Daniel Schürmann
6b4b044739
nir/opt_loop: add loop peeling optimization
...
This optimization turns:
loop {
do_work_1();
if (cond) {
break;
} else {
}
do_work_2();
}
into:
do_work_1();
if (cond) {
} else {
loop {
do_work_2();
do_work_1();
if (cond) {
break;
} else {
}
}
}
RADV GFX11:
Totals from 925 (1.17% of 79395) affected shaders:
MaxWaves: 20583 -> 20455 (-0.62%)
Instrs: 5260489 -> 5361418 (+1.92%); split: -0.63%, +2.55%
CodeSize: 26965388 -> 27501104 (+1.99%); split: -0.48%, +2.47%
VGPRs: 70304 -> 70712 (+0.58%)
SpillSGPRs: 2163 -> 2159 (-0.18%)
Scratch: 51200 -> 69632 (+36.00%)
Latency: 36404844 -> 34542213 (-5.12%); split: -5.51%, +0.39%
InvThroughput: 6628474 -> 6384249 (-3.68%); split: -4.19%, +0.50%
VClause: 124997 -> 127008 (+1.61%); split: -0.43%, +2.04%
SClause: 121774 -> 120799 (-0.80%); split: -3.21%, +2.40%
Copies: 357048 -> 360850 (+1.06%); split: -0.62%, +1.68%
Branches: 171985 -> 168082 (-2.27%); split: -3.61%, +1.34%
PreSGPRs: 59812 -> 60088 (+0.46%); split: -0.20%, +0.66%
PreVGPRs: 60325 -> 60586 (+0.43%); split: -0.29%, +0.72%
VALU: 2882263 -> 2951373 (+2.40%); split: -0.37%, +2.77%
SALU: 636373 -> 640091 (+0.58%); split: -0.87%, +1.46%
VMEM: 200059 -> 204612 (+2.28%); split: -0.09%, +2.36%
SMEM: 173328 -> 174343 (+0.59%); split: -2.34%, +2.92%
VOPD: 1064 -> 898 (-15.60%); split: +0.09%, -15.70%
Reviewed-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/28150 >
2024-05-03 13:01:29 +00:00
Daniel Schürmann
bf43af984a
nir/opt_loop_cf: generalize removal of "trivial" continues
...
So that is also handles break statements and works in
arbitrarily nested control flow.
Totals from 905 (1.18% of 76636) affected shaders: (RADV, GFX11)
Instrs: 605164 -> 605548 (+0.06%); split: -0.01%, +0.08%
CodeSize: 3162036 -> 3163472 (+0.05%); split: -0.01%, +0.06%
Latency: 2045559 -> 1387622 (-32.16%)
InvThroughput: 352344 -> 231676 (-34.25%)
SClause: 16092 -> 16088 (-0.02%); split: -0.04%, +0.02%
Copies: 41286 -> 41297 (+0.03%); split: -0.02%, +0.05%
Branches: 19949 -> 19929 (-0.10%)
PreSGPRs: 33413 -> 33385 (-0.08%)
PreVGPRs: 19177 -> 19135 (-0.22%)
Reviewed-by: Konstantin Seurer <konstantin.seurer@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24940 >
2024-01-03 20:48:05 +00:00
Daniel Schürmann
9808ef0349
nir/opt_loop: move loop control-flow optimizations into separate pass
...
This new pass aims to simplify loop control-flow by reducing the number
of break and continue statements. It also supersedes nir_opt_trivial_continues().
For this purpose, it implements 3 optimizations:
- opt_loop_terminator(), as previously
- opt_loop_merge_break_continue(), similar to opt_merge_breaks() incl. continues
- opt_loop_last_block(), a generalization of opt_if_loop_last_continue()
Reviewed-by: Konstantin Seurer <konstantin.seurer@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24940 >
2024-01-03 20:48:04 +00:00
