The-Homebrew-Cloud/skyline
2
0
Fork 0
mirror of https://github.com/skyline-emu/skyline.git synced 2024-12-24 18:41:52 +01:00
Code Issues Packages Projects Releases Wiki Activity

Use raw pointers to hold constant buffer views

Browse Source 
The constant destruction and creation of `BufferView`s in cbuf-heavy games showed up as a large chunk of the profiler. Fix this by taking advantage of the fact that constant buffer `BufferView`s are never deleted and always kept around in the cache to just return a pointer to them in the cache.
This commit is contained in:
Billy Laws 2022-07-31 13:57:11 +01:00
parent 6b2e84712b
commit 460e6c9c84
1 changed files with 36 additions and 31 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

67
app/src/main/cpp/skyline/gpu/interconnect/graphics_context.h
View File

@ -617,7 +617,7 @@ namespace skyline::gpu::interconnect {
struct ConstantBuffer { struct ConstantBuffer {
IOVA iova; IOVA iova;
u32 size; u32 size;
BufferView view; BufferView *view;
/** /**
* @brief Reads an object from the supplied offset in the constant buffer * @brief Reads an object from the supplied offset in the constant buffer
@ -626,8 +626,8 @@ namespace skyline::gpu::interconnect {
template<typename T> template<typename T>
T Read(CommandExecutor &pExecutor, size_t dstOffset) const { T Read(CommandExecutor &pExecutor, size_t dstOffset) const {
T object; T object;
ContextLock lock{pExecutor.tag, view}; ContextLock lock{pExecutor.tag, *view};
view.Read(lock.IsFirstUsage(), []() { view->Read(lock.IsFirstUsage(), []() {
// TODO: here we should trigger a SubmitWithFlush, however that doesn't currently work due to Read being called mid-draw and attached objects not handling this case // TODO: here we should trigger a SubmitWithFlush, however that doesn't currently work due to Read being called mid-draw and attached objects not handling this case
Logger::Warn("GPU dirty buffer reads for attached buffers are unimplemented"); Logger::Warn("GPU dirty buffer reads for attached buffers are unimplemented");
}, span<T>(object).template cast<u8>(), dstOffset); }, span<T>(object).template cast<u8>(), dstOffset);
@ -642,28 +642,29 @@ namespace skyline::gpu::interconnect {
void Write(CommandExecutor &pExecutor, MegaBufferAllocator &megaBufferAllocator, span<T> buf, size_t dstOffset) { void Write(CommandExecutor &pExecutor, MegaBufferAllocator &megaBufferAllocator, span<T> buf, size_t dstOffset) {
auto srcCpuBuf{buf.template cast<u8>()}; auto srcCpuBuf{buf.template cast<u8>()};
ContextLock lock{pExecutor.tag, view}; ContextLock lock{pExecutor.tag, *view};
view.Write(lock.IsFirstUsage(), pExecutor.cycle, []() { view->Write(lock.IsFirstUsage(), pExecutor.cycle, []() {
// TODO: see Read() // TODO: see Read()
Logger::Warn("GPU dirty buffer reads for attached buffers are unimplemented"); Logger::Warn("GPU dirty buffer reads for attached buffers are unimplemented");
}, [&megaBufferAllocator, &pExecutor, srcCpuBuf, dstOffset, &view = this->view, &lock]() { }, [&megaBufferAllocator, &pExecutor, srcCpuBuf, dstOffset, &view = this->view, &lock]() {
pExecutor.AttachLockedBufferView(view, std::move(lock)); pExecutor.AttachLockedBufferView(*view, std::move(lock));
// This will prevent any CPU accesses to backing for the duration of the usage // This will prevent any CPU accesses to backing for the duration of the usage
// ONLY in this specific case is it fine to access the backing buffer directly since the flag will be propagated with recreations // ONLY in this specific case is it fine to access the backing buffer directly since the flag will be propagated with recreations
view->buffer->BlockAllCpuBackingWrites(); (*view)->buffer->BlockAllCpuBackingWrites();
auto srcGpuAllocation{megaBufferAllocator.Push(pExecutor.cycle, srcCpuBuf)}; auto srcGpuAllocation{megaBufferAllocator.Push(pExecutor.cycle, srcCpuBuf)};
pExecutor.AddOutsideRpCommand([=](vk::raii::CommandBuffer &commandBuffer, const std::shared_ptr<FenceCycle> &, GPU &) { pExecutor.AddOutsideRpCommand([=](vk::raii::CommandBuffer &commandBuffer, const std::shared_ptr<FenceCycle> &, GPU &) {
vk::BufferCopy copyRegion{ vk::BufferCopy copyRegion{
.size = srcCpuBuf.size_bytes(), .size = srcCpuBuf.size_bytes(),
.srcOffset = srcGpuAllocation.offset, .srcOffset = srcGpuAllocation.offset,
.dstOffset = view->view->offset + dstOffset .dstOffset = (*view)->view->offset + dstOffset
}; };
commandBuffer.copyBuffer(srcGpuAllocation.buffer, view->buffer->GetBacking(), copyRegion); commandBuffer.copyBuffer(srcGpuAllocation.buffer, (*view)->buffer->GetBacking(), copyRegion);
}); });
}, srcCpuBuf, dstOffset); }, srcCpuBuf, dstOffset);
} }
}; };
ConstantBuffer constantBufferSelector{}; //!< The constant buffer selector is used to bind a constant buffer to a stage or update data in it ConstantBuffer constantBufferSelector{}; //!< The constant buffer selector is used to bind a constant buffer to a stage or update data in it
public: public:
@ -708,40 +709,44 @@ namespace skyline::gpu::interconnect {
std::unordered_map<Key, BufferView, KeyHash> cache; std::unordered_map<Key, BufferView, KeyHash> cache;
public: public:
std::optional<BufferView> Lookup(u32 size, u64 iova) { BufferView *Lookup(u32 size, u64 iova) {
if (auto it{cache.find({size, iova})}; it != cache.end()) if (auto it{cache.find({size, iova})}; it != cache.end())
return it->second; return &it->second;
return std::nullopt; return nullptr;
} }
void Insert(u32 size, u64 iova, BufferView &view) { BufferView *Insert(u32 size, u64 iova, BufferView &&view) {
cache[Key{size, iova}] = view; return &cache.emplace(Key{size, iova}, view).first->second;
} }
} constantBufferCache; } constantBufferCache;
std::optional<ConstantBuffer> GetConstantBufferSelector() { ConstantBuffer *GetConstantBufferSelector() {
if (constantBufferSelector.size == 0) if (constantBufferSelector.size == 0)
return std::nullopt; return nullptr;
else if (constantBufferSelector.view) else if (constantBufferSelector.view)
return constantBufferSelector; return &constantBufferSelector;
auto view{constantBufferCache.Lookup(constantBufferSelector.size, constantBufferSelector.iova)}; auto view{constantBufferCache.Lookup(constantBufferSelector.size, constantBufferSelector.iova)};
if (!view) { if (!view) {
auto mappings{channelCtx.asCtx->gmmu.TranslateRange(constantBufferSelector.iova, constantBufferSelector.size)}; auto mappings{channelCtx.asCtx->gmmu.TranslateRange(constantBufferSelector.iova, constantBufferSelector.size)};
view = executor.AcquireBufferManager().FindOrCreate(mappings.front(), executor.tag, [this](std::shared_ptr<Buffer> buffer, ContextLock<Buffer> &&lock) { view = constantBufferCache.Insert(constantBufferSelector.size, constantBufferSelector.iova,
executor.AttachLockedBuffer(buffer, std::move(lock)); executor.AcquireBufferManager().FindOrCreate(mappings.front(), executor.tag, [this](std::shared_ptr<Buffer> buffer, ContextLock<Buffer> &&lock) {
}); executor.AttachLockedBuffer(buffer, std::move(lock));
constantBufferCache.Insert(constantBufferSelector.size, constantBufferSelector.iova, *view); })
);
} }
constantBufferSelector.view = *view; constantBufferSelector.view = view;
return constantBufferSelector; return &constantBufferSelector;
} }
void ConstantBufferUpdate(std::vector<u32> data, u32 offset) { void ConstantBufferUpdate(span<u32> data, u32 offset) {
auto constantBuffer{GetConstantBufferSelector().value()}; auto constantBuffer{GetConstantBufferSelector()};
constantBuffer.Write<u32>(executor, executor.AcquireMegaBufferAllocator(), data, offset); if (constantBuffer)
constantBuffer->Write<u32>(executor, executor.AcquireMegaBufferAllocator(), data, offset);
else
throw exception("Attempting to write to invalid constant buffer!");
} }
/* Shader Program */ /* Shader Program */
@ -1112,7 +1117,7 @@ namespace skyline::gpu::interconnect {
.stageFlags = pipelineStage.vkStage, .stageFlags = pipelineStage.vkStage,
}); });
auto view{pipelineStage.constantBuffers[constantBuffer.index].view}; auto &view{*pipelineStage.constantBuffers[constantBuffer.index].view};
executor.AttachBuffer(view); executor.AttachBuffer(view);
if (auto megaBufferAllocation{view.AcquireMegaBuffer(executor.cycle, executor.AcquireMegaBufferAllocator())}) { if (auto megaBufferAllocation{view.AcquireMegaBuffer(executor.cycle, executor.AcquireMegaBufferAllocator())}) {
// If the buffer is megabuffered then since we don't get out data from the underlying buffer, rather the megabuffer which stays consistent throughout a single execution, we can skip registering usage // If the buffer is megabuffered then since we don't get out data from the underlying buffer, rather the megabuffer which stays consistent throughout a single execution, we can skip registering usage
@ -1268,12 +1273,12 @@ namespace skyline::gpu::interconnect {
} }
void BindPipelineConstantBuffer(maxwell3d::PipelineStage stage, bool enable, u32 index) { void BindPipelineConstantBuffer(maxwell3d::PipelineStage stage, bool enable, u32 index) {
auto &constantBuffer{pipelineStages[stage].constantBuffers[index]}; auto &targetConstantBuffer{pipelineStages[stage].constantBuffers[index]};
if (enable) if (auto selector{GetConstantBufferSelector()}; selector && enable)
constantBuffer = GetConstantBufferSelector().value(); targetConstantBuffer = *selector;
else else
constantBuffer = {}; targetConstantBuffer = {};
} }
/* Rasterizer State */ /* Rasterizer State */