Modify ImageTransportSurfaceOverlayMac to allow pipelining

gpu/ipc/service/image_transport_surface_overlay_mac.mm

 // Copyright 2015 The Chromium Authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
 #include "gpu/ipc/service/image_transport_surface_overlay_mac.h"
 
 #include <CoreGraphics/CoreGraphics.h>
 #include <IOSurface/IOSurface.h>
 #include <OpenGL/CGLRenderers.h>
 #include <OpenGL/CGLTypes.h>
@@ skipping 139 matching lines @@
   params.scale_factor = scale_factor;
   params.latency_info = std::move(latency_info);
   params.result = gfx::SwapResult::SWAP_ACK;
   stub_->SendSwapBuffersCompleted(params);
 }
 
 gfx::SwapResult ImageTransportSurfaceOverlayMac::SwapBuffersInternal(
     const gfx::Rect& pixel_damage_rect) {
   TRACE_EVENT0("gpu", "ImageTransportSurfaceOverlayMac::SwapBuffersInternal");
 
-  // A glFlush is necessary to ensure correct content appears. A glFinish
-  // appears empirically to be the best way to get maximum performance when
-  // GPU bound.
-  {
+  // If supported, use GLFence to ensure that we haven't gotten more than one
+  // frame ahead of GL.
+  if (gl::GLFence::IsSupported()) {
+    // If we have gotten more than one frame ahead of GL, wait for the previous
+    // frame to complete.
+    if (previous_frame_fence_) {
+      TRACE_EVENT0("gpu", "ImageTransportSurfaceOverlayMac::ClientWait");
+      previous_frame_fence_->ClientWait();

[ccameron, 2016/06/01 21:26:37]

This will cause the CPU to spin up to 100% on Mac. See:
https://bugs.chromium.org/p/chromium/issues/detail?id=611805#c1

Maybe we can just do a
  whilee (!passed)
     Sleep(1msec)

1msec is a long time, but this only happens when we have >16msec of GPU work, so
it shouldn't hurt too much.

+      previous_frame_fence_ = nullptr;
+    }
+
+    // A glFlush is necessary to ensure correct content appears.
+    {
+      gl::ScopedSetGLToRealGLApi scoped_set_gl_api;
+      CheckGLErrors("Before flush");
+      glFlush();
+      CheckGLErrors("After flush");
+    }
+
+    // Create a new fence after flushing work from the current frame.
+    previous_frame_fence_.reset(gl::GLFence::Create());
+  } else {
+    // GLFence isn't supported - issue a glFinish on each frame to ensure
+    // there is backpressure from GL.
     gl::ScopedSetGLToRealGLApi scoped_set_gl_api;
     TRACE_EVENT0("gpu", "ImageTransportSurfaceOverlayMac::glFinish");
     CheckGLErrors("Before finish");
     glFinish();
     CheckGLErrors("After finish");
   }
 
   base::TimeTicks finish_time = base::TimeTicks::Now();
 
   bool fullscreen_low_power_layer_valid = false;
@@ skipping 142 matching lines @@
   gl_renderer_id_ = context_renderer_id & kCGLRendererIDMatchingMask;
 
   // Post a task holding a reference to the new GL context. The reason for
   // this is to avoid creating-then-destroying the context for every image
   // transport surface that is observing the GPU switch.
   base::MessageLoop::current()->PostTask(
       FROM_HERE, base::Bind(&IOSurfaceContextNoOp, context_on_new_gpu));
 }
 
 }  // namespace gpu