OpenHarmony/third_party_mesa3d
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

nir: change 16bit image dest folding option to per type

Browse Source 
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21404>
This commit is contained in:
Georg Lehmann
2023-02-18 13:37:48 +01:00
committed by Marge Bot
parent a14d46fde2
commit a00b50d820
5 changed files with 19 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
src/amd/vulkan/radv_pipeline.c
View File

@@ -3297,7 +3297,8 @@ radv_postprocess_nir(struct radv_pipeline *pipeline,
struct nir_fold_16bit_tex_image_options fold_16bit_options = { struct nir_fold_16bit_tex_image_options fold_16bit_options = {
.rounding_mode = nir_rounding_mode_rtne, .rounding_mode = nir_rounding_mode_rtne,
.fold_tex_dest_types = nir_type_float | nir_type_uint | nir_type_int, .fold_tex_dest_types = nir_type_float | nir_type_uint | nir_type_int,
.fold_image_load_store_data = true, .fold_image_dest_types = nir_type_float | nir_type_uint | nir_type_int,
.fold_image_store_data = true,
.fold_image_srcs = !radv_use_llvm_for_stage(device, stage->stage), .fold_image_srcs = !radv_use_llvm_for_stage(device, stage->stage),
.fold_srcs_options_count = separate_g16 ? 2 : 1, .fold_srcs_options_count = separate_g16 ? 2 : 1,
.fold_srcs_options = fold_srcs_options, .fold_srcs_options = fold_srcs_options,

3
src/compiler/nir/nir.h
View File

@@ -5566,7 +5566,8 @@ struct nir_fold_tex_srcs_options {
struct nir_fold_16bit_tex_image_options { struct nir_fold_16bit_tex_image_options {
nir_rounding_mode rounding_mode; nir_rounding_mode rounding_mode;
nir_alu_type fold_tex_dest_types; nir_alu_type fold_tex_dest_types;
bool fold_image_load_store_data; nir_alu_type fold_image_dest_types;
bool fold_image_store_data;
bool fold_image_srcs; bool fold_image_srcs;
unsigned fold_srcs_options_count; unsigned fold_srcs_options_count;
struct nir_fold_tex_srcs_options *fold_srcs_options; struct nir_fold_tex_srcs_options *fold_srcs_options;

15
src/compiler/nir/nir_lower_mediump.c
View File

@@ -886,11 +886,14 @@ fold_16bit_destination(nir_ssa_def *ssa, nir_alu_type dest_type,
} }
static bool static bool
fold_16bit_load_data(nir_builder *b, nir_intrinsic_instr *instr, fold_16bit_image_dest(nir_intrinsic_instr *instr, unsigned exec_mode,
unsigned exec_mode, nir_rounding_mode rdm) nir_alu_type allowed_types, nir_rounding_mode rdm)
{ {
nir_alu_type dest_type = nir_intrinsic_dest_type(instr); nir_alu_type dest_type = nir_intrinsic_dest_type(instr);
if (!(nir_alu_type_get_base_type(dest_type) & allowed_types))
return false;
if (!fold_16bit_destination(&instr->dest.ssa, dest_type, exec_mode, rdm)) if (!fold_16bit_destination(&instr->dest.ssa, dest_type, exec_mode, rdm))
return false; return false;
@@ -1016,7 +1019,7 @@ fold_16bit_tex_image(nir_builder *b, nir_instr *instr, void *params)
case nir_intrinsic_bindless_image_store: case nir_intrinsic_bindless_image_store:
case nir_intrinsic_image_deref_store: case nir_intrinsic_image_deref_store:
case nir_intrinsic_image_store: case nir_intrinsic_image_store:
if (options->fold_image_load_store_data) if (options->fold_image_store_data)
progress |= fold_16bit_store_data(b, intrinsic); progress |= fold_16bit_store_data(b, intrinsic);
if (options->fold_image_srcs) if (options->fold_image_srcs)
progress |= fold_16bit_image_srcs(b, intrinsic, 4); progress |= fold_16bit_image_srcs(b, intrinsic, 4);
@@ -1024,8 +1027,10 @@ fold_16bit_tex_image(nir_builder *b, nir_instr *instr, void *params)
case nir_intrinsic_bindless_image_load: case nir_intrinsic_bindless_image_load:
case nir_intrinsic_image_deref_load: case nir_intrinsic_image_deref_load:
case nir_intrinsic_image_load: case nir_intrinsic_image_load:
if (options->fold_image_load_store_data) if (options->fold_image_dest_types)
progress |= fold_16bit_load_data(b, intrinsic, exec_mode, options->rounding_mode); progress |= fold_16bit_image_dest(intrinsic, exec_mode,
options->fold_image_dest_types,
options->rounding_mode);
if (options->fold_image_srcs) if (options->fold_image_srcs)
progress |= fold_16bit_image_srcs(b, intrinsic, 3); progress |= fold_16bit_image_srcs(b, intrinsic, 3);
break; break;

4
src/freedreno/ir3/ir3_nir.c
View File

@@ -783,7 +783,9 @@ ir3_nir_lower_variant(struct ir3_shader_variant *so, nir_shader *s)
.rounding_mode = nir_rounding_mode_rtz, .rounding_mode = nir_rounding_mode_rtz,
.fold_tex_dest_types = nir_type_float, .fold_tex_dest_types = nir_type_float,
/* blob dumps have no half regs on pixel 2's ldib or stib, so only enable for a6xx+. */ /* blob dumps have no half regs on pixel 2's ldib or stib, so only enable for a6xx+. */
.fold_image_load_store_data = so->compiler->gen >= 6, .fold_image_dest_types = so->compiler->gen >= 6 ?
nir_type_float | nir_type_uint | nir_type_int : 0,
.fold_image_store_data = so->compiler->gen >= 6,
.fold_srcs_options_count = 1, .fold_srcs_options_count = 1,
.fold_srcs_options = &fold_srcs_options, .fold_srcs_options = &fold_srcs_options,
}; };

3
src/gallium/drivers/radeonsi/si_shader_nir.c
View File

@@ -198,7 +198,8 @@ static void si_late_optimize_16bit_samplers(struct si_screen *sscreen, nir_shade
struct nir_fold_16bit_tex_image_options fold_16bit_options = { struct nir_fold_16bit_tex_image_options fold_16bit_options = {
.rounding_mode = nir_rounding_mode_rtne, .rounding_mode = nir_rounding_mode_rtne,
.fold_tex_dest_types = nir_type_float | nir_type_uint | nir_type_int, .fold_tex_dest_types = nir_type_float | nir_type_uint | nir_type_int,
.fold_image_load_store_data = true, .fold_image_dest_types = nir_type_float | nir_type_uint | nir_type_int,
.fold_image_store_data = true,
.fold_srcs_options_count = has_g16 ? 2 : 1, .fold_srcs_options_count = has_g16 ? 2 : 1,
.fold_srcs_options = fold_srcs_options, .fold_srcs_options = fold_srcs_options,
}; };