Commit graph

14085 commits

Author SHA1 Message Date
Samuel Pitoiset
76ccf71587 radv: simplify binding the GS copy shader with ESO
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27699>
2024-02-22 13:05:28 +00:00
Samuel Pitoiset
59b0f7b6b7 radv: free NIR shaders when creating linked shaders with ESO
This was leaking memory too.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27715>
2024-02-22 07:48:02 +00:00
Samuel Pitoiset
77f036d574 radv: fix a big memleak with VK_EXT_shader_object
Shaders and binaries weren't freed at all!

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27715>
2024-02-22 07:48:02 +00:00
Samuel Pitoiset
e88a5842b2 radv: remove the union in radv_shader_object
It's actually more annoying to deal with.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27715>
2024-02-22 07:48:02 +00:00
Samuel Pitoiset
9ae8f0f9d7 radv: fix indirect dispatches on compute queue with conditional rendering on GFX7
COND_EXEC needs to happen right before PKT3_DISPATCH_INDIRECT.

Like this combination will probably never happen but better to have
it fixed anyways.

Fixes: 5c03cdbd02 ("radv: fix indirect dispatches on the compute queue on GFX7")
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27655>
2024-02-22 07:23:43 +00:00
Samuel Pitoiset
776f4523f9 radv: fix binary shaders compatibility with ESO
RADV needs to assume that all features that affect shaders might be
enabled. Note that disable_trunc_coord is specific to DXVK and
mesh_shader_queries is still disabled by default.

This fixes dEQP-VK.shader_object.binary.device_features.* on GFX10+.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27632>
2024-02-22 07:03:04 +00:00
Samuel Pitoiset
f123a04256 radv: introduce a per physical device cache key
This introduces a new level cache key for physical device. The main
motivation is for shader object because the Vulkan spec says:

"Guaranteed compatibility of shader binaries is expressed through a
 combination of the shaderBinaryUUID and shaderBinaryVersion members of
 the VkPhysicalDeviceShaderObjectPropertiesEXT structure queried from a
 physical device. Binary shaders retrieved from a physical device with
 a certain shaderBinaryUUID are guaranteed to be compatible with all
 other physical devices reporting the same shaderBinaryUUID and the
 same or higher shaderBinaryVersion."

Meaning that with ESO, the driver needs to compile shaders for the
worst case with every possible logical device features enabled.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27632>
2024-02-22 07:03:04 +00:00
Samuel Pitoiset
c303d399ff radv: initialize disk cache slightly later when creating a physical device
This will allow us to use a per physical device cache key.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27632>
2024-02-22 07:03:04 +00:00
Samuel Pitoiset
0543394bfa radv: move mesh_fast_launch_2 to radv_physical_device
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27632>
2024-02-22 07:03:04 +00:00
Rhys Perry
d0595e4805 aco/ra: fix GFX9- writelane
061b8bfd29 moved handling of fixed operands earlier, but it should have
moved the fixing of writelane operands earlier too.

This fixes Crucible's func.uniform-subgroup.exclusive.imin64 on GFX8.

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Fixes: 061b8bfd29 ("aco/ra: rework fixed operands")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27583>
2024-02-21 12:40:14 +00:00
Rhys Perry
53800191a3 aco/ra: don't initialize assigned in initializer list
According to Valgrind, vcc/m0 are uninitialized and this fixes it.

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27583>
2024-02-21 12:40:14 +00:00
Samuel Pitoiset
a5bdc4b19a radv: fix a compilation warning in radv_bind_graphics_shaders()
../src/amd/vulkan/radv_cmd_buffer.c: In function ‘radv_bind_graphics_shaders’:
../src/amd/vulkan/radv_cmd_buffer.c:9366:65: warning: array subscript -1 is below array bounds of ‘struct radv_shader *[14]’ [-Warray-bounds=]
 9366 |    cmd_buffer->state.last_vgt_shader = cmd_buffer->state.shaders[last_vgt_api_stage];

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27670>
2024-02-20 08:47:18 +00:00
Samuel Pitoiset
2deaf632e2 radv: use radv_shader_info for computing NGG LDS layout
This should be similar.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27575>
2024-02-19 16:30:08 +00:00
Samuel Pitoiset
21577bfacb radv: store the number of outputs for VS/TES as NGG
To compute NGG info without relying on NIR.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27575>
2024-02-19 16:30:08 +00:00
Samuel Pitoiset
5e1da1d10f radv: determine the number of invocations only for VS/TES as NGG
To make it clear that this doesn't depend on GS.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27575>
2024-02-19 16:30:08 +00:00
Samuel Pitoiset
12b5c3f00e radv: set radv_vgt_shader_key::ngg_streamout for ESO
Not used yet, but will be needed.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27575>
2024-02-19 16:30:08 +00:00
Samuel Pitoiset
9f660d8c72 radv: prefix radv_vgt_shader_key::streamout with ngg
This is only for NGG streamout (GFX11).

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27575>
2024-02-19 16:30:08 +00:00
Samuel Pitoiset
b98a7abcd6 radv: determine the ES stage for merged NGG shaders compiled separately
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27575>
2024-02-19 16:30:08 +00:00
Samuel Pitoiset
eea75777be radv: pass the ES stage when emitting geometry shader with ESO
Not used yet, but will be needed for NGG support.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27575>
2024-02-19 16:30:08 +00:00
Samuel Pitoiset
10b8e52780 radv: disable conditional rendering if enabled when preprocessing DGC
vkCmdPreprocessGeneratedCommandsNV() shouldn't be affected by
conditional rendering. Confirmed with Ricardo after writing some
tests against NVIDIA.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27658>
2024-02-19 08:07:41 +01:00
Samuel Pitoiset
5b0e4dd6e8 radv: stop using conditional rendering internally when preprocessing DGC
Preprocess and execute can be executed on different queue families. If
preprocess is executed on GFX and the indirect count optimization is
enabled, this will generate an empty cmdbuf. Expect GPU hangs if this
DGC IB is then executed on ACE.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27658>
2024-02-19 08:05:13 +01:00
Bas Nieuwenhuizen
c7b2ac3377 radv: Remove ray_launch_size_addr_amd system value.
Not used anymore, so clean it up.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27664>
2024-02-17 11:08:16 +00:00
Chia-I Wu
ec5d0ffb04 radv: fix pipeline stats mask
Left-shifting by 11*8 or 14*8 is undefined.  This fixes many
dEQP-VK.query_pool.statistics_query.* failures (but not pre-existing
flakes) for release builds using clang.

Fixes: 48aabaf225 ("radv: do not harcode the pipeline stats mask for query resolves")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27651>
2024-02-16 17:04:30 +00:00
Samuel Pitoiset
7009f00ae0 radv: re-enable GS_FAST_LAUNCH=2 by default on GFX11
The performance issue is fixed. This adds RADV_DEBUG=nogsfastlaunch2
to disable it in case users still report issues with it.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27636>
2024-02-16 15:58:31 +00:00
Samuel Pitoiset
3713f27260 radv: only configure {XYZ_DIM,DRAW_INDEX}_REG for mesh shaders if enabled
This seems to be the root cause of the huge performance issue with
GS_FAST_LAUNCH=2 on RDNA3 dGPUs. Probably a firmware bug, but on dGPUs
only, because on RDNA3 APUs this is just working...

With that GS_FAST_LAUNCH=2 can be re-enabled by default on GFX11.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27636>
2024-02-16 15:58:31 +00:00
Martin Roukala (né Peres)
dce2069054 radv/ci: switch vkcts-polaris10 from mupuf to KWS' farm
The DUT that currently hosts the Polaris10 will be used for NVK
testing, so let's move the polaris10 job to another farm.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27652>
2024-02-16 14:08:33 +00:00
Samuel Pitoiset
6894692d27 radv: enable GS_FAST_LAUNCH=2 by default for RDNA3 APUs (Phoenix)
GS_FAST_LAUNCH=1 shouldn't be used on GFX11 but it's still needed for
dGPUs (eg. NAVI31) because it destroys performance for unknown reasons.

On RDNA3 APUs, GS_FAST_LAUNCH=2 seems to be required for working
mesh shaders and performance is fine. There is possibly a firmware bug
on APUs that would explain why GS_FAST_LAUNCH=1 doesn't work on Phoenix.

Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/10583
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/10397
Cc: mesa-stable
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27618>
2024-02-16 07:32:18 +00:00
Samuel Pitoiset
0c2213cbbd radv: fix setting the rasterized primitive for ESO
This was missing.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27479>
2024-02-16 07:08:36 +00:00
Samuel Pitoiset
b199aa1bda radv: fix selecting shader variants with ESO
The shader BO for variants (like as LS) wasn't added to the BO list.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27479>
2024-02-16 07:08:36 +00:00
Samuel Pitoiset
a47de04ff3 radv: make sure to reset the GS copy shader with ESO
It should be NULL when not used. This is also slightly more optimal
than doing it in the loop.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27479>
2024-02-16 07:08:36 +00:00
Samuel Pitoiset
b1d4a7a95e radv: fix re-emitting DB_RENDER_CONTROL when resetting gfx pipeline state
This is only used by internal operations but the state wasn't
re-emitted.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27479>
2024-02-16 07:08:36 +00:00
Samuel Pitoiset
64e18b04d9 radv: clear the custom blend mode when resetting gfx pipeline state
This custom blend mode is only used by internal operations and they
don't use shader objects.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27479>
2024-02-16 07:08:35 +00:00
Samuel Pitoiset
b0a31a62e4 radv: clear RADV_CMD_DIRTY_SHADERS when resetting the shader object state
Binding a pipeline resets the shader object state and it's more optimal
to clear this flag here.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27479>
2024-02-16 07:08:35 +00:00
Georg Lehmann
ad5fbc4407 aco: use fmamk/ak instead of fma with inline constant for more VOPD
Foz-DB navi31, forced wave32:
Totals from 24438 (31.29% of 78112) affected shaders:
Instrs: 21632788 -> 21551766 (-0.37%); split: -0.38%, +0.01%
CodeSize: 126181860 -> 126083848 (-0.08%); split: -0.10%, +0.02%
Latency: 162491062 -> 162516234 (+0.02%); split: -0.05%, +0.07%
InvThroughput: 31121194 -> 31002125 (-0.38%); split: -0.40%, +0.02%
VClause: 420176 -> 420169 (-0.00%); split: -0.00%, +0.00%
SClause: 791844 -> 791762 (-0.01%); split: -0.01%, +0.00%

Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27523>
2024-02-15 12:38:55 +01:00
Hans-Kristian Arntzen
786c1b8322 radv: export multiview in VS/TES/GS for depth-only rendering
For depth-only rendering, a VS would not export layer properly.

Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/10606
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
cc: mesa-stable

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27615>
2024-02-15 09:07:42 +00:00
Georg Lehmann
e136a0629d radv/gfx11+: add rtwave32 perftest option
Useful for testing compiler changes and performance considerations.

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27584>
2024-02-14 17:11:01 +00:00
Samuel Pitoiset
32c1e45718 radv: fix emitting VS prologs for merged shaders compiled separately on GFX10+
RSRC1 isn't equal to the VS RSRC1 and both config registers need to
be re-emitted.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27574>
2024-02-13 14:01:42 +00:00
Samuel Pitoiset
6762307698 radv: cleanup radv_shader_combine_cfg_vs_tcs()
To match radv_shader_combine_cfg_vs_gs().

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27574>
2024-02-13 14:01:42 +00:00
Georg Lehmann
6121497228 aco/gfx11+: limit hard clauses to 32 instructions
https://github.com/llvm/llvm-project/pull/81287

Foz-DB Navi31:
Totals from 406 (0.52% of 78112) affected shaders:
Instrs: 585342 -> 585750 (+0.07%)
CodeSize: 3077856 -> 3079456 (+0.05%); split: -0.00%, +0.05%
Latency: 3263165 -> 3263326 (+0.00%); split: -0.00%, +0.01%
InvThroughput: 664092 -> 664114 (+0.00%); split: -0.00%, +0.00%
VClause: 11143 -> 11537 (+3.54%)
SClause: 11878 -> 11884 (+0.05%)
Copies: 39807 -> 39815 (+0.02%)

Cc: mesa-stable
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27569>
2024-02-13 13:40:52 +00:00
Samuel Pitoiset
0c05bdf1c1 radv/ci: enable RADV_PERFTEST=shader_object on VEGA10
Renoir currently hangs in Mesa CI. Needs to be investigated.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27433>
2024-02-13 09:14:21 +00:00
Samuel Pitoiset
bead3f2ec3 radv: allow RADV_PERFTEST=shader_object on GFX9/VEGA10
It's passing VKCTS on VEGA10 but for some reasons RENOIR currently
hangs.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27433>
2024-02-13 09:14:21 +00:00
Rhys Perry
926d9f1cef radv: support minmax filter for more formats
Support should be the same as AMDVLK, except for these formats:
- VK_FORMAT_R4G4_UNORM_PACK8
- VK_FORMAT_A4R4G4B4_UNORM_PACK16_EXT
- VK_FORMAT_A4B4G4R4_UNORM_PACK16_EXT
- VK_FORMAT_A1B5G5R5_UNORM_PACK16_KHR
- VK_FORMAT_A8_UNORM_KHR
- VK_FORMAT_X8_D24_UNORM_PACK32
- VK_FORMAT_D24_UNORM_S8_UINT
And the various emulated compressed formats.

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/27551>
2024-02-12 20:05:27 +00:00
Konstantin Seurer
fb62bffcda radv: Wire up ac_gather_context_rolls
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27322>
2024-02-12 14:04:24 +00:00
Konstantin Seurer
ba6d6e5ee1 amd/common: Use the correct register table for GFX10_3
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27322>
2024-02-12 14:04:24 +00:00
Samuel Pitoiset
6cab5559f9 radv: add support for emitting TES+GS compiled separately on GFX9+
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27432>
2024-02-12 08:09:28 +00:00
Samuel Pitoiset
dd92f5f664 radv: bind the vertex input SGPR only for relevant stages
Otherwise, user_data_0 is wrong if merged shaders are compiled
separately and if we have GS.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27432>
2024-02-12 08:09:28 +00:00
Samuel Pitoiset
d64d7373f3 radv: declare AC_UD_TES_STATE for separate compilation of GS on GFX9+
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27432>
2024-02-12 08:09:28 +00:00
Samuel Pitoiset
e15d1ed7cb radv: declare streamout buffers for TES+GS compiled separately on GFX9+
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27432>
2024-02-12 08:09:28 +00:00
Samuel Pitoiset
83bc7e27a5 radv: force GS stage for TES as ES compiled separately on GFX9+
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27432>
2024-02-12 08:09:28 +00:00
Samuel Pitoiset
b58de424f4 radv: fix RGP barrier reason for RP barriers inserted by the runtime
Without that, RGP is confused and it's reporting CmdPipelineBarrier()
instead of CmdRenderPassSync().

Cc: mesa-stable
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27546>
2024-02-12 07:50:16 +00:00