diff options
| author | Rhys Perry <pendingchaos02@gmail.com> | 2020-09-01 11:55:58 +0100 |
|---|---|---|
| committer | Marge Bot <eric+marge@anholt.net> | 2021-01-13 18:54:18 +0000 |
| commit | dfe429eb414511170f3dfc960d247c4aa295f924 (patch) | |
| tree | 42b2af52228489f694c8efa8fb18a2d1bdba5728 /src/amd | |
| parent | 74748f16c969c59096a0bf4ce9f86fc92c797905 (diff) | |
nir/loop_unroll: unroll more aggressively if it can improve load scheduling
Significantly improves performance of a Control compute shader. Also seems
to increase FPS at the very start of the game by ~5% (RX 580, 1080p,
medium settings, no MSAA).
fossil-db (Sienna):
Totals from 81 (0.06% of 139391) affected shaders:
SGPRs: 3848 -> 4362 (+13.36%); split: -0.99%, +14.35%
VGPRs: 4132 -> 4648 (+12.49%)
CodeSize: 275532 -> 659188 (+139.24%)
MaxWaves: 986 -> 906 (-8.11%)
Instrs: 54422 -> 126865 (+133.11%)
Cycles: 1057240 -> 750464 (-29.02%); split: -42.61%, +13.60%
VMEM: 26507 -> 61829 (+133.26%); split: +135.56%, -2.30%
SMEM: 4748 -> 5895 (+24.16%); split: +31.47%, -7.31%
VClause: 1933 -> 6802 (+251.89%); split: -0.72%, +252.61%
SClause: 1179 -> 1810 (+53.52%); split: -3.14%, +56.66%
Branches: 1174 -> 1157 (-1.45%); split: -23.94%, +22.49%
PreVGPRs: 3219 -> 3387 (+5.22%); split: -0.96%, +6.18%
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6538>
Diffstat (limited to 'src/amd')
| -rw-r--r-- | src/amd/vulkan/radv_shader.c | 1 |
1 files changed, 1 insertions, 0 deletions
diff --git a/src/amd/vulkan/radv_shader.c b/src/amd/vulkan/radv_shader.c index 846913875dd..e25ac04c29e 100644 --- a/src/amd/vulkan/radv_shader.c +++ b/src/amd/vulkan/radv_shader.c @@ -83,6 +83,7 @@ static const struct nir_shader_compiler_options nir_options = { .has_isub = true, .use_scoped_barrier = true, .max_unroll_iterations = 32, + .max_unroll_iterations_aggressive = 128, .use_interpolated_input_intrinsics = true, .vectorize_vec2_16bit = true, /* nir_lower_int64() isn't actually called for the LLVM backend, but |