The-Homebrew-Cloud/skyline
2
0
Fork 0
mirror of https://github.com/skyline-emu/skyline.git synced 2024-11-30 07:54:15 +01:00
Code Issues Packages Projects Releases Wiki Activity

Implement descriptor update batching and push descriptors

Browse Source 
Batching helps to avoid the need to attach so many objects to the fence cycle, which ends up taking a fair bit of time due to the allocation required.
This commit is contained in:
Billy Laws 2022-10-09 13:58:50 +01:00
parent 62a165b51e
commit 9ce848d4e0
3 changed files with 47 additions and 23 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

29
app/src/main/cpp/skyline/gpu/interconnect/maxwell_3d/maxwell_3d.cpp
View File

@ -26,15 +26,23 @@ namespace skyline::gpu::interconnect::maxwell3d {
textures{manager, registerBundle.texturePoolRegisters}, textures{manager, registerBundle.texturePoolRegisters},
directState{activeState.directState} { directState{activeState.directState} {
executor.AddFlushCallback([this] { executor.AddFlushCallback([this] {
if (attachedDescriptorSets) {
ctx.executor.AttachDependency(attachedDescriptorSets);
attachedDescriptorSets = nullptr;
activeDescriptorSet = nullptr;
}
activeState.MarkAllDirty(); activeState.MarkAllDirty();
constantBuffers.MarkAllDirty(); constantBuffers.MarkAllDirty();
samplers.MarkAllDirty(); samplers.MarkAllDirty();
textures.MarkAllDirty(); textures.MarkAllDirty();
quadConversionBufferAttached = false; quadConversionBufferAttached = false;
constantBuffers.DisableQuickBind();
}); });
executor.AddPipelineChangeCallback([this] { executor.AddPipelineChangeCallback([this] {
activeState.MarkAllDirty(); activeState.MarkAllDirty();
activeDescriptorSet = nullptr;
}); });
} }
@ -232,14 +240,23 @@ namespace skyline::gpu::interconnect::maxwell3d {
builder.SetPipeline(pipeline->compiledPipeline.pipeline); builder.SetPipeline(pipeline->compiledPipeline.pipeline);
if (descUpdateInfo) { if (descUpdateInfo) {
auto newSet{std::make_shared<DescriptorAllocator::ActiveDescriptorSet>(ctx.gpu.descriptor.AllocateSet(descUpdateInfo->descriptorSetLayout))}; if (ctx.gpu.traits.supportsPushDescriptors) {
ctx.executor.cycle->AttachObject(newSet); builder.SetDescriptorSetWithPush(descUpdateInfo);
} else {
if (!attachedDescriptorSets)
attachedDescriptorSets = std::make_shared<boost::container::static_vector<DescriptorAllocator::ActiveDescriptorSet, DescriptorBatchSize>>();
// Descriptor set lifetime is bound to the current cycle so we can safely use a raw pointer from now on auto newSet{&attachedDescriptorSets->emplace_back(ctx.gpu.descriptor.AllocateSet(descUpdateInfo->descriptorSetLayout))};
auto *oldSet{activeDescriptorSet}; auto *oldSet{activeDescriptorSet};
activeDescriptorSet = newSet.get(); activeDescriptorSet = newSet;
builder.SetDescriptorSetWithUpdate(descUpdateInfo, activeDescriptorSet, oldSet); builder.SetDescriptorSetWithUpdate(descUpdateInfo, activeDescriptorSet, oldSet);
if (attachedDescriptorSets->size() == DescriptorBatchSize) {
ctx.executor.AttachDependency(attachedDescriptorSets);
attachedDescriptorSets.reset();
}
}
} }
auto stateUpdater{builder.Build()}; auto stateUpdater{builder.Build()};

2
app/src/main/cpp/skyline/gpu/interconnect/maxwell_3d/maxwell_3d.h
View File

@ -48,6 +48,8 @@ namespace skyline::gpu::interconnect::maxwell3d {
std::shared_ptr<memory::Buffer> quadConversionBuffer{}; std::shared_ptr<memory::Buffer> quadConversionBuffer{};
bool quadConversionBufferAttached{}; bool quadConversionBufferAttached{};
static constexpr size_t DescriptorBatchSize{0x100};
std::shared_ptr<boost::container::static_vector<DescriptorAllocator::ActiveDescriptorSet, DescriptorBatchSize>> attachedDescriptorSets;
DescriptorAllocator::ActiveDescriptorSet *activeDescriptorSet{}; DescriptorAllocator::ActiveDescriptorSet *activeDescriptorSet{};
size_t UpdateQuadConversionBuffer(u32 count, u32 firstVertex); size_t UpdateQuadConversionBuffer(u32 count, u32 firstVertex);

39
app/src/main/cpp/skyline/gpu/interconnect/maxwell_3d/state_updater.h
View File

@ -205,31 +205,36 @@ namespace skyline::gpu::interconnect::maxwell3d {
} }
} }
// Set the destination/(source) descriptor set(s) for all writes/(copies) if constexpr (PushDescriptor) {
for (auto &write : updateInfo->writes) commandBuffer.pushDescriptorSetKHR(updateInfo->bindPoint, updateInfo->pipelineLayout, updateInfo->descriptorSetIndex, updateInfo->writes);
write.dstSet = **dstSet; } else {
// Set the destination/(source) descriptor set(s) for all writes/(copies)
for (auto &write : updateInfo->writes)
write.dstSet = **dstSet;
for (auto &copy : updateInfo->copies) { for (auto &copy : updateInfo->copies) {
copy.dstSet = **dstSet; copy.dstSet = **dstSet;
copy.srcSet = **srcSet; copy.srcSet = **srcSet;
}
// Perform the updates, doing copies first to avoid overwriting
if (!updateInfo->copies.empty())
gpu.vkDevice.updateDescriptorSets({}, updateInfo->copies);
if (!updateInfo->writes.empty())
gpu.vkDevice.updateDescriptorSets(updateInfo->writes, {});
// Bind the updated descriptor set and we're done!
commandBuffer.bindDescriptorSets(updateInfo->bindPoint, updateInfo->pipelineLayout, updateInfo->descriptorSetIndex, **dstSet, {});
} }
// Perform the updates, doing copies first to avoid overwriting
if (!updateInfo->copies.empty())
gpu.vkDevice.updateDescriptorSets({}, updateInfo->copies);
if (!updateInfo->writes.empty())
gpu.vkDevice.updateDescriptorSets(updateInfo->writes, {});
// Bind the updated descriptor set and we're done!
commandBuffer.bindDescriptorSets(updateInfo->bindPoint, updateInfo->pipelineLayout, updateInfo->descriptorSetIndex, **dstSet, {});
} }
DescriptorUpdateInfo *updateInfo; DescriptorUpdateInfo *updateInfo;
DescriptorAllocator::ActiveDescriptorSet *srcSet; DescriptorAllocator::ActiveDescriptorSet *srcSet;
DescriptorAllocator::ActiveDescriptorSet *dstSet; DescriptorAllocator::ActiveDescriptorSet *dstSet;
}; };
using SetDescriptorSetWithUpdateCmd = CmdHolder<SetDescriptorSetWithUpdateCmdImpl>; using SetDescriptorSetWithUpdateCmd = CmdHolder<SetDescriptorSetCmdImpl<false>>;
using SetDescriptorSetWithPushCmd = CmdHolder<SetDescriptorSetCmdImpl<true>>;
struct SetPipelineCmdImpl { struct SetPipelineCmdImpl {
void Record(GPU &gpu, vk::raii::CommandBuffer &commandBuffer) { void Record(GPU &gpu, vk::raii::CommandBuffer &commandBuffer) {