fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2025-12-24 13:10:10 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

nir: Add a new has_fmulz_no_denorms flag

Browse source 
Reviewed-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26569>
This commit is contained in:
Faith Ekstrand 2023-12-07 10:10:34 -06:00 • committed by Marge Bot
parent a8b68badd5
commit aac1e3f595
2 changed files with 15 additions and 5 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

6
src/compiler/nir/nir.h
View file

@ -3900,6 +3900,12 @@ typedef struct nir_shader_compiler_options {
/** Backend supports fmulz (and ffmaz if lower_ffma32=false) */
bool has_fmulz;
/**
* Backend supports fmulz (and ffmaz if lower_ffma32=false) but only if
* FLOAT_CONTROLS_DENORM_PRESERVE_FP32 is not set
*/
bool has_fmulz_no_denorms;
/** Backend supports 32bit ufind_msb_rev and ifind_msb_rev. */
bool has_find_msb_rev;

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

@ -46,6 +46,10 @@ signed_zero_nan_preserve_32 = ('(nir_is_float_control_signed_zero_preserve(info-
signed_zero_inf_nan_preserve_16 = 'nir_is_float_control_signed_zero_inf_nan_preserve(info->float_controls_execution_mode, 16)'
signed_zero_inf_nan_preserve_32 = 'nir_is_float_control_signed_zero_inf_nan_preserve(info->float_controls_execution_mode, 32)'
has_fmulz = '(options->has_fmulz || \
(options->has_fmulz_no_denorms && \
!nir_is_denorm_preserve(info->float_controls_execution_mode, 32)))'
ignore_exact = nir_algebraic.ignore_exact
# Written in the form (<search>, <replace>) where <search> is an expression
@ -274,20 +278,20 @@ optimizations = [
# Optimize open-coded fmulz.
# (b==0.0 ? 0.0 : a) * (a==0.0 ? 0.0 : b) -> fmulz(a, b)
(('fmul@32', ('bcsel', ignore_exact('feq', b, 0.0), 0.0, a), ('bcsel', ignore_exact('feq', a, 0.0), 0.0, b)),
('fmulz', a, b), 'options->has_fmulz && !'+signed_zero_preserve_32),
('fmulz', a, b), has_fmulz+' && !'+signed_zero_preserve_32),
(('fmul@32', a, ('bcsel', ignore_exact('feq', a, 0.0), 0.0, '#b(is_not_const_zero)')),
('fmulz', a, b), 'options->has_fmulz && !'+signed_zero_preserve_32),
('fmulz', a, b), has_fmulz+' && !'+signed_zero_preserve_32),
# ffma(b==0.0 ? 0.0 : a, a==0.0 ? 0.0 : b, c) -> ffmaz(a, b, c)
(('ffma@32', ('bcsel', ignore_exact('feq', b, 0.0), 0.0, a), ('bcsel', ignore_exact('feq', a, 0.0), 0.0, b), c),
('ffmaz', a, b, c), 'options->has_fmulz && !'+signed_zero_preserve_32),
('ffmaz', a, b, c), has_fmulz+' && !'+signed_zero_preserve_32),
(('ffma@32', a, ('bcsel', ignore_exact('feq', a, 0.0), 0.0, '#b(is_not_const_zero)'), c),
('ffmaz', a, b, c), 'options->has_fmulz && !'+signed_zero_preserve_32),
('ffmaz', a, b, c), has_fmulz+' && !'+signed_zero_preserve_32),
# b == 0.0 ? 1.0 : fexp2(fmul(a, b)) -> fexp2(fmulz(a, b))
(('bcsel', ignore_exact('feq', b, 0.0), 1.0, ('fexp2', ('fmul@32', a, b))),
('fexp2', ('fmulz', a, b)),
'options->has_fmulz && !'+signed_zero_inf_nan_preserve_32),
has_fmulz+' && !'+signed_zero_inf_nan_preserve_32),
]
# Shorthand for the expansion of just the dot product part of the [iu]dp4a