This will create code that is easier to combine into MADs/FMA when the
last component of the vector is 1.0.
nir_opt_algebraic_late has an optimization to do something similar but it
only works for inexact code, if the multiplication-by-1 optimization is
done before it and if the backend enables fuse_ffma.
fossil-db (Navi):
Totals from 4296 (3.75% of 114665) affected shaders:
SGPRs: 283468 -> 283764 (+0.10%); split: -0.02%, +0.12%
VGPRs: 172868 -> 172904 (+0.02%); split: -0.09%, +0.11%
CodeSize:
14045312 ->
14027128 (-0.13%); split: -0.15%, +0.02%
MaxWaves: 59285 -> 59282 (-0.01%); split: +0.04%, -0.05%
Instrs: 2703507 -> 2683187 (-0.75%); split: -0.76%, +0.00%
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5631>
for (unsigned i = 0; i < src1_columns; i++) {
/* dest[i] = sum(src0[j] * src1[i][j] for all j) */
dest->elems[i]->def =
- nir_fmul(&b->nb, src0->elems[0]->def,
- nir_channel(&b->nb, src1->elems[i]->def, 0));
- for (unsigned j = 1; j < src0_columns; j++) {
+ nir_fmul(&b->nb, src0->elems[src0_columns - 1]->def,
+ nir_channel(&b->nb, src1->elems[i]->def, src0_columns - 1));
+ for (int j = src0_columns - 2; j >= 0; j--) {
dest->elems[i]->def =
nir_fadd(&b->nb, dest->elems[i]->def,
nir_fmul(&b->nb, src0->elems[j]->def,