units="pixels" />
-        <event offset="17" counter="JS1_TASKS" title="Job Manager Tasks" name="JS1 tasks" description="Number of Tasks completed in JS1" units="tasks" />
-        <event offset="25" counter="JS2_TASKS" title="Job Manager Tasks" name="JS2 tasks" description="Number of Tasks completed in JS2" units="tasks" />
+        <event offset="6" counter="GPU_ACTIVE" title="Job Manager Cycles" name="GPU cycles" description="Number of cycles GPU active" units="cycles"/>
+        <event offset="7" counter="IRQ_ACTIVE" title="Job Manager Cycles" name="IRQ cycles" description="Number of cycles GPU interrupt pending" units="cycles"/>
+        <event offset="10" counter="JS0_ACTIVE" title="Job Manager Cycles" name="JS0 cycles" description="Number of cycles JS0 (fragment) active" units="cycles"/>
+        <event offset="18" counter="JS1_ACTIVE" title="Job Manager Cycles" name="JS1 cycles" description="Number of cycles JS1 (vertex/tiler/compute) active" units="cycles"/>
+        <event offset="26" counter="JS2_ACTIVE" title="Job Manager Cycles" name="JS2 cycles" description="Number of cycles JS2 (vertex/compute) active" units="cycles"/>
+        <event offset="8" counter="JS0_JOBS" title="Job Manager Jobs" name="JS0 jobs" description="Number of Jobs (fragment) completed in JS0" units="jobs"/>
+        <event offset="16" counter="JS1_JOBS" title="Job Manager Jobs" name="JS1 jobs" description="Number of Jobs (vertex/tiler/compute) completed in JS1" units="jobs"/>
+        <event offset="24" counter="JS2_JOBS" title="Job Manager Jobs" name="JS2 jobs" description="Number of Jobs (vertex/compute) completed in JS2" units="jobs"/>
+        <event offset="9" counter="JS0_TASKS" title="Job Manager" name="Pixels" description="Number of Pixels completed in JS0" multiplier="1024" units="pixels"/>
+        <event offset="17" counter="JS1_TASKS" title="Job Manager Tasks" name="JS1 tasks" description="Number of Tasks completed in JS1" units="tasks"/>
+        <event offset="25" counter="JS2_TASKS" title="Job Manager Tasks" name="JS2 tasks" description="Number of Tasks completed in JS2" units="tasks"/>
     </category>
     <category name="Tiler" per_cpu="no">
-        <event offset="109" counter="TI_ACTIVE" title="Tiler Cycles" name="Tiler cycles" description="Number of cycles Tiler active" units="cycles" />
-        <event offset="70" counter="TI_POLYGONS" title="Tiler Primitives" name="Polygons" description="Number of polygons processed" units="primitives" />
-        <event offset="69" counter="TI_QUADS" title="Tiler Primitives" name="Quads" description="Number of quads processed" units="primitives" />
-        <event offset="68" counter="TI_TRIANGLES" title="Tiler Primitives" name="Triangles" description="Number of triangles processed" units="primitives" />
-        <event offset="72" counter="TI_LINES" title="Tiler Primitives" name="Lines" description="Number of lines processed" units="primitives" />
-        <event offset="71" counter="TI_POINTS" title="Tiler Primitives" name="Points" description="Number of points processed" units="primitives" />
-        <event offset="75" counter="TI_FRONT_FACING" title="Tiler Culling" name="Front facing prims" description="Number of front facing primitives" units="primitives" />
-        <event offset="76" counter="TI_BACK_FACING" title="Tiler Culling" name="Back facing prims" description="Number of back facing primitives" units="primitives" />
-        <event offset="77" counter="TI_PRIM_VISIBLE" title="Tiler Culling" name="Visible prims" description="Number of visible primitives" units="primitives" />
-        <event offset="78" counter="TI_PRIM_CULLED" title="Tiler Culling" name="Culled prims" description="Number of culled primitives" units="primitives" />
-        <event offset="79" counter="TI_PRIM_CLIPPED" title="Tiler Culling" name="Clipped prims" description="Number of clipped primitives" units="primitives" />
-        <event offset="80" counter="TI_LEVEL0" title="Tiler Hierarchy" name="L0 prims" description="Number of primitives in hierarchy level 0" units="primitives" />
-        <event offset="81" counter="TI_LEVEL1" title="Tiler Hierarchy" name="L1 prims" description="Number of primitives in hierarchy level 1" units="primitives" />
-        <event offset="82" counter="TI_LEVEL2" title="Tiler Hierarchy" name="L2 prims" description="Number of primitives in hierarchy level 2" units="primitives" />
-        <event offset="83" counter="TI_LEVEL3" title="Tiler Hierarchy" name="L3 prims" description="Number of primitives in hierarchy level 3" units="primitives" />
-        <event offset="84" counter="TI_LEVEL4" title="Tiler Hierarchy" name="L4 prims" description="Number of primitives in hierarchy level 4" units="primitives" />
-        <event offset="85" counter="TI_LEVEL5" title="Tiler Hierarchy" name="L5 prims" description="Number of primitives in hierarchy level 5" units="primitives" />
-        <event offset="86" counter="TI_LEVEL6" title="Tiler Hierarchy" name="L6 prims" description="Number of primitives in hierarchy level 6" units="primitives" />
-        <event offset="87" counter="TI_LEVEL7" title="Tiler Hierarchy" name="L7 prims" description="Number of primitives in hierarchy level 7" units="primitives" />
-    </category>
-    <category name="Shader Core" per_cpu="no">
-        <event offset="218" counter="TRIPIPE_ACTIVE" title="Core Cycles" name="Tripipe cycles" description="Number of cycles tripipe was active" units="cycles" />
-        <event offset="196" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment cycles" description="Number of cycles fragment processing was active" units="cycles" />
-        <event offset="214" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Compute cycles" description="Number of cycles vertex\compute processing was active" units="cycles" />
-        <event offset="211" counter="FRAG_CYCLES_NO_TILE" title="Core Cycles" name="Fragment cycles waiting for tile" description="Number of cycles spent waiting for a physical tile buffer" units="cycles" />
-        <event offset="200" counter="FRAG_CYCLES_FPKQ_ACTIVE" title="Core Cycles" name="Fragment cycles pre-pipe buffer not empty" description="Number of cycles the pre-pipe queue contains quads" units="cycles" />
-        <event offset="204" counter="FRAG_THREADS" title="Fragment Threads" name="Fragment threads" description="Number of fragment threads started" units="threads" />
-        <event offset="205" counter="FRAG_DUMMY_THREADS" title="Fragment Threads" name="Dummy fragment threads" description="Number of dummy fragment threads started" units="threads" />
-        <event offset="209" counter="FRAG_THREADS_LZS_TEST" title="Fragment Threads" name="Fragment threads doing late ZS" description="Number of threads doing late ZS test" units="threads" />
-        <event offset="210" counter="FRAG_THREADS_LZS_KILLED" title="Fragment Threads" name="Fragment threads killed late ZS" description="Number of threads killed by late ZS test" units="threads" />
-        <event offset="215" counter="COMPUTE_TASKS" title="Compute Tasks" name="Compute tasks" description="Number of compute tasks" units="tasks" />
-        <event offset="216" counter="COMPUTE_THREADS" title="Compute Threads" name="Compute threads" description="Number of compute threads started" units="threads" />
-        <event offset="197" counter="FRAG_PRIMITIVES" title="Fragment Primitives" name="Primitives loaded" description="Number of primitives loaded from tiler" units="primitives" />
-        <event offset="198" counter="FRAG_PRIMITIVES_DROPPED" title="Fragment Primitives" name="Primitives dropped" description="Number of primitives dropped because out of tile" units="primitives" />
-        <event offset="206" counter="FRAG_QUADS_RAST" title="Fragment Quads" name="Quads rasterized" description="Number of quads rasterized" units="quads" />
-        <event offset="207" counter="FRAG_QUADS_EZS_TEST" title="Fragment Quads" name="Quads doing early ZS" description="Number of quads doing early ZS test" units="quads" />
-        <event offset="208" counter="FRAG_QUADS_EZS_KILLED" title="Fragment Quads" name="Quads killed early Z" description="Number of quads killed by early ZS test" units="quads" />
-        <event offset="212" counter="FRAG_NUM_TILES" title="Fragment" name="Pixels" description="Number of pixels rendered" multiplier="1024" units="pixels" />
-        <event offset="213" counter="FRAG_TRANS_ELIM" title="Fragment Tasks" name="Tile writes killed by TE" description="Number of tile writes skipped by transaction elimination" units="tiles" />
-        <event offset="219" counter="ARITH_WORDS" title="Arithmetic Pipe" name="A instructions" description="Number of instructions completed by the the A-pipe (normalized per pipeline)" units="instructions" />
-        <event offset="223" counter="LS_WORDS" title="Load/Store Pipe" name="LS instructions" description="Number of instructions completed by the LS-pipe" units="instructions" />
-        <event offset="224" counter="LS_ISSUES" title="Load/Store Pipe" name="LS instruction issues" description="Number of instructions issued to the LS-pipe, including restarts" units="instructions" />
-        <event offset="230" counter="TEX_WORDS" title="Texture Pipe" name="T instructions" description="Number of instructions completed by the T-pipe" units="instructions" />
-        <event offset="234" counter="TEX_ISSUES" title="Texture Pipe" name="T instruction issues" description="Number of threads through loop 2 address calculation" units="instructions" />
-        <event offset="235" counter="TEX_RECIRC_FMISS" title="Texture Pipe" name="Cache misses" description="Number of instructions in the T-pipe, recirculated due to cache miss" units="requests" />
-        <event offset="241" counter="LSC_READ_OP" title="Load/Store Cache Reads" name="Read operations" description="Number of read operations in the Load/Store cache" units="requests" />
-        <event offset="240" counter="LSC_READ_HITS" title="Load/Store Cache Reads" name="Read hits" description="Number of read hits in the Load/Store cache" units="requests" />
-        <event offset="243" counter="LSC_WRITE_OP" title="Load/Store Cache Writes" name="Write operations" description="Number of write operations in the Load/Store cache" units="requests" />
-        <event offset="242" counter="LSC_WRITE_HITS" title="Load/Store Cache Writes" name="Write hits" description="Number of write hits in the Load/Store cache" units="requests" />
-        <event offset="245" counter="LSC_ATOMIC_OP" title="Load/Store Cache Atomics" name="Atomic operations" description="Number of atomic operations in the Load/Store cache" units="requests" />
-        <event offset="244" counter="LSC_ATOMIC_HITS" title="Load/Store Cache Atomics" name="Atomic hits" description="Number of atomic hits in the Load/Store cache" units="requests" />
-        <event offset="246" counter="LSC_LINE_FETCHES" title="Load/Store Cache Bus" name="Line fetches" description="Number of line fetches in the Load/Store cache" units="requests" />
-        <event offset="247" counter="LSC_DIRTY_LINE" title="Load/Store Cache Bus" name="Dirty line evictions" description="Number of dirty line evictions in the Load/Store cache" units="requests" />
-        <event offset="248" counter="LSC_SNOOPS" title="Load/Store Cache Bus" name="Snoops in to LSC" description="Number of coherent memory snoops in to the Load/Store cache" units="requests" />
+        <event offset="45" counter="TI_ACTIVE" title="Tiler Cycles" name="Tiler cycles" description="Number of cycles Tiler active" units="cycles"/>
+        <event offset="6" counter="TI_POLYGONS" title="Tiler Primitives" name="Polygons" description="Number of polygons processed" units="primitives"/>
+        <event offset="5" counter="TI_QUADS" title="Tiler Primitives" name="Quads" description="Number of quads processed" units="primitives"/>
+        <event offset="4" counter="TI_TRIANGLES" title="Tiler Primitives" name="Triangles" description="Number of triangles processed" units="primitives"/>
+        <event offset="8" counter="TI_LINES" title="Tiler Primitives" name="Lines" description="Number of lines processed" units="primitives"/>
+        <event offset="7" counter="TI_POINTS" title="Tiler Primitives" name="Points" description="Number of points processed" units="primitives"/>
+        <event offset="11" counter="TI_FRONT_FACING" title="Tiler Culling" name="Front facing prims" description="Number of front facing primitives" units="primitives"/>
+        <event offset="12" counter="TI_BACK_FACING" title="Tiler Culling" name="Back facing prims" description="Number of back facing primitives" units="primitives"/>
+        <event offset="13" counter="TI_PRIM_VISIBLE" title="Tiler Culling" name="Visible prims" description="Number of visible primitives" units="primitives"/>
+        <event offset="14" counter="TI_PRIM_CULLED" title="Tiler Culling" name="Culled prims" description="Number of culled primitives" units="primitives"/>
+        <event offset="15" counter="TI_PRIM_CLIPPED" title="Tiler Culling" name="Clipped prims" description="Number of clipped primitives" units="primitives"/>
+        <event offset="16" counter="TI_LEVEL0" title="Tiler Hierarchy" name="L0 prims" description="Number of primitives in hierarchy level 0" units="primitives"/>
+        <event offset="17" counter="TI_LEVEL1" title="Tiler Hierarchy" name="L1 prims" description="Number of primitives in hierarchy level 1" units="primitives"/>
+        <event offset="18" counter="TI_LEVEL2" title="Tiler Hierarchy" name="L2 prims" description="Number of primitives in hierarchy level 2" units="primitives"/>
+        <event offset="19" counter="TI_LEVEL3" title="Tiler Hierarchy" name="L3 prims" description="Number of primitives in hierarchy level 3" units="primitives"/>
+        <event offset="20" counter="TI_LEVEL4" title="Tiler Hierarchy" name="L4 prims" description="Number of primitives in hierarchy level 4" units="primitives"/>
+        <event offset="21" counter="TI_LEVEL5" title="Tiler Hierarchy" name="L5 prims" description="Number of primitives in hierarchy level 5" units="primitives"/>
+        <event offset="22" counter="TI_LEVEL6" title="Tiler Hierarchy" name="L6 prims" description="Number of primitives in hierarchy level 6" units="primitives"/>
+        <event offset="23" counter="TI_LEVEL7" title="Tiler Hierarchy" name="L7 prims" description="Number of primitives in hierarchy level 7" units="primitives"/>
     </category>
     <category name="L2 Cache" per_cpu="no">
-        <event offset="161" counter="L2_READ_LOOKUP" title="L2 Cache Reads" name="L2 read lookups" description="Number of reads into the L2 cache" units="requests" />
-        <event offset="164" counter="L2_READ_SNOOP" title="L2 Cache Reads" name="Read snoops" description="Number of read transaction snoops" units="requests" />
-        <event offset="165" counter="L2_READ_HIT" title="L2 Cache Reads" name="L2 read hits" description="Number of reads hitting in the L2 cache" units="requests" />
-        <event offset="170" counter="L2_WRITE_SNOOP" title="L2 Cache Writes" name="Write snoops" description="Number of write transaction snoops" units="requests" />
-        <event offset="171" counter="L2_WRITE_HIT" title="L2 Cache Writes" name="L2 write hits" description="Number of writes hitting in the L2 cache" units="requests" />
-        <event offset="167" counter="L2_WRITE_LOOKUP" title="L2 Cache Writes" name="L2 write lookups" description="Number of writes into the L2 cache" units="requests" />
-        <event offset="159" counter="L2_EXT_READ_BEATS" title="L2 Cache Ext Reads" name="External read bytes" description="Number of external bus read bytes" multiplier="16" units="bytes" />
-        <event offset="182" counter="L2_EXT_AR_STALL" title="L2 Cache Ext Reads" name="External bus stalls (AR)" description="Number of cycles a valid read address (AR) is stalled by the external interconnect" units="cycles" />
-        <event offset="188" counter="L2_EXT_R_BUF_FULL" title="L2 Cache Ext Reads" name="External bus response buffer full" description="Number of cycles a valid request is blocked by a full response buffer" units="cycles" />
-        <event offset="184" counter="L2_EXT_RD_BUF_FULL" title="L2 Cache Ext Reads" name="External bus read data buffer full" description="Number of cycles a valid request is blocked by a full read data buffer" units="cycles" />
-        <event offset="158" counter="L2_EXT_WRITE_BEATS" title="L2 Cache Ext Writes" name="External write bytes" description="Number of external bus write bytes" multiplier="16" units="bytes" />
-        <event offset="186" counter="L2_EXT_W_STALL" title="L2 Cache Ext Writes" name="External bus stalls (W)" description="Number of cycles a valid write data (W channel) is stalled by the external interconnect" units="cycles" />
-        <event offset="187" counter="L2_EXT_W_BUF_FULL" title="L2 Cache Ext Writes" name="External bus write buffer full" description="Number of cycles a valid request is blocked by a full write buffer" units="cycles" />
+        <event offset="33" counter="L2_READ_LOOKUP" title="L2 Cache Reads" name="L2 read lookups" description="Number of reads into the L2 cache" units="requests"/>
+        <event offset="36" counter="L2_READ_SNOOP" title="L2 Cache Reads" name="Read snoops" description="Number of read transaction snoops" units="requests"/>
+        <event offset="37" counter="L2_READ_HIT" title="L2 Cache Reads" name="L2 read hits" description="Number of reads hitting in the L2 cache" units="requests"/>
+        <event offset="42" counter="L2_WRITE_SNOOP" title="L2 Cache Writes" name="Write snoops" description="Number of write transaction snoops" units="requests"/>
+        <event offset="43" counter="L2_WRITE_HIT" title="L2 Cache Writes" name="L2 write hits" description="Number of writes hitting in the L2 cache" units="requests"/>
+        <event offset="39" counter="L2_WRITE_LOOKUP" title="L2 Cache Writes" name="L2 write lookups" description="Number of writes into the L2 cache" units="requests"/>
+        <event offset="31" counter="L2_EXT_READ_BEATS" title="L2 Cache Ext Reads" name="External read bytes" description="Number of external bus read bytes" multiplier="16" units="bytes"/>
+        <event offset="54" counter="L2_EXT_AR_STALL" title="L2 Cache Ext Reads" name="External bus stalls (AR)" description="Number of cycles a valid read address (AR) is stalled by the external interconnect" units="cycles"/>
+        <event offset="60" counter="L2_EXT_R_BUF_FULL" title="L2 Cache Ext Reads" name="External bus response buffer full" description="Number of cycles a valid request is blocked by a full response buffer" units="cycles"/>
+        <event offset="56" counter="L2_EXT_RD_BUF_FULL" title="L2 Cache Ext Reads" name="External bus read data buffer full" description="Number of cycles a valid request is blocked by a full read data buffer" units="cycles"/>
+        <event offset="30" counter="L2_EXT_WRITE_BEATS" title="L2 Cache Ext Writes" name="External write bytes" description="Number of external bus write bytes" multiplier="16" units="bytes"/>
+        <event offset="58" counter="L2_EXT_W_STALL" title="L2 Cache Ext Writes" name="External bus stalls (W)" description="Number of cycles a valid write data (W channel) is stalled by the external interconnect" units="cycles"/>
+        <event offset="59" counter="L2_EXT_W_BUF_FULL" title="L2 Cache Ext Writes" name="External bus write buffer full" description="Number of cycles a valid request is blocked by a full write buffer" units="cycles"/>
+    </category>
+    <category name="Shader Core" per_cpu="no">
+        <event offset="26" counter="TRIPIPE_ACTIVE" title="Core Cycles" name="Tripipe cycles" description="Number of cycles tripipe was active" units="cycles"/>
+        <event offset="4" counter="FRAG_ACTIVE" title="Core Cycles" name="Fragment cycles" description="Number of cycles fragment processing was active" units="cycles"/>
+        <event offset="22" counter="COMPUTE_ACTIVE" title="Core Cycles" name="Compute cycles" description="Number of cycles vertex\compute processing was active" units="cycles"/>
+        <event offset="19" counter="FRAG_CYCLES_NO_TILE" title="Core Cycles" name="Fragment cycles waiting for tile" description="Number of cycles spent waiting for a physical tile buffer" units="cycles"/>
+        <event offset="8" counter="FRAG_CYCLES_FPKQ_ACTIVE" title="Core Cycles" name="Fragment cycles pre-pipe buffer not empty" description="Number of cycles the pre-pipe queue contains quads" units="cycles"/>
+        <event offset="12" counter="FRAG_THREADS" title="Fragment Threads" name="Fragment threads" description="Number of fragment threads started" units="threads"/>
+        <event offset="13" counter="FRAG_DUMMY_THREADS" title="Fragment Threads" name="Dummy fragment threads" description="Number of dummy fragment threads started" units="threads"/>
+        <event offset="17" counter="FRAG_THREADS_LZS_TEST" title="Fragment Threads" name="Fragment threads doing late ZS" description="Number of threads doing late ZS test" units="threads"/>
+        <event offset="18" counter="FRAG_THREADS_LZS_KILLED" title="Fragment Threads" name="Fragment threads killed late ZS" description="Number of threads killed by late ZS test" units="threads"/>
+        <event offset="23" counter="COMPUTE_TASKS" title="Compute Tasks" name="Compute tasks" description="Number of compute tasks" units="tasks"/>
+        <event offset="24" counter="COMPUTE_THREADS" title="Compute Threads" name="Compute threads" description="Number of compute threads started" units="threads"/>
+        <event offset="5" counter="FRAG_PRIMITIVES" title="Fragment Primitives" name="Primitives loaded" description="Number of primitives loaded from tiler" units="primitives"/>
+        <event offset="6" counter="FRAG_PRIMITIVES_DROPPED" title="Fragment Primitives" name="Primitives dropped" description="Number of primitives dropped because out of tile" units="primitives"/>
+        <event offset="14" counter="FRAG_QUADS_RAST" title="Fragment Quads" name="Quads rasterized" description="Number of quads rasterized" units="quads"/>
+        <event offset="15" counter="FRAG_QUADS_EZS_TEST" title="Fragment Quads" name="Quads doing early ZS" description="Number of quads doing early ZS test" units="quads"/>
+        <event offset="16" counter="FRAG_QUADS_EZS_KILLED" title="Fragment Quads" name="Quads killed early Z" description="Number of quads killed by early ZS test" units="quads"/>
+        <event offset="20" counter="FRAG_NUM_TILES" title="Fragment" name="Pixels" description="Number of pixels rendered" multiplier="1024" units="pixels"/>
+        <event offset="21" counter="FRAG_TRANS_ELIM" title="Fragment Tasks" name="Tile writes killed by TE" description="Number of tile writes skipped by transaction elimination" units="tiles"/>
+        <event offset="27" counter="ARITH_WORDS" title="Arithmetic Pipe" name="A instructions" description="Number of instructions completed by the the A-pipe (normalized per pipeline)" units="instructions"/>
+        <event offset="31" counter="LS_WORDS" title="Load/Store Pipe" name="LS instructions" description="Number of instructions completed by the LS-pipe" units="instructions"/>
+        <event offset="32" counter="LS_ISSUES" title="Load/Store Pipe" name="LS instruction issues" description="Number of instructions issued to the LS-pipe, including restarts" units="instructions"/>
+        <event offset="38" counter="TEX_WORDS" title="Texture Pipe" name="T instructions" description="Number of instructions completed by the T-pipe" units="instructions"/>
+        <event offset="42" counter="TEX_ISSUES" title="Texture Pipe" name="T instruction issues" description="Number of threads through loop 2 address calculation" units="instructions"/>
+        <event offset="43" counter="TEX_RECIRC_FMISS" title="Texture Pipe" name="Cache misses" description="Number of instructions in the T-pipe, recirculated due to cache miss" units="requests"/>
+        <event offset="49" counter="LSC_READ_OP" title="Load/Store Cache Reads" name="Read operations" description="Number of read operations in the Load/Store cache" units="requests"/>
+        <event offset="48" counter="LSC_READ_HITS" title="Load/Store Cache Reads" name="Read hits" description="Number of read hits in the Load/Store cache" units="requests"/>
+        <event offset="51" counter="LSC_WRITE_OP" title="Load/Store Cache Writes" name="Write operations" description="Number of write operations in the Load/Store cache" units="requests"/>
+        <event offset="50" counter="LSC_WRITE_HITS" title="Load/Store Cache Writes" name="Write hits" description="Number of write hits in the Load/Store cache" units="requests"/>
+        <event offset="53" counter="LSC_ATOMIC_OP" title="Load/Store Cache Atomics" name="Atomic operations" description="Number of atomic operations in the Load/Store cache" units="requests"/>
+        <event offset="52" counter="LSC_ATOMIC_HITS" title="Load/Store Cache Atomics" name="Atomic hits" description="Number of atomic hits in the Load/Store cache" units="requests"/>
+        <event offset="54" counter="LSC_LINE_FETCHES" title="Load/Store Cache Bus" name="Line fetches" description="Number of line fetches in the Load/Store cache" units="requests"/>
+        <event offset="55" counter="LSC_DIRTY_LINE" title="Load/Store Cache Bus" name="Dirty line evictions" description="Number of dirty line evictions in the Load/Store cache" units="requests"/>
+        <event offset="56" counter="LSC_SNOOPS" title="Load/Store Cache Bus" name="Snoops in to LSC" description="Number of coherent memory snoops in to the Load/Store cache" units="requests"/>
     </category>
 </metrics>
\ No newline at end of file
diff --git a/src/panfrost/perf/pan_gen_perf.py b/src/panfrost/perf/pan_gen_perf.py
index bfcb8ea9a22..de0672c8392 100644
--- a/src/panfrost/perf/pan_gen_perf.py
+++ b/src/panfrost/perf/pan_gen_perf.py
@@ -200,7 +200,7 @@ def main():
             c.write(".symbol_name = \"%s\"," % (counter.underscore_name))
             c.write(".units = PAN_PERF_COUNTER_UNITS_%s," % (counter.units.upper()))
             c.write(".offset = %u," % (counter.offset))
-            c.write(".category = &%s.categories[%u]," % (current_struct_name, i))
+            c.write(".category_index = %u," % i)
 
             c.outdent(tab_size)
             c.write("}, // counter")
diff --git a/src/panfrost/perf/pan_perf.c b/src/panfrost/perf/pan_perf.c
index 06f81cb505d..930eae78d11 100644
--- a/src/panfrost/perf/pan_perf.c
+++ b/src/panfrost/perf/pan_perf.c
@@ -28,19 +28,22 @@
 #include <drm-uapi/panfrost_drm.h>
 
 #define PAN_COUNTERS_PER_CATEGORY 64
-#define PAN_SHADER_CORE_INDEX 2
+#define PAN_SHADER_CORE_INDEX 3
 
 uint32_t
 panfrost_perf_counter_read(const struct panfrost_perf_counter *counter,
                            const struct panfrost_perf *perf)
 {
-   assert(counter->offset < perf->n_counter_values);
-   uint32_t ret = perf->counter_values[counter->offset];
+   unsigned offset = perf->category_offset[counter->category_index];
+   offset += counter->offset;
+   assert(offset < perf->n_counter_values);
+
+   uint32_t ret = perf->counter_values[offset];
 
    // If counter belongs to shader core, accumulate values for all other cores
-   if (counter->category == &perf->cfg->categories[PAN_SHADER_CORE_INDEX]) {
+   if (counter->category_index == PAN_SHADER_CORE_INDEX) {
       for (uint32_t core = 1; core < perf->dev->core_count; ++core) {
-         ret += perf->counter_values[counter->offset + PAN_COUNTERS_PER_CATEGORY * core];
+         ret += perf->counter_values[offset + PAN_COUNTERS_PER_CATEGORY * core];
       }
    }
 
@@ -72,10 +75,17 @@ panfrost_perf_init(struct panfrost_perf *perf, struct panfrost_device *dev)
            unreachable("Performance counters missing!");
 
    // Generally counter blocks are laid out in the following order:
-   // Job manager, tiler, L2 cache, and one or more shader cores.
-   uint32_t n_blocks = 3 + dev->core_count;
+   // Job manager, tiler, one or more L2 caches, and one or more shader cores.
+   unsigned l2_slices = panfrost_query_l2_slices(dev);
+   uint32_t n_blocks = 2 + l2_slices + dev->core_count;
    perf->n_counter_values = PAN_COUNTERS_PER_CATEGORY * n_blocks;
    perf->counter_values = ralloc_array(perf, uint32_t, perf->n_counter_values);
+
+   /* Setup the layout */
+   perf->category_offset[0] = PAN_COUNTERS_PER_CATEGORY * 0;
+   perf->category_offset[1] = PAN_COUNTERS_PER_CATEGORY * 1;
+   perf->category_offset[2] = PAN_COUNTERS_PER_CATEGORY * 2;
+   perf->category_offset[3] = PAN_COUNTERS_PER_CATEGORY * (2 + l2_slices);
 }
 
 static int
diff --git a/src/panfrost/perf/pan_perf.h b/src/panfrost/perf/pan_perf.h
index 1a42b7f351d..592dce71cea 100644
--- a/src/panfrost/perf/pan_perf.h
+++ b/src/panfrost/perf/pan_perf.h
@@ -60,10 +60,9 @@ struct panfrost_perf_counter {
    const char *desc;
    const char *symbol_name;
    enum panfrost_perf_counter_units units;
-   // Offset of this counter's value within the counters memory block
+   // Offset of this counter's value within the category
    uint32_t offset;
-
-   const struct panfrost_perf_category *category;
+   unsigned category_index;
 };
 
 struct panfrost_perf_category {
@@ -71,6 +70,9 @@ struct panfrost_perf_category {
 
    struct panfrost_perf_counter counters[PAN_PERF_MAX_COUNTERS];
    uint32_t n_counters;
+
+   /* Offset of this category within the counters memory block */
+   unsigned offset;
 };
 
 struct panfrost_perf_config {
@@ -88,6 +90,9 @@ struct panfrost_perf {
    // Memory where to dump counter values
    uint32_t *counter_values;
    uint32_t n_counter_values;
+
+   /* Offsets of categories */
+   unsigned category_offset[PAN_PERF_MAX_CATEGORIES];
 };
 
 uint32_t