OLD | NEW |
| (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 // This file contains an implementation of VideoDecodeAccelerator | |
6 // that utilizes hardware video decoders, which expose Video4Linux 2 API | |
7 // (http://linuxtv.org/downloads/v4l-dvb-apis/). | |
8 | |
9 #ifndef CONTENT_COMMON_GPU_MEDIA_V4L2_VIDEO_DECODE_ACCELERATOR_H_ | |
10 #define CONTENT_COMMON_GPU_MEDIA_V4L2_VIDEO_DECODE_ACCELERATOR_H_ | |
11 | |
12 #include <stddef.h> | |
13 #include <stdint.h> | |
14 | |
15 #include <memory> | |
16 #include <queue> | |
17 #include <vector> | |
18 | |
19 #include "base/callback_forward.h" | |
20 #include "base/macros.h" | |
21 #include "base/memory/linked_ptr.h" | |
22 #include "base/memory/ref_counted.h" | |
23 #include "base/synchronization/waitable_event.h" | |
24 #include "base/threading/thread.h" | |
25 #include "content/common/content_export.h" | |
26 #include "content/common/gpu/media/gpu_video_decode_accelerator_helpers.h" | |
27 #include "content/common/gpu/media/v4l2_device.h" | |
28 #include "content/common/gpu/media/v4l2_image_processor.h" | |
29 #include "media/base/limits.h" | |
30 #include "media/base/video_decoder_config.h" | |
31 #include "media/video/picture.h" | |
32 #include "media/video/video_decode_accelerator.h" | |
33 #include "ui/gfx/geometry/size.h" | |
34 #include "ui/gl/gl_bindings.h" | |
35 | |
36 namespace media { | |
37 class H264Parser; | |
38 } // namespace media | |
39 | |
40 namespace content { | |
41 // This class handles video accelerators directly through a V4L2 device exported | |
42 // by the hardware blocks. | |
43 // | |
44 // The threading model of this class is driven by the fact that it needs to | |
45 // interface two fundamentally different event queues -- the one Chromium | |
46 // provides through MessageLoop, and the one driven by the V4L2 devices which | |
47 // is waited on with epoll(). There are three threads involved in this class: | |
48 // | |
49 // * The child thread, which is the main GPU process thread which calls the | |
50 // media::VideoDecodeAccelerator entry points. Calls from this thread | |
51 // generally do not block (with the exception of Initialize() and Destroy()). | |
52 // They post tasks to the decoder_thread_, which actually services the task | |
53 // and calls back when complete through the | |
54 // media::VideoDecodeAccelerator::Client interface. | |
55 // * The decoder_thread_, owned by this class. It services API tasks, through | |
56 // the *Task() routines, as well as V4L2 device events, through | |
57 // ServiceDeviceTask(). Almost all state modification is done on this thread | |
58 // (this doesn't include buffer (re)allocation sequence, see below). | |
59 // * The device_poll_thread_, owned by this class. All it does is epoll() on | |
60 // the V4L2 in DevicePollTask() and schedule a ServiceDeviceTask() on the | |
61 // decoder_thread_ when something interesting happens. | |
62 // TODO(sheu): replace this thread with an TYPE_IO decoder_thread_. | |
63 // | |
64 // Note that this class has (almost) no locks, apart from the pictures_assigned_ | |
65 // WaitableEvent. Everything (apart from buffer (re)allocation) is serviced on | |
66 // the decoder_thread_, so there are no synchronization issues. | |
67 // ... well, there are, but it's a matter of getting messages posted in the | |
68 // right order, not fiddling with locks. | |
69 // Buffer creation is a two-step process that is serviced partially on the | |
70 // Child thread, because we need to wait for the client to provide textures | |
71 // for the buffers we allocate. We cannot keep the decoder thread running while | |
72 // the client allocates Pictures for us, because we need to REQBUFS first to get | |
73 // the required number of output buffers from the device and that cannot be done | |
74 // unless we free the previous set of buffers, leaving the decoding in a | |
75 // inoperable state for the duration of the wait for Pictures. So to prevent | |
76 // subtle races (esp. if we get Reset() in the meantime), we block the decoder | |
77 // thread while we wait for AssignPictureBuffers from the client. | |
78 // | |
79 // V4L2VideoDecodeAccelerator may use image processor to convert the output. | |
80 // There are three cases: | |
81 // Flush: V4L2VDA should wait until image processor returns all processed | |
82 // frames. | |
83 // Reset: V4L2VDA doesn't need to wait for image processor. When image processor | |
84 // returns an old frame, drop it. | |
85 // Resolution change: V4L2VDA destroy image processor when destroying output | |
86 // buffrers. We cannot drop any frame during resolution change. So V4L2VDA | |
87 // should destroy output buffers after image processor returns all the frames. | |
88 class CONTENT_EXPORT V4L2VideoDecodeAccelerator | |
89 : public media::VideoDecodeAccelerator { | |
90 public: | |
91 V4L2VideoDecodeAccelerator( | |
92 EGLDisplay egl_display, | |
93 const GetGLContextCallback& get_gl_context_cb, | |
94 const MakeGLContextCurrentCallback& make_context_current_cb, | |
95 const scoped_refptr<V4L2Device>& device); | |
96 ~V4L2VideoDecodeAccelerator() override; | |
97 | |
98 // media::VideoDecodeAccelerator implementation. | |
99 // Note: Initialize() and Destroy() are synchronous. | |
100 bool Initialize(const Config& config, Client* client) override; | |
101 void Decode(const media::BitstreamBuffer& bitstream_buffer) override; | |
102 void AssignPictureBuffers( | |
103 const std::vector<media::PictureBuffer>& buffers) override; | |
104 void ReusePictureBuffer(int32_t picture_buffer_id) override; | |
105 void Flush() override; | |
106 void Reset() override; | |
107 void Destroy() override; | |
108 bool TryToSetupDecodeOnSeparateThread( | |
109 const base::WeakPtr<Client>& decode_client, | |
110 const scoped_refptr<base::SingleThreadTaskRunner>& decode_task_runner) | |
111 override; | |
112 media::VideoPixelFormat GetOutputFormat() const override; | |
113 | |
114 static media::VideoDecodeAccelerator::SupportedProfiles | |
115 GetSupportedProfiles(); | |
116 | |
117 private: | |
118 // These are rather subjectively tuned. | |
119 enum { | |
120 kInputBufferCount = 8, | |
121 // TODO(posciak): determine input buffer size based on level limits. | |
122 // See http://crbug.com/255116. | |
123 // Input bitstream buffer size for up to 1080p streams. | |
124 kInputBufferMaxSizeFor1080p = 1024 * 1024, | |
125 // Input bitstream buffer size for up to 4k streams. | |
126 kInputBufferMaxSizeFor4k = 4 * kInputBufferMaxSizeFor1080p, | |
127 // Number of output buffers to use for each VDA stage above what's required | |
128 // by the decoder (e.g. DPB size, in H264). We need | |
129 // media::limits::kMaxVideoFrames to fill up the GpuVideoDecode pipeline, | |
130 // and +1 for a frame in transit. | |
131 kDpbOutputBufferExtraCount = media::limits::kMaxVideoFrames + 1, | |
132 }; | |
133 | |
134 // Internal state of the decoder. | |
135 enum State { | |
136 kUninitialized, // Initialize() not yet called. | |
137 kInitialized, // Initialize() returned true; ready to start decoding. | |
138 kDecoding, // DecodeBufferInitial() successful; decoding frames. | |
139 kResetting, // Presently resetting. | |
140 kAfterReset, // After Reset(), ready to start decoding again. | |
141 kChangingResolution, // Performing resolution change, all remaining | |
142 // pre-change frames decoded and processed. | |
143 kError, // Error in kDecoding state. | |
144 }; | |
145 | |
146 enum OutputRecordState { | |
147 kFree, // Ready to be queued to the device. | |
148 kAtDevice, // Held by device. | |
149 kAtProcessor, // Held by image processor. | |
150 kAtClient, // Held by client of V4L2VideoDecodeAccelerator. | |
151 }; | |
152 | |
153 enum BufferId { | |
154 kFlushBufferId = -2 // Buffer id for flush buffer, queued by FlushTask(). | |
155 }; | |
156 | |
157 // Auto-destruction reference for BitstreamBuffer, for message-passing from | |
158 // Decode() to DecodeTask(). | |
159 struct BitstreamBufferRef; | |
160 | |
161 // Auto-destruction reference for EGLSync (for message-passing). | |
162 struct EGLSyncKHRRef; | |
163 | |
164 // Record for decoded pictures that can be sent to PictureReady. | |
165 struct PictureRecord; | |
166 | |
167 // Record for input buffers. | |
168 struct InputRecord { | |
169 InputRecord(); | |
170 ~InputRecord(); | |
171 bool at_device; // held by device. | |
172 void* address; // mmap() address. | |
173 size_t length; // mmap() length. | |
174 off_t bytes_used; // bytes filled in the mmap() segment. | |
175 int32_t input_id; // triggering input_id as given to Decode(). | |
176 }; | |
177 | |
178 // Record for output buffers. | |
179 struct OutputRecord { | |
180 OutputRecord(); | |
181 OutputRecord(OutputRecord&&) = default; | |
182 ~OutputRecord(); | |
183 OutputRecordState state; | |
184 EGLImageKHR egl_image; // EGLImageKHR for the output buffer. | |
185 EGLSyncKHR egl_sync; // sync the compositor's use of the EGLImage. | |
186 int32_t picture_id; // picture buffer id as returned to PictureReady(). | |
187 bool cleared; // Whether the texture is cleared and safe to render | |
188 // from. See TextureManager for details. | |
189 // Exported fds for image processor to import. | |
190 std::vector<base::ScopedFD> fds; | |
191 }; | |
192 | |
193 // | |
194 // Decoding tasks, to be run on decode_thread_. | |
195 // | |
196 | |
197 // Enqueue a BitstreamBuffer to decode. This will enqueue a buffer to the | |
198 // decoder_input_queue_, then queue a DecodeBufferTask() to actually decode | |
199 // the buffer. | |
200 void DecodeTask(const media::BitstreamBuffer& bitstream_buffer); | |
201 | |
202 // Decode from the buffers queued in decoder_input_queue_. Calls | |
203 // DecodeBufferInitial() or DecodeBufferContinue() as appropriate. | |
204 void DecodeBufferTask(); | |
205 // Advance to the next fragment that begins a frame. | |
206 bool AdvanceFrameFragment(const uint8_t* data, size_t size, size_t* endpos); | |
207 // Schedule another DecodeBufferTask() if we're behind. | |
208 void ScheduleDecodeBufferTaskIfNeeded(); | |
209 | |
210 // Return true if we should continue to schedule DecodeBufferTask()s after | |
211 // completion. Store the amount of input actually consumed in |endpos|. | |
212 bool DecodeBufferInitial(const void* data, size_t size, size_t* endpos); | |
213 bool DecodeBufferContinue(const void* data, size_t size); | |
214 | |
215 // Accumulate data for the next frame to decode. May return false in | |
216 // non-error conditions; for example when pipeline is full and should be | |
217 // retried later. | |
218 bool AppendToInputFrame(const void* data, size_t size); | |
219 // Flush data for one decoded frame. | |
220 bool FlushInputFrame(); | |
221 | |
222 // Service I/O on the V4L2 devices. This task should only be scheduled from | |
223 // DevicePollTask(). If |event_pending| is true, one or more events | |
224 // on file descriptor are pending. | |
225 void ServiceDeviceTask(bool event_pending); | |
226 // Handle the various device queues. | |
227 void Enqueue(); | |
228 void Dequeue(); | |
229 | |
230 // Return true if there is a resolution change event pending. | |
231 bool DequeueResolutionChangeEvent(); | |
232 | |
233 // Enqueue a buffer on the corresponding queue. | |
234 bool EnqueueInputRecord(); | |
235 bool EnqueueOutputRecord(); | |
236 | |
237 // Process a ReusePictureBuffer() API call. The API call create an EGLSync | |
238 // object on the main (GPU process) thread; we will record this object so we | |
239 // can wait on it before reusing the buffer. | |
240 void ReusePictureBufferTask(int32_t picture_buffer_id, | |
241 std::unique_ptr<EGLSyncKHRRef> egl_sync_ref); | |
242 | |
243 // Flush() task. Child thread should not submit any more buffers until it | |
244 // receives the NotifyFlushDone callback. This task will schedule an empty | |
245 // BitstreamBufferRef (with input_id == kFlushBufferId) to perform the flush. | |
246 void FlushTask(); | |
247 // Notify the client of a flush completion, if required. This should be | |
248 // called any time a relevant queue could potentially be emptied: see | |
249 // function definition. | |
250 void NotifyFlushDoneIfNeeded(); | |
251 | |
252 // Reset() task. This task will schedule a ResetDoneTask() that will send | |
253 // the NotifyResetDone callback, then set the decoder state to kResetting so | |
254 // that all intervening tasks will drain. | |
255 void ResetTask(); | |
256 // ResetDoneTask() will set the decoder state back to kAfterReset, so | |
257 // subsequent decoding can continue. | |
258 void ResetDoneTask(); | |
259 | |
260 // Device destruction task. | |
261 void DestroyTask(); | |
262 | |
263 // Start |device_poll_thread_|. | |
264 bool StartDevicePoll(); | |
265 | |
266 // Stop |device_poll_thread_|. | |
267 bool StopDevicePoll(); | |
268 | |
269 bool StopInputStream(); | |
270 bool StopOutputStream(); | |
271 | |
272 void StartResolutionChange(); | |
273 void FinishResolutionChange(); | |
274 | |
275 // Try to get output format and visible size, detected after parsing the | |
276 // beginning of the stream. Sets |again| to true if more parsing is needed. | |
277 // |visible_size| could be nullptr and ignored. | |
278 bool GetFormatInfo(struct v4l2_format* format, | |
279 gfx::Size* visible_size, | |
280 bool* again); | |
281 // Create output buffers for the given |format| and |visible_size|. | |
282 bool CreateBuffersForFormat(const struct v4l2_format& format, | |
283 const gfx::Size& visible_size); | |
284 | |
285 // Try to get |visible_size|. Return visible size, or, if querying it is not | |
286 // supported or produces invalid size, return |coded_size| instead. | |
287 gfx::Size GetVisibleSize(const gfx::Size& coded_size); | |
288 | |
289 // | |
290 // Device tasks, to be run on device_poll_thread_. | |
291 // | |
292 | |
293 // The device task. | |
294 void DevicePollTask(bool poll_device); | |
295 | |
296 // | |
297 // Safe from any thread. | |
298 // | |
299 | |
300 // Error notification (using PostTask() to child thread, if necessary). | |
301 void NotifyError(Error error); | |
302 | |
303 // Set the decoder_state_ to kError and notify the client (if necessary). | |
304 void SetErrorState(Error error); | |
305 | |
306 // | |
307 // Other utility functions. Called on decoder_thread_, unless | |
308 // decoder_thread_ is not yet started, in which case the child thread can call | |
309 // these (e.g. in Initialize() or Destroy()). | |
310 // | |
311 | |
312 // Create the buffers we need. | |
313 bool CreateInputBuffers(); | |
314 bool CreateOutputBuffers(); | |
315 | |
316 // Set input and output formats before starting decode. | |
317 bool SetupFormats(); | |
318 // Return a usable input format of image processor. Return 0 if not found. | |
319 uint32_t FindImageProcessorInputFormat(); | |
320 // Return a usable output format of image processor. Return 0 if not found. | |
321 uint32_t FindImageProcessorOutputFormat(); | |
322 | |
323 // | |
324 // Methods run on child thread. | |
325 // | |
326 | |
327 // Destroy buffers. | |
328 void DestroyInputBuffers(); | |
329 // In contrast to DestroyInputBuffers, which is called only from destructor, | |
330 // we call DestroyOutputBuffers also during playback, on resolution change. | |
331 // Even if anything fails along the way, we still want to go on and clean | |
332 // up as much as possible, so return false if this happens, so that the | |
333 // caller can error out on resolution change. | |
334 bool DestroyOutputBuffers(); | |
335 void ResolutionChangeDestroyBuffers(); | |
336 | |
337 // Send decoded pictures to PictureReady. | |
338 void SendPictureReady(); | |
339 | |
340 // Callback that indicates a picture has been cleared. | |
341 void PictureCleared(); | |
342 | |
343 // Image processor returns a processed frame. Its id is |bitstream_buffer_id| | |
344 // and stored in |output_buffer_index| buffer of image processor. | |
345 void FrameProcessed(int32_t bitstream_buffer_id, int output_buffer_index); | |
346 | |
347 // Image processor notifies an error. | |
348 void ImageProcessorError(); | |
349 | |
350 // Our original calling task runner for the child thread. | |
351 scoped_refptr<base::SingleThreadTaskRunner> child_task_runner_; | |
352 | |
353 // Task runner Decode() and PictureReady() run on. | |
354 scoped_refptr<base::SingleThreadTaskRunner> decode_task_runner_; | |
355 | |
356 // WeakPtr<> pointing to |this| for use in posting tasks from the decoder or | |
357 // device worker threads back to the child thread. Because the worker threads | |
358 // are members of this class, any task running on those threads is guaranteed | |
359 // that this object is still alive. As a result, tasks posted from the child | |
360 // thread to the decoder or device thread should use base::Unretained(this), | |
361 // and tasks posted the other way should use |weak_this_|. | |
362 base::WeakPtr<V4L2VideoDecodeAccelerator> weak_this_; | |
363 | |
364 // To expose client callbacks from VideoDecodeAccelerator. | |
365 // NOTE: all calls to these objects *MUST* be executed on | |
366 // child_task_runner_. | |
367 std::unique_ptr<base::WeakPtrFactory<Client>> client_ptr_factory_; | |
368 base::WeakPtr<Client> client_; | |
369 // Callbacks to |decode_client_| must be executed on |decode_task_runner_|. | |
370 base::WeakPtr<Client> decode_client_; | |
371 | |
372 // | |
373 // Decoder state, owned and operated by decoder_thread_. | |
374 // Before decoder_thread_ has started, the decoder state is managed by | |
375 // the child (main) thread. After decoder_thread_ has started, the decoder | |
376 // thread should be the only one managing these. | |
377 // | |
378 | |
379 // This thread services tasks posted from the VDA API entry points by the | |
380 // child thread and device service callbacks posted from the device thread. | |
381 base::Thread decoder_thread_; | |
382 // Decoder state machine state. | |
383 State decoder_state_; | |
384 // BitstreamBuffer we're presently reading. | |
385 std::unique_ptr<BitstreamBufferRef> decoder_current_bitstream_buffer_; | |
386 // The V4L2Device this class is operating upon. | |
387 scoped_refptr<V4L2Device> device_; | |
388 // FlushTask() and ResetTask() should not affect buffers that have been | |
389 // queued afterwards. For flushing or resetting the pipeline then, we will | |
390 // delay these buffers until after the flush or reset completes. | |
391 int decoder_delay_bitstream_buffer_id_; | |
392 // Input buffer we're presently filling. | |
393 int decoder_current_input_buffer_; | |
394 // We track the number of buffer decode tasks we have scheduled, since each | |
395 // task execution should complete one buffer. If we fall behind (due to | |
396 // resource backpressure, etc.), we'll have to schedule more to catch up. | |
397 int decoder_decode_buffer_tasks_scheduled_; | |
398 // Picture buffers held by the client. | |
399 int decoder_frames_at_client_; | |
400 // Are we flushing? | |
401 bool decoder_flushing_; | |
402 // Got a reset request while we were performing resolution change. | |
403 bool resolution_change_reset_pending_; | |
404 // Input queue for decoder_thread_: BitstreamBuffers in. | |
405 std::queue<linked_ptr<BitstreamBufferRef> > decoder_input_queue_; | |
406 // For H264 decode, hardware requires that we send it frame-sized chunks. | |
407 // We'll need to parse the stream. | |
408 std::unique_ptr<media::H264Parser> decoder_h264_parser_; | |
409 // Set if the decoder has a pending incomplete frame in an input buffer. | |
410 bool decoder_partial_frame_pending_; | |
411 | |
412 // | |
413 // Hardware state and associated queues. Since decoder_thread_ services | |
414 // the hardware, decoder_thread_ owns these too. | |
415 // output_buffer_map_, free_output_buffers_ and output_planes_count_ are an | |
416 // exception during the buffer (re)allocation sequence, when the | |
417 // decoder_thread_ is blocked briefly while the Child thread manipulates | |
418 // them. | |
419 // | |
420 | |
421 // Completed decode buffers. | |
422 std::queue<int> input_ready_queue_; | |
423 | |
424 // Input buffer state. | |
425 bool input_streamon_; | |
426 // Input buffers enqueued to device. | |
427 int input_buffer_queued_count_; | |
428 // Input buffers ready to use, as a LIFO since we don't care about ordering. | |
429 std::vector<int> free_input_buffers_; | |
430 // Mapping of int index to input buffer record. | |
431 std::vector<InputRecord> input_buffer_map_; | |
432 | |
433 // Output buffer state. | |
434 bool output_streamon_; | |
435 // Output buffers enqueued to device. | |
436 int output_buffer_queued_count_; | |
437 // Output buffers ready to use, as a FIFO since we want oldest-first to hide | |
438 // synchronization latency with GL. | |
439 std::queue<int> free_output_buffers_; | |
440 // Mapping of int index to output buffer record. | |
441 std::vector<OutputRecord> output_buffer_map_; | |
442 // Required size of DPB for decoding. | |
443 int output_dpb_size_; | |
444 | |
445 // Number of planes (i.e. separate memory buffers) for output. | |
446 size_t output_planes_count_; | |
447 | |
448 // Pictures that are ready but not sent to PictureReady yet. | |
449 std::queue<PictureRecord> pending_picture_ready_; | |
450 | |
451 // The number of pictures that are sent to PictureReady and will be cleared. | |
452 int picture_clearing_count_; | |
453 | |
454 // Used by the decoder thread to wait for AssignPictureBuffers to arrive | |
455 // to avoid races with potential Reset requests. | |
456 base::WaitableEvent pictures_assigned_; | |
457 | |
458 // Output picture coded size. | |
459 gfx::Size coded_size_; | |
460 | |
461 // Output picture visible size. | |
462 gfx::Size visible_size_; | |
463 | |
464 // | |
465 // The device polling thread handles notifications of V4L2 device changes. | |
466 // | |
467 | |
468 // The thread. | |
469 base::Thread device_poll_thread_; | |
470 | |
471 // | |
472 // Other state, held by the child (main) thread. | |
473 // | |
474 | |
475 // EGL state | |
476 EGLDisplay egl_display_; | |
477 | |
478 // Callback to get current GLContext. | |
479 GetGLContextCallback get_gl_context_cb_; | |
480 // Callback to set the correct gl context. | |
481 MakeGLContextCurrentCallback make_context_current_cb_; | |
482 | |
483 // The codec we'll be decoding for. | |
484 media::VideoCodecProfile video_profile_; | |
485 // Chosen output format. | |
486 uint32_t output_format_fourcc_; | |
487 | |
488 // Image processor device, if one is in use. | |
489 scoped_refptr<V4L2Device> image_processor_device_; | |
490 // Image processor. Created and destroyed on child thread. | |
491 std::unique_ptr<V4L2ImageProcessor> image_processor_; | |
492 | |
493 // The V4L2Device EGLImage is created from. | |
494 scoped_refptr<V4L2Device> egl_image_device_; | |
495 // The format of EGLImage. | |
496 uint32_t egl_image_format_fourcc_; | |
497 // The logical dimensions of EGLImage buffer in pixels. | |
498 gfx::Size egl_image_size_; | |
499 // Number of planes for EGLImage. | |
500 size_t egl_image_planes_count_; | |
501 | |
502 // IDs of bitstream buffers sent to image processor to process. After a | |
503 // buffer is processed, it will sent to render if the id is in this | |
504 // queue. If the id is not in this queue, the buffer will be dropped. | |
505 std::queue<int> image_processor_bitstream_buffer_ids_; | |
506 | |
507 // Input format V4L2 fourccs this class supports. | |
508 static const uint32_t supported_input_fourccs_[]; | |
509 | |
510 // The WeakPtrFactory for |weak_this_|. | |
511 base::WeakPtrFactory<V4L2VideoDecodeAccelerator> weak_this_factory_; | |
512 | |
513 DISALLOW_COPY_AND_ASSIGN(V4L2VideoDecodeAccelerator); | |
514 }; | |
515 | |
516 } // namespace content | |
517 | |
518 #endif // CONTENT_COMMON_GPU_MEDIA_V4L2_VIDEO_DECODE_ACCELERATOR_H_ | |
OLD | NEW |