content/common/gpu/media/vt_video_decode_accelerator.cc

   1 // Copyright 2014 The Chromium Authors. All rights reserved.
   2 // Use of this source code is governed by a BSD-style license that can be
   3 // found in the LICENSE file.
   4
   5 #include <algorithm>
   6
   7 #include <CoreVideo/CoreVideo.h>
   8 #include <OpenGL/CGLIOSurface.h>
   9 #include <OpenGL/gl.h>
  10
  11 #include "base/bind.h"
  12 #include "base/command_line.h"
  13 #include "base/logging.h"
  14 #include "base/mac/mac_logging.h"
  15 #include "base/metrics/histogram_macros.h"
  16 #include "base/sys_byteorder.h"
  17 #include "base/sys_info.h"
  18 #include "base/thread_task_runner_handle.h"
  19 #include "base/version.h"
  20 #include "content/common/gpu/media/vt_video_decode_accelerator.h"
  21 #include "content/public/common/content_switches.h"
  22 #include "media/base/limits.h"
  23 #include "ui/gl/gl_context.h"
  24 #include "ui/gl/scoped_binders.h"
  25
  26 using content_common_gpu_media::kModuleVt;
  27 using content_common_gpu_media::InitializeStubs;
  28 using content_common_gpu_media::IsVtInitialized;
  29 using content_common_gpu_media::StubPathMap;
  30
  31 #define NOTIFY_STATUS(name, status, session_failure)   \
  32     do {                                               \
  33       OSSTATUS_DLOG(ERROR, status) << name;            \
  34       NotifyError(PLATFORM_FAILURE, session_failure);  \
  35     } while (0)
  36
  37 namespace content {
  38
  39 // Only H.264 with 4:2:0 chroma sampling is supported.
  40 static const media::VideoCodecProfile kSupportedProfiles[] = {
  41   media::H264PROFILE_BASELINE,
  42   media::H264PROFILE_MAIN,
  43   media::H264PROFILE_EXTENDED,
  44   media::H264PROFILE_HIGH,
  45   media::H264PROFILE_HIGH10PROFILE,
  46   media::H264PROFILE_SCALABLEBASELINE,
  47   media::H264PROFILE_SCALABLEHIGH,
  48   media::H264PROFILE_STEREOHIGH,
  49   media::H264PROFILE_MULTIVIEWHIGH,
  50 };
  51
  52 // Size to use for NALU length headers in AVC format (can be 1, 2, or 4).
  53 static const int kNALUHeaderLength = 4;
  54
  55 // We request 5 picture buffers from the client, each of which has a texture ID
  56 // that we can bind decoded frames to. We need enough to satisfy preroll, and
  57 // enough to avoid unnecessary stalling, but no more than that. The resource
  58 // requirements are low, as we don't need the textures to be backed by storage.
  59 static const int kNumPictureBuffers = media::limits::kMaxVideoFrames + 1;
  60
  61 // Maximum number of frames to queue for reordering before we stop asking for
  62 // more. (NotifyEndOfBitstreamBuffer() is called when frames are moved into the
  63 // reorder queue.)
  64 static const int kMaxReorderQueueSize = 16;
  65
  66 // When set to false, always create a new decoder instead of reusing the
  67 // existing configuration when the configuration changes. This works around a
  68 // bug in VideoToolbox that results in corruption before Mac OS X 10.10.3. The
  69 // value is set in InitializeVideoToolbox().
  70 static bool g_enable_compatible_configuration_reuse = true;
  71
  72 // Build an |image_config| dictionary for VideoToolbox initialization.
  73 static base::ScopedCFTypeRef<CFMutableDictionaryRef>
  74 BuildImageConfig(CMVideoDimensions coded_dimensions) {
  75   base::ScopedCFTypeRef<CFMutableDictionaryRef> image_config;
  76
  77   // 4:2:2 is used over the native 4:2:0 because only 4:2:2 can be directly
  78   // bound to a texture by CGLTexImageIOSurface2D().
  79   int32_t pixel_format = kCVPixelFormatType_422YpCbCr8;
  80 #define CFINT(i) CFNumberCreate(kCFAllocatorDefault, kCFNumberSInt32Type, &i)
  81   base::ScopedCFTypeRef<CFNumberRef> cf_pixel_format(CFINT(pixel_format));
  82   base::ScopedCFTypeRef<CFNumberRef> cf_width(CFINT(coded_dimensions.width));
  83   base::ScopedCFTypeRef<CFNumberRef> cf_height(CFINT(coded_dimensions.height));
  84 #undef CFINT
  85   if (!cf_pixel_format.get() || !cf_width.get() || !cf_height.get())
  86     return image_config;
  87
  88   image_config.reset(
  89       CFDictionaryCreateMutable(
  90           kCFAllocatorDefault,
  91           4,  // capacity
  92           &kCFTypeDictionaryKeyCallBacks,
  93           &kCFTypeDictionaryValueCallBacks));
  94   if (!image_config.get())
  95     return image_config;
  96
  97   CFDictionarySetValue(image_config, kCVPixelBufferPixelFormatTypeKey,
  98                        cf_pixel_format);
  99   CFDictionarySetValue(image_config, kCVPixelBufferWidthKey, cf_width);
 100   CFDictionarySetValue(image_config, kCVPixelBufferHeightKey, cf_height);
 101   CFDictionarySetValue(image_config, kCVPixelBufferOpenGLCompatibilityKey,
 102                        kCFBooleanTrue);
 103
 104   return image_config;
 105 }
 106
 107 // Create a VTDecompressionSession using the provided |pps| and |sps|. If
 108 // |require_hardware| is true, the session must uses real hardware decoding
 109 // (as opposed to software decoding inside of VideoToolbox) to be considered
 110 // successful.
 111 //
 112 // TODO(sandersd): Merge with ConfigureDecoder(), as the code is very similar.
 113 static bool CreateVideoToolboxSession(const uint8_t* sps, size_t sps_size,
 114                                       const uint8_t* pps, size_t pps_size,
 115                                       bool require_hardware) {
 116   const uint8_t* data_ptrs[] = {sps, pps};
 117   const size_t data_sizes[] = {sps_size, pps_size};
 118
 119   base::ScopedCFTypeRef<CMFormatDescriptionRef> format;
 120   OSStatus status = CMVideoFormatDescriptionCreateFromH264ParameterSets(
 121       kCFAllocatorDefault,
 122       2,                    // parameter_set_count
 123       data_ptrs,            // &parameter_set_pointers
 124       data_sizes,           // &parameter_set_sizes
 125       kNALUHeaderLength,    // nal_unit_header_length
 126       format.InitializeInto());
 127   if (status) {
 128     OSSTATUS_DLOG(WARNING, status)
 129         << "Failed to create CMVideoFormatDescription.";
 130     return false;
 131   }
 132
 133   base::ScopedCFTypeRef<CFMutableDictionaryRef> decoder_config(
 134       CFDictionaryCreateMutable(
 135           kCFAllocatorDefault,
 136           1,  // capacity
 137           &kCFTypeDictionaryKeyCallBacks,
 138           &kCFTypeDictionaryValueCallBacks));
 139   if (!decoder_config.get())
 140     return false;
 141
 142   if (require_hardware) {
 143     CFDictionarySetValue(
 144         decoder_config,
 145         // kVTVideoDecoderSpecification_RequireHardwareAcceleratedVideoDecoder
 146         CFSTR("RequireHardwareAcceleratedVideoDecoder"),
 147         kCFBooleanTrue);
 148   }
 149
 150   base::ScopedCFTypeRef<CFMutableDictionaryRef> image_config(
 151       BuildImageConfig(CMVideoFormatDescriptionGetDimensions(format)));
 152   if (!image_config.get())
 153     return false;
 154
 155   VTDecompressionOutputCallbackRecord callback = {0};
 156
 157   base::ScopedCFTypeRef<VTDecompressionSessionRef> session;
 158   status = VTDecompressionSessionCreate(
 159       kCFAllocatorDefault,
 160       format,               // video_format_description
 161       decoder_config,       // video_decoder_specification
 162       image_config,         // destination_image_buffer_attributes
 163       &callback,            // output_callback
 164       session.InitializeInto());
 165   if (status) {
 166     OSSTATUS_DLOG(WARNING, status) << "Failed to create VTDecompressionSession";
 167     return false;
 168   }
 169
 170   return true;
 171 }
 172
 173 // The purpose of this function is to preload the generic and hardware-specific
 174 // libraries required by VideoToolbox before the GPU sandbox is enabled.
 175 // VideoToolbox normally loads the hardware-specific libraries lazily, so we
 176 // must actually create a decompression session. If creating a decompression
 177 // session fails, hardware decoding will be disabled (Initialize() will always
 178 // return false).
 179 static bool InitializeVideoToolboxInternal() {
 180   if (base::CommandLine::ForCurrentProcess()->HasSwitch(
 181           switches::kDisableAcceleratedVideoDecode)) {
 182     return false;
 183   }
 184
 185   if (!IsVtInitialized()) {
 186     // CoreVideo is also required, but the loader stops after the first path is
 187     // loaded. Instead we rely on the transitive dependency from VideoToolbox to
 188     // CoreVideo.
 189     // TODO(sandersd): Fallback to PrivateFrameworks to support OS X < 10.8.
 190     StubPathMap paths;
 191     paths[kModuleVt].push_back(FILE_PATH_LITERAL(
 192         "/System/Library/Frameworks/VideoToolbox.framework/VideoToolbox"));
 193     if (!InitializeStubs(paths)) {
 194       LOG(WARNING) << "Failed to initialize VideoToolbox framework. "
 195                    << "Hardware accelerated video decoding will be disabled.";
 196       return false;
 197     }
 198   }
 199
 200   // Create a hardware decoding session.
 201   // SPS and PPS data are taken from a 480p sample (buck2.mp4).
 202   const uint8_t sps_normal[] = {0x67, 0x64, 0x00, 0x1e, 0xac, 0xd9, 0x80, 0xd4,
 203                                 0x3d, 0xa1, 0x00, 0x00, 0x03, 0x00, 0x01, 0x00,
 204                                 0x00, 0x03, 0x00, 0x30, 0x8f, 0x16, 0x2d, 0x9a};
 205   const uint8_t pps_normal[] = {0x68, 0xe9, 0x7b, 0xcb};
 206   if (!CreateVideoToolboxSession(sps_normal, arraysize(sps_normal), pps_normal,
 207                                  arraysize(pps_normal), true)) {
 208     LOG(WARNING) << "Failed to create hardware VideoToolbox session. "
 209                  << "Hardware accelerated video decoding will be disabled.";
 210     return false;
 211   }
 212
 213   // Create a software decoding session.
 214   // SPS and PPS data are taken from a 18p sample (small2.mp4).
 215   const uint8_t sps_small[] = {0x67, 0x64, 0x00, 0x0a, 0xac, 0xd9, 0x89, 0x7e,
 216                                0x22, 0x10, 0x00, 0x00, 0x3e, 0x90, 0x00, 0x0e,
 217                                0xa6, 0x08, 0xf1, 0x22, 0x59, 0xa0};
 218   const uint8_t pps_small[] = {0x68, 0xe9, 0x79, 0x72, 0xc0};
 219   if (!CreateVideoToolboxSession(sps_small, arraysize(sps_small), pps_small,
 220                                  arraysize(pps_small), false)) {
 221     LOG(WARNING) << "Failed to create software VideoToolbox session. "
 222                  << "Hardware accelerated video decoding will be disabled.";
 223     return false;
 224   }
 225
 226   // Set |g_enable_compatible_configuration_reuse| to false on
 227   // Mac OS X < 10.10.3.
 228   base::Version os_x_version(base::SysInfo::OperatingSystemVersion());
 229   if (os_x_version.IsOlderThan("10.10.3"))
 230     g_enable_compatible_configuration_reuse = false;
 231
 232   return true;
 233 }
 234
 235 bool InitializeVideoToolbox() {
 236   // InitializeVideoToolbox() is called only from the GPU process main thread;
 237   // once for sandbox warmup, and then once each time a VTVideoDecodeAccelerator
 238   // is initialized.
 239   static bool attempted = false;
 240   static bool succeeded = false;
 241
 242   if (!attempted) {
 243     attempted = true;
 244     succeeded = InitializeVideoToolboxInternal();
 245   }
 246
 247   return succeeded;
 248 }
 249
 250 // Route decoded frame callbacks back into the VTVideoDecodeAccelerator.
 251 static void OutputThunk(
 252     void* decompression_output_refcon,
 253     void* source_frame_refcon,
 254     OSStatus status,
 255     VTDecodeInfoFlags info_flags,
 256     CVImageBufferRef image_buffer,
 257     CMTime presentation_time_stamp,
 258     CMTime presentation_duration) {
 259   VTVideoDecodeAccelerator* vda =
 260       reinterpret_cast<VTVideoDecodeAccelerator*>(decompression_output_refcon);
 261   vda->Output(source_frame_refcon, status, image_buffer);
 262 }
 263
 264 VTVideoDecodeAccelerator::Task::Task(TaskType type) : type(type) {
 265 }
 266
 267 VTVideoDecodeAccelerator::Task::~Task() {
 268 }
 269
 270 VTVideoDecodeAccelerator::Frame::Frame(int32_t bitstream_id)
 271     : bitstream_id(bitstream_id), pic_order_cnt(0), reorder_window(0) {
 272 }
 273
 274 VTVideoDecodeAccelerator::Frame::~Frame() {
 275 }
 276
 277 bool VTVideoDecodeAccelerator::FrameOrder::operator()(
 278     const linked_ptr<Frame>& lhs,
 279     const linked_ptr<Frame>& rhs) const {
 280   if (lhs->pic_order_cnt != rhs->pic_order_cnt)
 281     return lhs->pic_order_cnt > rhs->pic_order_cnt;
 282   // If |pic_order_cnt| is the same, fall back on using the bitstream order.
 283   // TODO(sandersd): Assign a sequence number in Decode() and use that instead.
 284   // TODO(sandersd): Using the sequence number, ensure that frames older than
 285   // |kMaxReorderQueueSize| are ordered first, regardless of |pic_order_cnt|.
 286   return lhs->bitstream_id > rhs->bitstream_id;
 287 }
 288
 289 VTVideoDecodeAccelerator::VTVideoDecodeAccelerator(
 290     const base::Callback<bool(void)>& make_context_current)
 291     : make_context_current_(make_context_current),
 292       client_(nullptr),
 293       state_(STATE_DECODING),
 294       format_(nullptr),
 295       session_(nullptr),
 296       last_sps_id_(-1),
 297       last_pps_id_(-1),
 298       gpu_task_runner_(base::ThreadTaskRunnerHandle::Get()),
 299       decoder_thread_("VTDecoderThread"),
 300       weak_this_factory_(this) {
 301   DCHECK(!make_context_current_.is_null());
 302   callback_.decompressionOutputCallback = OutputThunk;
 303   callback_.decompressionOutputRefCon = this;
 304   weak_this_ = weak_this_factory_.GetWeakPtr();
 305 }
 306
 307 VTVideoDecodeAccelerator::~VTVideoDecodeAccelerator() {
 308 }
 309
 310 bool VTVideoDecodeAccelerator::Initialize(
 311     media::VideoCodecProfile profile,
 312     Client* client) {
 313   DCHECK(gpu_thread_checker_.CalledOnValidThread());
 314   client_ = client;
 315
 316   if (!InitializeVideoToolbox())
 317     return false;
 318
 319   bool profile_supported = false;
 320   for (const auto& supported_profile : kSupportedProfiles) {
 321     if (profile == supported_profile) {
 322       profile_supported = true;
 323       break;
 324     }
 325   }
 326   if (!profile_supported)
 327     return false;
 328
 329   // Spawn a thread to handle parsing and calling VideoToolbox.
 330   if (!decoder_thread_.Start())
 331     return false;
 332
 333   // Count the session as successfully initialized.
 334   UMA_HISTOGRAM_ENUMERATION("Media.VTVDA.SessionFailureReason",
 335                             SFT_SUCCESSFULLY_INITIALIZED,
 336                             SFT_MAX + 1);
 337   return true;
 338 }
 339
 340 bool VTVideoDecodeAccelerator::FinishDelayedFrames() {
 341   DCHECK(decoder_thread_.task_runner()->BelongsToCurrentThread());
 342   if (session_) {
 343     OSStatus status = VTDecompressionSessionWaitForAsynchronousFrames(session_);
 344     if (status) {
 345       NOTIFY_STATUS("VTDecompressionSessionWaitForAsynchronousFrames()",
 346                     status, SFT_PLATFORM_ERROR);
 347       return false;
 348     }
 349   }
 350   return true;
 351 }
 352
 353 bool VTVideoDecodeAccelerator::ConfigureDecoder() {
 354   DCHECK(decoder_thread_.task_runner()->BelongsToCurrentThread());
 355   DCHECK(!last_sps_.empty());
 356   DCHECK(!last_pps_.empty());
 357
 358   // Build the configuration records.
 359   std::vector<const uint8_t*> nalu_data_ptrs;
 360   std::vector<size_t> nalu_data_sizes;
 361   nalu_data_ptrs.reserve(3);
 362   nalu_data_sizes.reserve(3);
 363   nalu_data_ptrs.push_back(&last_sps_.front());
 364   nalu_data_sizes.push_back(last_sps_.size());
 365   if (!last_spsext_.empty()) {
 366     nalu_data_ptrs.push_back(&last_spsext_.front());
 367     nalu_data_sizes.push_back(last_spsext_.size());
 368   }
 369   nalu_data_ptrs.push_back(&last_pps_.front());
 370   nalu_data_sizes.push_back(last_pps_.size());
 371
 372   // Construct a new format description from the parameter sets.
 373   // TODO(sandersd): Replace this with custom code to support OS X < 10.9.
 374   format_.reset();
 375   OSStatus status = CMVideoFormatDescriptionCreateFromH264ParameterSets(
 376       kCFAllocatorDefault,
 377       nalu_data_ptrs.size(),      // parameter_set_count
 378       &nalu_data_ptrs.front(),    // &parameter_set_pointers
 379       &nalu_data_sizes.front(),   // &parameter_set_sizes
 380       kNALUHeaderLength,          // nal_unit_header_length
 381       format_.InitializeInto());
 382   if (status) {
 383     NOTIFY_STATUS("CMVideoFormatDescriptionCreateFromH264ParameterSets()",
 384                   status, SFT_PLATFORM_ERROR);
 385     return false;
 386   }
 387
 388   // Store the new configuration data.
 389   CMVideoDimensions coded_dimensions =
 390       CMVideoFormatDescriptionGetDimensions(format_);
 391   coded_size_.SetSize(coded_dimensions.width, coded_dimensions.height);
 392
 393   // If the session is compatible, there's nothing else to do.
 394   if (g_enable_compatible_configuration_reuse && session_ &&
 395       VTDecompressionSessionCanAcceptFormatDescription(session_, format_)) {
 396     return true;
 397   }
 398
 399   // Prepare VideoToolbox configuration dictionaries.
 400   base::ScopedCFTypeRef<CFMutableDictionaryRef> decoder_config(
 401       CFDictionaryCreateMutable(
 402           kCFAllocatorDefault,
 403           1,  // capacity
 404           &kCFTypeDictionaryKeyCallBacks,
 405           &kCFTypeDictionaryValueCallBacks));
 406   if (!decoder_config.get()) {
 407     DLOG(ERROR) << "Failed to create CFMutableDictionary.";
 408     NotifyError(PLATFORM_FAILURE, SFT_PLATFORM_ERROR);
 409     return false;
 410   }
 411
 412   CFDictionarySetValue(
 413       decoder_config,
 414       // kVTVideoDecoderSpecification_EnableHardwareAcceleratedVideoDecoder
 415       CFSTR("EnableHardwareAcceleratedVideoDecoder"),
 416       kCFBooleanTrue);
 417
 418   base::ScopedCFTypeRef<CFMutableDictionaryRef> image_config(
 419       BuildImageConfig(coded_dimensions));
 420   if (!image_config.get()) {
 421     DLOG(ERROR) << "Failed to create decoder image configuration.";
 422     NotifyError(PLATFORM_FAILURE, SFT_PLATFORM_ERROR);
 423     return false;
 424   }
 425
 426   // Ensure that the old decoder emits all frames before the new decoder can
 427   // emit any.
 428   if (!FinishDelayedFrames())
 429     return false;
 430
 431   session_.reset();
 432   status = VTDecompressionSessionCreate(
 433       kCFAllocatorDefault,
 434       format_,              // video_format_description
 435       decoder_config,       // video_decoder_specification
 436       image_config,         // destination_image_buffer_attributes
 437       &callback_,           // output_callback
 438       session_.InitializeInto());
 439   if (status) {
 440     NOTIFY_STATUS("VTDecompressionSessionCreate()", status,
 441                   SFT_UNSUPPORTED_STREAM_PARAMETERS);
 442     return false;
 443   }
 444
 445   // Report whether hardware decode is being used.
 446   bool using_hardware = false;
 447   base::ScopedCFTypeRef<CFBooleanRef> cf_using_hardware;
 448   if (VTSessionCopyProperty(
 449           session_,
 450           // kVTDecompressionPropertyKey_UsingHardwareAcceleratedVideoDecoder
 451           CFSTR("UsingHardwareAcceleratedVideoDecoder"),
 452           kCFAllocatorDefault,
 453           cf_using_hardware.InitializeInto()) == 0) {
 454     using_hardware = CFBooleanGetValue(cf_using_hardware);
 455   }
 456   UMA_HISTOGRAM_BOOLEAN("Media.VTVDA.HardwareAccelerated", using_hardware);
 457
 458   return true;
 459 }
 460
 461 void VTVideoDecodeAccelerator::DecodeTask(
 462     const media::BitstreamBuffer& bitstream,
 463     Frame* frame) {
 464   DCHECK(decoder_thread_.task_runner()->BelongsToCurrentThread());
 465
 466   // Map the bitstream buffer.
 467   base::SharedMemory memory(bitstream.handle(), true);
 468   size_t size = bitstream.size();
 469   if (!memory.Map(size)) {
 470     DLOG(ERROR) << "Failed to map bitstream buffer";
 471     NotifyError(PLATFORM_FAILURE, SFT_PLATFORM_ERROR);
 472     return;
 473   }
 474   const uint8_t* buf = static_cast<uint8_t*>(memory.memory());
 475
 476   // NALUs are stored with Annex B format in the bitstream buffer (start codes),
 477   // but VideoToolbox expects AVC format (length headers), so we must rewrite
 478   // the data.
 479   //
 480   // Locate relevant NALUs and compute the size of the rewritten data. Also
 481   // record any parameter sets for VideoToolbox initialization.
 482   std::vector<uint8_t> sps;
 483   std::vector<uint8_t> spsext;
 484   std::vector<uint8_t> pps;
 485   bool has_slice = false;
 486   size_t data_size = 0;
 487   std::vector<media::H264NALU> nalus;
 488   parser_.SetStream(buf, size);
 489   media::H264NALU nalu;
 490   while (true) {
 491     media::H264Parser::Result result = parser_.AdvanceToNextNALU(&nalu);
 492     if (result == media::H264Parser::kEOStream)
 493       break;
 494     if (result == media::H264Parser::kUnsupportedStream) {
 495       DLOG(ERROR) << "Unsupported H.264 stream";
 496       NotifyError(PLATFORM_FAILURE, SFT_UNSUPPORTED_STREAM);
 497       return;
 498     }
 499     if (result != media::H264Parser::kOk) {
 500       DLOG(ERROR) << "Failed to parse H.264 stream";
 501       NotifyError(UNREADABLE_INPUT, SFT_INVALID_STREAM);
 502       return;
 503     }
 504     switch (nalu.nal_unit_type) {
 505       case media::H264NALU::kSPS:
 506         result = parser_.ParseSPS(&last_sps_id_);
 507         if (result == media::H264Parser::kUnsupportedStream) {
 508           DLOG(ERROR) << "Unsupported SPS";
 509           NotifyError(PLATFORM_FAILURE, SFT_UNSUPPORTED_STREAM);
 510           return;
 511         }
 512         if (result != media::H264Parser::kOk) {
 513           DLOG(ERROR) << "Could not parse SPS";
 514           NotifyError(UNREADABLE_INPUT, SFT_INVALID_STREAM);
 515           return;
 516         }
 517         sps.assign(nalu.data, nalu.data + nalu.size);
 518         spsext.clear();
 519         break;
 520
 521       case media::H264NALU::kSPSExt:
 522         // TODO(sandersd): Check that the previous NALU was an SPS.
 523         spsext.assign(nalu.data, nalu.data + nalu.size);
 524         break;
 525
 526       case media::H264NALU::kPPS:
 527         result = parser_.ParsePPS(&last_pps_id_);
 528         if (result == media::H264Parser::kUnsupportedStream) {
 529           DLOG(ERROR) << "Unsupported PPS";
 530           NotifyError(PLATFORM_FAILURE, SFT_UNSUPPORTED_STREAM);
 531           return;
 532         }
 533         if (result != media::H264Parser::kOk) {
 534           DLOG(ERROR) << "Could not parse PPS";
 535           NotifyError(UNREADABLE_INPUT, SFT_INVALID_STREAM);
 536           return;
 537         }
 538         pps.assign(nalu.data, nalu.data + nalu.size);
 539         break;
 540
 541       case media::H264NALU::kSliceDataA:
 542       case media::H264NALU::kSliceDataB:
 543       case media::H264NALU::kSliceDataC:
 544       case media::H264NALU::kNonIDRSlice:
 545         // TODO(sandersd): Check that there has been an IDR slice since the
 546         // last reset.
 547       case media::H264NALU::kIDRSlice:
 548         // Compute the |pic_order_cnt| for the picture from the first slice.
 549         // TODO(sandersd): Make sure that any further slices are part of the
 550         // same picture or a redundant coded picture.
 551         if (!has_slice) {
 552           media::H264SliceHeader slice_hdr;
 553           result = parser_.ParseSliceHeader(nalu, &slice_hdr);
 554           if (result == media::H264Parser::kUnsupportedStream) {
 555             DLOG(ERROR) << "Unsupported slice header";
 556             NotifyError(PLATFORM_FAILURE, SFT_UNSUPPORTED_STREAM);
 557             return;
 558           }
 559           if (result != media::H264Parser::kOk) {
 560             DLOG(ERROR) << "Could not parse slice header";
 561             NotifyError(UNREADABLE_INPUT, SFT_INVALID_STREAM);
 562             return;
 563           }
 564
 565           // TODO(sandersd): Maintain a cache of configurations and reconfigure
 566           // when a slice references a new config.
 567           DCHECK_EQ(slice_hdr.pic_parameter_set_id, last_pps_id_);
 568           const media::H264PPS* pps =
 569               parser_.GetPPS(slice_hdr.pic_parameter_set_id);
 570           if (!pps) {
 571             DLOG(ERROR) << "Mising PPS referenced by slice";
 572             NotifyError(UNREADABLE_INPUT, SFT_INVALID_STREAM);
 573             return;
 574           }
 575
 576           DCHECK_EQ(pps->seq_parameter_set_id, last_sps_id_);
 577           const media::H264SPS* sps = parser_.GetSPS(pps->seq_parameter_set_id);
 578           if (!sps) {
 579             DLOG(ERROR) << "Mising SPS referenced by PPS";
 580             NotifyError(UNREADABLE_INPUT, SFT_INVALID_STREAM);
 581             return;
 582           }
 583
 584           if (!poc_.ComputePicOrderCnt(sps, slice_hdr, &frame->pic_order_cnt)) {
 585             DLOG(ERROR) << "Unable to compute POC";
 586             NotifyError(UNREADABLE_INPUT, SFT_INVALID_STREAM);
 587             return;
 588           }
 589
 590           if (sps->vui_parameters_present_flag &&
 591               sps->bitstream_restriction_flag) {
 592             frame->reorder_window = std::min(sps->max_num_reorder_frames,
 593                                              kMaxReorderQueueSize - 1);
 594           }
 595         }
 596         has_slice = true;
 597       default:
 598         nalus.push_back(nalu);
 599         data_size += kNALUHeaderLength + nalu.size;
 600         break;
 601     }
 602   }
 603
 604   // Initialize VideoToolbox.
 605   bool config_changed = false;
 606   if (!sps.empty() && sps != last_sps_) {
 607     last_sps_.swap(sps);
 608     last_spsext_.swap(spsext);
 609     config_changed = true;
 610   }
 611   if (!pps.empty() && pps != last_pps_) {
 612     last_pps_.swap(pps);
 613     config_changed = true;
 614   }
 615   if (config_changed) {
 616     if (last_sps_.empty()) {
 617       DLOG(ERROR) << "Invalid configuration; no SPS";
 618       NotifyError(INVALID_ARGUMENT, SFT_INVALID_STREAM);
 619       return;
 620     }
 621     if (last_pps_.empty()) {
 622       DLOG(ERROR) << "Invalid configuration; no PPS";
 623       NotifyError(INVALID_ARGUMENT, SFT_INVALID_STREAM);
 624       return;
 625     }
 626     if (!ConfigureDecoder())
 627       return;
 628   }
 629
 630   // If there are no image slices, drop the bitstream buffer by returning an
 631   // empty frame.
 632   if (!has_slice) {
 633     if (!FinishDelayedFrames())
 634       return;
 635     gpu_task_runner_->PostTask(FROM_HERE, base::Bind(
 636         &VTVideoDecodeAccelerator::DecodeDone, weak_this_, frame));
 637     return;
 638   }
 639
 640   // If the session is not configured by this point, fail.
 641   if (!session_) {
 642     DLOG(ERROR) << "Cannot decode without configuration";
 643     NotifyError(INVALID_ARGUMENT, SFT_INVALID_STREAM);
 644     return;
 645   }
 646
 647   // Update the frame metadata with configuration data.
 648   frame->coded_size = coded_size_;
 649
 650   // Create a memory-backed CMBlockBuffer for the translated data.
 651   // TODO(sandersd): Pool of memory blocks.
 652   base::ScopedCFTypeRef<CMBlockBufferRef> data;
 653   OSStatus status = CMBlockBufferCreateWithMemoryBlock(
 654       kCFAllocatorDefault,
 655       nullptr,              // &memory_block
 656       data_size,            // block_length
 657       kCFAllocatorDefault,  // block_allocator
 658       nullptr,              // &custom_block_source
 659       0,                    // offset_to_data
 660       data_size,            // data_length
 661       0,                    // flags
 662       data.InitializeInto());
 663   if (status) {
 664     NOTIFY_STATUS("CMBlockBufferCreateWithMemoryBlock()", status,
 665                   SFT_PLATFORM_ERROR);
 666     return;
 667   }
 668
 669   // Make sure that the memory is actually allocated.
 670   // CMBlockBufferReplaceDataBytes() is documented to do this, but prints a
 671   // message each time starting in Mac OS X 10.10.
 672   status = CMBlockBufferAssureBlockMemory(data);
 673   if (status) {
 674     NOTIFY_STATUS("CMBlockBufferAssureBlockMemory()", status,
 675                   SFT_PLATFORM_ERROR);
 676     return;
 677   }
 678
 679   // Copy NALU data into the CMBlockBuffer, inserting length headers.
 680   size_t offset = 0;
 681   for (size_t i = 0; i < nalus.size(); i++) {
 682     media::H264NALU& nalu = nalus[i];
 683     uint32_t header = base::HostToNet32(static_cast<uint32_t>(nalu.size));
 684     status = CMBlockBufferReplaceDataBytes(
 685         &header, data, offset, kNALUHeaderLength);
 686     if (status) {
 687       NOTIFY_STATUS("CMBlockBufferReplaceDataBytes()", status,
 688                     SFT_PLATFORM_ERROR);
 689       return;
 690     }
 691     offset += kNALUHeaderLength;
 692     status = CMBlockBufferReplaceDataBytes(nalu.data, data, offset, nalu.size);
 693     if (status) {
 694       NOTIFY_STATUS("CMBlockBufferReplaceDataBytes()", status,
 695                     SFT_PLATFORM_ERROR);
 696       return;
 697     }
 698     offset += nalu.size;
 699   }
 700
 701   // Package the data in a CMSampleBuffer.
 702   base::ScopedCFTypeRef<CMSampleBufferRef> sample;
 703   status = CMSampleBufferCreate(
 704       kCFAllocatorDefault,
 705       data,                 // data_buffer
 706       true,                 // data_ready
 707       nullptr,              // make_data_ready_callback
 708       nullptr,              // make_data_ready_refcon
 709       format_,              // format_description
 710       1,                    // num_samples
 711       0,                    // num_sample_timing_entries
 712       nullptr,              // &sample_timing_array
 713       1,                    // num_sample_size_entries
 714       &data_size,           // &sample_size_array
 715       sample.InitializeInto());
 716   if (status) {
 717     NOTIFY_STATUS("CMSampleBufferCreate()", status, SFT_PLATFORM_ERROR);
 718     return;
 719   }
 720
 721   // Send the frame for decoding.
 722   // Asynchronous Decompression allows for parallel submission of frames
 723   // (without it, DecodeFrame() does not return until the frame has been
 724   // decoded). We don't enable Temporal Processing so that frames are always
 725   // returned in decode order; this makes it easier to avoid deadlock.
 726   VTDecodeFrameFlags decode_flags =
 727       kVTDecodeFrame_EnableAsynchronousDecompression;
 728   status = VTDecompressionSessionDecodeFrame(
 729       session_,
 730       sample,                                 // sample_buffer
 731       decode_flags,                           // decode_flags
 732       reinterpret_cast<void*>(frame),         // source_frame_refcon
 733       nullptr);                               // &info_flags_out
 734   if (status) {
 735     NOTIFY_STATUS("VTDecompressionSessionDecodeFrame()", status,
 736                   SFT_DECODE_ERROR);
 737     return;
 738   }
 739 }
 740
 741 // This method may be called on any VideoToolbox thread.
 742 void VTVideoDecodeAccelerator::Output(
 743     void* source_frame_refcon,
 744     OSStatus status,
 745     CVImageBufferRef image_buffer) {
 746   if (status) {
 747     NOTIFY_STATUS("Decoding", status, SFT_DECODE_ERROR);
 748     return;
 749   }
 750
 751   // The type of |image_buffer| is CVImageBuffer, but we only handle
 752   // CVPixelBuffers. This should be guaranteed as we set
 753   // kCVPixelBufferOpenGLCompatibilityKey in |image_config|.
 754   //
 755   // Sometimes, for unknown reasons (http://crbug.com/453050), |image_buffer| is
 756   // NULL, which causes CFGetTypeID() to crash. While the rest of the code would
 757   // smoothly handle NULL as a dropped frame, we choose to fail permanantly here
 758   // until the issue is better understood.
 759   if (!image_buffer || CFGetTypeID(image_buffer) != CVPixelBufferGetTypeID()) {
 760     DLOG(ERROR) << "Decoded frame is not a CVPixelBuffer";
 761     NotifyError(PLATFORM_FAILURE, SFT_DECODE_ERROR);
 762     return;
 763   }
 764
 765   Frame* frame = reinterpret_cast<Frame*>(source_frame_refcon);
 766   frame->image.reset(image_buffer, base::scoped_policy::RETAIN);
 767   gpu_task_runner_->PostTask(FROM_HERE, base::Bind(
 768       &VTVideoDecodeAccelerator::DecodeDone, weak_this_, frame));
 769 }
 770
 771 void VTVideoDecodeAccelerator::DecodeDone(Frame* frame) {
 772   DCHECK(gpu_thread_checker_.CalledOnValidThread());
 773   DCHECK_EQ(1u, pending_frames_.count(frame->bitstream_id));
 774   Task task(TASK_FRAME);
 775   task.frame = pending_frames_[frame->bitstream_id];
 776   pending_frames_.erase(frame->bitstream_id);
 777   task_queue_.push(task);
 778   ProcessWorkQueues();
 779 }
 780
 781 void VTVideoDecodeAccelerator::FlushTask(TaskType type) {
 782   DCHECK(decoder_thread_.task_runner()->BelongsToCurrentThread());
 783   FinishDelayedFrames();
 784
 785   // Always queue a task, even if FinishDelayedFrames() fails, so that
 786   // destruction always completes.
 787   gpu_task_runner_->PostTask(FROM_HERE, base::Bind(
 788       &VTVideoDecodeAccelerator::FlushDone, weak_this_, type));
 789 }
 790
 791 void VTVideoDecodeAccelerator::FlushDone(TaskType type) {
 792   DCHECK(gpu_thread_checker_.CalledOnValidThread());
 793   task_queue_.push(Task(type));
 794   ProcessWorkQueues();
 795 }
 796
 797 void VTVideoDecodeAccelerator::Decode(const media::BitstreamBuffer& bitstream) {
 798   DCHECK(gpu_thread_checker_.CalledOnValidThread());
 799   DCHECK_EQ(0u, assigned_bitstream_ids_.count(bitstream.id()));
 800   assigned_bitstream_ids_.insert(bitstream.id());
 801   Frame* frame = new Frame(bitstream.id());
 802   pending_frames_[frame->bitstream_id] = make_linked_ptr(frame);
 803   decoder_thread_.task_runner()->PostTask(
 804       FROM_HERE, base::Bind(&VTVideoDecodeAccelerator::DecodeTask,
 805                             base::Unretained(this), bitstream, frame));
 806 }
 807
 808 void VTVideoDecodeAccelerator::AssignPictureBuffers(
 809     const std::vector<media::PictureBuffer>& pictures) {
 810   DCHECK(gpu_thread_checker_.CalledOnValidThread());
 811
 812   for (const media::PictureBuffer& picture : pictures) {
 813     DCHECK(!texture_ids_.count(picture.id()));
 814     assigned_picture_ids_.insert(picture.id());
 815     available_picture_ids_.push_back(picture.id());
 816     texture_ids_[picture.id()] = picture.texture_id();
 817   }
 818
 819   // Pictures are not marked as uncleared until after this method returns, and
 820   // they will be broken if they are used before that happens. So, schedule
 821   // future work after that happens.
 822   gpu_task_runner_->PostTask(FROM_HERE, base::Bind(
 823       &VTVideoDecodeAccelerator::ProcessWorkQueues, weak_this_));
 824 }
 825
 826 void VTVideoDecodeAccelerator::ReusePictureBuffer(int32_t picture_id) {
 827   DCHECK(gpu_thread_checker_.CalledOnValidThread());
 828   DCHECK_EQ(CFGetRetainCount(picture_bindings_[picture_id]), 1);
 829   picture_bindings_.erase(picture_id);
 830   if (assigned_picture_ids_.count(picture_id) != 0) {
 831     available_picture_ids_.push_back(picture_id);
 832     ProcessWorkQueues();
 833   } else {
 834     client_->DismissPictureBuffer(picture_id);
 835   }
 836 }
 837
 838 void VTVideoDecodeAccelerator::ProcessWorkQueues() {
 839   DCHECK(gpu_thread_checker_.CalledOnValidThread());
 840   switch (state_) {
 841     case STATE_DECODING:
 842       // TODO(sandersd): Batch where possible.
 843       while (state_ == STATE_DECODING) {
 844         if (!ProcessReorderQueue() && !ProcessTaskQueue())
 845           break;
 846       }
 847       return;
 848
 849     case STATE_ERROR:
 850       // Do nothing until Destroy() is called.
 851       return;
 852
 853     case STATE_DESTROYING:
 854       // Drop tasks until we are ready to destruct.
 855       while (!task_queue_.empty()) {
 856         if (task_queue_.front().type == TASK_DESTROY) {
 857           delete this;
 858           return;
 859         }
 860         task_queue_.pop();
 861       }
 862       return;
 863   }
 864 }
 865
 866 bool VTVideoDecodeAccelerator::ProcessTaskQueue() {
 867   DCHECK(gpu_thread_checker_.CalledOnValidThread());
 868   DCHECK_EQ(state_, STATE_DECODING);
 869
 870   if (task_queue_.empty())
 871     return false;
 872
 873   const Task& task = task_queue_.front();
 874   switch (task.type) {
 875     case TASK_FRAME:
 876       // TODO(sandersd): Signal IDR explicitly (not using pic_order_cnt == 0).
 877       if (reorder_queue_.size() < kMaxReorderQueueSize &&
 878           (task.frame->pic_order_cnt != 0 || reorder_queue_.empty())) {
 879         assigned_bitstream_ids_.erase(task.frame->bitstream_id);
 880         client_->NotifyEndOfBitstreamBuffer(task.frame->bitstream_id);
 881         reorder_queue_.push(task.frame);
 882         task_queue_.pop();
 883         return true;
 884       }
 885       return false;
 886
 887     case TASK_FLUSH:
 888       DCHECK_EQ(task.type, pending_flush_tasks_.front());
 889       if (reorder_queue_.size() == 0) {
 890         pending_flush_tasks_.pop();
 891         client_->NotifyFlushDone();
 892         task_queue_.pop();
 893         return true;
 894       }
 895       return false;
 896
 897     case TASK_RESET:
 898       DCHECK_EQ(task.type, pending_flush_tasks_.front());
 899       if (reorder_queue_.size() == 0) {
 900         last_sps_id_ = -1;
 901         last_pps_id_ = -1;
 902         last_sps_.clear();
 903         last_spsext_.clear();
 904         last_pps_.clear();
 905         poc_.Reset();
 906         pending_flush_tasks_.pop();
 907         client_->NotifyResetDone();
 908         task_queue_.pop();
 909         return true;
 910       }
 911       return false;
 912
 913     case TASK_DESTROY:
 914       NOTREACHED() << "Can't destroy while in STATE_DECODING.";
 915       NotifyError(ILLEGAL_STATE, SFT_PLATFORM_ERROR);
 916       return false;
 917   }
 918 }
 919
 920 bool VTVideoDecodeAccelerator::ProcessReorderQueue() {
 921   DCHECK(gpu_thread_checker_.CalledOnValidThread());
 922   DCHECK_EQ(state_, STATE_DECODING);
 923
 924   if (reorder_queue_.empty())
 925     return false;
 926
 927   // If the next task is a flush (because there is a pending flush or becuase
 928   // the next frame is an IDR), then we don't need a full reorder buffer to send
 929   // the next frame.
 930   bool flushing = !task_queue_.empty() &&
 931                   (task_queue_.front().type != TASK_FRAME ||
 932                    task_queue_.front().frame->pic_order_cnt == 0);
 933
 934   size_t reorder_window = std::max(0, reorder_queue_.top()->reorder_window);
 935   if (flushing || reorder_queue_.size() > reorder_window) {
 936     if (ProcessFrame(*reorder_queue_.top())) {
 937       reorder_queue_.pop();
 938       return true;
 939     }
 940   }
 941
 942   return false;
 943 }
 944
 945 bool VTVideoDecodeAccelerator::ProcessFrame(const Frame& frame) {
 946   DCHECK(gpu_thread_checker_.CalledOnValidThread());
 947   DCHECK_EQ(state_, STATE_DECODING);
 948
 949   // If the next pending flush is for a reset, then the frame will be dropped.
 950   bool resetting = !pending_flush_tasks_.empty() &&
 951                    pending_flush_tasks_.front() == TASK_RESET;
 952
 953   if (!resetting && frame.image.get()) {
 954     // If the |coded_size| has changed, request new picture buffers and then
 955     // wait for them.
 956     // TODO(sandersd): If GpuVideoDecoder didn't specifically check the size of
 957     // textures, this would be unnecessary, as the size is actually a property
 958     // of the texture binding, not the texture. We rebind every frame, so the
 959     // size passed to ProvidePictureBuffers() is meaningless.
 960     if (picture_size_ != frame.coded_size) {
 961       // Dismiss current pictures.
 962       for (int32_t picture_id : assigned_picture_ids_)
 963         client_->DismissPictureBuffer(picture_id);
 964       assigned_picture_ids_.clear();
 965       available_picture_ids_.clear();
 966
 967       // Request new pictures.
 968       picture_size_ = frame.coded_size;
 969       client_->ProvidePictureBuffers(
 970           kNumPictureBuffers, coded_size_, GL_TEXTURE_RECTANGLE_ARB);
 971       return false;
 972     }
 973     if (!SendFrame(frame))
 974       return false;
 975   }
 976
 977   return true;
 978 }
 979
 980 bool VTVideoDecodeAccelerator::SendFrame(const Frame& frame) {
 981   DCHECK(gpu_thread_checker_.CalledOnValidThread());
 982   DCHECK_EQ(state_, STATE_DECODING);
 983
 984   if (available_picture_ids_.empty())
 985     return false;
 986
 987   int32_t picture_id = available_picture_ids_.back();
 988   IOSurfaceRef surface = CVPixelBufferGetIOSurface(frame.image.get());
 989
 990   if (!make_context_current_.Run()) {
 991     DLOG(ERROR) << "Failed to make GL context current";
 992     NotifyError(PLATFORM_FAILURE, SFT_PLATFORM_ERROR);
 993     return false;
 994   }
 995
 996   glEnable(GL_TEXTURE_RECTANGLE_ARB);
 997   gfx::ScopedTextureBinder
 998       texture_binder(GL_TEXTURE_RECTANGLE_ARB, texture_ids_[picture_id]);
 999   CGLContextObj cgl_context =
1000       static_cast<CGLContextObj>(gfx::GLContext::GetCurrent()->GetHandle());
1001   CGLError status = CGLTexImageIOSurface2D(
1002       cgl_context,                  // ctx
1003       GL_TEXTURE_RECTANGLE_ARB,     // target
1004       GL_RGB,                       // internal_format
1005       frame.coded_size.width(),     // width
1006       frame.coded_size.height(),    // height
1007       GL_YCBCR_422_APPLE,           // format
1008       GL_UNSIGNED_SHORT_8_8_APPLE,  // type
1009       surface,                      // io_surface
1010       0);                           // plane
1011   glDisable(GL_TEXTURE_RECTANGLE_ARB);
1012   if (status != kCGLNoError) {
1013     NOTIFY_STATUS("CGLTexImageIOSurface2D()", status, SFT_PLATFORM_ERROR);
1014     return false;
1015   }
1016
1017   available_picture_ids_.pop_back();
1018   picture_bindings_[picture_id] = frame.image;
1019   client_->PictureReady(media::Picture(picture_id, frame.bitstream_id,
1020                                        gfx::Rect(frame.coded_size), false));
1021   return true;
1022 }
1023
1024 void VTVideoDecodeAccelerator::NotifyError(
1025     Error vda_error_type,
1026     VTVDASessionFailureType session_failure_type) {
1027   DCHECK_LT(session_failure_type, SFT_MAX + 1);
1028   if (!gpu_thread_checker_.CalledOnValidThread()) {
1029     gpu_task_runner_->PostTask(FROM_HERE, base::Bind(
1030         &VTVideoDecodeAccelerator::NotifyError, weak_this_, vda_error_type,
1031         session_failure_type));
1032   } else if (state_ == STATE_DECODING) {
1033     state_ = STATE_ERROR;
1034     UMA_HISTOGRAM_ENUMERATION("Media.VTVDA.SessionFailureReason",
1035                               session_failure_type,
1036                               SFT_MAX + 1);
1037     client_->NotifyError(vda_error_type);
1038   }
1039 }
1040
1041 void VTVideoDecodeAccelerator::QueueFlush(TaskType type) {
1042   DCHECK(gpu_thread_checker_.CalledOnValidThread());
1043   pending_flush_tasks_.push(type);
1044   decoder_thread_.task_runner()->PostTask(
1045       FROM_HERE, base::Bind(&VTVideoDecodeAccelerator::FlushTask,
1046                             base::Unretained(this), type));
1047
1048   // If this is a new flush request, see if we can make progress.
1049   if (pending_flush_tasks_.size() == 1)
1050     ProcessWorkQueues();
1051 }
1052
1053 void VTVideoDecodeAccelerator::Flush() {
1054   DCHECK(gpu_thread_checker_.CalledOnValidThread());
1055   QueueFlush(TASK_FLUSH);
1056 }
1057
1058 void VTVideoDecodeAccelerator::Reset() {
1059   DCHECK(gpu_thread_checker_.CalledOnValidThread());
1060   QueueFlush(TASK_RESET);
1061 }
1062
1063 void VTVideoDecodeAccelerator::Destroy() {
1064   DCHECK(gpu_thread_checker_.CalledOnValidThread());
1065
1066   // In a forceful shutdown, the decoder thread may be dead already.
1067   if (!decoder_thread_.IsRunning()) {
1068     delete this;
1069     return;
1070   }
1071
1072   // For a graceful shutdown, return assigned buffers and flush before
1073   // destructing |this|.
1074   // TODO(sandersd): Make sure the decoder won't try to read the buffers again
1075   // before discarding them.
1076   for (int32_t bitstream_id : assigned_bitstream_ids_)
1077     client_->NotifyEndOfBitstreamBuffer(bitstream_id);
1078   assigned_bitstream_ids_.clear();
1079   state_ = STATE_DESTROYING;
1080   QueueFlush(TASK_DESTROY);
1081 }
1082
1083 bool VTVideoDecodeAccelerator::CanDecodeOnIOThread() {
1084   return false;
1085 }
1086
1087 // static
1088 media::VideoDecodeAccelerator::SupportedProfiles
1089 VTVideoDecodeAccelerator::GetSupportedProfiles() {
1090   SupportedProfiles profiles;
1091   for (const auto& supported_profile : kSupportedProfiles) {
1092     SupportedProfile profile;
1093     profile.profile = supported_profile;
1094     profile.min_resolution.SetSize(16, 16);
1095     profile.max_resolution.SetSize(4096, 2160);
1096     profiles.push_back(profile);
1097   }
1098   return profiles;
1099 }
1100
1101 }  // namespace content