MVKCommandBuffer: Optimize management of command storage.

The main motivation behind this commit is to reduce the time needed to free or
reset a command buffer. With the current scheme, we need to loop over all
recorded commands and return them to their pools. This can take a significant
amount of time in games, of the order of a millisecond.

With this change, instead of pooling individual commands by their type, we
instead keep chunks of memory in each command buffer, from which we bump
allocate memory for the commands to be recorded. MVKCommandVector is introduced
to use the same memory pool for dynamic arrays of data associated with the
commands.

Pointers to commands that actually have non-trivial destructors (i.e. they keep
references to other objects) are stored separately, so that we can call their
destructors when reseting the command buffer. All the memory simply gets
discarded by setting the current chunk's index and offset to zero.

As a side note, specific applications might opt to not retain pipeline layouts
in bind/push descriptor commands (introduced by
30113fc) if it's known not to be needed, to
further reduce time spent resetting command buffers.

MoltenVK/MoltenVK.xcodeproj/project.pbxproj

@@ -524,7 +524,6 @@
 		A9B51BD2225E986A00AC74D2 /* MVKOSExtensions.mm */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.cpp.objcpp; path = MVKOSExtensions.mm; sourceTree = "<group>"; };
 		A9B51BD6225E986A00AC74D2 /* MVKOSExtensions.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; path = MVKOSExtensions.h; sourceTree = "<group>"; };
 		A9B8EE0A1A98D796009C5A02 /* libMoltenVK.a */ = {isa = PBXFileReference; explicitFileType = archive.ar; includeInIndex = 0; path = libMoltenVK.a; sourceTree = BUILT_PRODUCTS_DIR; };
-		A9C83DCD24533E22003E5261 /* MVKCommandTypePools.def */ = {isa = PBXFileReference; explicitFileType = sourcecode.cpp.h; fileEncoding = 4; path = MVKCommandTypePools.def; sourceTree = "<group>"; };
 		A9C86CB61C55B8350096CAF2 /* MoltenVKShaderConverter.xcodeproj */ = {isa = PBXFileReference; lastKnownFileType = "wrapper.pb-project"; name = MoltenVKShaderConverter.xcodeproj; path = ../MoltenVKShaderConverter/MoltenVKShaderConverter.xcodeproj; sourceTree = "<group>"; };
 		A9C96DCE1DDC20C20053187F /* MVKMTLBufferAllocation.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; path = MVKMTLBufferAllocation.h; sourceTree = "<group>"; };
 		A9C96DCF1DDC20C20053187F /* MVKMTLBufferAllocation.mm */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.cpp.objcpp; path = MVKMTLBufferAllocation.mm; sourceTree = "<group>"; };
@@ -605,7 +604,6 @@
 				A94FB77B1C7DFB4800632CA3 /* MVKCommandPool.mm */,
 				A95870F61C90D29F009EB096 /* MVKCommandResourceFactory.h */,
 				A95870F71C90D29F009EB096 /* MVKCommandResourceFactory.mm */,
-				A9C83DCD24533E22003E5261 /* MVKCommandTypePools.def */,
 				A9C96DCE1DDC20C20053187F /* MVKMTLBufferAllocation.h */,
 				A9C96DCF1DDC20C20053187F /* MVKMTLBufferAllocation.mm */,
 				A9E4B7881E1D8AF10046A4CE /* MVKMTLResourceBindings.h */,

MoltenVK/MoltenVK/Commands/MVKCmdDebug.h

@@ -47,10 +47,6 @@ class MVKCmdDebugMarkerBegin : public MVKCmdDebugMarker {
 
 public:
 	void encode(MVKCommandEncoder* cmdEncoder) override;
-
-protected:
-	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;
-
 };
 
 
@@ -64,10 +60,6 @@ class MVKCmdDebugMarkerEnd : public MVKCommand {
 	VkResult setContent(MVKCommandBuffer* cmdBuff);
 
 	void encode(MVKCommandEncoder* cmdEncoder) override;
-
-protected:
-	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;
-
 };
 
 
@@ -79,10 +71,6 @@ class MVKCmdDebugMarkerInsert : public MVKCmdDebugMarker {
 
 public:
 	void encode(MVKCommandEncoder* cmdEncoder) override;
-
-protected:
-	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;
-
 };
 
 

MoltenVK/MoltenVK/Commands/MVKCmdDebug.mm

@@ -28,7 +28,7 @@
 									   const float color[4]) {
 	[_markerName release];
 	_markerName = [[NSString alloc] initWithUTF8String: pMarkerName];	// retained
-
+    cmdBuff->destroyList.push_back(this);
 	return VK_SUCCESS;
 }
 

MoltenVK/MoltenVK/Commands/MVKCmdDispatch.h

@@ -37,9 +37,7 @@ class MVKCmdDispatch : public MVKCommand {
     void encode(MVKCommandEncoder* cmdEncoder) override;
 
 protected:
-	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;
-
-	uint32_t _baseGroupX;
+    uint32_t _baseGroupX;
 	uint32_t _baseGroupY;
 	uint32_t _baseGroupZ;
 	uint32_t _groupCountX;
@@ -60,8 +58,6 @@ class MVKCmdDispatchIndirect : public MVKCommand {
 	void encode(MVKCommandEncoder* cmdEncoder) override;
 
 protected:
-	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;
-
 	id<MTLBuffer> _mtlIndirectBuffer;
 	VkDeviceSize _mtlIndirectBufferOffset;
 };

MoltenVK/MoltenVK/Commands/MVKCmdDraw.h

@@ -30,9 +30,7 @@
 
 /**
  * Vulkan command to bind buffers containing vertex content.
- * Template class to balance vector pre-allocations between very common low counts and fewer larger counts.
  */
-template <size_t N>
 class MVKCmdBindVertexBuffers : public MVKCommand {
 
 public:
@@ -45,16 +43,9 @@ class MVKCmdBindVertexBuffers : public MVKCommand {
     void encode(MVKCommandEncoder* cmdEncoder) override;
 
 protected:
-	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;
-
-    MVKSmallVector<MVKMTLBufferBinding, N> _bindings;
+    MVKCommandVector<MVKMTLBufferBinding> _bindings;
 };
 
-// Concrete template class implementations.
-typedef MVKCmdBindVertexBuffers<1> MVKCmdBindVertexBuffers1;
-typedef MVKCmdBindVertexBuffers<2> MVKCmdBindVertexBuffers2;
-typedef MVKCmdBindVertexBuffers<8> MVKCmdBindVertexBuffersMulti;
-
 
 #pragma mark -
 #pragma mark MVKCmdBindIndexBuffer
@@ -71,8 +62,6 @@ class MVKCmdBindIndexBuffer : public MVKCommand {
 	void encode(MVKCommandEncoder* cmdEncoder) override;
 
 protected:
-	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;
-
     MVKIndexMTLBufferBinding _binding;
 };
 
@@ -93,8 +82,6 @@ class MVKCmdDraw : public MVKCommand {
     void encode(MVKCommandEncoder* cmdEncoder) override;
 
 protected:
-	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;
-
 	uint32_t _firstVertex;
 	uint32_t _vertexCount;
 	uint32_t _firstInstance;
@@ -119,8 +106,6 @@ class MVKCmdDrawIndexed : public MVKCommand {
 	void encode(MVKCommandEncoder* cmdEncoder) override;
 
 protected:
-	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;
-
 	uint32_t _firstIndex;
 	uint32_t _indexCount;
 	int32_t	_vertexOffset;
@@ -145,8 +130,6 @@ class MVKCmdDrawIndirect : public MVKCommand {
 	void encode(MVKCommandEncoder* cmdEncoder) override;
 
 protected:
-	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;
-
 	id<MTLBuffer> _mtlIndirectBuffer;
 	VkDeviceSize _mtlIndirectBufferOffset;
 	uint32_t _mtlIndirectBufferStride;
@@ -170,8 +153,6 @@ class MVKCmdDrawIndexedIndirect : public MVKCommand {
 	void encode(MVKCommandEncoder* cmdEncoder) override;
 
 protected:
-	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;
-
 	id<MTLBuffer> _mtlIndirectBuffer;
 	VkDeviceSize _mtlIndirectBufferOffset;
 	uint32_t _mtlIndirectBufferStride;

MoltenVK/MoltenVK/Commands/MVKCmdDraw.mm

@@ -28,14 +28,14 @@
 #pragma mark -
 #pragma mark MVKCmdBindVertexBuffers
 
-template <size_t N>
-VkResult MVKCmdBindVertexBuffers<N>::setContent(MVKCommandBuffer* cmdBuff,
+VkResult MVKCmdBindVertexBuffers::setContent(MVKCommandBuffer* cmdBuff,
 												uint32_t startBinding,
 												uint32_t bindingCount,
 												const VkBuffer* pBuffers,
 												const VkDeviceSize* pOffsets) {
 
 	MVKDevice* mvkDvc = cmdBuff->getDevice();
+    _bindings.alc.cmdBuffer = cmdBuff;
 	_bindings.clear();	// Clear for reuse
     _bindings.reserve(bindingCount);
     MVKMTLBufferBinding b;
@@ -50,15 +50,10 @@
 	return VK_SUCCESS;
 }
 
-template <size_t N>
-void MVKCmdBindVertexBuffers<N>::encode(MVKCommandEncoder* cmdEncoder) {
+void MVKCmdBindVertexBuffers::encode(MVKCommandEncoder* cmdEncoder) {
     for (auto& b : _bindings) { cmdEncoder->_graphicsResourcesState.bindBuffer(kMVKShaderStageVertex, b); }
 }
 
-template class MVKCmdBindVertexBuffers<1>;
-template class MVKCmdBindVertexBuffers<2>;
-template class MVKCmdBindVertexBuffers<8>;
-
 
 #pragma mark -
 #pragma mark MVKCmdBindIndexBuffer

MoltenVK/MoltenVK/Commands/MVKCmdPipeline.h

@@ -35,9 +35,7 @@ class MVKDescriptorUpdateTemplate;
 
 /**
  * Vulkan command to add a pipeline barrier.
- * Template class to balance vector pre-allocations between very common low counts and fewer larger counts.
  */
-template <size_t N>
 class MVKCmdPipelineBarrier : public MVKCommand {
 
 public:
@@ -55,20 +53,14 @@ class MVKCmdPipelineBarrier : public MVKCommand {
 	void encode(MVKCommandEncoder* cmdEncoder) override;
 
 protected:
-	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;
 	bool coversTextures();
 
-	MVKSmallVector<MVKPipelineBarrier, N> _barriers;
+	MVKCommandVector<MVKPipelineBarrier> _barriers;
 	VkPipelineStageFlags _srcStageMask;
 	VkPipelineStageFlags _dstStageMask;
 	VkDependencyFlags _dependencyFlags;
 };
 
-// Concrete template class implementations.
-typedef MVKCmdPipelineBarrier<1> MVKCmdPipelineBarrier1;
-typedef MVKCmdPipelineBarrier<4> MVKCmdPipelineBarrier4;
-typedef MVKCmdPipelineBarrier<32> MVKCmdPipelineBarrierMulti;
-
 
 #pragma mark -
 #pragma mark MVKCmdBindPipeline
@@ -97,10 +89,6 @@ class MVKCmdBindGraphicsPipeline : public MVKCmdBindPipeline {
 	void encode(MVKCommandEncoder* cmdEncoder) override;
 
 	bool isTessellationPipeline() override;
-
-protected:
-	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;
-
 };
 
 
@@ -112,10 +100,6 @@ class MVKCmdBindComputePipeline : public MVKCmdBindPipeline {
 
 public:
 	void encode(MVKCommandEncoder* cmdEncoder) override;
-
-protected:
-	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;
-
 };
 
 
@@ -124,9 +108,7 @@ class MVKCmdBindComputePipeline : public MVKCmdBindPipeline {
 
 /**
  * Vulkan command to bind descriptor sets without dynamic offsets.
- * Template class to balance vector pre-allocations between very common low counts and fewer larger counts.
  */
-template <size_t N>
 class MVKCmdBindDescriptorSetsStatic : public MVKCommand {
 
 public:
@@ -142,30 +124,22 @@ class MVKCmdBindDescriptorSetsStatic : public MVKCommand {
 	~MVKCmdBindDescriptorSetsStatic() override;
 
 protected:
-	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;
 	void encode(MVKCommandEncoder* cmdEncoder, MVKArrayRef<uint32_t> dynamicOffsets);
 
-	MVKSmallVector<MVKDescriptorSet*, N> _descriptorSets;
+	MVKCommandVector<MVKDescriptorSet*> _descriptorSets;
 	MVKPipelineLayout* _pipelineLayout = nullptr;
 	VkPipelineBindPoint _pipelineBindPoint;
 	uint32_t _firstSet;
 };
 
-// Concrete template class implementations.
-typedef MVKCmdBindDescriptorSetsStatic<1> MVKCmdBindDescriptorSetsStatic1;
-typedef MVKCmdBindDescriptorSetsStatic<4> MVKCmdBindDescriptorSetsStatic4;
-typedef MVKCmdBindDescriptorSetsStatic<8> MVKCmdBindDescriptorSetsStaticMulti;
-
 
 #pragma mark -
 #pragma mark MVKCmdBindDescriptorSetsDynamic
 
 /**
  * Vulkan command to bind descriptor sets with dynamic offsets.
- * Template class to balance vector pre-allocations between very common low counts and fewer larger counts.
  */
-template <size_t N>
-class MVKCmdBindDescriptorSetsDynamic : public MVKCmdBindDescriptorSetsStatic<N> {
+class MVKCmdBindDescriptorSetsDynamic : public MVKCmdBindDescriptorSetsStatic {
 
 public:
 	VkResult setContent(MVKCommandBuffer* cmdBuff,
@@ -180,24 +154,16 @@ class MVKCmdBindDescriptorSetsDynamic : public MVKCmdBindDescriptorSetsStatic<N>
 	void encode(MVKCommandEncoder* cmdEncoder) override;
 
 protected:
-	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;
-
-	MVKSmallVector<uint32_t, N> _dynamicOffsets;
+    MVKCommandVector<uint32_t> _dynamicOffsets;
 };
 
-// Concrete template class implementations.
-typedef MVKCmdBindDescriptorSetsDynamic<4> MVKCmdBindDescriptorSetsDynamic4;
-typedef MVKCmdBindDescriptorSetsDynamic<8> MVKCmdBindDescriptorSetsDynamicMulti;
-
 
 #pragma mark -
 #pragma mark MVKCmdPushConstants
 
 /**
  * Vulkan command to bind push constants.
- * Template class to balance vector pre-allocations between very common low counts and fewer larger counts.
  */
-template <size_t N>
 class MVKCmdPushConstants : public MVKCommand {
 
 public:
@@ -211,19 +177,11 @@ class MVKCmdPushConstants : public MVKCommand {
 	void encode(MVKCommandEncoder* cmdEncoder) override;
 
 protected:
-	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;
-
-	MVKSmallVector<char, N> _pushConstants;
+	MVKCommandVector<char> _pushConstants;
 	VkShaderStageFlags _stageFlags;
 	uint32_t _offset;
 };
 
-// Concrete template class implementations.
-typedef MVKCmdPushConstants<64> MVKCmdPushConstants64;
-typedef MVKCmdPushConstants<128> MVKCmdPushConstants128;
-typedef MVKCmdPushConstants<512> MVKCmdPushConstantsMulti;
-
-
 #pragma mark -
 #pragma mark MVKCmdPushDescriptorSet
 
@@ -243,10 +201,9 @@ class MVKCmdPushDescriptorSet : public MVKCommand {
 	~MVKCmdPushDescriptorSet() override;
 
 protected:
-	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;
 	void clearDescriptorWrites();
 
-	MVKSmallVector<VkWriteDescriptorSet, 1> _descriptorWrites;
+	MVKCommandVector<VkWriteDescriptorSet> _descriptorWrites;
 	MVKPipelineLayout* _pipelineLayout = nullptr;
 	VkPipelineBindPoint _pipelineBindPoint;
 	uint32_t _set;
@@ -271,8 +228,6 @@ class MVKCmdPushDescriptorSetWithTemplate : public MVKCommand {
 	~MVKCmdPushDescriptorSetWithTemplate() override;
 
 protected:
-	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;
-
 	MVKDescriptorUpdateTemplate* _descUpdateTemplate;
 	MVKPipelineLayout* _pipelineLayout = nullptr;
 	void* _pData = nullptr;
@@ -305,10 +260,6 @@ class MVKCmdSetEvent : public MVKCmdSetResetEvent {
 
 public:
 	void encode(MVKCommandEncoder* cmdEncoder) override;
-
-protected:
-	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;
-
 };
 
 
@@ -320,10 +271,6 @@ class MVKCmdResetEvent : public MVKCmdSetResetEvent {
 
 public:
 	void encode(MVKCommandEncoder* cmdEncoder) override;
-
-protected:
-	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;
-
 };
 
 
@@ -333,9 +280,7 @@ class MVKCmdResetEvent : public MVKCmdSetResetEvent {
 /** Vulkan command to wait for an event to be signaled. */
 /**
  * Vulkan command to wait for an event to be signaled.
- * Template class to balance vector pre-allocations between very common low counts and fewer larger counts.
  */
-template <size_t N>
 class MVKCmdWaitEvents : public MVKCommand {
 
 public:
@@ -354,12 +299,5 @@ class MVKCmdWaitEvents : public MVKCommand {
 	void encode(MVKCommandEncoder* cmdEncoder) override;
 
 protected:
-	MVKCommandTypePool<MVKCommand>* getTypePool(MVKCommandPool* cmdPool) override;
-
-	MVKSmallVector<MVKEvent*, N> _mvkEvents;
-
+	MVKCommandVector<MVKEvent*> _mvkEvents;
 };
-
-// Concrete template class implementations.
-typedef MVKCmdWaitEvents<1> MVKCmdWaitEvents1;
-typedef MVKCmdWaitEvents<8> MVKCmdWaitEventsMulti;