[dxvk] Implement lifetime tracking for graphics pipelines

src/dxvk/dxvk_cmdlist.cpp

@@ -219,11 +219,18 @@ namespace dxvk {
     m_signalTracker.reset();
     m_statCounters.reset();
 
+    // Recycle descriptor pools
     for (const auto& descriptorPools : m_descriptorPools)
       descriptorPools.second->recycleDescriptorPool(descriptorPools.first);
 
     m_descriptorPools.clear();
 
+    // Release pipelines
+    for (auto pipeline : m_pipelines)
+      pipeline->releasePipeline();
+
+    m_pipelines.clear();
+
     m_waitSemaphores.clear();
     m_signalSemaphores.clear();
   }

src/dxvk/dxvk_cmdlist.h

@@ -8,6 +8,7 @@
 #include "dxvk_fence.h"
 #include "dxvk_gpu_event.h"
 #include "dxvk_gpu_query.h"
+#include "dxvk_graphics.h"
 #include "dxvk_lifetime.h"
 #include "dxvk_limits.h"
 #include "dxvk_pipelayout.h"
@@ -30,7 +31,7 @@ namespace dxvk {
   };
   
   using DxvkCmdBufferFlags = Flags<DxvkCmdBuffer>;
-  
+
   /**
    * \brief Queue submission info
    *
@@ -176,6 +177,15 @@ namespace dxvk {
       m_gpuQueryTracker.trackQuery(handle);
     }
     
+    /**
+     * \brief Tracks a graphics pipeline
+     * \param [in] pipeline Pipeline
+     */
+    void trackGraphicsPipeline(DxvkGraphicsPipeline* pipeline) {
+      pipeline->acquirePipeline();
+      m_pipelines.push_back(pipeline);
+    }
+
     /**
      * \brief Queues signal
      * 
@@ -822,6 +832,8 @@ namespace dxvk {
       Rc<DxvkDescriptorPool>,
       Rc<DxvkDescriptorManager>>> m_descriptorPools;
 
+    std::vector<DxvkGraphicsPipeline*> m_pipelines;
+
     VkCommandBuffer getCmdBuffer(DxvkCmdBuffer cmdBuffer) const {
       if (cmdBuffer == DxvkCmdBuffer::ExecBuffer) return m_execBuffer;
       if (cmdBuffer == DxvkCmdBuffer::InitBuffer) return m_initBuffer;

src/dxvk/dxvk_context.cpp

@@ -49,6 +49,11 @@ namespace dxvk {
 
     if (m_device->features().extTransformFeedback.transformFeedback)
       m_globalRwGraphicsBarrier.access |= VK_ACCESS_TRANSFORM_FEEDBACK_COUNTER_READ_BIT_EXT;
+
+    // Store the lifetime tracking bit as a context feature so
+    // that we don't have to scan device features at draw time
+    if (m_device->mustTrackPipelineLifetime())
+      m_features.set(DxvkContextFeature::TrackGraphicsPipeline);
   }
   
   
@@ -4484,6 +4489,9 @@ namespace dxvk {
       return false;
     }
 
+    if (m_features.test(DxvkContextFeature::TrackGraphicsPipeline))
+      m_cmd->trackGraphicsPipeline(newPipeline);
+
     if (unlikely(newPipeline->getSpecConstantMask() != m_state.gp.constants.mask))
       this->resetSpecConstants<VK_PIPELINE_BIND_POINT_GRAPHICS>(newPipeline->getSpecConstantMask());
 

src/dxvk/dxvk_context_state.h

@@ -60,7 +60,8 @@ namespace dxvk {
   /**
    * \brief Context feature bits
    */
-  enum class DxvkContextFeature {
+  enum class DxvkContextFeature : uint32_t {
+    TrackGraphicsPipeline,
     FeatureCount
   };
 

src/dxvk/dxvk_graphics.cpp

@@ -880,15 +880,8 @@ namespace dxvk {
   
   
   DxvkGraphicsPipeline::~DxvkGraphicsPipeline() {
-    for (const auto& instance : m_fastPipelines) {
+    this->destroyBasePipelines();
-      this->destroyPipeline(instance.second);
+    this->destroyOptimizedPipelines();
-
-      m_vsLibrary->releasePipelineHandle();
-      m_fsLibrary->releasePipelineHandle();
-    }
-
-    for (const auto& instance : m_basePipelines)
-      this->destroyPipeline(instance.second);
   }
   
   
@@ -915,7 +908,7 @@ namespace dxvk {
         return std::make_pair(VK_NULL_HANDLE, DxvkGraphicsPipelineType::FastPipeline);
 
       // Prevent other threads from adding new instances and check again
-      std::lock_guard<dxvk::mutex> lock(m_mutex);
+      std::unique_lock<dxvk::mutex> lock(m_mutex);
       instance = this->findInstance(state);
 
       if (!instance) {
@@ -924,6 +917,10 @@ namespace dxvk {
         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);
@@ -965,7 +962,7 @@ namespace dxvk {
         return;
 
       // Prevent other threads from adding new instances and check again
-      std::lock_guard<dxvk::mutex> lock(m_mutex);
+      std::unique_lock<dxvk::mutex> lock(m_mutex);
       instance = this->findInstance(state);
 
       if (!instance)
@@ -987,6 +984,45 @@ namespace dxvk {
   }
 
 
+  void DxvkGraphicsPipeline::acquirePipeline() {
+    if (!m_device->mustTrackPipelineLifetime())
+      return;
+
+    // We need to lock here to make sure that any ongoing pipeline
+    // destruction finishes before the calling thread can access the
+    // pipeline, and that no pipelines get destroyed afterwards.
+    std::unique_lock<dxvk::mutex> lock(m_mutex);
+    m_useCount += 1;
+  }
+
+
+  void DxvkGraphicsPipeline::releasePipeline() {
+    if (!m_device->mustTrackPipelineLifetime())
+      return;
+
+    std::unique_lock<dxvk::mutex> lock(m_mutex);
+
+    if (!(--m_useCount)) {
+      // Don't destroy base pipelines if that's all we're going to
+      // use, since that would pretty much ruin the experience.
+      if (m_device->config().enableGraphicsPipelineLibrary == Tristate::True)
+        return;
+
+      // Exit early if there's nothing to do
+      if (m_basePipelines.empty())
+        return;
+
+      // Remove any base pipeline references, but
+      // keep the optimized pipelines around.
+      for (auto& entry : m_pipelines)
+        entry.baseHandle.store(VK_NULL_HANDLE);
+
+      // Destroy the actual Vulkan pipelines
+      this->destroyBasePipelines();
+    }
+  }
+
+
   DxvkGraphicsPipelineInstance* DxvkGraphicsPipeline::createInstance(
     const DxvkGraphicsPipelineStateInfo& state,
           bool                           doCreateBasePipeline) {
@@ -1220,7 +1256,27 @@ namespace dxvk {
   }
   
   
-  void DxvkGraphicsPipeline::destroyPipeline(VkPipeline pipeline) const {
+  void DxvkGraphicsPipeline::destroyBasePipelines() {
+    for (const auto& instance : m_basePipelines) {
+      this->destroyVulkanPipeline(instance.second);
+
+      m_vsLibrary->releasePipelineHandle();
+      m_fsLibrary->releasePipelineHandle();
+    }
+
+    m_basePipelines.clear();
+  }
+
+
+  void DxvkGraphicsPipeline::destroyOptimizedPipelines() {
+    for (const auto& instance : m_fastPipelines)
+      this->destroyVulkanPipeline(instance.second);
+
+    m_fastPipelines.clear();
+  }
+
+
+  void DxvkGraphicsPipeline::destroyVulkanPipeline(VkPipeline pipeline) const {
     auto vk = m_device->vkd();
 
     vk->vkDestroyPipeline(vk->device(), pipeline, nullptr);

src/dxvk/dxvk_graphics.h

@@ -535,7 +535,24 @@ namespace dxvk {
      */
     void compilePipeline(
       const DxvkGraphicsPipelineStateInfo&    state);
-    
+
+    /**
+     * \brief Acquires the pipeline
+     *
+     * Increments the use count by one and prevents any Vulkan
+     * pipelines from being destroyed. Must be called before
+     * \ref getPipelineHandle or \ref compilePipeline.
+     */
+    void acquirePipeline();
+
+    /**
+     * \brief Releases the pipeline
+     *
+     * Decrements the use count by one. If the counter reaches
+     * zero, any Vulkan pipeline objects may be destroyed.
+     */
+    void releasePipeline();
+
   private:
 
     DxvkDevice*                 m_device;    
@@ -560,11 +577,12 @@ namespace dxvk {
     alignas(CACHE_LINE_SIZE)
     dxvk::mutex                                   m_mutex;
     sync::List<DxvkGraphicsPipelineInstance>      m_pipelines;
+    uint32_t                                      m_useCount = 0;
 
     std::unordered_map<
       DxvkGraphicsPipelineBaseInstanceKey,
       VkPipeline, DxvkHash, DxvkEq>               m_basePipelines;
-    
+
     alignas(CACHE_LINE_SIZE)
     dxvk::mutex                                   m_fastMutex;
     std::unordered_map<
@@ -595,7 +613,11 @@ namespace dxvk {
       const DxvkGraphicsPipelineFastInstanceKey& key,
             VkPipelineCreateFlags          flags) const;
 
-    void destroyPipeline(
+    void destroyBasePipelines();
+
+    void destroyOptimizedPipelines();
+
+    void destroyVulkanPipeline(
             VkPipeline                     pipeline) const;
     
     SpirvCodeBuffer getShaderCode(

src/dxvk/dxvk_pipemanager.cpp

@@ -64,6 +64,7 @@ namespace dxvk {
     std::unique_lock lock(m_queueLock);
     this->startWorkers();
 
+    pipeline->acquirePipeline();
     m_pendingTasks += 1;
 
     PipelineEntry e = { };
@@ -148,10 +149,12 @@ namespace dxvk {
       }
 
       if (p) {
-        if (p->computePipeline)
+        if (p->computePipeline) {
           p->computePipeline->compilePipeline(p->computeState);
-        else if (p->graphicsPipeline)
+        } else if (p->graphicsPipeline) {
           p->graphicsPipeline->compilePipeline(p->graphicsState);
+          p->graphicsPipeline->releasePipeline();
+        }
 
         m_pendingTasks -= 1;
       }