This reduces the penalty the heuristic gives to SIMD32 shaders
relative to SIMD16 in presence of discard control flow on Xe2+. The
penalty was meant to account for the inefficient divergence behavior
of SIMD32 shaders on Gfx12.x platforms, since Gfx12 hardware had EUs
bundled in groups of two, and each pair shared control flow logic so
both EUs could only execute instructions in lockstep, which meant that
SIMD32 shaders had an effective warp size of 64 on Gfx12.x.
This change switches back to more optimistic modelling of discard
divergence. With it we gain about 6% performance in a Shadow of the
Tomb Raider trace (tested on BMG).
One may wonder if there are still workloads that would suffer
materially from enabling SIMD32 for all pixel shaders on Xe2 instead
of using this heuristic, since Xe2 EUs have twice the GRF space, twice
the FPU throughput and better divergence behavior than Xe, but the
answer seems to be yes unfortunately: E.g. Superposition has some
pixel shaders where SIMD32 has substantially worse scheduling due to
the increased number of false dependencies due to higher register
pressure, and using SIMD32 for them reduces performance significantly.
The heuristic seems to model this correctly so it doesn't look like we
can do without it at least right now on Xe2.
Acked-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/31697>
This very test was working until the commit 4da147a02b
("mesa: remove fallback for GL_DEPTH_STENCIL"). Indeed this
commit lets the driver handles this path and this was
failing on evergreen r600.
The test was processed through r600_blit() which loads the
fragment shader util_make_fs_blit_zs(). This fragment shader
loads two textures the stencil and depth. The texture depth
was processed properly but the other texture was generating
incorrect values. This issue, which seems to be related to
the hardware configuration, disappears when the underlying
surface is allocated using a width multiple of 32.
This change was tested on cayman and palm with the normal test:
"piglit/bin/ext_packed_depth_stencil-getteximage -auto -fb" and
the test was modified to test all the relevant width and height
values. The gpu rv770 was not affected by this issue. Here is
the result:
spec/ext_packed_depth_stencil/getteximage: fail pass
Cc: mesa-stable
Signed-off-by: Patrick Lerda <patrick9876@free.fr>
Acked-by: Gert Wollny <gert.wollny@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/31757>
According to pandecode, r32 is global attribute offset and r36 is vertex
offset. Follow panfrost to use r36 instead of r32 for both non-indexed
firstVertex and indexed vertexOffset.
With this, gl_VertexIndex stops being zero-based which is incorrect.
Fixes: 5544d39f44 ("panvk: Add a CSF backend for panvk_queue/cmd_buffer")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/31810>
We don't need to bend over backwards to try to repair the consequences
of something that happened earlier when we can just prevent that from
happening in the first palce instead.
While we're at it, also move the cmd_dispatch initialization into the
conditional block, because that's only used by the secondary-buffer
emulation code.
This fixes WSI, because there's more differences than just secondary
command buffers between the device-level dispatch-table and the
cmd_dispatch table.
Fixes: c2299b6642 ("panvk/csf: Implement vkCmdExecuteCommands")
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/31819>
We don't dirty the CS state, so if a 3D blit comes between binding a
compute pipeline and executing a dispatch then we won't re-emit the
pipeline and invalidating CS state causes immediates emitted via
CP_LOAD_STATE to disappear. Fixes
dEQP-VK.binding_model.descriptor_buffer.ycbcr_sampler.compute_comp.
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/31764>
This GPU seems to have half the compute constlen of other a7xx GPUs,
because there are sporadic hangs in dEQP-VK.robustness.robustness2.* and
other tests unless we limit the constlen. This does *not* happen on
SM8550-HDK, so it does seem to be specific to the GPU in x1e laptops.
Fixes: b0d22461b9 ("freedreno: Enable the X1-85")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/31764>
Setting the surface index is optimal for performance in case of
multiple MRTs because addrlib rotates tiles differently.
But this should be disabled when the image is used for descriptor
buffers capture&replay because the descriptor isn't immutable
(ie. tile_swizzle can be different).
This fixes dEQP-VK.binding_model.descriptor_buffer.capture_replay.*
on some GPUs where tile_swizzle is non-zero.
Fixes: 3b57a35ece ("radv: Enable descriptorBufferCaptureReplay.")
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/31818>
Application should send the full DPB state in each pic to keep
the references alive. Ideally the surfaces would be evicted immediately,
but unfortunately this breaks some applications. Add evict flag and
only evict surfaces if not present in reference frames array for two
consecutive frames.
DPB buffers are not destroyed upon eviction, but instead they are kept
around and reused next time a new surface is added to DPB.
Fixes: cc14724d73 ("frontends/va: Implement DPB management for H264/5 encode")
Reviewed-by: Ruijing Dong <ruijing.dong@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/31741>
