fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2025-12-24 19:40:10 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

r300: reorder for easier presubtract 1-x pattern recognition

Browse source 
It is much easier to just add a simple late algebraic pass than actually
trying to teach the backend to recognize all the different patterns.

RV530 shader-db:
total instructions in shared programs: 129643 -> 129468 (-0.13%)
instructions in affected programs: 17665 -> 17490 (-0.99%)
helped: 176
HURT: 39
total presub in shared programs: 4912 -> 5411 (10.16%)
presub in affected programs: 1651 -> 2150 (30.22%)
helped: 0
HURT: 287
total temps in shared programs: 16904 -> 16918 (0.08%)
temps in affected programs: 812 -> 826 (1.72%)
helped: 25
HURT: 37
total cycles in shared programs: 194771 -> 194675 (-0.05%)
cycles in affected programs: 28096 -> 28000 (-0.34%)
helped: 146
HURT: 41

Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/9364
Reviewed-by: Filip Gawin <filip.gawin@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24830>
This commit is contained in:
Pavel Ondračka 2023-07-17 13:17:04 +02:00 • committed by Marge Bot
parent 939845e424
commit 24d35a56d9
3 changed files with 32 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

6
src/gallium/drivers/r300/compiler/nir_to_rc.c
View file

@ -26,6 +26,7 @@
#include "compiler/nir/nir_legacy.h"
#include "compiler/nir/nir_worklist.h"
#include "nir_to_rc.h"
#include "r300_nir.h"
#include "pipe/p_screen.h"
#include "pipe/p_state.h"
#include "tgsi/tgsi_dump.h"
@ -2433,6 +2434,11 @@ const void *nir_to_rc_options(struct nir_shader *s,
}
} while (progress);
if (s->info.stage == MESA_SHADER_FRAGMENT) {
NIR_PASS_V(s, r300_nir_prepare_presubtract);
NIR_PASS_V(s, r300_nir_clean_double_fneg);
}
NIR_PASS_V(s, nir_lower_int_to_float);
NIR_PASS_V(s, nir_lower_bool_to_float,
!options->lower_cmp && !options->lower_fabs);

4
src/gallium/drivers/r300/compiler/r300_nir.h
View file

@ -36,4 +36,8 @@ extern bool r300_nir_fuse_fround_d3d9(struct nir_shader *shader);
extern bool r300_nir_lower_bool_to_float(struct nir_shader *shader);
extern bool r300_nir_prepare_presubtract(struct nir_shader *shader);
extern bool r300_nir_clean_double_fneg(struct nir_shader *shader);
#endif /* R300_NIR_H */

22
src/gallium/drivers/r300/compiler/r300_nir_algebraic.py
View file

@ -63,6 +63,22 @@ r300_nir_fuse_fround_d3d9 = [
('fround_even', 'a'))
]
# Here are some specific optimizations for code reordering such that the backend
# has easier task of recognizing output modifiers and presubtract patterns.
r300_nir_prepare_presubtract = [
# Backend can only recognize 1 - x pattern.
(('fadd', ('fneg', a), 1.0), ('fadd', 1.0, ('fneg', a))),
(('fadd', a, -1.0), ('fneg', ('fadd', 1.0, ('fneg', a)))),
(('fadd', -1.0, a), ('fneg', ('fadd', 1.0, ('fneg', a)))),
]
# Previous prepare_presubtract pass can sometimes produce double fneg patterns.
# The backend copy propagate could handle it, but the nir to tgsi translation
# does not and blows up. Just run a simple pass to clean it up.
r300_nir_clean_double_fneg = [
(('fneg', ('fneg', a)), a)
]
def main():
parser = argparse.ArgumentParser()
parser.add_argument('-p', '--import-path', required=True)
@ -103,5 +119,11 @@ def main():
f.write(nir_algebraic.AlgebraicPass("r300_nir_lower_bool_to_float",
r300_nir_lower_bool_to_float).render())
f.write(nir_algebraic.AlgebraicPass("r300_nir_prepare_presubtract",
r300_nir_prepare_presubtract).render())
f.write(nir_algebraic.AlgebraicPass("r300_nir_clean_double_fneg",
r300_nir_clean_double_fneg).render())
if __name__ == '__main__':
main()