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

nir/algebraic: optimize open-coded uadd_sat/usub_sat

Browse source 
fossil-db (navi21):
Totals from 19 (0.01% of 135636) affected shaders:
Instrs: 40730 -> 40688 (-0.10%)
CodeSize: 217708 -> 217568 (-0.06%)
Latency: 261466 -> 261373 (-0.04%)
InvThroughput: 74944 -> 74896 (-0.06%)

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Georg Lehmann <dadschoorse@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19473>
This commit is contained in:
Rhys Perry 2022-11-02 16:45:41 +00:00 • committed by Marge Bot
parent 6fc4a76057
commit e19584db2b
1 changed files with 26 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

26
src/compiler/nir/nir_opt_algebraic.py
View file

@ -2079,6 +2079,32 @@ optimizations.extend([
(('imul24', a, 0), (0)),
])
for bit_size in [8, 16, 32, 64]:
cond = '!options->lower_uadd_sat'
if bit_size == 64:
cond += ' && !(options->lower_int64_options & nir_lower_iadd64)'
add = 'iadd@' + str(bit_size)
optimizations += [
(('bcsel', ('ult', ('iadd', a, b), a), -1, (add, a, b)), ('uadd_sat', a, b), cond),
(('bcsel', ('uge', ('iadd', a, b), a), (add, a, b), -1), ('uadd_sat', a, b), cond),
(('bcsel', ('ieq', ('uadd_carry', a, b), 0), (add, a, b), -1), ('uadd_sat', a, b), cond),
(('bcsel', ('ine', ('uadd_carry', a, b), 0), -1, (add, a, b)), ('uadd_sat', a, b), cond),
]
for bit_size in [8, 16, 32, 64]:
cond = '!options->lower_usub_sat'
if bit_size == 64:
cond += ' && !(options->lower_int64_options & nir_lower_usub_sat64)'
add = 'iadd@' + str(bit_size)
optimizations += [
(('bcsel', ('ult', a, b), 0, (add, a, ('ineg', b))), ('usub_sat', a, b), cond),
(('bcsel', ('uge', a, b), (add, a, ('ineg', b)), 0), ('usub_sat', a, b), cond),
(('bcsel', ('ieq', ('usub_borrow', a, b), 0), (add, a, ('ineg', b)), 0), ('usub_sat', a, b), cond),
(('bcsel', ('ine', ('usub_borrow', a, b), 0), 0, (add, a, ('ineg', b))), ('usub_sat', a, b), cond),
]
# bit_size dependent lowerings
for bit_size in [8, 16, 32, 64]:
# convenience constants