We can provide better guidance on when to (un-)set this given that
everyone's on NIR now.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16539>
We want to use nir_opt_large_constants() instead (which is already
enabled), since that doesn't involve uploading the large immediate data
array again on each CB0 update. The downside is a bit of addressing math,
since constant_data is accessed using 64-bit global addresses.
The shader-db results are a bit all over:
All Iris driver platforms had similar results. (Ice Lake shown)
total instructions in shared programs: 19910185 -> 19913931 (0.02%)
instructions in affected programs: 225374 -> 229120 (1.66%)
helped: 3 / HURT: 348
total cycles in shared programs: 856004856 -> 855016808 (-0.12%)
cycles in affected programs: 22832422 -> 21844374 (-4.33%)
helped: 277 / HURT: 101
total spills in shared programs: 6580 -> 6609 (0.44%)
spills in affected programs: 516 -> 545 (5.62%)
helped: 1 / HURT: 4
total fills in shared programs: 8235 -> 8267 (0.39%)
fills in affected programs: 1022 -> 1054 (3.13%)
helped: 1 / HURT: 3
total sends in shared programs: 1039347 -> 1039095 (-0.02%)
sends in affected programs: 16367 -> 16115 (-1.54%)
helped: 251 / HURT: 0
LOST: 5
GAINED: 2
LOST:
- 3 SIMD16 fragment shaders (Superposition)
- 2 SIMD16 compute shaders (Aztec Ruins)
GAINED:
- fake news... 2 SIMD8 compute shaders that replace the lost SIMD16
compute shaders.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16539>
nir_opt_large_constants balks at seeing a store_deref of a variable
where the source is a vecN operation of multiple load_consts, and thinks
that isn't a constant, so it should not bother promoting it.
Unfortunately, we were running nir_lower_load_const_to_scalar before
nir_opt_large_constants, so this prevented a ton of constant promotion.
This commit /used to help/ some shaders in shader-db. Presumably since
!16770 landed, those shaders were already helped. Currently ther are
no shader-db changes on any Intel platform.
Fossil-db results:
All Intel platforms had similar results. (Ice Lake shown)
Instructions in all programs: 141998227 -> 141421756 (-0.4%)
Instructions helped: 12515
Instructions hurt: 237
SENDs in all programs: 7437925 -> 7468033 (+0.4%)
SENDs hurt: 12806
Cycles in all programs: 9161655753 -> 9132869800 (-0.3%)
Cycles helped: 10163
Cycles hurt: 2637
Spills in all programs: 19977 -> 18678 (-6.5%)
Spills helped: 384
Spills hurt: 40
Fills in all programs: 32863 -> 31396 (-4.5%)
Fills helped: 385
Fills hurt: 42
Lost: 1
Lots of Shadow of the Tomb Raider fragment shaders and Batman Arkham
Origins vertex shaders were hurt for SENDs in this commit. A couple
Aztec Ruins compute shaders and Spaceship shaders (multiple stages)
were also hurt.
All of the shaders hurt for spills or fills were Spaceship compute
shaders. Nearly all of the shaders helped were Shadow of the Tomb
Raider fragmenet shaders. One Spaceship shader was reall, REALLY helped:
Spills helped fossils/fossil-db/Spaceship.run.9f90a2a226fcc57f.1.foz/0b507d3abe2e3c28/compute: 321 -> 13 (-96.0%)
Fills helped fossils/fossil-db/Spaceship.run.9f90a2a226fcc57f.1.foz/0b507d3abe2e3c28/compute: 279 -> 21 (-92.5%)
Overall this seems like an improvement, but we may want to actually
run these few benchmarks before landing.
Reviewed-by: Emma Anholt <emma@anholt.net>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16539>
We were using the old load_global intrinsic still, which can't be
reordered, limiting optimization opportunities. We know the data here
is constant, so we can use the newer load_global_constant intrinsic.
This doesn't seem to have any impact on shader-db or fossil-db on any
Intel platform.
Reviewed-by: Emma Anholt <emma@anholt.net>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16539>
Mainly to get rid of alloc/free in collect_countables. Sampling at 1ms,
perf becomes
- 22.75% pps::FreedrenoDriver::collect_countables
22.59% pps::FreedrenoDriver::Countable::collect
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19427>
It is always possible on a5xx+ and allows Countable::collect to do 1 ldr rather
than 2.
Sampling at 1ms, perf goes from
- 34.44% pps::FreedrenoDriver::collect_countables
25.36% pps::FreedrenoDriver::Countable::collect
3.92% cfree
+ 2.28% operator new
to
- 29.60% pps::FreedrenoDriver::collect_countables
20.70% pps::FreedrenoDriver::Countable::collect
4.01% cfree
+ 2.35% operator new
1.09% memcpy
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19427>
Otherwise, queries might return invalid data because they used
the same offsets as NGG streamout.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19442>
Foz-DB Navi21:
Totals from 2403 (1.78% of 134913) affected shaders:
CodeSize: 25329156 -> 25311244 (-0.07%)
Signed-off-by: Georg Lehmann <dadschoorse@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19413>
The kernel keeps returning -ENOMEM if multiple processes allocate GDS,
this always happen while running VKCTS. This solution is loosely based
on RadeonSI, except that it includes a timeout of 1s to exit the loop.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19440>
NGG streamout lowering requires nir_shader::num_outputs to be set to
the total number of outputs in order to compute the pervertex LDS size
correctly. This is wasting LDS memory but it's currently the only viable
solution.
This fixes a bunch of dEQP-VK.transform_feedback.* failures.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19436>
Given that we no longer need the slot_to_register mapping, it's
useless to use this function.
This also fixes a bunch of failures with
dEQP-VK.transform_feedback.*omit_write* on RADV because in Vulkan
the spec requires XFB query counters to be incremented even if XFB
outputs aren't written to.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Qiang Yu <yuq825@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19437>
[problem]
When decoding an av1 bitstream, it shows image corruption
in the middle of the bitstream around key frames.
[analysis]
in av1_spec.pdf page 38/669, there is a sentence below:
if ( frame_type == KEY_FRAME && show_frame ) {
for ( i = 0; i < NUM_REF_FRAMES; i++) {
RefValid[ i ] = 0
......
}
......
}
This shows that the condition of invalidating current
DPB frames should be the coming frame_type is KEY_FRAME plus
show_frame is equal to 1. Otherwise, some of the frames
in sequence after KEY_FRAME still refer to the reference frames
before KEY_FRAME, and if these before KEY_FRAME reference
frames were invalidated, these frames could not find their
reference frames, and it could cause image corruption.
[solution]
Add condition of show_frame, with the corresponding fix
in ffmpeg, we cannot see this issue any longer.
Reviewed-by: Leo Liu <leo.liu@amd.com>
Signed-off-by: Ruijing Dong <ruijing.dong@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19386>
Allows to make Clover devices appearing as cpu, gpu or accelerator
by setting the CLOVER_DEVICE_TYPE environment variable like
the RUSTICL_DEVICE_TYPE environment variable does.
For example it can make the CPU llvmpipe device appear as GPU or GPU devices
appear as CPU. This is useful for testing OpenCL with applications that may
use different code path given the OpenCL device is a CPU or a GPU.
The initial motivation for RUSTICL_DEVICE_TYPE implementation was to test
rusticl with llvmipe on applications ignoring CPU devices.
This brings Clover on par with rusticl on that topic.
CL_DEVICE_TYPE_CUSTOM isn't implemented or applications may crash when
iterating devices because CL_DEVICE_TYPE_CUSTOM is OpenCL 1.2 and Clover
is OpenCL 1.1.
Signed-off-by: Thomas Debesse <dev@illwieckz.net>
Reviewed-by: Mihai Preda <mhpreda@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18931>
If the size is passed through CmdBindTransformFeedback() uses that.
This partially fixes dEQP-VK.transform_feedback.simple.multiquery_1
by reporting the correct number of primitives written (the computation
is based on the buffer size). There is still a bug around GDS offsets
that will be fixed later.
Tested on GFX10.3 by forcing NGG streamout.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19433>
According to the AMD registers database, SAMPLE_STREAMOUTSTATS no
longer exists on GFX11. This fixes primitives generated query if only
the NGG path is used. Tested on GFX10.3 by forcing NGG everywhere.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19410>
With NGG only, only the GDS query counter will be incremented.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19410>
Removed warning about deprecated usage that was not necessary.
Also added information on whether support is enabled or not to the
summary so it is easier to check when building.
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19332>
Supported in all hardware and software drivers. Only that don't support
are virgl and svga, depending on host capabilities. I don't think
there's anything to be done there. This does give fewer places to screw
up the CAPs, though, because everyone wants ARB_buffer_storage.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Emma Anholt <emma@anholt.net>
Reviewed-by: Marek Ol<C5><A1><C3><A1>k <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19392>
This CAP only concerns mapping of buffers, not textures. For buffers,
etnaviv does not seem to use any staging resources. In neither case does
etnaviv use explicit syncs or flushes, so persistent/coherent mapping
should come as the implicit default and the cap should be almost free.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Emma Anholt <emma@anholt.net>
Reviewed-by: Christian Gmeiner <christian.gmeiner@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19392>
Currently buffer_map returns the raw pointer, there is no syncing or
flushing or staging buffers or resource shadowing anywhere. That means
if the driver is currently correct, it is implementing persistent and
coherent behaviour already, so the CAP is trivial.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19392>
It's a software rasterizer, all buffers are necessarily persistent and
coherent. There are no staging buffers in transfer_map and no sync/flush
operations, so this is trivial.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19392>
Implement Wa_1508744258:
Disable RHWO by setting 0x7010[14] by default except during resolve
pass.
Disable the RCC RHWO optimization at all times except when resolving
single sampled color surfaces. MCS partial resolves are done via
software (i.e., not via a HW bit) and so are not expected to need this
workaround.
Reviewed-by: Mark Janes <markjanes@swizzler.org>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19360>
This required fixing up the disk cache to hash in stuff that was missing
(like the proper pipeline cache uuid rather than just PCI IDs, and also
zink's build ID). Once we do that, we can just hand the frontend our disk
cache and avoid GLSL->NIR recompiles.
tu asphalt_9 trace perf +3.14859% +/- 1.36891% (n=4)
Fixes: #7513
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19124>
It's useful for catching regressions, e.g. it would have caught
https://gitlab.freedesktop.org/mesa/mesa/-/issues/7567 .
For the radeonsi-stoney-gl:amd64 job, we can use the "gpu" profile,
which contains quick_gl and quick_shader.
For the radeonsi-raven-piglit-gl:amd64 job, that would almost double
the runtime, and bring it dangerously close to the 30 minutes timeout.
Instead, split it up into radeonsi-raven-piglit-quick_gl:amd64 and
radeonsi-raven-piglit-quick_shader:amd64 jobs.
v2:
* Rebase, drop spec@arb_tessellation_shader@execution flakes and add
one crash instead.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19378>
FMask doesn't exist on GFX11. Have txf_ms take the fragment_fetch_amd
path.
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19375>
