Implement descriptor update batching and push descriptors

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.

app/src/main/cpp/skyline/gpu/interconnect/maxwell_3d/maxwell_3d.cpp

@@ -26,15 +26,23 @@ namespace skyline::gpu::interconnect::maxwell3d {
           textures{manager, registerBundle.texturePoolRegisters},
           directState{activeState.directState} {
         executor.AddFlushCallback([this] {
+            if (attachedDescriptorSets) {
+                ctx.executor.AttachDependency(attachedDescriptorSets);
+                attachedDescriptorSets = nullptr;
+                activeDescriptorSet = nullptr;
+            }
+
             activeState.MarkAllDirty();
             constantBuffers.MarkAllDirty();
             samplers.MarkAllDirty();
             textures.MarkAllDirty();
             quadConversionBufferAttached = false;
+            constantBuffers.DisableQuickBind();
         });
 
         executor.AddPipelineChangeCallback([this] {
             activeState.MarkAllDirty();
+            activeDescriptorSet = nullptr;
         });
     }
 
@@ -232,14 +240,23 @@ namespace skyline::gpu::interconnect::maxwell3d {
             builder.SetPipeline(pipeline->compiledPipeline.pipeline);
 
         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};
-            activeDescriptorSet = newSet.get();
+                activeDescriptorSet = newSet;
 
                 builder.SetDescriptorSetWithUpdate(descUpdateInfo, activeDescriptorSet, oldSet);
+
+                if (attachedDescriptorSets->size() == DescriptorBatchSize) {
+                    ctx.executor.AttachDependency(attachedDescriptorSets);
+                    attachedDescriptorSets.reset();
+                }
+            }
         }
 
         auto stateUpdater{builder.Build()};

app/src/main/cpp/skyline/gpu/interconnect/maxwell_3d/maxwell_3d.h

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

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

@@ -205,6 +205,9 @@ namespace skyline::gpu::interconnect::maxwell3d {
                 }
             }
 
+            if constexpr (PushDescriptor) {
+                commandBuffer.pushDescriptorSetKHR(updateInfo->bindPoint, updateInfo->pipelineLayout, updateInfo->descriptorSetIndex, updateInfo->writes);
+            } else {
                 // Set the destination/(source) descriptor set(s) for all writes/(copies)
                 for (auto &write : updateInfo->writes)
                     write.dstSet = **dstSet;
@@ -224,12 +227,14 @@ namespace skyline::gpu::interconnect::maxwell3d {
                 // Bind the updated descriptor set and we're done!
                 commandBuffer.bindDescriptorSets(updateInfo->bindPoint, updateInfo->pipelineLayout, updateInfo->descriptorSetIndex, **dstSet, {});
             }
+        }
 
         DescriptorUpdateInfo *updateInfo;
         DescriptorAllocator::ActiveDescriptorSet *srcSet;
         DescriptorAllocator::ActiveDescriptorSet *dstSet;
     };
-    using SetDescriptorSetWithUpdateCmd = CmdHolder<SetDescriptorSetWithUpdateCmdImpl>;
+    using SetDescriptorSetWithUpdateCmd = CmdHolder<SetDescriptorSetCmdImpl<false>>;
+    using SetDescriptorSetWithPushCmd = CmdHolder<SetDescriptorSetCmdImpl<true>>;
 
     struct SetPipelineCmdImpl {
         void Record(GPU &gpu, vk::raii::CommandBuffer &commandBuffer) {