NeroReflex/mesa
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

gallium/u_simple_shaders: Optimize out ffloors

Browse source 
ffloor(f2i(x)) can't be optimized to f2i(x) due to differing behaviour for
negative x, but u_blitter only uses this with nonnegative x so we can instead
use ftrunc(f2i(x)) which NIR will optimize to f2i(x) for us. This gets rid of
the silly ffloor instructions in blit shaders.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24319>
This commit is contained in:
Alyssa Rosenzweig 2023-07-29 18:49:41 -04:00 • committed by Marge Bot
parent ff9423b693
commit 18b2daa136
1 changed files with 4 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

6
src/gallium/auxiliary/util/u_simple_shaders.c
View file

@ -211,7 +211,9 @@ ureg_load_tex(struct ureg_program *ureg, struct ureg_dst out,
/* Nearest filtering floors and then converts to integer, and then
* applies clamp to edge as clamp(coord, 0, dim - 1).
* u_blitter only uses this when the coordinates are in bounds,
* so no clamping is needed.
* so no clamping is needed and we can use trunc instead of floor. trunc
* with f2i will get optimized out in NIR where f2i has round-to-zero
* behaviour already.
*/
unsigned wrmask = tex_target == TGSI_TEXTURE_1D ||
tex_target == TGSI_TEXTURE_1D_ARRAY ? TGSI_WRITEMASK_X :
@ -219,7 +221,7 @@ ureg_load_tex(struct ureg_program *ureg, struct ureg_dst out,
TGSI_WRITEMASK_XY;
ureg_MOV(ureg, temp, coord);
ureg_FLR(ureg, ureg_writemask(temp, wrmask), ureg_src(temp));
ureg_TRUNC(ureg, ureg_writemask(temp, wrmask), ureg_src(temp));
ureg_F2I(ureg, temp, ureg_src(temp));
if (load_level_zero)