fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2025-12-31 12:00:12 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

nir: guard nir_def_as_alu()

Browse source 
We will potentially create load_const_instr instead of ALU.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/37195>
This commit is contained in:
Daniel Schürmann 2025-09-05 10:57:27 +02:00 • committed by Marge Bot
parent 3180656bbc
commit d1f2f1222e
3 changed files with 8 additions and 58 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

3
src/compiler/nir/nir_builder.h
View file

@ -1033,7 +1033,8 @@ static inline nir_def *
nir_iadd_nuw(nir_builder *b, nir_def *x, nir_def *y)
{
nir_def *d = nir_iadd(b, x, y);
nir_def_as_alu(d)->no_unsigned_wrap = true;
if (d->parent_instr->type == nir_instr_type_alu)
nir_def_as_alu(d)->no_unsigned_wrap = true;
return d;
}

57
src/compiler/nir/nir_lower_flrp.c
View file

@ -52,16 +52,8 @@ replace_with_strict_ffma(struct nir_builder *bld, struct u_vector *dead_flrp,
nir_def *const c = nir_ssa_for_alu_src(bld, alu, 2);
nir_def *const neg_a = nir_fneg(bld, a);
nir_def_as_alu(neg_a)->exact = alu->exact;
nir_def_as_alu(neg_a)->fp_fast_math = alu->fp_fast_math;
nir_def *const inner_ffma = nir_ffma(bld, neg_a, c, a);
nir_def_as_alu(inner_ffma)->exact = alu->exact;
nir_def_as_alu(inner_ffma)->fp_fast_math = alu->fp_fast_math;
nir_def *const outer_ffma = nir_ffma(bld, b, c, inner_ffma);
nir_def_as_alu(outer_ffma)->exact = alu->exact;
nir_def_as_alu(outer_ffma)->fp_fast_math = alu->fp_fast_math;
nir_def_rewrite_uses(&alu->def, outer_ffma);
@ -84,21 +76,10 @@ replace_with_single_ffma(struct nir_builder *bld, struct u_vector *dead_flrp,
nir_def *const c = nir_ssa_for_alu_src(bld, alu, 2);
nir_def *const neg_c = nir_fneg(bld, c);
nir_def_as_alu(neg_c)->exact = alu->exact;
nir_def_as_alu(neg_c)->fp_fast_math = alu->fp_fast_math;
nir_def *const one_minus_c =
nir_fadd(bld, nir_imm_floatN_t(bld, 1.0f, c->bit_size), neg_c);
nir_def_as_alu(one_minus_c)->exact = alu->exact;
nir_def_as_alu(one_minus_c)->fp_fast_math = alu->fp_fast_math;
nir_def *const b_times_c = nir_fmul(bld, b, c);
nir_def_as_alu(b_times_c)->exact = alu->exact;
nir_def_as_alu(b_times_c)->fp_fast_math = alu->fp_fast_math;
nir_def *const final_ffma = nir_ffma(bld, a, one_minus_c, b_times_c);
nir_def_as_alu(final_ffma)->exact = alu->exact;
nir_def_as_alu(final_ffma)->fp_fast_math = alu->fp_fast_math;
nir_def_rewrite_uses(&alu->def, final_ffma);
@ -121,25 +102,11 @@ replace_with_strict(struct nir_builder *bld, struct u_vector *dead_flrp,
nir_def *const c = nir_ssa_for_alu_src(bld, alu, 2);
nir_def *const neg_c = nir_fneg(bld, c);
nir_def_as_alu(neg_c)->exact = alu->exact;
nir_def_as_alu(neg_c)->fp_fast_math = alu->fp_fast_math;
nir_def *const one_minus_c =
nir_fadd(bld, nir_imm_floatN_t(bld, 1.0f, c->bit_size), neg_c);
nir_def_as_alu(one_minus_c)->exact = alu->exact;
nir_def_as_alu(one_minus_c)->fp_fast_math = alu->fp_fast_math;
nir_def *const first_product = nir_fmul(bld, a, one_minus_c);
nir_def_as_alu(first_product)->exact = alu->exact;
nir_def_as_alu(first_product)->fp_fast_math = alu->fp_fast_math;
nir_def *const second_product = nir_fmul(bld, b, c);
nir_def_as_alu(second_product)->exact = alu->exact;
nir_def_as_alu(second_product)->fp_fast_math = alu->fp_fast_math;
nir_def *const sum = nir_fadd(bld, first_product, second_product);
nir_def_as_alu(sum)->exact = alu->exact;
nir_def_as_alu(sum)->fp_fast_math = alu->fp_fast_math;
nir_def_rewrite_uses(&alu->def, sum);
@ -162,20 +129,9 @@ replace_with_fast(struct nir_builder *bld, struct u_vector *dead_flrp,
nir_def *const c = nir_ssa_for_alu_src(bld, alu, 2);
nir_def *const neg_a = nir_fneg(bld, a);
nir_def_as_alu(neg_a)->exact = alu->exact;
nir_def_as_alu(neg_a)->fp_fast_math = alu->fp_fast_math;
nir_def *const b_minus_a = nir_fadd(bld, b, neg_a);
nir_def_as_alu(b_minus_a)->exact = alu->exact;
nir_def_as_alu(b_minus_a)->fp_fast_math = alu->fp_fast_math;
nir_def *const product = nir_fmul(bld, c, b_minus_a);
nir_def_as_alu(product)->exact = alu->exact;
nir_def_as_alu(product)->fp_fast_math = alu->fp_fast_math;
nir_def *const sum = nir_fadd(bld, a, product);
nir_def_as_alu(sum)->exact = alu->exact;
nir_def_as_alu(sum)->fp_fast_math = alu->fp_fast_math;
nir_def_rewrite_uses(&alu->def, sum);
@ -201,27 +157,16 @@ replace_with_expanded_ffma_and_add(struct nir_builder *bld,
nir_def *const c = nir_ssa_for_alu_src(bld, alu, 2);
nir_def *const b_times_c = nir_fmul(bld, b, c);
nir_def_as_alu(b_times_c)->exact = alu->exact;
nir_def_as_alu(b_times_c)->fp_fast_math = alu->fp_fast_math;
nir_def *inner_sum;
if (subtract_c) {
nir_def *const neg_c = nir_fneg(bld, c);
nir_def_as_alu(neg_c)->exact = alu->exact;
nir_def_as_alu(neg_c)->fp_fast_math = alu->fp_fast_math;
inner_sum = nir_fadd(bld, a, neg_c);
} else {
inner_sum = nir_fadd(bld, a, c);
}
nir_def_as_alu(inner_sum)->exact = alu->exact;
nir_def_as_alu(inner_sum)->fp_fast_math = alu->fp_fast_math;
nir_def *const outer_sum = nir_fadd(bld, inner_sum, b_times_c);
nir_def_as_alu(outer_sum)->exact = alu->exact;
nir_def_as_alu(outer_sum)->fp_fast_math = alu->fp_fast_math;
nir_def_rewrite_uses(&alu->def, outer_sum);
@ -399,6 +344,8 @@ convert_flrp_instruction(nir_builder *bld,
UNREACHABLE("invalid bit_size");
bld->cursor = nir_before_instr(&alu->instr);
bld->exact = alu->exact;
bld->fp_fast_math = alu->fp_fast_math;
/* There are two methods to implement flrp(x, y, t). The strictly correct
* implementation according to the GLSL spec is:

6
src/compiler/spirv/vtn_alu.c
View file

@ -1087,8 +1087,10 @@ vtn_handle_alu(struct vtn_builder *b, SpvOp opcode,
case SpvOpShiftLeftLogical:
case SpvOpSNegate: {
nir_alu_instr *alu = nir_def_as_alu(dest->def);
alu->no_signed_wrap |= vtn_has_decoration(b, dest_val, SpvDecorationNoSignedWrap);
alu->no_unsigned_wrap |= vtn_has_decoration(b, dest_val, SpvDecorationNoUnsignedWrap);
if (alu) {
alu->no_signed_wrap |= vtn_has_decoration(b, dest_val, SpvDecorationNoSignedWrap);
alu->no_unsigned_wrap |= vtn_has_decoration(b, dest_val, SpvDecorationNoUnsignedWrap);
}
break;
}
default: