fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-01-06 17:30:20 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 2

agx: optimize b2x(inot)

Browse source 
easier to do as algebraic rules.

total instructions in shared programs: 2915611 -> 2910477 (-0.18%)
instructions in affected programs: 285251 -> 280117 (-1.80%)
helped: 1245
HURT: 5
Instructions are helped.

total bytes in shared programs: 19439752 -> 19400010 (-0.20%)
bytes in affected programs: 1881556 -> 1841814 (-2.11%)
helped: 1244
HURT: 6
Bytes are helped.

total halfregs in shared programs: 519253 -> 519553 (0.06%)
halfregs in affected programs: 5509 -> 5809 (5.45%)
helped: 24
HURT: 107
Halfregs are HURT.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27616>
This commit is contained in:
Alyssa Rosenzweig 2024-01-19 18:02:00 -04:00 • committed by Marge Bot
parent 34ca925064
commit b45d54a388
2 changed files with 13 additions and 19 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

18
src/asahi/compiler/agx_compile.c
View file

@ -1576,24 +1576,6 @@ agx_emit_alu(agx_builder *b, nir_alu_instr *instr)
case nir_op_bcsel:
return agx_icmpsel_to(b, dst, s0, i0, s2, s1, AGX_ICOND_UEQ);
case nir_op_b2i32:
case nir_op_b2i16:
case nir_op_b2i8:
return agx_icmpsel_to(b, dst, s0, i0, i0, i1, AGX_ICOND_UEQ);
case nir_op_b2b32:
return agx_icmpsel_to(b, dst, s0, i0, i0, agx_mov_imm(b, 32, 0xFFFFFFFF),
AGX_ICOND_UEQ);
case nir_op_b2f16:
case nir_op_b2f32: {
/* At this point, boolean is just zero/nonzero, so compare with zero */
agx_index f1 = (sz == 16) ? agx_mov_imm(b, 16, _mesa_float_to_half(1.0))
: agx_mov_imm(b, 32, fui(1.0));
return agx_fcmpsel_to(b, dst, s0, i0, i0, f1, AGX_FCOND_EQ);
}
case nir_op_i2i32: {
if (src_sz == 8) {
/* Sign extend in software, NIR likes 8-bit conversions */

14
src/asahi/compiler/agx_nir_algebraic.py
View file

@ -83,6 +83,17 @@ lower_pack = [
('isub', 32, 'bits'))),
]
lower_selects = []
for T, sizes, one in [('f', [16, 32], 1.0),
('i', [8, 16, 32], 1),
('b', [32], -1)]:
for size in sizes:
lower_selects.extend([
((f'b2{T}{size}', ('inot', 'a@1')), ('bcsel', a, 0, one)),
((f'b2{T}{size}', 'a@1'), ('bcsel', a, one, 0)),
])
fuse_extr = []
for start in range(32):
fuse_extr.extend([
@ -170,7 +181,8 @@ def run():
print('#include "agx_nir.h"')
print(nir_algebraic.AlgebraicPass("agx_nir_lower_algebraic_late",
lower_sm5_shift + lower_pack).render())
lower_sm5_shift + lower_pack +
lower_selects).render())
print(nir_algebraic.AlgebraicPass("agx_nir_fuse_algebraic_late",
fuse_extr + fuse_ubfe + fuse_imad).render())
print(nir_algebraic.AlgebraicPass("agx_nir_opt_ixor_bcsel",