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

gpu: store large compute shared memory in device buffers

Browse Source
This commit is contained in:
nmzik
2026-10-04 22:07:07 +04:00
parent 0ee8800905
commit b0bbaef1eb
17 changed files with 269 additions and 69 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/host_gpu/renderer/pipeline/descriptors.cpp
+6 -1
View File
@@ -68,7 +68,8 @@ vk::DescriptorType NativeDescriptorType(BindingKind kind) {
case BindingKind::BdaPagetable:
case BindingKind::FaultBuffer:
case BindingKind::FlattenedSrt:
case BindingKind::ShaderData: return vk::DescriptorType::eStorageBuffer;
case BindingKind::ShaderData:
case BindingKind::SharedMemory: return vk::DescriptorType::eStorageBuffer;
case BindingKind::Count: EXIT("invalid native descriptor binding kind");
}
EXIT("invalid native descriptor binding kind");
@@ -779,6 +780,7 @@ void RenderExecutor::PrepareBindings(const ShaderStageRuntime& runtime,
prepared.gds = {nullptr, 0, VK_WHOLE_SIZE};
prepared.flattened_srt = {};
prepared.shader_data_buffer = {};
prepared.shared_memory = {};
prepared.images.resize(program.info.images.size());
prepared.samplers.clear();
prepared.shader_data.clear();
@@ -1120,12 +1122,15 @@ void RenderExecutor::CommitBindings(CommandBuffer& buffer,
}
case BindingKind::FlattenedSrt:
case BindingKind::ShaderData:
case BindingKind::SharedMemory:
case BindingKind::Gds: {
const vk::DescriptorBufferInfo* view = &descriptors.gds;
if (binding.kind == BindingKind::FlattenedSrt) {
view = &descriptors.flattened_srt;
} else if (binding.kind == BindingKind::ShaderData) {
view = &descriptors.shader_data_buffer;
} else if (binding.kind == BindingKind::SharedMemory) {
view = &descriptors.shared_memory;
}
EXIT_IF(view->buffer == nullptr);
m_descriptor_buffers.push_back(*view);
src/graphics/host_gpu/renderer/pipeline/descriptors.h
+1
View File
@@ -44,6 +44,7 @@ struct PreparedBindings {
vk::DescriptorBufferInfo gds {nullptr, 0, VK_WHOLE_SIZE};
vk::DescriptorBufferInfo flattened_srt;
vk::DescriptorBufferInfo shader_data_buffer;
vk::DescriptorBufferInfo shared_memory;
std::vector<uint32_t> shader_data;
};
src/graphics/host_gpu/renderer/pipeline/pipelineCache.cpp
+2 -13
View File
@@ -680,19 +680,8 @@ ShaderProgram PipelineCache::GetComputeProgram(const HW::ComputeShaderInfo& regs
ShaderComputeInputInfo& input_info) {
input_info.host_subgroup_size = m_graphics.SupportsComputeWave64() ? 64u : 32u;
const auto params = PrepareProgram(regs, sh, input_info);
// Use one effective size for the cache key, LDS declaration, and access bounds.
const auto max_lds_dwords =
m_graphics.GetPhysicalDeviceProperties().limits.maxComputeSharedMemorySize / 4u;
if (input_info.lds_size_dwords > max_lds_dwords) {
static std::atomic_bool warned = false;
if (!warned.exchange(true, std::memory_order_relaxed)) {
PipelineCacheLog("GPU warning: game compute shader requests {} bytes of LDS, but "
"the Vulkan device limit is {} bytes. Clamping LDS; rendering may "
"be incorrect.",
input_info.lds_size_dwords * 4u, max_lds_dwords * 4u);
}
}
input_info.lds_size_dwords = std::min(input_info.lds_size_dwords, max_lds_dwords);
input_info.lds_storage = input_info.lds_size_dwords * 4u >
m_graphics.GetPhysicalDeviceProperties().limits.maxComputeSharedMemorySize;
uint32_t push_data_cursor = 0;
return m_program_cache->Get(params, input_info, push_data_cursor);
}
src/graphics/host_gpu/renderer/renderCompute.cpp
+36 -2
View File
@@ -20,6 +20,7 @@
#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"
@@ -195,6 +196,35 @@ bool RenderExecutor::TryConsumeComputeImageClear(const ShaderComputeInputInfo& i
return true;
}
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) {
return;
}
auto& cache = context.GetBufferCache();
if (indirect_args != 0) {
cache.ReadMemory(indirect_args, sizeof(vk::DispatchIndirectCommand));
std::memcpy(input.workgroup_counts, reinterpret_cast<const void*>(indirect_args),
sizeof(input.workgroup_counts));
}
// LDS has no contents to preserve between dispatches. The existing shader hazard
// barriers also order other users of this GPU-only utility buffer.
auto& storage = cache.GetUtilityBuffer(MemoryUsage::DeviceLocal);
const auto limit = std::min<uint64_t>(storage.Size(),
context.GetGraphics().GetPhysicalDeviceProperties().limits.maxStorageBufferRange);
uint64_t size = sizeof(uint32_t);
if (std::ranges::find(input.workgroup_counts, 0u) == std::end(input.workgroup_counts)) {
size = uint64_t {input.lds_size_dwords} * sizeof(uint32_t);
EXIT_IF(size == 0 || size > limit);
for (const auto count: input.workgroup_counts) {
EXIT_IF(size > limit / count);
size *= count;
}
}
bindings.shared_memory = {storage.Handle(), 0, size};
}
void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
uint32_t thread_group_x, uint32_t thread_group_y,
uint32_t thread_group_z, uint32_t mode) {
@@ -317,13 +347,15 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
m_context.PrepareBda();
}
RebindImages(bindings);
BindSharedMemory(m_context, input_info, bindings);
RebindBuffers(bindings);
auto vk_buffer = buffer.Handle();
PreparedBindings* descriptor_stage = &bindings;
CommitBindings(buffer, vk::PipelineBindPoint::eCompute, pipeline,
std::span {&descriptor_stage, 1u});
bool has_storage_writes = HasShaderBufferWrites(input_info.stage);
bool has_storage_writes = bindings.shared_memory.buffer != nullptr ||
HasShaderBufferWrites(input_info.stage);
has_storage_writes =
std::any_of(program.info.images.begin(), program.info.images.end(),
[](const auto& image) {
@@ -430,6 +462,7 @@ void RenderExecutor::DispatchIndirect(uint64_t submit_id, CommandBuffer& buffer,
if (program.info.uses_dma) {
m_context.PrepareBda();
}
BindSharedMemory(m_context, input_info, bindings, args_addr);
RebindImages(bindings);
// Acquiring arguments can merge cache buffers; finalize shader bindings afterward.
const auto [args_buffer, args_offset] = m_context.GetBufferCache().ObtainBuffer(
@@ -440,7 +473,8 @@ void RenderExecutor::DispatchIndirect(uint64_t submit_id, CommandBuffer& buffer,
CommitBindings(buffer, vk::PipelineBindPoint::eCompute, pipeline,
std::span {&descriptor_stage, 1u});
const auto vk_buffer = buffer.Handle();
const bool has_storage_writes = HasShaderBufferWrites(input_info.stage) ||
const bool has_storage_writes = bindings.shared_memory.buffer != nullptr ||
HasShaderBufferWrites(input_info.stage) ||
std::any_of(program.info.images.begin(), program.info.images.end(), [](const auto& image) {
return image.written && image.resource_class ==
ShaderRecompiler::IR::ImageResourceClass::Storage;