AlexIndustrial/mesa
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

r300: lower comparison ops early in NIR

Browse Source 
The HW can't do comparison opcodes in the FS, so the old way was to
lower all to ADD and CMP in the backend. However, the comparison bool
result is used either in if (where we can encode the comparions in the
aluresults calculation) or is uses in another bcsel. Therefore, we now
end with two CMPs, one from the original bcsel and one from the lowered
comparison.

This patch fixes it by doing the comparison lowering early for some nice
shader-db gains. Similarly to vertex shaders, we need some special
passes for this, because we can't nir_lower_bool_to_float too early, so
we just manually lower the most common patterns in the main opt loop and
than clean up the rest later.

Shader-db RV530:
total instructions in shared programs: 130797 -> 130400 (-0.30%)
instructions in affected programs: 34591 -> 34194 (-1.15%)
helped: 203
HURT: 133
total presub in shared programs: 8175 -> 8220 (0.55%)
presub in affected programs: 1738 -> 1783 (2.59%)
helped: 62
HURT: 53
total omod in shared programs: 414 -> 412 (-0.48%)
omod in affected programs: 4 -> 2 (-50.00%)
helped: 2
HURT: 0
total temps in shared programs: 17570 -> 17566 (-0.02%)
temps in affected programs: 1122 -> 1118 (-0.36%)
helped: 41
HURT: 43
total consts in shared programs: 94362 -> 94359 (<.01%)
consts in affected programs: 381 -> 378 (-0.79%)
helped: 13
HURT: 10
total lits in shared programs: 2951 -> 2961 (0.34%)
lits in affected programs: 104 -> 114 (9.62%)
helped: 3
HURT: 9
total cycles in shared programs: 198965 -> 198744 (-0.11%)
cycles in affected programs: 55784 -> 55563 (-0.40%)
helped: 177
HURT: 144

LOST:   0
GAINED: 1

Shader-db RV380:
total instructions in shared programs: 84224 -> 84109 (-0.14%)
instructions in affected programs: 6039 -> 5924 (-1.90%)
helped: 106
HURT: 38
total presub in shared programs: 1401 -> 1372 (-2.07%)
presub in affected programs: 113 -> 84 (-25.66%)
helped: 27
HURT: 10
total temps in shared programs: 13231 -> 13224 (-0.05%)
temps in affected programs: 303 -> 296 (-2.31%)
helped: 22
HURT: 12
total consts in shared programs: 82484 -> 82505 (0.03%)
consts in affected programs: 271 -> 292 (7.75%)
helped: 4
HURT: 25
total cycles in shared programs: 132957 -> 132840 (-0.09%)
cycles in affected programs: 12696 -> 12579 (-0.92%)
helped: 101
HURT: 43

LOST:   0
GAINED: 1

Signed-off-by: Pavel Ondračka <pavel.ondracka@gmail.com>
Reviewed-by: Filip Gawin <filip.gawin@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27089>
This commit is contained in:
Pavel Ondračka
2024-01-14 20:40:15 +01:00
committed by Marge Bot
parent 99abd9ad30
commit 5115cebd5d
4 changed files with 65 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/gallium/drivers/r300/compiler/nir_to_rc.c
+2
View File
@@ -2465,6 +2465,8 @@ const void *nir_to_rc_options(struct nir_shader *s,
else
NIR_PASS_V(s, r300_nir_lower_fcsel_r300);
NIR_PASS_V(s, r300_nir_lower_flrp);
} else {
NIR_PASS_V(s, r300_nir_lower_comparison_fs);
}
NIR_PASS_V(s, r300_nir_opt_algebraic_late);
NIR_PASS_V(s, nir_opt_dce);
src/gallium/drivers/r300/compiler/r300_nir.c
+3
View File
@@ -137,6 +137,9 @@ r300_optimize_nir(struct nir_shader *s, struct pipe_screen *screen)
nir_metadata_block_index |
nir_metadata_dominance, NULL);
NIR_PASS(progress, s, nir_opt_peephole_select, is_r500 ? 8 : ~0, true, true);
if (s->info.stage == MESA_SHADER_FRAGMENT) {
NIR_PASS(progress, s, r300_nir_lower_bool_to_float_fs);
}
NIR_PASS(progress, s, nir_opt_algebraic);
NIR_PASS(progress, s, nir_opt_constant_folding);
nir_load_store_vectorize_options vectorize_opts = {
src/gallium/drivers/r300/compiler/r300_nir.h
+17
View File
@@ -47,6 +47,19 @@ is_ubo_or_input(UNUSED struct hash_table *ht, const nir_alu_instr *instr,
}
}
static inline bool
is_not_used_in_single_if(const nir_alu_instr *instr)
{
unsigned if_uses = 0;
nir_foreach_use(src, &instr->def) {
if (nir_src_is_if(src))
if_uses++;
else
return true;
}
return if_uses != 1;
}
static inline bool
is_only_used_by_load_ubo_vec4(const nir_alu_instr *instr)
{
@@ -77,6 +90,8 @@ 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_lower_bool_to_float_fs(struct nir_shader *shader);
extern bool r300_nir_prepare_presubtract(struct nir_shader *shader);
extern bool r300_nir_opt_algebraic_late(struct nir_shader *shader);
@@ -89,4 +104,6 @@ extern bool r300_nir_lower_fcsel_r300(nir_shader *shader);
extern bool r300_nir_lower_flrp(nir_shader *shader);
extern bool r300_nir_lower_comparison_fs(nir_shader *shader);
#endif /* R300_NIR_H */
src/gallium/drivers/r300/compiler/r300_nir_algebraic.py
+43 -1
View File
@@ -93,6 +93,9 @@ r300_nir_opt_algebraic_late = [
# does not and blows up. Clean this up.
(('fneg', ('fneg', a)), a),
(('fabs', ('fneg', a)), ('fabs', a)),
# Some cleanups after comparison lowering if one of the operands is 0.
(('fadd', a, 0.0), a),
(('fadd', a, ('fneg', 0.0)), a)
]
# This is very late flrp lowering to clean up after bcsel->fcsel->flrp.
@@ -105,6 +108,20 @@ r300_nir_lower_fcsel_r300 = [
(('fcsel_ge', a, b, c), ('flrp', c, b, ('sge', a, 0.0)))
]
# Fragment shaders have no comparison opcodes. However, we can encode the comparison
# in the aluresults operation, which is than used by next if. So if the comparison result
# is used only in a single if, we can handle it just fine on R500.
r300_nir_lower_comparison_fs = [
(('seq(is_not_used_in_single_if)', 'a@32', 'b@32'),
('fcsel_ge', ('fneg', ('fabs', ('fadd', a, ('fneg', b)))), 1.0, 0.0)),
(('sne(is_not_used_in_single_if)', 'a@32', 'b@32'),
('fcsel_ge', ('fneg', ('fabs', ('fadd', a, ('fneg', b)))), 0.0, 1.0)),
(('slt(is_not_used_in_single_if)', 'a@32', 'b@32'),
('fcsel_ge', ('fadd', a, ('fneg', b)), 0.0, 1.0)),
(('sge(is_not_used_in_single_if)', 'a@32', 'b@32'),
('fcsel_ge', ('fadd', a, ('fneg', b)), 1.0, 0.0)),
]
r300_nir_post_integer_lowering = [
# If ffloor result is used only for indirect constant load, we can get rid of it
# completelly as ntt emits ARL by default which already does the flooring.
@@ -148,7 +165,26 @@ def main():
('fcsel', ('slt', a, b), c, d), "options->has_fused_comp_and_csel"),
(('bcsel@32(is_only_used_as_float)', ('fge', 'a@32', 'b@32'), c, d),
('fcsel', ('sge', a, b), c, d), "options->has_fused_comp_and_csel"),
]
]
r300_nir_lower_bool_to_float_fs = [
(('bcsel@32(r300_is_only_used_as_float)', ignore_exact('feq', 'a@32', 'b@32'), c, d),
('fcsel_ge', ('fneg', ('fabs', ('fadd', a, ('fneg', b)))), c, d)),
(('bcsel@32(r300_is_only_used_as_float)', ignore_exact('fneu', 'a@32', 'b@32'), c, d),
('fcsel_ge', ('fneg', ('fabs', ('fadd', a, ('fneg', b)))), d, c)),
(('bcsel@32(r300_is_only_used_as_float)', ignore_exact('flt', 'a@32', 'b@32'), c, d),
('fcsel_ge', ('fadd', a, ('fneg', b)), d, c)),
(('bcsel@32(r300_is_only_used_as_float)', ignore_exact('fge', 'a@32', 'b@32'), c, d),
('fcsel_ge', ('fadd', a, ('fneg', b)), c, d)),
(('b2f32', ('feq', 'a@32', 'b@32')),
('fcsel_ge', ('fneg', ('fabs', ('fadd', a, ('fneg', b)))), 1.0, 0.0)),
(('b2f32', ('fneu', 'a@32', 'b@32')),
('fcsel_ge', ('fneg', ('fabs', ('fadd', a, ('fneg', b)))), 0.0, 1.0)),
(('b2f32', ('flt', 'a@32', 'b@32')),
('fcsel_ge', ('fadd', a, ('fneg', b)), 0.0, 1.0)),
(('b2f32', ('fge', 'a@32', 'b@32')),
('fcsel_ge', ('fadd', a, ('fneg', b)), 1.0, 0.0)),
]
with open(args.output, 'w') as f:
f.write('#include "compiler/r300_nir.h"')
@@ -165,6 +201,9 @@ 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_lower_bool_to_float_fs",
r300_nir_lower_bool_to_float_fs).render())
f.write(nir_algebraic.AlgebraicPass("r300_nir_prepare_presubtract",
r300_nir_prepare_presubtract).render())
@@ -180,5 +219,8 @@ def main():
f.write(nir_algebraic.AlgebraicPass("r300_nir_lower_fcsel_r300",
r300_nir_lower_fcsel_r300).render())
f.write(nir_algebraic.AlgebraicPass("r300_nir_lower_comparison_fs",
r300_nir_lower_comparison_fs).render())
if __name__ == '__main__':
main()