Updating XTBs based on .GRDs from branch master
[chromium-blink-merge.git] / media / filters / gpu_video_decoder.cc
blob82a7221de05450fbf6cc1d7a502d850686cc7701
1 // Copyright (c) 2012 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 "media/filters/gpu_video_decoder.h"
7 #include <algorithm>
9 #include "base/bind.h"
10 #include "base/callback_helpers.h"
11 #include "base/command_line.h"
12 #include "base/cpu.h"
13 #include "base/message_loop/message_loop.h"
14 #include "base/metrics/histogram.h"
15 #include "base/stl_util.h"
16 #include "base/task_runner_util.h"
17 #include "gpu/command_buffer/common/mailbox_holder.h"
18 #include "media/base/bind_to_current_loop.h"
19 #include "media/base/decoder_buffer.h"
20 #include "media/base/media_switches.h"
21 #include "media/base/pipeline.h"
22 #include "media/base/video_decoder_config.h"
23 #include "media/renderers/gpu_video_accelerator_factories.h"
24 #include "third_party/skia/include/core/SkBitmap.h"
26 namespace media {
28 const char GpuVideoDecoder::kDecoderName[] = "GpuVideoDecoder";
30 // Maximum number of concurrent VDA::Decode() operations GVD will maintain.
31 // Higher values allow better pipelining in the GPU, but also require more
32 // resources.
33 enum { kMaxInFlightDecodes = 4 };
35 // Size of shared-memory segments we allocate. Since we reuse them we let them
36 // be on the beefy side.
37 static const size_t kSharedMemorySegmentBytes = 100 << 10;
39 GpuVideoDecoder::SHMBuffer::SHMBuffer(scoped_ptr<base::SharedMemory> m,
40 size_t s)
41 : shm(m.Pass()), size(s) {
44 GpuVideoDecoder::SHMBuffer::~SHMBuffer() {}
46 GpuVideoDecoder::PendingDecoderBuffer::PendingDecoderBuffer(
47 SHMBuffer* s,
48 const scoped_refptr<DecoderBuffer>& b,
49 const DecodeCB& done_cb)
50 : shm_buffer(s), buffer(b), done_cb(done_cb) {
53 GpuVideoDecoder::PendingDecoderBuffer::~PendingDecoderBuffer() {}
55 GpuVideoDecoder::BufferData::BufferData(
56 int32 bbid, base::TimeDelta ts, const gfx::Rect& vr, const gfx::Size& ns)
57 : bitstream_buffer_id(bbid), timestamp(ts), visible_rect(vr),
58 natural_size(ns) {
61 GpuVideoDecoder::BufferData::~BufferData() {}
63 GpuVideoDecoder::GpuVideoDecoder(
64 const scoped_refptr<GpuVideoAcceleratorFactories>& factories)
65 : needs_bitstream_conversion_(false),
66 factories_(factories),
67 state_(kNormal),
68 decoder_texture_target_(0),
69 next_picture_buffer_id_(0),
70 next_bitstream_buffer_id_(0),
71 available_pictures_(0),
72 weak_factory_(this) {
73 DCHECK(factories_.get());
76 void GpuVideoDecoder::Reset(const base::Closure& closure) {
77 DVLOG(3) << "Reset()";
78 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
80 if (state_ == kDrainingDecoder) {
81 base::MessageLoop::current()->PostTask(
82 FROM_HERE,
83 base::Bind(
84 &GpuVideoDecoder::Reset, weak_factory_.GetWeakPtr(), closure));
85 return;
88 if (!vda_) {
89 base::MessageLoop::current()->PostTask(FROM_HERE, closure);
90 return;
93 DCHECK(pending_reset_cb_.is_null());
94 pending_reset_cb_ = BindToCurrentLoop(closure);
96 vda_->Reset();
99 static bool IsCodedSizeSupported(const gfx::Size& coded_size,
100 const gfx::Size& min_resolution,
101 const gfx::Size& max_resolution) {
102 return (coded_size.width() <= max_resolution.width() &&
103 coded_size.height() <= max_resolution.height() &&
104 coded_size.width() >= min_resolution.width() &&
105 coded_size.height() >= min_resolution.height());
108 // Report |success| to UMA and run |cb| with it. This is super-specific to the
109 // UMA stat reported because the UMA_HISTOGRAM_ENUMERATION API requires a
110 // callsite to always be called with the same stat name (can't parameterize it).
111 static void ReportGpuVideoDecoderInitializeStatusToUMAAndRunCB(
112 const VideoDecoder::InitCB& cb,
113 bool success) {
114 // TODO(xhwang): Report |success| directly.
115 PipelineStatus status = success ? PIPELINE_OK : DECODER_ERROR_NOT_SUPPORTED;
116 UMA_HISTOGRAM_ENUMERATION(
117 "Media.GpuVideoDecoderInitializeStatus", status, PIPELINE_STATUS_MAX + 1);
118 cb.Run(success);
121 std::string GpuVideoDecoder::GetDisplayName() const {
122 return kDecoderName;
125 void GpuVideoDecoder::Initialize(const VideoDecoderConfig& config,
126 bool /* low_delay */,
127 const InitCB& init_cb,
128 const OutputCB& output_cb) {
129 DVLOG(3) << "Initialize()";
130 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
131 DCHECK(config.IsValidConfig());
132 DCHECK(!config.is_encrypted());
134 InitCB bound_init_cb =
135 base::Bind(&ReportGpuVideoDecoderInitializeStatusToUMAAndRunCB,
136 BindToCurrentLoop(init_cb));
138 bool previously_initialized = config_.IsValidConfig();
139 DVLOG(1) << "(Re)initializing GVD with config: "
140 << config.AsHumanReadableString();
142 // TODO(posciak): destroy and create a new VDA on codec/profile change
143 // (http://crbug.com/260224).
144 if (previously_initialized && (config_.profile() != config.profile())) {
145 DVLOG(1) << "Codec or profile changed, cannot reinitialize.";
146 bound_init_cb.Run(false);
147 return;
150 if (!IsProfileSupported(config.profile(), config.coded_size())) {
151 bound_init_cb.Run(false);
152 return;
155 config_ = config;
156 needs_bitstream_conversion_ = (config.codec() == kCodecH264);
157 output_cb_ = BindToCurrentLoop(output_cb);
159 if (previously_initialized) {
160 // Reinitialization with a different config (but same codec and profile).
161 // VDA should handle it by detecting this in-stream by itself,
162 // no need to notify it.
163 bound_init_cb.Run(true);
164 return;
167 vda_ = factories_->CreateVideoDecodeAccelerator().Pass();
168 if (!vda_ || !vda_->Initialize(config.profile(), this)) {
169 bound_init_cb.Run(false);
170 return;
173 DVLOG(3) << "GpuVideoDecoder::Initialize() succeeded.";
174 bound_init_cb.Run(true);
177 void GpuVideoDecoder::DestroyPictureBuffers(PictureBufferMap* buffers) {
178 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
179 for (PictureBufferMap::iterator it = buffers->begin(); it != buffers->end();
180 ++it) {
181 factories_->DeleteTexture(it->second.texture_id());
184 buffers->clear();
187 void GpuVideoDecoder::DestroyVDA() {
188 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
190 vda_.reset();
192 // Not destroying PictureBuffers in |picture_buffers_at_display_| yet, since
193 // their textures may still be in use by the user of this GpuVideoDecoder.
194 for (PictureBufferTextureMap::iterator it =
195 picture_buffers_at_display_.begin();
196 it != picture_buffers_at_display_.end();
197 ++it) {
198 assigned_picture_buffers_.erase(it->first);
200 DestroyPictureBuffers(&assigned_picture_buffers_);
203 void GpuVideoDecoder::Decode(const scoped_refptr<DecoderBuffer>& buffer,
204 const DecodeCB& decode_cb) {
205 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
206 DCHECK(pending_reset_cb_.is_null());
208 DecodeCB bound_decode_cb = BindToCurrentLoop(decode_cb);
210 if (state_ == kError || !vda_) {
211 bound_decode_cb.Run(kDecodeError);
212 return;
215 switch (state_) {
216 case kDecoderDrained:
217 state_ = kNormal;
218 // Fall-through.
219 case kNormal:
220 break;
221 case kDrainingDecoder:
222 case kError:
223 NOTREACHED();
224 return;
227 DCHECK_EQ(state_, kNormal);
229 if (buffer->end_of_stream()) {
230 state_ = kDrainingDecoder;
231 eos_decode_cb_ = bound_decode_cb;
232 vda_->Flush();
233 return;
236 size_t size = buffer->data_size();
237 scoped_ptr<SHMBuffer> shm_buffer = GetSHM(size);
238 if (!shm_buffer) {
239 bound_decode_cb.Run(kDecodeError);
240 return;
243 memcpy(shm_buffer->shm->memory(), buffer->data(), size);
244 BitstreamBuffer bitstream_buffer(
245 next_bitstream_buffer_id_, shm_buffer->shm->handle(), size);
246 // Mask against 30 bits, to avoid (undefined) wraparound on signed integer.
247 next_bitstream_buffer_id_ = (next_bitstream_buffer_id_ + 1) & 0x3FFFFFFF;
248 DCHECK(!ContainsKey(bitstream_buffers_in_decoder_, bitstream_buffer.id()));
249 bitstream_buffers_in_decoder_.insert(std::make_pair(
250 bitstream_buffer.id(),
251 PendingDecoderBuffer(shm_buffer.release(), buffer, decode_cb)));
252 DCHECK_LE(static_cast<int>(bitstream_buffers_in_decoder_.size()),
253 kMaxInFlightDecodes);
254 RecordBufferData(bitstream_buffer, *buffer.get());
256 vda_->Decode(bitstream_buffer);
259 void GpuVideoDecoder::RecordBufferData(const BitstreamBuffer& bitstream_buffer,
260 const DecoderBuffer& buffer) {
261 input_buffer_data_.push_front(BufferData(bitstream_buffer.id(),
262 buffer.timestamp(),
263 config_.visible_rect(),
264 config_.natural_size()));
265 // Why this value? Because why not. avformat.h:MAX_REORDER_DELAY is 16, but
266 // that's too small for some pathological B-frame test videos. The cost of
267 // using too-high a value is low (192 bits per extra slot).
268 static const size_t kMaxInputBufferDataSize = 128;
269 // Pop from the back of the list, because that's the oldest and least likely
270 // to be useful in the future data.
271 if (input_buffer_data_.size() > kMaxInputBufferDataSize)
272 input_buffer_data_.pop_back();
275 void GpuVideoDecoder::GetBufferData(int32 id, base::TimeDelta* timestamp,
276 gfx::Rect* visible_rect,
277 gfx::Size* natural_size) {
278 for (std::list<BufferData>::const_iterator it =
279 input_buffer_data_.begin(); it != input_buffer_data_.end();
280 ++it) {
281 if (it->bitstream_buffer_id != id)
282 continue;
283 *timestamp = it->timestamp;
284 *visible_rect = it->visible_rect;
285 *natural_size = it->natural_size;
286 return;
288 NOTREACHED() << "Missing bitstreambuffer id: " << id;
291 bool GpuVideoDecoder::NeedsBitstreamConversion() const {
292 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
293 return needs_bitstream_conversion_;
296 bool GpuVideoDecoder::CanReadWithoutStalling() const {
297 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
298 return
299 next_picture_buffer_id_ == 0 || // Decode() will ProvidePictureBuffers().
300 available_pictures_ > 0;
303 int GpuVideoDecoder::GetMaxDecodeRequests() const {
304 return kMaxInFlightDecodes;
307 void GpuVideoDecoder::ProvidePictureBuffers(uint32 count,
308 const gfx::Size& size,
309 uint32 texture_target) {
310 DVLOG(3) << "ProvidePictureBuffers(" << count << ", "
311 << size.width() << "x" << size.height() << ")";
312 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
314 std::vector<uint32> texture_ids;
315 std::vector<gpu::Mailbox> texture_mailboxes;
316 decoder_texture_target_ = texture_target;
317 if (!factories_->CreateTextures(count,
318 size,
319 &texture_ids,
320 &texture_mailboxes,
321 decoder_texture_target_)) {
322 NotifyError(VideoDecodeAccelerator::PLATFORM_FAILURE);
323 return;
325 DCHECK_EQ(count, texture_ids.size());
326 DCHECK_EQ(count, texture_mailboxes.size());
328 if (!vda_)
329 return;
331 std::vector<PictureBuffer> picture_buffers;
332 for (size_t i = 0; i < texture_ids.size(); ++i) {
333 picture_buffers.push_back(PictureBuffer(
334 next_picture_buffer_id_++, size, texture_ids[i], texture_mailboxes[i]));
335 bool inserted = assigned_picture_buffers_.insert(std::make_pair(
336 picture_buffers.back().id(), picture_buffers.back())).second;
337 DCHECK(inserted);
340 available_pictures_ += count;
342 vda_->AssignPictureBuffers(picture_buffers);
345 void GpuVideoDecoder::DismissPictureBuffer(int32 id) {
346 DVLOG(3) << "DismissPictureBuffer(" << id << ")";
347 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
349 PictureBufferMap::iterator it = assigned_picture_buffers_.find(id);
350 if (it == assigned_picture_buffers_.end()) {
351 NOTREACHED() << "Missing picture buffer: " << id;
352 return;
355 PictureBuffer buffer_to_dismiss = it->second;
356 assigned_picture_buffers_.erase(it);
358 if (!picture_buffers_at_display_.count(id)) {
359 // We can delete the texture immediately as it's not being displayed.
360 factories_->DeleteTexture(buffer_to_dismiss.texture_id());
361 CHECK_GT(available_pictures_, 0);
362 --available_pictures_;
364 // Not destroying a texture in display in |picture_buffers_at_display_|.
365 // Postpone deletion until after it's returned to us.
368 void GpuVideoDecoder::PictureReady(const media::Picture& picture) {
369 DVLOG(3) << "PictureReady()";
370 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
372 PictureBufferMap::iterator it =
373 assigned_picture_buffers_.find(picture.picture_buffer_id());
374 if (it == assigned_picture_buffers_.end()) {
375 NOTREACHED() << "Missing picture buffer: " << picture.picture_buffer_id();
376 NotifyError(VideoDecodeAccelerator::PLATFORM_FAILURE);
377 return;
379 const PictureBuffer& pb = it->second;
381 // Validate picture rectangle from GPU. This is for sanity/security check
382 // even the rectangle is not used in this class.
383 if (picture.visible_rect().IsEmpty() ||
384 !gfx::Rect(pb.size()).Contains(picture.visible_rect())) {
385 NOTREACHED() << "Invalid picture size from VDA: "
386 << picture.visible_rect().ToString() << " should fit in "
387 << pb.size().ToString();
388 NotifyError(media::VideoDecodeAccelerator::PLATFORM_FAILURE);
389 return;
392 // Update frame's timestamp.
393 base::TimeDelta timestamp;
394 // Some of the VDAs don't support and thus don't provide us with visible
395 // size in picture.size, passing coded size instead, so always drop it and
396 // use config information instead.
397 gfx::Rect visible_rect;
398 gfx::Size natural_size;
399 GetBufferData(picture.bitstream_buffer_id(), &timestamp, &visible_rect,
400 &natural_size);
401 DCHECK(decoder_texture_target_);
403 scoped_refptr<VideoFrame> frame(VideoFrame::WrapNativeTexture(
404 PIXEL_FORMAT_ARGB,
405 gpu::MailboxHolder(pb.texture_mailbox(), decoder_texture_target_,
406 0 /* sync_point */),
407 BindToCurrentLoop(base::Bind(
408 &GpuVideoDecoder::ReleaseMailbox, weak_factory_.GetWeakPtr(),
409 factories_, picture.picture_buffer_id(), pb.texture_id())),
410 pb.size(), visible_rect, natural_size, timestamp));
411 if (picture.allow_overlay())
412 frame->metadata()->SetBoolean(VideoFrameMetadata::ALLOW_OVERLAY, true);
413 CHECK_GT(available_pictures_, 0);
414 --available_pictures_;
415 bool inserted =
416 picture_buffers_at_display_.insert(std::make_pair(
417 picture.picture_buffer_id(),
418 pb.texture_id())).second;
419 DCHECK(inserted);
421 DeliverFrame(frame);
424 void GpuVideoDecoder::DeliverFrame(
425 const scoped_refptr<VideoFrame>& frame) {
426 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
428 // During a pending vda->Reset(), we don't accumulate frames. Drop it on the
429 // floor and return.
430 if (!pending_reset_cb_.is_null())
431 return;
433 output_cb_.Run(frame);
436 // static
437 void GpuVideoDecoder::ReleaseMailbox(
438 base::WeakPtr<GpuVideoDecoder> decoder,
439 const scoped_refptr<media::GpuVideoAcceleratorFactories>& factories,
440 int64 picture_buffer_id,
441 uint32 texture_id,
442 uint32 release_sync_point) {
443 DCHECK(factories->GetTaskRunner()->BelongsToCurrentThread());
444 factories->WaitSyncPoint(release_sync_point);
446 if (decoder) {
447 decoder->ReusePictureBuffer(picture_buffer_id);
448 return;
450 // It's the last chance to delete the texture after display,
451 // because GpuVideoDecoder was destructed.
452 factories->DeleteTexture(texture_id);
455 void GpuVideoDecoder::ReusePictureBuffer(int64 picture_buffer_id) {
456 DVLOG(3) << "ReusePictureBuffer(" << picture_buffer_id << ")";
457 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
459 DCHECK(!picture_buffers_at_display_.empty());
460 PictureBufferTextureMap::iterator display_iterator =
461 picture_buffers_at_display_.find(picture_buffer_id);
462 uint32 texture_id = display_iterator->second;
463 DCHECK(display_iterator != picture_buffers_at_display_.end());
464 picture_buffers_at_display_.erase(display_iterator);
466 if (!assigned_picture_buffers_.count(picture_buffer_id)) {
467 // This picture was dismissed while in display, so we postponed deletion.
468 factories_->DeleteTexture(texture_id);
469 return;
472 ++available_pictures_;
474 // DestroyVDA() might already have been called.
475 if (vda_)
476 vda_->ReusePictureBuffer(picture_buffer_id);
479 scoped_ptr<GpuVideoDecoder::SHMBuffer> GpuVideoDecoder::GetSHM(
480 size_t min_size) {
481 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
482 if (available_shm_segments_.empty() ||
483 available_shm_segments_.back()->size < min_size) {
484 size_t size_to_allocate = std::max(min_size, kSharedMemorySegmentBytes);
485 scoped_ptr<base::SharedMemory> shm =
486 factories_->CreateSharedMemory(size_to_allocate);
487 // CreateSharedMemory() can return NULL during Shutdown.
488 if (!shm)
489 return NULL;
490 return make_scoped_ptr(new SHMBuffer(shm.Pass(), size_to_allocate));
492 scoped_ptr<SHMBuffer> ret(available_shm_segments_.back());
493 available_shm_segments_.pop_back();
494 return ret.Pass();
497 void GpuVideoDecoder::PutSHM(scoped_ptr<SHMBuffer> shm_buffer) {
498 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
499 available_shm_segments_.push_back(shm_buffer.release());
502 void GpuVideoDecoder::NotifyEndOfBitstreamBuffer(int32 id) {
503 DVLOG(3) << "NotifyEndOfBitstreamBuffer(" << id << ")";
504 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
506 std::map<int32, PendingDecoderBuffer>::iterator it =
507 bitstream_buffers_in_decoder_.find(id);
508 if (it == bitstream_buffers_in_decoder_.end()) {
509 NotifyError(VideoDecodeAccelerator::PLATFORM_FAILURE);
510 NOTREACHED() << "Missing bitstream buffer: " << id;
511 return;
514 PutSHM(make_scoped_ptr(it->second.shm_buffer));
515 it->second.done_cb.Run(state_ == kError ? kDecodeError : kOk);
516 bitstream_buffers_in_decoder_.erase(it);
519 GpuVideoDecoder::~GpuVideoDecoder() {
520 DVLOG(3) << __FUNCTION__;
521 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
523 if (vda_)
524 DestroyVDA();
525 DCHECK(assigned_picture_buffers_.empty());
527 for (size_t i = 0; i < available_shm_segments_.size(); ++i) {
528 delete available_shm_segments_[i];
530 available_shm_segments_.clear();
532 for (std::map<int32, PendingDecoderBuffer>::iterator it =
533 bitstream_buffers_in_decoder_.begin();
534 it != bitstream_buffers_in_decoder_.end(); ++it) {
535 delete it->second.shm_buffer;
536 it->second.done_cb.Run(kAborted);
538 bitstream_buffers_in_decoder_.clear();
540 if (!pending_reset_cb_.is_null())
541 base::ResetAndReturn(&pending_reset_cb_).Run();
544 void GpuVideoDecoder::NotifyFlushDone() {
545 DVLOG(3) << "NotifyFlushDone()";
546 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
547 DCHECK_EQ(state_, kDrainingDecoder);
548 state_ = kDecoderDrained;
549 base::ResetAndReturn(&eos_decode_cb_).Run(kOk);
552 void GpuVideoDecoder::NotifyResetDone() {
553 DVLOG(3) << "NotifyResetDone()";
554 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
555 DCHECK(bitstream_buffers_in_decoder_.empty());
557 // This needs to happen after the Reset() on vda_ is done to ensure pictures
558 // delivered during the reset can find their time data.
559 input_buffer_data_.clear();
561 if (!pending_reset_cb_.is_null())
562 base::ResetAndReturn(&pending_reset_cb_).Run();
565 void GpuVideoDecoder::NotifyError(media::VideoDecodeAccelerator::Error error) {
566 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
567 if (!vda_)
568 return;
570 state_ = kError;
572 DLOG(ERROR) << "VDA Error: " << error;
573 DestroyVDA();
576 bool GpuVideoDecoder::IsProfileSupported(VideoCodecProfile profile,
577 const gfx::Size& coded_size) {
578 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
579 VideoDecodeAccelerator::SupportedProfiles supported_profiles =
580 factories_->GetVideoDecodeAcceleratorSupportedProfiles();
581 for (const auto& supported_profile : supported_profiles) {
582 if (profile == supported_profile.profile) {
583 return IsCodedSizeSupported(coded_size,
584 supported_profile.min_resolution,
585 supported_profile.max_resolution);
588 return false;
591 void GpuVideoDecoder::DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent()
592 const {
593 DCHECK(factories_->GetTaskRunner()->BelongsToCurrentThread());
596 } // namespace media