On SM20 this gives:
total instructions in shared programs : 6299222 -> 6294240 (-0.08%)
total gprs used in shared programs : 944139 -> 944068 (-0.01%)
total local used in shared programs : 54116 -> 54116 (0.00%)
local gpr inst bytes
helped 0 126 2781 2781
hurt 0 55 11 11
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
This way $r1 = $r0 + 4; c1[$r1] becomes c1[$r0+4].
On SM35:
total instructions in shared programs : 6206257 -> 6185058 (-0.34%)
total gprs used in shared programs : 911045 -> 910722 (-0.04%)
total local used in shared programs : 39072 -> 39072 (0.00%)
local gpr inst bytes
helped 0 417 4195 4195
hurt 0 280 0 0
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
This works when the add also has an immediate. This often happens in
address calculations. These addresses can then be inlined as well.
On code targeted to SM35:
total instructions in shared programs : 6223346 -> 6206257 (-0.27%)
total gprs used in shared programs : 911075 -> 911045 (-0.00%)
total local used in shared programs : 39072 -> 39072 (0.00%)
local gpr inst bytes
helped 0 119 3664 3664
hurt 0 74 15 15
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Even if the rasterizer has scissor disabled, we'll have whatever
vc4->scissor bounds were last set when someone set up a scissor, so we
shouldn't clip to them in that case.
Fixes piglit fbo-blit-rect, and a lot of MSAA tests once they're enabled.
We could potentially handle scissored blits when they're tile aligned, but
it doesn't seem worth it. If you're doing a scissored blit, you're
probably a testcase.
Fixes piglit's fbo-scissor-blit fbo
This implements more performance metrics than the previous support,
but some other metrics still need to be figured out.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
These performance metrics will be re-introduced in an upcoming
patch that will follow the same design as Fermi.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
inst_issued is performance metric not a hardware event on Kepler (SM30).
It will be re-introduced in an upcoming patch.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
SM30 is the compute capability version for GK104/GK106/GK107.
This also introduces a new signal group selection called UNK0F.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Move these to 'disasm' instead of the more verbose 'optmsgs' since, like
the tgsi dumps, it is useful without the more verbose compiler logging
enabled.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Fixes assertion in vkCreateImage when VkFormat is combined depthstencil.
Fixed many vulkancts tests that use combined depthstencil. For example,
fixes dEQP-VK.pipeline.depth.format.d16_unorm_s8_uint.compare_ops.\
not_equal_less_or_equal_not_equal_greater.
We're required to expose a host-visible, coherent memory type. On big
core GPUs that share, LLC, we can expose one such memory type that's
also cached. However, on non-LLC GPUs we can't both be cached and
coherent. Thus, we expose both the required coherent type and the cached
but non-coherent combination.
Regular objects are created I915_CACHING_CACHED on LLC platforms and
I915_CACHING_NONE on non-LLC platforms. However, userptr objects are
always created as I915_CACHING_CACHED, which on non-LLC means
snooped. That can be useful but comes with a bit of overheard. Since
we're eplicitly clflushing and don't want the overhead we need to turn
it off.
For MSAA, we store full resolution tile buffer contents, which have their
own tiling format. Since they're full resolution buffers, we have to
align their size to full tiles.
We were checking that the blit started at 0 and was 1:1, but not that it
went to the full width of the surface, or that the width was aligned to a
tile. We then told it to blit to the full width/height of the surface,
causing contents to be stomped in a bunch of MSAA tests that happen to
include half-screen-width blits to 0,0.
I've played with a few different approaches to tweak instruction
priority according to how much they increase/decrease register pressure,
etc. But nothing seems to change the fact that compared to original
(pre-multiple-block-support) scheduler, in some edge cases we are
generating shaders w/ 5-6x higher register usage.
The problem is that the priority queue approach completely looses the
dependency between instructions, and ends up scheduling all paths at the
same time.
Original reason for switching was that recursive approach relied on
starting from the shader outputs array. But we can achieve more or less
the same thing by starting from the depth-sorted list.
shader-db results:
total instructions in shared programs: 113350 -> 105183 (-7.21%)
total dwords in shared programs: 219328 -> 211168 (-3.72%)
total full registers used in shared programs: 7911 -> 7383 (-6.67%)
total half registers used in shader programs: 109 -> 109 (0.00%)
total const registers used in shared programs: 21294 -> 21294 (0.00%)
half full const instr dwords
helped 0 322 0 711 215
hurt 0 163 0 38 4
The shaders hurt tend to gain a register or two. While there are also a
lot of helped shaders that only loose a register or two, the more
complex ones tend to loose significanly more registers used. In some
more extreme cases, like glsl-fs-convolution-1.shader_test it is more
like 7 vs 34 registers!
Signed-off-by: Rob Clark <robclark@freedesktop.org>
It causes confusion in sched if we need to split_addr() since otherwise
we wouldn't easily know which block the new addr instr will be scheduled
in. So just side-step the whole situation.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
We'll need to add similar for ir3_instruction, but following the pattern
to use 'id' seems confusing. Let's just go w/ generic 'data' as the
name.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Undefining the NDEBUG is relevant for release build, as they are the
ones that set it.
[Emil Velikov: split from previous patch]
Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
This follows the src/util/u_atomic_test.c model of undefining NDEBUG
unconditionally throughouth the XvMC tests, to force asserts regardless
of debug mode.
The comment on u_atomic_test.c is also fixed (read 'debug' where it
should have been 'release').
v2: s/debug/release/ in relevant comments
Signed-off-by: Giuseppe Bilotta <giuseppe.bilotta@gmail.com>
[Emil Velikov: keep the src/util/ hunk as separate patch]
Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
Since we're using nir_lower_outputs_to_temporaries to shadow all our
outputs, it's impossible to actually get an indirect store. The code we
had to "handle" this was pretty bogus as it created a register with a
reladdr and then stuffed it in a fixed varying slot without so much as a
MOV. Not only does this not do the MOV, it also puts the indirect on the
wrong side of the transaction. Let's just delete the broken dead code.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
It's not really buying us anything at this point. It's just a way of
remapping one offset namespace onto another. We can just use the location
namespace the whole way through.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The original change to put zeroes directly into instructions created
conditional mov's with the zero immediate. However that can't be
emitted, so make sure to replace the zero with r63.
Fixes: 52a800a68 (nv50/ir: allow immediate 0 to be loaded anywhere)
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
A situation where there's a 128-bit load where the last component gets
DCE'd causes a 96-bit load to be generated, which no GPU can actually
emit. Avoid generating such instructions by scaling back to 64-bit on
the first load when splitting.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: "11.0 11.1" <mesa-stable@lists.freedesktop.org>
Edward O'Callaghan