Use PPS/SPS NALUs to initialize a VTDecompressionSession.

content/common/gpu/media/vt_video_decode_accelerator.cc

Index: content/common/gpu/media/vt_video_decode_accelerator.cc
diff --git a/content/common/gpu/media/vt_video_decode_accelerator.cc b/content/common/gpu/media/vt_video_decode_accelerator.cc
index 59251d20465c7249396029833cf575007046efbd..ab6e31f9722d20503f05cbf7763a39f588656f06 100644
--- a/content/common/gpu/media/vt_video_decode_accelerator.cc
+++ b/content/common/gpu/media/vt_video_decode_accelerator.cc
@@ -2,37 +2,212 @@
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
+#include <CoreVideo/CoreVideo.h>
+#include <OpenGL/CGLIOSurface.h>
+
+#include "base/bind.h"
+#include "base/thread_task_runner_handle.h"
 #include "content/common/gpu/media/vt_video_decode_accelerator.h"
+#include "media/filters/h264_parser.h"
+
+using content_common_gpu_media::kModuleVt;
+using content_common_gpu_media::InitializeStubs;
+using content_common_gpu_media::IsVtInitialized;
+using content_common_gpu_media::StubPathMap;
 
 namespace content {
 
+// Route decoded frame callbacks back into the VTVideoDecodeAccelerator.
+static void OutputThunk(
+    void* decompression_output_refcon,
+    void* source_frame_refcon,
+    OSStatus status,
+    VTDecodeInfoFlags info_flags,
+    CVImageBufferRef image_buffer,
+    CMTime presentation_time_stamp,
+    CMTime presentation_duration) {
+  VTVideoDecodeAccelerator* vda =
+      reinterpret_cast<VTVideoDecodeAccelerator*>(decompression_output_refcon);
+  int32_t* bitstream_id_ptr = reinterpret_cast<int32_t*>(source_frame_refcon);
+  int32_t bitstream_id = *bitstream_id_ptr;
+  delete bitstream_id_ptr;
+  CFRetain(image_buffer);
+  vda->Output(bitstream_id, status, info_flags, image_buffer);
+}
+
 VTVideoDecodeAccelerator::VTVideoDecodeAccelerator(CGLContextObj cgl_context)
-    : loop_proxy_(base::MessageLoopProxy::current()),
+    : task_runner_(base::ThreadTaskRunnerHandle::Get()),
       cgl_context_(cgl_context),
       client_(NULL),
-      weak_this_factory_(this) {
+      format_(NULL),
+      session_(NULL),
+      coded_width_(0),
+      coded_height_(0),
+      weak_factory_(this) {
+  callback_.decompressionOutputCallback = OutputThunk;
+  callback_.decompressionOutputRefCon = this;
 }
 
 VTVideoDecodeAccelerator::~VTVideoDecodeAccelerator() {
+  // TODO(sandersd): Move these into scoped pointers.
+  if (format_)
+    CFRelease(format_);
+  if (session_)
+    CFRelease(session_);
 }
 
 bool VTVideoDecodeAccelerator::Initialize(
     media::VideoCodecProfile profile,
     Client* client) {
   DCHECK(CalledOnValidThread());
-  DVLOG(2) << __FUNCTION__;
   client_ = client;
 
   // Only H.264 is supported.
   if (profile < media::H264PROFILE_MIN || profile > media::H264PROFILE_MAX)
     return false;
 
-  // Prevent anyone from using VTVideoDecoder for now. http://crbug.com/133828
-  return false;
+  // TODO(sandersd): Move VideoToolbox library loading to sandbox startup;
+  // until then, --no-sandbox is required.
+  if (!IsVtInitialized()) {
+    StubPathMap paths;
+    // CoreVideo is also required, but the loader stops after the first
+    // path is loaded. Instead we rely on the transitive dependency from
+    // VideoToolbox to CoreVideo.
+    // TODO(sandersd): Fallback to PrivateFrameworks for VideoToolbox.
+    paths[kModuleVt].push_back(FILE_PATH_LITERAL(
+        "/System/Library/Frameworks/VideoToolbox.framework/VideoToolbox"));
+    if (!InitializeStubs(paths))
+      return false;
+  }
+
+  // Note that --ignore-gpu-blacklist is still required to get here.
+  return true;
+}
+
+// TODO(sandersd): Proper error reporting instead of CHECKs.
+void VTVideoDecodeAccelerator::ConfigureDecoder(
+    const std::vector<const uint8_t*>& nalu_data_ptrs,
+    const std::vector<size_t>& nalu_data_sizes) {
+  if (format_ != NULL) {
+    CFRelease(format_);
+    format_ = NULL;
+  }
+
+  CHECK(!CMVideoFormatDescriptionCreateFromH264ParameterSets(
+      kCFAllocatorDefault,
+      nalu_data_ptrs.size(),      // parameter_set_count
+      &nalu_data_ptrs.front(),    // &parameter_set_pointers
+      &nalu_data_sizes.front(),   // &parameter_set_sizes
+      4,                          // nal_unit_header_length
+      &format_
+  ));
+
+  // TODO(sandersd): Check if the size has changed and handle picture requests.
+  CMVideoDimensions coded_size = CMVideoFormatDescriptionGetDimensions(format_);
+  coded_width_ = coded_size.width;
+  coded_height_ = coded_size.height;
+
+  // TODO(sandersd): Scoped pointers.
+  CFMutableDictionaryRef decoder_config = CFDictionaryCreateMutable(
+      kCFAllocatorDefault,
+      1,                          // capacity
+      &kCFTypeDictionaryKeyCallBacks,
+      &kCFTypeDictionaryValueCallBacks);
+
+  CFDictionarySetValue(
+      decoder_config,
+      // kVTVideoDecoderSpecification_EnableHardwareAcceleratedVideoDecoder
+      CFSTR("EnableHardwareAcceleratedVideoDecoder"),
+      kCFBooleanTrue);
+
+  CFMutableDictionaryRef image_config = CFDictionaryCreateMutable(
+      kCFAllocatorDefault,
+      4,                          // capacity
+      &kCFTypeDictionaryKeyCallBacks,
+      &kCFTypeDictionaryValueCallBacks);
+
+  // TODO(sandersd): ARGB for video that is not 4:2:0.
+  int32_t pixel_format = '2vuy';
+#define CFINT(i) CFNumberCreate(kCFAllocatorDefault, kCFNumberSInt32Type, &i)
+  CFNumberRef cf_pixel_format = CFINT(pixel_format);
+  CFNumberRef cf_width = CFINT(coded_width_);
+  CFNumberRef cf_height = CFINT(coded_height_);
+  CFDictionarySetValue(
+      image_config, kCVPixelBufferPixelFormatTypeKey, cf_pixel_format);
+  CFDictionarySetValue(image_config, kCVPixelBufferWidthKey, cf_width);
+  CFDictionarySetValue(image_config, kCVPixelBufferHeightKey, cf_height);
+  CFDictionarySetValue(
+      image_config, kCVPixelBufferOpenGLCompatibilityKey, kCFBooleanTrue);
+  CFRelease(cf_pixel_format);
+  CFRelease(cf_width);
+  CFRelease(cf_height);
+
+  if (session_ != NULL) {
+    // TODO(sandersd): Check if the session is compatable first.
+    // TODO(sandersd): Flush frames.
+    CFRelease(session_);
+    session_ = NULL;
+  }
+
+  CHECK(!VTDecompressionSessionCreate(
+      kCFAllocatorDefault,
+      format_,              // video_format_description
+      decoder_config,       // video_decoder_specification
+      image_config,         // destination_image_buffer_attributes
+      &callback_,           // output_callback
+      &session_
+  ));
+
+  CFRelease(decoder_config);
+  CFRelease(image_config);
 }
 
 void VTVideoDecodeAccelerator::Decode(const media::BitstreamBuffer& bitstream) {
   DCHECK(CalledOnValidThread());
+
+  // Map the bitstream buffer.
+  base::SharedMemory memory(bitstream.handle(), true);
+  size_t size = bitstream.size();
+  CHECK(memory.Map(size));
+  const uint8_t* buf = static_cast<uint8_t*>(memory.memory());

[Pawel Osciak, 2014/07/09 00:56:05]

Please try to do as little as possible in Decode(), as this is executed on
GPUProcess ChildThread an blocks other GPU tasks. It'd be better to have a
separate thread for this.

[sandersd (OOO until July 31), 2014/07/09 02:16:01]

Done. I've moved this into a decoder thread, similar to how the VAAPI
implementation is done.

+
+  // Locate relevant NALUs in the buffer.
+  size_t slice_data_size = 0;
+  std::vector<media::H264NALU> slice_nalus;

[Pawel Osciak, 2014/07/09 00:56:05]

Is this to store the slice nalus that come before SPS and PPS to use them with
an SPS that comes after them? This would not be a correct stream. An SPS that
comes after a slice cannot be applied to slices before it.

[sandersd (OOO until July 31), 2014/07/09 02:16:01]

No, it's an implementation to take any that are at the start without having to
parse twice, and without all the error checking done yet.

+  std::vector<const uint8_t*> config_nalu_data_ptrs;
+  std::vector<size_t> config_nalu_data_sizes;
+  parser_.SetStream(buf, size);
+  media::H264NALU nalu;
+  while (true) {
+    media::H264Parser::Result result = parser_.AdvanceToNextNALU(&nalu);
+    if (result == media::H264Parser::kEOStream)
+      break;
+    CHECK(result == media::H264Parser::kOk);

[Pawel Osciak, 2014/07/09 00:56:05]

It's better not to crash the GPU process on a invalid stream. Please
client->NotifyError instead.

[sandersd (OOO until July 31), 2014/07/09 02:16:01]

Indeed, but I was hoping to keep the boilerplate low for the next few CLs and
then replace all the CHECKs in one go.

+    if (nalu.nal_unit_type >= 1 && nalu.nal_unit_type <= 5) {

[Pawel Osciak, 2014/07/09 00:56:05]

Please use nalu type defines instead of hardcoded numbers.

[sandersd (OOO until July 31), 2014/07/09 02:16:01]

Done. Changed to accept all types, as that is closer to the correct
implementation once CanAcceptFormatDescription is used.

+      slice_nalus.push_back(nalu);
+      slice_data_size += 4 + nalu.size;
+    } else if (nalu.nal_unit_type == media::H264NALU::kSPS ||
+               nalu.nal_unit_type == media::H264NALU::kPPS ||
+               nalu.nal_unit_type == media::H264NALU::kSPSExt) {
+      config_nalu_data_ptrs.push_back(nalu.data);
+      config_nalu_data_sizes.push_back(nalu.size);
+    }
+  }
+
+  if (!config_nalu_data_ptrs.empty())
+    ConfigureDecoder(config_nalu_data_ptrs, config_nalu_data_sizes);
+
+  // TODO(sandersd): Rewrite slice NALU headers and send for decoding.
+}
+
+// This method may be called on any VideoToolbox thread.
+void VTVideoDecodeAccelerator::Output(
+    int32_t bitstream_id,
+    OSStatus status,
+    VTDecodeInfoFlags info_flags,
+    CVImageBufferRef image_buffer) {
+  // TODO(sandersd): Store the frame in a queue.
+  CFRelease(image_buffer);
 }
 
 void VTVideoDecodeAccelerator::AssignPictureBuffers(
@@ -46,14 +221,18 @@ void VTVideoDecodeAccelerator::ReusePictureBuffer(int32_t picture_id) {
 
 void VTVideoDecodeAccelerator::Flush() {
   DCHECK(CalledOnValidThread());
+  // TODO(sandersd): Trigger flush, sending frames.
 }
 
 void VTVideoDecodeAccelerator::Reset() {
   DCHECK(CalledOnValidThread());
+  // TODO(sandersd): Trigger flush, discarding frames.
 }
 
 void VTVideoDecodeAccelerator::Destroy() {
   DCHECK(CalledOnValidThread());
+  // TODO(sandersd): Trigger flush, discarding frames, and wait for them.
+  delete this;
 }
 
 bool VTVideoDecodeAccelerator::CanDecodeOnIOThread() {