fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2025-12-28 06:00:10 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

amd/lower_mem_access_bit_sizes: also use SMEM for subdword loads

Browse source 
We can simply extract from the loaded dwords as per
nir_lower_mem_access_bit_sizes() lowering.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/37843>
This commit is contained in:
Daniel Schürmann 2025-10-08 19:15:50 +02:00 • committed by Marge Bot
parent fbf0399517
commit 8ff44f17ef
5 changed files with 5 additions and 13 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
src/amd/common/nir/ac_nir.h
View file

@ -413,7 +413,7 @@ bool
ac_nir_opt_shared_append(nir_shader *shader);
bool
ac_nir_flag_smem_for_loads(nir_shader *shader, enum amd_gfx_level gfx_level, bool use_llvm, bool after_lowering);
ac_nir_flag_smem_for_loads(nir_shader *shader, enum amd_gfx_level gfx_level, bool use_llvm);
bool
ac_nir_lower_mem_access_bit_sizes(nir_shader *shader, enum amd_gfx_level gfx_level, bool use_llvm);

10
src/amd/common/nir/ac_nir_lower_mem_access_bit_sizes.c
View file

@ -13,7 +13,6 @@
typedef struct {
enum amd_gfx_level gfx_level;
bool use_llvm;
bool after_lowering;
} mem_access_cb_data;
static bool
@ -39,12 +38,6 @@ use_smem_for_load(nir_builder *b, nir_intrinsic_instr *intrin, void *cb_data_)
if (intrin->def.divergent)
return false;
/* ACO doesn't support instruction selection for multi-component 8/16-bit SMEM loads. */
const bool supports_scalar_subdword = cb_data->gfx_level >= GFX12 && !cb_data->use_llvm;
if (cb_data->after_lowering && intrin->def.bit_size < 32 &&
(intrin->def.num_components > 1 || !supports_scalar_subdword))
return false;
enum gl_access_qualifier access = nir_intrinsic_access(intrin);
bool glc = access & (ACCESS_VOLATILE | ACCESS_COHERENT);
bool reorder = nir_intrinsic_can_reorder(intrin) || ((access & ACCESS_NON_WRITEABLE) && !(access & ACCESS_VOLATILE));
@ -169,12 +162,11 @@ lower_mem_access_cb(nir_intrinsic_op intrin, uint8_t bytes, uint8_t bit_size, ui
}
bool
ac_nir_flag_smem_for_loads(nir_shader *shader, enum amd_gfx_level gfx_level, bool use_llvm, bool after_lowering)
ac_nir_flag_smem_for_loads(nir_shader *shader, enum amd_gfx_level gfx_level, bool use_llvm)
{
mem_access_cb_data cb_data = {
.gfx_level = gfx_level,
.use_llvm = use_llvm,
.after_lowering = after_lowering,
};
return nir_shader_intrinsics_pass(shader, &use_smem_for_load, nir_metadata_all, &cb_data);
}

2
src/amd/compiler/instruction_selection/aco_isel_setup.cpp
View file

@ -381,7 +381,7 @@ init_context(isel_context* ctx, nir_shader* shader)
nir_divergence_analysis_impl(impl, (nir_divergence_options)options);
apply_nuw_to_offsets(ctx, impl);
ac_nir_flag_smem_for_loads(shader, ctx->program->gfx_level, false, true);
ac_nir_flag_smem_for_loads(shader, ctx->program->gfx_level, false);
if (shader->info.stage == MESA_SHADER_FRAGMENT) {
nir_opt_load_skip_helpers_options skip_helper_options = {};

2
src/amd/vulkan/radv_pipeline.c
View file

@ -354,7 +354,7 @@ radv_postprocess_nir(struct radv_device *device, const struct radv_graphics_stat
NIR_PASS(_, stage->nir, radv_nir_opt_tid_function, &tid_options);
nir_divergence_analysis(stage->nir);
NIR_PASS(_, stage->nir, ac_nir_flag_smem_for_loads, gfx_level, use_llvm, false);
NIR_PASS(_, stage->nir, ac_nir_flag_smem_for_loads, gfx_level, use_llvm);
NIR_PASS(_, stage->nir, radv_nir_opt_access_can_speculate);
NIR_PASS(_, stage->nir, nir_lower_memory_model);

2
src/gallium/drivers/radeonsi/si_shader.c
View file

@ -1627,7 +1627,7 @@ static void run_late_optimization_and_lowering_passes(struct si_nir_shader_ctx *
nir_divergence_analysis(nir); /* required by ac_nir_flag_smem_for_loads */
/* This is required by ac_nir_scalarize_overfetching_loads_callback. */
NIR_PASS(progress, nir, ac_nir_flag_smem_for_loads, sel->screen->info.gfx_level,
!sel->info.base.use_aco_amd, false);
!sel->info.base.use_aco_amd);
/* Scalarize overfetching loads, so that we don't load more components than necessary.
* Adjacent loads will be re-vectorized with a conservative overfetching limit.
*/