Add support for EXT_image_flush_external extension

ui/gl/gl_image_ozone_native_pixmap.cc

Index: ui/gl/gl_image_ozone_native_pixmap.cc
diff --git a/ui/gl/gl_image_ozone_native_pixmap.cc b/ui/gl/gl_image_ozone_native_pixmap.cc
index dd2fb3e1a4bbd9dc6f8f7fdf9d0b78af3e6bd1cb..91d27ec587ab03d4cb1d8cc6eacdfa8bf4459cb5 100644
--- a/ui/gl/gl_image_ozone_native_pixmap.cc
+++ b/ui/gl/gl_image_ozone_native_pixmap.cc
@@ -3,6 +3,7 @@
 // found in the LICENSE file.
 
 #include "ui/gfx/buffer_format_util.h"
+#include "ui/gl/gl_fence.h"
 #include "ui/gl/gl_image_ozone_native_pixmap.h"
 
 #define FOURCC(a, b, c, d)                                        \
@@ -242,4 +243,19 @@ unsigned GLImageOzoneNativePixmap::GetInternalFormatForTesting(
   return GL_NONE;
 }
 
+void GLImageOzoneNativePixmap::Flush() {
+  const EGLAttrib attribs[] = {EGL_IMAGE_EXTERNAL_TARGET_NVX,
+                               EGL_DECOMPRESSED_NVX, EGL_NONE};
+
+  EGLDisplay display = eglGetDisplay(EGL_DEFAULT_DISPLAY);

[dnicoara, 2016/06/29 14:21:56]

I believe you should be using "GLSurfaceEGL::GetHardwareDisplay()" here.

[vinceh, 2016/06/30 09:22:48]

OK, will fix.

+  if (!eglImageFlushExternalEXT(display, egl_image_, attribs)) {

[piman, 2016/06/29 17:34:50]

What if the extension is not supported?

[vinceh, 2016/06/30 09:22:48]

I should add check for the extension. Thanks.

+    LOG(ERROR) << "Failed to flush rendering";
+    return;
+  }
+
+  GLFence* fence = GLFence::Create();

[piman, 2016/06/29 17:34:50]

nit: use std::unique_ptr<GLFence>

+  fence->ClientWait();

[piman, 2016/06/29 17:34:50]

So, you're doing this on all drivers, even though all current ones don't need
this. This is a big hammer (which is not reflected in the "Flush" name nor the
documentation for the API), and will likely cause significant regressions on
those other drivers.

Even on NVIDIA, this looks like it would break pipelining and cause poor
performance.

Can we have a high level discussion about what's needed - maybe in a bug - and
how we might go around to solve it in a more efficient / more general way?

[vinceh, 2016/06/30 09:22:48]

I should do this only when the extension is supported.
Could you go to
https://docs.google.com/document/d/1Z3P00qyjI7R5LYA2T9qYzmXymm1NOTbcoFbEod_GD...
and comment there? If that's not convenient then I will file a bug?

[piman, 2016/06/30 18:18:58]

Commented there. I think this should integrate with the code in
GLSurfaceOzoneSurfaceless::SwapBuffersAsync to wait for the fence in a worker
thread

+  delete fence;
+}
+
 }  // namespace gl