fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-01-04 11:20:20 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

ac/llvm: lower vector load_const in NIR

Browse source 
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33046>
This commit is contained in:
Marek Olšák 2024-12-29 14:26:12 -05:00 • committed by Marge Bot
parent d160252270
commit 08a47fa05c
2 changed files with 9 additions and 42 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

40
src/amd/llvm/ac_nir_to_llvm.c
View file

@ -1269,40 +1269,13 @@ static bool visit_alu(struct ac_nir_context *ctx, const nir_alu_instr *instr)
return true;
}
static bool visit_load_const(struct ac_nir_context *ctx, const nir_load_const_instr *instr)
static void visit_load_const(struct ac_nir_context *ctx, const nir_load_const_instr *instr)
{
LLVMValueRef values[16], value = NULL;
LLVMTypeRef element_type = LLVMIntTypeInContext(ctx->ac.context, instr->def.bit_size);
assert(instr->def.num_components == 1);
for (unsigned i = 0; i < instr->def.num_components; ++i) {
switch (instr->def.bit_size) {
case 1:
values[i] = LLVMConstInt(element_type, instr->value[i].b, false);
break;
case 8:
values[i] = LLVMConstInt(element_type, instr->value[i].u8, false);
break;
case 16:
values[i] = LLVMConstInt(element_type, instr->value[i].u16, false);
break;
case 32:
values[i] = LLVMConstInt(element_type, instr->value[i].u32, false);
break;
case 64:
values[i] = LLVMConstInt(element_type, instr->value[i].u64, false);
break;
default:
fprintf(stderr, "unsupported nir load_const bit_size: %d\n", instr->def.bit_size);
return false;
}
}
if (instr->def.num_components > 1) {
value = LLVMConstVector(values, instr->def.num_components);
} else
value = values[0];
ctx->ssa_defs[instr->def.index] = value;
return true;
ctx->ssa_defs[instr->def.index] =
LLVMConstInt(LLVMIntTypeInContext(ctx->ac.context, instr->def.bit_size),
nir_const_value_as_uint(instr->value[0], instr->def.bit_size), false);
}
/* Gather4 should follow the same rules as bilinear filtering, but the hardware
@ -3868,8 +3841,7 @@ static bool visit_block(struct ac_nir_context *ctx, nir_block *block)
return false;
break;
case nir_instr_type_load_const:
if (!visit_load_const(ctx, nir_instr_as_load_const(instr)))
return false;
visit_load_const(ctx, nir_instr_as_load_const(instr));
break;
case nir_instr_type_intrinsic:
if (!visit_intrinsic(ctx, nir_instr_as_intrinsic(instr)))

11
src/gallium/drivers/radeonsi/si_shader.c
View file

@ -2694,11 +2694,8 @@ static struct nir_shader *si_get_nir_shader(struct si_shader *shader, struct si_
*/
NIR_PASS(progress, nir, nir_opt_move, nir_move_comparisons);
/* aco only accept scalar const, must be done after si_nir_late_opts()
* which may generate vec const.
*/
if (nir->info.use_aco_amd)
NIR_PASS_V(nir, nir_lower_load_const_to_scalar);
/* This must be done after si_nir_late_opts() because it may generate vec const. */
NIR_PASS(_, nir, nir_lower_load_const_to_scalar);
/* This helps LLVM form VMEM clauses and thus get more GPU cache hits.
* 200 is tuned for Viewperf. It should be done last.
@ -2802,9 +2799,7 @@ si_nir_generate_gs_copy_shader(struct si_screen *sscreen,
si_nir_opts(gs_selector->screen, nir, false);
/* aco only accept scalar const */
if (gs_nir->info.use_aco_amd)
NIR_PASS_V(nir, nir_lower_load_const_to_scalar);
NIR_PASS_V(nir, nir_lower_load_const_to_scalar);
if (si_can_dump_shader(sscreen, MESA_SHADER_GEOMETRY, SI_DUMP_NIR)) {
fprintf(stderr, "GS Copy Shader:\n");