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

vulkan_renderer: Async shader/graphics pipeline compilation

15 files changed, 210 insertions, 88 deletions

diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index cb284db77..4af5824cd 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -177,15 +177,7 @@ RasterizerOpenGL::RasterizerOpenGL(Core::System& system, Core::Frontend::EmuWind
     }
 
     if (device.UseAsynchronousShaders()) {
-        // Max worker threads we should allow
-        constexpr u32 MAX_THREADS = 4;
-        // Deduce how many threads we can use
-        const u32 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();
     }
 }
 
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;
 
diff --git a/src/video_core/shader/async_shaders.cpp b/src/video_core/shader/async_shaders.cpp
index 3f6d5a75b..f815584f7 100644
--- a/src/video_core/shader/async_shaders.cpp
+++ b/src/video_core/shader/async_shaders.cpp
@@ -2,7 +2,6 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
-#include <chrono>
 #include <condition_variable>
 #include <mutex>
 #include <thread>
@@ -20,9 +19,18 @@ AsyncShaders::~AsyncShaders() {
     KillWorkers();
 }
 
-void AsyncShaders::AllocateWorkers(std::size_t num_workers) {
-    // If we're already have workers queued or don't want to queue workers, ignore
-    if (num_workers == worker_threads.size() || num_workers == 0) {
+void AsyncShaders::AllocateWorkers() {
+    // Max worker threads we should allow
+    constexpr u32 MAX_THREADS = 4;
+    // Deduce how many threads we can use
+    const u32 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 num_workers = std::min(max_worker_count, MAX_THREADS);
+
+    // If we already have workers queued, ignore
+    if (num_workers == worker_threads.size()) {
         return;
     }
 
@@ -111,24 +119,50 @@ void AsyncShaders::QueueOpenGLShader(const OpenGL::Device& device,
                                      VideoCommon::Shader::CompilerSettings compiler_settings,
                                      const VideoCommon::Shader::Registry& registry,
                                      VAddr cpu_addr) {
-    WorkerParams params{device.UseAssemblyShaders() ? AsyncShaders::Backend::GLASM
-                                                    : AsyncShaders::Backend::OpenGL,
-                        device,
-                        shader_type,
-                        uid,
-                        std::move(code),
-                        std::move(code_b),
-                        main_offset,
-                        compiler_settings,
-                        registry,
-                        cpu_addr};
+    WorkerParams params{
+        .backend = device.UseAssemblyShaders() ? Backend::GLASM : Backend::OpenGL,
+        .device = &device,
+        .shader_type = shader_type,
+        .uid = uid,
+        .code = std::move(code),
+        .code_b = std::move(code_b),
+        .main_offset = main_offset,
+        .compiler_settings = compiler_settings,
+        .registry = registry,
+        .cpu_address = cpu_addr,
+    };
     std::unique_lock lock(queue_mutex);
-    pending_queue.push_back(std::move(params));
+    pending_queue.push(std::move(params));
+    cv.notify_one();
+}
+
+void AsyncShaders::QueueVulkanShader(Vulkan::VKPipelineCache* pp_cache,
+                                     const Vulkan::VKDevice& device, Vulkan::VKScheduler& scheduler,
+                                     Vulkan::VKDescriptorPool& descriptor_pool,
+                                     Vulkan::VKUpdateDescriptorQueue& update_descriptor_queue,
+                                     Vulkan::VKRenderPassCache& renderpass_cache,
+                                     std::vector<VkDescriptorSetLayoutBinding> bindings,
+                                     Vulkan::SPIRVProgram program,
+                                     Vulkan::GraphicsPipelineCacheKey key) {
+    WorkerParams params{
+        .backend = Backend::Vulkan,
+        .pp_cache = pp_cache,
+        .vk_device = &device,
+        .scheduler = &scheduler,
+        .descriptor_pool = &descriptor_pool,
+        .update_descriptor_queue = &update_descriptor_queue,
+        .renderpass_cache = &renderpass_cache,
+        .bindings = bindings,
+        .program = program,
+        .key = key,
+    };
+
+    std::unique_lock lock(queue_mutex);
+    pending_queue.push(std::move(params));
     cv.notify_one();
 }
 
 void AsyncShaders::ShaderCompilerThread(Core::Frontend::GraphicsContext* context) {
-    using namespace std::chrono_literals;
     while (!is_thread_exiting.load(std::memory_order_relaxed)) {
         std::unique_lock lock{queue_mutex};
         cv.wait(lock, [this] { return HasWorkQueued() || is_thread_exiting; });
@@ -144,18 +178,17 @@ void AsyncShaders::ShaderCompilerThread(Core::Frontend::GraphicsContext* context
         if (pending_queue.empty()) {
             continue;
         }
+
         // Pull work from queue
         WorkerParams work = std::move(pending_queue.front());
-        pending_queue.pop_front();
-
+        pending_queue.pop();
         lock.unlock();
 
-        if (work.backend == AsyncShaders::Backend::OpenGL ||
-            work.backend == AsyncShaders::Backend::GLASM) {
-            const ShaderIR ir(work.code, work.main_offset, work.compiler_settings, work.registry);
+        if (work.backend == Backend::OpenGL || work.backend == Backend::GLASM) {
+            const ShaderIR ir(work.code, work.main_offset, work.compiler_settings, *work.registry);
             const auto scope = context->Acquire();
             auto program =
-                OpenGL::BuildShader(work.device, work.shader_type, work.uid, ir, work.registry);
+                OpenGL::BuildShader(*work.device, work.shader_type, work.uid, ir, *work.registry);
             Result result{};
             result.backend = work.backend;
             result.cpu_address = work.cpu_address;
@@ -164,9 +197,9 @@ void AsyncShaders::ShaderCompilerThread(Core::Frontend::GraphicsContext* context
             result.code_b = std::move(work.code_b);
             result.shader_type = work.shader_type;
 
-            if (work.backend == AsyncShaders::Backend::OpenGL) {
+            if (work.backend == Backend::OpenGL) {
                 result.program.opengl = std::move(program->source_program);
-            } else if (work.backend == AsyncShaders::Backend::GLASM) {
+            } else if (work.backend == Backend::GLASM) {
                 result.program.glasm = std::move(program->assembly_program);
             }
 
@@ -174,6 +207,13 @@ void AsyncShaders::ShaderCompilerThread(Core::Frontend::GraphicsContext* context
                 std::unique_lock complete_lock(completed_mutex);
                 finished_work.push_back(std::move(result));
             }
+        } else if (work.backend == Backend::Vulkan) {
+            auto pipeline = std::make_unique<Vulkan::VKGraphicsPipeline>(
+                *work.vk_device, *work.scheduler, *work.descriptor_pool,
+                *work.update_descriptor_queue, *work.renderpass_cache, work.key, work.bindings,
+                work.program);
+
+            work.pp_cache->EmplacePipeline(std::move(pipeline));
         }
     }
 }
diff --git a/src/video_core/shader/async_shaders.h b/src/video_core/shader/async_shaders.h
index 2f5ee94ad..d5ae814d5 100644
--- a/src/video_core/shader/async_shaders.h
+++ b/src/video_core/shader/async_shaders.h
@@ -14,6 +14,10 @@
 #include "video_core/renderer_opengl/gl_device.h"
 #include "video_core/renderer_opengl/gl_resource_manager.h"
 #include "video_core/renderer_opengl/gl_shader_decompiler.h"
+#include "video_core/renderer_vulkan/vk_device.h"
+#include "video_core/renderer_vulkan/vk_pipeline_cache.h"
+#include "video_core/renderer_vulkan/vk_scheduler.h"
+#include "video_core/renderer_vulkan/vk_update_descriptor.h"
 
 namespace Core::Frontend {
 class EmuWindow;
@@ -24,6 +28,10 @@ namespace Tegra {
 class GPU;
 }
 
+namespace Vulkan {
+class VKPipelineCache;
+}
+
 namespace VideoCommon::Shader {
 
 class AsyncShaders {
@@ -31,6 +39,7 @@ public:
     enum class Backend {
         OpenGL,
         GLASM,
+        Vulkan,
     };
 
     struct ResultPrograms {
@@ -52,7 +61,7 @@ public:
     ~AsyncShaders();
 
     /// Start up shader worker threads
-    void AllocateWorkers(std::size_t num_workers);
+    void AllocateWorkers();
 
     /// Clear the shader queue and kill all worker threads
     void FreeWorkers();
@@ -76,6 +85,14 @@ public:
                            VideoCommon::Shader::CompilerSettings compiler_settings,
                            const VideoCommon::Shader::Registry& registry, VAddr cpu_addr);
 
+    void QueueVulkanShader(Vulkan::VKPipelineCache* pp_cache, const Vulkan::VKDevice& device,
+                           Vulkan::VKScheduler& scheduler,
+                           Vulkan::VKDescriptorPool& descriptor_pool,
+                           Vulkan::VKUpdateDescriptorQueue& update_descriptor_queue,
+                           Vulkan::VKRenderPassCache& renderpass_cache,
+                           std::vector<VkDescriptorSetLayoutBinding> bindings,
+                           Vulkan::SPIRVProgram program, Vulkan::GraphicsPipelineCacheKey key);
+
 private:
     void ShaderCompilerThread(Core::Frontend::GraphicsContext* context);
 
@@ -83,16 +100,28 @@ private:
     bool HasWorkQueued();
 
     struct WorkerParams {
-        AsyncShaders::Backend backend;
-        OpenGL::Device device;
+        Backend backend;
+        // For OGL
+        const OpenGL::Device* device;
         Tegra::Engines::ShaderType shader_type;
         u64 uid;
         std::vector<u64> code;
         std::vector<u64> code_b;
         u32 main_offset;
         VideoCommon::Shader::CompilerSettings compiler_settings;
-        VideoCommon::Shader::Registry registry;
+        std::optional<VideoCommon::Shader::Registry> registry;
         VAddr cpu_address;
+
+        // For Vulkan
+        Vulkan::VKPipelineCache* pp_cache;
+        const Vulkan::VKDevice* vk_device;
+        Vulkan::VKScheduler* scheduler;
+        Vulkan::VKDescriptorPool* descriptor_pool;
+        Vulkan::VKUpdateDescriptorQueue* update_descriptor_queue;
+        Vulkan::VKRenderPassCache* renderpass_cache;
+        std::vector<VkDescriptorSetLayoutBinding> bindings;
+        Vulkan::SPIRVProgram program;
+        Vulkan::GraphicsPipelineCacheKey key;
     };
 
     std::condition_variable cv;
@@ -101,7 +130,7 @@ private:
     std::atomic<bool> is_thread_exiting{};
     std::vector<std::unique_ptr<Core::Frontend::GraphicsContext>> context_list;
     std::vector<std::thread> worker_threads;
-    std::deque<WorkerParams> pending_queue;
+    std::queue<WorkerParams> pending_queue;
     std::vector<AsyncShaders::Result> finished_work;
     Core::Frontend::EmuWindow& emu_window;
 };
diff --git a/src/yuzu/configuration/configure_graphics_advanced.ui b/src/yuzu/configuration/configure_graphics_advanced.ui
index a793c803d..846a30586 100644
--- a/src/yuzu/configuration/configure_graphics_advanced.ui
+++ b/src/yuzu/configuration/configure_graphics_advanced.ui
@@ -92,7 +92,7 @@
            <string>Enables asynchronous shader compilation, which may reduce shader stutter. This feature is experimental.</string>
           </property>
           <property name="text">
-           <string>Use asynchronous shader building (experimental, OpenGL or Assembly shaders only)</string>
+           <string>Use asynchronous shader building (experimental)</string>
           </property>
          </widget>
         </item>