Vulkan: Optimize glMemoryBarrier

src/libANGLE/renderer/vulkan/ContextVk.h

@@ -594,6 +594,8 @@ class ContextVk : public ContextImpl, public vk::Context, public MultisampleText
     // Dirty bits.
     enum DirtyBitType : size_t
     {
+        // A glMemoryBarrier has been called and command buffers may need flushing.
+        DIRTY_BIT_MEMORY_BARRIER,
         // Dirty bits that must be processed before the render pass is started.  The handlers for
         // these dirty bits don't record any commands.
         DIRTY_BIT_DEFAULT_ATTRIBS,
@@ -781,6 +783,8 @@ class ContextVk : public ContextImpl, public vk::Context, public MultisampleText
     void invalidateDriverUniforms();
 
     // Handlers for graphics pipeline dirty bits.
+    angle::Result handleDirtyGraphicsMemorybarrier(DirtyBits::Iterator *dirtyBitsIterator,
+                                                   DirtyBits dirtyBitMask);
     angle::Result handleDirtyGraphicsEventLog(DirtyBits::Iterator *dirtyBitsIterator,
                                               DirtyBits dirtyBitMask);
     angle::Result handleDirtyGraphicsDefaultAttribs(DirtyBits::Iterator *dirtyBitsIterator,
@@ -817,6 +821,7 @@ class ContextVk : public ContextImpl, public vk::Context, public MultisampleText
                                                     DirtyBits dirtyBitMask);
 
     // Handlers for compute pipeline dirty bits.
+    angle::Result handleDirtyComputeMemoryBarrier();
     angle::Result handleDirtyComputeEventLog();
     angle::Result handleDirtyComputePipelineDesc();
     angle::Result handleDirtyComputePipelineBinding();
@@ -827,6 +832,8 @@ class ContextVk : public ContextImpl, public vk::Context, public MultisampleText
     angle::Result handleDirtyComputeDescriptorSets();
 
     // Common parts of the common dirty bit handlers.
+    angle::Result handleDirtyMemorybarrierImpl(DirtyBits::Iterator *dirtyBitsIterator,
+                                               DirtyBits dirtyBitMask);
     angle::Result handleDirtyEventLogImpl(vk::CommandBuffer *commandBuffer);
     angle::Result handleDirtyTexturesImpl(vk::CommandBufferHelper *commandBufferHelper);
     angle::Result handleDirtyShaderResourcesImpl(vk::CommandBufferHelper *commandBufferHelper);
@@ -890,6 +897,7 @@ class ContextVk : public ContextImpl, public vk::Context, public MultisampleText
 
     angle::Result onResourceAccess(const vk::CommandBufferAccess &access);
     angle::Result flushCommandBuffersIfNecessary(const vk::CommandBufferAccess &access);
+    bool renderPassUsesStorageResources() const;
 
     void outputCumulativePerfCounters();
 

src/libANGLE/renderer/vulkan/vk_helpers.cpp

@@ -793,6 +793,8 @@ CommandBufferHelper::CommandBufferHelper()
       mIsTransformFeedbackActiveUnpaused(false),
       mIsRenderPassCommandBuffer(false),
       mReadOnlyDepthStencilMode(false),
+      mHasShaderStorageOutput(false),
+      mHasGLMemoryBarrierIssued(false),
       mDepthAccess(ResourceAccess::Unused),
       mStencilAccess(ResourceAccess::Unused),
       mDepthCmdSizeInvalidated(kInfiniteCmdSize),
@@ -1553,6 +1555,8 @@ void CommandBufferHelper::reset()
         mRenderPassStarted                 = false;
         mValidTransformFeedbackBufferCount = 0;
         mRebindTransformFeedbackBuffers    = false;
+        mHasShaderStorageOutput            = false;
+        mHasGLMemoryBarrierIssued          = false;
         mDepthAccess                       = ResourceAccess::Unused;
         mStencilAccess                     = ResourceAccess::Unused;
         mDepthCmdSizeInvalidated           = kInfiniteCmdSize;

src/libANGLE/renderer/vulkan/vk_helpers.h

@@ -1170,6 +1170,18 @@ class CommandBufferHelper : angle::NonCopyable
 
     bool hasRenderPass() const { return mIsRenderPassCommandBuffer; }
 
+    void setHasShaderStorageOutput() { mHasShaderStorageOutput = true; }
+    bool hasShaderStorageOutput() const { return mHasShaderStorageOutput; }
+
+    void setGLMemoryBarrierIssued()
+    {
+        if (!empty())
+        {
+            mHasGLMemoryBarrierIssued = true;
+        }
+    }
+    bool hasGLMemoryBarrierIssued() const { return mHasGLMemoryBarrierIssued; }
+
   private:
     bool onDepthStencilAccess(ResourceAccess access,
                               uint32_t *cmdCountInvalidated,
@@ -1207,6 +1219,15 @@ class CommandBufferHelper : angle::NonCopyable
     bool mIsRenderPassCommandBuffer;
     bool mReadOnlyDepthStencilMode;
 
+    // Whether the command buffers contains any draw/dispatch calls that possibly output data
+    // through storage buffers and images.  This is used to determine whether glMemoryBarrier*
+    // should flush the command buffer.
+    bool mHasShaderStorageOutput;
+    // Whether glMemoryBarrier has been called while commands are recorded in this command buffer.
+    // This is used to know when to check and potentially flush the command buffer if storage
+    // buffers and images are used in it.
+    bool mHasGLMemoryBarrierIssued;
+
     // State tracking for the maximum (Write been the highest) depth access during the entire
     // renderpass. Note that this does not include VK_ATTACHMENT_LOAD_OP_CLEAR which is tracked
     // separately. This is done this way to allow clear op to being optimized out when we find out