Chromium Code Reviews
chromiumcodereview-hr@appspot.gserviceaccount.com (chromiumcodereview-hr) | Please choose your nickname with Settings | Help | Chromium Project | Gerrit Changes | Sign out
(206)

Side by Side Diff: content/common/gpu/media/vt_video_decode_accelerator.cc

Issue 1474823002: content/common/gpu/media add platform suffix to some files (cleanup) (Closed) Base URL: https://chromium.googlesource.com/chromium/src.git@master
Patch Set: Reverted naming video_encode_accelerator_unittest.cc; A few content/content_tests.gypi alphabetical… Created 5 years ago
Use n/p to move between diff chunks; N/P to move between comments. Draft comments are only viewable by you.
Jump to:
View unified diff | Download patch
OLDNEW
(Empty)
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/gl_image_io_surface.h"
25 #include "ui/gl/gl_implementation.h"
26 #include "ui/gl/scoped_binders.h"
27
28 using content_common_gpu_media::kModuleVt;
29 using content_common_gpu_media::InitializeStubs;
30 using content_common_gpu_media::IsVtInitialized;
31 using content_common_gpu_media::StubPathMap;
32
33 #define NOTIFY_STATUS(name, status, session_failure) \
34 do { \
35 OSSTATUS_DLOG(ERROR, status) << name; \
36 NotifyError(PLATFORM_FAILURE, session_failure); \
37 } while (0)
38
39 namespace content {
40
41 // Only H.264 with 4:2:0 chroma sampling is supported.
42 static const media::VideoCodecProfile kSupportedProfiles[] = {
43 media::H264PROFILE_BASELINE,
44 media::H264PROFILE_MAIN,
45 media::H264PROFILE_EXTENDED,
46 media::H264PROFILE_HIGH,
47 media::H264PROFILE_HIGH10PROFILE,
48 media::H264PROFILE_SCALABLEBASELINE,
49 media::H264PROFILE_SCALABLEHIGH,
50 media::H264PROFILE_STEREOHIGH,
51 media::H264PROFILE_MULTIVIEWHIGH,
52 };
53
54 // Size to use for NALU length headers in AVC format (can be 1, 2, or 4).
55 static const int kNALUHeaderLength = 4;
56
57 // We request 5 picture buffers from the client, each of which has a texture ID
58 // that we can bind decoded frames to. We need enough to satisfy preroll, and
59 // enough to avoid unnecessary stalling, but no more than that. The resource
60 // requirements are low, as we don't need the textures to be backed by storage.
61 static const int kNumPictureBuffers = media::limits::kMaxVideoFrames + 1;
62
63 // Maximum number of frames to queue for reordering before we stop asking for
64 // more. (NotifyEndOfBitstreamBuffer() is called when frames are moved into the
65 // reorder queue.)
66 static const int kMaxReorderQueueSize = 16;
67
68 // Build an |image_config| dictionary for VideoToolbox initialization.
69 static base::ScopedCFTypeRef<CFMutableDictionaryRef>
70 BuildImageConfig(CMVideoDimensions coded_dimensions) {
71 base::ScopedCFTypeRef<CFMutableDictionaryRef> image_config;
72
73 // 4:2:2 is used over the native 4:2:0 because only 4:2:2 can be directly
74 // bound to a texture by CGLTexImageIOSurface2D().
75 int32_t pixel_format = kCVPixelFormatType_422YpCbCr8;
76 #define CFINT(i) CFNumberCreate(kCFAllocatorDefault, kCFNumberSInt32Type, &i)
77 base::ScopedCFTypeRef<CFNumberRef> cf_pixel_format(CFINT(pixel_format));
78 base::ScopedCFTypeRef<CFNumberRef> cf_width(CFINT(coded_dimensions.width));
79 base::ScopedCFTypeRef<CFNumberRef> cf_height(CFINT(coded_dimensions.height));
80 #undef CFINT
81 if (!cf_pixel_format.get() || !cf_width.get() || !cf_height.get())
82 return image_config;
83
84 image_config.reset(
85 CFDictionaryCreateMutable(
86 kCFAllocatorDefault,
87 4, // capacity
88 &kCFTypeDictionaryKeyCallBacks,
89 &kCFTypeDictionaryValueCallBacks));
90 if (!image_config.get())
91 return image_config;
92
93 CFDictionarySetValue(image_config, kCVPixelBufferPixelFormatTypeKey,
94 cf_pixel_format);
95 CFDictionarySetValue(image_config, kCVPixelBufferWidthKey, cf_width);
96 CFDictionarySetValue(image_config, kCVPixelBufferHeightKey, cf_height);
97 CFDictionarySetValue(image_config, kCVPixelBufferOpenGLCompatibilityKey,
98 kCFBooleanTrue);
99
100 return image_config;
101 }
102
103 // Create a VTDecompressionSession using the provided |pps| and |sps|. If
104 // |require_hardware| is true, the session must uses real hardware decoding
105 // (as opposed to software decoding inside of VideoToolbox) to be considered
106 // successful.
107 //
108 // TODO(sandersd): Merge with ConfigureDecoder(), as the code is very similar.
109 static bool CreateVideoToolboxSession(const uint8_t* sps, size_t sps_size,
110 const uint8_t* pps, size_t pps_size,
111 bool require_hardware) {
112 const uint8_t* data_ptrs[] = {sps, pps};
113 const size_t data_sizes[] = {sps_size, pps_size};
114
115 base::ScopedCFTypeRef<CMFormatDescriptionRef> format;
116 OSStatus status = CMVideoFormatDescriptionCreateFromH264ParameterSets(
117 kCFAllocatorDefault,
118 2, // parameter_set_count
119 data_ptrs, // &parameter_set_pointers
120 data_sizes, // &parameter_set_sizes
121 kNALUHeaderLength, // nal_unit_header_length
122 format.InitializeInto());
123 if (status) {
124 OSSTATUS_DLOG(WARNING, status)
125 << "Failed to create CMVideoFormatDescription.";
126 return false;
127 }
128
129 base::ScopedCFTypeRef<CFMutableDictionaryRef> decoder_config(
130 CFDictionaryCreateMutable(
131 kCFAllocatorDefault,
132 1, // capacity
133 &kCFTypeDictionaryKeyCallBacks,
134 &kCFTypeDictionaryValueCallBacks));
135 if (!decoder_config.get())
136 return false;
137
138 if (require_hardware) {
139 CFDictionarySetValue(
140 decoder_config,
141 // kVTVideoDecoderSpecification_RequireHardwareAcceleratedVideoDecoder
142 CFSTR("RequireHardwareAcceleratedVideoDecoder"),
143 kCFBooleanTrue);
144 }
145
146 base::ScopedCFTypeRef<CFMutableDictionaryRef> image_config(
147 BuildImageConfig(CMVideoFormatDescriptionGetDimensions(format)));
148 if (!image_config.get())
149 return false;
150
151 VTDecompressionOutputCallbackRecord callback = {0};
152
153 base::ScopedCFTypeRef<VTDecompressionSessionRef> session;
154 status = VTDecompressionSessionCreate(
155 kCFAllocatorDefault,
156 format, // video_format_description
157 decoder_config, // video_decoder_specification
158 image_config, // destination_image_buffer_attributes
159 &callback, // output_callback
160 session.InitializeInto());
161 if (status) {
162 OSSTATUS_DLOG(WARNING, status) << "Failed to create VTDecompressionSession";
163 return false;
164 }
165
166 return true;
167 }
168
169 // The purpose of this function is to preload the generic and hardware-specific
170 // libraries required by VideoToolbox before the GPU sandbox is enabled.
171 // VideoToolbox normally loads the hardware-specific libraries lazily, so we
172 // must actually create a decompression session. If creating a decompression
173 // session fails, hardware decoding will be disabled (Initialize() will always
174 // return false).
175 static bool InitializeVideoToolboxInternal() {
176 if (base::CommandLine::ForCurrentProcess()->HasSwitch(
177 switches::kDisableAcceleratedVideoDecode)) {
178 return false;
179 }
180
181 if (!IsVtInitialized()) {
182 // CoreVideo is also required, but the loader stops after the first path is
183 // loaded. Instead we rely on the transitive dependency from VideoToolbox to
184 // CoreVideo.
185 StubPathMap paths;
186 paths[kModuleVt].push_back(FILE_PATH_LITERAL(
187 "/System/Library/Frameworks/VideoToolbox.framework/VideoToolbox"));
188 if (!InitializeStubs(paths)) {
189 LOG(WARNING) << "Failed to initialize VideoToolbox framework. "
190 << "Hardware accelerated video decoding will be disabled.";
191 return false;
192 }
193 }
194
195 // Create a hardware decoding session.
196 // SPS and PPS data are taken from a 480p sample (buck2.mp4).
197 const uint8_t sps_normal[] = {0x67, 0x64, 0x00, 0x1e, 0xac, 0xd9, 0x80, 0xd4,
198 0x3d, 0xa1, 0x00, 0x00, 0x03, 0x00, 0x01, 0x00,
199 0x00, 0x03, 0x00, 0x30, 0x8f, 0x16, 0x2d, 0x9a};
200 const uint8_t pps_normal[] = {0x68, 0xe9, 0x7b, 0xcb};
201 if (!CreateVideoToolboxSession(sps_normal, arraysize(sps_normal), pps_normal,
202 arraysize(pps_normal), true)) {
203 LOG(WARNING) << "Failed to create hardware VideoToolbox session. "
204 << "Hardware accelerated video decoding will be disabled.";
205 return false;
206 }
207
208 // Create a software decoding session.
209 // SPS and PPS data are taken from a 18p sample (small2.mp4).
210 const uint8_t sps_small[] = {0x67, 0x64, 0x00, 0x0a, 0xac, 0xd9, 0x89, 0x7e,
211 0x22, 0x10, 0x00, 0x00, 0x3e, 0x90, 0x00, 0x0e,
212 0xa6, 0x08, 0xf1, 0x22, 0x59, 0xa0};
213 const uint8_t pps_small[] = {0x68, 0xe9, 0x79, 0x72, 0xc0};
214 if (!CreateVideoToolboxSession(sps_small, arraysize(sps_small), pps_small,
215 arraysize(pps_small), false)) {
216 LOG(WARNING) << "Failed to create software VideoToolbox session. "
217 << "Hardware accelerated video decoding will be disabled.";
218 return false;
219 }
220
221 return true;
222 }
223
224 bool InitializeVideoToolbox() {
225 // InitializeVideoToolbox() is called only from the GPU process main thread;
226 // once for sandbox warmup, and then once each time a VTVideoDecodeAccelerator
227 // is initialized.
228 static bool attempted = false;
229 static bool succeeded = false;
230
231 if (!attempted) {
232 attempted = true;
233 succeeded = InitializeVideoToolboxInternal();
234 }
235
236 return succeeded;
237 }
238
239 // Route decoded frame callbacks back into the VTVideoDecodeAccelerator.
240 static void OutputThunk(
241 void* decompression_output_refcon,
242 void* source_frame_refcon,
243 OSStatus status,
244 VTDecodeInfoFlags info_flags,
245 CVImageBufferRef image_buffer,
246 CMTime presentation_time_stamp,
247 CMTime presentation_duration) {
248 VTVideoDecodeAccelerator* vda =
249 reinterpret_cast<VTVideoDecodeAccelerator*>(decompression_output_refcon);
250 vda->Output(source_frame_refcon, status, image_buffer);
251 }
252
253 VTVideoDecodeAccelerator::Task::Task(TaskType type) : type(type) {
254 }
255
256 VTVideoDecodeAccelerator::Task::~Task() {
257 }
258
259 VTVideoDecodeAccelerator::Frame::Frame(int32_t bitstream_id)
260 : bitstream_id(bitstream_id),
261 pic_order_cnt(0),
262 is_idr(false),
263 reorder_window(0) {
264 }
265
266 VTVideoDecodeAccelerator::Frame::~Frame() {
267 }
268
269 VTVideoDecodeAccelerator::PictureInfo::PictureInfo(uint32_t client_texture_id,
270 uint32_t service_texture_id)
271 : client_texture_id(client_texture_id),
272 service_texture_id(service_texture_id) {}
273
274 VTVideoDecodeAccelerator::PictureInfo::~PictureInfo() {
275 if (gl_image)
276 gl_image->Destroy(false);
277 }
278
279 bool VTVideoDecodeAccelerator::FrameOrder::operator()(
280 const linked_ptr<Frame>& lhs,
281 const linked_ptr<Frame>& rhs) const {
282 if (lhs->pic_order_cnt != rhs->pic_order_cnt)
283 return lhs->pic_order_cnt > rhs->pic_order_cnt;
284 // If |pic_order_cnt| is the same, fall back on using the bitstream order.
285 // TODO(sandersd): Assign a sequence number in Decode() and use that instead.
286 // TODO(sandersd): Using the sequence number, ensure that frames older than
287 // |kMaxReorderQueueSize| are ordered first, regardless of |pic_order_cnt|.
288 return lhs->bitstream_id > rhs->bitstream_id;
289 }
290
291 VTVideoDecodeAccelerator::VTVideoDecodeAccelerator(
292 const base::Callback<bool(void)>& make_context_current,
293 const base::Callback<void(uint32, uint32, scoped_refptr<gl::GLImage>)>&
294 bind_image)
295 : make_context_current_(make_context_current),
296 bind_image_(bind_image),
297 client_(nullptr),
298 state_(STATE_DECODING),
299 format_(nullptr),
300 session_(nullptr),
301 last_sps_id_(-1),
302 last_pps_id_(-1),
303 gpu_task_runner_(base::ThreadTaskRunnerHandle::Get()),
304 decoder_thread_("VTDecoderThread"),
305 weak_this_factory_(this) {
306 DCHECK(!make_context_current_.is_null());
307 callback_.decompressionOutputCallback = OutputThunk;
308 callback_.decompressionOutputRefCon = this;
309 weak_this_ = weak_this_factory_.GetWeakPtr();
310 }
311
312 VTVideoDecodeAccelerator::~VTVideoDecodeAccelerator() {
313 DCHECK(gpu_thread_checker_.CalledOnValidThread());
314 }
315
316 bool VTVideoDecodeAccelerator::Initialize(
317 media::VideoCodecProfile profile,
318 Client* client) {
319 DCHECK(gpu_thread_checker_.CalledOnValidThread());
320 client_ = client;
321
322 if (!InitializeVideoToolbox())
323 return false;
324
325 bool profile_supported = false;
326 for (const auto& supported_profile : kSupportedProfiles) {
327 if (profile == supported_profile) {
328 profile_supported = true;
329 break;
330 }
331 }
332 if (!profile_supported)
333 return false;
334
335 // Spawn a thread to handle parsing and calling VideoToolbox.
336 if (!decoder_thread_.Start())
337 return false;
338
339 // Count the session as successfully initialized.
340 UMA_HISTOGRAM_ENUMERATION("Media.VTVDA.SessionFailureReason",
341 SFT_SUCCESSFULLY_INITIALIZED,
342 SFT_MAX + 1);
343 return true;
344 }
345
346 bool VTVideoDecodeAccelerator::FinishDelayedFrames() {
347 DCHECK(decoder_thread_.task_runner()->BelongsToCurrentThread());
348 if (session_) {
349 OSStatus status = VTDecompressionSessionWaitForAsynchronousFrames(session_);
350 if (status) {
351 NOTIFY_STATUS("VTDecompressionSessionWaitForAsynchronousFrames()",
352 status, SFT_PLATFORM_ERROR);
353 return false;
354 }
355 }
356 return true;
357 }
358
359 bool VTVideoDecodeAccelerator::ConfigureDecoder() {
360 DCHECK(decoder_thread_.task_runner()->BelongsToCurrentThread());
361 DCHECK(!last_sps_.empty());
362 DCHECK(!last_pps_.empty());
363
364 // Build the configuration records.
365 std::vector<const uint8_t*> nalu_data_ptrs;
366 std::vector<size_t> nalu_data_sizes;
367 nalu_data_ptrs.reserve(3);
368 nalu_data_sizes.reserve(3);
369 nalu_data_ptrs.push_back(&last_sps_.front());
370 nalu_data_sizes.push_back(last_sps_.size());
371 if (!last_spsext_.empty()) {
372 nalu_data_ptrs.push_back(&last_spsext_.front());
373 nalu_data_sizes.push_back(last_spsext_.size());
374 }
375 nalu_data_ptrs.push_back(&last_pps_.front());
376 nalu_data_sizes.push_back(last_pps_.size());
377
378 // Construct a new format description from the parameter sets.
379 format_.reset();
380 OSStatus status = CMVideoFormatDescriptionCreateFromH264ParameterSets(
381 kCFAllocatorDefault,
382 nalu_data_ptrs.size(), // parameter_set_count
383 &nalu_data_ptrs.front(), // &parameter_set_pointers
384 &nalu_data_sizes.front(), // &parameter_set_sizes
385 kNALUHeaderLength, // nal_unit_header_length
386 format_.InitializeInto());
387 if (status) {
388 NOTIFY_STATUS("CMVideoFormatDescriptionCreateFromH264ParameterSets()",
389 status, SFT_PLATFORM_ERROR);
390 return false;
391 }
392
393 // Store the new configuration data.
394 // TODO(sandersd): Despite the documentation, this seems to return the visible
395 // size. However, the output always appears to be top-left aligned, so it
396 // makes no difference. Re-verify this and update the variable name.
397 CMVideoDimensions coded_dimensions =
398 CMVideoFormatDescriptionGetDimensions(format_);
399 coded_size_.SetSize(coded_dimensions.width, coded_dimensions.height);
400
401 // Prepare VideoToolbox configuration dictionaries.
402 base::ScopedCFTypeRef<CFMutableDictionaryRef> decoder_config(
403 CFDictionaryCreateMutable(
404 kCFAllocatorDefault,
405 1, // capacity
406 &kCFTypeDictionaryKeyCallBacks,
407 &kCFTypeDictionaryValueCallBacks));
408 if (!decoder_config.get()) {
409 DLOG(ERROR) << "Failed to create CFMutableDictionary.";
410 NotifyError(PLATFORM_FAILURE, SFT_PLATFORM_ERROR);
411 return false;
412 }
413
414 CFDictionarySetValue(
415 decoder_config,
416 // kVTVideoDecoderSpecification_EnableHardwareAcceleratedVideoDecoder
417 CFSTR("EnableHardwareAcceleratedVideoDecoder"),
418 kCFBooleanTrue);
419
420 base::ScopedCFTypeRef<CFMutableDictionaryRef> image_config(
421 BuildImageConfig(coded_dimensions));
422 if (!image_config.get()) {
423 DLOG(ERROR) << "Failed to create decoder image configuration.";
424 NotifyError(PLATFORM_FAILURE, SFT_PLATFORM_ERROR);
425 return false;
426 }
427
428 // Ensure that the old decoder emits all frames before the new decoder can
429 // emit any.
430 if (!FinishDelayedFrames())
431 return false;
432
433 session_.reset();
434 status = VTDecompressionSessionCreate(
435 kCFAllocatorDefault,
436 format_, // video_format_description
437 decoder_config, // video_decoder_specification
438 image_config, // destination_image_buffer_attributes
439 &callback_, // output_callback
440 session_.InitializeInto());
441 if (status) {
442 NOTIFY_STATUS("VTDecompressionSessionCreate()", status,
443 SFT_UNSUPPORTED_STREAM_PARAMETERS);
444 return false;
445 }
446
447 // Report whether hardware decode is being used.
448 bool using_hardware = false;
449 base::ScopedCFTypeRef<CFBooleanRef> cf_using_hardware;
450 if (VTSessionCopyProperty(
451 session_,
452 // kVTDecompressionPropertyKey_UsingHardwareAcceleratedVideoDecoder
453 CFSTR("UsingHardwareAcceleratedVideoDecoder"),
454 kCFAllocatorDefault,
455 cf_using_hardware.InitializeInto()) == 0) {
456 using_hardware = CFBooleanGetValue(cf_using_hardware);
457 }
458 UMA_HISTOGRAM_BOOLEAN("Media.VTVDA.HardwareAccelerated", using_hardware);
459
460 return true;
461 }
462
463 void VTVideoDecodeAccelerator::DecodeTask(
464 const media::BitstreamBuffer& bitstream,
465 Frame* frame) {
466 DCHECK(decoder_thread_.task_runner()->BelongsToCurrentThread());
467
468 // Map the bitstream buffer.
469 base::SharedMemory memory(bitstream.handle(), true);
470 size_t size = bitstream.size();
471 if (!memory.Map(size)) {
472 DLOG(ERROR) << "Failed to map bitstream buffer";
473 NotifyError(PLATFORM_FAILURE, SFT_PLATFORM_ERROR);
474 return;
475 }
476 const uint8_t* buf = static_cast<uint8_t*>(memory.memory());
477
478 // NALUs are stored with Annex B format in the bitstream buffer (start codes),
479 // but VideoToolbox expects AVC format (length headers), so we must rewrite
480 // the data.
481 //
482 // Locate relevant NALUs and compute the size of the rewritten data. Also
483 // record any parameter sets for VideoToolbox initialization.
484 std::vector<uint8_t> sps;
485 std::vector<uint8_t> spsext;
486 std::vector<uint8_t> pps;
487 bool has_slice = false;
488 size_t data_size = 0;
489 std::vector<media::H264NALU> nalus;
490 parser_.SetStream(buf, size);
491 media::H264NALU nalu;
492 while (true) {
493 media::H264Parser::Result result = parser_.AdvanceToNextNALU(&nalu);
494 if (result == media::H264Parser::kEOStream)
495 break;
496 if (result == media::H264Parser::kUnsupportedStream) {
497 DLOG(ERROR) << "Unsupported H.264 stream";
498 NotifyError(PLATFORM_FAILURE, SFT_UNSUPPORTED_STREAM);
499 return;
500 }
501 if (result != media::H264Parser::kOk) {
502 DLOG(ERROR) << "Failed to parse H.264 stream";
503 NotifyError(UNREADABLE_INPUT, SFT_INVALID_STREAM);
504 return;
505 }
506 switch (nalu.nal_unit_type) {
507 case media::H264NALU::kSPS:
508 result = parser_.ParseSPS(&last_sps_id_);
509 if (result == media::H264Parser::kUnsupportedStream) {
510 DLOG(ERROR) << "Unsupported SPS";
511 NotifyError(PLATFORM_FAILURE, SFT_UNSUPPORTED_STREAM);
512 return;
513 }
514 if (result != media::H264Parser::kOk) {
515 DLOG(ERROR) << "Could not parse SPS";
516 NotifyError(UNREADABLE_INPUT, SFT_INVALID_STREAM);
517 return;
518 }
519 sps.assign(nalu.data, nalu.data + nalu.size);
520 spsext.clear();
521 break;
522
523 case media::H264NALU::kSPSExt:
524 // TODO(sandersd): Check that the previous NALU was an SPS.
525 spsext.assign(nalu.data, nalu.data + nalu.size);
526 break;
527
528 case media::H264NALU::kPPS:
529 result = parser_.ParsePPS(&last_pps_id_);
530 if (result == media::H264Parser::kUnsupportedStream) {
531 DLOG(ERROR) << "Unsupported PPS";
532 NotifyError(PLATFORM_FAILURE, SFT_UNSUPPORTED_STREAM);
533 return;
534 }
535 if (result != media::H264Parser::kOk) {
536 DLOG(ERROR) << "Could not parse PPS";
537 NotifyError(UNREADABLE_INPUT, SFT_INVALID_STREAM);
538 return;
539 }
540 pps.assign(nalu.data, nalu.data + nalu.size);
541 break;
542
543 case media::H264NALU::kSliceDataA:
544 case media::H264NALU::kSliceDataB:
545 case media::H264NALU::kSliceDataC:
546 case media::H264NALU::kNonIDRSlice:
547 case media::H264NALU::kIDRSlice:
548 // Compute the |pic_order_cnt| for the picture from the first slice.
549 if (!has_slice) {
550 media::H264SliceHeader slice_hdr;
551 result = parser_.ParseSliceHeader(nalu, &slice_hdr);
552 if (result == media::H264Parser::kUnsupportedStream) {
553 DLOG(ERROR) << "Unsupported slice header";
554 NotifyError(PLATFORM_FAILURE, SFT_UNSUPPORTED_STREAM);
555 return;
556 }
557 if (result != media::H264Parser::kOk) {
558 DLOG(ERROR) << "Could not parse slice header";
559 NotifyError(UNREADABLE_INPUT, SFT_INVALID_STREAM);
560 return;
561 }
562
563 // TODO(sandersd): Maintain a cache of configurations and reconfigure
564 // when a slice references a new config.
565 DCHECK_EQ(slice_hdr.pic_parameter_set_id, last_pps_id_);
566 const media::H264PPS* pps =
567 parser_.GetPPS(slice_hdr.pic_parameter_set_id);
568 if (!pps) {
569 DLOG(ERROR) << "Mising PPS referenced by slice";
570 NotifyError(UNREADABLE_INPUT, SFT_INVALID_STREAM);
571 return;
572 }
573
574 DCHECK_EQ(pps->seq_parameter_set_id, last_sps_id_);
575 const media::H264SPS* sps = parser_.GetSPS(pps->seq_parameter_set_id);
576 if (!sps) {
577 DLOG(ERROR) << "Mising SPS referenced by PPS";
578 NotifyError(UNREADABLE_INPUT, SFT_INVALID_STREAM);
579 return;
580 }
581
582 if (!poc_.ComputePicOrderCnt(sps, slice_hdr, &frame->pic_order_cnt)) {
583 DLOG(ERROR) << "Unable to compute POC";
584 NotifyError(UNREADABLE_INPUT, SFT_INVALID_STREAM);
585 return;
586 }
587
588 if (nalu.nal_unit_type == media::H264NALU::kIDRSlice)
589 frame->is_idr = true;
590
591 if (sps->vui_parameters_present_flag &&
592 sps->bitstream_restriction_flag) {
593 frame->reorder_window = std::min(sps->max_num_reorder_frames,
594 kMaxReorderQueueSize - 1);
595 }
596 }
597 has_slice = true;
598 default:
599 nalus.push_back(nalu);
600 data_size += kNALUHeaderLength + nalu.size;
601 break;
602 }
603 }
604
605 // Initialize VideoToolbox.
606 bool config_changed = false;
607 if (!sps.empty() && sps != last_sps_) {
608 last_sps_.swap(sps);
609 last_spsext_.swap(spsext);
610 config_changed = true;
611 }
612 if (!pps.empty() && pps != last_pps_) {
613 last_pps_.swap(pps);
614 config_changed = true;
615 }
616 if (config_changed) {
617 if (last_sps_.empty()) {
618 DLOG(ERROR) << "Invalid configuration; no SPS";
619 NotifyError(INVALID_ARGUMENT, SFT_INVALID_STREAM);
620 return;
621 }
622 if (last_pps_.empty()) {
623 DLOG(ERROR) << "Invalid configuration; no PPS";
624 NotifyError(INVALID_ARGUMENT, SFT_INVALID_STREAM);
625 return;
626 }
627
628 // If it's not an IDR frame, we can't reconfigure the decoder anyway. We
629 // assume that any config change not on an IDR must be compatible.
630 if (frame->is_idr && !ConfigureDecoder())
631 return;
632 }
633
634 // If there are no image slices, drop the bitstream buffer by returning an
635 // empty frame.
636 if (!has_slice) {
637 if (!FinishDelayedFrames())
638 return;
639 gpu_task_runner_->PostTask(FROM_HERE, base::Bind(
640 &VTVideoDecodeAccelerator::DecodeDone, weak_this_, frame));
641 return;
642 }
643
644 // If the session is not configured by this point, fail.
645 if (!session_) {
646 DLOG(ERROR) << "Cannot decode without configuration";
647 NotifyError(INVALID_ARGUMENT, SFT_INVALID_STREAM);
648 return;
649 }
650
651 // Update the frame metadata with configuration data.
652 frame->coded_size = coded_size_;
653
654 // Create a memory-backed CMBlockBuffer for the translated data.
655 // TODO(sandersd): Pool of memory blocks.
656 base::ScopedCFTypeRef<CMBlockBufferRef> data;
657 OSStatus status = CMBlockBufferCreateWithMemoryBlock(
658 kCFAllocatorDefault,
659 nullptr, // &memory_block
660 data_size, // block_length
661 kCFAllocatorDefault, // block_allocator
662 nullptr, // &custom_block_source
663 0, // offset_to_data
664 data_size, // data_length
665 0, // flags
666 data.InitializeInto());
667 if (status) {
668 NOTIFY_STATUS("CMBlockBufferCreateWithMemoryBlock()", status,
669 SFT_PLATFORM_ERROR);
670 return;
671 }
672
673 // Make sure that the memory is actually allocated.
674 // CMBlockBufferReplaceDataBytes() is documented to do this, but prints a
675 // message each time starting in Mac OS X 10.10.
676 status = CMBlockBufferAssureBlockMemory(data);
677 if (status) {
678 NOTIFY_STATUS("CMBlockBufferAssureBlockMemory()", status,
679 SFT_PLATFORM_ERROR);
680 return;
681 }
682
683 // Copy NALU data into the CMBlockBuffer, inserting length headers.
684 size_t offset = 0;
685 for (size_t i = 0; i < nalus.size(); i++) {
686 media::H264NALU& nalu = nalus[i];
687 uint32_t header = base::HostToNet32(static_cast<uint32_t>(nalu.size));
688 status = CMBlockBufferReplaceDataBytes(
689 &header, data, offset, kNALUHeaderLength);
690 if (status) {
691 NOTIFY_STATUS("CMBlockBufferReplaceDataBytes()", status,
692 SFT_PLATFORM_ERROR);
693 return;
694 }
695 offset += kNALUHeaderLength;
696 status = CMBlockBufferReplaceDataBytes(nalu.data, data, offset, nalu.size);
697 if (status) {
698 NOTIFY_STATUS("CMBlockBufferReplaceDataBytes()", status,
699 SFT_PLATFORM_ERROR);
700 return;
701 }
702 offset += nalu.size;
703 }
704
705 // Package the data in a CMSampleBuffer.
706 base::ScopedCFTypeRef<CMSampleBufferRef> sample;
707 status = CMSampleBufferCreate(
708 kCFAllocatorDefault,
709 data, // data_buffer
710 true, // data_ready
711 nullptr, // make_data_ready_callback
712 nullptr, // make_data_ready_refcon
713 format_, // format_description
714 1, // num_samples
715 0, // num_sample_timing_entries
716 nullptr, // &sample_timing_array
717 1, // num_sample_size_entries
718 &data_size, // &sample_size_array
719 sample.InitializeInto());
720 if (status) {
721 NOTIFY_STATUS("CMSampleBufferCreate()", status, SFT_PLATFORM_ERROR);
722 return;
723 }
724
725 // Send the frame for decoding.
726 // Asynchronous Decompression allows for parallel submission of frames
727 // (without it, DecodeFrame() does not return until the frame has been
728 // decoded). We don't enable Temporal Processing so that frames are always
729 // returned in decode order; this makes it easier to avoid deadlock.
730 VTDecodeFrameFlags decode_flags =
731 kVTDecodeFrame_EnableAsynchronousDecompression;
732 status = VTDecompressionSessionDecodeFrame(
733 session_,
734 sample, // sample_buffer
735 decode_flags, // decode_flags
736 reinterpret_cast<void*>(frame), // source_frame_refcon
737 nullptr); // &info_flags_out
738 if (status) {
739 NOTIFY_STATUS("VTDecompressionSessionDecodeFrame()", status,
740 SFT_DECODE_ERROR);
741 return;
742 }
743 }
744
745 // This method may be called on any VideoToolbox thread.
746 void VTVideoDecodeAccelerator::Output(
747 void* source_frame_refcon,
748 OSStatus status,
749 CVImageBufferRef image_buffer) {
750 if (status) {
751 NOTIFY_STATUS("Decoding", status, SFT_DECODE_ERROR);
752 return;
753 }
754
755 // The type of |image_buffer| is CVImageBuffer, but we only handle
756 // CVPixelBuffers. This should be guaranteed as we set
757 // kCVPixelBufferOpenGLCompatibilityKey in |image_config|.
758 //
759 // Sometimes, for unknown reasons (http://crbug.com/453050), |image_buffer| is
760 // NULL, which causes CFGetTypeID() to crash. While the rest of the code would
761 // smoothly handle NULL as a dropped frame, we choose to fail permanantly here
762 // until the issue is better understood.
763 if (!image_buffer || CFGetTypeID(image_buffer) != CVPixelBufferGetTypeID()) {
764 DLOG(ERROR) << "Decoded frame is not a CVPixelBuffer";
765 NotifyError(PLATFORM_FAILURE, SFT_DECODE_ERROR);
766 return;
767 }
768
769 Frame* frame = reinterpret_cast<Frame*>(source_frame_refcon);
770 frame->image.reset(image_buffer, base::scoped_policy::RETAIN);
771 gpu_task_runner_->PostTask(FROM_HERE, base::Bind(
772 &VTVideoDecodeAccelerator::DecodeDone, weak_this_, frame));
773 }
774
775 void VTVideoDecodeAccelerator::DecodeDone(Frame* frame) {
776 DCHECK(gpu_thread_checker_.CalledOnValidThread());
777 DCHECK_EQ(1u, pending_frames_.count(frame->bitstream_id));
778 Task task(TASK_FRAME);
779 task.frame = pending_frames_[frame->bitstream_id];
780 pending_frames_.erase(frame->bitstream_id);
781 task_queue_.push(task);
782 ProcessWorkQueues();
783 }
784
785 void VTVideoDecodeAccelerator::FlushTask(TaskType type) {
786 DCHECK(decoder_thread_.task_runner()->BelongsToCurrentThread());
787 FinishDelayedFrames();
788
789 // Always queue a task, even if FinishDelayedFrames() fails, so that
790 // destruction always completes.
791 gpu_task_runner_->PostTask(FROM_HERE, base::Bind(
792 &VTVideoDecodeAccelerator::FlushDone, weak_this_, type));
793 }
794
795 void VTVideoDecodeAccelerator::FlushDone(TaskType type) {
796 DCHECK(gpu_thread_checker_.CalledOnValidThread());
797 task_queue_.push(Task(type));
798 ProcessWorkQueues();
799 }
800
801 void VTVideoDecodeAccelerator::Decode(const media::BitstreamBuffer& bitstream) {
802 DCHECK(gpu_thread_checker_.CalledOnValidThread());
803 DCHECK_EQ(0u, assigned_bitstream_ids_.count(bitstream.id()));
804 assigned_bitstream_ids_.insert(bitstream.id());
805 Frame* frame = new Frame(bitstream.id());
806 pending_frames_[frame->bitstream_id] = make_linked_ptr(frame);
807 decoder_thread_.task_runner()->PostTask(
808 FROM_HERE, base::Bind(&VTVideoDecodeAccelerator::DecodeTask,
809 base::Unretained(this), bitstream, frame));
810 }
811
812 void VTVideoDecodeAccelerator::AssignPictureBuffers(
813 const std::vector<media::PictureBuffer>& pictures) {
814 DCHECK(gpu_thread_checker_.CalledOnValidThread());
815
816 for (const media::PictureBuffer& picture : pictures) {
817 DCHECK(!picture_info_map_.count(picture.id()));
818 assigned_picture_ids_.insert(picture.id());
819 available_picture_ids_.push_back(picture.id());
820 picture_info_map_.insert(std::make_pair(
821 picture.id(),
822 make_scoped_ptr(new PictureInfo(picture.internal_texture_id(),
823 picture.texture_id()))));
824 }
825
826 // Pictures are not marked as uncleared until after this method returns, and
827 // they will be broken if they are used before that happens. So, schedule
828 // future work after that happens.
829 gpu_task_runner_->PostTask(FROM_HERE, base::Bind(
830 &VTVideoDecodeAccelerator::ProcessWorkQueues, weak_this_));
831 }
832
833 void VTVideoDecodeAccelerator::ReusePictureBuffer(int32_t picture_id) {
834 DCHECK(gpu_thread_checker_.CalledOnValidThread());
835 DCHECK(picture_info_map_.count(picture_id));
836 PictureInfo* picture_info = picture_info_map_.find(picture_id)->second.get();
837 DCHECK_EQ(CFGetRetainCount(picture_info->cv_image), 1);
838 picture_info->cv_image.reset();
839 picture_info->gl_image->Destroy(false);
840 picture_info->gl_image = nullptr;
841
842 if (assigned_picture_ids_.count(picture_id) != 0) {
843 available_picture_ids_.push_back(picture_id);
844 ProcessWorkQueues();
845 } else {
846 client_->DismissPictureBuffer(picture_id);
847 }
848 }
849
850 void VTVideoDecodeAccelerator::ProcessWorkQueues() {
851 DCHECK(gpu_thread_checker_.CalledOnValidThread());
852 switch (state_) {
853 case STATE_DECODING:
854 // TODO(sandersd): Batch where possible.
855 while (state_ == STATE_DECODING) {
856 if (!ProcessReorderQueue() && !ProcessTaskQueue())
857 break;
858 }
859 return;
860
861 case STATE_ERROR:
862 // Do nothing until Destroy() is called.
863 return;
864
865 case STATE_DESTROYING:
866 // Drop tasks until we are ready to destruct.
867 while (!task_queue_.empty()) {
868 if (task_queue_.front().type == TASK_DESTROY) {
869 delete this;
870 return;
871 }
872 task_queue_.pop();
873 }
874 return;
875 }
876 }
877
878 bool VTVideoDecodeAccelerator::ProcessTaskQueue() {
879 DCHECK(gpu_thread_checker_.CalledOnValidThread());
880 DCHECK_EQ(state_, STATE_DECODING);
881
882 if (task_queue_.empty())
883 return false;
884
885 const Task& task = task_queue_.front();
886 switch (task.type) {
887 case TASK_FRAME:
888 if (reorder_queue_.size() < kMaxReorderQueueSize &&
889 (!task.frame->is_idr || reorder_queue_.empty())) {
890 assigned_bitstream_ids_.erase(task.frame->bitstream_id);
891 client_->NotifyEndOfBitstreamBuffer(task.frame->bitstream_id);
892 reorder_queue_.push(task.frame);
893 task_queue_.pop();
894 return true;
895 }
896 return false;
897
898 case TASK_FLUSH:
899 DCHECK_EQ(task.type, pending_flush_tasks_.front());
900 if (reorder_queue_.size() == 0) {
901 pending_flush_tasks_.pop();
902 client_->NotifyFlushDone();
903 task_queue_.pop();
904 return true;
905 }
906 return false;
907
908 case TASK_RESET:
909 DCHECK_EQ(task.type, pending_flush_tasks_.front());
910 if (reorder_queue_.size() == 0) {
911 last_sps_id_ = -1;
912 last_pps_id_ = -1;
913 last_sps_.clear();
914 last_spsext_.clear();
915 last_pps_.clear();
916 poc_.Reset();
917 pending_flush_tasks_.pop();
918 client_->NotifyResetDone();
919 task_queue_.pop();
920 return true;
921 }
922 return false;
923
924 case TASK_DESTROY:
925 NOTREACHED() << "Can't destroy while in STATE_DECODING.";
926 NotifyError(ILLEGAL_STATE, SFT_PLATFORM_ERROR);
927 return false;
928 }
929 }
930
931 bool VTVideoDecodeAccelerator::ProcessReorderQueue() {
932 DCHECK(gpu_thread_checker_.CalledOnValidThread());
933 DCHECK_EQ(state_, STATE_DECODING);
934
935 if (reorder_queue_.empty())
936 return false;
937
938 // If the next task is a flush (because there is a pending flush or becuase
939 // the next frame is an IDR), then we don't need a full reorder buffer to send
940 // the next frame.
941 bool flushing = !task_queue_.empty() &&
942 (task_queue_.front().type != TASK_FRAME ||
943 task_queue_.front().frame->is_idr);
944
945 size_t reorder_window = std::max(0, reorder_queue_.top()->reorder_window);
946 if (flushing || reorder_queue_.size() > reorder_window) {
947 if (ProcessFrame(*reorder_queue_.top())) {
948 reorder_queue_.pop();
949 return true;
950 }
951 }
952
953 return false;
954 }
955
956 bool VTVideoDecodeAccelerator::ProcessFrame(const Frame& frame) {
957 DCHECK(gpu_thread_checker_.CalledOnValidThread());
958 DCHECK_EQ(state_, STATE_DECODING);
959
960 // If the next pending flush is for a reset, then the frame will be dropped.
961 bool resetting = !pending_flush_tasks_.empty() &&
962 pending_flush_tasks_.front() == TASK_RESET;
963
964 if (!resetting && frame.image.get()) {
965 // If the |coded_size| has changed, request new picture buffers and then
966 // wait for them.
967 // TODO(sandersd): If GpuVideoDecoder didn't specifically check the size of
968 // textures, this would be unnecessary, as the size is actually a property
969 // of the texture binding, not the texture. We rebind every frame, so the
970 // size passed to ProvidePictureBuffers() is meaningless.
971 if (picture_size_ != frame.coded_size) {
972 // Dismiss current pictures.
973 for (int32_t picture_id : assigned_picture_ids_)
974 client_->DismissPictureBuffer(picture_id);
975 assigned_picture_ids_.clear();
976 available_picture_ids_.clear();
977
978 // Request new pictures.
979 picture_size_ = frame.coded_size;
980 client_->ProvidePictureBuffers(
981 kNumPictureBuffers, coded_size_, GL_TEXTURE_RECTANGLE_ARB);
982 return false;
983 }
984 if (!SendFrame(frame))
985 return false;
986 }
987
988 return true;
989 }
990
991 bool VTVideoDecodeAccelerator::SendFrame(const Frame& frame) {
992 DCHECK(gpu_thread_checker_.CalledOnValidThread());
993 DCHECK_EQ(state_, STATE_DECODING);
994
995 if (available_picture_ids_.empty())
996 return false;
997
998 int32_t picture_id = available_picture_ids_.back();
999 DCHECK(picture_info_map_.count(picture_id));
1000 PictureInfo* picture_info = picture_info_map_.find(picture_id)->second.get();
1001 DCHECK(!picture_info->cv_image);
1002 DCHECK(!picture_info->gl_image);
1003
1004 if (!make_context_current_.Run()) {
1005 DLOG(ERROR) << "Failed to make GL context current";
1006 NotifyError(PLATFORM_FAILURE, SFT_PLATFORM_ERROR);
1007 return false;
1008 }
1009
1010 IOSurfaceRef surface = CVPixelBufferGetIOSurface(frame.image.get());
1011 if (gfx::GetGLImplementation() != gfx::kGLImplementationDesktopGLCoreProfile)
1012 glEnable(GL_TEXTURE_RECTANGLE_ARB);
1013 gfx::ScopedTextureBinder texture_binder(GL_TEXTURE_RECTANGLE_ARB,
1014 picture_info->service_texture_id);
1015 CGLContextObj cgl_context =
1016 static_cast<CGLContextObj>(gfx::GLContext::GetCurrent()->GetHandle());
1017 CGLError status = CGLTexImageIOSurface2D(
1018 cgl_context, // ctx
1019 GL_TEXTURE_RECTANGLE_ARB, // target
1020 GL_RGB, // internal_format
1021 frame.coded_size.width(), // width
1022 frame.coded_size.height(), // height
1023 GL_YCBCR_422_APPLE, // format
1024 GL_UNSIGNED_SHORT_8_8_APPLE, // type
1025 surface, // io_surface
1026 0); // plane
1027 if (gfx::GetGLImplementation() != gfx::kGLImplementationDesktopGLCoreProfile)
1028 glDisable(GL_TEXTURE_RECTANGLE_ARB);
1029 if (status != kCGLNoError) {
1030 NOTIFY_STATUS("CGLTexImageIOSurface2D()", status, SFT_PLATFORM_ERROR);
1031 return false;
1032 }
1033
1034 bool allow_overlay = false;
1035 scoped_refptr<gl::GLImageIOSurface> gl_image(
1036 new gl::GLImageIOSurface(frame.coded_size, GL_BGRA_EXT));
1037 if (gl_image->Initialize(surface, gfx::GenericSharedMemoryId(),
1038 gfx::BufferFormat::BGRA_8888)) {
1039 allow_overlay = true;
1040 } else {
1041 gl_image = nullptr;
1042 }
1043 bind_image_.Run(picture_info->client_texture_id, GL_TEXTURE_RECTANGLE_ARB,
1044 gl_image);
1045
1046 // Assign the new image(s) to the the picture info.
1047 picture_info->gl_image = gl_image;
1048 picture_info->cv_image = frame.image;
1049 available_picture_ids_.pop_back();
1050
1051 // TODO(sandersd): Currently, the size got from
1052 // CMVideoFormatDescriptionGetDimensions is visible size. We pass it to
1053 // GpuVideoDecoder so that GpuVideoDecoder can use correct visible size in
1054 // resolution changed. We should find the correct API to get the real
1055 // coded size and fix it.
1056 client_->PictureReady(media::Picture(picture_id, frame.bitstream_id,
1057 gfx::Rect(frame.coded_size),
1058 allow_overlay));
1059 return true;
1060 }
1061
1062 void VTVideoDecodeAccelerator::NotifyError(
1063 Error vda_error_type,
1064 VTVDASessionFailureType session_failure_type) {
1065 DCHECK_LT(session_failure_type, SFT_MAX + 1);
1066 if (!gpu_thread_checker_.CalledOnValidThread()) {
1067 gpu_task_runner_->PostTask(FROM_HERE, base::Bind(
1068 &VTVideoDecodeAccelerator::NotifyError, weak_this_, vda_error_type,
1069 session_failure_type));
1070 } else if (state_ == STATE_DECODING) {
1071 state_ = STATE_ERROR;
1072 UMA_HISTOGRAM_ENUMERATION("Media.VTVDA.SessionFailureReason",
1073 session_failure_type,
1074 SFT_MAX + 1);
1075 client_->NotifyError(vda_error_type);
1076 }
1077 }
1078
1079 void VTVideoDecodeAccelerator::QueueFlush(TaskType type) {
1080 DCHECK(gpu_thread_checker_.CalledOnValidThread());
1081 pending_flush_tasks_.push(type);
1082 decoder_thread_.task_runner()->PostTask(
1083 FROM_HERE, base::Bind(&VTVideoDecodeAccelerator::FlushTask,
1084 base::Unretained(this), type));
1085
1086 // If this is a new flush request, see if we can make progress.
1087 if (pending_flush_tasks_.size() == 1)
1088 ProcessWorkQueues();
1089 }
1090
1091 void VTVideoDecodeAccelerator::Flush() {
1092 DCHECK(gpu_thread_checker_.CalledOnValidThread());
1093 QueueFlush(TASK_FLUSH);
1094 }
1095
1096 void VTVideoDecodeAccelerator::Reset() {
1097 DCHECK(gpu_thread_checker_.CalledOnValidThread());
1098 QueueFlush(TASK_RESET);
1099 }
1100
1101 void VTVideoDecodeAccelerator::Destroy() {
1102 DCHECK(gpu_thread_checker_.CalledOnValidThread());
1103
1104 // In a forceful shutdown, the decoder thread may be dead already.
1105 if (!decoder_thread_.IsRunning()) {
1106 delete this;
1107 return;
1108 }
1109
1110 // For a graceful shutdown, return assigned buffers and flush before
1111 // destructing |this|.
1112 // TODO(sandersd): Prevent the decoder from reading buffers before discarding
1113 // them.
1114 for (int32_t bitstream_id : assigned_bitstream_ids_)
1115 client_->NotifyEndOfBitstreamBuffer(bitstream_id);
1116 assigned_bitstream_ids_.clear();
1117 state_ = STATE_DESTROYING;
1118 QueueFlush(TASK_DESTROY);
1119 }
1120
1121 bool VTVideoDecodeAccelerator::CanDecodeOnIOThread() {
1122 return false;
1123 }
1124
1125 // static
1126 media::VideoDecodeAccelerator::SupportedProfiles
1127 VTVideoDecodeAccelerator::GetSupportedProfiles() {
1128 SupportedProfiles profiles;
1129 for (const auto& supported_profile : kSupportedProfiles) {
1130 SupportedProfile profile;
1131 profile.profile = supported_profile;
1132 profile.min_resolution.SetSize(16, 16);
1133 profile.max_resolution.SetSize(4096, 2160);
1134 profiles.push_back(profile);
1135 }
1136 return profiles;
1137 }
1138
1139 } // namespace content
OLDNEW
« no previous file with comments | « content/common/gpu/media/vt_video_decode_accelerator.h ('k') | content/common/gpu/media/vt_video_decode_accelerator_mac.h » ('j') | no next file with comments »

Powered by Google App Engine
This is Rietveld 408576698