Merge pull request #15589 from hrydgard/glsl-compilation-parallelization

Vulkan: Parallelize GLSL compilation

Common/GPU/Vulkan/VulkanQueueRunner.cpp

@@ -1182,18 +1182,11 @@ void VulkanQueueRunner::PerformRenderPass(const VKRStep &step, VkCommandBuffer c
 
 		case VKRRenderCommand::BIND_GRAPHICS_PIPELINE:
 		{
-			VKRGraphicsPipeline *pipeline = c.graphics_pipeline.pipeline;
-			if (pipeline->Pending()) {
-				// Stall processing, waiting for the compile queue to catch up.
-				std::unique_lock<std::mutex> lock(compileDoneMutex_);
-				while (!pipeline->pipeline) {
-					compileDone_.wait(lock);
-				}
-			}
-			if (pipeline->pipeline != lastGraphicsPipeline && pipeline->pipeline != VK_NULL_HANDLE) {
-				vkCmdBindPipeline(cmd, VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline->pipeline);
+			VkPipeline pipeline = c.graphics_pipeline.pipeline->BlockUntilReady();
+			if (pipeline != lastGraphicsPipeline && pipeline != VK_NULL_HANDLE) {
+				vkCmdBindPipeline(cmd, VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline);
 				pipelineLayout = c.pipeline.pipelineLayout;
-				lastGraphicsPipeline = pipeline->pipeline;
+				lastGraphicsPipeline = pipeline;
 				// Reset dynamic state so it gets refreshed with the new pipeline.
 				lastStencilWriteMask = -1;
 				lastStencilCompareMask = -1;
@@ -1204,18 +1197,11 @@ void VulkanQueueRunner::PerformRenderPass(const VKRStep &step, VkCommandBuffer c
 
 		case VKRRenderCommand::BIND_COMPUTE_PIPELINE:
 		{
-			VKRComputePipeline *pipeline = c.compute_pipeline.pipeline;
-			if (pipeline->Pending()) {
-				// Stall processing, waiting for the compile queue to catch up.
-				std::unique_lock<std::mutex> lock(compileDoneMutex_);
-				while (!pipeline->pipeline) {
-					compileDone_.wait(lock);
-				}
-			}
-			if (pipeline->pipeline != lastComputePipeline && pipeline->pipeline != VK_NULL_HANDLE) {
-				vkCmdBindPipeline(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, pipeline->pipeline);
+			VkPipeline pipeline = c.graphics_pipeline.pipeline->BlockUntilReady();
+			if (pipeline != lastComputePipeline && pipeline != VK_NULL_HANDLE) {
+				vkCmdBindPipeline(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, pipeline);
 				pipelineLayout = c.pipeline.pipelineLayout;
-				lastComputePipeline = pipeline->pipeline;
+				lastComputePipeline = pipeline;
 			}
 			break;
 		}
@@ -1335,7 +1321,6 @@ void VulkanQueueRunner::PerformRenderPass(const VKRStep &step, VkCommandBuffer c
 		}
 		default:
 			ERROR_LOG(G3D, "Unimpl queue command");
-			;
 		}
 	}
 	vkCmdEndRenderPass(cmd);

Common/GPU/Vulkan/VulkanQueueRunner.h

@@ -4,7 +4,7 @@
 #include <mutex>
 #include <condition_variable>
 
-
+#include "Common/Thread/Promise.h"
 #include "Common/Data/Collections/Hashmaps.h"
 #include "Common/GPU/Vulkan/VulkanContext.h"
 #include "Common/GPU/Vulkan/VulkanBarrier.h"
@@ -55,11 +55,11 @@ struct VkRenderData {
 			VkPipelineLayout pipelineLayout;
 		} pipeline;
 		struct {
-			VKRGraphicsPipeline *pipeline;
+			Promise<VkPipeline> *pipeline;
 			VkPipelineLayout pipelineLayout;
 		} graphics_pipeline;
 		struct {
-			VKRComputePipeline *pipeline;
+			Promise<VkPipeline> *pipeline;
 			VkPipelineLayout pipelineLayout;
 		} compute_pipeline;
 		struct {

Common/GPU/Vulkan/VulkanRenderManager.cpp

@@ -5,6 +5,7 @@
 
 #include "Common/Log.h"
 #include "Common/StringUtils.h"
+#include "Common/TimeUtil.h"
 
 #include "Common/GPU/Vulkan/VulkanAlloc.h"
 #include "Common/GPU/Vulkan/VulkanContext.h"
@@ -29,23 +30,54 @@ bool VKRGraphicsPipeline::Create(VulkanContext *vulkan) {
 		// Already failed to create this one.
 		return false;
 	}
+
+	// Fill in the last part of the desc since now it's time to block.
+	VkShaderModule vs = desc->vertexShader->BlockUntilReady();
+	VkShaderModule fs = desc->fragmentShader->BlockUntilReady();
+
+	if (!vs || !fs) {
+		ERROR_LOG(G3D, "Failed creating graphics pipeline - missing shader modules");
+		// We're kinda screwed here?
+		return false;
+	}
+
+	VkPipelineShaderStageCreateInfo ss[2]{};
+	ss[0].sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO;
+	ss[0].stage = VK_SHADER_STAGE_VERTEX_BIT;
+	ss[0].pSpecializationInfo = nullptr;
+	ss[0].module = vs;
+	ss[0].pName = "main";
+	ss[0].flags = 0;
+	ss[1].sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO;
+	ss[1].stage = VK_SHADER_STAGE_FRAGMENT_BIT;
+	ss[1].pSpecializationInfo = nullptr;
+	ss[1].module = fs;
+	ss[1].pName = "main";
+	ss[1].flags = 0;
+
+	desc->pipe.pStages = ss;
+	desc->pipe.stageCount = 2;
+
+	double start = time_now_d();
 	VkPipeline vkpipeline;
 	VkResult result = vkCreateGraphicsPipelines(vulkan->GetDevice(), desc->pipelineCache, 1, &desc->pipe, nullptr, &vkpipeline);
 
+	INFO_LOG(G3D, "Pipeline creation time: %0.2f ms", (time_now_d() - start) * 1000.0);
+
 	bool success = true;
 	if (result == VK_INCOMPLETE) {
 		// Bad (disallowed by spec) return value seen on Adreno in Burnout :(  Try to ignore?
 		// Would really like to log more here, we could probably attach more info to desc.
 		//
 		// At least create a null placeholder to avoid creating over and over if something is broken.
-		pipeline = VK_NULL_HANDLE;
+		pipeline->Post(VK_NULL_HANDLE);
 		success = false;
 	} else if (result != VK_SUCCESS) {
-		pipeline = VK_NULL_HANDLE;
+		pipeline->Post(VK_NULL_HANDLE);
 		ERROR_LOG(G3D, "Failed creating graphics pipeline! result='%s'", VulkanResultToString(result));
 		success = false;
 	} else {
-		pipeline = vkpipeline;
+		pipeline->Post(vkpipeline);
 	}
 
 	delete desc;
@@ -63,11 +95,11 @@ bool VKRComputePipeline::Create(VulkanContext *vulkan) {
 
 	bool success = true;
 	if (result != VK_SUCCESS) {
-		pipeline = VK_NULL_HANDLE;
+		pipeline->Post(VK_NULL_HANDLE);
 		ERROR_LOG(G3D, "Failed creating compute pipeline! result='%s'", VulkanResultToString(result));
 		success = false;
 	} else {
-		pipeline = vkpipeline;
+		pipeline->Post(vkpipeline);
 	}
 
 	delete desc;
@@ -449,6 +481,12 @@ void VulkanRenderManager::CompileThreadFunc() {
 		if (!run_) {
 			break;
 		}
+
+		INFO_LOG(G3D, "Compilation thread has %d pipelines to create", (int)toCompile.size());
+
+		// TODO: Here we can sort the pending pipelines by vertex and fragment shaders,
+		// and split up further.
+		// Those with the same pairs of shaders should be on the same thread.
 		for (auto &entry : toCompile) {
 			switch (entry.type) {
 			case CompileQueueEntry::Type::GRAPHICS:

Common/GPU/Vulkan/VulkanRenderManager.h

@@ -12,6 +12,7 @@
 #include <thread>
 #include <queue>
 
+#include "Common/Thread/Promise.h"
 #include "Common/System/Display.h"
 #include "Common/GPU/Vulkan/VulkanContext.h"
 #include "Common/Data/Convert/SmallDataConvert.h"
@@ -121,7 +122,11 @@ struct VKRGraphicsPipelineDesc {
 	VkPipelineDynamicStateCreateInfo ds{ VK_STRUCTURE_TYPE_PIPELINE_DYNAMIC_STATE_CREATE_INFO };
 	VkPipelineRasterizationStateCreateInfo rs{ VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO };
 	VkPipelineMultisampleStateCreateInfo ms{ VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO };
-	VkPipelineShaderStageCreateInfo shaderStageInfo[2]{};
+
+	// Replaced the ShaderStageInfo with promises here so we can wait for compiles to finish.
+	Promise<VkShaderModule> *vertexShader;
+	Promise<VkShaderModule> *fragmentShader;
+
 	VkPipelineInputAssemblyStateCreateInfo inputAssembly{ VK_STRUCTURE_TYPE_PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO };
 	VkVertexInputAttributeDescription attrs[8]{};
 	VkVertexInputBindingDescription ibd{};
@@ -139,10 +144,12 @@ struct VKRComputePipelineDesc {
 // Wrapped pipeline, which will later allow for background compilation while emulating the rest of the frame.
 struct VKRGraphicsPipeline {
 	VKRGraphicsPipeline() {
-		pipeline = VK_NULL_HANDLE;
+		pipeline = Promise<VkPipeline>::CreateEmpty();
 	}
+
 	VKRGraphicsPipelineDesc *desc = nullptr;  // While non-zero, is pending and pipeline isn't valid.
-	std::atomic<VkPipeline> pipeline;
+
+	Promise<VkPipeline> *pipeline;
 
 	bool Create(VulkanContext *vulkan);
 	bool Pending() const {
@@ -155,7 +162,7 @@ struct VKRComputePipeline {
 		pipeline = VK_NULL_HANDLE;
 	}
 	VKRComputePipelineDesc *desc = nullptr;
-	std::atomic<VkPipeline> pipeline;
+	Promise<VkPipeline> *pipeline;
 
 	bool Create(VulkanContext *vulkan);
 	bool Pending() const {
@@ -256,7 +263,7 @@ class VulkanRenderManager {
 		_dbg_assert_(curRenderStep_ && curRenderStep_->stepType == VKRStepType::RENDER);
 		_dbg_assert_(pipeline != nullptr);
 		VkRenderData data{ VKRRenderCommand::BIND_GRAPHICS_PIPELINE };
-		data.graphics_pipeline.pipeline = pipeline;
+		data.graphics_pipeline.pipeline = pipeline->pipeline;
 		data.graphics_pipeline.pipelineLayout = pipelineLayout;
 		curPipelineFlags_ |= flags;
 		curRenderStep_->commands.push_back(data);
@@ -266,7 +273,7 @@ class VulkanRenderManager {
 		_dbg_assert_(curRenderStep_ && curRenderStep_->stepType == VKRStepType::RENDER);
 		_dbg_assert_(pipeline != nullptr);
 		VkRenderData data{ VKRRenderCommand::BIND_COMPUTE_PIPELINE };
-		data.compute_pipeline.pipeline = pipeline;
+		data.compute_pipeline.pipeline = pipeline->pipeline;
 		data.compute_pipeline.pipelineLayout = pipelineLayout;
 		curPipelineFlags_ |= flags;
 		curRenderStep_->commands.push_back(data);

Common/Thread/Channel.h

@@ -19,7 +19,7 @@ struct Mailbox {
 
 	std::mutex mutex_;
 	std::condition_variable condvar_;
-	T data_ = nullptr;
+	T data_{};
 
 	T Wait() {
 		std::unique_lock<std::mutex> lock(mutex_);
@@ -43,7 +43,7 @@ struct Mailbox {
 		std::unique_lock<std::mutex> lock(mutex_);
 		if (!data_) {
 			data_ = data;
-			condvar_.notify_one();
+			condvar_.notify_all();
 			return true;
 		} else {
 			// Already has value.

Common/Thread/Promise.h

@@ -1,7 +1,9 @@
 #pragma once
 
 #include <functional>
+#include <mutex>
 
+#include "Common/Log.h"
 #include "Common/Thread/Channel.h"
 #include "Common/Thread/ThreadManager.h"
 
@@ -33,6 +35,7 @@ class PromiseTask : public Task {
 // Has ownership over the data. Single use.
 // TODO: Split Mailbox (rx_ and tx_) up into separate proxy objects.
 // NOTE: Poll/BlockUntilReady should only be used from one thread.
+// TODO: Make movable?
 template<class T>
 class Promise {
 public:
@@ -47,15 +50,36 @@ class Promise {
 		return promise;
 	}
 
+	static Promise<T> *AlreadyDone(T data) {
+		Promise<T> *promise = new Promise<T>();
+		promise->data_ = data;
+		promise->ready_ = true;
+		return promise;
+	}
+
+	static Promise<T> *CreateEmpty() {
+		Mailbox<T> *mailbox = new Mailbox<T>();
+		Promise<T> *promise = new Promise<T>();
+		promise->rx_ = mailbox;
+		return promise;
+	}
+
+	// Allow an empty promise to spawn, too, in case we want to delay it.
+	void SpawnEmpty(ThreadManager *threadman, std::function<T()> fun, TaskType taskType) {
+		PromiseTask<T> *task = new PromiseTask<T>(fun, rx_, taskType);
+		threadman->EnqueueTask(task);
+	}
+
 	~Promise() {
+		std::lock_guard<std::mutex> guard(readyMutex_);
 		// A promise should have been fulfilled before it's destroyed.
 		_assert_(ready_);
 		_assert_(!rx_);
-		delete data_;
 	}
 
 	// Returns T if the data is ready, nullptr if it's not.
 	T Poll() {
+		std::lock_guard<std::mutex> guard(readyMutex_);
 		if (ready_) {
 			return data_;
 		} else {
@@ -71,6 +95,7 @@ class Promise {
 	}
 
 	T BlockUntilReady() {
+		std::lock_guard<std::mutex> guard(readyMutex_);
 		if (ready_) {
 			return data_;
 		} else {
@@ -82,10 +107,17 @@ class Promise {
 		}
 	}
 
+	// For outside injection of data, when not using Spawn
+	void Post(T data) {
+		rx_->Send(data);
+	}
+
 private:
 	Promise() {}
 
-	T data_ = nullptr;
+	// Promise can only be constructed in Spawn (or AlreadyDone).
+	T data_{};
 	bool ready_ = false;
-	Mailbox<T> *rx_;
+	std::mutex readyMutex_;
+	Mailbox<T> *rx_ = nullptr;
 };

GPU/Vulkan/PipelineManagerVulkan.cpp

@@ -36,7 +36,7 @@ void PipelineManagerVulkan::Clear() {
 
 	pipelines_.Iterate([&](const VulkanPipelineKey &key, VulkanPipeline *value) {
 		if (value->pipeline) {
-			VkPipeline pipeline = value->pipeline->pipeline;
+			VkPipeline pipeline = value->pipeline->pipeline->BlockUntilReady();
 			vulkan_->Delete().QueueDeletePipeline(pipeline);
 			vulkan_->Delete().QueueCallback([](void *p) {
 				VKRGraphicsPipeline *pipeline = (VKRGraphicsPipeline *)p;
@@ -256,28 +256,8 @@ static VulkanPipeline *CreateVulkanPipeline(VulkanRenderManager *renderManager,
 	ms.pSampleMask = nullptr;
 	ms.rasterizationSamples = VK_SAMPLE_COUNT_1_BIT;
 
-	VkPipelineShaderStageCreateInfo *ss = &desc->shaderStageInfo[0];
-	ss[0].sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO;
-	ss[0].stage = VK_SHADER_STAGE_VERTEX_BIT;
-	ss[0].pSpecializationInfo = nullptr;
-	ss[0].module = vs->GetModule();
-	ss[0].pName = "main";
-	ss[0].flags = 0;
-	ss[1].sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO;
-	ss[1].stage = VK_SHADER_STAGE_FRAGMENT_BIT;
-	ss[1].pSpecializationInfo = nullptr;
-	ss[1].module = fs->GetModule();
-	ss[1].pName = "main";
-	ss[1].flags = 0;
-
-	if (!ss[0].module || !ss[1].module) {
-		ERROR_LOG(G3D, "Failed creating graphics pipeline - bad shaders");
-		// Create a placeholder to avoid creating over and over if shader compiler broken.
-		VulkanPipeline *nullPipeline = new VulkanPipeline();
-		nullPipeline->pipeline = VK_NULL_HANDLE;
-		nullPipeline->flags = 0;
-		return nullPipeline;
-	}
+	desc->fragmentShader = fs->GetModule();
+	desc->vertexShader = vs->GetModule();
 
 	VkPipelineInputAssemblyStateCreateInfo &inputAssembly = desc->inputAssembly;
 	inputAssembly.flags = 0;
@@ -321,7 +301,6 @@ static VulkanPipeline *CreateVulkanPipeline(VulkanRenderManager *renderManager,
 	VkGraphicsPipelineCreateInfo &pipe = desc->pipe;
 	pipe.flags = 0;
 	pipe.stageCount = 2;
-	pipe.pStages = ss;
 	pipe.basePipelineIndex = 0;
 
 	pipe.pColorBlendState = &desc->cbs;
@@ -633,8 +612,8 @@ void PipelineManagerVulkan::SaveCache(FILE *file, bool saveRawPipelineCache, Sha
 	pipelines_.Iterate([&](const VulkanPipelineKey &pkey, VulkanPipeline *value) {
 		if (failed)
 			return;
-		VulkanVertexShader *vshader = shaderManager->GetVertexShaderFromModule(pkey.vShader);
-		VulkanFragmentShader *fshader = shaderManager->GetFragmentShaderFromModule(pkey.fShader);
+		VulkanVertexShader *vshader = shaderManager->GetVertexShaderFromModule(pkey.vShader->BlockUntilReady());
+		VulkanFragmentShader *fshader = shaderManager->GetFragmentShaderFromModule(pkey.fShader->BlockUntilReady());
 		if (!vshader || !fshader) {
 			failed = true;
 			return;