Merge Chromium + Blink git repositories
[chromium-blink-merge.git] / content / common / gpu / media / vt_video_decode_accelerator.cc
blobc98853dd6b863ebbba3e390dc731f814dad927a6
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.
5 #include <algorithm>
7 #include <CoreVideo/CoreVideo.h>
8 #include <OpenGL/CGLIOSurface.h>
9 #include <OpenGL/gl.h>
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/gl_image_io_surface.h"
25 #include "ui/gl/scoped_binders.h"
27 using content_common_gpu_media::kModuleVt;
28 using content_common_gpu_media::InitializeStubs;
29 using content_common_gpu_media::IsVtInitialized;
30 using content_common_gpu_media::StubPathMap;
32 #define NOTIFY_STATUS(name, status, session_failure) \
33 do { \
34 OSSTATUS_DLOG(ERROR, status) << name; \
35 NotifyError(PLATFORM_FAILURE, session_failure); \
36 } while (0)
38 namespace content {
40 // Only H.264 with 4:2:0 chroma sampling is supported.
41 static const media::VideoCodecProfile kSupportedProfiles[] = {
42 media::H264PROFILE_BASELINE,
43 media::H264PROFILE_MAIN,
44 media::H264PROFILE_EXTENDED,
45 media::H264PROFILE_HIGH,
46 media::H264PROFILE_HIGH10PROFILE,
47 media::H264PROFILE_SCALABLEBASELINE,
48 media::H264PROFILE_SCALABLEHIGH,
49 media::H264PROFILE_STEREOHIGH,
50 media::H264PROFILE_MULTIVIEWHIGH,
53 // Size to use for NALU length headers in AVC format (can be 1, 2, or 4).
54 static const int kNALUHeaderLength = 4;
56 // We request 5 picture buffers from the client, each of which has a texture ID
57 // that we can bind decoded frames to. We need enough to satisfy preroll, and
58 // enough to avoid unnecessary stalling, but no more than that. The resource
59 // requirements are low, as we don't need the textures to be backed by storage.
60 static const int kNumPictureBuffers = media::limits::kMaxVideoFrames + 1;
62 // Maximum number of frames to queue for reordering before we stop asking for
63 // more. (NotifyEndOfBitstreamBuffer() is called when frames are moved into the
64 // reorder queue.)
65 static const int kMaxReorderQueueSize = 16;
67 // When set to false, always create a new decoder instead of reusing the
68 // existing configuration when the configuration changes. This works around a
69 // bug in VideoToolbox that results in corruption before Mac OS X 10.10.3. The
70 // value is set in InitializeVideoToolbox().
71 static bool g_enable_compatible_configuration_reuse = true;
73 // Build an |image_config| dictionary for VideoToolbox initialization.
74 static base::ScopedCFTypeRef<CFMutableDictionaryRef>
75 BuildImageConfig(CMVideoDimensions coded_dimensions) {
76 base::ScopedCFTypeRef<CFMutableDictionaryRef> image_config;
78 // 4:2:2 is used over the native 4:2:0 because only 4:2:2 can be directly
79 // bound to a texture by CGLTexImageIOSurface2D().
80 int32_t pixel_format = kCVPixelFormatType_422YpCbCr8;
81 #define CFINT(i) CFNumberCreate(kCFAllocatorDefault, kCFNumberSInt32Type, &i)
82 base::ScopedCFTypeRef<CFNumberRef> cf_pixel_format(CFINT(pixel_format));
83 base::ScopedCFTypeRef<CFNumberRef> cf_width(CFINT(coded_dimensions.width));
84 base::ScopedCFTypeRef<CFNumberRef> cf_height(CFINT(coded_dimensions.height));
85 #undef CFINT
86 if (!cf_pixel_format.get() || !cf_width.get() || !cf_height.get())
87 return image_config;
89 image_config.reset(
90 CFDictionaryCreateMutable(
91 kCFAllocatorDefault,
92 4, // capacity
93 &kCFTypeDictionaryKeyCallBacks,
94 &kCFTypeDictionaryValueCallBacks));
95 if (!image_config.get())
96 return image_config;
98 CFDictionarySetValue(image_config, kCVPixelBufferPixelFormatTypeKey,
99 cf_pixel_format);
100 CFDictionarySetValue(image_config, kCVPixelBufferWidthKey, cf_width);
101 CFDictionarySetValue(image_config, kCVPixelBufferHeightKey, cf_height);
102 CFDictionarySetValue(image_config, kCVPixelBufferOpenGLCompatibilityKey,
103 kCFBooleanTrue);
105 return image_config;
108 // Create a VTDecompressionSession using the provided |pps| and |sps|. If
109 // |require_hardware| is true, the session must uses real hardware decoding
110 // (as opposed to software decoding inside of VideoToolbox) to be considered
111 // successful.
113 // TODO(sandersd): Merge with ConfigureDecoder(), as the code is very similar.
114 static bool CreateVideoToolboxSession(const uint8_t* sps, size_t sps_size,
115 const uint8_t* pps, size_t pps_size,
116 bool require_hardware) {
117 const uint8_t* data_ptrs[] = {sps, pps};
118 const size_t data_sizes[] = {sps_size, pps_size};
120 base::ScopedCFTypeRef<CMFormatDescriptionRef> format;
121 OSStatus status = CMVideoFormatDescriptionCreateFromH264ParameterSets(
122 kCFAllocatorDefault,
123 2, // parameter_set_count
124 data_ptrs, // &parameter_set_pointers
125 data_sizes, // &parameter_set_sizes
126 kNALUHeaderLength, // nal_unit_header_length
127 format.InitializeInto());
128 if (status) {
129 OSSTATUS_DLOG(WARNING, status)
130 << "Failed to create CMVideoFormatDescription.";
131 return false;
134 base::ScopedCFTypeRef<CFMutableDictionaryRef> decoder_config(
135 CFDictionaryCreateMutable(
136 kCFAllocatorDefault,
137 1, // capacity
138 &kCFTypeDictionaryKeyCallBacks,
139 &kCFTypeDictionaryValueCallBacks));
140 if (!decoder_config.get())
141 return false;
143 if (require_hardware) {
144 CFDictionarySetValue(
145 decoder_config,
146 // kVTVideoDecoderSpecification_RequireHardwareAcceleratedVideoDecoder
147 CFSTR("RequireHardwareAcceleratedVideoDecoder"),
148 kCFBooleanTrue);
151 base::ScopedCFTypeRef<CFMutableDictionaryRef> image_config(
152 BuildImageConfig(CMVideoFormatDescriptionGetDimensions(format)));
153 if (!image_config.get())
154 return false;
156 VTDecompressionOutputCallbackRecord callback = {0};
158 base::ScopedCFTypeRef<VTDecompressionSessionRef> session;
159 status = VTDecompressionSessionCreate(
160 kCFAllocatorDefault,
161 format, // video_format_description
162 decoder_config, // video_decoder_specification
163 image_config, // destination_image_buffer_attributes
164 &callback, // output_callback
165 session.InitializeInto());
166 if (status) {
167 OSSTATUS_DLOG(WARNING, status) << "Failed to create VTDecompressionSession";
168 return false;
171 return true;
174 // The purpose of this function is to preload the generic and hardware-specific
175 // libraries required by VideoToolbox before the GPU sandbox is enabled.
176 // VideoToolbox normally loads the hardware-specific libraries lazily, so we
177 // must actually create a decompression session. If creating a decompression
178 // session fails, hardware decoding will be disabled (Initialize() will always
179 // return false).
180 static bool InitializeVideoToolboxInternal() {
181 if (base::CommandLine::ForCurrentProcess()->HasSwitch(
182 switches::kDisableAcceleratedVideoDecode)) {
183 return false;
186 if (!IsVtInitialized()) {
187 // CoreVideo is also required, but the loader stops after the first path is
188 // loaded. Instead we rely on the transitive dependency from VideoToolbox to
189 // CoreVideo.
190 // TODO(sandersd): Fallback to PrivateFrameworks to support OS X < 10.8.
191 StubPathMap paths;
192 paths[kModuleVt].push_back(FILE_PATH_LITERAL(
193 "/System/Library/Frameworks/VideoToolbox.framework/VideoToolbox"));
194 if (!InitializeStubs(paths)) {
195 LOG(WARNING) << "Failed to initialize VideoToolbox framework. "
196 << "Hardware accelerated video decoding will be disabled.";
197 return false;
201 // Create a hardware decoding session.
202 // SPS and PPS data are taken from a 480p sample (buck2.mp4).
203 const uint8_t sps_normal[] = {0x67, 0x64, 0x00, 0x1e, 0xac, 0xd9, 0x80, 0xd4,
204 0x3d, 0xa1, 0x00, 0x00, 0x03, 0x00, 0x01, 0x00,
205 0x00, 0x03, 0x00, 0x30, 0x8f, 0x16, 0x2d, 0x9a};
206 const uint8_t pps_normal[] = {0x68, 0xe9, 0x7b, 0xcb};
207 if (!CreateVideoToolboxSession(sps_normal, arraysize(sps_normal), pps_normal,
208 arraysize(pps_normal), true)) {
209 LOG(WARNING) << "Failed to create hardware VideoToolbox session. "
210 << "Hardware accelerated video decoding will be disabled.";
211 return false;
214 // Create a software decoding session.
215 // SPS and PPS data are taken from a 18p sample (small2.mp4).
216 const uint8_t sps_small[] = {0x67, 0x64, 0x00, 0x0a, 0xac, 0xd9, 0x89, 0x7e,
217 0x22, 0x10, 0x00, 0x00, 0x3e, 0x90, 0x00, 0x0e,
218 0xa6, 0x08, 0xf1, 0x22, 0x59, 0xa0};
219 const uint8_t pps_small[] = {0x68, 0xe9, 0x79, 0x72, 0xc0};
220 if (!CreateVideoToolboxSession(sps_small, arraysize(sps_small), pps_small,
221 arraysize(pps_small), false)) {
222 LOG(WARNING) << "Failed to create software VideoToolbox session. "
223 << "Hardware accelerated video decoding will be disabled.";
224 return false;
227 // Set |g_enable_compatible_configuration_reuse| to false on
228 // Mac OS X < 10.10.3.
229 base::Version os_x_version(base::SysInfo::OperatingSystemVersion());
230 if (os_x_version.IsOlderThan("10.10.3"))
231 g_enable_compatible_configuration_reuse = false;
233 return true;
236 bool InitializeVideoToolbox() {
237 // InitializeVideoToolbox() is called only from the GPU process main thread;
238 // once for sandbox warmup, and then once each time a VTVideoDecodeAccelerator
239 // is initialized.
240 static bool attempted = false;
241 static bool succeeded = false;
243 if (!attempted) {
244 attempted = true;
245 succeeded = InitializeVideoToolboxInternal();
248 return succeeded;
251 // Route decoded frame callbacks back into the VTVideoDecodeAccelerator.
252 static void OutputThunk(
253 void* decompression_output_refcon,
254 void* source_frame_refcon,
255 OSStatus status,
256 VTDecodeInfoFlags info_flags,
257 CVImageBufferRef image_buffer,
258 CMTime presentation_time_stamp,
259 CMTime presentation_duration) {
260 VTVideoDecodeAccelerator* vda =
261 reinterpret_cast<VTVideoDecodeAccelerator*>(decompression_output_refcon);
262 vda->Output(source_frame_refcon, status, image_buffer);
265 VTVideoDecodeAccelerator::Task::Task(TaskType type) : type(type) {
268 VTVideoDecodeAccelerator::Task::~Task() {
271 VTVideoDecodeAccelerator::Frame::Frame(int32_t bitstream_id)
272 : bitstream_id(bitstream_id), pic_order_cnt(0), reorder_window(0) {
275 VTVideoDecodeAccelerator::Frame::~Frame() {
278 VTVideoDecodeAccelerator::PictureInfo::PictureInfo(uint32_t client_texture_id,
279 uint32_t service_texture_id)
280 : client_texture_id(client_texture_id),
281 service_texture_id(service_texture_id) {}
283 VTVideoDecodeAccelerator::PictureInfo::~PictureInfo() {
284 if (gl_image)
285 gl_image->Destroy(false);
288 bool VTVideoDecodeAccelerator::FrameOrder::operator()(
289 const linked_ptr<Frame>& lhs,
290 const linked_ptr<Frame>& rhs) const {
291 if (lhs->pic_order_cnt != rhs->pic_order_cnt)
292 return lhs->pic_order_cnt > rhs->pic_order_cnt;
293 // If |pic_order_cnt| is the same, fall back on using the bitstream order.
294 // TODO(sandersd): Assign a sequence number in Decode() and use that instead.
295 // TODO(sandersd): Using the sequence number, ensure that frames older than
296 // |kMaxReorderQueueSize| are ordered first, regardless of |pic_order_cnt|.
297 return lhs->bitstream_id > rhs->bitstream_id;
300 VTVideoDecodeAccelerator::VTVideoDecodeAccelerator(
301 const base::Callback<bool(void)>& make_context_current,
302 const base::Callback<void(uint32, uint32, scoped_refptr<gfx::GLImage>)>&
303 bind_image)
304 : make_context_current_(make_context_current),
305 bind_image_(bind_image),
306 client_(nullptr),
307 state_(STATE_DECODING),
308 format_(nullptr),
309 session_(nullptr),
310 last_sps_id_(-1),
311 last_pps_id_(-1),
312 gpu_task_runner_(base::ThreadTaskRunnerHandle::Get()),
313 decoder_thread_("VTDecoderThread"),
314 weak_this_factory_(this) {
315 DCHECK(!make_context_current_.is_null());
316 callback_.decompressionOutputCallback = OutputThunk;
317 callback_.decompressionOutputRefCon = this;
318 weak_this_ = weak_this_factory_.GetWeakPtr();
321 VTVideoDecodeAccelerator::~VTVideoDecodeAccelerator() {
322 DCHECK(gpu_thread_checker_.CalledOnValidThread());
325 bool VTVideoDecodeAccelerator::Initialize(
326 media::VideoCodecProfile profile,
327 Client* client) {
328 DCHECK(gpu_thread_checker_.CalledOnValidThread());
329 client_ = client;
331 if (!InitializeVideoToolbox())
332 return false;
334 bool profile_supported = false;
335 for (const auto& supported_profile : kSupportedProfiles) {
336 if (profile == supported_profile) {
337 profile_supported = true;
338 break;
341 if (!profile_supported)
342 return false;
344 // Spawn a thread to handle parsing and calling VideoToolbox.
345 if (!decoder_thread_.Start())
346 return false;
348 // Count the session as successfully initialized.
349 UMA_HISTOGRAM_ENUMERATION("Media.VTVDA.SessionFailureReason",
350 SFT_SUCCESSFULLY_INITIALIZED,
351 SFT_MAX + 1);
352 return true;
355 bool VTVideoDecodeAccelerator::FinishDelayedFrames() {
356 DCHECK(decoder_thread_.task_runner()->BelongsToCurrentThread());
357 if (session_) {
358 OSStatus status = VTDecompressionSessionWaitForAsynchronousFrames(session_);
359 if (status) {
360 NOTIFY_STATUS("VTDecompressionSessionWaitForAsynchronousFrames()",
361 status, SFT_PLATFORM_ERROR);
362 return false;
365 return true;
368 bool VTVideoDecodeAccelerator::ConfigureDecoder() {
369 DCHECK(decoder_thread_.task_runner()->BelongsToCurrentThread());
370 DCHECK(!last_sps_.empty());
371 DCHECK(!last_pps_.empty());
373 // Build the configuration records.
374 std::vector<const uint8_t*> nalu_data_ptrs;
375 std::vector<size_t> nalu_data_sizes;
376 nalu_data_ptrs.reserve(3);
377 nalu_data_sizes.reserve(3);
378 nalu_data_ptrs.push_back(&last_sps_.front());
379 nalu_data_sizes.push_back(last_sps_.size());
380 if (!last_spsext_.empty()) {
381 nalu_data_ptrs.push_back(&last_spsext_.front());
382 nalu_data_sizes.push_back(last_spsext_.size());
384 nalu_data_ptrs.push_back(&last_pps_.front());
385 nalu_data_sizes.push_back(last_pps_.size());
387 // Construct a new format description from the parameter sets.
388 // TODO(sandersd): Replace this with custom code to support OS X < 10.9.
389 format_.reset();
390 OSStatus status = CMVideoFormatDescriptionCreateFromH264ParameterSets(
391 kCFAllocatorDefault,
392 nalu_data_ptrs.size(), // parameter_set_count
393 &nalu_data_ptrs.front(), // &parameter_set_pointers
394 &nalu_data_sizes.front(), // &parameter_set_sizes
395 kNALUHeaderLength, // nal_unit_header_length
396 format_.InitializeInto());
397 if (status) {
398 NOTIFY_STATUS("CMVideoFormatDescriptionCreateFromH264ParameterSets()",
399 status, SFT_PLATFORM_ERROR);
400 return false;
403 // Store the new configuration data.
404 CMVideoDimensions coded_dimensions =
405 CMVideoFormatDescriptionGetDimensions(format_);
406 coded_size_.SetSize(coded_dimensions.width, coded_dimensions.height);
408 // If the session is compatible, there's nothing else to do.
409 if (g_enable_compatible_configuration_reuse && session_ &&
410 VTDecompressionSessionCanAcceptFormatDescription(session_, format_)) {
411 return true;
414 // Prepare VideoToolbox configuration dictionaries.
415 base::ScopedCFTypeRef<CFMutableDictionaryRef> decoder_config(
416 CFDictionaryCreateMutable(
417 kCFAllocatorDefault,
418 1, // capacity
419 &kCFTypeDictionaryKeyCallBacks,
420 &kCFTypeDictionaryValueCallBacks));
421 if (!decoder_config.get()) {
422 DLOG(ERROR) << "Failed to create CFMutableDictionary.";
423 NotifyError(PLATFORM_FAILURE, SFT_PLATFORM_ERROR);
424 return false;
427 CFDictionarySetValue(
428 decoder_config,
429 // kVTVideoDecoderSpecification_EnableHardwareAcceleratedVideoDecoder
430 CFSTR("EnableHardwareAcceleratedVideoDecoder"),
431 kCFBooleanTrue);
433 base::ScopedCFTypeRef<CFMutableDictionaryRef> image_config(
434 BuildImageConfig(coded_dimensions));
435 if (!image_config.get()) {
436 DLOG(ERROR) << "Failed to create decoder image configuration.";
437 NotifyError(PLATFORM_FAILURE, SFT_PLATFORM_ERROR);
438 return false;
441 // Ensure that the old decoder emits all frames before the new decoder can
442 // emit any.
443 if (!FinishDelayedFrames())
444 return false;
446 session_.reset();
447 status = VTDecompressionSessionCreate(
448 kCFAllocatorDefault,
449 format_, // video_format_description
450 decoder_config, // video_decoder_specification
451 image_config, // destination_image_buffer_attributes
452 &callback_, // output_callback
453 session_.InitializeInto());
454 if (status) {
455 NOTIFY_STATUS("VTDecompressionSessionCreate()", status,
456 SFT_UNSUPPORTED_STREAM_PARAMETERS);
457 return false;
460 // Report whether hardware decode is being used.
461 bool using_hardware = false;
462 base::ScopedCFTypeRef<CFBooleanRef> cf_using_hardware;
463 if (VTSessionCopyProperty(
464 session_,
465 // kVTDecompressionPropertyKey_UsingHardwareAcceleratedVideoDecoder
466 CFSTR("UsingHardwareAcceleratedVideoDecoder"),
467 kCFAllocatorDefault,
468 cf_using_hardware.InitializeInto()) == 0) {
469 using_hardware = CFBooleanGetValue(cf_using_hardware);
471 UMA_HISTOGRAM_BOOLEAN("Media.VTVDA.HardwareAccelerated", using_hardware);
473 return true;
476 void VTVideoDecodeAccelerator::DecodeTask(
477 const media::BitstreamBuffer& bitstream,
478 Frame* frame) {
479 DCHECK(decoder_thread_.task_runner()->BelongsToCurrentThread());
481 // Map the bitstream buffer.
482 base::SharedMemory memory(bitstream.handle(), true);
483 size_t size = bitstream.size();
484 if (!memory.Map(size)) {
485 DLOG(ERROR) << "Failed to map bitstream buffer";
486 NotifyError(PLATFORM_FAILURE, SFT_PLATFORM_ERROR);
487 return;
489 const uint8_t* buf = static_cast<uint8_t*>(memory.memory());
491 // NALUs are stored with Annex B format in the bitstream buffer (start codes),
492 // but VideoToolbox expects AVC format (length headers), so we must rewrite
493 // the data.
495 // Locate relevant NALUs and compute the size of the rewritten data. Also
496 // record any parameter sets for VideoToolbox initialization.
497 std::vector<uint8_t> sps;
498 std::vector<uint8_t> spsext;
499 std::vector<uint8_t> pps;
500 bool has_slice = false;
501 size_t data_size = 0;
502 std::vector<media::H264NALU> nalus;
503 parser_.SetStream(buf, size);
504 media::H264NALU nalu;
505 while (true) {
506 media::H264Parser::Result result = parser_.AdvanceToNextNALU(&nalu);
507 if (result == media::H264Parser::kEOStream)
508 break;
509 if (result == media::H264Parser::kUnsupportedStream) {
510 DLOG(ERROR) << "Unsupported H.264 stream";
511 NotifyError(PLATFORM_FAILURE, SFT_UNSUPPORTED_STREAM);
512 return;
514 if (result != media::H264Parser::kOk) {
515 DLOG(ERROR) << "Failed to parse H.264 stream";
516 NotifyError(UNREADABLE_INPUT, SFT_INVALID_STREAM);
517 return;
519 switch (nalu.nal_unit_type) {
520 case media::H264NALU::kSPS:
521 result = parser_.ParseSPS(&last_sps_id_);
522 if (result == media::H264Parser::kUnsupportedStream) {
523 DLOG(ERROR) << "Unsupported SPS";
524 NotifyError(PLATFORM_FAILURE, SFT_UNSUPPORTED_STREAM);
525 return;
527 if (result != media::H264Parser::kOk) {
528 DLOG(ERROR) << "Could not parse SPS";
529 NotifyError(UNREADABLE_INPUT, SFT_INVALID_STREAM);
530 return;
532 sps.assign(nalu.data, nalu.data + nalu.size);
533 spsext.clear();
534 break;
536 case media::H264NALU::kSPSExt:
537 // TODO(sandersd): Check that the previous NALU was an SPS.
538 spsext.assign(nalu.data, nalu.data + nalu.size);
539 break;
541 case media::H264NALU::kPPS:
542 result = parser_.ParsePPS(&last_pps_id_);
543 if (result == media::H264Parser::kUnsupportedStream) {
544 DLOG(ERROR) << "Unsupported PPS";
545 NotifyError(PLATFORM_FAILURE, SFT_UNSUPPORTED_STREAM);
546 return;
548 if (result != media::H264Parser::kOk) {
549 DLOG(ERROR) << "Could not parse PPS";
550 NotifyError(UNREADABLE_INPUT, SFT_INVALID_STREAM);
551 return;
553 pps.assign(nalu.data, nalu.data + nalu.size);
554 break;
556 case media::H264NALU::kSliceDataA:
557 case media::H264NALU::kSliceDataB:
558 case media::H264NALU::kSliceDataC:
559 case media::H264NALU::kNonIDRSlice:
560 // TODO(sandersd): Check that there has been an IDR slice since the
561 // last reset.
562 case media::H264NALU::kIDRSlice:
563 // Compute the |pic_order_cnt| for the picture from the first slice.
564 // TODO(sandersd): Make sure that any further slices are part of the
565 // same picture or a redundant coded picture.
566 if (!has_slice) {
567 media::H264SliceHeader slice_hdr;
568 result = parser_.ParseSliceHeader(nalu, &slice_hdr);
569 if (result == media::H264Parser::kUnsupportedStream) {
570 DLOG(ERROR) << "Unsupported slice header";
571 NotifyError(PLATFORM_FAILURE, SFT_UNSUPPORTED_STREAM);
572 return;
574 if (result != media::H264Parser::kOk) {
575 DLOG(ERROR) << "Could not parse slice header";
576 NotifyError(UNREADABLE_INPUT, SFT_INVALID_STREAM);
577 return;
580 // TODO(sandersd): Maintain a cache of configurations and reconfigure
581 // when a slice references a new config.
582 DCHECK_EQ(slice_hdr.pic_parameter_set_id, last_pps_id_);
583 const media::H264PPS* pps =
584 parser_.GetPPS(slice_hdr.pic_parameter_set_id);
585 if (!pps) {
586 DLOG(ERROR) << "Mising PPS referenced by slice";
587 NotifyError(UNREADABLE_INPUT, SFT_INVALID_STREAM);
588 return;
591 DCHECK_EQ(pps->seq_parameter_set_id, last_sps_id_);
592 const media::H264SPS* sps = parser_.GetSPS(pps->seq_parameter_set_id);
593 if (!sps) {
594 DLOG(ERROR) << "Mising SPS referenced by PPS";
595 NotifyError(UNREADABLE_INPUT, SFT_INVALID_STREAM);
596 return;
599 if (!poc_.ComputePicOrderCnt(sps, slice_hdr, &frame->pic_order_cnt)) {
600 DLOG(ERROR) << "Unable to compute POC";
601 NotifyError(UNREADABLE_INPUT, SFT_INVALID_STREAM);
602 return;
605 if (sps->vui_parameters_present_flag &&
606 sps->bitstream_restriction_flag) {
607 frame->reorder_window = std::min(sps->max_num_reorder_frames,
608 kMaxReorderQueueSize - 1);
611 has_slice = true;
612 default:
613 nalus.push_back(nalu);
614 data_size += kNALUHeaderLength + nalu.size;
615 break;
619 // Initialize VideoToolbox.
620 bool config_changed = false;
621 if (!sps.empty() && sps != last_sps_) {
622 last_sps_.swap(sps);
623 last_spsext_.swap(spsext);
624 config_changed = true;
626 if (!pps.empty() && pps != last_pps_) {
627 last_pps_.swap(pps);
628 config_changed = true;
630 if (config_changed) {
631 if (last_sps_.empty()) {
632 DLOG(ERROR) << "Invalid configuration; no SPS";
633 NotifyError(INVALID_ARGUMENT, SFT_INVALID_STREAM);
634 return;
636 if (last_pps_.empty()) {
637 DLOG(ERROR) << "Invalid configuration; no PPS";
638 NotifyError(INVALID_ARGUMENT, SFT_INVALID_STREAM);
639 return;
641 if (!ConfigureDecoder())
642 return;
645 // If there are no image slices, drop the bitstream buffer by returning an
646 // empty frame.
647 if (!has_slice) {
648 if (!FinishDelayedFrames())
649 return;
650 gpu_task_runner_->PostTask(FROM_HERE, base::Bind(
651 &VTVideoDecodeAccelerator::DecodeDone, weak_this_, frame));
652 return;
655 // If the session is not configured by this point, fail.
656 if (!session_) {
657 DLOG(ERROR) << "Cannot decode without configuration";
658 NotifyError(INVALID_ARGUMENT, SFT_INVALID_STREAM);
659 return;
662 // Update the frame metadata with configuration data.
663 frame->coded_size = coded_size_;
665 // Create a memory-backed CMBlockBuffer for the translated data.
666 // TODO(sandersd): Pool of memory blocks.
667 base::ScopedCFTypeRef<CMBlockBufferRef> data;
668 OSStatus status = CMBlockBufferCreateWithMemoryBlock(
669 kCFAllocatorDefault,
670 nullptr, // &memory_block
671 data_size, // block_length
672 kCFAllocatorDefault, // block_allocator
673 nullptr, // &custom_block_source
674 0, // offset_to_data
675 data_size, // data_length
676 0, // flags
677 data.InitializeInto());
678 if (status) {
679 NOTIFY_STATUS("CMBlockBufferCreateWithMemoryBlock()", status,
680 SFT_PLATFORM_ERROR);
681 return;
684 // Make sure that the memory is actually allocated.
685 // CMBlockBufferReplaceDataBytes() is documented to do this, but prints a
686 // message each time starting in Mac OS X 10.10.
687 status = CMBlockBufferAssureBlockMemory(data);
688 if (status) {
689 NOTIFY_STATUS("CMBlockBufferAssureBlockMemory()", status,
690 SFT_PLATFORM_ERROR);
691 return;
694 // Copy NALU data into the CMBlockBuffer, inserting length headers.
695 size_t offset = 0;
696 for (size_t i = 0; i < nalus.size(); i++) {
697 media::H264NALU& nalu = nalus[i];
698 uint32_t header = base::HostToNet32(static_cast<uint32_t>(nalu.size));
699 status = CMBlockBufferReplaceDataBytes(
700 &header, data, offset, kNALUHeaderLength);
701 if (status) {
702 NOTIFY_STATUS("CMBlockBufferReplaceDataBytes()", status,
703 SFT_PLATFORM_ERROR);
704 return;
706 offset += kNALUHeaderLength;
707 status = CMBlockBufferReplaceDataBytes(nalu.data, data, offset, nalu.size);
708 if (status) {
709 NOTIFY_STATUS("CMBlockBufferReplaceDataBytes()", status,
710 SFT_PLATFORM_ERROR);
711 return;
713 offset += nalu.size;
716 // Package the data in a CMSampleBuffer.
717 base::ScopedCFTypeRef<CMSampleBufferRef> sample;
718 status = CMSampleBufferCreate(
719 kCFAllocatorDefault,
720 data, // data_buffer
721 true, // data_ready
722 nullptr, // make_data_ready_callback
723 nullptr, // make_data_ready_refcon
724 format_, // format_description
725 1, // num_samples
726 0, // num_sample_timing_entries
727 nullptr, // &sample_timing_array
728 1, // num_sample_size_entries
729 &data_size, // &sample_size_array
730 sample.InitializeInto());
731 if (status) {
732 NOTIFY_STATUS("CMSampleBufferCreate()", status, SFT_PLATFORM_ERROR);
733 return;
736 // Send the frame for decoding.
737 // Asynchronous Decompression allows for parallel submission of frames
738 // (without it, DecodeFrame() does not return until the frame has been
739 // decoded). We don't enable Temporal Processing so that frames are always
740 // returned in decode order; this makes it easier to avoid deadlock.
741 VTDecodeFrameFlags decode_flags =
742 kVTDecodeFrame_EnableAsynchronousDecompression;
743 status = VTDecompressionSessionDecodeFrame(
744 session_,
745 sample, // sample_buffer
746 decode_flags, // decode_flags
747 reinterpret_cast<void*>(frame), // source_frame_refcon
748 nullptr); // &info_flags_out
749 if (status) {
750 NOTIFY_STATUS("VTDecompressionSessionDecodeFrame()", status,
751 SFT_DECODE_ERROR);
752 return;
756 // This method may be called on any VideoToolbox thread.
757 void VTVideoDecodeAccelerator::Output(
758 void* source_frame_refcon,
759 OSStatus status,
760 CVImageBufferRef image_buffer) {
761 if (status) {
762 NOTIFY_STATUS("Decoding", status, SFT_DECODE_ERROR);
763 return;
766 // The type of |image_buffer| is CVImageBuffer, but we only handle
767 // CVPixelBuffers. This should be guaranteed as we set
768 // kCVPixelBufferOpenGLCompatibilityKey in |image_config|.
770 // Sometimes, for unknown reasons (http://crbug.com/453050), |image_buffer| is
771 // NULL, which causes CFGetTypeID() to crash. While the rest of the code would
772 // smoothly handle NULL as a dropped frame, we choose to fail permanantly here
773 // until the issue is better understood.
774 if (!image_buffer || CFGetTypeID(image_buffer) != CVPixelBufferGetTypeID()) {
775 DLOG(ERROR) << "Decoded frame is not a CVPixelBuffer";
776 NotifyError(PLATFORM_FAILURE, SFT_DECODE_ERROR);
777 return;
780 Frame* frame = reinterpret_cast<Frame*>(source_frame_refcon);
781 frame->image.reset(image_buffer, base::scoped_policy::RETAIN);
782 gpu_task_runner_->PostTask(FROM_HERE, base::Bind(
783 &VTVideoDecodeAccelerator::DecodeDone, weak_this_, frame));
786 void VTVideoDecodeAccelerator::DecodeDone(Frame* frame) {
787 DCHECK(gpu_thread_checker_.CalledOnValidThread());
788 DCHECK_EQ(1u, pending_frames_.count(frame->bitstream_id));
789 Task task(TASK_FRAME);
790 task.frame = pending_frames_[frame->bitstream_id];
791 pending_frames_.erase(frame->bitstream_id);
792 task_queue_.push(task);
793 ProcessWorkQueues();
796 void VTVideoDecodeAccelerator::FlushTask(TaskType type) {
797 DCHECK(decoder_thread_.task_runner()->BelongsToCurrentThread());
798 FinishDelayedFrames();
800 // Always queue a task, even if FinishDelayedFrames() fails, so that
801 // destruction always completes.
802 gpu_task_runner_->PostTask(FROM_HERE, base::Bind(
803 &VTVideoDecodeAccelerator::FlushDone, weak_this_, type));
806 void VTVideoDecodeAccelerator::FlushDone(TaskType type) {
807 DCHECK(gpu_thread_checker_.CalledOnValidThread());
808 task_queue_.push(Task(type));
809 ProcessWorkQueues();
812 void VTVideoDecodeAccelerator::Decode(const media::BitstreamBuffer& bitstream) {
813 DCHECK(gpu_thread_checker_.CalledOnValidThread());
814 DCHECK_EQ(0u, assigned_bitstream_ids_.count(bitstream.id()));
815 assigned_bitstream_ids_.insert(bitstream.id());
816 Frame* frame = new Frame(bitstream.id());
817 pending_frames_[frame->bitstream_id] = make_linked_ptr(frame);
818 decoder_thread_.task_runner()->PostTask(
819 FROM_HERE, base::Bind(&VTVideoDecodeAccelerator::DecodeTask,
820 base::Unretained(this), bitstream, frame));
823 void VTVideoDecodeAccelerator::AssignPictureBuffers(
824 const std::vector<media::PictureBuffer>& pictures) {
825 DCHECK(gpu_thread_checker_.CalledOnValidThread());
827 for (const media::PictureBuffer& picture : pictures) {
828 DCHECK(!picture_info_map_.count(picture.id()));
829 assigned_picture_ids_.insert(picture.id());
830 available_picture_ids_.push_back(picture.id());
831 picture_info_map_.insert(picture.id(), make_scoped_ptr(new PictureInfo(
832 picture.internal_texture_id(),
833 picture.texture_id())));
836 // Pictures are not marked as uncleared until after this method returns, and
837 // they will be broken if they are used before that happens. So, schedule
838 // future work after that happens.
839 gpu_task_runner_->PostTask(FROM_HERE, base::Bind(
840 &VTVideoDecodeAccelerator::ProcessWorkQueues, weak_this_));
843 void VTVideoDecodeAccelerator::ReusePictureBuffer(int32_t picture_id) {
844 DCHECK(gpu_thread_checker_.CalledOnValidThread());
845 DCHECK(picture_info_map_.count(picture_id));
846 PictureInfo* picture_info = picture_info_map_.find(picture_id)->second;
847 DCHECK_EQ(CFGetRetainCount(picture_info->cv_image), 1);
848 picture_info->cv_image.reset();
849 picture_info->gl_image->Destroy(false);
850 picture_info->gl_image = nullptr;
852 if (assigned_picture_ids_.count(picture_id) != 0) {
853 available_picture_ids_.push_back(picture_id);
854 ProcessWorkQueues();
855 } else {
856 client_->DismissPictureBuffer(picture_id);
860 void VTVideoDecodeAccelerator::ProcessWorkQueues() {
861 DCHECK(gpu_thread_checker_.CalledOnValidThread());
862 switch (state_) {
863 case STATE_DECODING:
864 // TODO(sandersd): Batch where possible.
865 while (state_ == STATE_DECODING) {
866 if (!ProcessReorderQueue() && !ProcessTaskQueue())
867 break;
869 return;
871 case STATE_ERROR:
872 // Do nothing until Destroy() is called.
873 return;
875 case STATE_DESTROYING:
876 // Drop tasks until we are ready to destruct.
877 while (!task_queue_.empty()) {
878 if (task_queue_.front().type == TASK_DESTROY) {
879 delete this;
880 return;
882 task_queue_.pop();
884 return;
888 bool VTVideoDecodeAccelerator::ProcessTaskQueue() {
889 DCHECK(gpu_thread_checker_.CalledOnValidThread());
890 DCHECK_EQ(state_, STATE_DECODING);
892 if (task_queue_.empty())
893 return false;
895 const Task& task = task_queue_.front();
896 switch (task.type) {
897 case TASK_FRAME:
898 // TODO(sandersd): Signal IDR explicitly (not using pic_order_cnt == 0).
899 if (reorder_queue_.size() < kMaxReorderQueueSize &&
900 (task.frame->pic_order_cnt != 0 || reorder_queue_.empty())) {
901 assigned_bitstream_ids_.erase(task.frame->bitstream_id);
902 client_->NotifyEndOfBitstreamBuffer(task.frame->bitstream_id);
903 reorder_queue_.push(task.frame);
904 task_queue_.pop();
905 return true;
907 return false;
909 case TASK_FLUSH:
910 DCHECK_EQ(task.type, pending_flush_tasks_.front());
911 if (reorder_queue_.size() == 0) {
912 pending_flush_tasks_.pop();
913 client_->NotifyFlushDone();
914 task_queue_.pop();
915 return true;
917 return false;
919 case TASK_RESET:
920 DCHECK_EQ(task.type, pending_flush_tasks_.front());
921 if (reorder_queue_.size() == 0) {
922 last_sps_id_ = -1;
923 last_pps_id_ = -1;
924 last_sps_.clear();
925 last_spsext_.clear();
926 last_pps_.clear();
927 poc_.Reset();
928 pending_flush_tasks_.pop();
929 client_->NotifyResetDone();
930 task_queue_.pop();
931 return true;
933 return false;
935 case TASK_DESTROY:
936 NOTREACHED() << "Can't destroy while in STATE_DECODING.";
937 NotifyError(ILLEGAL_STATE, SFT_PLATFORM_ERROR);
938 return false;
942 bool VTVideoDecodeAccelerator::ProcessReorderQueue() {
943 DCHECK(gpu_thread_checker_.CalledOnValidThread());
944 DCHECK_EQ(state_, STATE_DECODING);
946 if (reorder_queue_.empty())
947 return false;
949 // If the next task is a flush (because there is a pending flush or becuase
950 // the next frame is an IDR), then we don't need a full reorder buffer to send
951 // the next frame.
952 bool flushing = !task_queue_.empty() &&
953 (task_queue_.front().type != TASK_FRAME ||
954 task_queue_.front().frame->pic_order_cnt == 0);
956 size_t reorder_window = std::max(0, reorder_queue_.top()->reorder_window);
957 if (flushing || reorder_queue_.size() > reorder_window) {
958 if (ProcessFrame(*reorder_queue_.top())) {
959 reorder_queue_.pop();
960 return true;
964 return false;
967 bool VTVideoDecodeAccelerator::ProcessFrame(const Frame& frame) {
968 DCHECK(gpu_thread_checker_.CalledOnValidThread());
969 DCHECK_EQ(state_, STATE_DECODING);
971 // If the next pending flush is for a reset, then the frame will be dropped.
972 bool resetting = !pending_flush_tasks_.empty() &&
973 pending_flush_tasks_.front() == TASK_RESET;
975 if (!resetting && frame.image.get()) {
976 // If the |coded_size| has changed, request new picture buffers and then
977 // wait for them.
978 // TODO(sandersd): If GpuVideoDecoder didn't specifically check the size of
979 // textures, this would be unnecessary, as the size is actually a property
980 // of the texture binding, not the texture. We rebind every frame, so the
981 // size passed to ProvidePictureBuffers() is meaningless.
982 if (picture_size_ != frame.coded_size) {
983 // Dismiss current pictures.
984 for (int32_t picture_id : assigned_picture_ids_)
985 client_->DismissPictureBuffer(picture_id);
986 assigned_picture_ids_.clear();
987 available_picture_ids_.clear();
989 // Request new pictures.
990 picture_size_ = frame.coded_size;
991 client_->ProvidePictureBuffers(
992 kNumPictureBuffers, coded_size_, GL_TEXTURE_RECTANGLE_ARB);
993 return false;
995 if (!SendFrame(frame))
996 return false;
999 return true;
1002 bool VTVideoDecodeAccelerator::SendFrame(const Frame& frame) {
1003 DCHECK(gpu_thread_checker_.CalledOnValidThread());
1004 DCHECK_EQ(state_, STATE_DECODING);
1006 if (available_picture_ids_.empty())
1007 return false;
1009 int32_t picture_id = available_picture_ids_.back();
1010 DCHECK(picture_info_map_.count(picture_id));
1011 PictureInfo* picture_info = picture_info_map_.find(picture_id)->second;
1012 DCHECK(!picture_info->cv_image);
1013 DCHECK(!picture_info->gl_image);
1015 if (!make_context_current_.Run()) {
1016 DLOG(ERROR) << "Failed to make GL context current";
1017 NotifyError(PLATFORM_FAILURE, SFT_PLATFORM_ERROR);
1018 return false;
1021 IOSurfaceRef surface = CVPixelBufferGetIOSurface(frame.image.get());
1022 glEnable(GL_TEXTURE_RECTANGLE_ARB);
1023 gfx::ScopedTextureBinder texture_binder(GL_TEXTURE_RECTANGLE_ARB,
1024 picture_info->service_texture_id);
1025 CGLContextObj cgl_context =
1026 static_cast<CGLContextObj>(gfx::GLContext::GetCurrent()->GetHandle());
1027 CGLError status = CGLTexImageIOSurface2D(
1028 cgl_context, // ctx
1029 GL_TEXTURE_RECTANGLE_ARB, // target
1030 GL_RGB, // internal_format
1031 frame.coded_size.width(), // width
1032 frame.coded_size.height(), // height
1033 GL_YCBCR_422_APPLE, // format
1034 GL_UNSIGNED_SHORT_8_8_APPLE, // type
1035 surface, // io_surface
1036 0); // plane
1037 glDisable(GL_TEXTURE_RECTANGLE_ARB);
1038 if (status != kCGLNoError) {
1039 NOTIFY_STATUS("CGLTexImageIOSurface2D()", status, SFT_PLATFORM_ERROR);
1040 return false;
1043 bool allow_overlay = false;
1044 scoped_refptr<gfx::GLImageIOSurface> gl_image(new gfx::GLImageIOSurface(
1045 gfx::GenericSharedMemoryId(), frame.coded_size, GL_BGRA_EXT));
1046 if (gl_image->Initialize(surface, gfx::BufferFormat::BGRA_8888)) {
1047 allow_overlay = true;
1048 } else {
1049 gl_image = nullptr;
1051 bind_image_.Run(picture_info->client_texture_id, GL_TEXTURE_RECTANGLE_ARB,
1052 gl_image);
1054 // Assign the new image(s) to the the picture info.
1055 picture_info->gl_image = gl_image;
1056 picture_info->cv_image = frame.image;
1057 available_picture_ids_.pop_back();
1059 // TODO(sandersd): Currently, the size got from
1060 // CMVideoFormatDescriptionGetDimensions is visible size. We pass it to
1061 // GpuVideoDecoder so that GpuVideoDecoder can use correct visible size in
1062 // resolution changed. We should find the correct API to get the real
1063 // coded size and fix it.
1064 client_->PictureReady(media::Picture(picture_id, frame.bitstream_id,
1065 gfx::Rect(frame.coded_size),
1066 allow_overlay));
1067 return true;
1070 void VTVideoDecodeAccelerator::NotifyError(
1071 Error vda_error_type,
1072 VTVDASessionFailureType session_failure_type) {
1073 DCHECK_LT(session_failure_type, SFT_MAX + 1);
1074 if (!gpu_thread_checker_.CalledOnValidThread()) {
1075 gpu_task_runner_->PostTask(FROM_HERE, base::Bind(
1076 &VTVideoDecodeAccelerator::NotifyError, weak_this_, vda_error_type,
1077 session_failure_type));
1078 } else if (state_ == STATE_DECODING) {
1079 state_ = STATE_ERROR;
1080 UMA_HISTOGRAM_ENUMERATION("Media.VTVDA.SessionFailureReason",
1081 session_failure_type,
1082 SFT_MAX + 1);
1083 client_->NotifyError(vda_error_type);
1087 void VTVideoDecodeAccelerator::QueueFlush(TaskType type) {
1088 DCHECK(gpu_thread_checker_.CalledOnValidThread());
1089 pending_flush_tasks_.push(type);
1090 decoder_thread_.task_runner()->PostTask(
1091 FROM_HERE, base::Bind(&VTVideoDecodeAccelerator::FlushTask,
1092 base::Unretained(this), type));
1094 // If this is a new flush request, see if we can make progress.
1095 if (pending_flush_tasks_.size() == 1)
1096 ProcessWorkQueues();
1099 void VTVideoDecodeAccelerator::Flush() {
1100 DCHECK(gpu_thread_checker_.CalledOnValidThread());
1101 QueueFlush(TASK_FLUSH);
1104 void VTVideoDecodeAccelerator::Reset() {
1105 DCHECK(gpu_thread_checker_.CalledOnValidThread());
1106 QueueFlush(TASK_RESET);
1109 void VTVideoDecodeAccelerator::Destroy() {
1110 DCHECK(gpu_thread_checker_.CalledOnValidThread());
1112 // In a forceful shutdown, the decoder thread may be dead already.
1113 if (!decoder_thread_.IsRunning()) {
1114 delete this;
1115 return;
1118 // For a graceful shutdown, return assigned buffers and flush before
1119 // destructing |this|.
1120 // TODO(sandersd): Make sure the decoder won't try to read the buffers again
1121 // before discarding them.
1122 for (int32_t bitstream_id : assigned_bitstream_ids_)
1123 client_->NotifyEndOfBitstreamBuffer(bitstream_id);
1124 assigned_bitstream_ids_.clear();
1125 state_ = STATE_DESTROYING;
1126 QueueFlush(TASK_DESTROY);
1129 bool VTVideoDecodeAccelerator::CanDecodeOnIOThread() {
1130 return false;
1133 // static
1134 media::VideoDecodeAccelerator::SupportedProfiles
1135 VTVideoDecodeAccelerator::GetSupportedProfiles() {
1136 SupportedProfiles profiles;
1137 for (const auto& supported_profile : kSupportedProfiles) {
1138 SupportedProfile profile;
1139 profile.profile = supported_profile;
1140 profile.min_resolution.SetSize(16, 16);
1141 profile.max_resolution.SetSize(4096, 2160);
1142 profiles.push_back(profile);
1144 return profiles;
1147 } // namespace content