fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-08 04:48:08 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

nir/algebraic: Optimize away re-extracts of packs.

Browse source 
Turns this v2i16 OpSDotKHR code on ir3:

MESA: info: isam.v.base0 (u16)(xy)hr1.x, r0.x, s#0, t#0
MESA: info: isam.v.base0 (u16)(xy)hr1.z, r0.x, s#0, t#2
MESA: info: (sy)(rpt1)cov.u16u32 r48.w, (last)(r)hr1.x
MESA: info: (rpt1)cov.u16u32 r49.y, (last)(r)hr1.z
MESA: info: (ss)shlg r49.x, 16, r49.x, r48.w
MESA: info: shlg r49.z, 16, r49.z, r49.y
MESA: info: shl.b r49.w, r49.x, 16
MESA: info: shl.b r50.x, r49.z, 16
MESA: info: ashr.b r49.w, r49.w, 16
MESA: info: ashr.b r49.x, r49.x, 16
MESA: info: ashr.b r50.x, r50.x, 16
MESA: info: ashr.b r49.z, r49.z, 16
MESA: info: mul.u24 r49.w, r49.w, r50.x
MESA: info: mul.s24 r49.x, r49.x, r49.z
MESA: info: add.u r49.w, r49.w, r49.x

into:

MESA: info: isam.v.base0 (u16)(xy)hr1.x, r0.x, s#0, t#0
MESA: info: isam.v.base0 (u16)(xy)hr1.z, r0.x, s#0, t#2
MESA: info: (sy)(rpt1)cov.s16s32 r48.w, (last)(r)hr1.x
MESA: info: (rpt1)cov.s16s32 r49.y, (last)(r)hr1.z
MESA: info: (ss)mul.s24 r48.w, r48.w, r49.y
MESA: info: mul.s24 r49.x, r49.x, r49.z
MESA: info: add.u r48.w, r48.w, r49.x

No change on shader-db.
This commit is contained in:
Emma Anholt 2026-04-29 17:16:49 -07:00
parent 1ebc14bcb9
commit 3c66bcd25b
1 changed files with 14 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

14
src/compiler/nir/nir_opt_algebraic.py
View file

@ -2018,6 +2018,20 @@ for pack, bits, compbits in [('pack_64_2x32', 64, 32), ('pack_32_2x16', 32, 16)]
((pack, (unpack, a)), a),
]
# No-op extracts from packs.
for (bits, chans, compbits) in ((32, 4, 8), (32, 2, 16), (64, 4, 16)):
for chan in range(chans):
chanvar = 'abcd'[chan]
optimizations += [((f'extract_i{compbits}', (f'pack_{bits}_{chans}x{compbits}', a), chan), ('i2i', 'a.' + 'xyzw'[chan]))]
optimizations += [((f'extract_u{compbits}', (f'pack_{bits}_{chans}x{compbits}', a), chan), ('u2u', 'a.' + 'xyzw'[chan]))]
if not (bits == 64 and compbits == 16):
if chans == 4:
optimizations += [((f'extract_i{compbits}', (f'pack_{bits}_{chans}x{compbits}_split', a, b, c, d), chan), ('i2i', chanvar))]
optimizations += [((f'extract_u{compbits}', (f'pack_{bits}_{chans}x{compbits}_split', a, b, c, d), chan), ('u2u', chanvar))]
if chans == 2:
optimizations += [((f'extract_i{compbits}', (f'pack_{bits}_{chans}x{compbits}_split', a, b), chan), ('i2i', chanvar))]
optimizations += [((f'extract_u{compbits}', (f'pack_{bits}_{chans}x{compbits}_split', a, b), chan), ('u2u', chanvar))]
optimizations.extend([
(('unpack_64_2x32_split_y', ('u2u64', 'a@1')), 0),
(('unpack_64_2x32_split_y', ('u2u64', 'a@8')), 0),