Commit Graph

193069 Commits

Author SHA1 Message Date
Alyssa Rosenzweig 27d39bf544 agx: reassociate bcsel with ior/iand
fairly dramatic improvement to citra shaders, citra seems to be a % faster with
this.

total instructions in shared programs: 2910477 -> 2909192 (-0.04%)
instructions in affected programs: 983562 -> 982277 (-0.13%)
helped: 529
HURT: 1
Instructions are helped.

total bytes in shared programs: 19400010 -> 19392214 (-0.04%)
bytes in affected programs: 6785058 -> 6777262 (-0.11%)
helped: 529
HURT: 58
Bytes are helped.

total halfregs in shared programs: 519553 -> 519429 (-0.02%)
halfregs in affected programs: 1877 -> 1753 (-6.61%)
helped: 37
HURT: 12
Halfregs are helped.

total spills in shared programs: 135040 -> 135022 (-0.01%)
spills in affected programs: 135040 -> 135022 (-0.01%)
helped: 6
HURT: 0

total fills in shared programs: 342326 -> 342308 (<.01%)
fills in affected programs: 342326 -> 342308 (<.01%)
helped: 6
HURT: 0

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:30 +00:00
Alyssa Rosenzweig b45d54a388 agx: optimize b2x(inot)
easier to do as algebraic rules.

total instructions in shared programs: 2915611 -> 2910477 (-0.18%)
instructions in affected programs: 285251 -> 280117 (-1.80%)
helped: 1245
HURT: 5
Instructions are helped.

total bytes in shared programs: 19439752 -> 19400010 (-0.20%)
bytes in affected programs: 1881556 -> 1841814 (-2.11%)
helped: 1244
HURT: 6
Bytes are helped.

total halfregs in shared programs: 519253 -> 519553 (0.06%)
halfregs in affected programs: 5509 -> 5809 (5.45%)
helped: 24
HURT: 107
Halfregs are HURT.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:30 +00:00
Alyssa Rosenzweig 34ca925064 agx: don't inline imms into stack_store
we need opt_large_consts..

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:30 +00:00
Alyssa Rosenzweig f6c17cbe48 ail: use template for tiled memcpy
less macro hell

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:30 +00:00
Alyssa Rosenzweig 075e595058 ail: port tiling routines to c++
to fix macro hell

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:30 +00:00
Alyssa Rosenzweig 104ba72547 ail: add tests for linear<-->twiddled copies
based on the panfrost tests.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:30 +00:00
Alyssa Rosenzweig 2c9189434f docs/asahi: fix strided linear note
turns out linear 2d arrays work

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:30 +00:00
Alyssa Rosenzweig f836d573e4 agx: only run early tests if needed
if it'd be early anyway don't bother, speeds up dolphin with spec shaders

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:30 +00:00
Alyssa Rosenzweig 7825044c39 asahi: honour discard_whole | persistent
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:30 +00:00
Alyssa Rosenzweig 793ac04077 asahi: handle read-only SSBOs
reduces silliness in Yuzu.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:30 +00:00
Alyssa Rosenzweig 97a945246c asahi: fix valid buffer tracking for SSBO/image/XFB
smaller ranges.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:30 +00:00
Alyssa Rosenzweig 6689b083d8 asahi: don't sync for uninitialized buffer
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:30 +00:00
Alyssa Rosenzweig 9e8ce4550c asahi: enable tcs caching
works fine

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:30 +00:00
Alyssa Rosenzweig 19da2f6d30 asahi: sync with query mismatches
otherwise the following assertion can fail. fixes crash in yuzu trace.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:30 +00:00
Alyssa Rosenzweig 3b927567ac Revert "asahi: don't canonicalize nans/flush denorms when copying"
This reverts commit ab7fb3d400.

This breaks ./arb_copy_image-formats -auto --samples=4 since we need the blit
path for compression transitions. The correct thing to do is cast the formats.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:30 +00:00
Asahi Lina 74110ef21d asahi: batch: Trace before waiting for syncobj
If the render hangs, at least this way we get the trace.

Signed-off-by: Asahi Lina <lina@asahilina.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:30 +00:00
Asahi Lina ff900ceffe asahi: Enable scratch debugging
Via ASAHI_MESA_DEBUG=scratch. This will assert if enabled and the
scratch workgroup allocation count turns out not to be sufficient (that
is, there were failed allocations), to help debug the max occupancy
calculation.

Signed-off-by: Asahi Lina <lina@asahilina.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:30 +00:00
Asahi Lina 8845702f8c asahi: Allocate scratch for shaders
Finally, now that scratch is available, tell the hardware how much we
need and allocate the scratch buffers as necessary.

We only do the max subgroup calculation for compute right now, and it's
probably conservative (we could probably calculate a reduced occupancy
depending on register pressure). Getting this too low won't break, the
GPU will just be less efficient if there is not enough scratch
allocated. Vertex and fragment just get the max 96 for now.

Signed-off-by: Asahi Lina <lina@asahilina.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:30 +00:00
Asahi Lina bb5277aa3d asahi: Hook up scratch
Instantiate our 3 scratch managers for each shader type, and pass them
to the hardware if a batch uses scratch.

Signed-off-by: Asahi Lina <lina@asahilina.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:30 +00:00
Asahi Lina fa475c1b56 asahi: scratch: Add feature to debug core IDs
Define SCRATCH_DEBUG_CORES to a raw core count to bypass the mapping
machinery and allocate memory for many cores, so we can work out how the
core IDs are mapped.

Signed-off-by: Asahi Lina <lina@asahilina.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:30 +00:00
Asahi Lina 494399c65c asahi: Implement scratch allocation
Implement our helper program to map scratch blocks into stack memory,
and the driver side that allocates these blocks as necessary.

Alloction is grow-only right now. Drivers are expected to instantiate
scratch memory managers for each shader type (VS, FS, CS) and the same
buffers are reused across commands for each one, growing as necessary.

Signed-off-by: Asahi Lina <lina@asahilina.net>
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00
Asahi Lina da9677f14b asahi: cmdbuf: Fix scratch bucket offset/size
This is 4 bits. It is an opaque value that gets passed to the
helper, except 0 means no scratch.

Signed-off-by: Asahi Lina <lina@asahilina.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00
Asahi Lina d6cc820d1a asahi: libagx: Move PACKED and GLOBAL macros to libagx.h
Signed-off-by: Asahi Lina <lina@asahilina.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00
Asahi Lina eec7c4beb8 agx: compiler: Enable stack_adjust
We'll need this once we start using scratch.

Signed-off-by: Asahi Lina <lina@asahilina.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00
Asahi Lina 1f1751d258 agx: compiler: Export scratch size to the driver
So it can configure the pipeline correctly, and allocate the required
scratch space.

Signed-off-by: Asahi Lina <lina@asahilina.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00
Asahi Lina b89da92a5e agx: compiler: Add fence_helper_exit_agx barrier
This is used by the helper program on exit.

Signed-off-by: Asahi Lina <lina@asahilina.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00
Asahi Lina c406ce793b agx: Add scaffolding to build the helper shader at device init
Add the scaffolding to compile our helper program, and load it at device
init time.

Signed-off-by: Asahi Lina <lina@asahilina.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00
Asahi Lina 05c8b59f5b agx: Hook up helper intrinsics into CL
So we can use these intrinsicts from libagx.

Signed-off-by: Asahi Lina <lina@asahilina.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00
Asahi Lina 30862d7189 agx: Hook up AGX helper NIR intrinsics
Signed-off-by: Asahi Lina <lina@asahilina.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00
Asahi Lina b07dbf7b0f nir: Add AGX-specific helper opcodes
These opcodes are used by the helper program to fetch the current
operation info and core ID.

Signed-off-by: Asahi Lina <lina@asahilina.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00
Asahi Lina 5d0f1ca77f agx: Rename some SRs
I think these better represent what these SRs actually do.

Signed-off-by: Asahi Lina <lina@asahilina.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00
Asahi Lina 28d34f6352 asahi: libagx: introduce AGX_STATIC_ASSERT
Using the array size trick, this works in both OpenCL and C.

Signed-off-by: Asahi Lina <lina@asahilina.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00
Alyssa Rosenzweig 8661006ef0 agx: add some more bitop tests
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00
Mary Guillemard 923767a968 agx: Add a bitop optimizer pass
Signed-off-by: Mary Guillemard <mary@mary.zone>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00
Mary Guillemard 2b89eb979e agx: Fuse not into and/or/xor
Signed-off-by: Mary Guillemard <mary@mary.zone>
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00
Mary Guillemard ba508fe854 agx: Remove and/or/xor pseudo ops
Switch back to bitop while keeping the aliases on agx_print_instr.

Also add all variants of 2 args of agx_bitop_table.

Signed-off-by: Mary Guillemard <mary@mary.zone>
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00
Mary Guillemard a5b6ff3ccc agx: Add more bitops in agx_bitop_table
Signed-off-by: Mary Guillemard <mary@mary.zone>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00
Alyssa Rosenzweig 62001e175f asahi: fix UB in qbo's
upper bits may be uninitialized. fixes
KHR-GL46.transform_feedback_overflow_query_ARB when built with gcc (passed by
chance with clang).

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00
Alyssa Rosenzweig a06b51559b asahi: plug geometry heap leak
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00
Alyssa Rosenzweig 877c180677 asahi: plug glsl type leak
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00
Alyssa Rosenzweig 06cf222830 asahi: plug so target leak
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00
Alyssa Rosenzweig e2c5a93358 asahi: plug early_serialized_nir leak
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00
Alyssa Rosenzweig ffac4e7179 asahi: plug pre-gs leak
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00
Alyssa Rosenzweig 7b44d15e5f asahi: plug geometry shader leaks
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00
Alyssa Rosenzweig 1e973dd5a9 asahi: rework meta shader infra
this fixes the leaks, and is also a lot more pleasant to work with.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00
Alyssa Rosenzweig 7eda3da983 asahi: plug passthrough tcs leak
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00
Alyssa Rosenzweig e91c0b6423 asahi: delete bogus assertion
dEQP-GLES3.functional.fragment_ops.scissor.outside_render_line at certain render
sizes.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00
Alyssa Rosenzweig 5261962186 asahi: fix prim restart unrolling with indirects
need to account for draw->start.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00
Alyssa Rosenzweig c0d3bf3608 agx: optimize vote_eq
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00
Alyssa Rosenzweig 0de7018879 agx: optimize first_invocation
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
2024-02-14 21:02:29 +00:00