fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-24 04:08:10 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

radeonsi: advertize new float multadd options

Browse source 
ac already handles advertizing fmad and ffma support, so simply decide
when nir should fuse, because radeonsi contrary to radv doesn't let aco do
all the fusing itself.

Also unset splitting for force_use_fma32 handling.

Reviewed-by: Georg Lehmann <dadschoorse@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/41165>
This commit is contained in:
Karol Herbst 2026-04-24 15:35:07 +02:00 • committed by Marge Bot
parent 6cf3cba68c
commit eec83fd9f6
1 changed files with 14 additions and 6 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

20
src/gallium/drivers/radeonsi/gfx/si_gfx_screen.c
View file

@ -316,15 +316,10 @@ static void si_init_screen_nir_options(struct si_screen *sscreen)
* gfx9 and newer prefer FMA for F16 because of the packed instruction.
* gfx10 and older prefer MAD for F32 because of the legacy instruction.
*/
bool use_fma32 =
sscreen->info.gfx_level >= GFX10_3 ||
(sscreen->info.family >= CHIP_GFX940 && !sscreen->info.has_graphics) ||
/* fma32 is too slow for gpu < gfx9, so apply the option only for gpu >= gfx9 */
(sscreen->info.gfx_level >= GFX9 && sscreen->options.force_use_fma32);
nir_shader_compiler_options *options = sscreen->nir_options;
/* GFX8 has precision issues with 16-bit PS outputs. */
bool has_16bit_io = sscreen->info.gfx_level >= GFX9;
nir_shader_compiler_options *options = sscreen->nir_options;
ac_nir_set_options(&sscreen->info.compiler_info, !sscreen->use_aco, options);
options->ignore_none_interpolation_in_sysval_gathering = true;
@ -334,6 +329,19 @@ static void si_init_screen_nir_options(struct si_screen *sscreen)
options->fuse_ffma16 = sscreen->info.gfx_level >= GFX9;
options->fuse_ffma32 = use_fma32;
options->fuse_ffma64 = true;
bool use_fma32 = !(options->float_mul_add32 & nir_float_muladd_support_prefers_split) ||
(sscreen->info.gfx_level >= GFX9 && sscreen->options.force_use_fma32);
if (sscreen->info.gfx_level >= GFX9)
options->float_mul_add16 |= nir_float_muladd_support_fuse;
if (use_fma32) {
/* for force_use_fma32 handling */
options->float_mul_add32 &= ~nir_float_muladd_support_prefers_split;
options->float_mul_add32 |= nir_float_muladd_support_fuse;
}
options->float_mul_add64 |= nir_float_muladd_support_fuse;
options->lower_uniforms_to_ubo = true;
options->lower_to_scalar = true;
options->lower_to_scalar_filter =