Turn off preferSubmitAtFBOBoundary and add early submit logic

Disabling preferSubmitAtFBOBoundary for MALI GPUs and submit pending
commands at the draw call time if the number of write-commands in the
current render pass reaches a threshold to avoid delaying
the submission too much.

Bug: angleproject:447444701
Change-Id: I413a4beb7147afd4f064286433a76014b19a49c6
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/6994438
Reviewed-by: Charlie Lao <[email protected]>
Reviewed-by: Amirali Abdolrashidi <[email protected]>
Auto-Submit: Sean Feng <[email protected]>
Commit-Queue: Shahbaz Youssefi <[email protected]>

Author: Sean Feng

src/libANGLE/renderer/vulkan/ContextVk.cpp

@@ -1579,6 +1579,16 @@ angle::Result ContextVk::setupDraw(const gl::Context *context,
         mGraphicsPipelineDesc->updateTopology(&mGraphicsPipelineTransition, mCurrentDrawMode);
     }
 
+    // Submit pending commands if the number of write-commands in the current render pass reaches a
+    // threshold to avoid delaying the submission too much.
+    if (ANGLE_UNLIKELY(mRenderPassCommands->getCommandBuffer().getRenderPassWriteCommandCount() >
+                       mRenderer->getMinRenderPassWriteCommandCountToEarlySubmit()) &&
+        (mCommandsPendingSubmissionCount > 0))
+    {
+        ANGLE_TRY(submitCommands(nullptr, nullptr));
+        mCommandsPendingSubmissionCount = 0;
+    }
+
     // Must be called before the command buffer is started. Can call finish.
     VertexArrayVk *vertexArrayVk = getVertexArray();
     if (vertexArrayVk->getStreamingVertexAttribsMask().any())

src/libANGLE/renderer/vulkan/vk_renderer.cpp

@@ -1950,7 +1950,8 @@ Renderer::Renderer()
       mNativeVectorWidthHalf(0),
       mPreferredVectorWidthDouble(0),
       mPreferredVectorWidthHalf(0),
-      mMinCommandCountToSubmit(0)
+      mMinCommandCountToSubmit(0),
+      mMinRPWriteCommandCountToEarlySubmit(UINT32_MAX)
 {
     VkFormatProperties invalid = {0, 0, kInvalidFormatFeatureFlags};
     mFormatProperties.fill(invalid);
@@ -5709,10 +5710,18 @@ void Renderer::initFeatures(const vk::ExtensionNameList &deviceExtensionNames,
     // improves performance. Most app traces shows frame time reduced and manhattan 3.1 offscreen
     // score improves 7%.
     ANGLE_FEATURE_CONDITION(&mFeatures, preferSubmitAtFBOBoundary,
-                            isTileBasedRenderer || isSwiftShader);
+                            ((isTileBasedRenderer || isSwiftShader) && !isARMProprietary));
+
     ANGLE_FEATURE_CONDITION(&mFeatures, forceSubmitExceptionsAtFBOBoundary, !isQualcommProprietary);
     mMinCommandCountToSubmit = isQualcommProprietary ? 1024 : 32;
 
+    // The number of minimum write commands in the command buffer to trigger one submission of
+    // pending commands at draw call time
+    if (isARMProprietary)
+    {
+        mMinRPWriteCommandCountToEarlySubmit = 128;
+    }
+
     // In order to support immutable samplers tied to external formats, we need to overallocate
     // descriptor counts for such immutable samplers
     ANGLE_FEATURE_CONDITION(&mFeatures, useMultipleDescriptorsForExternalFormats, true);

src/libANGLE/renderer/vulkan/vk_renderer.h

@@ -727,6 +727,11 @@ class Renderer : angle::NonCopyable
 
     angle::Result onFrameBoundary(const gl::Context *contextGL);
 
+    uint32_t getMinRenderPassWriteCommandCountToEarlySubmit() const
+    {
+        return mMinRPWriteCommandCountToEarlySubmit;
+    }
+
   private:
     angle::Result setupDevice(vk::ErrorContext *context,
                               const angle::FeatureOverrides &featureOverrides,
@@ -1130,6 +1135,10 @@ class Renderer : angle::NonCopyable
     // The number of minimum commands in the command buffer to prefer submit at FBO boundary or
     // immediately submit when the device is idle after calling to flush.
     uint32_t mMinCommandCountToSubmit;
+
+    // The number of minimum write commands in the command buffer to trigger one submission of
+    // pending commands at draw call time
+    uint32_t mMinRPWriteCommandCountToEarlySubmit;
 };
 
 ANGLE_INLINE Serial Renderer::generateQueueSerial(SerialIndex index)