raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-20 10:56:58 +02:00
Code Issues Packages Projects Releases Wiki Activity

graphics: size compute LDS from shader state

Browse Source
This commit is contained in:
nmzik
2026-08-16 13:48:10 +02:00
parent 1375d650a7
commit 4e343408dc
8 changed files with 172 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/shader/recompiler/backend/spirv/spirvEmitterInternal.h
+4
View File
@@ -972,6 +972,10 @@ uint32_t EmitStorageBufferElementPointer(EmitterState& state, const IR::MemoryIn
uint32_t EmitLdsElementPointer(EmitterState& state, uint32_t index);
uint32_t LdsDwordCount(const EmitterState& state);
uint32_t EmitLdsElementInBounds(EmitterState& state, uint32_t index);
uint32_t EmitGdsElementInBounds(EmitterState& state, uint32_t index);
uint32_t EmitGdsElementPointer(EmitterState& state, uint32_t index);
src/graphics/shader/recompiler/backend/spirv/spirvEmitterMemory.cpp
+5 -2
View File
@@ -495,11 +495,14 @@ uint32_t EmitLdsElementPointer(EmitterState& state, uint32_t index) {
return pointer;
}
uint32_t LdsDwordCount(const EmitterState& state) {
return state.needs_function_lds ? 8192u : state.compute_input_info->lds_size_dwords;
}
uint32_t EmitLdsElementInBounds(EmitterState& state, uint32_t index) {
const auto in_bounds = state.builder.AllocateId();
const auto dwords = state.needs_function_lds ? 8192u : 1024u;
state.builder.AddFunction(
{OpULessThan, state.bool_type, in_bounds, index, ConstantU32(state, dwords)});
{OpULessThan, state.bool_type, in_bounds, index, ConstantU32(state, LdsDwordCount(state))});
return in_bounds;
}
src/graphics/shader/recompiler/backend/spirv/spirvEmitterModule.cpp
+4 -1
View File
@@ -1,5 +1,7 @@
#include "graphics/shader/recompiler/backend/spirv/spirvEmitterInternal.h"
#include <algorithm>
namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter {
const IR::DescriptorBinding* DescriptorBinding(const EmitterState& state,
@@ -736,7 +738,8 @@ void EmitHeaderAndTypes(EmitterState& state) {
if (state.stage == ShaderType::Compute || state.needs_function_lds) {
const auto storage_class =
state.needs_function_lds ? StorageClassFunction : StorageClassWorkgroup;
const auto lds_size = ConstantU32(state, state.needs_function_lds ? 8192u : 1024u);
const auto lds_dwords = std::max(LdsDwordCount(state), 1u);
const auto lds_size = ConstantU32(state, lds_dwords);
state.builder.AddType({OpTypeArray, state.lds_array_type, state.uint_type, lds_size});
state.builder.AddType(
{OpTypePointer, state.ptr_workgroup_array, storage_class, state.lds_array_type});
src/graphics/shader/recompiler/backend/spirv/spirvEmitterValueMemory.cpp
+3 -2
View File
@@ -158,7 +158,7 @@ uint32_t DwordIndex(ValueEmitContext& ctx, const IR::Inst& inst, const IR::Memor
uint32_t InBounds(ValueEmitContext& ctx, const IR::MemoryInfo& mem, uint32_t index, uint32_t pc) {
if (mem.kind == IR::ResourceKind::Lds) {
return 0;
return EmitLdsElementInBounds(ctx.state, index);
}
if (mem.kind == IR::ResourceKind::Gds) {
return EmitGdsElementInBounds(ctx.state, index);
@@ -546,7 +546,8 @@ uint32_t AppendConsume(ValueEmitContext& ctx, const IR::Inst& inst, bool gds, bo
ConstantU32(state, ScopeSubgroup), ballot});
const auto is_first =
Binary(state, OpIEqual, state.bool_type, EmitSubgroupLocalInvocationId(state), first);
const auto storage_bounds = gds ? EmitGdsElementInBounds(state, index) : 0u;
const auto storage_bounds =
gds ? EmitGdsElementInBounds(state, index) : EmitLdsElementInBounds(state, index);
const auto m0_bounds =
gds ? Binary(state, OpINotEqual, state.bool_type, size, ConstantU32(state, 0))
: Binary(state, OpULessThan, state.bool_type,
src/graphics/shader/shader.cpp
+15 -11
View File
@@ -894,15 +894,16 @@ static void ShaderGetStaticInputInfoCS(const HW::ComputeShaderInfo& regs,
ShaderComputeInputInfo& info) {
info = {};
info.threads_num[0] = regs.cs_regs.num_thread_x;
info.threads_num[1] = regs.cs_regs.num_thread_y;
info.threads_num[2] = regs.cs_regs.num_thread_z;
info.group_id[0] = regs.cs_regs.tgid_x_en != 0;
info.group_id[1] = regs.cs_regs.tgid_y_en != 0;
info.group_id[2] = regs.cs_regs.tgid_z_en != 0;
info.wave_size = regs.cs_regs.wave_size;
info.thread_ids_num = regs.cs_regs.tidig_comp_cnt + 1;
info.tg_size_en = regs.cs_regs.tg_size_en != 0;
info.threads_num[0] = regs.cs_regs.num_thread_x;
info.threads_num[1] = regs.cs_regs.num_thread_y;
info.threads_num[2] = regs.cs_regs.num_thread_z;
info.lds_size_dwords = static_cast<uint32_t>(regs.cs_regs.lds_size) * 128u;
info.group_id[0] = regs.cs_regs.tgid_x_en != 0;
info.group_id[1] = regs.cs_regs.tgid_y_en != 0;
info.group_id[2] = regs.cs_regs.tgid_z_en != 0;
info.wave_size = regs.cs_regs.wave_size;
info.thread_ids_num = regs.cs_regs.tidig_comp_cnt + 1;
info.tg_size_en = regs.cs_regs.tg_size_en != 0;
info.workgroup_register = regs.cs_regs.user_sgpr;
}
@@ -1330,10 +1331,12 @@ void ShaderDbgDumpInputInfo(const ShaderComputeInputInfo& info) {
LOGF("\t workgroup_register = %d\n"
"\t thread_ids_num = %d\n"
"\t wave_size = %u\n"
"\t lds_size_dwords = %u\n"
"\t threads_num = {%u, %u, %u}\n"
"\t tg_size_en = %s\n",
info.workgroup_register, info.thread_ids_num, info.wave_size, info.threads_num[0],
info.threads_num[1], info.threads_num[2], info.tg_size_en ? "true" : "false");
info.workgroup_register, info.thread_ids_num, info.wave_size, info.lds_size_dwords,
info.threads_num[0], info.threads_num[1], info.threads_num[2],
info.tg_size_en ? "true" : "false");
LOGF("\t threadgroup_id = {%s, %s, %s}\n", info.group_id[0] ? "true" : "false",
info.group_id[1] ? "true" : "false", info.group_id[2] ? "true" : "false");
}
@@ -1723,6 +1726,7 @@ ShaderId ShaderGetIdCS(const HW::ComputeShaderInfo& regs, const ShaderComputeInp
ret.ids.push_back(input_info.workgroup_register);
ret.ids.push_back(input_info.wave_size);
ret.ids.push_back(input_info.thread_ids_num);
ret.ids.push_back(input_info.lds_size_dwords);
for (int i = 0; i < 3; i++) {
ret.ids.push_back(input_info.threads_num[i]);
src/graphics/shader/shader.h
+1
View File
@@ -91,6 +91,7 @@ struct ShaderVertexInputInfo {
struct ShaderComputeInputInfo {
uint32_t threads_num[3] = {0, 0, 0};
uint32_t dispatch_threads_num[3] = {0, 0, 0};
uint32_t lds_size_dwords = 0;
bool group_id[3] = {false, false, false};
bool dispatch_thread_dimensions = false;
uint32_t wave_size = 64;