From 6ac97405df021d5d2bd9a529253bd5c5a418c1a9 Mon Sep 17 00:00:00 2001 From: ameerj Date: Tue, 28 Jul 2020 00:08:02 -0400 Subject: Vk Async pipeline compilation --- src/video_core/renderer_vulkan/vk_device.cpp | 2 ++ src/video_core/renderer_vulkan/vk_device.h | 5 ++++ .../renderer_vulkan/vk_fence_manager.cpp | 2 +- .../renderer_vulkan/vk_graphics_pipeline.cpp | 5 ++-- .../renderer_vulkan/vk_graphics_pipeline.h | 6 +++++ .../renderer_vulkan/vk_pipeline_cache.cpp | 24 +++++++++++++++---- src/video_core/renderer_vulkan/vk_pipeline_cache.h | 27 ++++++++++++++++++++-- src/video_core/renderer_vulkan/vk_rasterizer.cpp | 27 ++++++++++++++++++++-- src/video_core/renderer_vulkan/vk_rasterizer.h | 10 ++++++++ src/video_core/renderer_vulkan/wrapper.cpp | 2 +- src/video_core/renderer_vulkan/wrapper.h | 2 +- 11 files changed, 98 insertions(+), 14 deletions(-) (limited to 'src/video_core/renderer_vulkan') diff --git a/src/video_core/renderer_vulkan/vk_device.cpp b/src/video_core/renderer_vulkan/vk_device.cpp index 0c03e4d83..ebcfaa0e3 100644 --- a/src/video_core/renderer_vulkan/vk_device.cpp +++ b/src/video_core/renderer_vulkan/vk_device.cpp @@ -382,6 +382,8 @@ bool VKDevice::Create() { graphics_queue = logical.GetQueue(graphics_family); present_queue = logical.GetQueue(present_family); + + use_asynchronous_shaders = Settings::values.use_asynchronous_shaders.GetValue(); return true; } diff --git a/src/video_core/renderer_vulkan/vk_device.h b/src/video_core/renderer_vulkan/vk_device.h index 529744f2d..30cd3e189 100644 --- a/src/video_core/renderer_vulkan/vk_device.h +++ b/src/video_core/renderer_vulkan/vk_device.h @@ -202,6 +202,10 @@ public: return reported_extensions; } + bool UseAsynchronousShaders() const { + return use_asynchronous_shaders; + } + /// Checks if the physical device is suitable. static bool IsSuitable(vk::PhysicalDevice physical, VkSurfaceKHR surface); @@ -251,6 +255,7 @@ private: bool ext_custom_border_color{}; ///< Support for VK_EXT_custom_border_color. bool ext_extended_dynamic_state{}; ///< Support for VK_EXT_extended_dynamic_state. bool nv_device_diagnostics_config{}; ///< Support for VK_NV_device_diagnostics_config. + bool use_asynchronous_shaders{}; // Telemetry parameters std::string vendor_name; ///< Device's driver name. diff --git a/src/video_core/renderer_vulkan/vk_fence_manager.cpp b/src/video_core/renderer_vulkan/vk_fence_manager.cpp index a02be5487..d7f65d435 100644 --- a/src/video_core/renderer_vulkan/vk_fence_manager.cpp +++ b/src/video_core/renderer_vulkan/vk_fence_manager.cpp @@ -29,7 +29,7 @@ void InnerFence::Queue() { } ASSERT(!event); - event = device.GetLogical().CreateEvent(); + event = device.GetLogical().CreateNewEvent(); ticks = scheduler.Ticks(); scheduler.RequestOutsideRenderPassOperationContext(); diff --git a/src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp b/src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp index aaf930b90..7d51b9836 100644 --- a/src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp +++ b/src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp @@ -84,9 +84,8 @@ VKGraphicsPipeline::VKGraphicsPipeline(const VKDevice& device, VKScheduler& sche update_descriptor_queue{update_descriptor_queue}, layout{CreatePipelineLayout()}, descriptor_template{CreateDescriptorUpdateTemplate(program)}, modules{CreateShaderModules( program)}, - renderpass{renderpass_cache.GetRenderPass(key.renderpass_params)}, pipeline{CreatePipeline( - key.renderpass_params, - program)} {} + renderpass{renderpass_cache.GetRenderPass(key.renderpass_params)}, + pipeline{CreatePipeline(key.renderpass_params, program)}, m_key{key} {} VKGraphicsPipeline::~VKGraphicsPipeline() = default; diff --git a/src/video_core/renderer_vulkan/vk_graphics_pipeline.h b/src/video_core/renderer_vulkan/vk_graphics_pipeline.h index a1d699a6c..39c73a139 100644 --- a/src/video_core/renderer_vulkan/vk_graphics_pipeline.h +++ b/src/video_core/renderer_vulkan/vk_graphics_pipeline.h @@ -54,6 +54,10 @@ public: return renderpass; } + const GraphicsPipelineCacheKey& GetCacheKey() { + return m_key; + } + private: vk::DescriptorSetLayout CreateDescriptorSetLayout( vk::Span bindings) const; @@ -82,6 +86,8 @@ private: VkRenderPass renderpass; vk::Pipeline pipeline; + + const GraphicsPipelineCacheKey& m_key; }; } // namespace Vulkan diff --git a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp index 418c62bc4..45d4dcb8c 100644 --- a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp +++ b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp @@ -205,7 +205,8 @@ std::array VKPipelineCache::GetShaders() { return last_shaders = shaders; } -VKGraphicsPipeline& VKPipelineCache::GetGraphicsPipeline(const GraphicsPipelineCacheKey& key) { +VKGraphicsPipeline& VKPipelineCache::GetGraphicsPipeline( + const GraphicsPipelineCacheKey& key, VideoCommon::Shader::AsyncShaders& async_shaders) { MICROPROFILE_SCOPE(Vulkan_PipelineCache); if (last_graphics_pipeline && last_graphics_key == key) { @@ -213,11 +214,27 @@ VKGraphicsPipeline& VKPipelineCache::GetGraphicsPipeline(const GraphicsPipelineC } last_graphics_key = key; + if (device.UseAsynchronousShaders()) { + auto work = async_shaders.GetCompletedWork(); + for (std::size_t i = 0; i < work.size(); ++i) { + auto& entry = graphics_cache.at(work[i].pipeline->GetCacheKey()); + entry = std::move(work[i].pipeline); + } + const auto [pair, is_cache_miss] = graphics_cache.try_emplace(key); + if (is_cache_miss) { + LOG_INFO(Render_Vulkan, "Compile 0x{:016X}", key.Hash()); + const auto [program, bindings] = DecompileShaders(key.fixed_state); + async_shaders.QueueVulkanShader(this, bindings, program, key.renderpass_params, + key.padding, key.shaders, key.fixed_state); + } + return *(last_graphics_pipeline = graphics_cache.at(key).get()); + } + const auto [pair, is_cache_miss] = graphics_cache.try_emplace(key); auto& entry = pair->second; if (is_cache_miss) { LOG_INFO(Render_Vulkan, "Compile 0x{:016X}", key.Hash()); - const auto [program, bindings] = DecompileShaders(key); + const auto [program, bindings] = DecompileShaders(key.fixed_state); entry = std::make_unique(device, scheduler, descriptor_pool, update_descriptor_queue, renderpass_cache, key, bindings, program); @@ -312,8 +329,7 @@ void VKPipelineCache::OnShaderRemoval(Shader* shader) { } std::pair> -VKPipelineCache::DecompileShaders(const GraphicsPipelineCacheKey& key) { - const auto& fixed_state = key.fixed_state; +VKPipelineCache::DecompileShaders(const FixedPipelineState& fixed_state) { auto& memory_manager = system.GPU().MemoryManager(); const auto& gpu = system.GPU().Maxwell3D(); diff --git a/src/video_core/renderer_vulkan/vk_pipeline_cache.h b/src/video_core/renderer_vulkan/vk_pipeline_cache.h index 0a3fe65fb..c70da6da4 100644 --- a/src/video_core/renderer_vulkan/vk_pipeline_cache.h +++ b/src/video_core/renderer_vulkan/vk_pipeline_cache.h @@ -22,6 +22,7 @@ #include "video_core/renderer_vulkan/vk_renderpass_cache.h" #include "video_core/renderer_vulkan/vk_shader_decompiler.h" #include "video_core/renderer_vulkan/wrapper.h" +#include "video_core/shader/async_shaders.h" #include "video_core/shader/memory_util.h" #include "video_core/shader/registry.h" #include "video_core/shader/shader_ir.h" @@ -152,16 +153,37 @@ public: std::array GetShaders(); - VKGraphicsPipeline& GetGraphicsPipeline(const GraphicsPipelineCacheKey& key); + VKGraphicsPipeline& GetGraphicsPipeline(const GraphicsPipelineCacheKey& key, + VideoCommon::Shader::AsyncShaders& async_shaders); VKComputePipeline& GetComputePipeline(const ComputePipelineCacheKey& key); + const VKDevice& GetDevice() { + return device; + } + + VKScheduler& GetScheduler() { + return scheduler; + } + + VKDescriptorPool& GetDescriptorPool() { + return descriptor_pool; + } + + VKUpdateDescriptorQueue& GetUpdateDescriptorQueue() { + return update_descriptor_queue; + } + + VKRenderPassCache& GetRenderpassCache() { + return renderpass_cache; + } + protected: void OnShaderRemoval(Shader* shader) final; private: std::pair> DecompileShaders( - const GraphicsPipelineCacheKey& key); + const FixedPipelineState& fixed_state); Core::System& system; const VKDevice& device; @@ -177,6 +199,7 @@ private: GraphicsPipelineCacheKey last_graphics_key; VKGraphicsPipeline* last_graphics_pipeline = nullptr; + std::vector> duplicates; std::unordered_map> graphics_cache; diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp index 7500e8244..6310e898c 100644 --- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp +++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp @@ -400,8 +400,25 @@ RasterizerVulkan::RasterizerVulkan(Core::System& system, Core::Frontend::EmuWind buffer_cache(*this, system, device, memory_manager, scheduler, staging_pool), sampler_cache(device), fence_manager(system, *this, device, scheduler, texture_cache, buffer_cache, query_cache), - query_cache(system, *this, device, scheduler), wfi_event{device.GetLogical().CreateEvent()} { + query_cache(system, *this, device, scheduler), + wfi_event{device.GetLogical().CreateNewEvent()}, async_shaders{renderer} { scheduler.SetQueryCache(query_cache); + if (device.UseAsynchronousShaders()) { + // Max worker threads we should allow + constexpr auto MAX_THREADS = 2u; + // Amount of threads we should reserve for other parts of yuzu + constexpr auto RESERVED_THREADS = 6u; + // Get the amount of threads we can use(this can return zero) + const auto cpu_thread_count = + std::max(RESERVED_THREADS, std::thread::hardware_concurrency()); + // Deduce how many "extra" threads we have to use. + const auto max_threads_unused = cpu_thread_count - RESERVED_THREADS; + // Always allow at least 1 thread regardless of our settings + const auto max_worker_count = std::max(1u, max_threads_unused); + // Don't use more than MAX_THREADS + const auto worker_count = std::min(max_worker_count, MAX_THREADS); + async_shaders.AllocateWorkers(worker_count); + } } RasterizerVulkan::~RasterizerVulkan() = default; @@ -439,7 +456,13 @@ void RasterizerVulkan::Draw(bool is_indexed, bool is_instanced) { key.renderpass_params = GetRenderPassParams(texceptions); key.padding = 0; - auto& pipeline = pipeline_cache.GetGraphicsPipeline(key); + auto& pipeline = pipeline_cache.GetGraphicsPipeline(key, async_shaders); + if (&pipeline == nullptr || pipeline.GetHandle() == VK_NULL_HANDLE) { + // Async graphics pipeline was not ready. + system.GPU().TickWork(); + return; + } + scheduler.BindGraphicsPipeline(pipeline.GetHandle()); const auto renderpass = pipeline.GetRenderPass(); diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.h b/src/video_core/renderer_vulkan/vk_rasterizer.h index 923178b0b..27604b9a3 100644 --- a/src/video_core/renderer_vulkan/vk_rasterizer.h +++ b/src/video_core/renderer_vulkan/vk_rasterizer.h @@ -32,6 +32,7 @@ #include "video_core/renderer_vulkan/vk_texture_cache.h" #include "video_core/renderer_vulkan/vk_update_descriptor.h" #include "video_core/renderer_vulkan/wrapper.h" +#include "video_core/shader/async_shaders.h" namespace Core { class System; @@ -136,6 +137,14 @@ public: u32 pixel_stride) override; void SetupDirtyFlags() override; + VideoCommon::Shader::AsyncShaders& GetAsyncShaders() { + return async_shaders; + } + + const VideoCommon::Shader::AsyncShaders& GetAsyncShaders() const { + return async_shaders; + } + /// Maximum supported size that a constbuffer can have in bytes. static constexpr std::size_t MaxConstbufferSize = 0x10000; static_assert(MaxConstbufferSize % (4 * sizeof(float)) == 0, @@ -278,6 +287,7 @@ private: VKMemoryManager& memory_manager; StateTracker& state_tracker; VKScheduler& scheduler; + VideoCommon::Shader::AsyncShaders async_shaders; VKStagingBufferPool staging_pool; VKDescriptorPool descriptor_pool; diff --git a/src/video_core/renderer_vulkan/wrapper.cpp b/src/video_core/renderer_vulkan/wrapper.cpp index 14cac38ea..c43d60adf 100644 --- a/src/video_core/renderer_vulkan/wrapper.cpp +++ b/src/video_core/renderer_vulkan/wrapper.cpp @@ -644,7 +644,7 @@ ShaderModule Device::CreateShaderModule(const VkShaderModuleCreateInfo& ci) cons return ShaderModule(object, handle, *dld); } -Event Device::CreateEvent() const { +Event Device::CreateNewEvent() const { static constexpr VkEventCreateInfo ci{ .sType = VK_STRUCTURE_TYPE_EVENT_CREATE_INFO, .pNext = nullptr, diff --git a/src/video_core/renderer_vulkan/wrapper.h b/src/video_core/renderer_vulkan/wrapper.h index 31885ef42..b9d3fedc1 100644 --- a/src/video_core/renderer_vulkan/wrapper.h +++ b/src/video_core/renderer_vulkan/wrapper.h @@ -721,7 +721,7 @@ public: ShaderModule CreateShaderModule(const VkShaderModuleCreateInfo& ci) const; - Event CreateEvent() const; + Event CreateNewEvent() const; SwapchainKHR CreateSwapchainKHR(const VkSwapchainCreateInfoKHR& ci) const; -- cgit v1.2.3 From 4539073ce1d8fd6df03263e826d3805b4909e055 Mon Sep 17 00:00:00 2001 From: ameerj Date: Thu, 30 Jul 2020 15:41:11 -0400 Subject: Address feedback. Bruteforce delete duplicates --- src/video_core/renderer_vulkan/vk_graphics_pipeline.h | 2 +- src/video_core/renderer_vulkan/vk_pipeline_cache.cpp | 16 +++++++++------- src/video_core/renderer_vulkan/vk_pipeline_cache.h | 19 +++++++++++++++++-- src/video_core/renderer_vulkan/vk_rasterizer.cpp | 18 ++++++++++-------- src/video_core/renderer_vulkan/vk_rasterizer.h | 2 +- 5 files changed, 38 insertions(+), 19 deletions(-) (limited to 'src/video_core/renderer_vulkan') diff --git a/src/video_core/renderer_vulkan/vk_graphics_pipeline.h b/src/video_core/renderer_vulkan/vk_graphics_pipeline.h index 39c73a139..d50bd347c 100644 --- a/src/video_core/renderer_vulkan/vk_graphics_pipeline.h +++ b/src/video_core/renderer_vulkan/vk_graphics_pipeline.h @@ -54,7 +54,7 @@ public: return renderpass; } - const GraphicsPipelineCacheKey& GetCacheKey() { + const GraphicsPipelineCacheKey& GetCacheKey() const { return m_key; } diff --git a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp index 45d4dcb8c..a7ce621ca 100644 --- a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp +++ b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp @@ -205,20 +205,20 @@ std::array VKPipelineCache::GetShaders() { return last_shaders = shaders; } -VKGraphicsPipeline& VKPipelineCache::GetGraphicsPipeline( +VKGraphicsPipeline* VKPipelineCache::GetGraphicsPipeline( const GraphicsPipelineCacheKey& key, VideoCommon::Shader::AsyncShaders& async_shaders) { MICROPROFILE_SCOPE(Vulkan_PipelineCache); if (last_graphics_pipeline && last_graphics_key == key) { - return *last_graphics_pipeline; + return last_graphics_pipeline; } last_graphics_key = key; if (device.UseAsynchronousShaders()) { auto work = async_shaders.GetCompletedWork(); - for (std::size_t i = 0; i < work.size(); ++i) { - auto& entry = graphics_cache.at(work[i].pipeline->GetCacheKey()); - entry = std::move(work[i].pipeline); + for (auto& w : work) { + auto& entry = graphics_cache.at(w.pipeline->GetCacheKey()); + entry = std::move(w.pipeline); } const auto [pair, is_cache_miss] = graphics_cache.try_emplace(key); if (is_cache_miss) { @@ -227,7 +227,8 @@ VKGraphicsPipeline& VKPipelineCache::GetGraphicsPipeline( async_shaders.QueueVulkanShader(this, bindings, program, key.renderpass_params, key.padding, key.shaders, key.fixed_state); } - return *(last_graphics_pipeline = graphics_cache.at(key).get()); + last_graphics_pipeline = graphics_cache.at(key).get(); + return last_graphics_pipeline; } const auto [pair, is_cache_miss] = graphics_cache.try_emplace(key); @@ -239,7 +240,8 @@ VKGraphicsPipeline& VKPipelineCache::GetGraphicsPipeline( update_descriptor_queue, renderpass_cache, key, bindings, program); } - return *(last_graphics_pipeline = entry.get()); + last_graphics_pipeline = entry.get(); + return last_graphics_pipeline; } VKComputePipeline& VKPipelineCache::GetComputePipeline(const ComputePipelineCacheKey& key) { diff --git a/src/video_core/renderer_vulkan/vk_pipeline_cache.h b/src/video_core/renderer_vulkan/vk_pipeline_cache.h index c70da6da4..404f2b3f7 100644 --- a/src/video_core/renderer_vulkan/vk_pipeline_cache.h +++ b/src/video_core/renderer_vulkan/vk_pipeline_cache.h @@ -153,31 +153,46 @@ public: std::array GetShaders(); - VKGraphicsPipeline& GetGraphicsPipeline(const GraphicsPipelineCacheKey& key, + VKGraphicsPipeline* GetGraphicsPipeline(const GraphicsPipelineCacheKey& key, VideoCommon::Shader::AsyncShaders& async_shaders); VKComputePipeline& GetComputePipeline(const ComputePipelineCacheKey& key); - const VKDevice& GetDevice() { + const VKDevice& GetDevice() const { return device; } VKScheduler& GetScheduler() { return scheduler; } + const VKScheduler& GetScheduler() const { + return scheduler; + } VKDescriptorPool& GetDescriptorPool() { return descriptor_pool; } + const VKDescriptorPool& GetDescriptorPool() const { + return descriptor_pool; + } + VKUpdateDescriptorQueue& GetUpdateDescriptorQueue() { return update_descriptor_queue; } + const VKUpdateDescriptorQueue& GetUpdateDescriptorQueue() const { + return update_descriptor_queue; + } + VKRenderPassCache& GetRenderpassCache() { return renderpass_cache; } + const VKRenderPassCache& GetRenderpassCache() const { + return renderpass_cache; + } + protected: void OnShaderRemoval(Shader* shader) final; diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp index 6310e898c..fc1b51a96 100644 --- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp +++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp @@ -404,10 +404,12 @@ RasterizerVulkan::RasterizerVulkan(Core::System& system, Core::Frontend::EmuWind wfi_event{device.GetLogical().CreateNewEvent()}, async_shaders{renderer} { scheduler.SetQueryCache(query_cache); if (device.UseAsynchronousShaders()) { + // The following is subject to move into the allocate workers method, to be api agnostic + // Max worker threads we should allow - constexpr auto MAX_THREADS = 2u; + constexpr u32 MAX_THREADS = 4; // Amount of threads we should reserve for other parts of yuzu - constexpr auto RESERVED_THREADS = 6u; + constexpr u32 RESERVED_THREADS = 6; // Get the amount of threads we can use(this can return zero) const auto cpu_thread_count = std::max(RESERVED_THREADS, std::thread::hardware_concurrency()); @@ -456,16 +458,16 @@ void RasterizerVulkan::Draw(bool is_indexed, bool is_instanced) { key.renderpass_params = GetRenderPassParams(texceptions); key.padding = 0; - auto& pipeline = pipeline_cache.GetGraphicsPipeline(key, async_shaders); - if (&pipeline == nullptr || pipeline.GetHandle() == VK_NULL_HANDLE) { + auto pipeline = pipeline_cache.GetGraphicsPipeline(key, async_shaders); + if (pipeline == nullptr || pipeline->GetHandle() == VK_NULL_HANDLE) { // Async graphics pipeline was not ready. system.GPU().TickWork(); return; } - scheduler.BindGraphicsPipeline(pipeline.GetHandle()); + scheduler.BindGraphicsPipeline(pipeline->GetHandle()); - const auto renderpass = pipeline.GetRenderPass(); + const auto renderpass = pipeline->GetRenderPass(); const auto [framebuffer, render_area] = ConfigureFramebuffers(renderpass); scheduler.RequestRenderpass(renderpass, framebuffer, render_area); @@ -475,8 +477,8 @@ void RasterizerVulkan::Draw(bool is_indexed, bool is_instanced) { BeginTransformFeedback(); - const auto pipeline_layout = pipeline.GetLayout(); - const auto descriptor_set = pipeline.CommitDescriptorSet(); + const auto pipeline_layout = pipeline->GetLayout(); + const auto descriptor_set = pipeline->CommitDescriptorSet(); scheduler.Record([pipeline_layout, descriptor_set, draw_params](vk::CommandBuffer cmdbuf) { if (descriptor_set) { cmdbuf.BindDescriptorSets(VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline_layout, diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.h b/src/video_core/renderer_vulkan/vk_rasterizer.h index 27604b9a3..f640ba649 100644 --- a/src/video_core/renderer_vulkan/vk_rasterizer.h +++ b/src/video_core/renderer_vulkan/vk_rasterizer.h @@ -287,7 +287,6 @@ private: VKMemoryManager& memory_manager; StateTracker& state_tracker; VKScheduler& scheduler; - VideoCommon::Shader::AsyncShaders async_shaders; VKStagingBufferPool staging_pool; VKDescriptorPool descriptor_pool; @@ -307,6 +306,7 @@ private: vk::Buffer default_buffer; VKMemoryCommit default_buffer_commit; vk::Event wfi_event; + VideoCommon::Shader::AsyncShaders async_shaders; std::array color_attachments; View zeta_attachment; -- cgit v1.2.3 From c02464f64e302b8c9ea0f310e6fd85834d26cca5 Mon Sep 17 00:00:00 2001 From: ameerj Date: Fri, 31 Jul 2020 17:30:05 -0400 Subject: Vk Async Worker directly emplace in cache --- src/video_core/renderer_vulkan/vk_pipeline_cache.cpp | 18 +++++++++++++----- src/video_core/renderer_vulkan/vk_pipeline_cache.h | 3 +++ 2 files changed, 16 insertions(+), 5 deletions(-) (limited to 'src/video_core/renderer_vulkan') diff --git a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp index a7ce621ca..1a8b2c62b 100644 --- a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp +++ b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp @@ -215,11 +215,7 @@ VKGraphicsPipeline* VKPipelineCache::GetGraphicsPipeline( last_graphics_key = key; if (device.UseAsynchronousShaders()) { - auto work = async_shaders.GetCompletedWork(); - for (auto& w : work) { - auto& entry = graphics_cache.at(w.pipeline->GetCacheKey()); - entry = std::move(w.pipeline); - } + std::unique_lock lock{pipeline_cache}; const auto [pair, is_cache_miss] = graphics_cache.try_emplace(key); if (is_cache_miss) { LOG_INFO(Render_Vulkan, "Compile 0x{:016X}", key.Hash()); @@ -296,6 +292,18 @@ VKComputePipeline& VKPipelineCache::GetComputePipeline(const ComputePipelineCach return *entry; } +void VKPipelineCache::EmplacePipeline(std::unique_ptr pipeline) { + std::unique_lock lock{pipeline_cache}; + const auto [pair, is_cache_miss] = graphics_cache.try_emplace(pipeline->GetCacheKey()); + auto& entry = pair->second; + if (entry) { + LOG_INFO(Render_Vulkan, "Pipeline already here 0x{:016X}", pipeline->GetCacheKey().Hash()); + duplicates.push_back(std::move(pipeline)); + } else { + entry = std::move(pipeline); + } +} + void VKPipelineCache::OnShaderRemoval(Shader* shader) { bool finished = false; const auto Finish = [&] { diff --git a/src/video_core/renderer_vulkan/vk_pipeline_cache.h b/src/video_core/renderer_vulkan/vk_pipeline_cache.h index 404f2b3f7..777ef2038 100644 --- a/src/video_core/renderer_vulkan/vk_pipeline_cache.h +++ b/src/video_core/renderer_vulkan/vk_pipeline_cache.h @@ -193,6 +193,8 @@ public: return renderpass_cache; } + void EmplacePipeline(std::unique_ptr pipeline); + protected: void OnShaderRemoval(Shader* shader) final; @@ -216,6 +218,7 @@ private: VKGraphicsPipeline* last_graphics_pipeline = nullptr; std::vector> duplicates; + std::mutex pipeline_cache; std::unordered_map> graphics_cache; std::unordered_map> compute_cache; -- cgit v1.2.3 From 31a76410e8fa09462d960c10148c075125dc385a Mon Sep 17 00:00:00 2001 From: ameerj Date: Sun, 2 Aug 2020 13:05:41 -0400 Subject: Address feedback, add shader compile notifier, update setting text --- src/video_core/renderer_vulkan/vk_device.h | 5 +- .../renderer_vulkan/vk_graphics_pipeline.cpp | 9 ++-- .../renderer_vulkan/vk_graphics_pipeline.h | 30 ++++++++--- .../renderer_vulkan/vk_pipeline_cache.cpp | 23 ++++----- src/video_core/renderer_vulkan/vk_pipeline_cache.h | 58 ---------------------- src/video_core/renderer_vulkan/vk_rasterizer.cpp | 19 +++---- 6 files changed, 51 insertions(+), 93 deletions(-) (limited to 'src/video_core/renderer_vulkan') diff --git a/src/video_core/renderer_vulkan/vk_device.h b/src/video_core/renderer_vulkan/vk_device.h index 30cd3e189..26a233db1 100644 --- a/src/video_core/renderer_vulkan/vk_device.h +++ b/src/video_core/renderer_vulkan/vk_device.h @@ -202,6 +202,7 @@ public: return reported_extensions; } + /// Returns true if the setting for async shader compilation is enabled. bool UseAsynchronousShaders() const { return use_asynchronous_shaders; } @@ -255,7 +256,9 @@ private: bool ext_custom_border_color{}; ///< Support for VK_EXT_custom_border_color. bool ext_extended_dynamic_state{}; ///< Support for VK_EXT_extended_dynamic_state. bool nv_device_diagnostics_config{}; ///< Support for VK_NV_device_diagnostics_config. - bool use_asynchronous_shaders{}; + + // Asynchronous Graphics Pipeline setting + bool use_asynchronous_shaders{}; ///< Setting to use asynchronous shaders/graphics pipeline // Telemetry parameters std::string vendor_name; ///< Device's driver name. diff --git a/src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp b/src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp index 7d51b9836..5dc4cd5af 100644 --- a/src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp +++ b/src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp @@ -78,14 +78,15 @@ VKGraphicsPipeline::VKGraphicsPipeline(const VKDevice& device, VKScheduler& sche const GraphicsPipelineCacheKey& key, vk::Span bindings, const SPIRVProgram& program) - : device{device}, scheduler{scheduler}, fixed_state{key.fixed_state}, hash{key.Hash()}, + : device{device}, scheduler{scheduler}, hash{key.Hash()}, cache_key{key}, descriptor_set_layout{CreateDescriptorSetLayout(bindings)}, descriptor_allocator{descriptor_pool, *descriptor_set_layout}, update_descriptor_queue{update_descriptor_queue}, layout{CreatePipelineLayout()}, descriptor_template{CreateDescriptorUpdateTemplate(program)}, modules{CreateShaderModules( program)}, - renderpass{renderpass_cache.GetRenderPass(key.renderpass_params)}, - pipeline{CreatePipeline(key.renderpass_params, program)}, m_key{key} {} + renderpass{renderpass_cache.GetRenderPass(key.renderpass_params)}, pipeline{CreatePipeline( + key.renderpass_params, + program)} {} VKGraphicsPipeline::~VKGraphicsPipeline() = default; @@ -180,7 +181,7 @@ std::vector VKGraphicsPipeline::CreateShaderModules( vk::Pipeline VKGraphicsPipeline::CreatePipeline(const RenderPassParams& renderpass_params, const SPIRVProgram& program) const { - const auto& state = fixed_state; + const auto& state = cache_key.fixed_state; const auto& viewport_swizzles = state.viewport_swizzles; FixedPipelineState::DynamicState dynamic; diff --git a/src/video_core/renderer_vulkan/vk_graphics_pipeline.h b/src/video_core/renderer_vulkan/vk_graphics_pipeline.h index d50bd347c..9d462db0a 100644 --- a/src/video_core/renderer_vulkan/vk_graphics_pipeline.h +++ b/src/video_core/renderer_vulkan/vk_graphics_pipeline.h @@ -19,7 +19,27 @@ namespace Vulkan { using Maxwell = Tegra::Engines::Maxwell3D::Regs; -struct GraphicsPipelineCacheKey; +struct GraphicsPipelineCacheKey { + RenderPassParams renderpass_params; + u32 padding; + std::array shaders; + FixedPipelineState fixed_state; + + std::size_t Hash() const noexcept; + + bool operator==(const GraphicsPipelineCacheKey& rhs) const noexcept; + + bool operator!=(const GraphicsPipelineCacheKey& rhs) const noexcept { + return !operator==(rhs); + } + + std::size_t Size() const noexcept { + return sizeof(renderpass_params) + sizeof(padding) + sizeof(shaders) + fixed_state.Size(); + } +}; +static_assert(std::has_unique_object_representations_v); +static_assert(std::is_trivially_copyable_v); +static_assert(std::is_trivially_constructible_v); class VKDescriptorPool; class VKDevice; @@ -54,8 +74,8 @@ public: return renderpass; } - const GraphicsPipelineCacheKey& GetCacheKey() const { - return m_key; + GraphicsPipelineCacheKey GetCacheKey() const { + return cache_key; } private: @@ -74,8 +94,8 @@ private: const VKDevice& device; VKScheduler& scheduler; - const FixedPipelineState fixed_state; const u64 hash; + GraphicsPipelineCacheKey cache_key; vk::DescriptorSetLayout descriptor_set_layout; DescriptorAllocator descriptor_allocator; @@ -86,8 +106,6 @@ private: VkRenderPass renderpass; vk::Pipeline pipeline; - - const GraphicsPipelineCacheKey& m_key; }; } // namespace Vulkan diff --git a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp index 1a8b2c62b..20ffbeb38 100644 --- a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp +++ b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp @@ -28,6 +28,7 @@ #include "video_core/shader/compiler_settings.h" #include "video_core/shader/memory_util.h" #include "video_core/shader_cache.h" +#include "video_core/shader_notify.h" namespace Vulkan { @@ -214,27 +215,31 @@ VKGraphicsPipeline* VKPipelineCache::GetGraphicsPipeline( } last_graphics_key = key; - if (device.UseAsynchronousShaders()) { + if (device.UseAsynchronousShaders() && async_shaders.IsShaderAsync(system.GPU())) { std::unique_lock lock{pipeline_cache}; const auto [pair, is_cache_miss] = graphics_cache.try_emplace(key); if (is_cache_miss) { + system.GPU().ShaderNotify().MarkSharderBuilding(); LOG_INFO(Render_Vulkan, "Compile 0x{:016X}", key.Hash()); const auto [program, bindings] = DecompileShaders(key.fixed_state); - async_shaders.QueueVulkanShader(this, bindings, program, key.renderpass_params, - key.padding, key.shaders, key.fixed_state); + async_shaders.QueueVulkanShader(this, device, scheduler, descriptor_pool, + update_descriptor_queue, renderpass_cache, bindings, + program, key); } - last_graphics_pipeline = graphics_cache.at(key).get(); + last_graphics_pipeline = pair->second.get(); return last_graphics_pipeline; } const auto [pair, is_cache_miss] = graphics_cache.try_emplace(key); auto& entry = pair->second; if (is_cache_miss) { + system.GPU().ShaderNotify().MarkSharderBuilding(); LOG_INFO(Render_Vulkan, "Compile 0x{:016X}", key.Hash()); const auto [program, bindings] = DecompileShaders(key.fixed_state); entry = std::make_unique(device, scheduler, descriptor_pool, update_descriptor_queue, renderpass_cache, key, bindings, program); + system.GPU().ShaderNotify().MarkShaderComplete(); } last_graphics_pipeline = entry.get(); return last_graphics_pipeline; @@ -294,14 +299,8 @@ VKComputePipeline& VKPipelineCache::GetComputePipeline(const ComputePipelineCach void VKPipelineCache::EmplacePipeline(std::unique_ptr pipeline) { std::unique_lock lock{pipeline_cache}; - const auto [pair, is_cache_miss] = graphics_cache.try_emplace(pipeline->GetCacheKey()); - auto& entry = pair->second; - if (entry) { - LOG_INFO(Render_Vulkan, "Pipeline already here 0x{:016X}", pipeline->GetCacheKey().Hash()); - duplicates.push_back(std::move(pipeline)); - } else { - entry = std::move(pipeline); - } + graphics_cache.at(pipeline->GetCacheKey()) = std::move(pipeline); + system.GPU().ShaderNotify().MarkShaderComplete(); } void VKPipelineCache::OnShaderRemoval(Shader* shader) { diff --git a/src/video_core/renderer_vulkan/vk_pipeline_cache.h b/src/video_core/renderer_vulkan/vk_pipeline_cache.h index 777ef2038..c04829e77 100644 --- a/src/video_core/renderer_vulkan/vk_pipeline_cache.h +++ b/src/video_core/renderer_vulkan/vk_pipeline_cache.h @@ -44,28 +44,6 @@ class VKUpdateDescriptorQueue; using Maxwell = Tegra::Engines::Maxwell3D::Regs; -struct GraphicsPipelineCacheKey { - RenderPassParams renderpass_params; - u32 padding; - std::array shaders; - FixedPipelineState fixed_state; - - std::size_t Hash() const noexcept; - - bool operator==(const GraphicsPipelineCacheKey& rhs) const noexcept; - - bool operator!=(const GraphicsPipelineCacheKey& rhs) const noexcept { - return !operator==(rhs); - } - - std::size_t Size() const noexcept { - return sizeof(renderpass_params) + sizeof(padding) + sizeof(shaders) + fixed_state.Size(); - } -}; -static_assert(std::has_unique_object_representations_v); -static_assert(std::is_trivially_copyable_v); -static_assert(std::is_trivially_constructible_v); - struct ComputePipelineCacheKey { GPUVAddr shader; u32 shared_memory_size; @@ -158,41 +136,6 @@ public: VKComputePipeline& GetComputePipeline(const ComputePipelineCacheKey& key); - const VKDevice& GetDevice() const { - return device; - } - - VKScheduler& GetScheduler() { - return scheduler; - } - const VKScheduler& GetScheduler() const { - return scheduler; - } - - VKDescriptorPool& GetDescriptorPool() { - return descriptor_pool; - } - - const VKDescriptorPool& GetDescriptorPool() const { - return descriptor_pool; - } - - VKUpdateDescriptorQueue& GetUpdateDescriptorQueue() { - return update_descriptor_queue; - } - - const VKUpdateDescriptorQueue& GetUpdateDescriptorQueue() const { - return update_descriptor_queue; - } - - VKRenderPassCache& GetRenderpassCache() { - return renderpass_cache; - } - - const VKRenderPassCache& GetRenderpassCache() const { - return renderpass_cache; - } - void EmplacePipeline(std::unique_ptr pipeline); protected: @@ -216,7 +159,6 @@ private: GraphicsPipelineCacheKey last_graphics_key; VKGraphicsPipeline* last_graphics_pipeline = nullptr; - std::vector> duplicates; std::mutex pipeline_cache; std::unordered_map> diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp index fc1b51a96..720802ad5 100644 --- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp +++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp @@ -14,6 +14,7 @@ #include "common/assert.h" #include "common/logging/log.h" #include "common/microprofile.h" +#include "common/scope_exit.h" #include "core/core.h" #include "core/settings.h" #include "video_core/engines/kepler_compute.h" @@ -408,15 +409,10 @@ RasterizerVulkan::RasterizerVulkan(Core::System& system, Core::Frontend::EmuWind // Max worker threads we should allow constexpr u32 MAX_THREADS = 4; - // Amount of threads we should reserve for other parts of yuzu - constexpr u32 RESERVED_THREADS = 6; - // Get the amount of threads we can use(this can return zero) - const auto cpu_thread_count = - std::max(RESERVED_THREADS, std::thread::hardware_concurrency()); - // Deduce how many "extra" threads we have to use. - const auto max_threads_unused = cpu_thread_count - RESERVED_THREADS; + // Deduce how many threads we can use + const auto threads_used = std::thread::hardware_concurrency() / 4; // Always allow at least 1 thread regardless of our settings - const auto max_worker_count = std::max(1u, max_threads_unused); + const auto max_worker_count = std::max(1U, threads_used); // Don't use more than MAX_THREADS const auto worker_count = std::min(max_worker_count, MAX_THREADS); async_shaders.AllocateWorkers(worker_count); @@ -432,6 +428,8 @@ void RasterizerVulkan::Draw(bool is_indexed, bool is_instanced) { query_cache.UpdateCounters(); + SCOPE_EXIT({ system.GPU().TickWork(); }); + const auto& gpu = system.GPU().Maxwell3D(); GraphicsPipelineCacheKey key; key.fixed_state.Fill(gpu.regs, device.IsExtExtendedDynamicStateSupported()); @@ -458,10 +456,9 @@ void RasterizerVulkan::Draw(bool is_indexed, bool is_instanced) { key.renderpass_params = GetRenderPassParams(texceptions); key.padding = 0; - auto pipeline = pipeline_cache.GetGraphicsPipeline(key, async_shaders); + auto* pipeline = pipeline_cache.GetGraphicsPipeline(key, async_shaders); if (pipeline == nullptr || pipeline->GetHandle() == VK_NULL_HANDLE) { // Async graphics pipeline was not ready. - system.GPU().TickWork(); return; } @@ -488,8 +485,6 @@ void RasterizerVulkan::Draw(bool is_indexed, bool is_instanced) { }); EndTransformFeedback(); - - system.GPU().TickWork(); } void RasterizerVulkan::Clear() { -- cgit v1.2.3 From 1b829fbd7a36f9c2b553b04aa39bdf8135d30458 Mon Sep 17 00:00:00 2001 From: ameerj Date: Wed, 5 Aug 2020 12:53:26 -0400 Subject: move thread 1/4 count computation into allocate workers method --- src/video_core/renderer_vulkan/vk_rasterizer.cpp | 12 +----------- 1 file changed, 1 insertion(+), 11 deletions(-) (limited to 'src/video_core/renderer_vulkan') diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp index 720802ad5..936f76195 100644 --- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp +++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp @@ -405,17 +405,7 @@ RasterizerVulkan::RasterizerVulkan(Core::System& system, Core::Frontend::EmuWind wfi_event{device.GetLogical().CreateNewEvent()}, async_shaders{renderer} { scheduler.SetQueryCache(query_cache); if (device.UseAsynchronousShaders()) { - // The following is subject to move into the allocate workers method, to be api agnostic - - // Max worker threads we should allow - constexpr u32 MAX_THREADS = 4; - // Deduce how many threads we can use - const auto threads_used = std::thread::hardware_concurrency() / 4; - // Always allow at least 1 thread regardless of our settings - const auto max_worker_count = std::max(1U, threads_used); - // Don't use more than MAX_THREADS - const auto worker_count = std::min(max_worker_count, MAX_THREADS); - async_shaders.AllocateWorkers(worker_count); + async_shaders.AllocateWorkers(); } } -- cgit v1.2.3 From fde8102a415c546e88346258bf42de2a248113b1 Mon Sep 17 00:00:00 2001 From: ameerj Date: Sun, 16 Aug 2020 16:33:21 -0400 Subject: Remove unneeded newlines, optional Registry in shader params Addressing feedback from Rodrigo --- src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp | 7 +++---- src/video_core/renderer_vulkan/vk_graphics_pipeline.h | 2 +- src/video_core/renderer_vulkan/vk_pipeline_cache.cpp | 2 +- 3 files changed, 5 insertions(+), 6 deletions(-) (limited to 'src/video_core/renderer_vulkan') diff --git a/src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp b/src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp index 5dc4cd5af..2e46c6278 100644 --- a/src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp +++ b/src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp @@ -78,15 +78,14 @@ VKGraphicsPipeline::VKGraphicsPipeline(const VKDevice& device, VKScheduler& sche const GraphicsPipelineCacheKey& key, vk::Span bindings, const SPIRVProgram& program) - : device{device}, scheduler{scheduler}, hash{key.Hash()}, cache_key{key}, + : device{device}, scheduler{scheduler}, cache_key{key}, hash{cache_key.Hash()}, descriptor_set_layout{CreateDescriptorSetLayout(bindings)}, descriptor_allocator{descriptor_pool, *descriptor_set_layout}, update_descriptor_queue{update_descriptor_queue}, layout{CreatePipelineLayout()}, descriptor_template{CreateDescriptorUpdateTemplate(program)}, modules{CreateShaderModules( program)}, - renderpass{renderpass_cache.GetRenderPass(key.renderpass_params)}, pipeline{CreatePipeline( - key.renderpass_params, - program)} {} + renderpass{renderpass_cache.GetRenderPass(cache_key.renderpass_params)}, + pipeline{CreatePipeline(cache_key.renderpass_params, program)} {} VKGraphicsPipeline::~VKGraphicsPipeline() = default; diff --git a/src/video_core/renderer_vulkan/vk_graphics_pipeline.h b/src/video_core/renderer_vulkan/vk_graphics_pipeline.h index 9d462db0a..58aa35efd 100644 --- a/src/video_core/renderer_vulkan/vk_graphics_pipeline.h +++ b/src/video_core/renderer_vulkan/vk_graphics_pipeline.h @@ -94,8 +94,8 @@ private: const VKDevice& device; VKScheduler& scheduler; + const GraphicsPipelineCacheKey cache_key; const u64 hash; - GraphicsPipelineCacheKey cache_key; vk::DescriptorSetLayout descriptor_set_layout; DescriptorAllocator descriptor_allocator; diff --git a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp index 20ffbeb38..cfdcdd6ab 100644 --- a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp +++ b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp @@ -298,9 +298,9 @@ VKComputePipeline& VKPipelineCache::GetComputePipeline(const ComputePipelineCach } void VKPipelineCache::EmplacePipeline(std::unique_ptr pipeline) { + system.GPU().ShaderNotify().MarkShaderComplete(); std::unique_lock lock{pipeline_cache}; graphics_cache.at(pipeline->GetCacheKey()) = std::move(pipeline); - system.GPU().ShaderNotify().MarkShaderComplete(); } void VKPipelineCache::OnShaderRemoval(Shader* shader) { -- cgit v1.2.3