fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-07 04:58:05 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 4

intel/fs: add more UNDEFs around SEND messages

Browse source 
lower_find_live_channel() in particular is used a lot in control flow
to find the live channel for the surface/sampler handle. Adding UNDEFs
on the temporary registers used for finding the live channels helps
reduce the liveness of those temporary registers, especially in loops.

Some titles affected :

Rise Of The Tomb Raider:
Totals from 2780 (22.58% of 12311) affected shaders:
Instrs: 1294455 -> 1294592 (+0.01%); split: -0.15%, +0.16%
Cycles: 1473136441 -> 1471302617 (-0.12%); split: -1.52%, +1.40%
Max live registers: 144282 -> 143595 (-0.48%)
Max dispatch width: 22200 -> 22232 (+0.14%)

Red Dead Redemption 2:
Totals from 435 (7.28% of 5972) affected shaders:
Instrs: 488472 -> 487594 (-0.18%); split: -0.31%, +0.14%
Cycles: 11354732 -> 11384928 (+0.27%); split: -0.44%, +0.71%
Spill count: 1217 -> 1172 (-3.70%)
Fill count: 3521 -> 3447 (-2.10%)
Scratch Memory Size: 64512 -> 62464 (-3.17%)
Max live registers: 35997 -> 35798 (-0.55%)

Fallout 4:
Totals from 8 (0.49% of 1638) affected shaders:
Instrs: 41908 -> 40509 (-3.34%)
Cycles: 3638464 -> 3555680 (-2.28%); split: -2.67%, +0.39%
Spill count: 717 -> 665 (-7.25%)
Fill count: 2542 -> 2438 (-4.09%)
Scratch Memory Size: 32768 -> 16384 (-50.00%)
Max live registers: 567 -> 534 (-5.82%)

Cyberpunk 2077:
Totals from 2984 (28.97% of 10301) affected shaders:
Instrs: 3888874 -> 3891600 (+0.07%); split: -0.20%, +0.27%
Cycles: 67906489 -> 67767721 (-0.20%); split: -0.68%, +0.47%
Spill count: 200 -> 98 (-51.00%)
Fill count: 237 -> 90 (-62.03%)
Scratch Memory Size: 10240 -> 8192 (-20.00%)
Max live registers: 215715 -> 212727 (-1.39%)

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24282>
This commit is contained in:
Lionel Landwerlin 2023-07-25 10:09:18 +03:00 • committed by Marge Bot
parent 5c72724819
commit d099e47de0
2 changed files with 3 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
src/intel/compiler/brw_fs.cpp
View file

@ -5826,6 +5826,7 @@ fs_visitor::lower_find_live_channel()
*/
if (!(first && packed_dispatch)) {
fs_reg mask = ubld.vgrf(BRW_REGISTER_TYPE_UD);
ubld.UNDEF(mask);
ubld.emit(SHADER_OPCODE_READ_SR_REG, mask, brw_imm_ud(vmask ? 3 : 2));
/* Quarter control has the effect of magically shifting the value of
@ -5843,6 +5844,7 @@ fs_visitor::lower_find_live_channel()
ubld.FBL(inst->dst, exec_mask);
} else {
fs_reg tmp = ubld.vgrf(BRW_REGISTER_TYPE_UD, 1);
ubld.UNDEF(tmp);
ubld.LZD(tmp, exec_mask);
ubld.ADD(inst->dst, negate(tmp), brw_imm_uw(31));
}

1
src/intel/compiler/brw_lower_logical_sends.cpp
View file

@ -1344,6 +1344,7 @@ emit_predicate_on_vector_mask(const fs_builder &bld, fs_inst *inst)
const fs_visitor *v = static_cast<const fs_visitor *>(bld.shader);
const fs_reg vector_mask = ubld.vgrf(BRW_REGISTER_TYPE_UW);
ubld.UNDEF(vector_mask);
ubld.emit(SHADER_OPCODE_READ_SR_REG, vector_mask, brw_imm_ud(3));
const unsigned subreg = sample_mask_flag_subreg(v);