Integrate VDA with WebRTC.

content/common/gpu/media/exynos_video_decode_accelerator.h

 // Copyright (c) 2012 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.
 //
 // This file contains an implementation of VideoDecoderAccelerator
 // that utilizes the hardware video decoder present on the Exynos SoC.
 
 #ifndef CONTENT_COMMON_GPU_MEDIA_EXYNOS_VIDEO_DECODE_ACCELERATOR_H_
 #define CONTENT_COMMON_GPU_MEDIA_EXYNOS_VIDEO_DECODE_ACCELERATOR_H_
 
@@ skipping 67 matching lines @@
   // Do any necessary initialization before the sandbox is enabled.
   static void PreSandboxInitialization();
 
   // Lazily initialize static data after sandbox is enabled.  Return false on
   // init failure.
   static bool PostSandboxInitialization();
 
  private:
   // These are rather subjectively tuned.
   enum {
-    kMfcInputBufferCount = 8,
+    kMfcInputBufferCount = 12,

[wuchengli, 2013/06/10 12:33:43]

Increasing the buffer counts has much better decode performance. But the root
cause is pegged GPU child thread . Let me know if I should revert this.

[Ami GONE FROM CHROMIUM, 2013/06/11 23:48:05]

These numbers are arbitrary (as l.86 says) so no need to revert this but IWBN to
check w/ sheu@ that he doesn't know of a reason these numbers are bad.

     kMfcInputBufferMaxSize = 512 * 1024,
-    kGscInputBufferCount = 4,
+    kGscInputBufferCount = 8,
     // Number of output buffers to use for each VDA stage above what's required
     // by the decoder (e.g. DPB size, in H264).
-    kDpbOutputBufferExtraCount = 3,
+    kDpbOutputBufferExtraCount = 8,

[Pawel Osciak, 2013/06/12 23:38:22]

This adds almost 42MB to our memory footprint for each 1080p video. While it's
not prohibitively high, we still have to address the performance problems anyway
and hopefully those will not be needed anymore then. We might consider holding
up on this particular change for now...

[sheu, 2013/06/13 01:24:26]

I can give you two buffers for free by removing
kGscOutputBufferExtraForSyncCount below (since we fixed our sync problems).

[wuchengli, 2013/06/13 10:28:07]

That's great. Will you do it? Or should I remove it in a separate CL?

     // We're continuing to have issues with synchronization between Mali 3D and
     // Exynos video decode, so we buffer this many extra GSC output buffers in
     // the GSC free output buffers queue, to add a safety margin.
     // Must be lesser than or equal to kDpbOutputBufferExtraCount.
     // TODO(sheu): remove this hack (http://crbug.com/225563).
     kGscOutputBufferExtraForSyncCount = 2,
   };
 
   // Internal state of the decoder.
   enum State {
@@ skipping 323 matching lines @@
 
   // The codec we'll be decoding for.
   media::VideoCodecProfile video_profile_;
 
   DISALLOW_COPY_AND_ASSIGN(ExynosVideoDecodeAccelerator);
 };
 
 }  // namespace content
 
 #endif  // CONTENT_COMMON_GPU_MEDIA_EXYNOS_VIDEO_DECODE_ACCELERATOR_H_