platform/upstream/mesa.git
3 years agopan/bi: Allow negating constants
Alyssa Rosenzweig [Fri, 5 Mar 2021 02:18:48 +0000 (02:18 +0000)]
pan/bi: Allow negating constants

Useful for representing -0 in transcendental sequences matching the
blob.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9420>

3 years agopan/bi: Use replace_index in more places
Alyssa Rosenzweig [Fri, 5 Mar 2021 02:18:25 +0000 (02:18 +0000)]
pan/bi: Use replace_index in more places

Needed to respect abs/neg.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9420>

3 years agoradeonsi/sqtt: export shader code to RGP
Pierre-Eric Pelloux-Prayer [Tue, 23 Feb 2021 14:22:40 +0000 (15:22 +0100)]
radeonsi/sqtt: export shader code to RGP

With these changes the shader code is visible in RGP.

Vk pipeline feature is emulated using si_update_shaders: when shaders are
updated we compute a sha1 of their code and use it as a pipeline hash.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9277>

3 years agoradeonsi/sqtt: don't always use WGP 0
Pierre-Eric Pelloux-Prayer [Thu, 25 Feb 2021 09:15:17 +0000 (10:15 +0100)]
radeonsi/sqtt: don't always use WGP 0

Because it may be disabled. Instead use the cu mask to
pick the first active WGP.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9277>

3 years agoradeonsi/sqtt: remove duplicate token
Pierre-Eric Pelloux-Prayer [Tue, 23 Feb 2021 14:12:24 +0000 (15:12 +0100)]
radeonsi/sqtt: remove duplicate token

V_008D18_REG_INCLUDE_CONTEXT was set twice.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9277>

3 years agoradeonsi/sqtt: keep a copy of the uploaded shader code
Pierre-Eric Pelloux-Prayer [Tue, 23 Feb 2021 14:05:19 +0000 (15:05 +0100)]
radeonsi/sqtt: keep a copy of the uploaded shader code

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9277>

3 years agoac/rgp: move radv/sqtt functions to ac
Pierre-Eric Pelloux-Prayer [Tue, 23 Feb 2021 15:00:37 +0000 (16:00 +0100)]
ac/rgp: move radv/sqtt functions to ac

pso_correlation and code_object_loader don't depend on drivers
specific logic so move them to the shared code.

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9277>

3 years agoac/rtld: make ac_rtld_upload returns the code size
Pierre-Eric Pelloux-Prayer [Tue, 23 Feb 2021 14:03:59 +0000 (15:03 +0100)]
ac/rtld: make ac_rtld_upload returns the code size

This will be useful to keep a copy of the uploaded code.

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9277>

3 years agoac/rgp: make the max gap between shader code a warning
Pierre-Eric Pelloux-Prayer [Tue, 23 Feb 2021 14:02:05 +0000 (15:02 +0100)]
ac/rgp: make the max gap between shader code a warning

For radeonsi the shaders don't live in the same BOs, so they're
unlikely to be less that 0x1000 bytes apart.

So this commit bumps the threshold to 0x10000 and warns once
when hitting it.

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9277>

3 years agoradeonsi: properly set SPI_SHADER_PGM_HI_ES
Pierre-Eric Pelloux-Prayer [Tue, 23 Feb 2021 10:08:20 +0000 (11:08 +0100)]
radeonsi: properly set SPI_SHADER_PGM_HI_ES

When not using S_00B324_MEM_BASE the value isn't properly truncated.

Cc: mesa-stable
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9277>

3 years agobroadcom/compiler: fix flags check for ldvary merge
Iago Toral Quiroga [Fri, 5 Mar 2021 12:18:02 +0000 (13:18 +0100)]
broadcom/compiler: fix flags check for ldvary merge

We were checking that the previous instruction doesn't write flags,
but we also need to check it doesn't read them.

Fixes: 1784dd22a32 ('broadcom/compiler: pipeline smooth ldvary sequences')
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9431>

3 years agobroadcom/compiler: ldvary doesn't implicitly write to r3 since V3D 4.1
Iago Toral Quiroga [Fri, 5 Mar 2021 10:26:21 +0000 (11:26 +0100)]
broadcom/compiler: ldvary doesn't implicitly write to r3 since V3D 4.1

total instructions in shared programs: 13805979 -> 13786037 (-0.14%)
instructions in affected programs: 2263244 -> 2243302 (-0.88%)
helped: 10646
HURT: 1508
Instructions are helped.

total threads in shared programs: 412220 -> 412242 (<.01%)
threads in affected programs: 58 -> 80 (37.93%)
helped: 17
HURT: 6
Threads are helped.

total uniforms in shared programs: 3793200 -> 3790401 (-0.07%)
uniforms in affected programs: 131281 -> 128482 (-2.13%)
helped: 1547
HURT: 281
Uniforms are helped.

total max-temps in shared programs: 2326309 -> 2324834 (-0.06%)
max-temps in affected programs: 31836 -> 30361 (-4.63%)
helped: 1139
HURT: 153
Max-temps are helped.

total spills in shared programs: 5932 -> 5940 (0.13%)
spills in affected programs: 80 -> 88 (10.00%)
helped: 2
HURT: 3

total fills in shared programs: 13370 -> 13372 (0.01%)
fills in affected programs: 480 -> 482 (0.42%)
helped: 2
HURT: 3

total sfu-stalls in shared programs: 30829 -> 30685 (-0.47%)
sfu-stalls in affected programs: 2190 -> 2046 (-6.58%)
helped: 570
HURT: 533
Sfu-stalls are helped.

total inst-and-stalls in shared programs: 13836808 -> 13816722 (-0.15%)
inst-and-stalls in affected programs: 2276152 -> 2256066 (-0.88%)
helped: 10643
HURT: 1525
Inst-and-stalls are helped.

Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9430>

3 years agoradv: don't set sx_blend_opt_epsilon for V_028C70_COLOR_10_11_11
Rhys Perry [Fri, 5 Mar 2021 10:58:03 +0000 (10:58 +0000)]
radv: don't set sx_blend_opt_epsilon for V_028C70_COLOR_10_11_11

Matches radeonsi and PAL. From PAL:
// 1 is recommended, but doesn't provide sufficient precision

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/4394
Fixes: ed946381564 ("radv: Enable RB+ where possible.")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9427>

3 years agobroadcom/compiler: always restart ldvary pipelining when scheduling ldvary
Iago Toral Quiroga [Thu, 4 Mar 2021 08:21:53 +0000 (09:21 +0100)]
broadcom/compiler: always restart ldvary pipelining when scheduling ldvary

When we were only able to pipeline smooth varyings, if we had to disable
ldvary pipelining in the middle of a sequence it would stay disabled for
the rest of the program, to prevent us from prioritizing scheduling of
ldvary instructions that we would not be able to pipeline effectively.
Now that we can pipeline all ldvary sequences we can change this.

This change re-enables ldvary pipelining upon finding the next
ldvary in the program in the hopes that we can continue pipelining
succesfully. To do this, we track the number of ldvary instructions we
emitted so far and compare that to the number of inputs in the fragment
shader we are scheduling. This also allows us to simplify our ldvary
tracking at nir to vir time, since that is all now handled in the QPU
scheduler.

total instructions in shared programs: 13817048 -> 13810783 (-0.05%)
instructions in affected programs: 810114 -> 803849 (-0.77%)
helped: 4843
HURT: 591
Instructions are helped.

total max-temps in shared programs: 2326612 -> 2326300 (-0.01%)
max-temps in affected programs: 4689 -> 4377 (-6.65%)
helped: 285
HURT: 7
Max-temps are helped.

total sfu-stalls in shared programs: 30942 -> 30865 (-0.25%)
sfu-stalls in affected programs: 207 -> 130 (-37.20%)
helped: 120
HURT: 42
Sfu-stalls are helped.

total inst-and-stalls in shared programs: 13847990 -> 13841648 (-0.05%)
inst-and-stalls in affected programs: 825378 -> 819036 (-0.77%)
helped: 4899
HURT: 590
Inst-and-stalls are helped.

Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9404>

3 years agoradv: re-enable TC-compat HTILE for MSAA D32S8 images on GFX9+
Samuel Pitoiset [Wed, 24 Feb 2021 10:22:10 +0000 (11:22 +0100)]
radv: re-enable TC-compat HTILE for MSAA D32S8 images on GFX9+

Should help MSAA games. Note that it's broken on GFX8 because
the tiling doesn't match.

Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/3868
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/9284>

3 years agovirgl: use atomic operations when increase sub_ctx_id
Xin He [Thu, 4 Mar 2021 10:46:31 +0000 (18:46 +0800)]
virgl: use atomic operations when increase sub_ctx_id

Use atomic operations to avoid competition. In addition,
since sub_ctx_id 0 has been used by default, sub_ctx_id
should start from 1.

Signed-off-by: Xin He <hexin.op@bytedance.com>
Reviewed-by: Gert Wollny <gert.wollny@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9406>

3 years agoradv: skip useless FCE when fast-clearing MSAA images with DCC enabled
Samuel Pitoiset [Wed, 3 Mar 2021 16:16:41 +0000 (17:16 +0100)]
radv: skip useless FCE when fast-clearing MSAA images with DCC enabled

The clear code is 0xCC which means CMASK isn't fast-cleared.

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/9392>

3 years agoradv: remove useless check about mips+layers for TC-compat HTILE images
Samuel Pitoiset [Thu, 4 Mar 2021 07:56:39 +0000 (08:56 +0100)]
radv: remove useless check about mips+layers for TC-compat HTILE images

radv_use_htile_for_image() prevents it.

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/9405>

3 years agoradv: cleanup enabling TC-compat HTILE for depth surfaces
Samuel Pitoiset [Thu, 4 Mar 2021 07:51:15 +0000 (08:51 +0100)]
radv: cleanup enabling TC-compat HTILE for depth surfaces

It makes more sense to try to enable TC-compat if the image has HTILE.

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/9405>

3 years agozink: add vk/spirv caps/extension for shader LAYER variable
Mike Blumenkrantz [Fri, 11 Dec 2020 23:56:46 +0000 (18:56 -0500)]
zink: add vk/spirv caps/extension for shader LAYER variable

this is required if gl_Layer is used outside of GEOMETRY stage

Fixes: c77df59c9e6 ("zink: export PIPE_CAP_TGSI_VS_LAYER_VIEWPORT")

Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9410>

3 years agolavapipe: fix dynamic viewport/scissor pipeline emission
Dave Airlie [Fri, 5 Mar 2021 03:17:28 +0000 (13:17 +1000)]
lavapipe: fix dynamic viewport/scissor pipeline emission

Just fixup the tests for when the pipeline vp/scissors
are emitted.

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9422>

3 years agolavapipe: fix pipeline vp/scissor mixup.
Dave Airlie [Fri, 5 Mar 2021 03:16:31 +0000 (13:16 +1000)]
lavapipe: fix pipeline vp/scissor mixup.

Not copying all the scissors caused
dEQP-VK.pipeline.extended_dynamic_state.two_draws_dynamic.2_viewports
to fail but thah test pointlessly relies on KHR_multiview (cts issue
filed).

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Fixes: b38879f8c5f57 ("vallium: initial import of the vulkan frontend")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9422>

3 years agoanv: don't advertise mipmaps for linear 3D surfaces on BDW
Iván Briano [Thu, 4 Mar 2021 22:55:35 +0000 (14:55 -0800)]
anv: don't advertise mipmaps for linear 3D surfaces on BDW

Prior to SKL, the mipmaps for 3D surfaces are laid out in a way
that make it impossible to represent in the way that
VkSubresourceLayout expects. Since we can't tell users how to make
sense of them, don't report them as available.

"Fixes" dEQP-VK.image.subresource_layout.3d.*

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Nanley Chery <nanley.g.chery@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9419>

3 years agonir/algebraic: Apply addition property of equality to the other ordering too
Ian Romanick [Tue, 2 Feb 2021 18:18:42 +0000 (10:18 -0800)]
nir/algebraic: Apply addition property of equality to the other ordering too

Inequality comparison operations are not commutative, so `foo < bar` and
`bar < foo` both have to be explicitly listed.

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
All Intel GPUs had similar results. (Ice Lake shown)
total instructions in shared programs: 20027051 -> 20026899 (<.01%)
instructions in affected programs: 37181 -> 37029 (-0.41%)
helped: 85
HURT: 0
helped stats (abs) min: 1 max: 20 x̄: 1.79 x̃: 1
helped stats (rel) min: 0.05% max: 6.78% x̄: 0.92% x̃: 0.68%
95% mean confidence interval for instructions value: -2.42 -1.15
95% mean confidence interval for instructions %-change: -1.23% -0.61%
Instructions are helped.

total cycles in shared programs: 979762793 -> 979753527 (<.01%)
cycles in affected programs: 2653905 -> 2644639 (-0.35%)
helped: 104
HURT: 50
helped stats (abs) min: 1 max: 1048 x̄: 119.99 x̃: 11
helped stats (rel) min: <.01% max: 9.88% x̄: 0.77% x̃: 0.20%
HURT stats (abs)   min: 1 max: 734 x̄: 64.26 x̃: 8
HURT stats (rel)   min: <.01% max: 3.06% x̄: 0.36% x̃: 0.10%
95% mean confidence interval for cycles value: -98.65 -21.68
95% mean confidence interval for cycles %-change: -0.66% -0.15%
Cycles are helped.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9374>

3 years agonir/algebraic: Apply addition property of equality more conservatively
Ian Romanick [Thu, 11 Jan 2018 02:50:58 +0000 (18:50 -0800)]
nir/algebraic: Apply addition property of equality more conservatively

This allows a lot more CSE.  Depending on where the addition and the
comparison are scheduled, it may also reduce register pressure by
reducing the live range of the addends.

Across all the platforms, the shaders affected for spills or fills were
all fragment shaders from Dirt Rally.

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Tiger Lake and Ice Lake had similar results. (Tiger Lake shown)
total instructions in shared programs: 21043103 -> 21038804 (-0.02%)
instructions in affected programs: 892878 -> 888579 (-0.48%)
helped: 1549
HURT: 724
helped stats (abs) min: 1 max: 225 x̄: 4.14 x̃: 2
helped stats (rel) min: 0.05% max: 11.18% x̄: 1.04% x̃: 0.78%
HURT stats (abs)   min: 1 max: 71 x̄: 2.93 x̃: 1
HURT stats (rel)   min: 0.07% max: 6.90% x̄: 0.80% x̃: 0.56%
95% mean confidence interval for instructions value: -2.33 -1.45
95% mean confidence interval for instructions %-change: -0.50% -0.40%
Instructions are helped.

total cycles in shared programs: 855054155 -> 855757566 (0.08%)
cycles in affected programs: 58275918 -> 58979329 (1.21%)
helped: 1213
HURT: 1680
helped stats (abs) min: 1 max: 107405 x̄: 1684.00 x̃: 10
helped stats (rel) min: <.01% max: 38.09% x̄: 1.51% x̃: 0.25%
HURT stats (abs)   min: 1 max: 126632 x̄: 1634.59 x̃: 12
HURT stats (rel)   min: <.01% max: 85.91% x̄: 2.75% x̃: 0.49%
95% mean confidence interval for cycles value: -98.06 584.35
95% mean confidence interval for cycles %-change: 0.71% 1.22%
Inconclusive result (value mean confidence interval includes 0).

total spills in shared programs: 9843 -> 9771 (-0.73%)
spills in affected programs: 72 -> 0
helped: 5
HURT: 0

total fills in shared programs: 9600 -> 9451 (-1.55%)
fills in affected programs: 149 -> 0
helped: 5
HURT: 0

LOST:   14
GAINED: 9

Skylake
total instructions in shared programs: 18185074 -> 18183866 (<.01%)
instructions in affected programs: 575180 -> 573972 (-0.21%)
helped: 1286
HURT: 468
helped stats (abs) min: 1 max: 15 x̄: 1.55 x̃: 1
helped stats (rel) min: 0.03% max: 4.08% x̄: 0.67% x̃: 0.65%
HURT stats (abs)   min: 1 max: 8 x̄: 1.69 x̃: 1
HURT stats (rel)   min: 0.13% max: 7.69% x̄: 0.87% x̃: 0.45%
95% mean confidence interval for instructions value: -0.77 -0.60
95% mean confidence interval for instructions %-change: -0.30% -0.22%
Instructions are helped.

total cycles in shared programs: 960518105 -> 960608234 (<.01%)
cycles in affected programs: 42536073 -> 42626202 (0.21%)
helped: 1210
HURT: 1714
helped stats (abs) min: 1 max: 7015 x̄: 123.41 x̃: 10
helped stats (rel) min: <.01% max: 33.76% x̄: 1.32% x̃: 0.26%
HURT stats (abs)   min: 1 max: 14474 x̄: 139.71 x̃: 14
HURT stats (rel)   min: <.01% max: 58.94% x̄: 2.00% x̃: 0.44%
95% mean confidence interval for cycles value: 4.02 57.63
95% mean confidence interval for cycles %-change: 0.43% 0.82%
Cycles are HURT.

LOST:   16
GAINED: 42

Broadwell
total instructions in shared programs: 17856880 -> 17852158 (-0.03%)
instructions in affected programs: 564836 -> 560114 (-0.84%)
helped: 1243
HURT: 418
helped stats (abs) min: 1 max: 115 x̄: 4.36 x̃: 1
helped stats (rel) min: 0.03% max: 9.67% x̄: 0.90% x̃: 0.67%
HURT stats (abs)   min: 1 max: 8 x̄: 1.67 x̃: 1
HURT stats (rel)   min: 0.14% max: 7.69% x̄: 0.89% x̃: 0.46%
95% mean confidence interval for instructions value: -3.45 -2.23
95% mean confidence interval for instructions %-change: -0.51% -0.38%
Instructions are helped.

total cycles in shared programs: 1031140321 -> 1029856892 (-0.12%)
cycles in affected programs: 66986946 -> 65703517 (-1.92%)
helped: 1084
HURT: 1653
helped stats (abs) min: 1 max: 415168 x̄: 1835.32 x̃: 10
helped stats (rel) min: <.01% max: 57.16% x̄: 1.19% x̃: 0.28%
HURT stats (abs)   min: 1 max: 43930 x̄: 427.14 x̃: 12
HURT stats (rel)   min: <.01% max: 57.53% x̄: 1.32% x̃: 0.39%
95% mean confidence interval for cycles value: -915.76 -22.07
95% mean confidence interval for cycles %-change: 0.17% 0.47%
Inconclusive result (value mean confidence interval and %-change mean confidence interval disagree).

total spills in shared programs: 20891 -> 20335 (-2.66%)
spills in affected programs: 1567 -> 1011 (-35.48%)
helped: 70
HURT: 0

total fills in shared programs: 27307 -> 25905 (-5.13%)
fills in affected programs: 5381 -> 3979 (-26.05%)
helped: 71
HURT: 0

LOST:   17
GAINED: 20

Haswell
total instructions in shared programs: 16411850 -> 16409414 (-0.01%)
instructions in affected programs: 602666 -> 600230 (-0.40%)
helped: 1152
HURT: 781
helped stats (abs) min: 1 max: 103 x̄: 3.59 x̃: 1
helped stats (rel) min: 0.03% max: 8.61% x̄: 0.85% x̃: 0.65%
HURT stats (abs)   min: 1 max: 41 x̄: 2.18 x̃: 1
HURT stats (rel)   min: 0.12% max: 7.69% x̄: 0.88% x̃: 0.69%
95% mean confidence interval for instructions value: -1.74 -0.78
95% mean confidence interval for instructions %-change: -0.21% -0.10%
Instructions are helped.

total cycles in shared programs: 1035338781 -> 1036977801 (0.16%)
cycles in affected programs: 68961096 -> 70600116 (2.38%)
helped: 1246
HURT: 2206
helped stats (abs) min: 1 max: 392022 x̄: 1040.28 x̃: 14
helped stats (rel) min: <.01% max: 56.44% x̄: 2.32% x̃: 0.38%
HURT stats (abs)   min: 1 max: 68630 x̄: 1330.56 x̃: 18
HURT stats (rel)   min: <.01% max: 69.97% x̄: 3.31% x̃: 0.61%
95% mean confidence interval for cycles value: 90.43 859.17
95% mean confidence interval for cycles %-change: 1.02% 1.54%
Cycles are HURT.

total spills in shared programs: 17805 -> 17457 (-1.95%)
spills in affected programs: 1202 -> 854 (-28.95%)
helped: 34
HURT: 31

total fills in shared programs: 20939 -> 20387 (-2.64%)
fills in affected programs: 2702 -> 2150 (-20.43%)
helped: 34
HURT: 31

LOST:   24
GAINED: 45

Ivy Bridge and earlier Intel GPUs had similar results. (Ivy Bridge shown)
total instructions in shared programs: 15515912 -> 15516757 (<.01%)
instructions in affected programs: 396569 -> 397414 (0.21%)
helped: 578
HURT: 858
helped stats (abs) min: 1 max: 9 x̄: 1.32 x̃: 1
helped stats (rel) min: 0.04% max: 3.70% x̄: 0.65% x̃: 0.65%
HURT stats (abs)   min: 1 max: 11 x̄: 1.87 x̃: 1
HURT stats (rel)   min: 0.08% max: 12.90% x̄: 0.95% x̃: 0.53%
95% mean confidence interval for instructions value: 0.47 0.70
95% mean confidence interval for instructions %-change: 0.24% 0.37%
Instructions are HURT.

total cycles in shared programs: 584395455 -> 584466352 (0.01%)
cycles in affected programs: 20346570 -> 20417467 (0.35%)
helped: 1192
HURT: 1896
helped stats (abs) min: 1 max: 4108 x̄: 123.27 x̃: 14
helped stats (rel) min: <.01% max: 37.20% x̄: 2.27% x̃: 0.46%
HURT stats (abs)   min: 1 max: 3698 x̄: 114.89 x̃: 19
HURT stats (rel)   min: <.01% max: 70.28% x̄: 3.02% x̃: 0.71%
95% mean confidence interval for cycles value: 10.75 35.16
95% mean confidence interval for cycles %-change: 0.73% 1.23%
Cycles are HURT.

LOST:   20
GAINED: 12
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9374>

3 years agoiris: Enable u_threaded_context
Kenneth Graunke [Fri, 15 May 2020 18:23:03 +0000 (11:23 -0700)]
iris: Enable u_threaded_context

This implements most of the remaining u_threaded_context support.  Most
of the heavy lifting was done in the previous patches which fixed things
up for the new thread safety requirements.  Only a few things remain.

u_threaded_context support can be disabled via an environment variable:

   GALLIUM_THREAD=0

On Felix's Tigerlake with the GPU at fixed frequency, enabling
u_threaded_context improves performance of several games:

   - Civilization VI: +17%
   - Shadow of Mordor: +6%
   - Bioshock Infinite +6%
   - Xonotic: +6%

Various microbenchmarks improve substantially as well:

   - GfxBench5 gl_driver2: +58%
   - SynMark2 OglBatch6: +54%
   - Piglit drawoverhead: +25%

Reviewed-by: Zoltán Böszörményi <zboszor@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8964>

3 years agoiris: Use thread safe slab allocators in transfer_map handling
Kenneth Graunke [Wed, 10 Feb 2021 23:09:11 +0000 (15:09 -0800)]
iris: Use thread safe slab allocators in transfer_map handling

pipe->transfer_map can be called from u_threaded_context's thread
rather than the driver thread.  We need to use two different slab
allocators, one for each thread.  transfer_unmap, on the other hand,
is only ever called from the driver thread.

Reviewed-by: Zoltán Böszörményi <zboszor@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8964>

3 years agoiris: Make various classes inherit from u_threaded_context base classes
Kenneth Graunke [Tue, 9 Feb 2021 00:39:42 +0000 (16:39 -0800)]
iris: Make various classes inherit from u_threaded_context base classes

u_threaded_context requires various objects to inherit from a new
threaded_foo base class rather than directly from pipe_foo.  This
patch does most of the mechanical changes required for that.

It also initializes the new threaded_resource fields.

Reviewed-by: Zoltán Böszörményi <zboszor@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8964>

3 years agoiris: Use different shader uploaders for precompile vs. draw time
Kenneth Graunke [Tue, 9 Feb 2021 01:00:36 +0000 (17:00 -0800)]
iris: Use different shader uploaders for precompile vs. draw time

When we enable u_threaded_context, the pipe->create_*_state hooks
(precompile variants) are going to be called from one thread, while
iris_update_compiled_shaders (on-the-fly variants) are going to be
called from a driver thread.  BLORP shaders also happen from
clear, blit, and so on in the driver thread.

u_upload_mgr isn't thread-safe, so use an uploader for each purpose.

Reviewed-by: Zoltán Böszörményi <zboszor@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8964>

3 years agoiris: Support rebinding of stream output targets
Kenneth Graunke [Wed, 3 Feb 2021 02:42:41 +0000 (18:42 -0800)]
iris: Support rebinding of stream output targets

This enables us to replace the backing storage of resources that have
been used as stream output targets, in case we're invalidating their
entire contents.  This can avoid stalls.  We simply hadn't supported it
because it was going to be tricky to re-emit 3DSTATE_SO_BUFFER without
screwing up "reset offset to zero" vs. "keep appending".  But that
should be working fine with the previous patch's refactor.

Reviewed-by: Zoltán Böszörményi <zboszor@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8964>

3 years agoiris: Rework zeroing of stream output buffer offsets
Kenneth Graunke [Wed, 3 Feb 2021 01:02:05 +0000 (17:02 -0800)]
iris: Rework zeroing of stream output buffer offsets

The previous mechanism was a bit fragile.  We stored the zero offset
in the pre-baked packet, and used an flag to override 0xFFFFFFFF
(append) offsets until our first emit - then prohibited anyone from
trying to re-emit the packet by flagging IRIS_DIRTY_SO_BUFFERS,
because that would re-emit the version with the zeroing of the offset.

Now, we always store 0xFFFFFFFF in the pre-baked packet, and use a
flag to override it to zero on the first emit.  That way, we can
re-emit that packet at any time, and it'll just keep appending.

Reviewed-by: Zoltán Böszörményi <zboszor@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8964>

3 years agoiris: Defer stream output target space allocation until set time
Kenneth Graunke [Mon, 1 Feb 2021 13:12:30 +0000 (05:12 -0800)]
iris: Defer stream output target space allocation until set time

In the future, Marek is planning to make u_threaded_context call
create_stream_output_target() from a different thread than the main
driver thread, which means that we can't safely use uploaders there.

To prepare for this eventual future, just defer the allocation of
the offset BO 'til later.  It's a very small amount of overhead.

Reviewed-by: Zoltán Böszörményi <zboszor@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Nanley Chery <nanley.g.chery@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8964>

3 years agoiris: Defer uploading of surface states
Kenneth Graunke [Mon, 1 Feb 2021 12:51:11 +0000 (04:51 -0800)]
iris: Defer uploading of surface states

With u_threaded_context, create_surface and create_sampler_view will
be called from a different thread than the driver thread.  They aren't
allowed to access the context, which means that they can't use the
uploaders there to upload our SURFACE_STATE entries.

Thanks to backing-storage replacement and iris_rebind_buffer, we already
reworked things to maintain CPU-side copies of the SURFACE_STATE entries
and added the ability to upload or re-upload them later.  So we can skip
the upload at object creation time, and add a simple resource-is-NULL
check at binding table upload time to ensure that they get uploaded by
the time we need them.  (They might get uploaded earlier due to rebinds
or clear color updates, but this is the last moment to do so.)

Reviewed-by: Zoltán Böszörményi <zboszor@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Nanley Chery <nanley.g.chery@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8964>

3 years agolima: avoid stomping over bound shader state when creating new shaders
Eric Anholt [Tue, 16 Feb 2021 20:55:54 +0000 (12:55 -0800)]
lima: avoid stomping over bound shader state when creating new shaders

It shouldn't affect bound program state, and the current context state
shouldn't be relevant for shader creation precompiles anyway (level load
isn't going to have the eventual set of sampler views bound when you go to
draw with that shader).

Closes: #4306
Signed-off-by: Erico Nunes <nunes.erico@gmail.com>
Reviewed-by: Vasily Khoruzhick <anarsoul@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9089>

3 years agolima: upload the shader to a BO at shader creation
Eric Anholt [Tue, 16 Feb 2021 20:49:01 +0000 (12:49 -0800)]
lima: upload the shader to a BO at shader creation

No need to conditionally upload later.

Signed-off-by: Erico Nunes <nunes.erico@gmail.com>
Reviewed-by: Vasily Khoruzhick <anarsoul@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9089>

3 years agolima: don't look at dirty bits for setup of FS key
Eric Anholt [Tue, 16 Feb 2021 20:45:08 +0000 (12:45 -0800)]
lima: don't look at dirty bits for setup of FS key

You always have to populate the key with the right texture swizzles, even
if textures haven't changed since binding a new shader.

Signed-off-by: Erico Nunes <nunes.erico@gmail.com>
Reviewed-by: Vasily Khoruzhick <anarsoul@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9089>

3 years agolima: stop encoding the texture format in the shader key
Eric Anholt [Tue, 16 Feb 2021 20:42:22 +0000 (12:42 -0800)]
lima: stop encoding the texture format in the shader key

We can compose the swizzles at sampler view creation time, saving
recompiles on texture format changes.

Signed-off-by: Erico Nunes <nunes.erico@gmail.com>
Reviewed-by: Vasily Khoruzhick <anarsoul@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9089>

3 years agoanv: implement INTEL_DEBUG=submit
Lionel Landwerlin [Wed, 17 Jun 2020 12:37:33 +0000 (15:37 +0300)]
anv: implement INTEL_DEBUG=submit

Name all the BOs!

v2: Fix 32bit build issue (Thanks Marge!)

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Marcin Ślusarz <marcin.slusarz@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5736>

3 years agovirgl: Add support for querying detailed memory info
Rohan Garg [Thu, 25 Feb 2021 17:13:22 +0000 (18:13 +0100)]
virgl: Add support for querying detailed memory info

This allows for virgl guests to expose GL_NVX_gpu_memory_info and
GL_ATI_meminfo when the extensions are supported on the host.

Signed-off-by: Rohan Garg <rohan.garg@collabora.com>
Reviewed-by: Gert Wollny <gert.wollny@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9337>

3 years agointel/mi_builder: Drop the gen_ prefix
Jason Ekstrand [Wed, 3 Mar 2021 18:29:39 +0000 (12:29 -0600)]
intel/mi_builder: Drop the gen_ prefix

mi_ is already a unique prefix in Mesa so the gen_ isn't really gaining
us anything except extra characters.  It's possible that MI_ may
conflict a tiny bit with GenXML but it doesn't seem to be a problem
today and we can deal with that in the future if it's ever an issue.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9393>

3 years agointel: Rename gen_mi_builder.h to mi_builder.h
Jason Ekstrand [Wed, 3 Mar 2021 18:09:11 +0000 (12:09 -0600)]
intel: Rename gen_mi_builder.h to mi_builder.h

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9393>

3 years agoir3: disallow moving memory writes over discard
Danylo Piliaiev [Tue, 2 Mar 2021 12:30:01 +0000 (14:30 +0200)]
ir3: disallow moving memory writes over discard

Writes to global memory should not be moved over discard,
otherwise we could have unintended side-effects or lack of
side-effects where they should be observed.

Fixes tests:
 dEQP-VK.rasterization.frag_side_effects.color_at_beginning.kill
 dEQP-VK.rasterization.frag_side_effects.color_at_end.kill

Signed-off-by: Danylo Piliaiev <dpiliaiev@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9365>

3 years agoci: Bump deqp to vk-gl-cts 1.2.5.2
Juan A. Suarez Romero [Tue, 2 Mar 2021 10:45:27 +0000 (11:45 +0100)]
ci: Bump deqp to vk-gl-cts 1.2.5.2

Reviewed-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Juan A. Suarez Romero <jasuarez@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9369>

3 years agoir3: make mark_kill_path exit early if instr is already seen
Danylo Piliaiev [Wed, 3 Mar 2021 10:16:55 +0000 (12:16 +0200)]
ir3: make mark_kill_path exit early if instr is already seen

Would bring down its complexity in pathological cases.

Signed-off-by: Danylo Piliaiev <dpiliaiev@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9386>

3 years agoir3: prevent duplication of instruction's dependencies
Danylo Piliaiev [Wed, 3 Mar 2021 10:08:53 +0000 (12:08 +0200)]
ir3: prevent duplication of instruction's dependencies

Otherwise mark_kill_path() is happy to take exponential time to finish.

It was possible to have such chains:
  ...
 stib.base0 imm[0.000000,0,0x0], ssa_233, ssa_234, false-deps:ssa_231, ssa_231
 stib.base0 imm[0.000000,0,0x0], ssa_237, ssa_238, false-deps:ssa_235, ssa_235
 stib.base0 imm[0.000000,0,0x0], ssa_241, ssa_242, false-deps:ssa_239, ssa_239
 stib.base0 imm[0.000000,0,0x0], ssa_245, ssa_246, false-deps:ssa_243, ssa_243
 stib.base0 imm[0.000000,0,0x0], ssa_249, ssa_250, false-deps:ssa_247, ssa_247
 stib.base0 imm[0.000000,0,0x0], ssa_105, ssa_253, false-deps:ssa_251, ssa_251
 stib.base0 imm[0.000000,0,0x0], ssa_109, ssa_256, false-deps:ssa_254, ssa_254
 stib.base0 imm[0.000000,0,0x0], ssa_113, ssa_259, false-deps:ssa_257, ssa_257
 stib.base0 imm[0.000000,0,0x0], ssa_117, ssa_262, false-deps:ssa_260, ssa_260
 stib.base0 imm[0.000000,0,0x0], ssa_265, ssa_266, false-deps:ssa_263, ssa_263
 stib.base0 imm[0.000000,0,0x0], ssa_269, ssa_270, false-deps:ssa_267, ssa_267
 stib.base0 imm[0.000000,0,0x0], ssa_273, ssa_274, false-deps:ssa_271, ssa_271
  ...

Fixes tests:
 dEQP-VK.geometry.layered.cube_array.36_36_12.secondary_cmd_buffer_inherit_framebuffer
 dEQP-VK.geometry.layered.3d.64_64_8.secondary_cmd_buffer_inherit_framebuffer
 dEQP-VK.geometry.layered.cube_array.64_64_12.secondary_cmd_buffer_inherit_framebuffer

Signed-off-by: Danylo Piliaiev <dpiliaiev@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9386>

3 years agoRevert "radv: stop using VM_ALWAYS_VALID on APUs"
Samuel Pitoiset [Mon, 1 Mar 2021 16:21:04 +0000 (16:21 +0000)]
Revert "radv: stop using VM_ALWAYS_VALID on APUs"

Disabling VM_ALWAYS_VALID actually hurts more than it helps
after doing more testing. Managing the global BO list in userspace
is really costly and make a bunch of games CPU bound.

I think re-enabling VM_ALWAYS_VALID is a step in the right direction.

This reverts commit 6ac6e2fbfb47e737f2f823fec2931d80769acc33.

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/9341>

3 years agor600/sfn: lower intrinsic_load_tess_coord to driver version
Gert Wollny [Tue, 2 Mar 2021 19:23:59 +0000 (20:23 +0100)]
r600/sfn: lower intrinsic_load_tess_coord to driver version

Fixes
  KHR-GL45.tessellation_shader.tessellation_shader_tessellation.TCS_TES
  KHR-GL45.tessellation_shader.tessellation_shader_tessellation.TES

Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9373>

3 years agonir: Add r600 specific intrinsic for loading the tesselation coords
Gert Wollny [Tue, 2 Mar 2021 19:21:53 +0000 (20:21 +0100)]
nir: Add r600 specific intrinsic for loading the tesselation coords

Only the XY pair is provided directly, the Z value has to be deducted
from the primitive type.

Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9373>

3 years agovirgl: add astc 2d compressed formats
cheyang [Fri, 26 Feb 2021 09:47:27 +0000 (17:47 +0800)]
virgl: add astc 2d compressed formats

Signed-off-by: cheyang <cheyang@bytedance.com>
Signed-off-by: hexin <hexin.op@bytedance.com>
Reviewed-by: Gert Wollny <gert.wollny@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9306>

3 years agobroadcom/compiler: be more aggressive skipping unifa writes
Iago Toral Quiroga [Tue, 2 Mar 2021 12:05:09 +0000 (13:05 +0100)]
broadcom/compiler: be more aggressive skipping unifa writes

We had an optimization in place to skip a unifa write if the address
happens to be right after the last ldunifa read address, but we can
take this further and update the unifa address by emitting ldunifa
instructions if needed to skip a unifa write that is close enough.
This is because a unifa write involves 4 cycles: 1 for the write
and 3 delay slots before we can emit the first ldunifa.

So if we have code like this:

unifa addr + 0
ldunifa.r0
unifa addr + 12
ldunifa.r1

In practice we end up with QPU like this:

unifa addr + 0
nop
nop
nop
ldunifa.r0
unifa addr + 12
nop
nop
nop
ldunifa.r1

And with this patch we get:

unifa addr + 0
nop
nop
nop
ldunifa.r0  <--- reads offset 0
ldunifa.-   <--- reads offset 4
ldunifa.-   <--- reads offset 8
ldunifa.r1  <--- reads offset 12

Of course, QPU scheduling might find ways to fill the NOPs to some
extent and remove some of the gains, but generally speaking, this is
still usually a win.

Going by shader-db results, allowing the next unifa address to be up
to 12 bytes after the address resulting from the last ldunifa read
shows the best results:

total instructions in shared programs: 13817048 -> 13812202 (-0.04%)
instructions in affected programs: 602701 -> 597855 (-0.80%)
helped: 1750
HURT: 760
Instructions are helped.

total uniforms in shared programs: 3795485 -> 3793200 (-0.06%)
uniforms in affected programs: 43930 -> 41645 (-5.20%)
helped: 898
HURT: 0
Uniforms are helped.

total max-temps in shared programs: 2326612 -> 2326621 (<.01%)
max-temps in affected programs: 651 -> 660 (1.38%)
helped: 10
HURT: 21
Inconclusive result (value mean confidence interval includes 0).

total sfu-stalls in shared programs: 30942 -> 30906 (-0.12%)
sfu-stalls in affected programs: 627 -> 591 (-5.74%)
helped: 186
HURT: 158
Inconclusive result (value mean confidence interval includes 0).

total inst-and-stalls in shared programs: 13847990 -> 13843108 (-0.04%)
inst-and-stalls in affected programs: 601404 -> 596522 (-0.81%)
helped: 1747
HURT: 757
Inst-and-stalls are helped.

Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9384>

3 years agobroadcom/compiler: drop the destination for unused ldunifa
Iago Toral Quiroga [Tue, 2 Mar 2021 14:37:30 +0000 (15:37 +0100)]
broadcom/compiler: drop the destination for unused ldunifa

We can't remove unused ldunifa that are not the first or last
in a sequence, but we can still ignore their destination
to reduce register pressure.

Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9384>

3 years agoutil/disk_cache: make MESA_DISK_CACHE_READ_ONLY_FOZ_DBS a relative path
Timothy Arceri [Thu, 25 Feb 2021 10:15:22 +0000 (21:15 +1100)]
util/disk_cache: make MESA_DISK_CACHE_READ_ONLY_FOZ_DBS a relative path

Rather than passing in full paths this changes things so that we can
just pass in filenames relative to the current cache directory.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9279>

3 years agoci/turnip: Mark a flaky WSI test.
Eric Anholt [Wed, 3 Mar 2021 21:51:12 +0000 (13:51 -0800)]
ci/turnip: Mark a flaky WSI test.

This one has flaked many times at this point, and I've even seen it flake
locally.  No luck debugging it yet.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9397>

3 years agofreedreno: Remove dead-cells MBR workaround
Rob Clark [Wed, 3 Mar 2021 19:49:06 +0000 (11:49 -0800)]
freedreno: Remove dead-cells MBR workaround

With threaded-context we won't have a chance to apply the workaround in
the backend driver.  But the previous commit moves it to a driconf
configured workaround in mesa/st, so we can drop this now.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9316>

3 years agodriconf: Add ignore_map_unsynchronized option
Rob Clark [Sun, 21 Feb 2021 18:36:50 +0000 (10:36 -0800)]
driconf: Add ignore_map_unsynchronized option

Add an option to workaround incorrect unsynchronized VBO updates in
Dead-Cells.

See: https://gitlab.freedesktop.org/mesa/mesa/-/issues/4337
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9316>

3 years agozink: rewrite macro for getting KHR device functions
Mike Blumenkrantz [Wed, 3 Mar 2021 21:44:09 +0000 (16:44 -0500)]
zink: rewrite macro for getting KHR device functions

we have the technology. we can improve our our lives with better macros.

Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9398>

3 years agofreedreno/a6xx: Fix compile warning
Rob Clark [Wed, 3 Mar 2021 19:05:52 +0000 (11:05 -0800)]
freedreno/a6xx: Fix compile warning

Fixes: 79921b81bcf ("freedreno/a6xx: Document threadsize-related fields")
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9394>

3 years agofreedreno: Deduplicate fixup_shader_state()
Rob Clark [Wed, 3 Mar 2021 18:51:55 +0000 (10:51 -0800)]
freedreno: Deduplicate fixup_shader_state()

All the ir3 gens had the same thing, time to move it out into a shared
helper.

The keeping the storage in fdN_context is to avoid namespace clashes
between ir3 and ir2.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9394>

3 years agofreedreno/ir3: Add comments about shader key/gen
Rob Clark [Wed, 3 Mar 2021 18:24:17 +0000 (10:24 -0800)]
freedreno/ir3: Add comments about shader key/gen

I had forgotton on which gens these where used on (which is important if
you need to know which shader stages use these).. expand the comments a
bit.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9394>

3 years agoclover: fix array images view creation
Dave Airlie [Wed, 3 Mar 2021 02:44:46 +0000 (12:44 +1000)]
clover: fix array images view creation

Found this on top of Karol's patches but it seems like it can just be
applied to master.

Helps with some cases of
kernel_image_methods/test_kernel_image_methods 2Darray

Reviewed-by: Karol Herbst <kherbst@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9381>

3 years agoci/zink: Add another primitive restart flake.
Eric Anholt [Wed, 3 Mar 2021 21:08:26 +0000 (13:08 -0800)]
ci/zink: Add another primitive restart flake.

This one flaked all the way to a run failure in a recent MR of mine.

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9396>

3 years agoci/a5xx: Update piglit expectations.
Eric Anholt [Fri, 26 Feb 2021 21:47:12 +0000 (13:47 -0800)]
ci/a5xx: Update piglit expectations.

The mesa/st shader variants change fixed some fails for us.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9314>

3 years agoci/a5xx: Increase the gles3/31 coverage.
Eric Anholt [Fri, 26 Feb 2021 20:26:08 +0000 (12:26 -0800)]
ci/a5xx: Increase the gles3/31 coverage.

Now that there's more time available in our budget per board, we can run
all of gles31, and half of gles3, instead of 10%.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9314>

3 years agoci/a3xx: Run all of GLES3 dEQP.
Eric Anholt [Fri, 26 Feb 2021 20:55:49 +0000 (12:55 -0800)]
ci/a3xx: Run all of GLES3 dEQP.

We're not spending half our time booting any more, so run the other half.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9314>

3 years agoci/a5xx: Run all of gles2 in one job.
Eric Anholt [Fri, 26 Feb 2021 20:24:50 +0000 (12:24 -0800)]
ci/a5xx: Run all of gles2 in one job.

Now that we're not spending so much time on boot overhead, no need to
parallelize.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9314>

3 years agoci/freedreno: Switch the fastboot boards to using nfsroot.
Eric Anholt [Fri, 26 Feb 2021 18:24:31 +0000 (10:24 -0800)]
ci/freedreno: Switch the fastboot boards to using nfsroot.

This saves time in packing the rootfs, allows for larger rootfses, and
avoids the need for webdav.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9314>

3 years agoci/freedreno: Also retest when only CI configuration changes.
Eric Anholt [Fri, 26 Feb 2021 20:47:44 +0000 (12:47 -0800)]
ci/freedreno: Also retest when only CI configuration changes.

Fixes: dab845d457eb ("ci: Move specific driver testing to separate files in separate dirs.")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9314>

3 years agotgsi_exec: Jump over entirely non-taken THEN or ELSE branches.
Eric Anholt [Mon, 1 Mar 2021 20:23:56 +0000 (12:23 -0800)]
tgsi_exec: Jump over entirely non-taken THEN or ELSE branches.

TGSI has these nice labels for us for where to jump in this case, let's
use them.  Improves piglit arb_shader_image_load_store-shader-mem-barrier
runtime massively, though not enough to make the test really reasonable to
run.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9347>

3 years agotgsi_exec: Roll the loops for condmask handling.
Eric Anholt [Mon, 1 Mar 2021 20:19:48 +0000 (12:19 -0800)]
tgsi_exec: Roll the loops for condmask handling.

No need to hand-unroll this, the compiler will do it.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9347>

3 years agoi965: support GL_EXT_color_buffer_half_float
Ilia Mirkin [Tue, 2 Mar 2021 21:05:59 +0000 (16:05 -0500)]
i965: support GL_EXT_color_buffer_half_float

FP16 rendering is supported on all gen4 hardware.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
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/9379>

3 years agoci: Use lock file to build deqp-runner
Antonio Caggiano [Tue, 2 Mar 2021 15:01:05 +0000 (16:01 +0100)]
ci: Use lock file to build deqp-runner

Build deqp-runner with the `--locked` option to use dependencies
versions specified in `Cargo.lock`.

v2: Bump image tags.

Signed-off-by: Antonio Caggiano <antonio.caggiano@collabora.com>
Suggested-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9368>

3 years agoac/llvm: open code fpow on LLVM 12 using fmul.legacy
Marek Olšák [Fri, 12 Feb 2021 06:06:18 +0000 (01:06 -0500)]
ac/llvm: open code fpow on LLVM 12 using fmul.legacy

A quick look at the asm shows that this enables source modifiers
(neg, abs) for v_mul_legacy_f32.

Totals from affected shaders:
SGPRS: 110104 -> 110400 (0.27 %)
VGPRS: 57632 -> 57636 (0.01 %)
Spilled SGPRs: 66 -> 63 (-4.55 %)
Spilled VGPRs: 0 -> 0 (0.00 %)
Private memory VGPRs: 0 -> 0 (0.00 %)
Scratch size: 0 -> 0 (0.00 %) dwords per thread
Code Size: 3290412 -> 3283068 (-0.22 %) bytes
Max Waves: 32141 -> 32141 (0.00 %)

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9395>

3 years agoac/llvm: add type parameter into ac_build_buffer_load to fix 16-bit TES inputs
Marek Olšák [Mon, 15 Feb 2021 06:25:25 +0000 (01:25 -0500)]
ac/llvm: add type parameter into ac_build_buffer_load to fix 16-bit TES inputs

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9395>

3 years agoac/llvm: fix visit_load_ubo_buffer to use SMEM for 16 bits instead of VMEM
Marek Olšák [Fri, 12 Feb 2021 08:54:19 +0000 (03:54 -0500)]
ac/llvm: fix visit_load_ubo_buffer to use SMEM for 16 bits instead of VMEM

This has 3 advantages:
- It's SMEM.
- Multiple single component loads are merged into 1 multi-dword load by LLVM.
- The result is always packed for packed instructions.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9395>

3 years agoac/llvm: implement 16-bit and 64-bit fpow correctly
Marek Olšák [Fri, 12 Feb 2021 06:05:19 +0000 (01:05 -0500)]
ac/llvm: implement 16-bit and 64-bit fpow correctly

LLVM converts to 32 bits and back for llvm.pow, so we can't use it.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9395>

3 years agoac/llvm: add support for 16-bit source operands for samplers
Marek Olšák [Thu, 11 Feb 2021 08:22:00 +0000 (03:22 -0500)]
ac/llvm: add support for 16-bit source operands for samplers

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9395>

3 years agonir/search: Constify instruction parameter to search helpers
Ian Romanick [Tue, 12 May 2020 19:48:17 +0000 (12:48 -0700)]
nir/search: Constify instruction parameter to search helpers

The search helps must *never* modify the instruction passed in, so let
the compiler enforce this.

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9378>

3 years agoanv: fix missing general state pool in validation list
Lionel Landwerlin [Wed, 3 Mar 2021 13:21:55 +0000 (15:21 +0200)]
anv: fix missing general state pool in validation list

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Fixes: 83fee30e85965c ("anv: allow multiple command buffers in anv_queue_submit")
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9388>

3 years agoci/lava: Move the driver expectation files to the per-driver CI dir.
Eric Anholt [Mon, 1 Mar 2021 20:53:37 +0000 (12:53 -0800)]
ci/lava: Move the driver expectation files to the per-driver CI dir.

This will cause less retesting of other drivers when changing the dEQP
results for a driver.

Reviewed-by: Michel Dänzer <mdaenzer@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9353>

3 years agoci/lava: Move the per-driver gitlab-ci.yml to each driver.
Eric Anholt [Mon, 1 Mar 2021 20:50:51 +0000 (12:50 -0800)]
ci/lava: Move the per-driver gitlab-ci.yml to each driver.

Follow-up to !9139, will cause less testing of other drivers when changing
the CI configuration for a single driver.

Reviewed-by: Michel Dänzer <mdaenzer@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9353>

3 years agoci: Move deqp-default-skips.txt back to .gitlab-ci/
Eric Anholt [Mon, 1 Mar 2021 20:57:30 +0000 (12:57 -0800)]
ci: Move deqp-default-skips.txt back to .gitlab-ci/

Since we don't manually enumerate the drivers using it, we have to retest
all drivers when changing it (which basically never happens, anyway).

Reviewed-by: Michel Dänzer <mdaenzer@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9353>

3 years agoradv: fix RGP barrier layout transition for TC-compatible CMASK images
Samuel Pitoiset [Wed, 3 Mar 2021 16:11:15 +0000 (17:11 +0100)]
radv: fix RGP barrier layout transition for TC-compatible CMASK images

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/9390>

3 years agozink: Enable GL_EXT_depth_bounds_test
Adam Jackson [Tue, 2 Mar 2021 17:28:04 +0000 (12:28 -0500)]
zink: Enable GL_EXT_depth_bounds_test

Available since Vulkan 1.0, and in fact already wired up, just not
advertised. It looks like we could make this dynamic state but this
works for now.

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9371>

3 years agoradv: don't shrink image stores for The Surge 2
Rhys Perry [Tue, 23 Feb 2021 19:33:02 +0000 (19:33 +0000)]
radv: don't shrink image stores for The Surge 2

The game seems to declare the wrong format.

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Fixes: e4d75c22 ("nir/opt_shrink_vectors: shrink image stores using the format")
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/4347
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9229>

3 years agonir/opt_shrink_vectors: add option to skip shrinking image stores
Rhys Perry [Tue, 23 Feb 2021 19:31:48 +0000 (19:31 +0000)]
nir/opt_shrink_vectors: add option to skip shrinking image stores

Some games declare the wrong format, so we might want to disable this
optimization in that case.

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Fixes: e4d75c22 ("nir/opt_shrink_vectors: shrink image stores using the format")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9229>

3 years agoturnip: fix leak of tu_shader object during compute pipeline creation
Danylo Piliaiev [Tue, 2 Mar 2021 10:00:58 +0000 (12:00 +0200)]
turnip: fix leak of tu_shader object during compute pipeline creation

tu_shader should be freed after pipeline is successfully created.

Fixes tests:
 dEQP-VK.api.object_management.alloc_callback_fail.compute_pipeline
 dEQP-VK.api.object_management.alloc_callback_fail_multiple.compute_pipeline

Signed-off-by: Danylo Piliaiev <dpiliaiev@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9364>

3 years agoradv: bump the initial SQTT buffer size to 32MB per SE
Samuel Pitoiset [Tue, 2 Mar 2021 14:33:36 +0000 (15:33 +0100)]
radv: bump the initial SQTT buffer size to 32MB per SE

Most of the games need 32MB or more, but rarely less.

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/9367>

3 years agoradv: trigger a new SQTT capture automatically after resizing the buffer
Samuel Pitoiset [Tue, 2 Mar 2021 14:29:25 +0000 (15:29 +0100)]
radv: trigger a new SQTT capture automatically after resizing the buffer

It's way better.

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/9367>

3 years agoradv: double the SQTT buffer size when it is resized
Samuel Pitoiset [Tue, 2 Mar 2021 14:14:50 +0000 (15:14 +0100)]
radv: double the SQTT buffer size when it is resized

Computing the expected buffer size isn't reliable on GFX10+ because
DROPPED_CNTR returns weird results.

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/9367>

3 years agoac/sqtt: fix determining if the trace is complete on GFX10+
Samuel Pitoiset [Tue, 2 Mar 2021 14:13:24 +0000 (15:13 +0100)]
ac/sqtt: fix determining if the trace is complete on GFX10+

DROPPED_CNTR isn't reliable and might still report non-zero if the
SQTT buffer isn't full. Checking if the number of written bytes by
the hw is equal to the SQTT buffer size seems reliable.

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/9367>

3 years agoradv: do not trace inactive shader engines with SQTT
Samuel Pitoiset [Tue, 2 Mar 2021 15:42:39 +0000 (16:42 +0100)]
radv: do not trace inactive shader engines with SQTT

This fixes a GPU hang on my Sienna because the number of SE is
less than the maximum, and SE #1 is disabled.

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/9370>

3 years agozink: ci updates
Mike Blumenkrantz [Wed, 3 Mar 2021 01:21:00 +0000 (20:21 -0500)]
zink: ci updates

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9291>

3 years agozink: use staging resource for write transfer_map in order to not stall
Mike Blumenkrantz [Wed, 9 Sep 2020 20:08:54 +0000 (16:08 -0400)]
zink: use staging resource for write transfer_map in order to not stall

we can just give the user a staging resource and then flush the data back
later

Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9291>

3 years agoradeonsi: don't crash on NULL images in si_check_needs_implicit_sync
Marek Olšák [Tue, 2 Mar 2021 07:33:36 +0000 (02:33 -0500)]
radeonsi: don't crash on NULL images in si_check_needs_implicit_sync

This fixes CTS test: KHR-GL46.arrays_of_arrays_gl.AtomicUsage

Fixes: bddc0e023c "radeonsi: fix read from compute / write from draw sync"

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Reviewed-by: Zoltán Böszörményi <zboszor@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9361>

3 years agoac/llvm: fix ac_build_atomic_rmw with LLVM 13
Marek Olšák [Tue, 2 Mar 2021 07:59:23 +0000 (02:59 -0500)]
ac/llvm: fix ac_build_atomic_rmw with LLVM 13

Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/4383

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Reviewed-by: Zoltán Böszörményi <zboszor@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9361>

3 years agonir/vec_to_movs: Don't generate MOVs for undef channels.
Eric Anholt [Mon, 1 Mar 2021 19:57:19 +0000 (11:57 -0800)]
nir/vec_to_movs: Don't generate MOVs for undef channels.

This appeared in softpipe's image operations, since NIR always uses
4-component values for the coords, while the GLSL IR only has 2 components
for a 2D image (for example).
arb_shader_image_load_store-shader-mem-barrier (which times out in CI and
spends its time inside of tgsi_exec) was spending 4/51 of its instructions
on moving these undefs around.

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9345>

3 years agonir: Add a nir_src_is_undef() helper, like nir_src_is_const().
Eric Anholt [Mon, 1 Mar 2021 23:04:25 +0000 (15:04 -0800)]
nir: Add a nir_src_is_undef() helper, like nir_src_is_const().

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9345>

3 years agozink: export PIPE_CAP_TGSI_VS_LAYER_VIEWPORT
Mike Blumenkrantz [Mon, 14 Dec 2020 16:41:26 +0000 (11:41 -0500)]
zink: export PIPE_CAP_TGSI_VS_LAYER_VIEWPORT

Acked-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9283>

3 years agozink: enable PIPE_CAP_CLEAR_SCISSORED
Mike Blumenkrantz [Wed, 14 Oct 2020 13:10:17 +0000 (09:10 -0400)]
zink: enable PIPE_CAP_CLEAR_SCISSORED

Acked-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9283>

3 years agolavapipe: sort bindings before creating descriptor set
Dave Airlie [Tue, 2 Mar 2021 20:50:00 +0000 (06:50 +1000)]
lavapipe: sort bindings before creating descriptor set

This ensures the dynamic offsets are correct

Fixes: b38879f8c5f5 ("vallium: initial import of the vulkan frontend")
Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9359>