path: root/dxvk-async.patch

diff --git a/meson.build b/meson.build
index d2d1b6cd..aff9446c 100644
--- a/meson.build
+++ b/meson.build
@@ -90,7 +90,7 @@ else
 endif
 
 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 9f9a4c25..b9e34316 100644
--- a/src/dxvk/dxvk_context.cpp
+++ b/src/dxvk/dxvk_context.cpp
@@ -590,7 +590,7 @@ namespace dxvk {
     const Rc<DxvkImageView>&    imageView,
           VkImageAspectFlags    clearAspects,
     const VkClearValue&         clearValue) {
-    this->updateFramebuffer();
+    this->updateFramebuffer(false);
 
     // Prepare attachment ops
     DxvkColorAttachmentOps colorOp;
@@ -2237,7 +2237,7 @@ namespace dxvk {
           VkOffset3D            offset,
           VkExtent3D            extent,
           VkClearValue          value) {
-    this->updateFramebuffer();
+    this->updateFramebuffer(false);
 
     // Find out if the render target view is currently bound,
     // so that we can avoid spilling the render pass if it is.
@@ -3354,7 +3354,7 @@ namespace dxvk {
       // Retrieve and bind actual Vulkan pipeline handle
       m_gpActivePipeline = m_state.gp.pipeline != nullptr && m_state.om.framebuffer != nullptr
         ? m_state.gp.pipeline->getPipelineHandle(m_state.gp.state,
-            m_state.om.framebuffer->getRenderPass())
+            m_state.om.framebuffer->getRenderPass(), this->checkAsyncCompilationCompat())
         : VK_NULL_HANDLE;
       
       if (m_gpActivePipeline != VK_NULL_HANDLE) {
@@ -3623,7 +3623,7 @@ namespace dxvk {
   }
   
   
-  void DxvkContext::updateFramebuffer() {
+  void DxvkContext::updateFramebuffer(bool isDraw) {
     if (m_flags.test(DxvkContextFlag::GpDirtyFramebuffer)) {
       m_flags.clr(DxvkContextFlag::GpDirtyFramebuffer);
       
@@ -3642,6 +3642,11 @@ namespace dxvk {
           : VkComponentMapping();
       }
 
+      if (isDraw) {
+        for (uint32_t i = 0; i < fb->numAttachments(); i++)
+          fb->getAttachment(i).view->setRtBindingFrameId(m_device->getCurrentFrameId());
+      }
+
       m_flags.set(DxvkContextFlag::GpDirtyPipelineState);
     }
   }
@@ -3876,7 +3881,7 @@ namespace dxvk {
   
   void DxvkContext::commitGraphicsState(bool indexed) {
     if (m_flags.test(DxvkContextFlag::GpDirtyFramebuffer))
-      this->updateFramebuffer();
+      this->updateFramebuffer(true);
 
     if (!m_flags.test(DxvkContextFlag::GpRenderPassBound))
       this->startRenderPass();
@@ -4124,4 +4129,13 @@ namespace dxvk {
     }
   }
   
+
+  bool DxvkContext::checkAsyncCompilationCompat() {
+    bool fbCompat = true;
+    for (uint32_t i = 0; fbCompat && i < m_state.om.framebuffer->numAttachments(); i++) {
+      const auto& attachment = m_state.om.framebuffer->getAttachment(i);
+      fbCompat &= attachment.view->getRtBindingAsyncCompilationCompat();
+    }
+    return fbCompat;
+  }
 }
\ No newline at end of file
diff --git a/src/dxvk/dxvk_context.h b/src/dxvk/dxvk_context.h
index 68aa5542..b264f6ee 100644
--- a/src/dxvk/dxvk_context.h
+++ b/src/dxvk/dxvk_context.h
@@ -1076,7 +1076,7 @@ namespace dxvk {
             VkDescriptorSet         set,
       const DxvkPipelineLayout*     layout);
 
-    void updateFramebuffer();
+    void updateFramebuffer(bool isDraw);
     
     void updateIndexBufferBinding();
     void updateVertexBufferBindings();
@@ -1102,6 +1102,8 @@ namespace dxvk {
     
     void commitGraphicsPostBarriers();
 
+    bool checkAsyncCompilationCompat();
+
     void emitMemoryBarrier(
             VkPipelineStageFlags      srcStages,
             VkAccessFlags             srcAccess,
diff --git a/src/dxvk/dxvk_framebuffer.h b/src/dxvk/dxvk_framebuffer.h
index 49133845..ff80297b 100644
--- a/src/dxvk/dxvk_framebuffer.h
+++ b/src/dxvk/dxvk_framebuffer.h
@@ -120,8 +120,8 @@ namespace dxvk {
      * \brief Retrieves render pass
      * \returns Render pass reference
      */
-    const DxvkRenderPass& getRenderPass() const {
-      return *m_renderPass;
+    const Rc<DxvkRenderPass>& getRenderPass() const {
+      return m_renderPass;
     }
     
     /**
@@ -221,4 +221,4 @@ namespace dxvk {
     
   };
   
-}
\ No newline at end of file
+}
diff --git a/src/dxvk/dxvk_graphics.cpp b/src/dxvk/dxvk_graphics.cpp
index d50fb209..3e2132c0 100644
--- a/src/dxvk/dxvk_graphics.cpp
+++ b/src/dxvk/dxvk_graphics.cpp
@@ -4,6 +4,7 @@
 #include "dxvk_device.h"
 #include "dxvk_graphics.h"
 #include "dxvk_pipemanager.h"
+#include "dxvk_pipecompiler.h"
 #include "dxvk_spec_const.h"
 #include "dxvk_state_cache.h"
 
@@ -79,8 +80,6 @@ namespace dxvk {
   
   
   DxvkGraphicsPipeline::~DxvkGraphicsPipeline() {
-    for (const auto& instance : m_pipelines)
-      this->destroyPipeline(instance.pipeline());
   }
   
   
@@ -100,8 +99,9 @@ namespace dxvk {
 
   VkPipeline DxvkGraphicsPipeline::getPipelineHandle(
     const DxvkGraphicsPipelineStateInfo& state,
-    const DxvkRenderPass&                renderPass) {
-    VkRenderPass renderPassHandle = renderPass.getDefaultHandle();
+    const Rc<DxvkRenderPass>&            renderPass,
+          bool                           async) {
+    VkRenderPass renderPassHandle = renderPass->getDefaultHandle();
     
     VkPipeline newPipelineHandle = VK_NULL_HANDLE;
 
@@ -119,29 +119,49 @@ namespace dxvk {
       
       // If no pipeline instance exists with the given state
       // vector, create a new one and add it to the list.
-      newPipelineHandle = this->compilePipeline(state, renderPass, m_basePipeline);
+      Rc<DxvkGraphicsPipelineInstance> newPipeline =
+        new DxvkGraphicsPipelineInstance(m_pipeMgr->m_device->vkd(),
+          state, renderPassHandle, VK_NULL_HANDLE);
+
+      if (async && m_pipeMgr->m_compiler != nullptr)
+        m_pipeMgr->m_compiler->queueCompilation(this, newPipeline, renderPass);
+      else
+        newPipelineHandle = this->compileInstance(newPipeline, renderPass);
 
       // Add new pipeline to the set
-      m_pipelines.emplace_back(state, renderPassHandle, newPipelineHandle);
+      m_pipelines.push_back(newPipeline);
       m_pipeMgr->m_numGraphicsPipelines += 1;
-      
-      if (!m_basePipeline && newPipelineHandle)
-        m_basePipeline = newPipelineHandle;
     }
     
-    if (newPipelineHandle != VK_NULL_HANDLE)
-      this->writePipelineStateToCache(state, renderPass.format());
-    
     return newPipelineHandle;
   }
   
   
-  const DxvkGraphicsPipelineInstance* DxvkGraphicsPipeline::findInstance(
+  VkPipeline DxvkGraphicsPipeline::compileInstance(
+    const Rc<DxvkGraphicsPipelineInstance>& instance,
+    const Rc<DxvkRenderPass>&               renderPass) {
+    VkPipeline newPipelineHandle = this->compilePipeline(
+      instance->m_stateVector, *renderPass, m_basePipeline);
+
+    if (!instance->setPipeline(newPipelineHandle)) {
+      m_vkd->vkDestroyPipeline(m_vkd->device(), newPipelineHandle, nullptr);
+    } else {
+      if (!m_basePipeline && newPipelineHandle)
+        m_basePipeline = newPipelineHandle;
+      if (newPipelineHandle != VK_NULL_HANDLE)
+        this->writePipelineStateToCache(instance->m_stateVector, renderPass->format());
+    }
+
+    return newPipelineHandle;
+  }
+  
+  
+  DxvkGraphicsPipelineInstance* DxvkGraphicsPipeline::findInstance(
     const DxvkGraphicsPipelineStateInfo& state,
           VkRenderPass                   renderPass) const {
     for (const auto& instance : m_pipelines) {
-      if (instance.isCompatible(state, renderPass))
-        return &instance;
+      if (instance->isCompatible(state, renderPass))
+        return instance.ptr();
     }
     
     return nullptr;
diff --git a/src/dxvk/dxvk_graphics.h b/src/dxvk/dxvk_graphics.h
index e5686950..8811f10c 100644
--- a/src/dxvk/dxvk_graphics.h
+++ b/src/dxvk/dxvk_graphics.h
@@ -132,19 +132,25 @@ namespace dxvk {
    * Stores a state vector and the
    * corresponding pipeline handle.
    */
-  class DxvkGraphicsPipelineInstance {
-
+  class DxvkGraphicsPipelineInstance: public RcObject {
+    friend class DxvkGraphicsPipeline;
   public:
 
     DxvkGraphicsPipelineInstance() { }
     DxvkGraphicsPipelineInstance(
+      const Rc<vk::DeviceFn>&               vkd,
       const DxvkGraphicsPipelineStateInfo&  state,
             VkRenderPass                    rp,
             VkPipeline                      pipe)
-    : m_stateVector (state),
+    : m_vkd         (vkd),
+      m_stateVector (state),
       m_renderPass  (rp),
       m_pipeline    (pipe) { }
 
+    ~DxvkGraphicsPipelineInstance() {
+      m_vkd->vkDestroyPipeline(m_vkd->device(), m_pipeline, nullptr);
+    }
+
     /**
      * \brief Checks for matching pipeline state
      * 
@@ -159,23 +165,38 @@ namespace dxvk {
           && m_renderPass  == rp;
     }
 
+    /**
+     * \brief Sets the pipeline handle
+     *
+     * If a pipeline handle has already been
+     * set up, this method will fail and the new pipeline
+     * handle should be destroyed.
+     * \param [in] pipeline The pipeline
+     */
+    bool setPipeline(VkPipeline pipeline) {
+      VkPipeline expected = VK_NULL_HANDLE;
+      return m_pipeline.compare_exchange_strong(expected, pipeline);
+    }
+
     /**
      * \brief Retrieves pipeline
      * \returns The pipeline handle
      */
     VkPipeline pipeline() const {
-      return m_pipeline;
+      return m_pipeline.load();
     }
 
   private:
 
+    const Rc<vk::DeviceFn>        m_vkd;
     DxvkGraphicsPipelineStateInfo m_stateVector;
     VkRenderPass                  m_renderPass;
-    VkPipeline                    m_pipeline;
+    std::atomic<VkPipeline>       m_pipeline;
 
   };
 
   
+
   /**
    * \brief Graphics pipeline
    * 
@@ -234,20 +255,20 @@ namespace dxvk {
      * state. If necessary, a new pipeline will be created.
      * \param [in] state Pipeline state vector
      * \param [in] renderPass The render pass
+     * \param [in] async Compile asynchronously
      * \returns Pipeline handle
      */
     VkPipeline getPipelineHandle(
       const DxvkGraphicsPipelineStateInfo&    state,
-      const DxvkRenderPass&                   renderPass);
+      const Rc<DxvkRenderPass>&               renderPass,
+            bool                              async);
+    
+    VkPipeline compileInstance(
+      const Rc<DxvkGraphicsPipelineInstance>& instance,
+      const Rc<DxvkRenderPass>&               renderPass);
     
   private:
     
-    struct PipelineStruct {
-      DxvkGraphicsPipelineStateInfo stateVector;
-      VkRenderPass                  renderPass;
-      VkPipeline                    pipeline;
-    };
-    
     Rc<vk::DeviceFn>          m_vkd;
     DxvkPipelineManager*      m_pipeMgr;
 
@@ -267,13 +288,13 @@ namespace dxvk {
     DxvkGraphicsCommonPipelineStateInfo m_common;
     
     // List of pipeline instances, shared between threads
-    alignas(CACHE_LINE_SIZE) sync::Spinlock   m_mutex;
-    std::vector<DxvkGraphicsPipelineInstance> m_pipelines;
+    alignas(CACHE_LINE_SIZE) sync::Spinlock       m_mutex;
+    std::vector<Rc<DxvkGraphicsPipelineInstance>> m_pipelines;
     
     // Pipeline handles used for derivative pipelines
     VkPipeline m_basePipeline = VK_NULL_HANDLE;
     
-    const DxvkGraphicsPipelineInstance* findInstance(
+    DxvkGraphicsPipelineInstance* findInstance(
       const DxvkGraphicsPipelineStateInfo& state,
             VkRenderPass                   renderPass) const;
     
@@ -302,4 +323,4 @@ namespace dxvk {
     
   };
   
-}
\ No newline at end of file
+}
diff --git a/src/dxvk/dxvk_image.h b/src/dxvk/dxvk_image.h
index f55ae42d..1bd638f0 100644
--- a/src/dxvk/dxvk_image.h
+++ b/src/dxvk/dxvk_image.h
@@ -435,6 +435,37 @@ namespace dxvk {
       return result;
     }
 
+    /**
+     * \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;
@@ -443,6 +474,9 @@ 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_pipecompiler.cpp b/src/dxvk/dxvk_pipecompiler.cpp
new file mode 100644
index 00000000..a70c9908
--- /dev/null
+++ b/src/dxvk/dxvk_pipecompiler.cpp
@@ -0,0 +1,69 @@
+#include "dxvk_graphics.h"
+#include "dxvk_pipecompiler.h"
+
+namespace dxvk {
+
+  DxvkPipelineCompiler::DxvkPipelineCompiler() {
+    uint32_t sysCpuCount = dxvk::thread::hardware_concurrency();
+    uint32_t threadCount = sysCpuCount > 2 ? sysCpuCount - 2 : 1;
+
+    Logger::info(str::format(
+      "DxvkPipelineCompiler: Using ",
+      threadCount, " workers"));
+
+    // Start the compiler threads
+    m_compilerThreads.resize(threadCount);
+
+    for (uint32_t i = 0; i < threadCount; i++) {
+      m_compilerThreads.at(i) = dxvk::thread(
+        [this] { this->runCompilerThread(); });
+    }
+  }
+
+
+  DxvkPipelineCompiler::~DxvkPipelineCompiler() {
+    { std::unique_lock<std::mutex> lock(m_compilerLock);
+      m_compilerStop.store(true);
+    }
+
+    m_compilerCond.notify_all();
+    for (auto& thread : m_compilerThreads)
+      thread.join();
+  }
+
+
+  void DxvkPipelineCompiler::queueCompilation(
+    const Rc<DxvkGraphicsPipeline>&         pipeline,
+    const Rc<DxvkGraphicsPipelineInstance>& instance,
+    const Rc<DxvkRenderPass>&               renderPass) {
+    std::unique_lock<std::mutex> lock(m_compilerLock);
+    m_compilerQueue.push({ pipeline, instance, renderPass });
+    m_compilerCond.notify_one();
+  }
+
+
+  void DxvkPipelineCompiler::runCompilerThread() {
+    env::setThreadName("dxvk-pcompiler");
+
+    while (!m_compilerStop.load()) {
+      PipelineEntry entry;
+
+      { std::unique_lock<std::mutex> lock(m_compilerLock);
+
+        m_compilerCond.wait(lock, [this] {
+          return m_compilerStop.load()
+              || m_compilerQueue.size() != 0;
+        });
+
+        if (m_compilerQueue.size() != 0) {
+          entry = std::move(m_compilerQueue.front());
+          m_compilerQueue.pop();
+        }
+      }
+
+      if (entry.pipeline != nullptr && entry.instance != nullptr)
+        entry.pipeline->compileInstance(entry.instance, entry.renderPass);
+    }
+  }
+
+}
diff --git a/src/dxvk/dxvk_pipecompiler.h b/src/dxvk/dxvk_pipecompiler.h
new file mode 100644
index 00000000..26e983cc
--- /dev/null
+++ b/src/dxvk/dxvk_pipecompiler.h
@@ -0,0 +1,60 @@
+#pragma once
+
+#include <atomic>
+#include <condition_variable>
+#include <mutex>
+#include <queue>
+
+#include "../util/thread.h"
+#include "dxvk_include.h"
+
+namespace dxvk {
+
+  class DxvkGraphicsPipeline;
+  class DxvkGraphicsPipelineInstance;
+
+  /**
+   * \brief Pipeline compiler
+   *
+   * Asynchronous pipeline compiler
+   */
+  class DxvkPipelineCompiler : public RcObject {
+
+  public:
+
+    DxvkPipelineCompiler();
+    ~DxvkPipelineCompiler();
+
+    /**
+     * \brief Compiles a pipeline asynchronously
+     *
+     * This should be used to compile graphics
+     * pipeline instances asynchronously.
+     * \param [in] pipeline The pipeline object
+     * \param [in] instance The pipeline instance
+     * \param [in] renderPass
+     */
+    void queueCompilation(
+      const Rc<DxvkGraphicsPipeline>&         pipeline,
+      const Rc<DxvkGraphicsPipelineInstance>& instance,
+      const Rc<DxvkRenderPass>&               renderPass);
+
+  private:
+
+    struct PipelineEntry {
+      Rc<DxvkGraphicsPipeline>         pipeline;
+      Rc<DxvkGraphicsPipelineInstance> instance;
+      Rc<DxvkRenderPass>               renderPass;
+    };
+
+    std::atomic<bool>           m_compilerStop = { false };
+    std::mutex                  m_compilerLock;
+    std::condition_variable     m_compilerCond;
+    std::queue<PipelineEntry>   m_compilerQueue;
+    std::vector<dxvk::thread>   m_compilerThreads;
+
+    void runCompilerThread();
+
+  };
+
+}
diff --git a/src/dxvk/dxvk_pipemanager.cpp b/src/dxvk/dxvk_pipemanager.cpp
index 682db83d..2783c847 100644
--- a/src/dxvk/dxvk_pipemanager.cpp
+++ b/src/dxvk/dxvk_pipemanager.cpp
@@ -45,9 +45,13 @@ namespace dxvk {
   : m_device    (device),
     m_cache     (new DxvkPipelineCache(device->vkd())) {
     std::string useStateCache = env::getEnvVar("DXVK_STATE_CACHE");
+    std::string useAsync      = env::getEnvVar("DXVK_ASYNC");
     
     if (useStateCache != "0" && device->config().enableStateCache)
       m_stateCache = new DxvkStateCache(device, this, passManager);
+
+    if (useAsync != "0" || device->config().useAsync)
+      m_compiler = new DxvkPipelineCompiler();
   }
   
   
diff --git a/src/dxvk/dxvk_pipemanager.h b/src/dxvk/dxvk_pipemanager.h
index fb62cb9a..c3b846cf 100644
--- a/src/dxvk/dxvk_pipemanager.h
+++ b/src/dxvk/dxvk_pipemanager.h
@@ -5,6 +5,7 @@
 
 #include "dxvk_compute.h"
 #include "dxvk_graphics.h"
+#include "dxvk_pipecompiler.h"
 
 namespace dxvk {
 
@@ -139,6 +140,7 @@ namespace dxvk {
     const DxvkDevice*         m_device;
     Rc<DxvkPipelineCache>     m_cache;
     Rc<DxvkStateCache>        m_stateCache;
+    Rc<DxvkPipelineCompiler>  m_compiler;
 
     std::atomic<uint32_t>     m_numComputePipelines  = { 0 };
     std::atomic<uint32_t>     m_numGraphicsPipelines = { 0 };
diff --git a/src/dxvk/dxvk_state_cache.cpp b/src/dxvk/dxvk_state_cache.cpp
index a9249a3a..4d296c19 100644
--- a/src/dxvk/dxvk_state_cache.cpp
+++ b/src/dxvk/dxvk_state_cache.cpp
@@ -263,7 +263,7 @@ namespace dxvk {
         const auto& entry = m_entries[e->second];
 
         auto rp = m_passManager->getRenderPass(entry.format);
-        pipeline->getPipelineHandle(entry.gpState, *rp);
+        pipeline->getPipelineHandle(entry.gpState, rp, false);
       }
     } else {
       auto pipeline = m_pipeManager->createComputePipeline(item.cs);
diff --git a/src/dxvk/meson.build b/src/dxvk/meson.build
index 64f20998..bd11e390 100644
--- a/src/dxvk/meson.build
+++ b/src/dxvk/meson.build
@@ -78,6 +78,7 @@ dxvk_src = files([
   'dxvk_openvr.cpp',
   'dxvk_options.cpp',
   'dxvk_pipecache.cpp',
+  'dxvk_pipecompiler.cpp',
   'dxvk_pipelayout.cpp',
   'dxvk_pipemanager.cpp',
   'dxvk_queue.cpp',

diff --git a/src/dxvk/dxvk_options.cpp b/src/dxvk/dxvk_options.cpp
index e1da8a48..9d6b4ee9 100644
--- a/src/dxvk/dxvk_options.cpp
+++ b/src/dxvk/dxvk_options.cpp
@@ -4,6 +4,7 @@ namespace dxvk {
 
   DxvkOptions::DxvkOptions(const Config& config) {
     enableStateCache      = config.getOption<bool>    ("dxvk.enableStateCache",       true);
+    useAsync              = config.getOption<bool>    ("dxvk.useAsync",               false);
     numCompilerThreads    = config.getOption<int32_t> ("dxvk.numCompilerThreads",     0);
     useRawSsbo            = config.getOption<Tristate>("dxvk.useRawSsbo",             Tristate::Auto);
     useEarlyDiscard       = config.getOption<Tristate>("dxvk.useEarlyDiscard",        Tristate::Auto);
diff --git a/src/dxvk/dxvk_options.h b/src/dxvk/dxvk_options.h
index 79259875..db2f8a32 100644
--- a/src/dxvk/dxvk_options.h
+++ b/src/dxvk/dxvk_options.h
@@ -10,6 +10,7 @@ namespace dxvk {
 
     /// Enable state cache
     bool enableStateCache;
+    bool useAsync;
 
     /// Number of compiler threads
     /// when using the state cache
diff --git a/dxvk.conf b/dxvk.conf
index 25d27cb9..4e409075 100644
--- a/dxvk.conf
+++ b/dxvk.conf
@@ -159,6 +159,15 @@
 # dxvk.numCompilerThreads = 0
 
 
+# Compile pipelines asynchronously if possible. This may reduce stuttering
+# in some games, but may also introduce rendering issues that might become
+# apparent over time. Do not report bugs with this option enabled.
+#
+# Supported values: True, False
+
+# dxvk.useAsync = False
+
+
 # Toggles raw SSBO usage.
 # 
 # Uses storage buffers to implement raw and structured buffer