raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-27 05:16:56 +01:00
Code Issues Packages Projects Releases Wiki Activity

shader: bound storage buffers by pushed binding sizes instead of OpArrayLength

Browse Source
This commit is contained in:
nmzik
2026-10-25 20:04:37 +01:00
parent 7524e8f79e
commit ba7f9969ef
7 changed files with 96 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/shader/recompiler/backend/spirv/spirvEmitterInternal.h
+2 -1
View File
@@ -112,6 +112,7 @@ struct EmitterState {
uint32_t lane_half = 0;
std::array<BufferDefinition, 4> storage_buffers {};
std::vector<uint32_t> memory_byte_offsets;
std::vector<uint32_t> memory_byte_sizes;
uint32_t bda_pagetable_variable = 0;
uint32_t fault_buffer_variable = 0;
uint32_t bda_pointer_function = 0;
@@ -414,7 +415,7 @@ Prospero::BufferFormat StorageBufferFormat(const EmitterState& state, const IR::
uint32_t SharedIndirectBufferResource(const EmitterState& state, uint32_t resource);
void EmitMemoryOffsets(EmitterState& state);
void EmitMemoryBindings(EmitterState& state);
uint32_t LdsDwordCount(const EmitterState& state);
void EnsureLdsStorage(EmitterState& state);
src/graphics/shader/recompiler/backend/spirv/spirvEmitterMemoryHelpers.cpp
+16 -7
View File
@@ -154,9 +154,10 @@ uint32_t SharedIndirectBufferResource(const EmitterState& state, uint32_t resour
return representative;
}
void EmitMemoryOffsets(EmitterState& state) {
void EmitMemoryBindings(EmitterState& state) {
const auto count = state.program.bindings.descriptor_counts[static_cast<size_t>(IR::DescriptorBindingKind::Buffers)];
state.memory_byte_offsets.resize(count);
state.memory_byte_sizes.resize(count);
std::vector<bool> dynamic_offsets(count);
for (uint32_t resource = 0; resource < state.program.info.buffers.size(); ++resource) {
if (SharedIndirectBufferResource(state, resource) == UINT32_MAX) continue;
@@ -174,6 +175,8 @@ void EmitMemoryOffsets(EmitterState& state) {
if (dynamic_offsets[i + lane]) continue;
state.memory_byte_offsets[i + lane] = EmitNative<spv::OpBitFieldUExtract, IR::Type::U32>(
state, word, ConstantU32(state, lane * 8u), ConstantU32(state, 8u));
state.memory_byte_sizes[i + lane] =
EmitShaderDataDwordLoad(state, state.program.bindings.MemorySizeDword() + i + lane);
}
}
}
@@ -257,22 +260,28 @@ MemoryResourceAccess PrepareStorageBufferResourceAccess(EmitterState& state,
state.builder.RequireCapability(spv::CapabilityShaderNonUniform);
state.builder.RequireCapability(spv::CapabilityStorageBufferArrayNonUniformIndexing);
state.builder.AddAnnotation(spv::OpDecorate, access.object_pointer, spv::DecorationNonUniform);
const auto offsets_start = state.program.bindings.memory_offset_dword +
(state.program.bindings.UsesPushData() ? state.program.bindings.push_data_start_dword : 0u);
const auto& bindings = state.program.bindings;
const auto data_start = bindings.UsesPushData() ? bindings.push_data_start_dword : 0u;
const auto word = EmitShaderDataDwordLoadAt(state, EmitAddU32(
state, ConstantU32(state, offsets_start),
state, ConstantU32(state, bindings.memory_offset_dword + data_start),
Binary(state, spv::OpShiftRightLogical, TypeU32(state), array_index, ConstantU32(state, 2))));
const auto shift = Binary(state, spv::OpShiftLeftLogical, TypeU32(state),
Binary(state, spv::OpBitwiseAnd, TypeU32(state), array_index, ConstantU32(state, 3)),
ConstantU32(state, 3));
access.byte_offset = EmitNative<spv::OpBitFieldUExtract, IR::Type::U32>(
state, word, shift, ConstantU32(state, 8));
access.length = EmitShaderDataDwordLoadAt(state, EmitAddU32(
state, ConstantU32(state, bindings.MemorySizeDword() + data_start), array_index));
} else {
access.byte_offset = state.memory_byte_offsets[static_index];
access.length = state.memory_byte_sizes[static_index];
}
// storage_buffers[i] views the binding as 1 << i byte elements.
const auto shift = static_cast<uint32_t>(&buffer - state.storage_buffers.data());
if (shift != 0u) {
access.length = Binary(state, spv::OpShiftRightLogical, TypeU32(state), access.length,
ConstantU32(state, shift));
}
access.length = state.builder.AllocateId();
state.builder.AddFunction(spv::OpArrayLength, TypeU32(state), access.length,
access.object_pointer, 0);
return access;
}
src/graphics/shader/recompiler/backend/spirv/spirvEmitterProgram.cpp
+1 -1
View File
@@ -746,7 +746,7 @@ void EmitProgram(EmitterState& state) {
state.lds_base_dwords = EmitBinaryU32(state, spv::OpIMul, index,
ConstantU32(state, LdsDwordCount(state)));
}
EmitMemoryOffsets(state);
EmitMemoryBindings(state);
if (program.blocks.empty()) {
EmitReturn(ctx);
} else if (state.program.dispatcher_fallback) {
src/graphics/shader/recompiler/ir/ShaderIR.h
+7 -1
View File
@@ -495,10 +495,16 @@ struct Bindings {
uint32_t dispatch_thread_dword = PushData::NoStart;
uint32_t memory_offset_dword = 0;
[[nodiscard]] uint32_t ShaderDataDwords() const {
// Packed 8-bit binding offsets, then each binding's bound range in bytes: OpArrayLength
// reads 0 for ranges of 2 GiB or more on NVIDIA.
[[nodiscard]] uint32_t MemorySizeDword() const {
return memory_offset_dword +
(descriptor_counts[static_cast<size_t>(DescriptorBindingKind::Buffers)] + 3u) / 4u;
}
[[nodiscard]] uint32_t ShaderDataDwords() const {
return MemorySizeDword() +
descriptor_counts[static_cast<size_t>(DescriptorBindingKind::Buffers)];
}
[[nodiscard]] bool UsesPushData() const {
return push_data_start_dword != PushData::NoStart;
}