fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-04-30 23:28:06 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 5

nir/opt_algebraic: create 16bit fmin/fmax if only used by pack_half_2x16_rtz_split

Browse source 
Foz-DB Navi21:
Totals from 1842 (2.30% of 80066) affected shaders:
Instrs: 869152 -> 866751 (-0.28%)
CodeSize: 4687316 -> 4682496 (-0.10%); split: -0.14%, +0.03%
VGPRs: 75216 -> 75312 (+0.13%)
Latency: 7297749 -> 7297929 (+0.00%); split: -0.01%, +0.02%
InvThroughput: 1864933 -> 1860706 (-0.23%); split: -0.23%, +0.00%
Copies: 52679 -> 52463 (-0.41%)
VALU: 665076 -> 662890 (-0.33%)
SALU: 56226 -> 56010 (-0.38%)

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/36535>
This commit is contained in:
Georg Lehmann 2025-08-01 20:24:55 +02:00 • committed by Marge Bot
parent 22afe83473
commit 8512479097
1 changed files with 13 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

13
src/compiler/nir/nir_opt_algebraic.py
View file

@ -3015,6 +3015,19 @@ optimizations += [
('fmax', ('f2f16', ('vec2', a, c)), ('f2f16', ('vec2', b, d))), 'options->support_16bit_alu'),
(('f2f16', ('vec2(is_used_once)', ('fmin(is_used_once)', a, '#b'), ('fmin(is_used_once)', c, '#d'))),
('fmin', ('f2f16', ('vec2', a, c)), ('f2f16', ('vec2', b, d))), 'options->support_16bit_alu'),
(('pack_half_2x16_rtz_split', ('fmax(is_used_once)', a, '#b'), ('fmax(is_used_once)', c, '#d')),
('pack_32_2x16', ('fmax', ('unpack_32_2x16', ('pack_half_2x16_rtz_split', a, c)), ('unpack_32_2x16', ('pack_half_2x16_rtz_split', b, d)))),
'options->vectorize_vec2_16bit'),
(('pack_half_2x16_rtz_split', ('fmin(is_used_once)', a, '#b'), ('fmin(is_used_once)', c, '#d')),
('pack_32_2x16', ('fmin', ('unpack_32_2x16', ('pack_half_2x16_rtz_split', a, c)), ('unpack_32_2x16', ('pack_half_2x16_rtz_split', b, d)))),
'options->vectorize_vec2_16bit'),
(('pack_half_2x16_rtz_split', ('fneg(is_used_once)', ('fmax(is_used_once)', a, '#b')), ('fneg(is_used_once)', ('fmax(is_used_once)', c, '#d'))),
('pack_32_2x16', ('fmin', ('fneg', ('unpack_32_2x16', ('pack_half_2x16_rtz_split', a, c))), ('fneg', ('unpack_32_2x16', ('pack_half_2x16_rtz_split', b, d))))),
'options->vectorize_vec2_16bit'),
(('pack_half_2x16_rtz_split', ('fneg(is_used_once)', ('fmin(is_used_once)', a, '#b')), ('fneg(is_used_once)', ('fmin(is_used_once)', c, '#d'))),
('pack_32_2x16', ('fmax', ('fneg', ('unpack_32_2x16', ('pack_half_2x16_rtz_split', a, c))), ('fneg', ('unpack_32_2x16', ('pack_half_2x16_rtz_split', b, d))))),
'options->vectorize_vec2_16bit'),
]
# Some operations such as iadd have the property that the bottom N bits of the