raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-21 19:36:55 +02:00
Code Issues Packages Projects Releases Wiki Activity

shader: collect usage once and emit descriptor slots directly

Browse Source
This commit is contained in:
nmzik
2026-10-09 02:17:38 +04:00
parent b35cdc39b7
commit e1ac893a0c
28 changed files with 1068 additions and 1408 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/host_gpu/renderer/pipeline/descriptors.cpp
+67 -68
View File
@@ -26,7 +26,6 @@
#include "graphics/host_gpu/renderer/renderContext.h"
#include "graphics/host_gpu/vulkanCommon.h"
#include "graphics/shader/recompiler/ir/ShaderIR.h"
#include "graphics/shader/recompiler/ir/passes/BindingLayout.h"
#include "graphics/shader/recompiler/ir/passes/ResourceMaterialization.h"
#include "graphics/shader/shader.h"
#include "kernel/memory.h"
@@ -76,10 +75,6 @@ vk::DescriptorType NativeDescriptorType(BindingKind kind) {
EXIT("invalid native descriptor binding kind");
}
uint32_t NativeDescriptorCount(const ShaderRecompiler::IR::DescriptorBinding& binding) {
return binding.resources.empty() ? 1u : static_cast<uint32_t>(binding.resources.size());
}
vk::DescriptorImageInfo MakeImageInfo(const TextureBinding& texture, uint32_t element) {
vk::ImageView view = nullptr;
if (texture.mip_views.empty()) {
@@ -799,12 +794,11 @@ void RenderExecutor::PrepareBindings(const ShaderStageRuntime& runtime,
m_context, program, i, snapshot.samplers[program.info.samplers[i].snapshot_index]));
}
prepared.shader_data.reserve(program.bindings.ShaderDataDwords());
for (const auto reg: program.bindings.user_data_registers) {
for (const auto reg: program.info.user_data_registers) {
prepared.shader_data.push_back(snapshot.user_data[reg - program.user_data_base]);
}
prepared.shader_data.resize(program.bindings.ShaderDataDwords());
if (ShaderRecompiler::IR::FindBinding(
program.bindings, ShaderRecompiler::IR::DescriptorBindingKind::Gds) != nullptr) {
if (program.bindings.descriptor_counts[static_cast<size_t>(BindingKind::Gds)] != 0) {
prepared.gds.buffer = m_context.GetBufferCache().GetGdsBuffer()->Handle();
}
}
@@ -817,19 +811,20 @@ void RenderExecutor::FindBuffers(std::span<PreparedBindings* const> stages) {
EXIT_IF(prepared.runtime == nullptr || !*prepared.runtime);
const auto& program = *prepared.runtime->program;
const auto& snapshot = *prepared.runtime->resources;
prepared.buffer_sources.clear();
const auto& layout = program.bindings;
if (layout.memory_offset_count == 0) {
prepared.buffer_sources.resize(layout.descriptor_counts[static_cast<size_t>(BindingKind::Buffers)]);
if (layout.descriptor_counts[static_cast<size_t>(BindingKind::Buffers)] == 0) {
continue;
}
const auto& resources = layout.descriptors.front().resources;
prepared.buffer_sources.reserve(resources.size());
for (const auto resource: resources) {
for (uint32_t resource = 0; resource < program.info.buffers.size(); ++resource) {
const auto& metadata = program.info.buffers[resource];
if (metadata.descriptor_index == UINT32_MAX) continue;
auto& source = prepared.buffer_sources[metadata.descriptor_index];
const auto descriptor = DecodeNativeDescriptor<ShaderBufferResource>(snapshot.buffers[resource]);
const auto address = descriptor.Base48();
auto size = descriptor.GetSize();
if (address == 0 || size == 0) {
prepared.buffer_sources.push_back({});
source = {};
continue;
}
if (descriptor.NumRecords() == UINT32_MAX) {
@@ -853,7 +848,7 @@ void RenderExecutor::FindBuffers(std::span<PreparedBindings* const> stages) {
}
}
size = Libs::LibKernel::Memory::ClampRangeSize(address, size);
prepared.buffer_sources.push_back({address, size, cache.FindBuffer(address, size)});
source = {address, size, cache.FindBuffer(address, size)};
}
}
}
@@ -864,10 +859,9 @@ void RenderExecutor::RebindBuffers(PreparedBindings& prepared) {
const auto& program = *prepared.runtime->program;
const auto& snapshot = *prepared.runtime->resources;
const auto& layout = program.bindings;
EXIT_IF(prepared.buffer_sources.size() != layout.memory_offset_count);
EXIT_IF(prepared.buffer_sources.size() != layout.descriptor_counts[static_cast<size_t>(BindingKind::Buffers)]);
prepared.buffers.clear();
prepared.buffers.reserve(layout.memory_offset_count);
prepared.buffers.resize(layout.descriptor_counts[static_cast<size_t>(BindingKind::Buffers)]);
EXIT_IF(prepared.shader_data.size() != layout.ShaderDataDwords());
std::fill(prepared.shader_data.begin() + layout.memory_offset_dword,
prepared.shader_data.end(), 0);
@@ -876,20 +870,18 @@ void RenderExecutor::RebindBuffers(PreparedBindings& prepared) {
const auto shift = (index % 4u) * 8u;
prepared.shader_data[dword] |= offset << shift;
};
for (uint32_t i = 0; i < layout.memory_offset_count; i++) {
const auto resource = layout.descriptors.front().resources[i];
for (const auto& resource: program.info.buffers) {
const auto index = resource.descriptor_index;
if (index == UINT32_MAX) continue;
uint32_t buffer_offset = 0;
prepared.buffers.push_back(NativeStorageBuffer(m_context, prepared.buffer_sources[i],
program.info.buffers[resource],
buffer_offset));
pack_memory_offset(i, buffer_offset);
prepared.buffers[index] = NativeStorageBuffer(m_context, prepared.buffer_sources[index],
resource, buffer_offset);
pack_memory_offset(index, buffer_offset);
}
if (ShaderRecompiler::IR::FindBinding(
layout, ShaderRecompiler::IR::DescriptorBindingKind::FlattenedSrt) != nullptr) {
if (layout.descriptor_counts[static_cast<size_t>(BindingKind::FlattenedSrt)] != 0) {
prepared.flattened_srt = NativeUpload(m_context, snapshot.flattened_srt);
}
if (ShaderRecompiler::IR::FindBinding(
program.bindings, ShaderRecompiler::IR::DescriptorBindingKind::ShaderData) != nullptr) {
if (layout.descriptor_counts[static_cast<size_t>(BindingKind::ShaderData)] != 0) {
prepared.shader_data_buffer = NativeUpload(m_context, prepared.shader_data);
}
}
@@ -982,6 +974,7 @@ void RenderExecutor::CommitBindings(CommandBuffer& buffer,
KYTY_PROFILER_FUNCTION();
auto vk_buffer = buffer.Handle();
size_t descriptor_count = 0;
size_t image_descriptor_count = 0;
size_t write_count = 0;
ShaderRecompiler::IR::PushData push_data;
bool has_push_data = false;
@@ -995,9 +988,15 @@ void RenderExecutor::CommitBindings(CommandBuffer& buffer,
for (const auto* prepared: prepared_bindings) {
EXIT_IF(prepared == nullptr || prepared->runtime == nullptr || !*prepared->runtime);
const auto& program = *prepared->runtime->program;
write_count += program.bindings.descriptors.size();
for (const auto& binding: program.bindings.descriptors) {
descriptor_count += NativeDescriptorCount(binding);
write_count += std::popcount(program.bindings.descriptor_mask);
for (auto mask = program.bindings.descriptor_mask; mask != 0; mask &= mask - 1) {
const auto index = std::countr_zero(mask);
const auto count = program.bindings.descriptor_counts[index];
descriptor_count += count;
if (index >= ShaderRecompiler::IR::FirstImageBinding &&
index <= static_cast<size_t>(BindingKind::Samplers)) {
image_descriptor_count += count;
}
}
const auto shader_stage = NativeShaderStage(program.stage);
push_stages |= shader_stage;
@@ -1029,10 +1028,9 @@ void RenderExecutor::CommitBindings(CommandBuffer& buffer,
}
for (const auto* writer: prepared_bindings) {
const auto& program = *writer->runtime->program;
for (uint32_t i = 0; i < writer->buffer_sources.size(); ++i) {
const auto resource = program.bindings.descriptors.front().resources[i];
if (!program.info.buffers[resource].written) continue;
const auto& written = writer->buffer_sources[i];
for (const auto& resource: program.info.buffers) {
if (!resource.written || resource.descriptor_index == UINT32_MAX) continue;
const auto& written = writer->buffer_sources[resource.descriptor_index];
if (written.size != 0 && ImageRangeOverlaps(address, size,
written.address, written.size)) {
EXIT("scalar resource reads overlap a shader buffer write\n");
@@ -1042,11 +1040,11 @@ void RenderExecutor::CommitBindings(CommandBuffer& buffer,
}
}
m_descriptor_buffers.clear();
m_descriptor_images.clear();
m_descriptor_images.resize(image_descriptor_count);
m_descriptor_writes.clear();
m_descriptor_buffers.reserve(descriptor_count);
m_descriptor_images.reserve(descriptor_count);
m_descriptor_writes.reserve(write_count);
size_t image_cursor = 0;
for (auto* prepared: prepared_bindings) {
const auto& program = *prepared->runtime->program;
@@ -1111,24 +1109,24 @@ void RenderExecutor::CommitBindings(CommandBuffer& buffer,
binding.layout = image.backing.state.layout;
}
m_image_occurrences.assign(descriptors.images.size(), 0);
for (const auto& binding: program.bindings.descriptors) {
std::array<size_t, ShaderRecompiler::IR::ImageBindingCount> image_starts {};
for (auto mask = program.bindings.descriptor_mask; mask != 0; mask &= mask - 1) {
const auto index = std::countr_zero(mask);
const auto kind = static_cast<BindingKind>(index);
vk::WriteDescriptorSet write {};
write.dstBinding = ShaderRecompiler::IR::NativeBinding(program.stage, binding.kind);
write.descriptorType = NativeDescriptorType(binding.kind);
write.descriptorCount = NativeDescriptorCount(binding);
write.dstBinding = ShaderRecompiler::IR::NativeBinding(program.stage, kind);
write.descriptorType = NativeDescriptorType(kind);
write.descriptorCount = program.bindings.descriptor_counts[index];
const auto buffer_start = m_descriptor_buffers.size();
const auto image_start = m_descriptor_images.size();
if (ShaderRecompiler::IR::ImageBindingResourceClass(binding.kind) !=
const auto image_start = image_cursor;
if (ShaderRecompiler::IR::ImageBindingResourceClass(kind) !=
ShaderRecompiler::IR::ImageResourceClass::None) {
for (const auto resource: binding.resources) {
m_descriptor_images.push_back(MakeImageInfo(
descriptors.images.at(resource), m_image_occurrences.at(resource)++));
}
image_starts[ShaderRecompiler::IR::ImageBindingIndex(kind)] = image_start;
image_cursor += write.descriptorCount;
} else {
switch (binding.kind) {
switch (kind) {
case BindingKind::Buffers:
EXIT_IF(descriptors.buffers.size() != binding.resources.size());
EXIT_IF(descriptors.buffers.size() != write.descriptorCount);
for (const auto& view: descriptors.buffers) {
EXIT_IF(view.buffer == nullptr);
m_descriptor_buffers.push_back(view);
@@ -1136,12 +1134,11 @@ void RenderExecutor::CommitBindings(CommandBuffer& buffer,
break;
case BindingKind::BdaPagetable:
case BindingKind::FaultBuffer: {
auto& cache = m_context.GetBufferCache();
const auto* bda_buffer = binding.kind == BindingKind::BdaPagetable
auto& cache = m_context.GetBufferCache();
const auto* bda_buffer = kind == BindingKind::BdaPagetable
? cache.GetBdaPageTableBuffer()
: cache.GetFaultBuffer();
m_descriptor_buffers.emplace_back(bda_buffer->Handle(), 0,
bda_buffer->Size());
m_descriptor_buffers.emplace_back(bda_buffer->Handle(), 0, bda_buffer->Size());
break;
}
case BindingKind::FlattenedSrt:
@@ -1149,11 +1146,11 @@ void RenderExecutor::CommitBindings(CommandBuffer& buffer,
case BindingKind::SharedMemory:
case BindingKind::Gds: {
const vk::DescriptorBufferInfo* view = &descriptors.gds;
if (binding.kind == BindingKind::FlattenedSrt) {
if (kind == BindingKind::FlattenedSrt) {
view = &descriptors.flattened_srt;
} else if (binding.kind == BindingKind::ShaderData) {
} else if (kind == BindingKind::ShaderData) {
view = &descriptors.shader_data_buffer;
} else if (binding.kind == BindingKind::SharedMemory) {
} else if (kind == BindingKind::SharedMemory) {
view = &descriptors.shared_memory;
}
EXIT_IF(view->buffer == nullptr);
@@ -1161,11 +1158,9 @@ void RenderExecutor::CommitBindings(CommandBuffer& buffer,
break;
}
case BindingKind::Samplers:
for (const auto resource: binding.resources) {
const auto sampler = descriptors.samplers.at(resource);
for (const auto sampler: descriptors.samplers) {
EXIT_IF(sampler == nullptr);
m_descriptor_images.emplace_back(sampler, nullptr,
vk::ImageLayout::eUndefined);
m_descriptor_images[image_cursor++] = {sampler, nullptr, vk::ImageLayout::eUndefined};
}
break;
case BindingKind::Count: EXIT("invalid descriptor binding kind");
@@ -1174,17 +1169,21 @@ void RenderExecutor::CommitBindings(CommandBuffer& buffer,
if (m_descriptor_buffers.size() != buffer_start) {
write.pBufferInfo = m_descriptor_buffers.data() + buffer_start;
}
if (m_descriptor_images.size() != image_start) {
if (image_cursor != image_start) {
write.pImageInfo = m_descriptor_images.data() + image_start;
}
m_descriptor_writes.push_back(write);
}
for (uint32_t i = 0; i < descriptors.images.size(); i++) {
const auto expected =
descriptors.images[i].mip_views.empty()
? 1u
: static_cast<uint32_t>(descriptors.images[i].mip_views.size());
EXIT_IF(m_image_occurrences[i] != expected);
for (uint32_t i = 0; i < program.info.images.size(); ++i) {
const auto& resource = program.info.images[i];
const auto& texture = descriptors.images[i];
EXIT_IF(resource.mip_count != (texture.mip_views.empty() ? 1u : texture.mip_views.size()));
const auto kind = *ShaderRecompiler::IR::DescriptorBindingForImage(resource);
const auto start = image_starts[ShaderRecompiler::IR::ImageBindingIndex(kind)] +
resource.descriptor_index;
for (uint32_t mip = 0; mip < resource.mip_count; ++mip) {
m_descriptor_images[start + mip] = MakeImageInfo(texture, mip);
}
}
const auto shader_data_dwords = program.bindings.ShaderDataDwords();
src/graphics/host_gpu/renderer/pipeline/descriptors.h
-2
View File
@@ -50,8 +50,6 @@ struct PreparedBindings {
[[nodiscard]] vk::DescriptorType
NativeDescriptorType(ShaderRecompiler::IR::DescriptorBindingKind kind);
[[nodiscard]] uint32_t
NativeDescriptorCount(const ShaderRecompiler::IR::DescriptorBinding& binding);
[[nodiscard]] vk::DescriptorImageInfo MakeImageInfo(const TextureBinding& texture,
uint32_t element = 0);
src/graphics/host_gpu/renderer/pipeline/shaders.cpp
+6 -3
View File
@@ -19,6 +19,7 @@
#include "graphics/shader/shader.h"
#include <algorithm>
#include <bit>
#include <limits>
#include <span>
#include <vector>
@@ -180,10 +181,12 @@ static vk::BlendOp GetBlendOp(uint32_t op) {
static void AddLayoutBindings(std::vector<vk::DescriptorSetLayoutBinding>& descriptor_bindings,
const ShaderRecompiler::IR::CompiledShaderInfo& program,
vk::ShaderStageFlagBits stage) {
for (const auto& binding: program.bindings.descriptors) {
for (auto mask = program.bindings.descriptor_mask; mask != 0; mask &= mask - 1) {
const auto index = std::countr_zero(mask);
const auto kind = static_cast<ShaderRecompiler::IR::DescriptorBindingKind>(index);
descriptor_bindings.push_back(
{ShaderRecompiler::IR::NativeBinding(program.stage, binding.kind),
NativeDescriptorType(binding.kind), NativeDescriptorCount(binding), stage, nullptr});
{ShaderRecompiler::IR::NativeBinding(program.stage, kind), NativeDescriptorType(kind),
program.bindings.descriptor_counts[index], stage, nullptr});
}
}
src/graphics/host_gpu/renderer/render.h
-1
View File
@@ -216,7 +216,6 @@ private:
std::vector<vk::DescriptorBufferInfo> m_descriptor_buffers;
std::vector<vk::DescriptorImageInfo> m_descriptor_images;
std::vector<vk::WriteDescriptorSet> m_descriptor_writes;
std::vector<uint32_t> m_image_occurrences;
friend class CommandProcessor;
friend struct RenderExecutorTestAccess;
src/graphics/host_gpu/renderer/renderCompute.cpp
+2 -3
View File
@@ -20,7 +20,6 @@
#include "graphics/host_gpu/vulkanCommon.h"
#include "graphics/shader/recompiler/ir/ShaderIR.h"
#include "graphics/shader/recompiler/ir/passes/ResourceMaterialization.h"
#include "graphics/shader/recompiler/ir/passes/BindingLayout.h"
#include "graphics/shader/shader.h"
#include "kernel/eventQueue.h"
#include "kernel/pthread.h"
@@ -198,8 +197,8 @@ bool RenderExecutor::TryConsumeComputeImageClear(const ShaderComputeInputInfo& i
static void BindSharedMemory(RenderContext& context, ShaderComputeInputInfo& input,
PreparedBindings& bindings, uint64_t indirect_args = 0) {
if (ShaderRecompiler::IR::FindBinding(input.stage.program->bindings,
ShaderRecompiler::IR::DescriptorBindingKind::SharedMemory) == nullptr) {
if (input.stage.program->bindings.descriptor_counts[
static_cast<size_t>(ShaderRecompiler::IR::DescriptorBindingKind::SharedMemory)] == 0) {
return;
}
auto& cache = context.GetBufferCache();
src/graphics/shader/recompiler/ShaderRecompiler.cpp
+2 -6
View File
@@ -8,7 +8,6 @@
#include "graphics/shader/recompiler/frontend/decode/ShaderDecoder.h"
#include "graphics/shader/recompiler/frontend/translate/Translate.h"
#include "graphics/shader/recompiler/ir/ShaderIR.h"
#include "graphics/shader/recompiler/ir/passes/BindingLayout.h"
#include "graphics/shader/recompiler/ir/passes/ConstantPropagation.h"
#include "graphics/shader/recompiler/ir/passes/DeadCodeElimination.h"
#include "graphics/shader/recompiler/ir/passes/ReadLaneElimination.h"
@@ -670,20 +669,17 @@ CompileResult CompileProgram(TranslateResult translated, const CompileOptions& o
IR::EliminateDeadCode(ir.blocks);
IR::CollectShaderInfo(ir, options.input_info);
IR::AllocateBindings(ir, push_data_start_dword,
ir.stage == ShaderType::Compute && options.input_info.compute != nullptr &&
options.input_info.compute->lds_storage);
std::string ir_dump;
if (options.dump_ir) {
ir_dump = MakeIrDump(translated.cfg_dump, ir);
if (options.early_dump) {
LOGF("%s native IR and bindings (early):\n%s", GetDumpLabel(options), ir_dump.c_str());
LOGF("%s native IR (early):\n%s", GetDumpLabel(options), ir_dump.c_str());
}
}
LOGF("%s phase begin: stage=%s hash=0x%016" PRIx64 " SPIR-V EmitProgram\n",
GetDumpLabel(options), StageName(ir.stage), ir.shader_hash);
auto spirv = Spirv::EmitProgram(ir, options.input_info);
auto spirv = Spirv::EmitProgram(ir, options.input_info, push_data_start_dword);
LOGF("%s phase end: stage=%s hash=0x%016" PRIx64 " SPIR-V EmitProgram words=%" PRIu64
" elapsed_ms=%" PRIu64 "\n",
GetDumpLabel(options), StageName(ir.stage), ir.shader_hash,
src/graphics/shader/recompiler/backend/spirv/SpirvEmitter.cpp
+101 -242
View File
@@ -3,7 +3,6 @@
#include "common/assert.h"
#include "graphics/shader/recompiler/backend/spirv/spirvEmitterInternal.h"
#include "graphics/shader/recompiler/ir/ShaderIR.h"
#include "graphics/shader/recompiler/ir/passes/BindingLayout.h"
#include <algorithm>
#include <array>
@@ -19,92 +18,14 @@ namespace {
}
void ValidateNativeProgram(const IR::Program& program, bool lds_storage) {
using Kind = IR::DescriptorBindingKind;
constexpr auto KindCount = static_cast<size_t>(Kind::Count);
std::array<std::vector<uint32_t>, KindCount> expected;
std::array<bool, KindCount> present {};
const auto Dense = [](size_t size) {
std::vector<uint32_t> values(size);
for (uint32_t i = 0; i < values.size(); i++) {
values[i] = i;
}
return values;
uint64_t live_buffers = 0;
bool uses_lds = false;
bool uses_gds = false;
const auto uses_mapping = [](const auto& resource) {
return resource.indirect_root != UINT32_MAX;
};
auto Expect = [&](Kind kind, std::vector<uint32_t> resources = {}) {
const auto index = static_cast<size_t>(kind);
present[index] = true;
expected[index] = std::move(resources);
};
for (uint32_t i = 0; i < program.info.images.size(); i++) {
const auto kind = IR::DescriptorBindingForImage(program.info.images[i]);
if (!kind.has_value()) {
Fail(program, "native shader plan has an invalid image class");
}
present[static_cast<size_t>(*kind)] = true;
const auto dynamic = program.info.images[i].mip_mode == IR::ImageMipMode::Dynamic;
const auto count = dynamic ? program.info.images[i].mip_count : 1u;
if (count == 0u || (!dynamic && program.info.images[i].mip_count != 1u)) {
Fail(program, "native shader plan has an invalid image mip descriptor count");
}
expected[static_cast<size_t>(*kind)].insert(expected[static_cast<size_t>(*kind)].end(),
count, i);
}
if (!program.info.samplers.empty()) {
Expect(Kind::Samplers, Dense(program.info.samplers.size()));
}
auto& buffers = expected[static_cast<size_t>(Kind::Buffers)];
const auto shared = IR::CollectMemoryResources(program, buffers);
present[static_cast<size_t>(Kind::Buffers)] = !buffers.empty();
if (shared.gds) {
Expect(Kind::Gds);
}
if (shared.lds && lds_storage) {
Expect(Kind::SharedMemory);
}
if (program.info.uses_dma) {
Expect(Kind::BdaPagetable);
Expect(Kind::FaultBuffer);
}
if (IR::UsesFlattenedSrt(program)) {
Expect(Kind::FlattenedSrt);
}
if (program.bindings.ShaderDataDwords() != 0 && !program.bindings.UsesPushData()) {
Expect(Kind::ShaderData);
}
std::array<bool, KindCount> seen {};
for (const auto& binding: program.bindings.descriptors) {
const auto kind = static_cast<size_t>(binding.kind);
if (kind >= KindCount || seen[kind] || !present[kind] ||
binding.resources != expected[kind]) {
Fail(program, "native descriptor groups do not match shader topology");
}
seen[kind] = true;
}
for (size_t i = 0; i < KindCount; i++) {
if (present[i] != seen[i]) {
Fail(program, "native shader plan is missing a required descriptor group");
}
}
const auto has_shader_data_storage = present[static_cast<size_t>(Kind::ShaderData)];
const auto shader_data_dwords = program.bindings.ShaderDataDwords();
const auto user_data_dwords = program.bindings.user_data_registers.size();
const bool has_dispatch_threads = program.bindings.dispatch_thread_dword != IR::PushData::NoStart;
if ((program.bindings.UsesPushData() &&
!IR::PushData::CanFit(program.bindings.push_data_start_dword, shader_data_dwords)) ||
(has_dispatch_threads && (program.stage != ShaderType::Compute ||
program.bindings.dispatch_thread_dword != user_data_dwords)) ||
program.bindings.memory_offset_dword != user_data_dwords + (has_dispatch_threads ? 3u : 0u) ||
program.bindings.memory_offset_count != buffers.size() ||
has_shader_data_storage != (shader_data_dwords != 0 && !program.bindings.UsesPushData()) ||
!std::is_sorted(program.bindings.user_data_registers.begin(),
program.bindings.user_data_registers.end()) ||
std::adjacent_find(program.bindings.user_data_registers.begin(),
program.bindings.user_data_registers.end()) !=
program.bindings.user_data_registers.end()) {
Fail(program, "native shader-data layout is inconsistent");
}
bool uses_flattened_srt = std::ranges::any_of(program.info.buffers, uses_mapping) ||
std::ranges::any_of(program.info.images, uses_mapping);
const auto planning_only_handle = [&](const IR::Inst& handle) {
return !handle.Uses().empty() &&
std::ranges::all_of(handle.Uses(), [&](const IR::Use& use) {
@@ -141,6 +62,39 @@ void ValidateNativeProgram(const IR::Program& program, bool lds_storage) {
};
for (const auto* block: program.blocks) {
for (const auto& inst: *block) {
const auto op = inst.GetOpcode();
uses_flattened_srt |= op == IR::ValueOpcode::ReadConst;
if (IR::BufferAccessOf(op) != IR::BufferAccess::None ||
IR::SharedAccessOf(op) != IR::SharedAccess::None ||
IR::AddressOpcodeInfoOf(op).access != IR::AddressAccess::None) {
const auto index = inst.Flags<IR::MemoryFlags>().index;
if (index >= program.memory_info.size()) {
Fail(program, "typed shader contains invalid memory metadata");
}
const auto& memory = program.memory_info[index];
if (!memory.planning_only) {
uses_lds |= memory.kind == IR::ResourceKind::FlatLocal;
if (IR::SharedAccessOf(op) != IR::SharedAccess::None) {
if (memory.kind != IR::ResourceKind::Lds && memory.kind != IR::ResourceKind::Gds) {
Fail(program, "typed shader contains invalid shared-memory metadata");
}
uses_gds |= memory.kind == IR::ResourceKind::Gds;
uses_lds |= memory.kind == IR::ResourceKind::Lds;
} else if (memory.kind == IR::ResourceKind::Buffer ||
memory.kind == IR::ResourceKind::ScalarBuffer) {
if (memory.resource >= program.info.buffers.size()) {
Fail(program, "typed shader contains an invalid buffer resource");
}
live_buffers |= uint64_t{1} << memory.resource;
for (const auto child: program.info.buffers[memory.resource].indirect_resources) {
if (child >= program.info.buffers.size()) {
Fail(program, "typed shader contains an invalid indirect buffer resource");
}
live_buffers |= uint64_t{1} << child;
}
}
}
}
const auto dense = inst.Flags<uint32_t>();
switch (inst.GetOpcode()) {
case IR::ValueOpcode::GetBufferResource:
@@ -186,163 +140,64 @@ void ValidateNativeProgram(const IR::Program& program, bool lds_storage) {
}
}
}
using Kind = IR::DescriptorBindingKind;
std::array<uint32_t, static_cast<size_t>(Kind::Count)> expected {};
auto& buffer_count = expected[static_cast<size_t>(Kind::Buffers)];
for (uint32_t index = 0; index < program.info.buffers.size(); ++index) {
const auto slot = (live_buffers & (uint64_t{1} << index)) != 0 ? buffer_count++ : UINT32_MAX;
if (program.info.buffers[index].descriptor_index != slot) {
Fail(program, "native buffer descriptor index does not match shader topology");
}
}
for (const auto& image: program.info.images) {
const auto kind = IR::DescriptorBindingForImage(image);
if (!kind.has_value()) {
Fail(program, "native shader plan has an invalid image class");
}
if (image.mip_count == 0u ||
(image.mip_mode != IR::ImageMipMode::Dynamic && image.mip_count != 1u)) {
Fail(program, "native shader plan has an invalid image mip descriptor count");
}
auto& count = expected[static_cast<size_t>(*kind)];
if (image.descriptor_index != count) {
Fail(program, "native image descriptor index does not match shader topology");
}
count += image.mip_count;
}
expected[static_cast<size_t>(Kind::Samplers)] = static_cast<uint32_t>(program.info.samplers.size());
expected[static_cast<size_t>(Kind::Gds)] = uses_gds;
expected[static_cast<size_t>(Kind::SharedMemory)] = uses_lds && lds_storage;
expected[static_cast<size_t>(Kind::BdaPagetable)] = program.info.uses_dma;
expected[static_cast<size_t>(Kind::FaultBuffer)] = program.info.uses_dma;
expected[static_cast<size_t>(Kind::FlattenedSrt)] = uses_flattened_srt;
expected[static_cast<size_t>(Kind::ShaderData)] =
program.bindings.ShaderDataDwords() != 0 && !program.bindings.UsesPushData();
uint64_t descriptor_mask = 0;
for (uint32_t index = 0; index < expected.size(); ++index) {
if (expected[index] != 0) descriptor_mask |= uint64_t{1} << index;
}
if (program.bindings.descriptor_counts != expected ||
program.bindings.descriptor_mask != descriptor_mask) {
Fail(program, "native descriptor counts do not match shader topology");
}
const auto shader_data_dwords = program.bindings.ShaderDataDwords();
const auto& registers = program.info.user_data_registers;
const bool has_dispatch_threads = program.bindings.dispatch_thread_dword != IR::PushData::NoStart;
if ((program.bindings.UsesPushData() &&
!IR::PushData::CanFit(program.bindings.push_data_start_dword, shader_data_dwords)) ||
(has_dispatch_threads && (program.stage != ShaderType::Compute ||
program.bindings.dispatch_thread_dword != registers.size())) ||
program.bindings.memory_offset_dword != registers.size() + (has_dispatch_threads ? 3u : 0u) ||
!std::is_sorted(registers.begin(), registers.end()) ||
std::adjacent_find(registers.begin(), registers.end()) != registers.end()) {
Fail(program, "native shader-data layout is inconsistent");
}
}
} // namespace
Emitter::SpirvRequirements Emitter::AnalyzeProgramRequirements(const IR::Program& program) {
SpirvRequirements requirements {};
for (const auto* block: program.blocks) {
for (const auto& inst: *block) {
requirements.float64 |= inst.GetType() == IR::Type::F64;
if (IR::BufferAccessOf(inst.GetOpcode()) == IR::BufferAccess::Atomic &&
inst.GetType() == IR::Type::U64) {
requirements.buffer_int64_atomics = true;
}
const auto address_access = IR::AddressOpcodeInfoOf(inst.GetOpcode()).access;
if (address_access != IR::AddressAccess::None) {
const auto memory_index = inst.Flags<IR::MemoryFlags>().index;
if (memory_index >= program.memory_info.size()) {
Fail(program, "address operation has invalid memory metadata");
}
const auto& memory = program.memory_info[memory_index];
const auto kind = memory.kind;
if (kind == IR::ResourceKind::Scratch || kind == IR::ResourceKind::FlatLocal) {
if (program.scratch_dwords == 0) {
Fail(program, "scratch operation has no per-thread storage");
}
requirements.function_scratch = true;
if (kind == IR::ResourceKind::FlatLocal && program.stage != ShaderType::Compute &&
program.stage != ShaderType::Mesh) {
requirements.function_lds = true;
}
} else if (address_access == IR::AddressAccess::Write) {
Fail(program, "writable FLAT/GLOBAL addresses require GPU ownership tracking");
} else {
// Descriptor stores can alias coherent physical-address loads.
requirements.coherent_buffers |= memory.coherent;
}
}
if (IR::BufferAccessOf(inst.GetOpcode()) != IR::BufferAccess::None) {
const auto memory_index = inst.Flags<IR::MemoryFlags>().index;
if (memory_index >= program.memory_info.size()) {
Fail(program, "buffer operation has invalid memory metadata");
}
const auto& memory = program.memory_info[memory_index];
if (memory.kind == IR::ResourceKind::IndirectBuffer &&
inst.GetOpcode() != IR::ValueOpcode::ReadConstBuffer) {
requirements.subgroup_local_invocation_id = true;
}
if (memory.kind == IR::ResourceKind::Buffer) {
requirements.coherent_buffers |= memory.coherent;
if (memory.resource >= program.info.buffers.size()) {
Fail(program, "buffer operation has invalid resource metadata");
}
const auto bits = StorageBufferElementBits(program, memory);
requirements.buffer_u8 |= bits == 8u;
requirements.buffer_u16 |= bits == 16u;
if ((program.info.buffers[memory.resource].packed_stride & (1u << 20u)) != 0u) {
if (program.stage != ShaderType::Compute) {
Fail(program, "buffer ADD_TID is only valid for compute shaders");
}
requirements.subgroup_local_invocation_id = true;
}
}
}
const auto shared_access = IR::SharedAccessOf(inst.GetOpcode());
if (shared_access != IR::SharedAccess::None) {
const auto index = inst.Flags<IR::MemoryFlags>().index;
if (index >= program.memory_info.size()) {
Fail(program, "shared operation has invalid memory metadata");
}
const auto kind = program.memory_info[index].kind;
if (kind != IR::ResourceKind::Lds && kind != IR::ResourceKind::Gds) {
Fail(program, "shared operation has invalid resource kind");
}
if (shared_access == IR::SharedAccess::Atomic &&
IR::SharedComponentCount(inst.GetOpcode()) == 2u) {
if (kind != IR::ResourceKind::Lds || program.stage != ShaderType::Compute) {
Fail(program, "64-bit shared atomics require compute LDS");
}
requirements.shared_int64_atomics = true;
}
if (program.stage != ShaderType::Compute && program.stage != ShaderType::Mesh &&
kind == IR::ResourceKind::Lds) {
requirements.function_lds = true;
}
if (shared_access == IR::SharedAccess::Append ||
shared_access == IR::SharedAccess::Consume) {
requirements.subgroup_ballot = true;
requirements.subgroup_shuffle = true;
requirements.subgroup_local_invocation_id = true;
}
}
switch (inst.GetOpcode()) {
case IR::ValueOpcode::StoreCompletion: requirements.subgroup_barrier = true; break;
case IR::ValueOpcode::BvhIntersect: requirements.bvh = true; break;
case IR::ValueOpcode::ConditionRef:
requirements.subgroup_ballot |=
inst.Flags<CFG::BranchCondition>() != CFG::BranchCondition::ScalarInstruction;
break;
case IR::ValueOpcode::Ballot: requirements.subgroup_ballot = true; break;
case IR::ValueOpcode::DppMoveU32:
case IR::ValueOpcode::ReadFirstLane:
case IR::ValueOpcode::ReadLane: {
requirements.subgroup_ballot = true;
requirements.subgroup_shuffle = true;
if (inst.GetOpcode() == IR::ValueOpcode::DppMoveU32) {
requirements.subgroup_local_invocation_id = true;
}
break;
}
case IR::ValueOpcode::DppUpdateU32:
case IR::ValueOpcode::WriteLane: {
requirements.subgroup_ballot = true;
requirements.subgroup_local_invocation_id = true;
break;
}
case IR::ValueOpcode::Permlane16U32: {
requirements.subgroup_ballot = true;
requirements.subgroup_shuffle = true;
requirements.subgroup_local_invocation_id = true;
break;
}
case IR::ValueOpcode::SwizzleU32:
case IR::ValueOpcode::PermuteU32:
case IR::ValueOpcode::BpermuteU32: {
requirements.subgroup_ballot = true;
requirements.subgroup_shuffle = true;
requirements.subgroup_local_invocation_id = true;
break;
}
case IR::ValueOpcode::LaneId:
requirements.subgroup_local_invocation_id |=
program.stage != ShaderType::TessellationControl;
break;
case IR::ValueOpcode::ImageQueryLod: requirements.compute_derivatives = true; break;
case IR::ValueOpcode::ImageGatherRaw:
requirements.image_gather_extended = true;
break;
case IR::ValueOpcode::SetAttribute: {
const auto index = inst.Flags<IR::ExportFlags>().index;
if (index >= program.export_info.size()) {
Fail(program, "attribute export has invalid metadata");
}
if (program.stage == ShaderType::Pixel &&
program.export_info[index].vm) {
requirements.pixel_valid_mask = true;
}
break;
}
default: break;
}
}
}
return requirements;
}
std::vector<uint32_t> EmitProgram(const IR::Program& program,
ShaderStageInputInfo input_info) {
std::vector<uint32_t> EmitProgram(IR::Program& program, ShaderStageInputInfo input_info,
uint32_t push_data_start_dword) {
using namespace Emitter;
if (program.stage != ShaderType::Compute && program.stage != ShaderType::Vertex &&
@@ -352,12 +207,14 @@ std::vector<uint32_t> EmitProgram(const IR::Program& program,
Fail(program, "binary SPIR-V emitter received an unsupported shader stage");
}
if (!program.srt_plan_complete || !program.resource_tracking_complete ||
!program.shader_info_complete || !program.binding_layout_complete) {
!program.shader_info_complete) {
Fail(program, "SPIR-V emitter requires a fully planned native shader program");
}
ValidateNativeProgram(program, program.stage == ShaderType::Compute &&
input_info.compute != nullptr && input_info.compute->lds_storage);
if (program.info.buffers.size() > IR::ShaderInfo::MaxBuffers) {
Fail(program, "native shader exceeds the buffer resource limit");
}
IR::ValidateProgram(program, true);
program.bindings = {.push_data_start_dword = push_data_start_dword};
EmitterState state(program, input_info);
const auto* workgroup = ShaderWorkgroupInput(program.stage, input_info);
state.lane_count =
@@ -365,6 +222,8 @@ std::vector<uint32_t> EmitProgram(const IR::Program& program,
? 2u
: 1u;
DefineModule(state);
ValidateNativeProgram(program, program.stage == ShaderType::Compute &&
input_info.compute != nullptr && input_info.compute->lds_storage);
EmitProgram(state);
state.builder.AddEntryPoint(ExecutionModelForStage(state.program.stage), state.main_func,
"main", state.interface_variables);
src/graphics/shader/recompiler/backend/spirv/SpirvEmitter.h
+3 -3
View File
@@ -9,9 +9,9 @@
namespace Libs::Graphics::ShaderRecompiler::Spirv {
// Resets and writes IR backend IDs; do not emit the same program concurrently.
std::vector<uint32_t> EmitProgram(const IR::Program& program,
ShaderStageInputInfo input_info);
// Resets and writes backend IDs and descriptor indices; do not emit the same program concurrently.
std::vector<uint32_t> EmitProgram(IR::Program& program, ShaderStageInputInfo input_info,
uint32_t push_data_start_dword = 0);
} // namespace Libs::Graphics::ShaderRecompiler::Spirv
src/graphics/shader/recompiler/backend/spirv/spirvEmitterAnalysis.cpp
+6 -19
View File
@@ -59,21 +59,6 @@ uint32_t ConstantU32(EmitterState& state, uint32_t value);
std::abort();
}
uint32_t ResourceForDescriptor(const EmitterState& state, IR::DescriptorBindingKind kind,
uint32_t resource) {
const auto* descriptor = IR::FindBinding(state.program.bindings, kind);
if (descriptor == nullptr) {
ExitDescriptorBindingFailure(state, kind, resource, "descriptor group was not allocated");
}
const auto found =
std::find(descriptor->resources.begin(), descriptor->resources.end(), resource);
if (found == descriptor->resources.end()) {
ExitDescriptorBindingFailure(state, kind, resource,
"resource is absent from descriptor group");
}
return static_cast<uint32_t>(found - descriptor->resources.begin());
}
uint32_t DescriptorElementPointer(EmitterState& state, uint32_t result_ptr_type,
uint32_t variable_id, uint32_t array_index,
IR::DescriptorBindingKind kind, uint32_t resource,
@@ -162,13 +147,15 @@ uint32_t LoadImageDescriptor(EmitterState& state, uint32_t resource, uint32_t mi
}
uint32_t LoadSamplerDescriptor(EmitterState& state, uint32_t sampler) {
const auto array_index =
ResourceForDescriptor(state, IR::DescriptorBindingKind::Samplers, sampler);
if (sampler >= state.program.info.samplers.size()) {
ExitDescriptorBindingFailure(state, IR::DescriptorBindingKind::Samplers, sampler,
"sampler resource index is out of range");
}
const auto sampler_type = state.builder.Type(spv::OpTypeSampler);
const auto pointer_type =
state.builder.Type(spv::OpTypePointer, spv::StorageClassUniformConstant, sampler_type);
const auto pointer = DescriptorElementPointer(
state, pointer_type, state.sampler_variable, ConstantU32(state, array_index),
state, pointer_type, state.sampler_variable, ConstantU32(state, sampler),
IR::DescriptorBindingKind::Samplers, sampler, "sampler descriptor array was not emitted");
const auto sampler_id = state.builder.AllocateId();
state.builder.AddFunction(spv::OpLoad, sampler_type, sampler_id, pointer);
@@ -200,7 +187,7 @@ uint32_t ImageDescriptorPointer(EmitterState& state, uint32_t resource, uint32_t
const auto kind = IR::DescriptorBindingForImage(image);
EXIT_IF(!kind.has_value());
if (array_index == 0u) {
array_index = ConstantU32(state, ResourceForDescriptor(state, *kind, resource) + mip);
array_index = ConstantU32(state, image.descriptor_index + mip);
}
const auto pointer_type = state.builder.Type(
spv::OpTypePointer, spv::StorageClassUniformConstant, ImageType(state, image));
src/graphics/shader/recompiler/backend/spirv/spirvEmitterBvh.cpp
+1 -1
View File
@@ -223,7 +223,7 @@ uint32_t Boxes(EmitterState& s, uint32_t first, uint32_t second, uint32_t half,
} // namespace
void DefineBvhIntersect(EmitterState& s) {
if (!s.requirements.bvh) return;
if (!s.program.info.bvh) return;
const auto u = TypeU32(s), f = TypeF32(s), b = TypeBool(s), wide = TypeU64(s);
const auto vec4 = TypeU32Vector(s, 4), vec3 = TypeF32Vector(s, 3);
const auto signature = s.builder.Type(spv::OpTypeFunction, vec4, vec4, wide, f, vec3, vec3, vec3);
src/graphics/shader/recompiler/backend/spirv/spirvEmitterFlow.cpp
+2 -2
View File
@@ -10,7 +10,7 @@ namespace {
bool UserDataDwordIndex(const EmitterState& state, IR::ScalarReg reg, uint32_t& dword_index) {
const auto register_index = IR::RegIndex(reg);
const auto& registers = state.program.bindings.user_data_registers;
const auto& registers = state.program.info.user_data_registers;
const auto found = std::lower_bound(registers.begin(), registers.end(), register_index);
if (found == registers.end() || *found != register_index) {
return false;
@@ -472,7 +472,7 @@ void EmitSetAttribute(ValueEmitContext& ctx, const IR::Inst& inst) {
auto& state = ctx.state;
const auto& exp = ctx.Export(inst);
const auto exec = ctx.Arg(inst, 1);
if (state.program.stage == ShaderType::Pixel && exp.vm && state.requirements.pixel_valid_mask &&
if (state.program.stage == ShaderType::Pixel && exp.vm && state.program.info.pixel_valid_mask &&
state.pixel_valid_mask_variable != 0) {
const auto value = state.builder.AllocateId();
state.builder.AddFunction(spv::OpSelect, TypeU32(state), value, exec, ConstantU32(state, 1),
src/graphics/shader/recompiler/backend/spirv/spirvEmitterImage.cpp
+4 -7
View File
@@ -867,7 +867,6 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
const auto resource = image.indirect_resources[ordinal];
const auto& candidate = state.program.info.images[resource];
const auto kind = *IR::DescriptorBindingForImage(candidate);
const auto& binding = *IR::FindBinding(state.program.bindings, kind);
if (!runs.empty()) {
auto& run = runs.back();
const auto& first = state.program.info.images[run.resource];
@@ -875,17 +874,16 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
if (candidate.dimension == first.dimension && candidate.cube == first.cube &&
IR::DescriptorBindingForImage(first) == kind &&
candidate.mip_count == 1u && first.mip_count == 1u &&
(ordinal == 1u || (next_slot < binding.resources.size() &&
binding.resources[next_slot] == resource))) {
(ordinal == 1u || candidate.descriptor_index == next_slot)) {
// Native roots precede appended children; only the root may have a slot gap.
if (ordinal == 1u)
run.slot_bias = ResourceForDescriptor(state, kind, resource) - ordinal;
run.slot_bias = candidate.descriptor_index - ordinal;
++run.count;
continue;
}
}
runs.push_back({ordinal, 1u, resource,
ResourceForDescriptor(state, kind, resource) - ordinal});
candidate.descriptor_index - ordinal});
}
const auto EmitRun = [&](uint32_t index) {
const auto& run = runs[index];
@@ -893,8 +891,7 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
auto slot = Binary(state, spv::OpIAdd, TypeU32(state), selected,
ConstantU32(state, run.slot_bias));
if (run.first == 0u) {
const auto kind = *IR::DescriptorBindingForImage(image);
const auto root_slot = ResourceForDescriptor(state, kind, mem.resource);
const auto root_slot = image.descriptor_index;
if (root_slot != run.slot_bias) {
const auto is_root = Binary(state, spv::OpIEqual, TypeBool(state), selected,
ConstantU32(state, 0u));
src/graphics/shader/recompiler/backend/spirv/spirvEmitterInternal.h
+5 -32
View File
@@ -6,7 +6,6 @@
#include "graphics/shader/recompiler/BufferFormat.h"
#include "graphics/shader/recompiler/backend/spirv/SpirvBuilder.h"
#include "graphics/shader/recompiler/ir/ShaderIR.h"
#include "graphics/shader/recompiler/ir/passes/BindingLayout.h"
#include "graphics/shader/recompiler/ir/passes/ResourceMaterialization.h"
#include <algorithm>
@@ -58,42 +57,21 @@ constexpr std::array<ImageDimensionInfo, 7> ImageDimensions {{
const ImageDimensionInfo& ImageDimensionInfoFor(ImageDimension dimension);
struct SpirvRequirements {
bool bvh = false;
bool subgroup_ballot = false;
bool subgroup_barrier = false;
bool subgroup_shuffle = false;
bool subgroup_local_invocation_id = false;
bool compute_derivatives = false;
bool image_gather_extended = false;
bool function_lds = false;
bool function_scratch = false;
bool pixel_valid_mask = false;
bool buffer_int64_atomics = false;
bool buffer_u8 = false;
bool buffer_u16 = false;
bool shared_int64_atomics = false;
bool coherent_buffers = false;
bool float64 = false;
};
SpirvRequirements AnalyzeProgramRequirements(const IR::Program& program);
struct EmitterState {
EmitterState(const IR::Program& program_, ShaderStageInputInfo input_info_)
EmitterState(IR::Program& program_, ShaderStageInputInfo input_info_)
: builder(program_.stage == ShaderType::Mesh ? 0x00010400u : 0x00010300u),
program(program_), input_info(input_info_),
requirements(AnalyzeProgramRequirements(program_)) {
program(program_), input_info(input_info_) {
if (ShaderWorkgroupInput(program.stage, input_info) != nullptr) {
lds_storage_class = spv::StorageClassWorkgroup;
}
if (IR::FindBinding(program.bindings, IR::DescriptorBindingKind::SharedMemory) != nullptr) {
if (program.info.uses_lds && program.stage == ShaderType::Compute &&
input_info.compute != nullptr && input_info.compute->lds_storage) {
lds_storage_class = spv::StorageClassStorageBuffer;
}
}
Builder builder;
const IR::Program& program;
IR::Program& program;
ShaderStageInputInfo input_info;
uint32_t void_type = 0;
uint32_t bool_type = 0;
@@ -113,7 +91,6 @@ struct EmitterState {
uint32_t tess_inner_variable = 0;
uint32_t tess_patch_base = 0;
const SpirvRequirements requirements;
uint32_t lane_count = 1;
uint32_t lane_half = 0;
uint32_t storage_buffer_variable = 0;
@@ -313,9 +290,6 @@ uint32_t EmitSubgroupLocalInvocationId(EmitterState& state);
IR::DescriptorBindingKind kind, uint32_t resource,
const char* reason);
uint32_t ResourceForDescriptor(const EmitterState& state, IR::DescriptorBindingKind kind,
uint32_t resource);
uint32_t DescriptorElementPointer(EmitterState& state, uint32_t result_ptr_type,
uint32_t variable_id, uint32_t array_index,
IR::DescriptorBindingKind kind, uint32_t resource,
@@ -410,7 +384,6 @@ uint32_t EmitShaderDataDwordLoad(EmitterState& state, uint32_t dword_index);
uint32_t StorageBufferPackedStride(const EmitterState& state, const IR::MemoryInfo& mem);
Prospero::BufferFormat StorageBufferFormat(const EmitterState& state, const IR::MemoryInfo& mem);
uint32_t StorageBufferElementBits(const IR::Program& program, const IR::MemoryInfo& mem);
void EmitMemoryOffsets(EmitterState& state);
src/graphics/shader/recompiler/backend/spirv/spirvEmitterMemory.cpp
+1 -2
View File
@@ -251,8 +251,7 @@ uint32_t LoadBda(ValueEmitContext& ctx, uint32_t address, uint32_t active, uint3
uint32_t ByteAddress(ValueEmitContext& ctx, const IR::Inst& inst, const IR::MemoryInfo& mem) {
if (mem.kind == IR::ResourceKind::Buffer) {
const auto resource = ResourceForDescriptor(ctx.state, IR::DescriptorBindingKind::Buffers,
mem.resource);
const auto resource = ctx.state.program.info.buffers.at(mem.resource).descriptor_index;
const auto prefix = ctx.state.memory_byte_offsets[resource];
const auto address = Binary(ctx.state, spv::OpIAdd, TypeU32(ctx.state),
BufferByteAddress(ctx, inst, mem), prefix);
src/graphics/shader/recompiler/backend/spirv/spirvEmitterMemoryHelpers.cpp
+6 -14
View File
@@ -102,17 +102,9 @@ Prospero::BufferFormat StorageBufferFormat(const EmitterState& state, const IR::
return state.program.info.buffers[mem.resource].descriptor_format;
}
uint32_t StorageBufferElementBits(const IR::Program& program, const IR::MemoryInfo& mem) {
if (!mem.formatted) return mem.data_bits;
const auto format = mem.typed ? Format::DecodeTBufferFormat(mem.data_format, mem.number_format)
: program.info.buffers[mem.resource].descriptor_format;
const auto info = Format::GetFormatInfo(format);
return info.type == Format::ComponentType::Unknown || info.packed_bitfield
? 32u : info.component_bits[0];
}
void EmitMemoryOffsets(EmitterState& state) {
for (uint32_t i = 0; i < state.program.bindings.memory_offset_count; i++) {
const auto count = state.program.bindings.descriptor_counts[static_cast<size_t>(IR::DescriptorBindingKind::Buffers)];
for (uint32_t i = 0; i < count; i++) {
const auto word =
EmitShaderDataDwordLoad(state, state.program.bindings.memory_offset_dword + i / 4u);
const auto shift = ConstantU32(state, (i % 4u) * 8u);
@@ -146,7 +138,7 @@ void EnsureLdsStorage(EmitterState& state) {
state.builder.AddAnnotation(spv::OpDecorate, variable, spv::DecorationAliased);
return variable;
};
if (state.requirements.shared_int64_atomics) {
if (state.program.info.shared_int64_atomics) {
state.lds_variable = define(TypeU32(state), 4u);
state.lds_u64_variable = define(TypeU64(state), 8u);
state.builder.AddName(state.lds_u64_variable, "lds_qwords");
@@ -167,7 +159,7 @@ MemoryResourceAccess PrepareStorageBufferResourceAccess(EmitterState& state,
"storage buffer descriptor array was not emitted");
}
const auto array_index =
ResourceForDescriptor(state, IR::DescriptorBindingKind::Buffers, mem.resource);
state.program.info.buffers.at(mem.resource).descriptor_index;
MemoryResourceAccess access {
.kind = mem.kind,
.memory_access = mem.coherent ? spv::MemoryAccessVolatileMask : spv::MemoryAccessMaskNone};
@@ -214,7 +206,7 @@ MemoryResourceAccess PrepareMemoryResourceAccess(EmitterState& state, const IR::
case IR::ResourceKind::ScalarBuffer:
case IR::ResourceKind::Buffer: {
const auto bits = mem.kind == IR::ResourceKind::Buffer
? StorageBufferElementBits(state.program, mem) : 32u;
? IR::StorageBufferElementBits(state.program, mem) : 32u;
const auto variable = bits == 8u ? state.storage_buffer_u8_variable
: bits == 16u ? state.storage_buffer_u16_variable
: state.storage_buffer_variable;
@@ -245,7 +237,7 @@ uint32_t EmitMemoryElementPointer(EmitterState& state, const MemoryResourceAcces
TypeStorageBufferElementPointer(state, 32));
}
const auto pointer = state.builder.AllocateId();
if (access.kind == IR::ResourceKind::Lds && state.requirements.shared_int64_atomics) {
if (access.kind == IR::ResourceKind::Lds && state.program.info.shared_int64_atomics) {
state.builder.AddFunction(spv::OpAccessChain, TypeU32ElementPointer(state, storage_class),
pointer, access.object_pointer, ConstantU32(state, 0), index);
return pointer;
src/graphics/shader/recompiler/backend/spirv/spirvEmitterModule.cpp
+75 -35
View File
@@ -200,6 +200,43 @@ uint32_t F32ArrayType(EmitterState& state, uint32_t count) {
}
void DefineDescriptors(EmitterState& state) {
using Kind = IR::DescriptorBindingKind;
auto& info = state.program.info;
auto& bindings = state.program.bindings;
auto& counts = bindings.descriptor_counts;
auto& buffer_count = counts[static_cast<size_t>(Kind::Buffers)];
for (uint32_t index = 0; index < info.buffers.size(); ++index) {
info.buffers[index].descriptor_index = (info.live_buffers & (uint64_t{1} << index)) != 0
? buffer_count++ : UINT32_MAX;
}
std::array<const IR::ImageResource*, IR::ImageBindingCount> image_types {};
for (auto& image: info.images) {
const auto kind = IR::DescriptorBindingForImage(image);
if (!kind.has_value() || image.mip_count == 0u ||
(image.mip_mode != IR::ImageMipMode::Dynamic && image.mip_count != 1u)) {
EXIT("shader image has an invalid descriptor type or mip count");
}
auto& count = counts[static_cast<size_t>(*kind)];
image.descriptor_index = count;
if (count == 0) image_types[IR::ImageBindingIndex(*kind)] = &image;
count += image.mip_count;
}
counts[static_cast<size_t>(Kind::Samplers)] = static_cast<uint32_t>(info.samplers.size());
counts[static_cast<size_t>(Kind::Gds)] = info.uses_gds;
counts[static_cast<size_t>(Kind::SharedMemory)] =
info.uses_lds && state.lds_storage_class == spv::StorageClassStorageBuffer;
counts[static_cast<size_t>(Kind::BdaPagetable)] = info.uses_dma;
counts[static_cast<size_t>(Kind::FaultBuffer)] = info.uses_dma;
counts[static_cast<size_t>(Kind::FlattenedSrt)] = info.uses_flattened_srt;
bindings.memory_offset_dword = static_cast<uint32_t>(info.user_data_registers.size());
if (info.uses_dispatch_threads) {
bindings.dispatch_thread_dword = bindings.memory_offset_dword;
bindings.memory_offset_dword += 3u;
}
bindings.push_data_start_dword =
IR::PushData::StartFor(bindings.push_data_start_dword, bindings.ShaderDataDwords());
counts[static_cast<size_t>(Kind::ShaderData)] =
bindings.ShaderDataDwords() != 0 && !bindings.UsesPushData();
if (state.program.bindings.UsesPushData() || state.program.stage == ShaderType::Mesh) {
const auto type = PushConstantBlockType(state);
state.push_constant_variable = state.builder.DefineGlobalVariable(
@@ -207,7 +244,10 @@ void DefineDescriptors(EmitterState& state) {
state.builder.AddName(type, "BufferResource");
state.builder.AddName(state.push_constant_variable, "vsharp");
}
for (const auto& binding: state.program.bindings.descriptors) {
for (uint32_t index = 0; index < counts.size(); ++index) {
if (counts[index] == 0) continue;
bindings.descriptor_mask |= uint64_t{1} << index;
const auto kind = static_cast<Kind>(index);
const auto Define = [&](uint32_t type, const char* name,
spv::StorageClass storage = spv::StorageClassStorageBuffer) {
const auto variable =
@@ -215,27 +255,27 @@ void DefineDescriptors(EmitterState& state) {
state.builder.AddName(variable, name);
state.builder.AddAnnotation(spv::OpDecorate, variable, spv::DecorationDescriptorSet, 0);
state.builder.AddAnnotation(spv::OpDecorate, variable, spv::DecorationBinding,
IR::NativeBinding(state.program.stage, binding.kind));
IR::NativeBinding(state.program.stage, kind));
return variable;
};
const auto ArrayType = [&](uint32_t type) {
return state.builder.Type(
spv::OpTypeArray, type,
ConstantU32(state, static_cast<uint32_t>(binding.resources.size())));
ConstantU32(state, counts[index]));
};
switch (binding.kind) {
switch (kind) {
case IR::DescriptorBindingKind::Buffers:
state.storage_buffer_variable =
Define(ArrayType(StorageBufferType(state)), "buffers");
if (state.requirements.buffer_u8) {
if (state.program.info.buffer_u8) {
state.storage_buffer_u8_variable =
Define(ArrayType(StorageBufferType(state, 8)), "buffers_u8");
}
if (state.requirements.buffer_u16) {
if (state.program.info.buffer_u16) {
state.storage_buffer_u16_variable =
Define(ArrayType(StorageBufferType(state, 16)), "buffers_u16");
}
if (state.requirements.buffer_int64_atomics) {
if (state.program.info.buffer_int64_atomics) {
state.storage_buffer_u64_variable =
Define(ArrayType(StorageBufferType(state, 64)), "buffers_u64");
}
@@ -248,7 +288,7 @@ void DefineDescriptors(EmitterState& state) {
}
// RDNA2 stores publish to L2 even without GLC; every alias of the buffer
// must participate in visibility for cache-bypassing polling loads.
if (state.requirements.coherent_buffers) {
if (state.program.info.coherent_buffers) {
state.builder.AddAnnotation(spv::OpDecorate, variable, spv::DecorationCoherent);
}
}
@@ -276,7 +316,7 @@ void DefineDescriptors(EmitterState& state) {
case IR::DescriptorBindingKind::SharedMemory:
state.lds_variable = Define(StorageBufferType(state), "lds_dwords");
state.builder.AddAnnotation(spv::OpDecorate, state.lds_variable, spv::DecorationCoherent);
if (state.requirements.shared_int64_atomics) {
if (state.program.info.shared_int64_atomics) {
state.lds_u64_variable = Define(StorageBufferType(state, 64), "lds_qwords");
state.builder.AddAnnotation(spv::OpDecorate, state.lds_u64_variable, spv::DecorationCoherent);
for (const auto variable: {state.lds_variable, state.lds_u64_variable}) {
@@ -285,11 +325,11 @@ void DefineDescriptors(EmitterState& state) {
}
break;
default: {
EXIT_IF(IR::ImageBindingResourceClass(binding.kind) ==
EXIT_IF(IR::ImageBindingResourceClass(kind) ==
IR::ImageResourceClass::None);
const auto& image = state.program.info.images.at(binding.resources.front());
const auto name = "image_" + std::to_string(static_cast<uint32_t>(binding.kind));
state.image_variables[IR::ImageBindingIndex(binding.kind)] =
const auto& image = *image_types[IR::ImageBindingIndex(kind)];
const auto name = "image_" + std::to_string(static_cast<uint32_t>(kind));
state.image_variables[IR::ImageBindingIndex(kind)] =
Define(ArrayType(ImageType(state, image)), name.c_str(),
spv::StorageClassUniformConstant);
if (image.dimension == ImageDimension::Dim1D ||
@@ -437,7 +477,7 @@ void DefineInputs(EmitterState& state) {
state.inputs.push_back({{kind, 0, components, name}});
}
};
if (IR::FindBinding(state.program.bindings, IR::DescriptorBindingKind::SharedMemory) != nullptr) {
if (state.program.info.uses_lds && state.lds_storage_class == spv::StorageClassStorageBuffer) {
add_builtin(IR::StageInputKind::WorkgroupId, 3, "gl_WorkGroupID");
add_builtin(IR::StageInputKind::NumWorkgroups, 3, "gl_NumWorkGroups");
}
@@ -524,7 +564,7 @@ void DefineInputs(EmitterState& state) {
builtin);
}
}
if (state.requirements.subgroup_local_invocation_id) {
if (state.program.info.subgroup_local_invocation_id) {
const auto variable = DefineInterfaceVariable(state, TypeU32(state), spv::StorageClassInput,
"gl_SubgroupInvocationID");
state.subgroup_local_invocation_id_variable = variable;
@@ -647,10 +687,10 @@ void DefineModule(EmitterState& state) {
DefineOutputs(state);
DefineTessellationInterfaces(state);
DefineDescriptors(state);
if (state.requirements.function_lds) {
if (state.program.info.function_lds) {
state.lds_variable = state.builder.AllocateId();
}
if (state.requirements.function_scratch) {
if (state.program.info.function_scratch) {
for (uint32_t half = 0; half < state.lane_count; half++) {
state.scratch_variable[half] = state.builder.AllocateId();
}
@@ -675,11 +715,11 @@ void DefineModule(EmitterState& state) {
state.builder.RequireCapability(spv::CapabilityShader);
state.builder.RequireCapability(spv::CapabilitySignedZeroInfNanPreserve);
if (state.requirements.buffer_u8) {
if (state.program.info.buffer_u8) {
state.builder.RequireExtension("SPV_KHR_8bit_storage");
state.builder.RequireCapability(spv::CapabilityStorageBuffer8BitAccess);
}
if (state.requirements.buffer_u16) {
if (state.program.info.buffer_u16) {
state.builder.RequireCapability(spv::CapabilityStorageBuffer16BitAccess);
}
if (state.program.info.uses_dma) {
@@ -687,11 +727,11 @@ void DefineModule(EmitterState& state) {
state.builder.RequireCapability(spv::CapabilityPhysicalStorageBufferAddresses);
state.builder.RequireExtension("SPV_KHR_physical_storage_buffer");
}
if (state.requirements.buffer_int64_atomics || state.requirements.shared_int64_atomics) {
if (state.program.info.buffer_int64_atomics || state.program.info.shared_int64_atomics) {
state.builder.RequireCapability(spv::CapabilityInt64);
state.builder.RequireCapability(spv::CapabilityInt64Atomics);
}
if (state.requirements.shared_int64_atomics &&
if (state.program.info.shared_int64_atomics &&
state.lds_storage_class == spv::StorageClassWorkgroup) {
state.builder.RequireVersion(0x00010400u);
state.builder.RequireExtension("SPV_KHR_workgroup_memory_explicit_layout");
@@ -714,21 +754,21 @@ void DefineModule(EmitterState& state) {
if (InputVariableForKind(state, IR::StageInputKind::SampleId) != 0) {
state.builder.RequireCapability(spv::CapabilitySampleRateShading);
}
if (state.requirements.image_gather_extended) {
if (state.program.info.image_gather_extended) {
state.builder.RequireCapability(spv::CapabilityImageGatherExtended);
}
if (state.lane_count == 2 || state.requirements.subgroup_barrier ||
state.requirements.subgroup_ballot || state.requirements.subgroup_shuffle ||
state.requirements.subgroup_local_invocation_id) {
if (state.lane_count == 2 || state.program.info.subgroup_barrier ||
state.program.info.subgroup_ballot || state.program.info.subgroup_shuffle ||
state.program.info.subgroup_local_invocation_id) {
state.builder.RequireCapability(spv::CapabilityGroupNonUniform);
}
if (state.lane_count == 2 || state.requirements.subgroup_ballot) {
if (state.lane_count == 2 || state.program.info.subgroup_ballot) {
state.builder.RequireCapability(spv::CapabilityGroupNonUniformBallot);
}
if (state.requirements.subgroup_shuffle) {
if (state.program.info.subgroup_shuffle) {
state.builder.RequireCapability(spv::CapabilityGroupNonUniformShuffle);
}
if (state.requirements.compute_derivatives && state.program.stage == ShaderType::Compute) {
if (state.program.info.compute_derivatives && state.program.stage == ShaderType::Compute) {
state.builder.RequireCapability(spv::CapabilityComputeDerivativeGroupQuadsKHR);
state.builder.RequireExtension("SPV_KHR_compute_shader_derivatives");
}
@@ -751,7 +791,7 @@ void DefineModule(EmitterState& state) {
// contract prevents host compilers from treating synthesized IEEE values as finite.
state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeSignedZeroInfNanPreserve,
32u);
if (state.requirements.float64) {
if (state.program.info.float64) {
EXIT_NOT_IMPLEMENTED(state.program.stage == ShaderType::Compute &&
state.input_info.compute->float_mode != 0xc0);
// Use native rounding for MODE=0xc0, consistent with ordinary FP32 arithmetic.
@@ -763,8 +803,8 @@ void DefineModule(EmitterState& state) {
// state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeDenormPreserve, 64u);
}
if (const auto* cs = ShaderWorkgroupInput(state.program.stage, state.input_info)) {
uint32_t local_x = state.requirements.compute_derivatives ? 2u : 1u;
uint32_t local_y = state.requirements.compute_derivatives ? 2u : 1u;
uint32_t local_x = state.program.info.compute_derivatives ? 2u : 1u;
uint32_t local_y = state.program.info.compute_derivatives ? 2u : 1u;
uint32_t local_z = 1u;
local_x = cs->threads_num[0] != 0u ? cs->threads_num[0] : local_x;
local_y = cs->threads_num[1] != 0u ? cs->threads_num[1] : local_y;
@@ -772,7 +812,7 @@ void DefineModule(EmitterState& state) {
if (state.lane_count == 2) {
local_x = ((local_x * local_y * local_z + 63u) / 64u) * 32u;
local_y = local_z = 1u;
if (state.requirements.compute_derivatives) {
if (state.program.info.compute_derivatives) {
local_y = local_x / 2u;
local_x = 2u;
}
@@ -791,14 +831,14 @@ void DefineModule(EmitterState& state) {
state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeEarlyFragmentTests);
}
}
if (state.requirements.compute_derivatives && state.program.stage == ShaderType::Compute) {
if (state.program.info.compute_derivatives && state.program.stage == ShaderType::Compute) {
state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeDerivativeGroupQuadsKHR);
}
state.builder.AddName(state.main_func, "main");
if (state.requirements.function_lds) {
if (state.program.info.function_lds) {
state.builder.AddName(state.lds_variable, "lds_dwords");
}
if (state.requirements.function_scratch) {
if (state.program.info.function_scratch) {
for (uint32_t half = 0; half < state.lane_count; half++) {
state.builder.AddName(state.scratch_variable[half], "scratch_dwords");
}
src/graphics/shader/recompiler/backend/spirv/spirvEmitterProgram.cpp
+7 -12
View File
@@ -618,7 +618,7 @@ void EmitProgram(EmitterState& state) {
high.half = 1;
}
std::optional<DispatcherFunctionState> dispatcher;
if (state.program.stage == ShaderType::Pixel && state.requirements.pixel_valid_mask) {
if (state.program.stage == ShaderType::Pixel && state.program.info.pixel_valid_mask) {
state.pixel_valid_mask_variable = state.builder.AllocateId();
state.builder.AddName(state.pixel_valid_mask_variable, "pixel_valid_mask_active");
}
@@ -686,28 +686,23 @@ void EmitProgram(EmitterState& state) {
}
DefineGetBdaPointer(state);
DefineBvhIntersect(state);
for (const auto* block: program.blocks) {
if (std::ranges::any_of(*block, [](const IR::Inst& inst) {
return inst.GetOpcode() == IR::ValueOpcode::SwizzleU32;
})) {
ctx.scratch_u32_variable = state.builder.AllocateId();
if (state.lane_count == 2) {
high.scratch_u32_variable = state.builder.AllocateId();
}
break;
if (program.info.uses_swizzle) {
ctx.scratch_u32_variable = state.builder.AllocateId();
if (state.lane_count == 2) {
high.scratch_u32_variable = state.builder.AllocateId();
}
}
state.builder.AddFunction(spv::OpFunction, TypeVoid(state),
state.mesh_guest_func != 0 ? state.mesh_guest_func : state.main_func,
spv::FunctionControlMaskNone, TypeFunction(state));
EmitLabel(state, state.entry_label);
if (state.requirements.function_lds) {
if (state.program.info.function_lds) {
state.builder.AddFunction(
spv::OpVariable,
TypeU32ArrayPointer(state, spv::StorageClassFunction, LdsDwordCount(state)),
state.lds_variable, spv::StorageClassFunction);
}
if (state.requirements.function_scratch) {
if (state.program.info.function_scratch) {
for (uint32_t half = 0; half < state.lane_count; half++) {
state.builder.AddFunction(
spv::OpVariable,
src/graphics/shader/recompiler/ir/Program.cpp
+11
View File
@@ -1,5 +1,6 @@
#include "common/assert.h"
#include "graphics/shader/recompiler/ir/ShaderIR.h"
#include "graphics/shader/recompiler/BufferFormat.h"
#include <fmt/format.h>
#include <map>
@@ -202,6 +203,16 @@ Value ResolveActiveU32(Value value, Value active) {
return {};
}
uint32_t StorageBufferElementBits(const Program& program, const MemoryInfo& memory) {
if (!memory.formatted) return memory.data_bits;
const auto format = memory.typed
? Format::DecodeTBufferFormat(memory.data_format, memory.number_format)
: program.info.buffers[memory.resource].descriptor_format;
const auto info = Format::GetFormatInfo(format);
return info.type == Format::ComponentType::Unknown || info.packed_bitfield
? 32u : info.component_bits[0];
}
bool HasShaderMemoryWrites(const Program& program) {
for (const auto* block: program.blocks) {
for (const auto& inst: *block) {
src/graphics/shader/recompiler/ir/ShaderIR.h
+40 -21
View File
@@ -104,6 +104,7 @@ struct BufferResource {
static constexpr uint32_t NoIndirectBuffer = UINT32_MAX;
uint32_t source = 0;
uint32_t descriptor_index = UINT32_MAX;
uint32_t first_use_pc = 0;
uint32_t max_byte_extent = 0;
uint32_t packed_stride = 0;
@@ -131,6 +132,7 @@ struct ImageResource {
static constexpr uint32_t NoIndirectImage = UINT32_MAX;
uint32_t source = 0;
uint32_t descriptor_index = UINT32_MAX;
uint32_t first_use_pc = 0;
ImageResourceClass resource_class = ImageResourceClass::None;
Prospero::TextureNumericClass numeric_class = Prospero::TextureNumericClass::Unsupported;
@@ -424,23 +426,16 @@ DescriptorBindingForImage(const ImageResource& image) {
return static_cast<DescriptorBindingKind>(base + dimension);
}
struct DescriptorBinding {
DescriptorBindingKind kind = DescriptorBindingKind::Buffers;
std::vector<uint32_t> resources;
bool operator==(const DescriptorBinding& other) const = default;
};
struct BindingLayout {
uint32_t push_data_start_dword = PushData::NoStart;
uint32_t dispatch_thread_dword = PushData::NoStart;
uint32_t memory_offset_dword = 0;
uint32_t memory_offset_count = 0;
std::vector<uint32_t> user_data_registers;
std::vector<DescriptorBinding> descriptors;
struct Bindings {
std::array<uint32_t, static_cast<size_t>(DescriptorBindingKind::Count)> descriptor_counts {};
uint64_t descriptor_mask = 0;
uint32_t push_data_start_dword = PushData::NoStart;
uint32_t dispatch_thread_dword = PushData::NoStart;
uint32_t memory_offset_dword = 0;
[[nodiscard]] uint32_t ShaderDataDwords() const {
return memory_offset_dword + (memory_offset_count + 3u) / 4u;
return memory_offset_dword +
(descriptor_counts[static_cast<size_t>(DescriptorBindingKind::Buffers)] + 3u) / 4u;
}
[[nodiscard]] bool UsesPushData() const {
return push_data_start_dword != PushData::NoStart;
@@ -451,7 +446,7 @@ struct BindingLayout {
}
}
bool operator==(const BindingLayout& other) const = default;
bool operator==(const Bindings& other) const = default;
};
struct ShaderInfo {
@@ -466,11 +461,34 @@ struct ShaderInfo {
std::vector<SampledResourcePair> sampled_pairs;
std::vector<StageInput> inputs;
std::vector<StageOutput> outputs;
std::vector<uint32_t> user_data_registers;
uint64_t live_buffers = 0;
std::array<uint8_t, 32> vertex_fetch_components {};
int32_t vertex_offset_sgpr = -1;
int32_t instance_offset_sgpr = -1;
bool has_bitwise_xor = false;
bool uses_dma = false;
bool has_bitwise_xor = false;
bool uses_dma = false;
bool uses_swizzle = false;
bool uses_lds = false;
bool uses_gds = false;
bool uses_flattened_srt = false;
bool uses_dispatch_threads = false;
bool bvh = false;
bool subgroup_ballot = false;
bool subgroup_barrier = false;
bool subgroup_shuffle = false;
bool subgroup_local_invocation_id = false;
bool compute_derivatives = false;
bool image_gather_extended = false;
bool function_lds = false;
bool function_scratch = false;
bool pixel_valid_mask = false;
bool buffer_int64_atomics = false;
bool buffer_u8 = false;
bool buffer_u16 = false;
bool shared_int64_atomics = false;
bool coherent_buffers = false;
bool float64 = false;
bool operator==(const ShaderInfo& other) const = default;
};
@@ -542,7 +560,7 @@ struct CompiledShaderInfo {
uint32_t param_export_mask = 0;
bool has_address_writes = false;
ShaderInfo info;
BindingLayout bindings;
Bindings bindings;
};
struct UniformFillPlan {
@@ -622,11 +640,12 @@ struct Program: ResourcePlan {
std::vector<ExportInfo> export_info;
bool has_address_writes = false;
bool shader_info_complete = false;
BindingLayout bindings;
bool binding_layout_complete = false;
Bindings bindings;
};
uint32_t StorageBufferElementBits(const Program& program, const MemoryInfo& memory);
std::string ProgramToString(const Program& program);
bool HasShaderMemoryWrites(const Program& program);
src/graphics/shader/recompiler/ir/passes/BindingLayout.cpp
-197
View File
@@ -1,197 +0,0 @@
#include "graphics/shader/recompiler/ir/passes/BindingLayout.h"
#include "common/assert.h"
#include "graphics/shader/recompiler/ir/ShaderIR.h"
#include <algorithm>
#include <array>
#include <utility>
namespace Libs::Graphics::ShaderRecompiler::IR {
namespace {
[[noreturn]] void BindingFail(const char* message) {
EXIT("shader binding layout failed: %s", message);
std::abort();
}
void CollectShaderData(const Program& program, BindingLayout& layout) {
std::array<bool, NumScalarRegs> registers {};
bool uses_dispatch_threads = false;
for (const auto* block: program.blocks) {
for (const auto& inst: *block) {
if (!inst.HasUses()) {
continue;
}
uses_dispatch_threads |= inst.GetOpcode() == ValueOpcode::GetDispatchThreadExtent;
if (inst.GetOpcode() != ValueOpcode::GetUserData) {
continue;
}
if (inst.Arg(0).GetType() != Type::ScalarReg) {
BindingFail("typed shader contains an invalid user-data register");
}
const auto index = RegIndex(inst.Arg(0).ScalarRegister());
if (index >= NumScalarRegs) {
BindingFail("typed shader contains an invalid user-data register");
}
registers[index] = true;
}
}
for (uint32_t index = 0; index < registers.size(); index++) {
if (registers[index]) {
layout.user_data_registers.push_back(index);
}
}
layout.memory_offset_dword = static_cast<uint32_t>(layout.user_data_registers.size());
if (uses_dispatch_threads) {
layout.dispatch_thread_dword = layout.memory_offset_dword;
layout.memory_offset_dword += 3u;
}
}
void AddBinding(BindingLayout& layout, DescriptorBindingKind kind,
std::vector<uint32_t> resources = {}) {
layout.descriptors.push_back({kind, std::move(resources)});
}
} // namespace
SharedMemoryResources CollectMemoryResources(const Program& program, std::vector<uint32_t>& buffers) {
std::array<bool, ShaderInfo::MaxBuffers> live_buffers {};
SharedMemoryResources shared;
for (const auto* block: program.blocks) {
for (const auto& inst: *block) {
const auto op = inst.GetOpcode();
if (BufferAccessOf(op) == BufferAccess::None &&
SharedAccessOf(op) == SharedAccess::None &&
AddressOpcodeInfoOf(op).access == AddressAccess::None) {
continue;
}
const auto index = inst.Flags<MemoryFlags>().index;
if (index >= program.memory_info.size()) {
BindingFail("typed shader contains invalid memory metadata");
}
const auto& memory = program.memory_info[index];
if (memory.planning_only) {
continue;
}
shared.lds |= memory.kind == ResourceKind::FlatLocal;
if (SharedAccessOf(op) != SharedAccess::None) {
if (memory.kind != ResourceKind::Lds && memory.kind != ResourceKind::Gds) {
BindingFail("typed shader contains invalid shared-memory metadata");
}
shared.gds |= memory.kind == ResourceKind::Gds;
shared.lds |= memory.kind == ResourceKind::Lds;
} else if (memory.kind == ResourceKind::Buffer || memory.kind == ResourceKind::ScalarBuffer) {
EXIT_IF(memory.resource >= program.info.buffers.size());
live_buffers.at(memory.resource) = true;
for (const auto child: program.info.buffers[memory.resource].indirect_resources) {
live_buffers.at(child) = true;
}
}
}
}
for (uint32_t i = 0; i < program.info.buffers.size(); i++) {
if (live_buffers.at(i)) {
buffers.push_back(i);
}
}
return shared;
}
bool UsesFlattenedSrt(const Program& program) {
const auto uses_mapping = [](const auto& resource) {
return resource.indirect_root != UINT32_MAX;
};
return std::ranges::any_of(program.blocks, [](const Block* block) {
return std::ranges::any_of(*block, [](const Inst& inst) {
return inst.GetOpcode() == ValueOpcode::ReadConst;
});
}) || std::ranges::any_of(program.info.buffers, uses_mapping) ||
std::ranges::any_of(program.info.images, uses_mapping);
}
void AllocateBindings(Program& program, uint32_t push_data_start_dword, bool lds_storage) {
if (!program.shader_info_complete || program.binding_layout_complete) {
EXIT("shader binding layout failed: %s", !program.shader_info_complete
? "shader info is not ready"
: "binding layout already allocated");
}
BindingLayout next;
std::vector<uint32_t> buffers;
const auto shared = CollectMemoryResources(program, buffers);
CollectShaderData(program, next);
next.memory_offset_count = static_cast<uint32_t>(buffers.size());
next.push_data_start_dword =
PushData::StartFor(push_data_start_dword, next.ShaderDataDwords());
if (!buffers.empty()) {
// Draw binding accesses this first group directly when memory_offset_count is nonzero.
AddBinding(next, DescriptorBindingKind::Buffers, std::move(buffers));
}
std::array<std::vector<uint32_t>, ImageBindingCount> image_groups;
for (uint32_t i = 0; i < program.info.images.size(); i++) {
const auto kind = DescriptorBindingForImage(program.info.images[i]);
if (!kind.has_value()) {
EXIT("shader binding layout failed: image %u has an invalid binding class", i);
}
const auto group = ImageBindingIndex(*kind);
if (group >= image_groups.size()) {
EXIT("shader binding layout failed: image %u has an unmapped binding class", i);
}
auto& resources = image_groups[group];
const auto dynamic = program.info.images[i].mip_mode == ImageMipMode::Dynamic;
const auto count = dynamic ? program.info.images[i].mip_count : 1u;
if (count == 0u || (!dynamic && program.info.images[i].mip_count != 1u)) {
EXIT("shader binding layout failed: image %u has invalid specialized mip count %u", i,
program.info.images[i].mip_count);
}
resources.insert(resources.end(), count, i);
}
for (uint32_t i = 0; i < image_groups.size(); i++) {
if (!image_groups[i].empty()) {
AddBinding(next, static_cast<DescriptorBindingKind>(FirstImageBinding + i),
std::move(image_groups[i]));
}
}
if (!program.info.samplers.empty()) {
std::vector<uint32_t> resources(program.info.samplers.size());
for (uint32_t i = 0; i < resources.size(); i++) {
resources[i] = i;
}
AddBinding(next, DescriptorBindingKind::Samplers, std::move(resources));
}
if (shared.gds) {
AddBinding(next, DescriptorBindingKind::Gds);
}
if (shared.lds && lds_storage) {
AddBinding(next, DescriptorBindingKind::SharedMemory);
}
if (program.info.uses_dma) {
AddBinding(next, DescriptorBindingKind::BdaPagetable);
AddBinding(next, DescriptorBindingKind::FaultBuffer);
}
if (UsesFlattenedSrt(program)) {
AddBinding(next, DescriptorBindingKind::FlattenedSrt);
}
if (next.ShaderDataDwords() != 0 && !next.UsesPushData()) {
AddBinding(next, DescriptorBindingKind::ShaderData);
}
program.bindings = std::move(next);
program.binding_layout_complete = true;
}
const DescriptorBinding* FindBinding(const BindingLayout& layout, DescriptorBindingKind kind) {
for (const auto& binding: layout.descriptors) {
if (binding.kind == kind) {
return &binding;
}
}
return nullptr;
}
} // namespace Libs::Graphics::ShaderRecompiler::IR
src/graphics/shader/recompiler/ir/passes/BindingLayout.h
-23
View File
@@ -1,23 +0,0 @@
#ifndef EMULATOR_INCLUDE_EMULATOR_GRAPHICS_SHADER_RECOMPILER_BINDINGLAYOUT_H_
#define EMULATOR_INCLUDE_EMULATOR_GRAPHICS_SHADER_RECOMPILER_BINDINGLAYOUT_H_
#include "graphics/shader/recompiler/ir/ShaderIR.h"
namespace Libs::Graphics::ShaderRecompiler::IR {
void AllocateBindings(Program& program, uint32_t push_data_start_dword = 0,
bool lds_storage = false);
struct SharedMemoryResources {
bool lds = false;
bool gds = false;
};
SharedMemoryResources CollectMemoryResources(const Program& program, std::vector<uint32_t>& buffers);
bool UsesFlattenedSrt(const Program& program);
const DescriptorBinding* FindBinding(const BindingLayout& layout, DescriptorBindingKind kind);
} // namespace Libs::Graphics::ShaderRecompiler::IR
#endif /* EMULATOR_INCLUDE_EMULATOR_GRAPHICS_SHADER_RECOMPILER_BINDINGLAYOUT_H_ */
src/graphics/shader/recompiler/ir/passes/ResourceMaterialization.cpp
+1 -2
View File
@@ -1229,8 +1229,7 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
}
void ApplyResourceSpecialization(Program& program, const ResourceSpecialization& specialization) {
EXIT_IF(!program.resource_tracking_complete || program.shader_info_complete ||
program.binding_layout_complete);
EXIT_IF(!program.resource_tracking_complete || program.shader_info_complete);
EXIT_IF(program.info.buffers.size() > specialization.buffers.size() ||
program.info.images.size() > specialization.images.size());
src/graphics/shader/recompiler/ir/passes/ShaderInfoCollection.cpp
+437 -283
View File
@@ -75,128 +75,11 @@ void ValidateOptions(const Program& program, ShaderStageInputInfo input_info) {
}
}
void ValidateValueReferences(const Program& program, ShaderStageInputInfo input_info) {
for (const auto* block: program.blocks) {
for (const auto& inst: *block) {
switch (inst.GetOpcode()) {
case ValueOpcode::GetAttribute: {
if (!inst.Arg(0).IsImmediate() || inst.Arg(0).GetType() != Type::U32 ||
!inst.Arg(1).IsImmediate() || inst.Arg(1).GetType() != Type::U32) {
return Fail("typed attribute reference is not constant");
}
if ((program.stage == ShaderType::Vertex ||
program.stage == ShaderType::Local) &&
(inst.Arg(1).U32() >= 4u ||
inst.Arg(0).U32() >=
static_cast<uint32_t>(input_info.vertex->resources_num))) {
return Fail("vertex input reference is out of range");
}
if (program.stage == ShaderType::Pixel &&
(inst.Arg(1).U32() >= 4u ||
inst.Arg(0).U32() >= input_info.pixel->input_num)) {
return Fail("pixel input reference is out of range");
}
break;
}
case ValueOpcode::GetInterpolationParameter: {
if (program.stage != ShaderType::Pixel || !inst.Arg(0).IsImmediate() ||
inst.Arg(0).GetType() != Type::U32 || !inst.Arg(1).IsImmediate() ||
inst.Arg(1).GetType() != Type::U32 || !inst.Arg(2).IsImmediate() ||
inst.Arg(2).GetType() != Type::U32) {
return Fail("interpolation parameter reference is invalid");
}
if (inst.Arg(0).U32() >= input_info.pixel->input_num ||
inst.Arg(1).U32() >= 4u || inst.Arg(2).U32() >= 3u) {
return Fail("interpolation parameter reference is out of range");
}
break;
}
case ValueOpcode::GetBuiltin: {
if (!inst.Arg(0).IsImmediate() || inst.Arg(0).GetType() != Type::U32 ||
!inst.Arg(1).IsImmediate() || inst.Arg(1).GetType() != Type::U32) {
return Fail("typed builtin reference is not constant");
}
const auto kind = static_cast<StageInputKind>(inst.Arg(0).U32());
const auto component = inst.Arg(1).U32();
switch (kind) {
case StageInputKind::PackedAncillary:
return Fail("packed pixel ancillary input has an unsupported live use");
case StageInputKind::Layer:
case StageInputKind::SampleId:
if (program.stage != ShaderType::Pixel || component != 0u) {
return Fail("typed pixel scalar input is invalid");
}
break;
case StageInputKind::VertexIndex:
case StageInputKind::InvocationId:
case StageInputKind::PrimitiveId:
case StageInputKind::InstanceIndex:
case StageInputKind::FrontFacing:
case StageInputKind::LocalInvocationIndex:
if (component != 0u) {
return Fail("typed scalar builtin component is out of range");
}
break;
case StageInputKind::FragCoord:
if (component >= 4u) {
return Fail("typed fragment-coordinate component is out of range");
}
break;
case StageInputKind::BaryCoordSmooth:
case StageInputKind::BaryCoordSmoothSample:
case StageInputKind::BaryCoordSmoothCentroid:
case StageInputKind::BaryCoordNoPerspective:
if (component >= 2u) {
return Fail("typed barycentric component is out of range");
}
break;
case StageInputKind::TessCoord:
case StageInputKind::WorkgroupId:
case StageInputKind::LocalInvocationId:
case StageInputKind::GlobalInvocationId:
if (component >= 3u) {
return Fail("typed invocation builtin component is out of range");
}
break;
case StageInputKind::Parameter:
default: return Fail("typed builtin kind is invalid");
}
break;
}
case ValueOpcode::SetAttribute: {
const auto index = inst.Flags<ExportFlags>().index;
if (index >= program.export_info.size()) {
return Fail("typed export metadata index is out of range");
}
const auto& exp = program.export_info[index];
if (exp.kind == ExportTargetKind::Position && exp.index != 0 && exp.en != 0 &&
program.stage != ShaderType::Vertex && program.stage != ShaderType::Mesh &&
program.stage != ShaderType::TessellationEvaluation) {
return Fail("auxiliary position export requires a vertex, mesh, or "
"tessellation evaluation shader");
}
break;
}
default: break;
}
}
}
}
void CollectVertexInputs(const Program& program, const ShaderVertexInputInfo* vertex,
ShaderInfo& info) {
void CollectVertexInputs(const ShaderVertexInputInfo* vertex, ShaderInfo& info,
const std::array<uint8_t, 32>& used_components) {
AddInput(info, StageInputKind::VertexIndex, 0, 1, "gl_VertexIndex");
AddInput(info, StageInputKind::InstanceIndex, 0, 1, "gl_InstanceIndex");
uint32_t used_components[ShaderVertexInputInfo::RES_MAX] = {};
for (const auto* block: program.blocks) {
for (const auto& inst: *block) {
if (inst.GetOpcode() == ValueOpcode::GetAttribute) {
const auto attr = inst.Arg(0).U32();
const auto chan = inst.Arg(1).U32();
used_components[attr] = std::max(used_components[attr], chan + 1u);
}
}
}
for (uint32_t attr = 0; attr < static_cast<uint32_t>(vertex->resources_num); attr++) {
if (used_components[attr] != 0) {
AddInput(info, StageInputKind::Parameter, attr, used_components[attr],
@@ -205,28 +88,15 @@ void CollectVertexInputs(const Program& program, const ShaderVertexInputInfo* ve
}
}
void CollectPixelInputs(const Program& program, const ShaderPixelInputInfo* pixel,
ShaderInfo& info) {
void CollectPixelInputs(const ShaderPixelInputInfo* pixel, ShaderInfo& info,
std::array<bool, 32>& per_vertex,
const std::array<uint8_t, 32>& used_components) {
if (pixel->HasPositionInput()) {
AddInput(info, StageInputKind::FragCoord, 0, 4, "gl_FragCoord");
}
if (pixel->ps_front_face) {
AddInput(info, StageInputKind::FrontFacing, 0, 1, "gl_FrontFacing");
}
std::array<bool, 32> per_vertex {};
std::array<bool, 32> interpolated {};
for (const auto* block: program.blocks) {
for (const auto& inst: *block) {
if (inst.GetOpcode() == ValueOpcode::GetAttribute) {
interpolated[inst.Arg(0).U32()] = true;
} else if (inst.GetOpcode() == ValueOpcode::GetInterpolationParameter) {
const auto input = inst.Arg(0).U32();
const auto mode = inst.Arg(2).U32();
per_vertex[input] =
per_vertex[input] || mode < 2u || !ShaderPixelParameterIsFlat(*pixel, input);
}
}
}
// Mixed interpolation modes share raw vertices at their guest export slot.
// Rectangle expansion alone supplies separate flat and smooth outputs.
for (uint32_t input = 0; input < pixel->input_num; input++) {
@@ -245,7 +115,7 @@ void CollectPixelInputs(const Program& program, const ShaderPixelInputInfo* pixe
per_vertex[input]);
}
for (uint32_t input = 0; input < pixel->input_num; input++) {
if (interpolated[input] && per_vertex[input] &&
if (used_components[input] != 0 && per_vertex[input] &&
!ShaderPixelParameterIsFlat(*pixel, input)) {
const auto kind = pixel->ps_no_perspective ? StageInputKind::BaryCoordNoPerspective
: StageInputKind::BaryCoordSmooth;
@@ -271,147 +141,411 @@ void CollectComputeInputs(const ShaderComputeInputInfo* compute, ShaderInfo& inf
}
}
void CollectBuiltinInputs(const Program& program, ShaderInfo& info) {
for (const auto* block: program.blocks) {
for (const auto& inst: *block) {
if (program.stage == ShaderType::TessellationControl &&
inst.GetOpcode() == ValueOpcode::LaneId) {
AddInput(info, StageInputKind::InvocationId, 0, 1, "gl_InvocationID");
}
if (inst.GetOpcode() != ValueOpcode::GetBuiltin) {
continue;
}
const auto kind = static_cast<StageInputKind>(inst.Arg(0).U32());
switch (kind) {
case StageInputKind::VertexIndex:
AddInput(info, kind, 0, 1, "gl_VertexIndex");
break;
case StageInputKind::InstanceIndex:
AddInput(info, kind, 0, 1, "gl_InstanceIndex");
break;
case StageInputKind::InvocationId:
AddInput(info, kind, 0, 1, "gl_InvocationID");
break;
case StageInputKind::PrimitiveId:
AddInput(info, kind, 0, 1, "gl_PrimitiveID");
break;
case StageInputKind::TessCoord: AddInput(info, kind, 0, 3, "gl_TessCoord"); break;
case StageInputKind::FragCoord: AddInput(info, kind, 0, 4, "gl_FragCoord"); break;
case StageInputKind::FrontFacing:
AddInput(info, kind, 0, 1, "gl_FrontFacing");
break;
case StageInputKind::Layer: AddInput(info, kind, 0, 1, "gl_Layer"); break;
case StageInputKind::SampleId: AddInput(info, kind, 0, 1, "gl_SampleID"); break;
case StageInputKind::BaryCoordSmoothSample:
AddInput(info, StageInputKind::SampleId, 0, 1, "gl_SampleID");
[[fallthrough]];
case StageInputKind::BaryCoordSmooth:
case StageInputKind::BaryCoordSmoothCentroid:
AddInput(info, StageInputKind::BaryCoordSmooth, 0, 3, "gl_BaryCoordKHR");
break;
case StageInputKind::BaryCoordNoPerspective:
AddInput(info, kind, 0, 3, "gl_BaryCoordNoPerspKHR");
break;
case StageInputKind::WorkgroupId:
AddInput(info, kind, 0, 3, "gl_WorkGroupID");
break;
case StageInputKind::LocalInvocationId:
AddInput(info, kind, 0, 3, "gl_LocalInvocationID");
break;
case StageInputKind::LocalInvocationIndex:
AddInput(info, kind, 0, 1, "gl_LocalInvocationIndex");
break;
case StageInputKind::GlobalInvocationId:
AddInput(info, kind, 0, 3, "gl_GlobalInvocationID");
break;
case StageInputKind::PackedAncillary:
case StageInputKind::Parameter: break;
}
}
void AddBuiltinInput(ShaderInfo& info, StageInputKind kind) {
switch (kind) {
case StageInputKind::VertexIndex:
AddInput(info, kind, 0, 1, "gl_VertexIndex");
break;
case StageInputKind::InstanceIndex:
AddInput(info, kind, 0, 1, "gl_InstanceIndex");
break;
case StageInputKind::InvocationId:
AddInput(info, kind, 0, 1, "gl_InvocationID");
break;
case StageInputKind::PrimitiveId:
AddInput(info, kind, 0, 1, "gl_PrimitiveID");
break;
case StageInputKind::TessCoord: AddInput(info, kind, 0, 3, "gl_TessCoord"); break;
case StageInputKind::FragCoord: AddInput(info, kind, 0, 4, "gl_FragCoord"); break;
case StageInputKind::FrontFacing:
AddInput(info, kind, 0, 1, "gl_FrontFacing");
break;
case StageInputKind::Layer: AddInput(info, kind, 0, 1, "gl_Layer"); break;
case StageInputKind::SampleId: AddInput(info, kind, 0, 1, "gl_SampleID"); break;
case StageInputKind::BaryCoordSmoothSample:
AddInput(info, StageInputKind::SampleId, 0, 1, "gl_SampleID");
[[fallthrough]];
case StageInputKind::BaryCoordSmooth:
case StageInputKind::BaryCoordSmoothCentroid:
AddInput(info, StageInputKind::BaryCoordSmooth, 0, 3, "gl_BaryCoordKHR");
break;
case StageInputKind::BaryCoordNoPerspective:
AddInput(info, kind, 0, 3, "gl_BaryCoordNoPerspKHR");
break;
case StageInputKind::WorkgroupId:
AddInput(info, kind, 0, 3, "gl_WorkGroupID");
break;
case StageInputKind::LocalInvocationId:
AddInput(info, kind, 0, 3, "gl_LocalInvocationID");
break;
case StageInputKind::LocalInvocationIndex:
AddInput(info, kind, 0, 1, "gl_LocalInvocationIndex");
break;
case StageInputKind::GlobalInvocationId:
AddInput(info, kind, 0, 3, "gl_GlobalInvocationID");
break;
case StageInputKind::PackedAncillary:
case StageInputKind::Parameter: break;
}
}
void CollectOutputs(const Program& program, ShaderStageInputInfo input_info, ShaderInfo& info) {
void CollectOutput(const Program& program, ShaderStageInputInfo input_info,
ShaderInfo& info, const Inst& inst) {
const bool alpha_remap = program.stage == ShaderType::Pixel && input_info.pixel != nullptr &&
input_info.pixel->alpha_blend_source != ShaderAlphaBlendSource::None;
for (const auto* block: program.blocks) {
for (const auto& inst: *block) {
if (inst.GetOpcode() != ValueOpcode::SetAttribute) {
continue;
const auto& export_info = program.export_info[inst.Flags<ExportFlags>().index];
if (export_info.kind == ExportTargetKind::MrtZ) {
if (program.stage == ShaderType::Pixel && (export_info.en & 0x1u) != 0 &&
input_info.pixel->ps_depth_export_enable) {
AddOutput(info, StageOutputKind::Depth, 0, 0, "gl_FragDepth");
}
if (program.stage == ShaderType::Pixel && (export_info.en & 0x4u) != 0 &&
input_info.pixel->ps_sample_mask_export_enable) {
AddOutput(info, StageOutputKind::SampleMask, 0, 0, "gl_SampleMask");
}
return;
}
if (export_info.en == 0) {
return;
}
switch (export_info.kind) {
case ExportTargetKind::Position:
if (export_info.index == 0) {
AddOutput(info, StageOutputKind::Position, 0, 0, "out_position");
break;
}
const auto& export_info = program.export_info[inst.Flags<ExportFlags>().index];
if (export_info.kind == ExportTargetKind::MrtZ) {
if (program.stage == ShaderType::Pixel && (export_info.en & 0x1u) != 0 &&
input_info.pixel->ps_depth_export_enable) {
AddOutput(info, StageOutputKind::Depth, 0, 0, "gl_FragDepth");
}
if (program.stage == ShaderType::Pixel && (export_info.en & 0x4u) != 0 &&
input_info.pixel->ps_sample_mask_export_enable) {
AddOutput(info, StageOutputKind::SampleMask, 0, 0, "gl_SampleMask");
}
continue;
if (export_info.compr) {
return Fail("compressed auxiliary position export is unsupported");
}
if (export_info.en == 0) {
continue;
}
switch (export_info.kind) {
case ExportTargetKind::Position:
if (export_info.index == 0) {
AddOutput(info, StageOutputKind::Position, 0, 0, "out_position");
break;
}
if (export_info.compr) {
return Fail("compressed auxiliary position export is unsupported");
}
for (uint32_t component = 0; component < 4; component++) {
if ((export_info.en & (1u << component)) == 0) {
continue;
}
const auto output = DecodePositionExportComponent(
input_info.vertex->pa_cl_vs_out_cntl, export_info.index, component);
if (output.viewport) {
AddOutput(info, StageOutputKind::ViewportIndex, 0, 0, "gl_ViewportIndex");
}
if (output.point_size) {
AddOutput(info, StageOutputKind::PointSize, 0, 0, "gl_PointSize");
}
if (output.layer) {
AddOutput(info, StageOutputKind::Layer, 0, 0, "gl_Layer");
}
if (output.clip_distance != UINT32_MAX) {
AddOutput(info, StageOutputKind::ClipDistance, output.clip_distance, 0,
"gl_ClipDistance");
}
if (output.cull_distance != UINT32_MAX) {
AddOutput(info, StageOutputKind::CullDistance, output.cull_distance, 0,
"gl_CullDistance");
}
}
break;
case ExportTargetKind::Parameter:
AddOutput(info, StageOutputKind::Parameter, export_info.index,
export_info.index, fmt::format("out_param_{}", export_info.index));
break;
case ExportTargetKind::Mrt: {
if (alpha_remap && export_info.index != 0) {
break;
}
const auto slot = input_info.pixel->dual_source_blending
? 0u : ShaderPixelExportTarget(input_info.pixel->target_shader_mask,
export_info.index);
if (slot >= 8) {
break;
}
AddOutput(info, StageOutputKind::Mrt, export_info.index, slot,
fmt::format("out_mrt_{}", export_info.index));
if (alpha_remap) {
AddOutput(info, StageOutputKind::Mrt, 1, 0, "out_mrt_1");
}
break;
for (uint32_t component = 0; component < 4; component++) {
if ((export_info.en & (1u << component)) == 0) {
continue;
}
const auto output = DecodePositionExportComponent(
input_info.vertex->pa_cl_vs_out_cntl, export_info.index, component);
if (output.viewport) {
AddOutput(info, StageOutputKind::ViewportIndex, 0, 0, "gl_ViewportIndex");
}
if (output.point_size) {
AddOutput(info, StageOutputKind::PointSize, 0, 0, "gl_PointSize");
}
if (output.layer) {
AddOutput(info, StageOutputKind::Layer, 0, 0, "gl_Layer");
}
if (output.clip_distance != UINT32_MAX) {
AddOutput(info, StageOutputKind::ClipDistance, output.clip_distance, 0,
"gl_ClipDistance");
}
if (output.cull_distance != UINT32_MAX) {
AddOutput(info, StageOutputKind::CullDistance, output.cull_distance, 0,
"gl_CullDistance");
}
}
break;
case ExportTargetKind::Parameter:
AddOutput(info, StageOutputKind::Parameter, export_info.index,
export_info.index, fmt::format("out_param_{}", export_info.index));
break;
case ExportTargetKind::Mrt: {
if (alpha_remap && export_info.index != 0) {
break;
}
const auto slot = input_info.pixel->dual_source_blending
? 0u : ShaderPixelExportTarget(input_info.pixel->target_shader_mask,
export_info.index);
if (slot >= 8) {
break;
}
AddOutput(info, StageOutputKind::Mrt, export_info.index, slot,
fmt::format("out_mrt_{}", export_info.index));
if (alpha_remap) {
AddOutput(info, StageOutputKind::Mrt, 1, 0, "out_mrt_1");
}
break;
}
default: break;
}
}
struct InputUsage {
std::array<uint8_t, 32> components {};
std::array<bool, 32> per_vertex {};
std::array<bool, NumScalarRegs> user_data {};
std::array<StageInputKind, static_cast<size_t>(StageInputKind::Parameter)> builtins {};
uint32_t builtin_mask = 0;
uint32_t builtin_count = 0;
void AddBuiltin(StageInputKind kind) {
const auto bit = 1u << static_cast<uint32_t>(kind);
if ((builtin_mask & bit) == 0) {
builtin_mask |= bit;
builtins[builtin_count++] = kind;
}
}
};
void Visit(Program& program, ShaderStageInputInfo input_info, InputUsage& inputs,
const Inst& inst) {
auto& info = program.info;
const auto op = inst.GetOpcode();
const auto buffer_access = BufferAccessOf(op);
const auto address_access = AddressOpcodeInfoOf(op).access;
const auto shared_access = SharedAccessOf(op);
info.float64 |= inst.GetType() == Type::F64;
info.buffer_int64_atomics |= buffer_access == BufferAccess::Atomic &&
inst.GetType() == Type::U64;
if (buffer_access != BufferAccess::None || address_access != AddressAccess::None ||
shared_access != SharedAccess::None) {
const auto index = inst.Flags<MemoryFlags>().index;
if (index >= program.memory_info.size()) {
return Fail("typed shader contains invalid memory metadata");
}
const auto& memory = program.memory_info[index];
const auto kind = memory.kind;
if (!memory.planning_only) {
info.uses_lds |= kind == ResourceKind::FlatLocal;
if (shared_access != SharedAccess::None) {
info.uses_lds |= kind == ResourceKind::Lds;
info.uses_gds |= kind == ResourceKind::Gds;
} else if (kind == ResourceKind::Buffer || kind == ResourceKind::ScalarBuffer) {
if (memory.resource >= info.buffers.size() || memory.resource >= ShaderInfo::MaxBuffers) {
return Fail("buffer operation has invalid resource metadata");
}
info.live_buffers |= uint64_t{1} << memory.resource;
for (const auto child: info.buffers[memory.resource].indirect_resources) {
if (child >= info.buffers.size()) {
return Fail("indirect buffer operation has invalid resource metadata");
}
info.live_buffers |= uint64_t{1} << child;
}
default: break;
}
}
if (address_access != AddressAccess::None) {
if (kind == ResourceKind::Scratch || kind == ResourceKind::FlatLocal) {
if (program.scratch_dwords == 0) {
return Fail("scratch operation has no per-thread storage");
}
info.function_scratch = true;
info.function_lds |= kind == ResourceKind::FlatLocal &&
program.stage != ShaderType::Compute && program.stage != ShaderType::Mesh;
} else if (address_access == AddressAccess::Write) {
return Fail("writable FLAT/GLOBAL addresses require GPU ownership tracking");
} else {
info.coherent_buffers |= memory.coherent;
}
}
if (buffer_access != BufferAccess::None) {
if (kind == ResourceKind::IndirectBuffer && op != ValueOpcode::ReadConstBuffer) {
info.subgroup_local_invocation_id = true;
}
if (kind == ResourceKind::Buffer) {
info.coherent_buffers |= memory.coherent;
if (memory.resource >= info.buffers.size()) {
return Fail("buffer operation has invalid resource metadata");
}
const auto bits = StorageBufferElementBits(program, memory);
info.buffer_u8 |= bits == 8u;
info.buffer_u16 |= bits == 16u;
if ((info.buffers[memory.resource].packed_stride & (1u << 20u)) != 0u) {
if (program.stage != ShaderType::Compute) {
return Fail("buffer ADD_TID is only valid for compute shaders");
}
info.subgroup_local_invocation_id = true;
}
}
}
if (shared_access != SharedAccess::None) {
if (kind != ResourceKind::Lds && kind != ResourceKind::Gds) {
return Fail("shared operation has invalid resource kind");
}
if (shared_access == SharedAccess::Atomic && SharedComponentCount(op) == 2u) {
if (kind != ResourceKind::Lds || program.stage != ShaderType::Compute) {
return Fail("64-bit shared atomics require compute LDS");
}
info.shared_int64_atomics = true;
}
info.function_lds |= kind == ResourceKind::Lds &&
program.stage != ShaderType::Compute && program.stage != ShaderType::Mesh;
if (shared_access == SharedAccess::Append || shared_access == SharedAccess::Consume) {
info.subgroup_ballot = true;
info.subgroup_shuffle = true;
info.subgroup_local_invocation_id = true;
}
}
}
switch (inst.GetOpcode()) {
case ValueOpcode::GetAttribute: {
if (!inst.Arg(0).IsImmediate() || inst.Arg(0).GetType() != Type::U32 ||
!inst.Arg(1).IsImmediate() || inst.Arg(1).GetType() != Type::U32) {
return Fail("typed attribute reference is not constant");
}
if ((program.stage == ShaderType::Vertex ||
program.stage == ShaderType::Local) &&
(inst.Arg(1).U32() >= 4u ||
inst.Arg(0).U32() >=
static_cast<uint32_t>(input_info.vertex->resources_num))) {
return Fail("vertex input reference is out of range");
}
if (program.stage == ShaderType::Pixel &&
(inst.Arg(1).U32() >= 4u ||
inst.Arg(0).U32() >= input_info.pixel->input_num)) {
return Fail("pixel input reference is out of range");
}
if (program.stage == ShaderType::Vertex || program.stage == ShaderType::Local ||
program.stage == ShaderType::Pixel) {
auto& components = inputs.components[inst.Arg(0).U32()];
components = std::max(components, static_cast<uint8_t>(inst.Arg(1).U32() + 1u));
}
break;
}
case ValueOpcode::GetInterpolationParameter: {
if (program.stage != ShaderType::Pixel || !inst.Arg(0).IsImmediate() ||
inst.Arg(0).GetType() != Type::U32 || !inst.Arg(1).IsImmediate() ||
inst.Arg(1).GetType() != Type::U32 || !inst.Arg(2).IsImmediate() ||
inst.Arg(2).GetType() != Type::U32) {
return Fail("interpolation parameter reference is invalid");
}
if (inst.Arg(0).U32() >= input_info.pixel->input_num ||
inst.Arg(1).U32() >= 4u || inst.Arg(2).U32() >= 3u) {
return Fail("interpolation parameter reference is out of range");
}
const auto input = inst.Arg(0).U32();
inputs.per_vertex[input] |= inst.Arg(2).U32() < 2u ||
!ShaderPixelParameterIsFlat(*input_info.pixel, input);
break;
}
case ValueOpcode::GetBuiltin: {
if (!inst.Arg(0).IsImmediate() || inst.Arg(0).GetType() != Type::U32 ||
!inst.Arg(1).IsImmediate() || inst.Arg(1).GetType() != Type::U32) {
return Fail("typed builtin reference is not constant");
}
const auto kind = static_cast<StageInputKind>(inst.Arg(0).U32());
const auto component = inst.Arg(1).U32();
switch (kind) {
case StageInputKind::PackedAncillary:
return Fail("packed pixel ancillary input has an unsupported live use");
case StageInputKind::Layer:
case StageInputKind::SampleId:
if (program.stage != ShaderType::Pixel || component != 0u) {
return Fail("typed pixel scalar input is invalid");
}
break;
case StageInputKind::VertexIndex:
case StageInputKind::InvocationId:
case StageInputKind::PrimitiveId:
case StageInputKind::InstanceIndex:
case StageInputKind::FrontFacing:
case StageInputKind::LocalInvocationIndex:
if (component != 0u) {
return Fail("typed scalar builtin component is out of range");
}
break;
case StageInputKind::FragCoord:
if (component >= 4u) {
return Fail("typed fragment-coordinate component is out of range");
}
break;
case StageInputKind::BaryCoordSmooth:
case StageInputKind::BaryCoordSmoothSample:
case StageInputKind::BaryCoordSmoothCentroid:
case StageInputKind::BaryCoordNoPerspective:
if (component >= 2u) {
return Fail("typed barycentric component is out of range");
}
break;
case StageInputKind::TessCoord:
case StageInputKind::WorkgroupId:
case StageInputKind::LocalInvocationId:
case StageInputKind::GlobalInvocationId:
if (component >= 3u) {
return Fail("typed invocation builtin component is out of range");
}
break;
case StageInputKind::Parameter:
default: return Fail("typed builtin kind is invalid");
}
inputs.AddBuiltin(kind);
break;
}
case ValueOpcode::SetAttribute: {
const auto index = inst.Flags<ExportFlags>().index;
if (index >= program.export_info.size()) {
return Fail("typed export metadata index is out of range");
}
const auto& exp = program.export_info[index];
if (exp.kind == ExportTargetKind::Position && exp.index != 0 && exp.en != 0 &&
program.stage != ShaderType::Vertex && program.stage != ShaderType::Mesh &&
program.stage != ShaderType::TessellationEvaluation) {
return Fail("auxiliary position export requires a vertex, mesh, or "
"tessellation evaluation shader");
}
info.pixel_valid_mask |= program.stage == ShaderType::Pixel && exp.vm;
CollectOutput(program, input_info, info, inst);
break;
}
case ValueOpcode::GetUserData: {
if (!inst.HasUses()) break;
if (inst.Arg(0).GetType() != Type::ScalarReg) {
return Fail("typed shader contains an invalid user-data register");
}
const auto index = RegIndex(inst.Arg(0).ScalarRegister());
if (index >= NumScalarRegs) {
return Fail("typed shader contains an invalid user-data register");
}
inputs.user_data[index] = true;
break;
}
case ValueOpcode::GetDispatchThreadExtent:
info.uses_dispatch_threads |= inst.HasUses();
break;
case ValueOpcode::ReadConst: info.uses_flattened_srt = true; break;
case ValueOpcode::BitwiseXor32: info.has_bitwise_xor = true; break;
case ValueOpcode::StoreCompletion: info.subgroup_barrier = true; break;
case ValueOpcode::BvhIntersect: info.bvh = true; break;
case ValueOpcode::ConditionRef:
info.subgroup_ballot |=
inst.Flags<CFG::BranchCondition>() != CFG::BranchCondition::ScalarInstruction;
break;
case ValueOpcode::Ballot: info.subgroup_ballot = true; break;
case ValueOpcode::DppMoveU32:
case ValueOpcode::ReadFirstLane:
case ValueOpcode::ReadLane: {
info.subgroup_ballot = true;
info.subgroup_shuffle = true;
if (inst.GetOpcode() == ValueOpcode::DppMoveU32) {
info.subgroup_local_invocation_id = true;
}
break;
}
case ValueOpcode::DppUpdateU32:
case ValueOpcode::WriteLane: {
info.subgroup_ballot = true;
info.subgroup_local_invocation_id = true;
break;
}
case ValueOpcode::SwizzleU32:
info.uses_swizzle = true;
[[fallthrough]];
case ValueOpcode::Permlane16U32:
case ValueOpcode::PermuteU32:
case ValueOpcode::BpermuteU32: {
info.subgroup_ballot = true;
info.subgroup_shuffle = true;
info.subgroup_local_invocation_id = true;
break;
}
case ValueOpcode::LaneId:
if (program.stage == ShaderType::TessellationControl) {
inputs.AddBuiltin(StageInputKind::InvocationId);
} else {
info.subgroup_local_invocation_id = true;
}
break;
case ValueOpcode::ImageQueryLod: info.compute_derivatives = true; break;
case ValueOpcode::ImageGatherRaw:
info.image_gather_extended = true;
break;
default: break;
}
}
@@ -423,29 +557,49 @@ void CollectShaderInfo(Program& program, ShaderStageInputInfo input_info) {
: "shader info already collected");
}
ValidateOptions(program, input_info);
ValidateValueReferences(program, input_info);
auto& next = program.info;
next.inputs.clear();
next.outputs.clear();
next.has_bitwise_xor =
std::any_of(program.blocks.begin(), program.blocks.end(), [](const auto* block) {
return std::any_of(block->begin(), block->end(), [](const auto& inst) {
return inst.GetOpcode() == ValueOpcode::BitwiseXor32;
});
});
auto& info = program.info;
if (info.buffers.size() > ShaderInfo::MaxBuffers) {
return Fail("buffer resource limit exceeded");
}
info.inputs.clear();
info.outputs.clear();
info.user_data_registers.clear();
info.live_buffers = 0;
info.has_bitwise_xor = info.uses_lds = info.uses_gds = info.uses_flattened_srt =
info.uses_dispatch_threads = info.uses_swizzle = false;
info.bvh = info.subgroup_ballot = info.subgroup_barrier = info.subgroup_shuffle =
info.subgroup_local_invocation_id = info.compute_derivatives = info.image_gather_extended =
info.function_lds = info.function_scratch = info.pixel_valid_mask = info.buffer_int64_atomics =
info.buffer_u8 = info.buffer_u16 = info.shared_int64_atomics = info.coherent_buffers = info.float64 = false;
InputUsage inputs;
for (const auto* block: program.blocks) {
for (const auto& inst: *block) {
Visit(program, input_info, inputs, inst);
}
}
for (uint32_t index = 0; index < inputs.user_data.size(); index++) {
if (inputs.user_data[index]) info.user_data_registers.push_back(index);
}
const auto uses_mapping = [](const auto& resource) {
return resource.indirect_root != UINT32_MAX;
};
info.uses_flattened_srt |= std::ranges::any_of(info.buffers, uses_mapping) ||
std::ranges::any_of(info.images, uses_mapping);
switch (program.stage) {
case ShaderType::Vertex:
case ShaderType::Local: CollectVertexInputs(program, input_info.vertex, next); break;
case ShaderType::Local: CollectVertexInputs(input_info.vertex, info, inputs.components); break;
case ShaderType::TessellationControl:
case ShaderType::TessellationEvaluation:
case ShaderType::Mesh: break;
case ShaderType::Pixel: CollectPixelInputs(program, input_info.pixel, next); break;
case ShaderType::Compute: CollectComputeInputs(input_info.compute, next); break;
case ShaderType::Pixel:
CollectPixelInputs(input_info.pixel, info, inputs.per_vertex, inputs.components);
break;
case ShaderType::Compute: CollectComputeInputs(input_info.compute, info); break;
default: return Fail("unsupported shader stage for info collection");
}
CollectBuiltinInputs(program, next);
CollectOutputs(program, input_info, next);
for (uint32_t i = 0; i < inputs.builtin_count; i++) {
AddBuiltinInput(info, inputs.builtins[i]);
}
program.shader_info_complete = true;
}
src/graphics/shader/recompiler/ir/passes/ShaderInfoCollection.h
+1 -2
View File
@@ -5,8 +5,7 @@
namespace Libs::Graphics::ShaderRecompiler::IR {
// Completes the immutable shader interface after resource tracking. On failure Program::info and
// all completion state remain unchanged.
// Collects the shader interface, resource usage and instruction features after resource tracking.
void CollectShaderInfo(Program& program, ShaderStageInputInfo input_info);
} // namespace Libs::Graphics::ShaderRecompiler::IR