dxvk-async.patch

diff --git a/meson.build b/meson.build
index f044c8f4..36b22d39 100644
--- a/meson.build
+++ b/meson.build
@@ -122,7 +122,7 @@ glsl_generator = generator(glsl_compiler,
 )
 
 dxvk_version = vcs_tag(
-  command: ['git', 'describe', '--dirty=+'],
+  command: ['git', 'describe', '--dirty=-async'],
   input:  'version.h.in',
   output: 'version.h',
 )
diff --git a/src/dxvk/dxvk_context.cpp b/src/dxvk/dxvk_context.cpp
index 6486d267..cc59297e 100644
--- a/src/dxvk/dxvk_context.cpp
+++ b/src/dxvk/dxvk_context.cpp
@@ -4540,7 +4540,8 @@ namespace dxvk {
       : DxvkContextFlag::GpDirtyRasterizerState);
 
     // Retrieve and bind actual Vulkan pipeline handle
-    auto pipelineInfo = m_state.gp.pipeline->getPipelineHandle(m_state.gp.state);
+    auto pipelineInfo = m_state.gp.pipeline->getPipelineHandle(
+      m_state.gp.state, this->checkAsyncCompilationCompat());
 
     if (unlikely(!pipelineInfo.first))
       return false;
@@ -4890,7 +4891,7 @@ namespace dxvk {
   }
 
 
-  void DxvkContext::updateFramebuffer() {
+  void DxvkContext::updateFramebuffer(bool isDraw) {
     if (m_flags.test(DxvkContextFlag::GpDirtyFramebuffer)) {
       m_flags.clr(DxvkContextFlag::GpDirtyFramebuffer);
 
@@ -4914,6 +4915,11 @@ namespace dxvk {
         m_state.gp.state.omSwizzle[i] = DxvkOmAttachmentSwizzle(mapping);
       }
 
+      if (isDraw) {
+        for (uint32_t i = 0; i < fbInfo.numAttachments(); i++)
+          fbInfo.getAttachment(i).view->setRtBindingFrameId(m_device->getCurrentFrameId());
+      }
+
       m_flags.set(DxvkContextFlag::GpDirtyPipelineState);
     }
   }
@@ -5363,7 +5369,7 @@ namespace dxvk {
     }
     
     if (m_flags.test(DxvkContextFlag::GpDirtyFramebuffer))
-      this->updateFramebuffer();
+      this->updateFramebuffer(true);
 
     if (!m_flags.test(DxvkContextFlag::GpRenderPassBound))
       this->startRenderPass();
@@ -5774,6 +5780,14 @@ namespace dxvk {
     return true;
   }
   
+  bool DxvkContext::checkAsyncCompilationCompat() {
+    bool fbCompat = true;
+    for (uint32_t i = 0; fbCompat && i < m_state.om.framebufferInfo.numAttachments(); i++) {
+      const auto& attachment = m_state.om.framebufferInfo.getAttachment(i);
+      fbCompat &= attachment.view->getRtBindingAsyncCompilationCompat();
+    }
+    return fbCompat;
+  }
 
   DxvkGraphicsPipeline* DxvkContext::lookupGraphicsPipeline(
     const DxvkGraphicsPipelineShaders&  shaders) {
diff --git a/src/dxvk/dxvk_context.h b/src/dxvk/dxvk_context.h
index fe24c556..0930bd81 100644
--- a/src/dxvk/dxvk_context.h
+++ b/src/dxvk/dxvk_context.h
@@ -1493,7 +1493,7 @@ namespace dxvk {
     DxvkFramebufferInfo makeFramebufferInfo(
       const DxvkRenderTargets&      renderTargets);
 
-    void updateFramebuffer();
+    void updateFramebuffer(bool isDraw = false);
     
     void applyRenderTargetLoadLayouts();
 
@@ -1576,6 +1576,8 @@ namespace dxvk {
       const Rc<DxvkBuffer>&           buffer,
             VkDeviceSize              copySize);
 
+    bool checkAsyncCompilationCompat();
+
     DxvkGraphicsPipeline* lookupGraphicsPipeline(
       const DxvkGraphicsPipelineShaders&  shaders);
 
diff --git a/src/dxvk/dxvk_graphics.cpp b/src/dxvk/dxvk_graphics.cpp
index c0ecbbb4..f4e4fd4f 100644
--- a/src/dxvk/dxvk_graphics.cpp
+++ b/src/dxvk/dxvk_graphics.cpp
@@ -899,7 +899,8 @@ namespace dxvk {
 
 
   std::pair<VkPipeline, DxvkGraphicsPipelineType> DxvkGraphicsPipeline::getPipelineHandle(
-    const DxvkGraphicsPipelineStateInfo& state) {
+    const DxvkGraphicsPipelineStateInfo& state,
+          bool                           async) {
     DxvkGraphicsPipelineInstance* instance = this->findInstance(state);
 
     if (unlikely(!instance)) {
@@ -907,28 +908,36 @@ namespace dxvk {
       if (!this->validatePipelineState(state, true))
         return std::make_pair(VK_NULL_HANDLE, DxvkGraphicsPipelineType::FastPipeline);
 
+      bool useAsync = m_device->config().enableAsync && async;
+
       // Prevent other threads from adding new instances and check again
-      std::unique_lock<dxvk::mutex> lock(m_mutex);
+      std::unique_lock<dxvk::mutex> lock(useAsync ? m_asyncMutex : m_mutex);
       instance = this->findInstance(state);
 
       if (!instance) {
-        // Keep pipeline object locked, at worst we're going to stall
-        // a state cache worker and the current thread needs priority.
-        bool canCreateBasePipeline = this->canCreateBasePipeline(state);
-        instance = this->createInstance(state, canCreateBasePipeline);
-
-        // Unlock here since we may dispatch the pipeline to a worker,
-        // which will then acquire it to increment the use counter.
-        lock.unlock();
-
-        // If necessary, compile an optimized pipeline variant
-        if (!instance->fastHandle.load())
+        if (useAsync) {
+          lock.unlock();
           m_workers->compileGraphicsPipeline(this, state);
-
-        // Only store pipelines in the state cache that cannot benefit
-        // from pipeline libraries, or if that feature is disabled.
-        if (!canCreateBasePipeline)
-          this->writePipelineStateToCache(state);
+          return std::make_pair(VK_NULL_HANDLE, DxvkGraphicsPipelineType::FastPipeline);
+        } else {
+          // Keep pipeline object locked, at worst we're going to stall
+          // a state cache worker and the current thread needs priority.
+          bool canCreateBasePipeline = this->canCreateBasePipeline(state);
+          instance = this->createInstance(state, canCreateBasePipeline);
+
+          // Unlock here since we may dispatch the pipeline to a worker,
+          // which will then acquire it to increment the use counter.
+          lock.unlock();
+
+          // If necessary, compile an optimized pipeline variant
+          if (!instance->fastHandle.load())
+            m_workers->compileGraphicsPipeline(this, state);
+
+          // Only store pipelines in the state cache that cannot benefit
+          // from pipeline libraries, or if that feature is disabled.
+          if (!canCreateBasePipeline)
+            this->writePipelineStateToCache(state);
+        }
       }
     }
 
diff --git a/src/dxvk/dxvk_graphics.h b/src/dxvk/dxvk_graphics.h
index cfc424cd..437c73a9 100644
--- a/src/dxvk/dxvk_graphics.h
+++ b/src/dxvk/dxvk_graphics.h
@@ -521,10 +521,12 @@ namespace dxvk {
      * Retrieves a pipeline handle for the given pipeline
      * state. If necessary, a new pipeline will be created.
      * \param [in] state Pipeline state vector
+     * \param [in] async Compile asynchronously
      * \returns Pipeline handle and handle type
      */
     std::pair<VkPipeline, DxvkGraphicsPipelineType> getPipelineHandle(
-      const DxvkGraphicsPipelineStateInfo&    state);
+      const DxvkGraphicsPipelineStateInfo&    state,
+            bool                              async);
     
     /**
      * \brief Compiles a pipeline
@@ -576,6 +578,8 @@ namespace dxvk {
 
     alignas(CACHE_LINE_SIZE)
     dxvk::mutex                                   m_mutex;
+    alignas(CACHE_LINE_SIZE)
+    dxvk::mutex                                   m_asyncMutex;
     sync::List<DxvkGraphicsPipelineInstance>      m_pipelines;
     uint32_t                                      m_useCount = 0;
 
diff --git a/src/dxvk/dxvk_image.h b/src/dxvk/dxvk_image.h
index 285c5e9e..989a240f 100644
--- a/src/dxvk/dxvk_image.h
+++ b/src/dxvk/dxvk_image.h
@@ -547,6 +547,37 @@ namespace dxvk {
         view->imageSubresources());
     }
 
+    /**
+     * \brief Sets render target usage frame number
+     *
+     * The image view will track internally when
+     * it was last used as a render target. This
+     * info is used for async shader compilation.
+     * \param [in] frameId Frame number
+     */
+    void setRtBindingFrameId(uint32_t frameId) {
+      if (frameId != m_rtBindingFrameId) {
+        if (frameId == m_rtBindingFrameId + 1)
+          m_rtBindingFrameCount += 1;
+        else
+          m_rtBindingFrameCount = 0;
+
+        m_rtBindingFrameId = frameId;
+      }
+    }
+
+    /**
+     * \brief Checks for async pipeline compatibility
+     *
+     * Asynchronous pipeline compilation may be enabled if the
+     * render target has been drawn to in the previous frames.
+     * \param [in] frameId Current frame ID
+     * \returns \c true if async compilation is supported
+     */
+    bool getRtBindingAsyncCompilationCompat() const {
+      return m_rtBindingFrameCount >= 5;
+    }
+
   private:
     
     Rc<vk::DeviceFn>  m_vkd;
@@ -555,6 +586,9 @@ namespace dxvk {
     DxvkImageViewCreateInfo m_info;
     VkImageView             m_views[ViewCount];
 
+    uint32_t m_rtBindingFrameId    = 0;
+    uint32_t m_rtBindingFrameCount = 0;
+
     void createView(VkImageViewType type, uint32_t numLayers);
     
   };
diff --git a/src/dxvk/dxvk_options.cpp b/src/dxvk/dxvk_options.cpp
index 00f284aa..f13429eb 100644
--- a/src/dxvk/dxvk_options.cpp
+++ b/src/dxvk/dxvk_options.cpp
@@ -10,6 +10,14 @@ namespace dxvk {
     trackPipelineLifetime = config.getOption<Tristate>("dxvk.trackPipelineLifetime",  Tristate::Auto);
     useRawSsbo            = config.getOption<Tristate>("dxvk.useRawSsbo",             Tristate::Auto);
     hud                   = config.getOption<std::string>("dxvk.hud", "");
+
+    if (env::getEnvVar("DXVK_ASYNC") == "1")
+      enableAsync = true;
+    else
+      enableAsync = config.getOption<bool>("dxvk.enableAsync", false);
+
+    if (enableAsync)
+      enableGraphicsPipelineLibrary = Tristate::False;
   }
 
 }
diff --git a/src/dxvk/dxvk_options.h b/src/dxvk/dxvk_options.h
index da5bf406..05a7823e 100644
--- a/src/dxvk/dxvk_options.h
+++ b/src/dxvk/dxvk_options.h
@@ -2,6 +2,8 @@
 
 #include "../util/config/config.h"
 
+#include "dxvk_include.h"
+
 namespace dxvk {
 
   struct DxvkOptions {
@@ -24,6 +26,9 @@ namespace dxvk {
     /// Enables pipeline lifetime tracking
     Tristate trackPipelineLifetime;
 
+    // Enable async pipelines
+    bool enableAsync;
+
     /// Shader-related options
     Tristate useRawSsbo;