Traktor: Async compute support in Vulkan renderer.

code/Core/Containers/StaticVector.h

@@ -8,6 +8,7 @@
  */
 #pragma once
 
+#include <initializer_list>
 #include <iterator>
 #include "Core/Config.h"
 #include "Core/Containers/VectorConstructor.h"
@@ -270,6 +271,13 @@ class StaticVector
 
 	StaticVector() = default;
 
+	StaticVector(const std::initializer_list< ItemType >& iv) noexcept
+	:	m_size(0)
+	{
+		for (auto it = std::begin(iv); it != std::end(iv); ++it)
+			push_back(*it);
+	}
+
 	explicit StaticVector(size_t size)
 	:	m_size(size)
 	{

code/Render/Context/RenderBlock.cpp

@@ -36,7 +36,8 @@ void ComputeRenderBlock::render(IRenderView* renderView) const
 
 	renderView->compute(
 		program,
-		workSize
+		workSize,
+		asynchronous
 	);
 
 	T_CONTEXT_POP_MARKER(renderView);

code/Render/Context/RenderBlock.h

@@ -69,6 +69,7 @@ class T_DLLCLASS ComputeRenderBlock : public RenderBlock
 	IProgram* program = nullptr;
 	ProgramParameters* programParams = nullptr;
 	int32_t workSize[3] = { 1, 1, 1 };
+	bool asynchronous = false;
 
 	virtual void render(IRenderView* renderView) const override final;
 };

code/Render/IRenderView.h

@@ -175,7 +175,7 @@ class T_DLLCLASS IRenderView : public Object
 	 *
 	 * \param workSize Work size, 3 dimensional size.
 	 */
-	virtual void compute(IProgram* program, const int32_t* workSize) = 0;
+	virtual void compute(IProgram* program, const int32_t* workSize, bool asynchronous) = 0;
 
 	/*! Enqueue indirect compute task.
 	 *

code/Render/Vrfy/RenderViewVrfy.cpp

@@ -325,7 +325,7 @@ void RenderViewVrfy::drawIndirect(const IBufferView* vertexBuffer, const IVertex
 	m_renderView->drawIndirect(wrappedVertexView, vl->getWrappedVertexLayout(), wrappedIndexView, indexType, programVrfy->m_program, primitiveType, wrappedDrawView, drawOffset, drawCount);
 }
 
-void RenderViewVrfy::compute(IProgram* program, const int32_t* workSize)
+void RenderViewVrfy::compute(IProgram* program, const int32_t* workSize, bool asynchronous)
 {
 	T_CAPTURE_TRACE(L"compute");
 	T_CAPTURE_ASSERT(m_insideFrame, L"Cannot compute outside of beginFrame/endFrame.");
@@ -340,7 +340,7 @@ void RenderViewVrfy::compute(IProgram* program, const int32_t* workSize)
 
 	programVrfy->verify();
 
-	m_renderView->compute(programVrfy->m_program, workSize);
+	m_renderView->compute(programVrfy->m_program, workSize, asynchronous);
 }
 
 void RenderViewVrfy::computeIndirect(IProgram* program, const IBufferView* workBuffer, uint32_t workOffset)

code/Render/Vrfy/RenderViewVrfy.h

@@ -93,7 +93,7 @@ class RenderViewVrfy : public IRenderView
 
 	virtual void drawIndirect(const IBufferView* vertexBuffer, const IVertexLayout* vertexLayout, const IBufferView* indexBuffer, IndexType indexType, IProgram* program, PrimitiveType primitiveType, const IBufferView* drawBuffer, uint32_t drawOffset, uint32_t drawCount) override final;
 
-	virtual void compute(IProgram* program, const int32_t* workSize) override final;
+	virtual void compute(IProgram* program, const int32_t* workSize, bool asynchronous) override final;
 
 	virtual void computeIndirect(IProgram* program, const IBufferView* workBuffer, uint32_t workOffset) override final;
 

code/Render/Vulkan/AccelerationStructureVk.cpp

@@ -192,7 +192,7 @@ Ref< AccelerationStructureVk > AccelerationStructureVk::createTopLevel(Context*
 
 Ref< AccelerationStructureVk > AccelerationStructureVk::createBottomLevel(Context* context, const Buffer* vertexBuffer, const IVertexLayout* vertexLayout, const Buffer* indexBuffer, IndexType indexType, const AlignedVector< Primitives >& primitives)
 {
-	auto commandBuffer = context->getGraphicsQueue()->acquireCommandBuffer(T_FILE_LINE_W);
+	auto commandBuffer = context->getGraphicsQueue()->acquireCommandBuffer(L"AccelerationStructureVk::createBottomLevel");
 
 	Ref< AccelerationStructureVk > as = new AccelerationStructureVk(context);
 	as->m_scratchAlignment = getScratchAlignment(context);

code/Render/Vulkan/BufferStaticVk.cpp

@@ -65,7 +65,7 @@ void BufferStaticVk::unlock()
 {
 	m_stageBuffer->unlock();
 
-	auto commandBuffer = m_context->getGraphicsQueue()->acquireCommandBuffer(T_FILE_LINE_W);
+	auto commandBuffer = m_context->getGraphicsQueue()->acquireCommandBuffer(L"BufferStaticVk::unlock");
 
 	const VkBufferCopy bc =
 	{

code/Render/Vulkan/Private/CommandBuffer.cpp

@@ -48,7 +48,7 @@ bool CommandBuffer::reset()
 	return true;
 }
 
-bool CommandBuffer::submit(VkSemaphore waitSemaphore, VkPipelineStageFlags waitStageFlags, VkSemaphore signalSemaphore)
+bool CommandBuffer::submit(const StaticVector< VkSemaphore, 2 >& waitSemaphores, const StaticVector< VkPipelineStageFlags, 2 >& waitStageFlags, VkSemaphore signalSemaphore)
 {
 	T_ASSERT(ThreadManager::getInstance().getCurrentThread() == m_thread);
 	T_ASSERT(!m_submitted);
@@ -59,11 +59,12 @@ bool CommandBuffer::submit(VkSemaphore waitSemaphore, VkPipelineStageFlags waitS
 	VkSubmitInfo si = {};
 	si.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO;
 
-	if (waitSemaphore != VK_NULL_HANDLE)
+	if (!waitSemaphores.empty())
 	{
-		si.waitSemaphoreCount = 1;
-		si.pWaitSemaphores = &waitSemaphore;
-		si.pWaitDstStageMask = &waitStageFlags;
+		T_ASSERT(waitSemaphores.size() == waitStageFlags.size());
+		si.waitSemaphoreCount = (uint32_t)waitSemaphores.size();
+		si.pWaitSemaphores = waitSemaphores.c_ptr();
+		si.pWaitDstStageMask = waitStageFlags.c_ptr();
 	}
 
 	si.commandBufferCount = 1;
@@ -101,7 +102,7 @@ bool CommandBuffer::wait()
 
 bool CommandBuffer::submitAndWait()
 {
-	if (!submit(VK_NULL_HANDLE, 0, VK_NULL_HANDLE))
+	if (!submit({}, {}, VK_NULL_HANDLE))
 		return false;
 	if (!wait())
 		return false;

code/Render/Vulkan/Private/CommandBuffer.h

@@ -1,6 +1,6 @@
 /*
  * TRAKTOR
- * Copyright (c) 2022 Anders Pistol.
+ * Copyright (c) 2022-2024 Anders Pistol.
  *
  * This Source Code Form is subject to the terms of the Mozilla Public
  * License, v. 2.0. If a copy of the MPL was not distributed with this
@@ -9,6 +9,7 @@
 #pragma once
 
 #include "Core/Object.h"
+#include "Core/Containers/StaticVector.h"
 #include "Render/Vulkan/Private/ApiHeader.h"
 
 namespace traktor
@@ -34,7 +35,7 @@ class CommandBuffer : public Object
 
 	bool reset();
 
-	bool submit(VkSemaphore waitSemaphore, VkPipelineStageFlags waitStageFlags, VkSemaphore signalSemaphore);
+	bool submit(const StaticVector< VkSemaphore, 2 >& waitSemaphores, const StaticVector< VkPipelineStageFlags, 2 >& waitStageFlags, VkSemaphore signalSemaphore);
 
 	bool wait();
 

code/Render/Vulkan/Private/Context.cpp

@@ -30,12 +30,14 @@ Context::Context(
 	VkPhysicalDevice physicalDevice,
 	VkDevice logicalDevice,
 	VmaAllocator allocator,
-	uint32_t graphicsQueueIndex
+	uint32_t graphicsQueueIndex,
+	uint32_t computeQueueIndex
 )
 :	m_physicalDevice(physicalDevice)
 ,	m_logicalDevice(logicalDevice)
 ,	m_allocator(allocator)
 ,	m_graphicsQueueIndex(graphicsQueueIndex)
+,	m_computeQueueIndex(computeQueueIndex)
 ,	m_sampledResourceIndexAllocator(0, MaxBindlessResources - 1)
 ,	m_storageResourceIndexAllocator(0, MaxBindlessResources - 1)
 ,	m_bufferResourceIndexAllocator(0, MaxBindlessResources - 1)
@@ -65,6 +67,7 @@ bool Context::create()
 
 	// Create queues.
 	m_graphicsQueue = Queue::create(this, m_graphicsQueueIndex);
+	m_computeQueue = Queue::create(this, m_computeQueueIndex);
 
 	// Create pipeline cache.
 	VkPipelineCacheCreateInfo pcci = {};
@@ -118,13 +121,15 @@ bool Context::create()
 	dps[5].type = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER;
 	dps[5].descriptorCount = MaxBindlessResources;
 
-	VkDescriptorPoolCreateInfo dpci = {};
-	dpci.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO;
-	dpci.pNext = nullptr;
-	dpci.flags = VK_DESCRIPTOR_POOL_CREATE_FREE_DESCRIPTOR_SET_BIT | VK_DESCRIPTOR_POOL_CREATE_UPDATE_AFTER_BIND_BIT_EXT;
-	dpci.maxSets = 32000;
-	dpci.poolSizeCount = sizeof_array(dps);
-	dpci.pPoolSizes = dps;
+	const VkDescriptorPoolCreateInfo dpci =
+	{
+		.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO,
+		.pNext = nullptr,
+		.flags = VK_DESCRIPTOR_POOL_CREATE_FREE_DESCRIPTOR_SET_BIT | VK_DESCRIPTOR_POOL_CREATE_UPDATE_AFTER_BIND_BIT_EXT,
+		.maxSets = 32000,
+		.poolSizeCount = sizeof_array(dps),
+		.pPoolSizes = dps
+	};
 
 	vkCreateDescriptorPool(m_logicalDevice, &dpci, nullptr, &m_descriptorPool);
 

code/Render/Vulkan/Private/Context.h

@@ -49,7 +49,8 @@ class Context : public Object
 		VkPhysicalDevice physicalDevice,
 		VkDevice logicalDevice,
 		VmaAllocator allocator,
-		uint32_t graphicsQueueIndex
+		uint32_t graphicsQueueIndex,
+		uint32_t computeQueueIndex
 	);
 
 	virtual ~Context();
@@ -90,6 +91,8 @@ class Context : public Object
 
 	Queue* getGraphicsQueue() const { return m_graphicsQueue; }
 
+	Queue* getComputeQueue() const { return m_computeQueue; }
+
 	UniformBufferPool* getUniformBufferPool(int32_t index) const { return m_uniformBufferPools[index]; }
 
 	VkDescriptorSetLayout getBindlessTexturesSetLayout() const { return m_bindlessTexturesDescriptorLayout; }
@@ -121,10 +124,12 @@ class Context : public Object
 	VkDevice m_logicalDevice;
 	VmaAllocator m_allocator;
 	uint32_t m_graphicsQueueIndex;
+	uint32_t m_computeQueueIndex;
 	VkPipelineCache m_pipelineCache = 0;
 	VkDescriptorPool m_descriptorPool = 0;
 	int32_t m_views = 0;
 	Ref< Queue > m_graphicsQueue;
+	Ref< Queue > m_computeQueue;
 	Ref< UniformBufferPool > m_uniformBufferPools[3];
 	Semaphore m_cleanupLock;
 	Semaphore m_resourceIndexLock;

code/Render/Vulkan/Private/Queue.cpp

@@ -18,18 +18,19 @@ namespace traktor::render
 
 T_IMPLEMENT_RTTI_CLASS(L"traktor.render.Queue", Queue, Object)
 
-thread_local VkCommandPool Queue::ms_commandPool;
+thread_local VkCommandPool Queue::ms_commandPools[32];
 
 Ref< Queue > Queue::create(Context* context, uint32_t queueIndex)
 {
 	VkQueue queue;
+	T_FATAL_ASSERT(queueIndex < sizeof_array(ms_commandPools));
 	vkGetDeviceQueue(context->getLogicalDevice(), queueIndex, 0, &queue);
 	return new Queue(context, queue, queueIndex);
 }
 
 Ref< CommandBuffer > Queue::acquireCommandBuffer(const wchar_t* const tag)
 {
-	if (!ms_commandPool)
+	if (!ms_commandPools[m_queueIndex])
 	{
 		VkCommandPool commandPool;
 
@@ -40,14 +41,14 @@ Ref< CommandBuffer > Queue::acquireCommandBuffer(const wchar_t* const tag)
 		if (vkCreateCommandPool(m_context->getLogicalDevice(), &cpci, 0, &commandPool) != VK_SUCCESS)
 			return nullptr;
 
-		ms_commandPool = commandPool;
+		ms_commandPools[m_queueIndex] = commandPool;
 	}
 
 	VkCommandBuffer commandBuffer = 0;
 
 	VkCommandBufferAllocateInfo cbai = {};
 	cbai.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO;
-	cbai.commandPool = ms_commandPool;
+	cbai.commandPool = ms_commandPools[m_queueIndex];
 	cbai.level = VK_COMMAND_BUFFER_LEVEL_PRIMARY;
 	cbai.commandBufferCount = 1;
 	if (vkAllocateCommandBuffers(m_context->getLogicalDevice(), &cbai, &commandBuffer) != VK_SUCCESS)
@@ -61,7 +62,7 @@ Ref< CommandBuffer > Queue::acquireCommandBuffer(const wchar_t* const tag)
 	if (vkBeginCommandBuffer(commandBuffer, &cbbi) != VK_SUCCESS)
 		return nullptr;
 
-	return new CommandBuffer(m_context, this, ms_commandPool, commandBuffer);
+	return new CommandBuffer(m_context, this, ms_commandPools[m_queueIndex], commandBuffer);
 }
 
 VkResult Queue::submit(const VkSubmitInfo& si, VkFence fence)

code/Render/Vulkan/Private/Queue.h

@@ -39,16 +39,14 @@ class Queue : public Object
 
 	uint32_t getQueueIndex() const { return m_queueIndex; }
 
-	VkCommandPool getCommandPool() const { return ms_commandPool; }
-
 private:
 	friend class Context;
 
 	Context* m_context;
 	VkQueue m_queue;
 	uint32_t m_queueIndex;
 	Semaphore m_lock;
-	static thread_local VkCommandPool ms_commandPool;
+	static thread_local VkCommandPool ms_commandPools[32];
 
 	Queue(Context* context, VkQueue queue, uint32_t queueIndex);
 };