From 5115cebd5d5f7980dabd1f0ff5c994a371f1a0b6 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Pavel=20Ondra=C4=8Dka?= Date: Sun, 14 Jan 2024 20:40:15 +0100 Subject: [PATCH] r300: lower comparison ops early in NIR MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 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 Reviewed-by: Filip Gawin Part-of: --- src/gallium/drivers/r300/compiler/nir_to_rc.c | 2 + src/gallium/drivers/r300/compiler/r300_nir.c | 3 ++ src/gallium/drivers/r300/compiler/r300_nir.h | 17 +++++++ .../r300/compiler/r300_nir_algebraic.py | 44 ++++++++++++++++++- 4 files changed, 65 insertions(+), 1 deletion(-) diff --git a/src/gallium/drivers/r300/compiler/nir_to_rc.c b/src/gallium/drivers/r300/compiler/nir_to_rc.c index bfc65970429..d909389f03e 100644 --- a/src/gallium/drivers/r300/compiler/nir_to_rc.c +++ b/src/gallium/drivers/r300/compiler/nir_to_rc.c @@ -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); diff --git a/src/gallium/drivers/r300/compiler/r300_nir.c b/src/gallium/drivers/r300/compiler/r300_nir.c index 3cb2faf2098..b137dc02742 100644 --- a/src/gallium/drivers/r300/compiler/r300_nir.c +++ b/src/gallium/drivers/r300/compiler/r300_nir.c @@ -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 = { diff --git a/src/gallium/drivers/r300/compiler/r300_nir.h b/src/gallium/drivers/r300/compiler/r300_nir.h index 2a49864af96..2739b4539d5 100644 --- a/src/gallium/drivers/r300/compiler/r300_nir.h +++ b/src/gallium/drivers/r300/compiler/r300_nir.h @@ -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 */ diff --git a/src/gallium/drivers/r300/compiler/r300_nir_algebraic.py b/src/gallium/drivers/r300/compiler/r300_nir_algebraic.py index fc6a8519429..f77b3b49d4a 100644 --- a/src/gallium/drivers/r300/compiler/r300_nir_algebraic.py +++ b/src/gallium/drivers/r300/compiler/r300_nir_algebraic.py @@ -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()