cc: Added inline to Tile::IsReadyToDraw
[chromium-blink-merge.git] / media / filters / gpu_video_decoder.cc
blob273542e85ee22a1623e560a9895ba1bf42aa8307
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/cpu.h"
12 #include "base/message_loop/message_loop.h"
13 #include "base/stl_util.h"
14 #include "base/task_runner_util.h"
15 #include "media/base/bind_to_loop.h"
16 #include "media/base/decoder_buffer.h"
17 #include "media/base/media_log.h"
18 #include "media/base/pipeline.h"
19 #include "media/base/pipeline_status.h"
20 #include "media/base/video_decoder_config.h"
21 #include "media/filters/gpu_video_accelerator_factories.h"
23 namespace media {
25 // Maximum number of concurrent VDA::Decode() operations GVD will maintain.
26 // Higher values allow better pipelining in the GPU, but also require more
27 // resources.
28 enum { kMaxInFlightDecodes = 4 };
30 // Size of shared-memory segments we allocate. Since we reuse them we let them
31 // be on the beefy side.
32 static const size_t kSharedMemorySegmentBytes = 100 << 10;
34 GpuVideoDecoder::SHMBuffer::SHMBuffer(base::SharedMemory* m, size_t s)
35 : shm(m), size(s) {
38 GpuVideoDecoder::SHMBuffer::~SHMBuffer() {}
40 GpuVideoDecoder::BufferPair::BufferPair(
41 SHMBuffer* s, const scoped_refptr<DecoderBuffer>& b)
42 : shm_buffer(s), buffer(b) {
45 GpuVideoDecoder::BufferPair::~BufferPair() {}
47 GpuVideoDecoder::BufferData::BufferData(
48 int32 bbid, base::TimeDelta ts, const gfx::Rect& vr, const gfx::Size& ns)
49 : bitstream_buffer_id(bbid), timestamp(ts), visible_rect(vr),
50 natural_size(ns) {
53 GpuVideoDecoder::BufferData::~BufferData() {}
55 GpuVideoDecoder::GpuVideoDecoder(
56 const scoped_refptr<GpuVideoAcceleratorFactories>& factories,
57 const scoped_refptr<MediaLog>& media_log)
58 : needs_bitstream_conversion_(false),
59 gvd_loop_proxy_(factories->GetMessageLoop()),
60 weak_factory_(this),
61 factories_(factories),
62 state_(kNormal),
63 media_log_(media_log),
64 decoder_texture_target_(0),
65 next_picture_buffer_id_(0),
66 next_bitstream_buffer_id_(0),
67 available_pictures_(0) {
68 DCHECK(factories_.get());
71 void GpuVideoDecoder::Reset(const base::Closure& closure) {
72 DVLOG(3) << "Reset()";
73 DCHECK(gvd_loop_proxy_->BelongsToCurrentThread());
75 if (state_ == kDrainingDecoder && !factories_->IsAborted()) {
76 gvd_loop_proxy_->PostTask(FROM_HERE, base::Bind(
77 &GpuVideoDecoder::Reset, weak_this_, closure));
78 // NOTE: if we're deferring Reset() until a Flush() completes, return
79 // queued pictures to the VDA so they can be used to finish that Flush().
80 if (pending_decode_cb_.is_null())
81 ready_video_frames_.clear();
82 return;
85 // Throw away any already-decoded, not-yet-delivered frames.
86 ready_video_frames_.clear();
88 if (!vda_) {
89 gvd_loop_proxy_->PostTask(FROM_HERE, closure);
90 return;
93 if (!pending_decode_cb_.is_null())
94 EnqueueFrameAndTriggerFrameDelivery(VideoFrame::CreateEmptyFrame());
96 DCHECK(pending_reset_cb_.is_null());
97 pending_reset_cb_ = BindToCurrentLoop(closure);
99 vda_->Reset();
102 void GpuVideoDecoder::Stop(const base::Closure& closure) {
103 DCHECK(gvd_loop_proxy_->BelongsToCurrentThread());
104 if (vda_)
105 DestroyVDA();
106 if (!pending_decode_cb_.is_null())
107 EnqueueFrameAndTriggerFrameDelivery(VideoFrame::CreateEmptyFrame());
108 if (!pending_reset_cb_.is_null())
109 base::ResetAndReturn(&pending_reset_cb_).Run();
110 BindToCurrentLoop(closure).Run();
113 static bool IsCodedSizeSupported(const gfx::Size& coded_size) {
114 // Only non-Windows, Ivy Bridge+ platforms can support more than 1920x1080.
115 // We test against 1088 to account for 16x16 macroblocks.
116 if (coded_size.width() <= 1920 && coded_size.height() <= 1088)
117 return true;
119 base::CPU cpu;
120 bool hw_large_video_support =
121 (cpu.vendor_name() == "GenuineIntel") && cpu.model() >= 58;
122 bool os_large_video_support = true;
123 #if defined(OS_WIN)
124 os_large_video_support = false;
125 #endif
126 return os_large_video_support && hw_large_video_support;
129 void GpuVideoDecoder::Initialize(const VideoDecoderConfig& config,
130 const PipelineStatusCB& orig_status_cb) {
131 DVLOG(3) << "Initialize()";
132 DCHECK(gvd_loop_proxy_->BelongsToCurrentThread());
133 DCHECK(config.IsValidConfig());
134 DCHECK(!config.is_encrypted());
136 weak_this_ = weak_factory_.GetWeakPtr();
138 PipelineStatusCB status_cb = CreateUMAReportingPipelineCB(
139 "Media.GpuVideoDecoderInitializeStatus",
140 BindToCurrentLoop(orig_status_cb));
142 bool previously_initialized = config_.IsValidConfig();
143 #if !defined(OS_CHROMEOS) && !defined(OS_WIN)
144 if (previously_initialized) {
145 // TODO(xhwang): Make GpuVideoDecoder reinitializable.
146 // See http://crbug.com/233608
147 DVLOG(1) << "GpuVideoDecoder reinitialization not supported.";
148 status_cb.Run(DECODER_ERROR_NOT_SUPPORTED);
149 return;
151 #endif
152 DVLOG(1) << "(Re)initializing GVD with config: "
153 << config.AsHumanReadableString();
155 // TODO(posciak): destroy and create a new VDA on codec/profile change
156 // (http://crbug.com/260224).
157 if (previously_initialized && (config_.profile() != config.profile())) {
158 DVLOG(1) << "Codec or profile changed, cannot reinitialize.";
159 status_cb.Run(DECODER_ERROR_NOT_SUPPORTED);
160 return;
163 if (!IsCodedSizeSupported(config.coded_size())) {
164 status_cb.Run(DECODER_ERROR_NOT_SUPPORTED);
165 return;
168 config_ = config;
169 needs_bitstream_conversion_ = (config.codec() == kCodecH264);
171 if (previously_initialized) {
172 // Reinitialization with a different config (but same codec and profile).
173 // VDA should handle it by detecting this in-stream by itself,
174 // no need to notify it.
175 status_cb.Run(PIPELINE_OK);
176 return;
179 vda_ =
180 factories_->CreateVideoDecodeAccelerator(config.profile(), this).Pass();
181 if (!vda_) {
182 status_cb.Run(DECODER_ERROR_NOT_SUPPORTED);
183 return;
186 DVLOG(3) << "GpuVideoDecoder::Initialize() succeeded.";
187 media_log_->SetStringProperty("video_decoder", "gpu");
188 status_cb.Run(PIPELINE_OK);
191 void GpuVideoDecoder::DestroyTextures() {
192 std::map<int32, PictureBuffer>::iterator it;
194 for (it = assigned_picture_buffers_.begin();
195 it != assigned_picture_buffers_.end(); ++it) {
196 factories_->DeleteTexture(it->second.texture_id());
198 assigned_picture_buffers_.clear();
200 for (it = dismissed_picture_buffers_.begin();
201 it != dismissed_picture_buffers_.end(); ++it) {
202 factories_->DeleteTexture(it->second.texture_id());
204 dismissed_picture_buffers_.clear();
207 void GpuVideoDecoder::DestroyVDA() {
208 DCHECK(gvd_loop_proxy_->BelongsToCurrentThread());
210 if (vda_)
211 vda_.release()->Destroy();
213 DestroyTextures();
216 void GpuVideoDecoder::Decode(const scoped_refptr<DecoderBuffer>& buffer,
217 const DecodeCB& decode_cb) {
218 DCHECK(gvd_loop_proxy_->BelongsToCurrentThread());
219 DCHECK(pending_reset_cb_.is_null());
220 DCHECK(pending_decode_cb_.is_null());
222 pending_decode_cb_ = BindToCurrentLoop(decode_cb);
224 if (state_ == kError || !vda_) {
225 base::ResetAndReturn(&pending_decode_cb_).Run(kDecodeError, NULL);
226 return;
229 switch (state_) {
230 case kDecoderDrained:
231 if (!ready_video_frames_.empty()) {
232 EnqueueFrameAndTriggerFrameDelivery(NULL);
233 return;
235 state_ = kNormal;
236 // Fall-through.
237 case kNormal:
238 break;
239 case kDrainingDecoder:
240 DCHECK(buffer->end_of_stream());
241 // Do nothing. Will be satisfied either by a PictureReady or
242 // NotifyFlushDone below.
243 return;
244 case kError:
245 NOTREACHED();
246 return;
249 if (buffer->end_of_stream()) {
250 if (state_ == kNormal) {
251 state_ = kDrainingDecoder;
252 vda_->Flush();
253 // If we have ready frames, go ahead and process them to ensure that the
254 // Flush operation does not block in the VDA due to lack of picture
255 // buffers.
256 if (!ready_video_frames_.empty())
257 EnqueueFrameAndTriggerFrameDelivery(NULL);
259 return;
262 size_t size = buffer->data_size();
263 SHMBuffer* shm_buffer = GetSHM(size);
264 if (!shm_buffer) {
265 base::ResetAndReturn(&pending_decode_cb_).Run(kDecodeError, NULL);
266 return;
269 memcpy(shm_buffer->shm->memory(), buffer->data(), size);
270 BitstreamBuffer bitstream_buffer(
271 next_bitstream_buffer_id_, shm_buffer->shm->handle(), size);
272 // Mask against 30 bits, to avoid (undefined) wraparound on signed integer.
273 next_bitstream_buffer_id_ = (next_bitstream_buffer_id_ + 1) & 0x3FFFFFFF;
274 bool inserted = bitstream_buffers_in_decoder_.insert(std::make_pair(
275 bitstream_buffer.id(), BufferPair(shm_buffer, buffer))).second;
276 DCHECK(inserted);
277 RecordBufferData(bitstream_buffer, *buffer.get());
279 vda_->Decode(bitstream_buffer);
281 if (!ready_video_frames_.empty()) {
282 EnqueueFrameAndTriggerFrameDelivery(NULL);
283 return;
286 if (CanMoreDecodeWorkBeDone())
287 base::ResetAndReturn(&pending_decode_cb_).Run(kNotEnoughData, NULL);
290 bool GpuVideoDecoder::CanMoreDecodeWorkBeDone() {
291 return bitstream_buffers_in_decoder_.size() < kMaxInFlightDecodes;
294 void GpuVideoDecoder::RecordBufferData(const BitstreamBuffer& bitstream_buffer,
295 const DecoderBuffer& buffer) {
296 input_buffer_data_.push_front(BufferData(bitstream_buffer.id(),
297 buffer.timestamp(),
298 config_.visible_rect(),
299 config_.natural_size()));
300 // Why this value? Because why not. avformat.h:MAX_REORDER_DELAY is 16, but
301 // that's too small for some pathological B-frame test videos. The cost of
302 // using too-high a value is low (192 bits per extra slot).
303 static const size_t kMaxInputBufferDataSize = 128;
304 // Pop from the back of the list, because that's the oldest and least likely
305 // to be useful in the future data.
306 if (input_buffer_data_.size() > kMaxInputBufferDataSize)
307 input_buffer_data_.pop_back();
310 void GpuVideoDecoder::GetBufferData(int32 id, base::TimeDelta* timestamp,
311 gfx::Rect* visible_rect,
312 gfx::Size* natural_size) {
313 for (std::list<BufferData>::const_iterator it =
314 input_buffer_data_.begin(); it != input_buffer_data_.end();
315 ++it) {
316 if (it->bitstream_buffer_id != id)
317 continue;
318 *timestamp = it->timestamp;
319 *visible_rect = it->visible_rect;
320 *natural_size = it->natural_size;
321 return;
323 NOTREACHED() << "Missing bitstreambuffer id: " << id;
326 bool GpuVideoDecoder::HasAlpha() const {
327 DCHECK(gvd_loop_proxy_->BelongsToCurrentThread());
328 return true;
331 bool GpuVideoDecoder::NeedsBitstreamConversion() const {
332 DCHECK(gvd_loop_proxy_->BelongsToCurrentThread());
333 return needs_bitstream_conversion_;
336 bool GpuVideoDecoder::CanReadWithoutStalling() const {
337 DCHECK(gvd_loop_proxy_->BelongsToCurrentThread());
338 return available_pictures_ > 0 || !ready_video_frames_.empty();
341 void GpuVideoDecoder::NotifyInitializeDone() {
342 NOTREACHED() << "GpuVideoDecodeAcceleratorHost::Initialize is synchronous!";
345 void GpuVideoDecoder::ProvidePictureBuffers(uint32 count,
346 const gfx::Size& size,
347 uint32 texture_target) {
348 DVLOG(3) << "ProvidePictureBuffers(" << count << ", "
349 << size.width() << "x" << size.height() << ")";
350 DCHECK(gvd_loop_proxy_->BelongsToCurrentThread());
352 std::vector<uint32> texture_ids;
353 std::vector<gpu::Mailbox> texture_mailboxes;
354 decoder_texture_target_ = texture_target;
355 // Discards the sync point returned here since PictureReady will imply that
356 // the produce has already happened, and the texture is ready for use.
357 if (!factories_->CreateTextures(count,
358 size,
359 &texture_ids,
360 &texture_mailboxes,
361 decoder_texture_target_)) {
362 NotifyError(VideoDecodeAccelerator::PLATFORM_FAILURE);
363 return;
365 DCHECK_EQ(count, texture_ids.size());
366 DCHECK_EQ(count, texture_mailboxes.size());
368 if (!vda_)
369 return;
371 std::vector<PictureBuffer> picture_buffers;
372 for (size_t i = 0; i < texture_ids.size(); ++i) {
373 picture_buffers.push_back(PictureBuffer(
374 next_picture_buffer_id_++, size, texture_ids[i], texture_mailboxes[i]));
375 bool inserted = assigned_picture_buffers_.insert(std::make_pair(
376 picture_buffers.back().id(), picture_buffers.back())).second;
377 DCHECK(inserted);
380 available_pictures_ += count;
382 vda_->AssignPictureBuffers(picture_buffers);
385 void GpuVideoDecoder::DismissPictureBuffer(int32 id) {
386 DVLOG(3) << "DismissPictureBuffer(" << id << ")";
387 DCHECK(gvd_loop_proxy_->BelongsToCurrentThread());
389 std::map<int32, PictureBuffer>::iterator it =
390 assigned_picture_buffers_.find(id);
391 if (it == assigned_picture_buffers_.end()) {
392 NOTREACHED() << "Missing picture buffer: " << id;
393 return;
396 PictureBuffer buffer_to_dismiss = it->second;
397 assigned_picture_buffers_.erase(it);
399 std::set<int32>::iterator at_display_it =
400 picture_buffers_at_display_.find(id);
402 if (at_display_it == picture_buffers_at_display_.end()) {
403 // We can delete the texture immediately as it's not being displayed.
404 factories_->DeleteTexture(buffer_to_dismiss.texture_id());
405 CHECK_GT(available_pictures_, 0);
406 --available_pictures_;
407 } else {
408 // Texture in display. Postpone deletion until after it's returned to us.
409 bool inserted = dismissed_picture_buffers_.insert(std::make_pair(
410 id, buffer_to_dismiss)).second;
411 DCHECK(inserted);
415 void GpuVideoDecoder::PictureReady(const media::Picture& picture) {
416 DVLOG(3) << "PictureReady()";
417 DCHECK(gvd_loop_proxy_->BelongsToCurrentThread());
419 std::map<int32, PictureBuffer>::iterator it =
420 assigned_picture_buffers_.find(picture.picture_buffer_id());
421 if (it == assigned_picture_buffers_.end()) {
422 NOTREACHED() << "Missing picture buffer: " << picture.picture_buffer_id();
423 NotifyError(VideoDecodeAccelerator::PLATFORM_FAILURE);
424 return;
426 const PictureBuffer& pb = it->second;
428 // Update frame's timestamp.
429 base::TimeDelta timestamp;
430 gfx::Rect visible_rect;
431 gfx::Size natural_size;
432 GetBufferData(picture.bitstream_buffer_id(), &timestamp, &visible_rect,
433 &natural_size);
434 DCHECK(decoder_texture_target_);
436 scoped_refptr<VideoFrame> frame(VideoFrame::WrapNativeTexture(
437 new VideoFrame::MailboxHolder(
438 pb.texture_mailbox(),
439 0, // sync_point
440 BindToCurrentLoop(base::Bind(&GpuVideoDecoder::ReusePictureBuffer,
441 weak_this_,
442 picture.picture_buffer_id()))),
443 decoder_texture_target_,
444 pb.size(),
445 visible_rect,
446 natural_size,
447 timestamp,
448 base::Bind(&GpuVideoAcceleratorFactories::ReadPixels,
449 factories_,
450 pb.texture_id(),
451 decoder_texture_target_,
452 gfx::Size(visible_rect.width(), visible_rect.height())),
453 base::Closure()));
454 CHECK_GT(available_pictures_, 0);
455 --available_pictures_;
456 bool inserted =
457 picture_buffers_at_display_.insert(picture.picture_buffer_id()).second;
458 DCHECK(inserted);
460 EnqueueFrameAndTriggerFrameDelivery(frame);
463 void GpuVideoDecoder::EnqueueFrameAndTriggerFrameDelivery(
464 const scoped_refptr<VideoFrame>& frame) {
465 DCHECK(gvd_loop_proxy_->BelongsToCurrentThread());
467 // During a pending vda->Reset(), we don't accumulate frames. Drop it on the
468 // floor and return.
469 if (!pending_reset_cb_.is_null())
470 return;
472 if (frame.get())
473 ready_video_frames_.push_back(frame);
474 else
475 DCHECK(!ready_video_frames_.empty());
477 if (pending_decode_cb_.is_null())
478 return;
480 base::ResetAndReturn(&pending_decode_cb_)
481 .Run(kOk, ready_video_frames_.front());
482 ready_video_frames_.pop_front();
485 void GpuVideoDecoder::ReusePictureBuffer(int64 picture_buffer_id,
486 uint32 sync_point) {
487 DVLOG(3) << "ReusePictureBuffer(" << picture_buffer_id << ")";
488 DCHECK(gvd_loop_proxy_->BelongsToCurrentThread());
490 if (!vda_)
491 return;
493 CHECK(!picture_buffers_at_display_.empty());
495 size_t num_erased = picture_buffers_at_display_.erase(picture_buffer_id);
496 DCHECK(num_erased);
498 std::map<int32, PictureBuffer>::iterator it =
499 assigned_picture_buffers_.find(picture_buffer_id);
501 if (it == assigned_picture_buffers_.end()) {
502 // This picture was dismissed while in display, so we postponed deletion.
503 it = dismissed_picture_buffers_.find(picture_buffer_id);
504 DCHECK(it != dismissed_picture_buffers_.end());
505 factories_->DeleteTexture(it->second.texture_id());
506 dismissed_picture_buffers_.erase(it);
507 return;
510 factories_->WaitSyncPoint(sync_point);
511 ++available_pictures_;
513 vda_->ReusePictureBuffer(picture_buffer_id);
516 GpuVideoDecoder::SHMBuffer* GpuVideoDecoder::GetSHM(size_t min_size) {
517 DCHECK(gvd_loop_proxy_->BelongsToCurrentThread());
518 if (available_shm_segments_.empty() ||
519 available_shm_segments_.back()->size < min_size) {
520 size_t size_to_allocate = std::max(min_size, kSharedMemorySegmentBytes);
521 base::SharedMemory* shm = factories_->CreateSharedMemory(size_to_allocate);
522 // CreateSharedMemory() can return NULL during Shutdown.
523 if (!shm)
524 return NULL;
525 return new SHMBuffer(shm, size_to_allocate);
527 SHMBuffer* ret = available_shm_segments_.back();
528 available_shm_segments_.pop_back();
529 return ret;
532 void GpuVideoDecoder::PutSHM(SHMBuffer* shm_buffer) {
533 DCHECK(gvd_loop_proxy_->BelongsToCurrentThread());
534 available_shm_segments_.push_back(shm_buffer);
537 void GpuVideoDecoder::NotifyEndOfBitstreamBuffer(int32 id) {
538 DVLOG(3) << "NotifyEndOfBitstreamBuffer(" << id << ")";
539 DCHECK(gvd_loop_proxy_->BelongsToCurrentThread());
541 std::map<int32, BufferPair>::iterator it =
542 bitstream_buffers_in_decoder_.find(id);
543 if (it == bitstream_buffers_in_decoder_.end()) {
544 NotifyError(VideoDecodeAccelerator::PLATFORM_FAILURE);
545 NOTREACHED() << "Missing bitstream buffer: " << id;
546 return;
549 PutSHM(it->second.shm_buffer);
550 bitstream_buffers_in_decoder_.erase(it);
552 if (pending_reset_cb_.is_null() && state_ != kDrainingDecoder &&
553 CanMoreDecodeWorkBeDone() && !pending_decode_cb_.is_null()) {
554 base::ResetAndReturn(&pending_decode_cb_).Run(kNotEnoughData, NULL);
558 GpuVideoDecoder::~GpuVideoDecoder() {
559 DCHECK(!vda_.get()); // Stop should have been already called.
560 DCHECK(pending_decode_cb_.is_null());
561 for (size_t i = 0; i < available_shm_segments_.size(); ++i) {
562 available_shm_segments_[i]->shm->Close();
563 delete available_shm_segments_[i];
565 available_shm_segments_.clear();
566 for (std::map<int32, BufferPair>::iterator it =
567 bitstream_buffers_in_decoder_.begin();
568 it != bitstream_buffers_in_decoder_.end(); ++it) {
569 it->second.shm_buffer->shm->Close();
571 bitstream_buffers_in_decoder_.clear();
573 DestroyTextures();
576 void GpuVideoDecoder::NotifyFlushDone() {
577 DVLOG(3) << "NotifyFlushDone()";
578 DCHECK(gvd_loop_proxy_->BelongsToCurrentThread());
579 DCHECK_EQ(state_, kDrainingDecoder);
580 state_ = kDecoderDrained;
581 EnqueueFrameAndTriggerFrameDelivery(VideoFrame::CreateEmptyFrame());
584 void GpuVideoDecoder::NotifyResetDone() {
585 DVLOG(3) << "NotifyResetDone()";
586 DCHECK(gvd_loop_proxy_->BelongsToCurrentThread());
587 DCHECK(ready_video_frames_.empty());
589 // This needs to happen after the Reset() on vda_ is done to ensure pictures
590 // delivered during the reset can find their time data.
591 input_buffer_data_.clear();
593 if (!pending_reset_cb_.is_null())
594 base::ResetAndReturn(&pending_reset_cb_).Run();
596 if (!pending_decode_cb_.is_null())
597 EnqueueFrameAndTriggerFrameDelivery(VideoFrame::CreateEmptyFrame());
600 void GpuVideoDecoder::NotifyError(media::VideoDecodeAccelerator::Error error) {
601 DCHECK(gvd_loop_proxy_->BelongsToCurrentThread());
602 if (!vda_)
603 return;
605 DLOG(ERROR) << "VDA Error: " << error;
606 DestroyVDA();
608 state_ = kError;
610 if (!pending_decode_cb_.is_null()) {
611 base::ResetAndReturn(&pending_decode_cb_).Run(kDecodeError, NULL);
612 return;
616 } // namespace media