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"
10 #include "base/callback_helpers.h"
11 #include "base/command_line.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/pipeline_status.h"
23 #include "media/base/video_decoder_config.h"
24 #include "media/renderers/gpu_video_accelerator_factories.h"
25 #include "third_party/skia/include/core/SkBitmap.h"
29 const char GpuVideoDecoder::kDecoderName
[] = "GpuVideoDecoder";
31 // Maximum number of concurrent VDA::Decode() operations GVD will maintain.
32 // Higher values allow better pipelining in the GPU, but also require more
34 enum { kMaxInFlightDecodes
= 4 };
36 // Size of shared-memory segments we allocate. Since we reuse them we let them
37 // be on the beefy side.
38 static const size_t kSharedMemorySegmentBytes
= 100 << 10;
40 GpuVideoDecoder::SHMBuffer::SHMBuffer(scoped_ptr
<base::SharedMemory
> m
,
42 : shm(m
.Pass()), size(s
) {
45 GpuVideoDecoder::SHMBuffer::~SHMBuffer() {}
47 GpuVideoDecoder::PendingDecoderBuffer::PendingDecoderBuffer(
49 const scoped_refptr
<DecoderBuffer
>& b
,
50 const DecodeCB
& done_cb
)
51 : shm_buffer(s
), buffer(b
), done_cb(done_cb
) {
54 GpuVideoDecoder::PendingDecoderBuffer::~PendingDecoderBuffer() {}
56 GpuVideoDecoder::BufferData::BufferData(
57 int32 bbid
, base::TimeDelta ts
, const gfx::Rect
& vr
, const gfx::Size
& ns
)
58 : bitstream_buffer_id(bbid
), timestamp(ts
), visible_rect(vr
),
62 GpuVideoDecoder::BufferData::~BufferData() {}
64 GpuVideoDecoder::GpuVideoDecoder(
65 const scoped_refptr
<GpuVideoAcceleratorFactories
>& factories
)
66 : needs_bitstream_conversion_(false),
67 factories_(factories
),
69 decoder_texture_target_(0),
70 next_picture_buffer_id_(0),
71 next_bitstream_buffer_id_(0),
72 available_pictures_(0),
74 DCHECK(factories_
.get());
77 void GpuVideoDecoder::Reset(const base::Closure
& closure
) {
78 DVLOG(3) << "Reset()";
79 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
81 if (state_
== kDrainingDecoder
) {
82 base::MessageLoop::current()->PostTask(
85 &GpuVideoDecoder::Reset
, weak_factory_
.GetWeakPtr(), closure
));
90 base::MessageLoop::current()->PostTask(FROM_HERE
, closure
);
94 DCHECK(pending_reset_cb_
.is_null());
95 pending_reset_cb_
= BindToCurrentLoop(closure
);
100 static bool IsCodedSizeSupported(const gfx::Size
& coded_size
,
101 VideoCodecProfile profile
) {
103 // Windows Media Foundation H.264 decoding does not support decoding videos
104 // with any dimension smaller than 48 pixels:
105 // http://msdn.microsoft.com/en-us/library/windows/desktop/dd797815
106 if (coded_size
.width() < 48 || coded_size
.height() < 48)
110 // Only non-Windows, Ivy Bridge+ platforms can support more than 1920x1080.
111 // We test against 1088 to account for 16x16 macroblocks.
112 if (coded_size
.width() <= 1920 && coded_size
.height() <= 1088)
115 // NOTE: additional autodetection logic may require updating input buffer size
116 // selection in platform-specific implementations, such as
117 // V4L2VideoDecodeAccelerator.
119 bool hw_large_video_support
=
120 base::CommandLine::ForCurrentProcess()->HasSwitch(
121 switches::kIgnoreResolutionLimitsForAcceleratedVideoDecode
) ||
122 ((cpu
.vendor_name() == "GenuineIntel") && cpu
.model() >= 55 &&
123 // TODO(posciak, henryhsu): Remove this once we can query in runtime.
124 profile
>= H264PROFILE_MIN
&& profile
<= H264PROFILE_MAX
);
125 bool os_large_video_support
= true;
127 os_large_video_support
= false;
129 return os_large_video_support
&& hw_large_video_support
;
132 // Report |status| to UMA and run |cb| with it. This is super-specific to the
133 // UMA stat reported because the UMA_HISTOGRAM_ENUMERATION API requires a
134 // callsite to always be called with the same stat name (can't parameterize it).
135 static void ReportGpuVideoDecoderInitializeStatusToUMAAndRunCB(
136 const PipelineStatusCB
& cb
,
137 PipelineStatus status
) {
138 UMA_HISTOGRAM_ENUMERATION(
139 "Media.GpuVideoDecoderInitializeStatus", status
, PIPELINE_STATUS_MAX
+ 1);
143 std::string
GpuVideoDecoder::GetDisplayName() const {
147 void GpuVideoDecoder::Initialize(const VideoDecoderConfig
& config
,
148 bool /* low_delay */,
149 const PipelineStatusCB
& orig_status_cb
,
150 const OutputCB
& output_cb
) {
151 DVLOG(3) << "Initialize()";
152 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
153 DCHECK(config
.IsValidConfig());
154 DCHECK(!config
.is_encrypted());
156 PipelineStatusCB status_cb
=
157 base::Bind(&ReportGpuVideoDecoderInitializeStatusToUMAAndRunCB
,
158 BindToCurrentLoop(orig_status_cb
));
160 bool previously_initialized
= config_
.IsValidConfig();
161 DVLOG(1) << "(Re)initializing GVD with config: "
162 << config
.AsHumanReadableString();
164 // TODO(posciak): destroy and create a new VDA on codec/profile change
165 // (http://crbug.com/260224).
166 if (previously_initialized
&& (config_
.profile() != config
.profile())) {
167 DVLOG(1) << "Codec or profile changed, cannot reinitialize.";
168 status_cb
.Run(DECODER_ERROR_NOT_SUPPORTED
);
172 if (!IsCodedSizeSupported(config
.coded_size(), config
.profile())) {
173 status_cb
.Run(DECODER_ERROR_NOT_SUPPORTED
);
178 needs_bitstream_conversion_
= (config
.codec() == kCodecH264
);
179 output_cb_
= BindToCurrentLoop(output_cb
);
181 if (previously_initialized
) {
182 // Reinitialization with a different config (but same codec and profile).
183 // VDA should handle it by detecting this in-stream by itself,
184 // no need to notify it.
185 status_cb
.Run(PIPELINE_OK
);
189 vda_
= factories_
->CreateVideoDecodeAccelerator().Pass();
190 if (!vda_
|| !vda_
->Initialize(config
.profile(), this)) {
191 status_cb
.Run(DECODER_ERROR_NOT_SUPPORTED
);
195 DVLOG(3) << "GpuVideoDecoder::Initialize() succeeded.";
196 status_cb
.Run(PIPELINE_OK
);
199 void GpuVideoDecoder::DestroyPictureBuffers(PictureBufferMap
* buffers
) {
200 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
201 for (PictureBufferMap::iterator it
= buffers
->begin(); it
!= buffers
->end();
203 factories_
->DeleteTexture(it
->second
.texture_id());
209 void GpuVideoDecoder::DestroyVDA() {
210 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
214 // Not destroying PictureBuffers in |picture_buffers_at_display_| yet, since
215 // their textures may still be in use by the user of this GpuVideoDecoder.
216 for (PictureBufferTextureMap::iterator it
=
217 picture_buffers_at_display_
.begin();
218 it
!= picture_buffers_at_display_
.end();
220 assigned_picture_buffers_
.erase(it
->first
);
222 DestroyPictureBuffers(&assigned_picture_buffers_
);
225 void GpuVideoDecoder::Decode(const scoped_refptr
<DecoderBuffer
>& buffer
,
226 const DecodeCB
& decode_cb
) {
227 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
228 DCHECK(pending_reset_cb_
.is_null());
230 DecodeCB bound_decode_cb
= BindToCurrentLoop(decode_cb
);
232 if (state_
== kError
|| !vda_
) {
233 bound_decode_cb
.Run(kDecodeError
);
238 case kDecoderDrained
:
243 case kDrainingDecoder
:
249 DCHECK_EQ(state_
, kNormal
);
251 if (buffer
->end_of_stream()) {
252 state_
= kDrainingDecoder
;
253 eos_decode_cb_
= bound_decode_cb
;
258 size_t size
= buffer
->data_size();
259 scoped_ptr
<SHMBuffer
> shm_buffer
= GetSHM(size
);
261 bound_decode_cb
.Run(kDecodeError
);
265 memcpy(shm_buffer
->shm
->memory(), buffer
->data(), size
);
266 BitstreamBuffer
bitstream_buffer(
267 next_bitstream_buffer_id_
, shm_buffer
->shm
->handle(), size
);
268 // Mask against 30 bits, to avoid (undefined) wraparound on signed integer.
269 next_bitstream_buffer_id_
= (next_bitstream_buffer_id_
+ 1) & 0x3FFFFFFF;
270 DCHECK(!ContainsKey(bitstream_buffers_in_decoder_
, bitstream_buffer
.id()));
271 bitstream_buffers_in_decoder_
.insert(std::make_pair(
272 bitstream_buffer
.id(),
273 PendingDecoderBuffer(shm_buffer
.release(), buffer
, decode_cb
)));
274 DCHECK_LE(static_cast<int>(bitstream_buffers_in_decoder_
.size()),
275 kMaxInFlightDecodes
);
276 RecordBufferData(bitstream_buffer
, *buffer
.get());
278 vda_
->Decode(bitstream_buffer
);
281 void GpuVideoDecoder::RecordBufferData(const BitstreamBuffer
& bitstream_buffer
,
282 const DecoderBuffer
& buffer
) {
283 input_buffer_data_
.push_front(BufferData(bitstream_buffer
.id(),
285 config_
.visible_rect(),
286 config_
.natural_size()));
287 // Why this value? Because why not. avformat.h:MAX_REORDER_DELAY is 16, but
288 // that's too small for some pathological B-frame test videos. The cost of
289 // using too-high a value is low (192 bits per extra slot).
290 static const size_t kMaxInputBufferDataSize
= 128;
291 // Pop from the back of the list, because that's the oldest and least likely
292 // to be useful in the future data.
293 if (input_buffer_data_
.size() > kMaxInputBufferDataSize
)
294 input_buffer_data_
.pop_back();
297 void GpuVideoDecoder::GetBufferData(int32 id
, base::TimeDelta
* timestamp
,
298 gfx::Rect
* visible_rect
,
299 gfx::Size
* natural_size
) {
300 for (std::list
<BufferData
>::const_iterator it
=
301 input_buffer_data_
.begin(); it
!= input_buffer_data_
.end();
303 if (it
->bitstream_buffer_id
!= id
)
305 *timestamp
= it
->timestamp
;
306 *visible_rect
= it
->visible_rect
;
307 *natural_size
= it
->natural_size
;
310 NOTREACHED() << "Missing bitstreambuffer id: " << id
;
313 bool GpuVideoDecoder::NeedsBitstreamConversion() const {
314 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
315 return needs_bitstream_conversion_
;
318 bool GpuVideoDecoder::CanReadWithoutStalling() const {
319 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
321 next_picture_buffer_id_
== 0 || // Decode() will ProvidePictureBuffers().
322 available_pictures_
> 0;
325 int GpuVideoDecoder::GetMaxDecodeRequests() const {
326 return kMaxInFlightDecodes
;
329 void GpuVideoDecoder::ProvidePictureBuffers(uint32 count
,
330 const gfx::Size
& size
,
331 uint32 texture_target
) {
332 DVLOG(3) << "ProvidePictureBuffers(" << count
<< ", "
333 << size
.width() << "x" << size
.height() << ")";
334 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
336 std::vector
<uint32
> texture_ids
;
337 std::vector
<gpu::Mailbox
> texture_mailboxes
;
338 decoder_texture_target_
= texture_target
;
339 if (!factories_
->CreateTextures(count
,
343 decoder_texture_target_
)) {
344 NotifyError(VideoDecodeAccelerator::PLATFORM_FAILURE
);
347 DCHECK_EQ(count
, texture_ids
.size());
348 DCHECK_EQ(count
, texture_mailboxes
.size());
353 std::vector
<PictureBuffer
> picture_buffers
;
354 for (size_t i
= 0; i
< texture_ids
.size(); ++i
) {
355 picture_buffers
.push_back(PictureBuffer(
356 next_picture_buffer_id_
++, size
, texture_ids
[i
], texture_mailboxes
[i
]));
357 bool inserted
= assigned_picture_buffers_
.insert(std::make_pair(
358 picture_buffers
.back().id(), picture_buffers
.back())).second
;
362 available_pictures_
+= count
;
364 vda_
->AssignPictureBuffers(picture_buffers
);
367 void GpuVideoDecoder::DismissPictureBuffer(int32 id
) {
368 DVLOG(3) << "DismissPictureBuffer(" << id
<< ")";
369 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
371 PictureBufferMap::iterator it
= assigned_picture_buffers_
.find(id
);
372 if (it
== assigned_picture_buffers_
.end()) {
373 NOTREACHED() << "Missing picture buffer: " << id
;
377 PictureBuffer buffer_to_dismiss
= it
->second
;
378 assigned_picture_buffers_
.erase(it
);
380 if (!picture_buffers_at_display_
.count(id
)) {
381 // We can delete the texture immediately as it's not being displayed.
382 factories_
->DeleteTexture(buffer_to_dismiss
.texture_id());
383 CHECK_GT(available_pictures_
, 0);
384 --available_pictures_
;
386 // Not destroying a texture in display in |picture_buffers_at_display_|.
387 // Postpone deletion until after it's returned to us.
390 void GpuVideoDecoder::PictureReady(const media::Picture
& picture
) {
391 DVLOG(3) << "PictureReady()";
392 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
394 PictureBufferMap::iterator it
=
395 assigned_picture_buffers_
.find(picture
.picture_buffer_id());
396 if (it
== assigned_picture_buffers_
.end()) {
397 NOTREACHED() << "Missing picture buffer: " << picture
.picture_buffer_id();
398 NotifyError(VideoDecodeAccelerator::PLATFORM_FAILURE
);
401 const PictureBuffer
& pb
= it
->second
;
403 // Validate picture rectangle from GPU. This is for sanity/security check
404 // even the rectangle is not used in this class.
405 if (picture
.visible_rect().IsEmpty() ||
406 !gfx::Rect(pb
.size()).Contains(picture
.visible_rect())) {
407 NOTREACHED() << "Invalid picture size from VDA: "
408 << picture
.visible_rect().ToString() << " should fit in "
409 << pb
.size().ToString();
410 NotifyError(media::VideoDecodeAccelerator::PLATFORM_FAILURE
);
414 // Update frame's timestamp.
415 base::TimeDelta timestamp
;
416 // Some of the VDAs don't support and thus don't provide us with visible
417 // size in picture.size, passing coded size instead, so always drop it and
418 // use config information instead.
419 gfx::Rect visible_rect
;
420 gfx::Size natural_size
;
421 GetBufferData(picture
.bitstream_buffer_id(), ×tamp
, &visible_rect
,
423 DCHECK(decoder_texture_target_
);
425 scoped_refptr
<VideoFrame
> frame(VideoFrame::WrapNativeTexture(
426 make_scoped_ptr(new gpu::MailboxHolder(
427 pb
.texture_mailbox(), decoder_texture_target_
, 0 /* sync_point */)),
428 BindToCurrentLoop(base::Bind(
429 &GpuVideoDecoder::ReleaseMailbox
, weak_factory_
.GetWeakPtr(),
430 factories_
, picture
.picture_buffer_id(), pb
.texture_id())),
431 pb
.size(), visible_rect
, natural_size
, timestamp
,
432 picture
.allow_overlay()));
433 CHECK_GT(available_pictures_
, 0);
434 --available_pictures_
;
436 picture_buffers_at_display_
.insert(std::make_pair(
437 picture
.picture_buffer_id(),
438 pb
.texture_id())).second
;
444 void GpuVideoDecoder::DeliverFrame(
445 const scoped_refptr
<VideoFrame
>& frame
) {
446 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
448 // During a pending vda->Reset(), we don't accumulate frames. Drop it on the
450 if (!pending_reset_cb_
.is_null())
453 output_cb_
.Run(frame
);
457 void GpuVideoDecoder::ReleaseMailbox(
458 base::WeakPtr
<GpuVideoDecoder
> decoder
,
459 const scoped_refptr
<media::GpuVideoAcceleratorFactories
>& factories
,
460 int64 picture_buffer_id
,
462 uint32 release_sync_point
) {
463 DCHECK(factories
->GetTaskRunner()->BelongsToCurrentThread());
464 factories
->WaitSyncPoint(release_sync_point
);
467 decoder
->ReusePictureBuffer(picture_buffer_id
);
470 // It's the last chance to delete the texture after display,
471 // because GpuVideoDecoder was destructed.
472 factories
->DeleteTexture(texture_id
);
475 void GpuVideoDecoder::ReusePictureBuffer(int64 picture_buffer_id
) {
476 DVLOG(3) << "ReusePictureBuffer(" << picture_buffer_id
<< ")";
477 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
479 DCHECK(!picture_buffers_at_display_
.empty());
480 PictureBufferTextureMap::iterator display_iterator
=
481 picture_buffers_at_display_
.find(picture_buffer_id
);
482 uint32 texture_id
= display_iterator
->second
;
483 DCHECK(display_iterator
!= picture_buffers_at_display_
.end());
484 picture_buffers_at_display_
.erase(display_iterator
);
486 if (!assigned_picture_buffers_
.count(picture_buffer_id
)) {
487 // This picture was dismissed while in display, so we postponed deletion.
488 factories_
->DeleteTexture(texture_id
);
492 ++available_pictures_
;
494 // DestroyVDA() might already have been called.
496 vda_
->ReusePictureBuffer(picture_buffer_id
);
499 scoped_ptr
<GpuVideoDecoder::SHMBuffer
> GpuVideoDecoder::GetSHM(
501 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
502 if (available_shm_segments_
.empty() ||
503 available_shm_segments_
.back()->size
< min_size
) {
504 size_t size_to_allocate
= std::max(min_size
, kSharedMemorySegmentBytes
);
505 scoped_ptr
<base::SharedMemory
> shm
=
506 factories_
->CreateSharedMemory(size_to_allocate
);
507 // CreateSharedMemory() can return NULL during Shutdown.
510 return make_scoped_ptr(new SHMBuffer(shm
.Pass(), size_to_allocate
));
512 scoped_ptr
<SHMBuffer
> ret(available_shm_segments_
.back());
513 available_shm_segments_
.pop_back();
517 void GpuVideoDecoder::PutSHM(scoped_ptr
<SHMBuffer
> shm_buffer
) {
518 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
519 available_shm_segments_
.push_back(shm_buffer
.release());
522 void GpuVideoDecoder::NotifyEndOfBitstreamBuffer(int32 id
) {
523 DVLOG(3) << "NotifyEndOfBitstreamBuffer(" << id
<< ")";
524 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
526 std::map
<int32
, PendingDecoderBuffer
>::iterator it
=
527 bitstream_buffers_in_decoder_
.find(id
);
528 if (it
== bitstream_buffers_in_decoder_
.end()) {
529 NotifyError(VideoDecodeAccelerator::PLATFORM_FAILURE
);
530 NOTREACHED() << "Missing bitstream buffer: " << id
;
534 PutSHM(make_scoped_ptr(it
->second
.shm_buffer
));
535 it
->second
.done_cb
.Run(state_
== kError
? kDecodeError
: kOk
);
536 bitstream_buffers_in_decoder_
.erase(it
);
539 GpuVideoDecoder::~GpuVideoDecoder() {
540 DVLOG(3) << __FUNCTION__
;
541 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
545 DCHECK(assigned_picture_buffers_
.empty());
547 for (size_t i
= 0; i
< available_shm_segments_
.size(); ++i
) {
548 delete available_shm_segments_
[i
];
550 available_shm_segments_
.clear();
552 for (std::map
<int32
, PendingDecoderBuffer
>::iterator it
=
553 bitstream_buffers_in_decoder_
.begin();
554 it
!= bitstream_buffers_in_decoder_
.end(); ++it
) {
555 delete it
->second
.shm_buffer
;
556 it
->second
.done_cb
.Run(kAborted
);
558 bitstream_buffers_in_decoder_
.clear();
560 if (!pending_reset_cb_
.is_null())
561 base::ResetAndReturn(&pending_reset_cb_
).Run();
564 void GpuVideoDecoder::NotifyFlushDone() {
565 DVLOG(3) << "NotifyFlushDone()";
566 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
567 DCHECK_EQ(state_
, kDrainingDecoder
);
568 state_
= kDecoderDrained
;
569 base::ResetAndReturn(&eos_decode_cb_
).Run(kOk
);
572 void GpuVideoDecoder::NotifyResetDone() {
573 DVLOG(3) << "NotifyResetDone()";
574 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
575 DCHECK(bitstream_buffers_in_decoder_
.empty());
577 // This needs to happen after the Reset() on vda_ is done to ensure pictures
578 // delivered during the reset can find their time data.
579 input_buffer_data_
.clear();
581 if (!pending_reset_cb_
.is_null())
582 base::ResetAndReturn(&pending_reset_cb_
).Run();
585 void GpuVideoDecoder::NotifyError(media::VideoDecodeAccelerator::Error error
) {
586 DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent();
592 DLOG(ERROR
) << "VDA Error: " << error
;
596 void GpuVideoDecoder::DCheckGpuVideoAcceleratorFactoriesTaskRunnerIsCurrent()
598 DCHECK(factories_
->GetTaskRunner()->BelongsToCurrentThread());