dsg/mesa
2
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

radv: Use wave32 for RT on gfx11+

Browse source 
ACO got a lot better at forming VOPD instructions, and testing
feedback seems to point in a slightly positive direction for this.

gfx12 will also start requiring wave32 for dynamic VGPR allocation at
some point.

Measurements on navi31:

Cyberpunk 2077:
Difference at 95.0% confidence
        1.12333 +/- 0.42876
        1.88216% +/- 0.718391%
        (Student's t, pooled s = 0.189165)

Black Myth Wukong benchmark:
Difference at 95.0% confidence
        4 +/- 1.30862
        13.9535% +/- 4.56495%
        (Student's t, pooled s = 0.57735)

Portal with RTX:
66.2ms->61.5ms (~7.64% improvement)

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/39275>
This commit is contained in:
Natalie Vock 2026-01-12 21:14:29 +01:00 • committed by Marge Bot
parent fcf53988c4
commit 6f076cdfda
4 changed files with 10 additions and 17 deletions
Download patch file Download diff file Expand all files Collapse all files

4
docs/envvars.rst
View file

@ -1577,10 +1577,8 @@ RADV driver environment variables
disable optimizations that get enabled when all VRAM is CPU visible.
``pswave32``
enable wave32 for pixel shaders (GFX10+)
``rtwave32``
enable wave32 for ray tracing shaders (GFX11+)
``rtwave64``
enable wave64 for ray tracing shaders (GFX10-10.3)
enable wave64 for ray tracing shaders (GFX10+)
``sam``
enable optimizations to move more driver internal objects to VRAM.
``sparse``

11
src/amd/vulkan/radv_debug.h
View file

@ -95,12 +95,11 @@ enum {
RADV_PERFTEST_DMA_SHADERS = 1u << 11,
RADV_PERFTEST_TRANSFER_QUEUE = 1u << 12,
RADV_PERFTEST_NIR_CACHE = 1u << 13,
RADV_PERFTEST_RT_WAVE_32 = 1u << 14,
RADV_PERFTEST_VIDEO_ENCODE = 1u << 15,
RADV_PERFTEST_NO_GTT_SPILL = 1u << 16,
RADV_PERFTEST_HIC = 1u << 17,
RADV_PERFTEST_SPARSE = 1u << 18,
RADV_PERFTEST_RT_CPS = 1u << 19,
RADV_PERFTEST_VIDEO_ENCODE = 1u << 14,
RADV_PERFTEST_NO_GTT_SPILL = 1u << 15,
RADV_PERFTEST_HIC = 1u << 16,
RADV_PERFTEST_SPARSE = 1u << 17,
RADV_PERFTEST_RT_CPS = 1u << 18,
};
enum {

1
src/amd/vulkan/radv_instance.c
View file

@ -120,7 +120,6 @@ static const struct debug_control radv_perftest_options[] = {
{"dmashaders", RADV_PERFTEST_DMA_SHADERS},
{"transfer_queue", RADV_PERFTEST_TRANSFER_QUEUE},
{"nircache", RADV_PERFTEST_NIR_CACHE},
{"rtwave32", RADV_PERFTEST_RT_WAVE_32},
{"video_encode", RADV_PERFTEST_VIDEO_ENCODE},
{"nogttspill", RADV_PERFTEST_NO_GTT_SPILL},
{"hic", RADV_PERFTEST_HIC},

11
src/amd/vulkan/radv_physical_device.c
View file

@ -2489,16 +2489,13 @@ radv_physical_device_try_create(struct radv_instance *instance, drmDevicePtr drm
if (instance->perftest_flags & RADV_PERFTEST_GE_WAVE_32)
pdev->ge_wave_size = 32;
/* Default to 32 on RDNA1-2 as that gives better perf due to less issues with divergence.
* However, on RDNA3+ default to wave64 as implicit dual issuing is likely better than
* wave32 VOPD for VALU dependent code.
* (as well as the SALU count becoming more problematic with wave32)
/* Default to 32 on RDNA as that gives better perf due to less issues with divergence.
* On GFX12+, wave32 will also be required for a future dynamic VGPR allocation implementation.
*/
if (instance->perftest_flags & RADV_PERFTEST_RT_WAVE_32 || pdev->info.gfx_level < GFX11)
pdev->rt_wave_size = 32;
if (radv_is_rt_wave64_enabled(instance))
pdev->rt_wave_size = 64;
else
pdev->rt_wave_size = 32;
}
radv_probe_video_decode(pdev);