raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-20 02:47:02 +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();