Merge pull request #4443 from ameerj/vk-async-shaders

vulkan_renderer: Async shader/graphics pipeline compilation

11 files changed, 109 insertions, 48 deletions

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..26a233db1 100644
--- a/src/video_core/renderer_vulkan/vk_device.h
+++ b/src/video_core/renderer_vulkan/vk_device.h
@@ -202,6 +202,11 @@ public:
         return reported_extensions;
     }
 
+    /// Returns true if the setting for async shader compilation is enabled.
+    bool UseAsynchronousShaders() const {
+        return use_asynchronous_shaders;
+    }
+
     /// Checks if the physical device is suitable.
     static bool IsSuitable(vk::PhysicalDevice physical, VkSurfaceKHR surface);
 
@@ -252,6 +257,9 @@ private:
     bool ext_extended_dynamic_state{};         ///< Support for VK_EXT_extended_dynamic_state.
     bool nv_device_diagnostics_config{};       ///< Support for VK_NV_device_diagnostics_config.
 
+    // 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.
     std::vector<std::string> reported_extensions; ///< Reported Vulkan extensions.
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..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<VkDescriptorSetLayoutBinding> bindings,
                                        const SPIRVProgram& program)
-    : device{device}, scheduler{scheduler}, fixed_state{key.fixed_state}, hash{key.Hash()},
+    : 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;
 
@@ -181,7 +180,7 @@ std::vector<vk::ShaderModule> 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 a1d699a6c..58aa35efd 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<GPUVAddr, Maxwell::MaxShaderProgram> 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<GraphicsPipelineCacheKey>);
+static_assert(std::is_trivially_copyable_v<GraphicsPipelineCacheKey>);
+static_assert(std::is_trivially_constructible_v<GraphicsPipelineCacheKey>);
 
 class VKDescriptorPool;
 class VKDevice;
@@ -54,6 +74,10 @@ public:
         return renderpass;
     }
 
+    GraphicsPipelineCacheKey GetCacheKey() const {
+        return cache_key;
+    }
+
 private:
     vk::DescriptorSetLayout CreateDescriptorSetLayout(
         vk::Span<VkDescriptorSetLayoutBinding> bindings) const;
@@ -70,7 +94,7 @@ private:
 
     const VKDevice& device;
     VKScheduler& scheduler;
-    const FixedPipelineState fixed_state;
+    const GraphicsPipelineCacheKey cache_key;
     const u64 hash;
 
     vk::DescriptorSetLayout descriptor_set_layout;
diff --git a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
index 418c62bc4..cfdcdd6ab 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 {
 
@@ -205,24 +206,43 @@ std::array<Shader*, Maxwell::MaxShaderProgram> 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) {
-        return *last_graphics_pipeline;
+        return last_graphics_pipeline;
     }
     last_graphics_key = key;
 
+    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, device, scheduler, descriptor_pool,
+                                            update_descriptor_queue, renderpass_cache, bindings,
+                                            program, key);
+        }
+        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);
+        const auto [program, bindings] = DecompileShaders(key.fixed_state);
         entry = std::make_unique<VKGraphicsPipeline>(device, scheduler, descriptor_pool,
                                                      update_descriptor_queue, renderpass_cache, key,
                                                      bindings, program);
+        system.GPU().ShaderNotify().MarkShaderComplete();
     }
-    return *(last_graphics_pipeline = entry.get());
+    last_graphics_pipeline = entry.get();
+    return last_graphics_pipeline;
 }
 
 VKComputePipeline& VKPipelineCache::GetComputePipeline(const ComputePipelineCacheKey& key) {
@@ -277,6 +297,12 @@ VKComputePipeline& VKPipelineCache::GetComputePipeline(const ComputePipelineCach
     return *entry;
 }
 
+void VKPipelineCache::EmplacePipeline(std::unique_ptr<VKGraphicsPipeline> pipeline) {
+    system.GPU().ShaderNotify().MarkShaderComplete();
+    std::unique_lock lock{pipeline_cache};
+    graphics_cache.at(pipeline->GetCacheKey()) = std::move(pipeline);
+}
+
 void VKPipelineCache::OnShaderRemoval(Shader* shader) {
     bool finished = false;
     const auto Finish = [&] {
@@ -312,8 +338,7 @@ void VKPipelineCache::OnShaderRemoval(Shader* shader) {
 }
 
 std::pair<SPIRVProgram, std::vector<VkDescriptorSetLayoutBinding>>
-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..c04829e77 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"
@@ -43,28 +44,6 @@ class VKUpdateDescriptorQueue;
 
 using Maxwell = Tegra::Engines::Maxwell3D::Regs;
 
-struct GraphicsPipelineCacheKey {
-    RenderPassParams renderpass_params;
-    u32 padding;
-    std::array<GPUVAddr, Maxwell::MaxShaderProgram> 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<GraphicsPipelineCacheKey>);
-static_assert(std::is_trivially_copyable_v<GraphicsPipelineCacheKey>);
-static_assert(std::is_trivially_constructible_v<GraphicsPipelineCacheKey>);
-
 struct ComputePipelineCacheKey {
     GPUVAddr shader;
     u32 shared_memory_size;
@@ -152,16 +131,19 @@ public:
 
     std::array<Shader*, Maxwell::MaxShaderProgram> GetShaders();
 
-    VKGraphicsPipeline& GetGraphicsPipeline(const GraphicsPipelineCacheKey& key);
+    VKGraphicsPipeline* GetGraphicsPipeline(const GraphicsPipelineCacheKey& key,
+                                            VideoCommon::Shader::AsyncShaders& async_shaders);
 
     VKComputePipeline& GetComputePipeline(const ComputePipelineCacheKey& key);
 
+    void EmplacePipeline(std::unique_ptr<VKGraphicsPipeline> pipeline);
+
 protected:
     void OnShaderRemoval(Shader* shader) final;
 
 private:
     std::pair<SPIRVProgram, std::vector<VkDescriptorSetLayoutBinding>> DecompileShaders(
-        const GraphicsPipelineCacheKey& key);
+        const FixedPipelineState& fixed_state);
 
     Core::System& system;
     const VKDevice& device;
@@ -178,6 +160,7 @@ private:
     GraphicsPipelineCacheKey last_graphics_key;
     VKGraphicsPipeline* last_graphics_pipeline = nullptr;
 
+    std::mutex pipeline_cache;
     std::unordered_map<GraphicsPipelineCacheKey, std::unique_ptr<VKGraphicsPipeline>>
         graphics_cache;
     std::unordered_map<ComputePipelineCacheKey, std::unique_ptr<VKComputePipeline>> compute_cache;
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
index 7500e8244..936f76195 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"
@@ -400,8 +401,12 @@ 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()) {
+        async_shaders.AllocateWorkers();
+    }
 }
 
 RasterizerVulkan::~RasterizerVulkan() = default;
@@ -413,6 +418,8 @@ 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());
@@ -439,10 +446,15 @@ void RasterizerVulkan::Draw(bool is_indexed, bool is_instanced) {
     key.renderpass_params = GetRenderPassParams(texceptions);
     key.padding = 0;
 
-    auto& pipeline = pipeline_cache.GetGraphicsPipeline(key);
-    scheduler.BindGraphicsPipeline(pipeline.GetHandle());
+    auto* pipeline = pipeline_cache.GetGraphicsPipeline(key, async_shaders);
+    if (pipeline == nullptr || pipeline->GetHandle() == VK_NULL_HANDLE) {
+        // Async graphics pipeline was not ready.
+        return;
+    }
+
+    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);
 
@@ -452,8 +464,8 @@ 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,
@@ -463,8 +475,6 @@ void RasterizerVulkan::Draw(bool is_indexed, bool is_instanced) {
     });
 
     EndTransformFeedback();
-
-    system.GPU().TickWork();
 }
 
 void RasterizerVulkan::Clear() {
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.h b/src/video_core/renderer_vulkan/vk_rasterizer.h
index 923178b0b..f640ba649 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,
@@ -297,6 +306,7 @@ private:
     vk::Buffer default_buffer;
     VKMemoryCommit default_buffer_commit;
     vk::Event wfi_event;
+    VideoCommon::Shader::AsyncShaders async_shaders;
 
     std::array<View, Maxwell::NumRenderTargets> color_attachments;
     View zeta_attachment;
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;