fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-01-19 17:40:32 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

ir3/shared_ra: don't reuse src of different halfness

Browse source 
Avoid reusing the src of an ALU/SFU instruction when its halfness is
different from the dst. A difference in halfness would introduce a (ss)
sync and unnecessarily increase ss-stall.

Totals from 8261 (5.02% of 164705) affected shaders:
Instrs: 10044160 -> 10044390 (+0.00%); split: -0.04%, +0.04%
CodeSize: 19875094 -> 19879238 (+0.02%); split: -0.08%, +0.10%
NOPs: 2249893 -> 2249273 (-0.03%); split: -0.14%, +0.11%
MOVs: 426644 -> 426565 (-0.02%); split: -0.24%, +0.23%
COVs: 134819 -> 134810 (-0.01%); split: -0.06%, +0.05%
(ss): 264012 -> 260680 (-1.26%); split: -1.34%, +0.08%
(sy): 122711 -> 122851 (+0.11%); split: -0.07%, +0.18%
(ss)-stall: 1111161 -> 1100625 (-0.95%); split: -1.07%, +0.12%
(sy)-stall: 3650422 -> 3651422 (+0.03%); split: -0.15%, +0.17%
STPs: 8693 -> 8701 (+0.09%); split: -0.08%, +0.17%
LDPs: 16814 -> 16815 (+0.01%); split: -0.10%, +0.11%
Preamble Instrs: 2346201 -> 2351217 (+0.21%); split: -0.25%, +0.46%
Last helper: 3417842 -> 3417889 (+0.00%); split: -0.07%, +0.08%
Cat0: 2486420 -> 2485758 (-0.03%); split: -0.13%, +0.10%
Cat1: 613403 -> 614417 (+0.17%); split: -0.21%, +0.37%
Cat2: 3742181 -> 3742109 (-0.00%); split: -0.00%, +0.00%
Cat6: 81897 -> 81906 (+0.01%); split: -0.03%, +0.04%
Cat7: 254014 -> 253955 (-0.02%); split: -0.11%, +0.09%

Note that the slight increase in Cat1 is mostly from lowered copies (to
swz) and seems to be RA bad luck.

Signed-off-by: Job Noorman <jnoorman@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/36376>
This commit is contained in:
Job Noorman 2025-07-25 15:09:27 +02:00 • committed by Marge Bot
parent b12db991eb
commit 04a182417e
1 changed files with 6 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

6
src/freedreno/ir3/ir3_shared_ra.c
View file

@ -693,6 +693,12 @@ get_reg(struct ra_ctx *ctx, struct ir3_register *reg, bool src)
struct ir3_register *src = reg->instr->srcs[i];
if (!ra_reg_is_src(src))
continue;
/* When src and dst are overlapping registers with different halfness,
* a (ss) sync is necessary. Avoid this to not unnecessarily increase
* ss-stall.
*/
if ((reg->flags & IR3_REG_HALF) != (src->flags & IR3_REG_HALF))
continue;
if ((src->flags & IR3_REG_SHARED) && reg_size(src) >= size) {
struct ra_interval *src_interval = ra_interval_get(ctx, src->def);
physreg_t src_physreg = ra_interval_get_physreg(src_interval);