fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-01-06 08:50:09 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

radv/nir/lower_cmat: tightly pack 8bit gfx11 acc matrix
Some checks are pending
macOS-CI / macOS-CI (dri) (push) Waiting to run
Details
macOS-CI / macOS-CI (xlib) (push) Waiting to run
Details

Browse source 
Invalid for now, but used by vkd3d-proton, where the use case is to convert
a result matrix to lower precision, followed by a store.

For 16bit accumulation matrices, GFX11 only uses 16bits per 32bit register.
RADV's coop matrix code pads the unused space with undefs and uses a vector
with twice as many elements as the matrix length. Extending that to 8bit by
leaving 24 bits unused is unnecessary as these matrices as there
is no hw unit that requires it. And in wave32, it would also result in
vectors larger than NIR's limit.
So tightly pack 8bit matrices without any undef padding.

Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34382>
This commit is contained in:
Georg Lehmann 2025-04-04 13:53:54 +02:00 • committed by Marge Bot
parent bbc9bc9d24
commit 6d2190300a
1 changed files with 7 additions and 7 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

14
src/amd/vulkan/nir/radv_nir_lower_cooperative_matrix.c
View file

@ -72,25 +72,25 @@ radv_nir_cmat_length(struct glsl_cmat_description desc, const lower_cmat_params
if (params->gfx_level >= GFX12) {
assert(desc.cols == 16 && desc.rows == 16);
return 256 / params->wave_size;
} else if (desc.use != GLSL_CMAT_USE_ACCUMULATOR) {
return 16;
} else {
return desc.use != GLSL_CMAT_USE_ACCUMULATOR
? 16
: (desc.cols * desc.rows / params->wave_size * 32 / radv_nir_cmat_bits(desc));
return desc.cols * desc.rows / params->wave_size * (radv_nir_cmat_bits(desc) == 16 ? 2 : 1);
}
}
static unsigned
radv_nir_cmat_length_mul(struct glsl_cmat_description desc, const lower_cmat_params *params)
{
if (params->gfx_level >= GFX12) {
if (params->gfx_level >= GFX12 || desc.use != GLSL_CMAT_USE_ACCUMULATOR) {
return 1;
} else {
/* For C matrices we have 1 VGPR per element even if the element type is
* < 32 bits. So with 8 fp16 elements we implement that with a f16vec16.
/* For GFX11 C matrices we have 1 VGPR per element even if the element type is
* 16bits. So with 8 fp16 elements we implement that with a f16vec16.
* We then use the coefficient generated by this function to figure out
* how many elements we really have.
*/
return desc.use == GLSL_CMAT_USE_ACCUMULATOR ? (32 / radv_nir_cmat_bits(desc)) : 1;
return radv_nir_cmat_bits(desc) == 16 ? 2 : 1;
}
}