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.
4
5 #include <algorithm>
6 #include <limits>
7
8 #include "base/bind.h"
9 #include "base/bind_helpers.h"
10 #include "base/numerics/safe_conversions.h"
11 #include "base/stl_util.h"
12 #include "content/common/gpu/media/vaapi_h264_decoder.h"
13
14 namespace content {
15
16 // Decode surface, used for decoding and reference. input_id comes from client
17 // and is associated with the surface that was produced as the result
18 // of decoding a bitstream buffer with that id.
19 class VaapiH264Decoder::DecodeSurface {
20 public:
21 DecodeSurface(int poc,
22 int32 input_id,
23 const scoped_refptr<VASurface>& va_surface);
24 DecodeSurface(int poc, const scoped_refptr<DecodeSurface>& dec_surface);
25 ~DecodeSurface();
26
poc()27 int poc() {
28 return poc_;
29 }
30
va_surface()31 scoped_refptr<VASurface> va_surface() {
32 return va_surface_;
33 }
34
input_id()35 int32 input_id() {
36 return input_id_;
37 }
38
39 private:
40 int poc_;
41 int32 input_id_;
42 scoped_refptr<VASurface> va_surface_;
43 };
44
DecodeSurface(int poc,int32 input_id,const scoped_refptr<VASurface> & va_surface)45 VaapiH264Decoder::DecodeSurface::DecodeSurface(
46 int poc,
47 int32 input_id,
48 const scoped_refptr<VASurface>& va_surface)
49 : poc_(poc),
50 input_id_(input_id),
51 va_surface_(va_surface) {
52 DCHECK(va_surface_.get());
53 }
54
~DecodeSurface()55 VaapiH264Decoder::DecodeSurface::~DecodeSurface() {
56 }
57
VaapiH264Decoder(VaapiWrapper * vaapi_wrapper,const OutputPicCB & output_pic_cb,const ReportErrorToUmaCB & report_error_to_uma_cb)58 VaapiH264Decoder::VaapiH264Decoder(
59 VaapiWrapper* vaapi_wrapper,
60 const OutputPicCB& output_pic_cb,
61 const ReportErrorToUmaCB& report_error_to_uma_cb)
62 : max_pic_order_cnt_lsb_(0),
63 max_frame_num_(0),
64 max_pic_num_(0),
65 max_long_term_frame_idx_(0),
66 max_num_reorder_frames_(0),
67 curr_sps_id_(-1),
68 curr_pps_id_(-1),
69 vaapi_wrapper_(vaapi_wrapper),
70 output_pic_cb_(output_pic_cb),
71 report_error_to_uma_cb_(report_error_to_uma_cb) {
72 Reset();
73 state_ = kNeedStreamMetadata;
74 }
75
~VaapiH264Decoder()76 VaapiH264Decoder::~VaapiH264Decoder() {
77 }
78
Reset()79 void VaapiH264Decoder::Reset() {
80 curr_pic_.reset();
81
82 curr_input_id_ = -1;
83 frame_num_ = 0;
84 prev_frame_num_ = -1;
85 prev_frame_num_offset_ = -1;
86
87 prev_ref_has_memmgmnt5_ = false;
88 prev_ref_top_field_order_cnt_ = -1;
89 prev_ref_pic_order_cnt_msb_ = -1;
90 prev_ref_pic_order_cnt_lsb_ = -1;
91 prev_ref_field_ = H264Picture::FIELD_NONE;
92
93 vaapi_wrapper_->DestroyPendingBuffers();
94
95 ref_pic_list0_.clear();
96 ref_pic_list1_.clear();
97
98 for (DecSurfacesInUse::iterator it = decode_surfaces_in_use_.begin();
99 it != decode_surfaces_in_use_.end(); ) {
100 int poc = it->second->poc();
101 // Must be incremented before UnassignSurfaceFromPoC as this call
102 // invalidates |it|.
103 ++it;
104 UnassignSurfaceFromPoC(poc);
105 }
106 DCHECK(decode_surfaces_in_use_.empty());
107
108 dpb_.Clear();
109 parser_.Reset();
110 last_output_poc_ = std::numeric_limits<int>::min();
111
112 // If we are in kDecoding, we can resume without processing an SPS.
113 if (state_ == kDecoding)
114 state_ = kAfterReset;
115 }
116
ReuseSurface(const scoped_refptr<VASurface> & va_surface)117 void VaapiH264Decoder::ReuseSurface(
118 const scoped_refptr<VASurface>& va_surface) {
119 available_va_surfaces_.push_back(va_surface);
120 }
121
122 // Fill |va_pic| with default/neutral values.
InitVAPicture(VAPictureH264 * va_pic)123 static void InitVAPicture(VAPictureH264* va_pic) {
124 memset(va_pic, 0, sizeof(*va_pic));
125 va_pic->picture_id = VA_INVALID_ID;
126 va_pic->flags = VA_PICTURE_H264_INVALID;
127 }
128
FillVAPicture(VAPictureH264 * va_pic,H264Picture * pic)129 void VaapiH264Decoder::FillVAPicture(VAPictureH264 *va_pic, H264Picture* pic) {
130 DCHECK(pic);
131
132 DecodeSurface* dec_surface = DecodeSurfaceByPoC(pic->pic_order_cnt);
133 if (!dec_surface) {
134 // Cannot provide a ref picture, will corrupt output, but may be able
135 // to recover.
136 InitVAPicture(va_pic);
137 return;
138 }
139
140 va_pic->picture_id = dec_surface->va_surface()->id();
141 va_pic->frame_idx = pic->frame_num;
142 va_pic->flags = 0;
143
144 switch (pic->field) {
145 case H264Picture::FIELD_NONE:
146 break;
147 case H264Picture::FIELD_TOP:
148 va_pic->flags |= VA_PICTURE_H264_TOP_FIELD;
149 break;
150 case H264Picture::FIELD_BOTTOM:
151 va_pic->flags |= VA_PICTURE_H264_BOTTOM_FIELD;
152 break;
153 }
154
155 if (pic->ref) {
156 va_pic->flags |= pic->long_term ? VA_PICTURE_H264_LONG_TERM_REFERENCE
157 : VA_PICTURE_H264_SHORT_TERM_REFERENCE;
158 }
159
160 va_pic->TopFieldOrderCnt = pic->top_field_order_cnt;
161 va_pic->BottomFieldOrderCnt = pic->bottom_field_order_cnt;
162 }
163
FillVARefFramesFromDPB(VAPictureH264 * va_pics,int num_pics)164 int VaapiH264Decoder::FillVARefFramesFromDPB(VAPictureH264 *va_pics,
165 int num_pics) {
166 H264DPB::Pictures::reverse_iterator rit;
167 int i;
168
169 // Return reference frames in reverse order of insertion.
170 // Libva does not document this, but other implementations (e.g. mplayer)
171 // do it this way as well.
172 for (rit = dpb_.rbegin(), i = 0; rit != dpb_.rend() && i < num_pics; ++rit) {
173 if ((*rit)->ref)
174 FillVAPicture(&va_pics[i++], *rit);
175 }
176
177 return i;
178 }
179
DecodeSurfaceByPoC(int poc)180 VaapiH264Decoder::DecodeSurface* VaapiH264Decoder::DecodeSurfaceByPoC(int poc) {
181 DecSurfacesInUse::iterator iter = decode_surfaces_in_use_.find(poc);
182 if (iter == decode_surfaces_in_use_.end()) {
183 DVLOG(1) << "Could not find surface assigned to POC: " << poc;
184 return NULL;
185 }
186
187 return iter->second.get();
188 }
189
AssignSurfaceToPoC(int32 input_id,int poc)190 bool VaapiH264Decoder::AssignSurfaceToPoC(int32 input_id, int poc) {
191 if (available_va_surfaces_.empty()) {
192 DVLOG(1) << "No VA Surfaces available";
193 return false;
194 }
195
196 linked_ptr<DecodeSurface> dec_surface(new DecodeSurface(
197 poc, input_id, available_va_surfaces_.back()));
198 available_va_surfaces_.pop_back();
199
200 DVLOG(4) << "POC " << poc
201 << " will use surface " << dec_surface->va_surface()->id();
202
203 bool inserted = decode_surfaces_in_use_.insert(
204 std::make_pair(poc, dec_surface)).second;
205 DCHECK(inserted);
206
207 return true;
208 }
209
UnassignSurfaceFromPoC(int poc)210 void VaapiH264Decoder::UnassignSurfaceFromPoC(int poc) {
211 DecSurfacesInUse::iterator it = decode_surfaces_in_use_.find(poc);
212 if (it == decode_surfaces_in_use_.end()) {
213 DVLOG(1) << "Asked to unassign an unassigned POC " << poc;
214 return;
215 }
216
217 DVLOG(4) << "POC " << poc << " no longer using VA surface "
218 << it->second->va_surface()->id();
219
220 decode_surfaces_in_use_.erase(it);
221 }
222
SendPPS()223 bool VaapiH264Decoder::SendPPS() {
224 const media::H264PPS* pps = parser_.GetPPS(curr_pps_id_);
225 DCHECK(pps);
226
227 const media::H264SPS* sps = parser_.GetSPS(pps->seq_parameter_set_id);
228 DCHECK(sps);
229
230 DCHECK(curr_pic_.get());
231
232 VAPictureParameterBufferH264 pic_param;
233 memset(&pic_param, 0, sizeof(VAPictureParameterBufferH264));
234
235 #define FROM_SPS_TO_PP(a) pic_param.a = sps->a;
236 #define FROM_SPS_TO_PP2(a, b) pic_param.b = sps->a;
237 FROM_SPS_TO_PP2(pic_width_in_mbs_minus1, picture_width_in_mbs_minus1);
238 // This assumes non-interlaced video
239 FROM_SPS_TO_PP2(pic_height_in_map_units_minus1,
240 picture_height_in_mbs_minus1);
241 FROM_SPS_TO_PP(bit_depth_luma_minus8);
242 FROM_SPS_TO_PP(bit_depth_chroma_minus8);
243 #undef FROM_SPS_TO_PP
244 #undef FROM_SPS_TO_PP2
245
246 #define FROM_SPS_TO_PP_SF(a) pic_param.seq_fields.bits.a = sps->a;
247 #define FROM_SPS_TO_PP_SF2(a, b) pic_param.seq_fields.bits.b = sps->a;
248 FROM_SPS_TO_PP_SF(chroma_format_idc);
249 FROM_SPS_TO_PP_SF2(separate_colour_plane_flag,
250 residual_colour_transform_flag);
251 FROM_SPS_TO_PP_SF(gaps_in_frame_num_value_allowed_flag);
252 FROM_SPS_TO_PP_SF(frame_mbs_only_flag);
253 FROM_SPS_TO_PP_SF(mb_adaptive_frame_field_flag);
254 FROM_SPS_TO_PP_SF(direct_8x8_inference_flag);
255 pic_param.seq_fields.bits.MinLumaBiPredSize8x8 = (sps->level_idc >= 31);
256 FROM_SPS_TO_PP_SF(log2_max_frame_num_minus4);
257 FROM_SPS_TO_PP_SF(pic_order_cnt_type);
258 FROM_SPS_TO_PP_SF(log2_max_pic_order_cnt_lsb_minus4);
259 FROM_SPS_TO_PP_SF(delta_pic_order_always_zero_flag);
260 #undef FROM_SPS_TO_PP_SF
261 #undef FROM_SPS_TO_PP_SF2
262
263 #define FROM_PPS_TO_PP(a) pic_param.a = pps->a;
264 FROM_PPS_TO_PP(num_slice_groups_minus1);
265 pic_param.slice_group_map_type = 0;
266 pic_param.slice_group_change_rate_minus1 = 0;
267 FROM_PPS_TO_PP(pic_init_qp_minus26);
268 FROM_PPS_TO_PP(pic_init_qs_minus26);
269 FROM_PPS_TO_PP(chroma_qp_index_offset);
270 FROM_PPS_TO_PP(second_chroma_qp_index_offset);
271 #undef FROM_PPS_TO_PP
272
273 #define FROM_PPS_TO_PP_PF(a) pic_param.pic_fields.bits.a = pps->a;
274 #define FROM_PPS_TO_PP_PF2(a, b) pic_param.pic_fields.bits.b = pps->a;
275 FROM_PPS_TO_PP_PF(entropy_coding_mode_flag);
276 FROM_PPS_TO_PP_PF(weighted_pred_flag);
277 FROM_PPS_TO_PP_PF(weighted_bipred_idc);
278 FROM_PPS_TO_PP_PF(transform_8x8_mode_flag);
279
280 pic_param.pic_fields.bits.field_pic_flag = 0;
281 FROM_PPS_TO_PP_PF(constrained_intra_pred_flag);
282 FROM_PPS_TO_PP_PF2(bottom_field_pic_order_in_frame_present_flag,
283 pic_order_present_flag);
284 FROM_PPS_TO_PP_PF(deblocking_filter_control_present_flag);
285 FROM_PPS_TO_PP_PF(redundant_pic_cnt_present_flag);
286 pic_param.pic_fields.bits.reference_pic_flag = curr_pic_->ref;
287 #undef FROM_PPS_TO_PP_PF
288 #undef FROM_PPS_TO_PP_PF2
289
290 pic_param.frame_num = curr_pic_->frame_num;
291
292 InitVAPicture(&pic_param.CurrPic);
293 FillVAPicture(&pic_param.CurrPic, curr_pic_.get());
294
295 // Init reference pictures' array.
296 for (int i = 0; i < 16; ++i)
297 InitVAPicture(&pic_param.ReferenceFrames[i]);
298
299 // And fill it with picture info from DPB.
300 FillVARefFramesFromDPB(pic_param.ReferenceFrames,
301 arraysize(pic_param.ReferenceFrames));
302
303 pic_param.num_ref_frames = sps->max_num_ref_frames;
304
305 return vaapi_wrapper_->SubmitBuffer(VAPictureParameterBufferType,
306 sizeof(VAPictureParameterBufferH264),
307 &pic_param);
308 }
309
SendIQMatrix()310 bool VaapiH264Decoder::SendIQMatrix() {
311 const media::H264PPS* pps = parser_.GetPPS(curr_pps_id_);
312 DCHECK(pps);
313
314 VAIQMatrixBufferH264 iq_matrix_buf;
315 memset(&iq_matrix_buf, 0, sizeof(VAIQMatrixBufferH264));
316
317 if (pps->pic_scaling_matrix_present_flag) {
318 for (int i = 0; i < 6; ++i) {
319 for (int j = 0; j < 16; ++j)
320 iq_matrix_buf.ScalingList4x4[i][j] = pps->scaling_list4x4[i][j];
321 }
322
323 for (int i = 0; i < 2; ++i) {
324 for (int j = 0; j < 64; ++j)
325 iq_matrix_buf.ScalingList8x8[i][j] = pps->scaling_list8x8[i][j];
326 }
327 } else {
328 const media::H264SPS* sps = parser_.GetSPS(pps->seq_parameter_set_id);
329 DCHECK(sps);
330 for (int i = 0; i < 6; ++i) {
331 for (int j = 0; j < 16; ++j)
332 iq_matrix_buf.ScalingList4x4[i][j] = sps->scaling_list4x4[i][j];
333 }
334
335 for (int i = 0; i < 2; ++i) {
336 for (int j = 0; j < 64; ++j)
337 iq_matrix_buf.ScalingList8x8[i][j] = sps->scaling_list8x8[i][j];
338 }
339 }
340
341 return vaapi_wrapper_->SubmitBuffer(VAIQMatrixBufferType,
342 sizeof(VAIQMatrixBufferH264),
343 &iq_matrix_buf);
344 }
345
SendVASliceParam(media::H264SliceHeader * slice_hdr)346 bool VaapiH264Decoder::SendVASliceParam(media::H264SliceHeader* slice_hdr) {
347 const media::H264PPS* pps = parser_.GetPPS(slice_hdr->pic_parameter_set_id);
348 DCHECK(pps);
349
350 const media::H264SPS* sps = parser_.GetSPS(pps->seq_parameter_set_id);
351 DCHECK(sps);
352
353 VASliceParameterBufferH264 slice_param;
354 memset(&slice_param, 0, sizeof(VASliceParameterBufferH264));
355
356 slice_param.slice_data_size = slice_hdr->nalu_size;
357 slice_param.slice_data_offset = 0;
358 slice_param.slice_data_flag = VA_SLICE_DATA_FLAG_ALL;
359 slice_param.slice_data_bit_offset = slice_hdr->header_bit_size;
360
361 #define SHDRToSP(a) slice_param.a = slice_hdr->a;
362 SHDRToSP(first_mb_in_slice);
363 slice_param.slice_type = slice_hdr->slice_type % 5;
364 SHDRToSP(direct_spatial_mv_pred_flag);
365
366 // TODO posciak: make sure parser sets those even when override flags
367 // in slice header is off.
368 SHDRToSP(num_ref_idx_l0_active_minus1);
369 SHDRToSP(num_ref_idx_l1_active_minus1);
370 SHDRToSP(cabac_init_idc);
371 SHDRToSP(slice_qp_delta);
372 SHDRToSP(disable_deblocking_filter_idc);
373 SHDRToSP(slice_alpha_c0_offset_div2);
374 SHDRToSP(slice_beta_offset_div2);
375
376 if (((slice_hdr->IsPSlice() || slice_hdr->IsSPSlice()) &&
377 pps->weighted_pred_flag) ||
378 (slice_hdr->IsBSlice() && pps->weighted_bipred_idc == 1)) {
379 SHDRToSP(luma_log2_weight_denom);
380 SHDRToSP(chroma_log2_weight_denom);
381
382 SHDRToSP(luma_weight_l0_flag);
383 SHDRToSP(luma_weight_l1_flag);
384
385 SHDRToSP(chroma_weight_l0_flag);
386 SHDRToSP(chroma_weight_l1_flag);
387
388 for (int i = 0; i <= slice_param.num_ref_idx_l0_active_minus1; ++i) {
389 slice_param.luma_weight_l0[i] =
390 slice_hdr->pred_weight_table_l0.luma_weight[i];
391 slice_param.luma_offset_l0[i] =
392 slice_hdr->pred_weight_table_l0.luma_offset[i];
393
394 for (int j = 0; j < 2; ++j) {
395 slice_param.chroma_weight_l0[i][j] =
396 slice_hdr->pred_weight_table_l0.chroma_weight[i][j];
397 slice_param.chroma_offset_l0[i][j] =
398 slice_hdr->pred_weight_table_l0.chroma_offset[i][j];
399 }
400 }
401
402 if (slice_hdr->IsBSlice()) {
403 for (int i = 0; i <= slice_param.num_ref_idx_l1_active_minus1; ++i) {
404 slice_param.luma_weight_l1[i] =
405 slice_hdr->pred_weight_table_l1.luma_weight[i];
406 slice_param.luma_offset_l1[i] =
407 slice_hdr->pred_weight_table_l1.luma_offset[i];
408
409 for (int j = 0; j < 2; ++j) {
410 slice_param.chroma_weight_l1[i][j] =
411 slice_hdr->pred_weight_table_l1.chroma_weight[i][j];
412 slice_param.chroma_offset_l1[i][j] =
413 slice_hdr->pred_weight_table_l1.chroma_offset[i][j];
414 }
415 }
416 }
417 }
418
419 for (int i = 0; i < 32; ++i) {
420 InitVAPicture(&slice_param.RefPicList0[i]);
421 InitVAPicture(&slice_param.RefPicList1[i]);
422 }
423
424 int i;
425 H264Picture::PtrVector::iterator it;
426 for (it = ref_pic_list0_.begin(), i = 0; it != ref_pic_list0_.end() && *it;
427 ++it, ++i)
428 FillVAPicture(&slice_param.RefPicList0[i], *it);
429 for (it = ref_pic_list1_.begin(), i = 0; it != ref_pic_list1_.end() && *it;
430 ++it, ++i)
431 FillVAPicture(&slice_param.RefPicList1[i], *it);
432
433 return vaapi_wrapper_->SubmitBuffer(VASliceParameterBufferType,
434 sizeof(VASliceParameterBufferH264),
435 &slice_param);
436 }
437
SendSliceData(const uint8 * ptr,size_t size)438 bool VaapiH264Decoder::SendSliceData(const uint8* ptr, size_t size) {
439 // Can't help it, blame libva...
440 void* non_const_ptr = const_cast<uint8*>(ptr);
441 return vaapi_wrapper_->SubmitBuffer(VASliceDataBufferType, size,
442 non_const_ptr);
443 }
444
PrepareRefPicLists(media::H264SliceHeader * slice_hdr)445 bool VaapiH264Decoder::PrepareRefPicLists(media::H264SliceHeader* slice_hdr) {
446 ref_pic_list0_.clear();
447 ref_pic_list1_.clear();
448
449 // Fill reference picture lists for B and S/SP slices.
450 if (slice_hdr->IsPSlice() || slice_hdr->IsSPSlice()) {
451 ConstructReferencePicListsP(slice_hdr);
452 return ModifyReferencePicList(slice_hdr, 0);
453 }
454
455 if (slice_hdr->IsBSlice()) {
456 ConstructReferencePicListsB(slice_hdr);
457 return ModifyReferencePicList(slice_hdr, 0) &&
458 ModifyReferencePicList(slice_hdr, 1);
459 }
460
461 return true;
462 }
463
QueueSlice(media::H264SliceHeader * slice_hdr)464 bool VaapiH264Decoder::QueueSlice(media::H264SliceHeader* slice_hdr) {
465 DCHECK(curr_pic_.get());
466
467 if (!PrepareRefPicLists(slice_hdr))
468 return false;
469
470 if (!SendVASliceParam(slice_hdr))
471 return false;
472
473 if (!SendSliceData(slice_hdr->nalu_data, slice_hdr->nalu_size))
474 return false;
475
476 return true;
477 }
478
479 // TODO(posciak) start using vaMapBuffer instead of vaCreateBuffer wherever
480 // possible.
DecodePicture()481 bool VaapiH264Decoder::DecodePicture() {
482 DCHECK(curr_pic_.get());
483
484 DVLOG(4) << "Decoding POC " << curr_pic_->pic_order_cnt;
485 DecodeSurface* dec_surface = DecodeSurfaceByPoC(curr_pic_->pic_order_cnt);
486 if (!dec_surface) {
487 DVLOG(1) << "Asked to decode an invalid POC " << curr_pic_->pic_order_cnt;
488 return false;
489 }
490
491 if (!vaapi_wrapper_->ExecuteAndDestroyPendingBuffers(
492 dec_surface->va_surface()->id())) {
493 DVLOG(1) << "Failed decoding picture";
494 return false;
495 }
496
497 return true;
498 }
499
InitCurrPicture(media::H264SliceHeader * slice_hdr)500 bool VaapiH264Decoder::InitCurrPicture(media::H264SliceHeader* slice_hdr) {
501 DCHECK(curr_pic_.get());
502
503 memset(curr_pic_.get(), 0, sizeof(H264Picture));
504
505 curr_pic_->idr = slice_hdr->idr_pic_flag;
506
507 if (slice_hdr->field_pic_flag) {
508 curr_pic_->field = slice_hdr->bottom_field_flag ? H264Picture::FIELD_BOTTOM
509 : H264Picture::FIELD_TOP;
510 } else {
511 curr_pic_->field = H264Picture::FIELD_NONE;
512 }
513
514 curr_pic_->ref = slice_hdr->nal_ref_idc != 0;
515 // This assumes non-interlaced stream.
516 curr_pic_->frame_num = curr_pic_->pic_num = slice_hdr->frame_num;
517
518 if (!CalculatePicOrderCounts(slice_hdr))
519 return false;
520
521 // Try to get an empty surface to decode this picture to.
522 if (!AssignSurfaceToPoC(curr_input_id_, curr_pic_->pic_order_cnt)) {
523 DVLOG(1) << "Failed getting a free surface for a picture";
524 return false;
525 }
526
527 curr_pic_->long_term_reference_flag = slice_hdr->long_term_reference_flag;
528 curr_pic_->adaptive_ref_pic_marking_mode_flag =
529 slice_hdr->adaptive_ref_pic_marking_mode_flag;
530
531 // If the slice header indicates we will have to perform reference marking
532 // process after this picture is decoded, store required data for that
533 // purpose.
534 if (slice_hdr->adaptive_ref_pic_marking_mode_flag) {
535 COMPILE_ASSERT(sizeof(curr_pic_->ref_pic_marking) ==
536 sizeof(slice_hdr->ref_pic_marking),
537 ref_pic_marking_array_sizes_do_not_match);
538 memcpy(curr_pic_->ref_pic_marking, slice_hdr->ref_pic_marking,
539 sizeof(curr_pic_->ref_pic_marking));
540 }
541
542 return true;
543 }
544
CalculatePicOrderCounts(media::H264SliceHeader * slice_hdr)545 bool VaapiH264Decoder::CalculatePicOrderCounts(
546 media::H264SliceHeader* slice_hdr) {
547 DCHECK_NE(curr_sps_id_, -1);
548 const media::H264SPS* sps = parser_.GetSPS(curr_sps_id_);
549
550 int pic_order_cnt_lsb = slice_hdr->pic_order_cnt_lsb;
551 curr_pic_->pic_order_cnt_lsb = pic_order_cnt_lsb;
552
553 switch (sps->pic_order_cnt_type) {
554 case 0:
555 // See spec 8.2.1.1.
556 int prev_pic_order_cnt_msb, prev_pic_order_cnt_lsb;
557 if (slice_hdr->idr_pic_flag) {
558 prev_pic_order_cnt_msb = prev_pic_order_cnt_lsb = 0;
559 } else {
560 if (prev_ref_has_memmgmnt5_) {
561 if (prev_ref_field_ != H264Picture::FIELD_BOTTOM) {
562 prev_pic_order_cnt_msb = 0;
563 prev_pic_order_cnt_lsb = prev_ref_top_field_order_cnt_;
564 } else {
565 prev_pic_order_cnt_msb = 0;
566 prev_pic_order_cnt_lsb = 0;
567 }
568 } else {
569 prev_pic_order_cnt_msb = prev_ref_pic_order_cnt_msb_;
570 prev_pic_order_cnt_lsb = prev_ref_pic_order_cnt_lsb_;
571 }
572 }
573
574 DCHECK_NE(max_pic_order_cnt_lsb_, 0);
575 if ((pic_order_cnt_lsb < prev_pic_order_cnt_lsb) &&
576 (prev_pic_order_cnt_lsb - pic_order_cnt_lsb >=
577 max_pic_order_cnt_lsb_ / 2)) {
578 curr_pic_->pic_order_cnt_msb = prev_pic_order_cnt_msb +
579 max_pic_order_cnt_lsb_;
580 } else if ((pic_order_cnt_lsb > prev_pic_order_cnt_lsb) &&
581 (pic_order_cnt_lsb - prev_pic_order_cnt_lsb >
582 max_pic_order_cnt_lsb_ / 2)) {
583 curr_pic_->pic_order_cnt_msb = prev_pic_order_cnt_msb -
584 max_pic_order_cnt_lsb_;
585 } else {
586 curr_pic_->pic_order_cnt_msb = prev_pic_order_cnt_msb;
587 }
588
589 if (curr_pic_->field != H264Picture::FIELD_BOTTOM) {
590 curr_pic_->top_field_order_cnt = curr_pic_->pic_order_cnt_msb +
591 pic_order_cnt_lsb;
592 }
593
594 if (curr_pic_->field != H264Picture::FIELD_TOP) {
595 // TODO posciak: perhaps replace with pic->field?
596 if (!slice_hdr->field_pic_flag) {
597 curr_pic_->bottom_field_order_cnt = curr_pic_->top_field_order_cnt +
598 slice_hdr->delta_pic_order_cnt_bottom;
599 } else {
600 curr_pic_->bottom_field_order_cnt = curr_pic_->pic_order_cnt_msb +
601 pic_order_cnt_lsb;
602 }
603 }
604 break;
605
606 case 1: {
607 // See spec 8.2.1.2.
608 if (prev_has_memmgmnt5_)
609 prev_frame_num_offset_ = 0;
610
611 if (slice_hdr->idr_pic_flag)
612 curr_pic_->frame_num_offset = 0;
613 else if (prev_frame_num_ > slice_hdr->frame_num)
614 curr_pic_->frame_num_offset = prev_frame_num_offset_ + max_frame_num_;
615 else
616 curr_pic_->frame_num_offset = prev_frame_num_offset_;
617
618 int abs_frame_num = 0;
619 if (sps->num_ref_frames_in_pic_order_cnt_cycle != 0)
620 abs_frame_num = curr_pic_->frame_num_offset + slice_hdr->frame_num;
621 else
622 abs_frame_num = 0;
623
624 if (slice_hdr->nal_ref_idc == 0 && abs_frame_num > 0)
625 --abs_frame_num;
626
627 int expected_pic_order_cnt = 0;
628 if (abs_frame_num > 0) {
629 if (sps->num_ref_frames_in_pic_order_cnt_cycle == 0) {
630 DVLOG(1) << "Invalid num_ref_frames_in_pic_order_cnt_cycle "
631 << "in stream";
632 return false;
633 }
634
635 int pic_order_cnt_cycle_cnt = (abs_frame_num - 1) /
636 sps->num_ref_frames_in_pic_order_cnt_cycle;
637 int frame_num_in_pic_order_cnt_cycle = (abs_frame_num - 1) %
638 sps->num_ref_frames_in_pic_order_cnt_cycle;
639
640 expected_pic_order_cnt = pic_order_cnt_cycle_cnt *
641 sps->expected_delta_per_pic_order_cnt_cycle;
642 // frame_num_in_pic_order_cnt_cycle is verified < 255 in parser
643 for (int i = 0; i <= frame_num_in_pic_order_cnt_cycle; ++i)
644 expected_pic_order_cnt += sps->offset_for_ref_frame[i];
645 }
646
647 if (!slice_hdr->nal_ref_idc)
648 expected_pic_order_cnt += sps->offset_for_non_ref_pic;
649
650 if (!slice_hdr->field_pic_flag) {
651 curr_pic_->top_field_order_cnt = expected_pic_order_cnt +
652 slice_hdr->delta_pic_order_cnt[0];
653 curr_pic_->bottom_field_order_cnt = curr_pic_->top_field_order_cnt +
654 sps->offset_for_top_to_bottom_field +
655 slice_hdr->delta_pic_order_cnt[1];
656 } else if (!slice_hdr->bottom_field_flag) {
657 curr_pic_->top_field_order_cnt = expected_pic_order_cnt +
658 slice_hdr->delta_pic_order_cnt[0];
659 } else {
660 curr_pic_->bottom_field_order_cnt = expected_pic_order_cnt +
661 sps->offset_for_top_to_bottom_field +
662 slice_hdr->delta_pic_order_cnt[0];
663 }
664 break;
665 }
666
667 case 2:
668 // See spec 8.2.1.3.
669 if (prev_has_memmgmnt5_)
670 prev_frame_num_offset_ = 0;
671
672 if (slice_hdr->idr_pic_flag)
673 curr_pic_->frame_num_offset = 0;
674 else if (prev_frame_num_ > slice_hdr->frame_num)
675 curr_pic_->frame_num_offset = prev_frame_num_offset_ + max_frame_num_;
676 else
677 curr_pic_->frame_num_offset = prev_frame_num_offset_;
678
679 int temp_pic_order_cnt;
680 if (slice_hdr->idr_pic_flag) {
681 temp_pic_order_cnt = 0;
682 } else if (!slice_hdr->nal_ref_idc) {
683 temp_pic_order_cnt =
684 2 * (curr_pic_->frame_num_offset + slice_hdr->frame_num) - 1;
685 } else {
686 temp_pic_order_cnt = 2 * (curr_pic_->frame_num_offset +
687 slice_hdr->frame_num);
688 }
689
690 if (!slice_hdr->field_pic_flag) {
691 curr_pic_->top_field_order_cnt = temp_pic_order_cnt;
692 curr_pic_->bottom_field_order_cnt = temp_pic_order_cnt;
693 } else if (slice_hdr->bottom_field_flag) {
694 curr_pic_->bottom_field_order_cnt = temp_pic_order_cnt;
695 } else {
696 curr_pic_->top_field_order_cnt = temp_pic_order_cnt;
697 }
698 break;
699
700 default:
701 DVLOG(1) << "Invalid pic_order_cnt_type: " << sps->pic_order_cnt_type;
702 return false;
703 }
704
705 switch (curr_pic_->field) {
706 case H264Picture::FIELD_NONE:
707 curr_pic_->pic_order_cnt = std::min(curr_pic_->top_field_order_cnt,
708 curr_pic_->bottom_field_order_cnt);
709 break;
710 case H264Picture::FIELD_TOP:
711 curr_pic_->pic_order_cnt = curr_pic_->top_field_order_cnt;
712 break;
713 case H264Picture::FIELD_BOTTOM:
714 curr_pic_->pic_order_cnt = curr_pic_->bottom_field_order_cnt;
715 break;
716 }
717
718 return true;
719 }
720
UpdatePicNums()721 void VaapiH264Decoder::UpdatePicNums() {
722 for (H264DPB::Pictures::iterator it = dpb_.begin(); it != dpb_.end(); ++it) {
723 H264Picture* pic = *it;
724 DCHECK(pic);
725 if (!pic->ref)
726 continue;
727
728 // Below assumes non-interlaced stream.
729 DCHECK_EQ(pic->field, H264Picture::FIELD_NONE);
730 if (pic->long_term) {
731 pic->long_term_pic_num = pic->long_term_frame_idx;
732 } else {
733 if (pic->frame_num > frame_num_)
734 pic->frame_num_wrap = pic->frame_num - max_frame_num_;
735 else
736 pic->frame_num_wrap = pic->frame_num;
737
738 pic->pic_num = pic->frame_num_wrap;
739 }
740 }
741 }
742
743 struct PicNumDescCompare {
operator ()content::PicNumDescCompare744 bool operator()(const H264Picture* a, const H264Picture* b) const {
745 return a->pic_num > b->pic_num;
746 }
747 };
748
749 struct LongTermPicNumAscCompare {
operator ()content::LongTermPicNumAscCompare750 bool operator()(const H264Picture* a, const H264Picture* b) const {
751 return a->long_term_pic_num < b->long_term_pic_num;
752 }
753 };
754
ConstructReferencePicListsP(media::H264SliceHeader * slice_hdr)755 void VaapiH264Decoder::ConstructReferencePicListsP(
756 media::H264SliceHeader* slice_hdr) {
757 // RefPicList0 (8.2.4.2.1) [[1] [2]], where:
758 // [1] shortterm ref pics sorted by descending pic_num,
759 // [2] longterm ref pics by ascending long_term_pic_num.
760 DCHECK(ref_pic_list0_.empty() && ref_pic_list1_.empty());
761 // First get the short ref pics...
762 dpb_.GetShortTermRefPicsAppending(ref_pic_list0_);
763 size_t num_short_refs = ref_pic_list0_.size();
764
765 // and sort them to get [1].
766 std::sort(ref_pic_list0_.begin(), ref_pic_list0_.end(), PicNumDescCompare());
767
768 // Now get long term pics and sort them by long_term_pic_num to get [2].
769 dpb_.GetLongTermRefPicsAppending(ref_pic_list0_);
770 std::sort(ref_pic_list0_.begin() + num_short_refs, ref_pic_list0_.end(),
771 LongTermPicNumAscCompare());
772
773 // Cut off if we have more than requested in slice header.
774 ref_pic_list0_.resize(slice_hdr->num_ref_idx_l0_active_minus1 + 1);
775 }
776
777 struct POCAscCompare {
operator ()content::POCAscCompare778 bool operator()(const H264Picture* a, const H264Picture* b) const {
779 return a->pic_order_cnt < b->pic_order_cnt;
780 }
781 };
782
783 struct POCDescCompare {
operator ()content::POCDescCompare784 bool operator()(const H264Picture* a, const H264Picture* b) const {
785 return a->pic_order_cnt > b->pic_order_cnt;
786 }
787 };
788
ConstructReferencePicListsB(media::H264SliceHeader * slice_hdr)789 void VaapiH264Decoder::ConstructReferencePicListsB(
790 media::H264SliceHeader* slice_hdr) {
791 // RefPicList0 (8.2.4.2.3) [[1] [2] [3]], where:
792 // [1] shortterm ref pics with POC < curr_pic's POC sorted by descending POC,
793 // [2] shortterm ref pics with POC > curr_pic's POC by ascending POC,
794 // [3] longterm ref pics by ascending long_term_pic_num.
795 DCHECK(ref_pic_list0_.empty() && ref_pic_list1_.empty());
796 dpb_.GetShortTermRefPicsAppending(ref_pic_list0_);
797 size_t num_short_refs = ref_pic_list0_.size();
798
799 // First sort ascending, this will put [1] in right place and finish [2].
800 std::sort(ref_pic_list0_.begin(), ref_pic_list0_.end(), POCAscCompare());
801
802 // Find first with POC > curr_pic's POC to get first element in [2]...
803 H264Picture::PtrVector::iterator iter;
804 iter = std::upper_bound(ref_pic_list0_.begin(), ref_pic_list0_.end(),
805 curr_pic_.get(), POCAscCompare());
806
807 // and sort [1] descending, thus finishing sequence [1] [2].
808 std::sort(ref_pic_list0_.begin(), iter, POCDescCompare());
809
810 // Now add [3] and sort by ascending long_term_pic_num.
811 dpb_.GetLongTermRefPicsAppending(ref_pic_list0_);
812 std::sort(ref_pic_list0_.begin() + num_short_refs, ref_pic_list0_.end(),
813 LongTermPicNumAscCompare());
814
815 // RefPicList1 (8.2.4.2.4) [[1] [2] [3]], where:
816 // [1] shortterm ref pics with POC > curr_pic's POC sorted by ascending POC,
817 // [2] shortterm ref pics with POC < curr_pic's POC by descending POC,
818 // [3] longterm ref pics by ascending long_term_pic_num.
819
820 dpb_.GetShortTermRefPicsAppending(ref_pic_list1_);
821 num_short_refs = ref_pic_list1_.size();
822
823 // First sort by descending POC.
824 std::sort(ref_pic_list1_.begin(), ref_pic_list1_.end(), POCDescCompare());
825
826 // Find first with POC < curr_pic's POC to get first element in [2]...
827 iter = std::upper_bound(ref_pic_list1_.begin(), ref_pic_list1_.end(),
828 curr_pic_.get(), POCDescCompare());
829
830 // and sort [1] ascending.
831 std::sort(ref_pic_list1_.begin(), iter, POCAscCompare());
832
833 // Now add [3] and sort by ascending long_term_pic_num
834 dpb_.GetShortTermRefPicsAppending(ref_pic_list1_);
835 std::sort(ref_pic_list1_.begin() + num_short_refs, ref_pic_list1_.end(),
836 LongTermPicNumAscCompare());
837
838 // If lists identical, swap first two entries in RefPicList1 (spec 8.2.4.2.3)
839 if (ref_pic_list1_.size() > 1 &&
840 std::equal(ref_pic_list0_.begin(), ref_pic_list0_.end(),
841 ref_pic_list1_.begin()))
842 std::swap(ref_pic_list1_[0], ref_pic_list1_[1]);
843
844 // Per 8.2.4.2 it's possible for num_ref_idx_lX_active_minus1 to indicate
845 // there should be more ref pics on list than we constructed.
846 // Those superfluous ones should be treated as non-reference.
847 ref_pic_list0_.resize(slice_hdr->num_ref_idx_l0_active_minus1 + 1);
848 ref_pic_list1_.resize(slice_hdr->num_ref_idx_l1_active_minus1 + 1);
849 }
850
851 // See 8.2.4
PicNumF(H264Picture * pic)852 int VaapiH264Decoder::PicNumF(H264Picture *pic) {
853 if (!pic)
854 return -1;
855
856 if (!pic->long_term)
857 return pic->pic_num;
858 else
859 return max_pic_num_;
860 }
861
862 // See 8.2.4
LongTermPicNumF(H264Picture * pic)863 int VaapiH264Decoder::LongTermPicNumF(H264Picture *pic) {
864 if (pic->ref && pic->long_term)
865 return pic->long_term_pic_num;
866 else
867 return 2 * (max_long_term_frame_idx_ + 1);
868 }
869
870 // Shift elements on the |v| starting from |from| to |to|, inclusive,
871 // one position to the right and insert pic at |from|.
ShiftRightAndInsert(H264Picture::PtrVector * v,int from,int to,H264Picture * pic)872 static void ShiftRightAndInsert(H264Picture::PtrVector *v,
873 int from,
874 int to,
875 H264Picture* pic) {
876 // Security checks, do not disable in Debug mode.
877 CHECK(from <= to);
878 CHECK(to <= std::numeric_limits<int>::max() - 2);
879 // Additional checks. Debug mode ok.
880 DCHECK(v);
881 DCHECK(pic);
882 DCHECK((to + 1 == static_cast<int>(v->size())) ||
883 (to + 2 == static_cast<int>(v->size())));
884
885 v->resize(to + 2);
886
887 for (int i = to + 1; i > from; --i)
888 (*v)[i] = (*v)[i - 1];
889
890 (*v)[from] = pic;
891 }
892
ModifyReferencePicList(media::H264SliceHeader * slice_hdr,int list)893 bool VaapiH264Decoder::ModifyReferencePicList(media::H264SliceHeader* slice_hdr,
894 int list) {
895 int num_ref_idx_lX_active_minus1;
896 H264Picture::PtrVector* ref_pic_listx;
897 media::H264ModificationOfPicNum* list_mod;
898
899 // This can process either ref_pic_list0 or ref_pic_list1, depending on
900 // the list argument. Set up pointers to proper list to be processed here.
901 if (list == 0) {
902 if (!slice_hdr->ref_pic_list_modification_flag_l0)
903 return true;
904
905 list_mod = slice_hdr->ref_list_l0_modifications;
906 num_ref_idx_lX_active_minus1 = ref_pic_list0_.size() - 1;
907
908 ref_pic_listx = &ref_pic_list0_;
909 } else {
910 if (!slice_hdr->ref_pic_list_modification_flag_l1)
911 return true;
912
913 list_mod = slice_hdr->ref_list_l1_modifications;
914 num_ref_idx_lX_active_minus1 = ref_pic_list1_.size() - 1;
915
916 ref_pic_listx = &ref_pic_list1_;
917 }
918
919 DCHECK_GE(num_ref_idx_lX_active_minus1, 0);
920
921 // Spec 8.2.4.3:
922 // Reorder pictures on the list in a way specified in the stream.
923 int pic_num_lx_pred = curr_pic_->pic_num;
924 int ref_idx_lx = 0;
925 int pic_num_lx_no_wrap;
926 int pic_num_lx;
927 bool done = false;
928 H264Picture* pic;
929 for (int i = 0; i < media::H264SliceHeader::kRefListModSize && !done; ++i) {
930 switch (list_mod->modification_of_pic_nums_idc) {
931 case 0:
932 case 1:
933 // Modify short reference picture position.
934 if (list_mod->modification_of_pic_nums_idc == 0) {
935 // Subtract given value from predicted PicNum.
936 pic_num_lx_no_wrap = pic_num_lx_pred -
937 (static_cast<int>(list_mod->abs_diff_pic_num_minus1) + 1);
938 // Wrap around max_pic_num_ if it becomes < 0 as result
939 // of subtraction.
940 if (pic_num_lx_no_wrap < 0)
941 pic_num_lx_no_wrap += max_pic_num_;
942 } else {
943 // Add given value to predicted PicNum.
944 pic_num_lx_no_wrap = pic_num_lx_pred +
945 (static_cast<int>(list_mod->abs_diff_pic_num_minus1) + 1);
946 // Wrap around max_pic_num_ if it becomes >= max_pic_num_ as result
947 // of the addition.
948 if (pic_num_lx_no_wrap >= max_pic_num_)
949 pic_num_lx_no_wrap -= max_pic_num_;
950 }
951
952 // For use in next iteration.
953 pic_num_lx_pred = pic_num_lx_no_wrap;
954
955 if (pic_num_lx_no_wrap > curr_pic_->pic_num)
956 pic_num_lx = pic_num_lx_no_wrap - max_pic_num_;
957 else
958 pic_num_lx = pic_num_lx_no_wrap;
959
960 DCHECK_LT(num_ref_idx_lX_active_minus1 + 1,
961 media::H264SliceHeader::kRefListModSize);
962 pic = dpb_.GetShortRefPicByPicNum(pic_num_lx);
963 if (!pic) {
964 DVLOG(1) << "Malformed stream, no pic num " << pic_num_lx;
965 return false;
966 }
967 ShiftRightAndInsert(ref_pic_listx, ref_idx_lx,
968 num_ref_idx_lX_active_minus1, pic);
969 ref_idx_lx++;
970
971 for (int src = ref_idx_lx, dst = ref_idx_lx;
972 src <= num_ref_idx_lX_active_minus1 + 1; ++src) {
973 if (PicNumF((*ref_pic_listx)[src]) != pic_num_lx)
974 (*ref_pic_listx)[dst++] = (*ref_pic_listx)[src];
975 }
976 break;
977
978 case 2:
979 // Modify long term reference picture position.
980 DCHECK_LT(num_ref_idx_lX_active_minus1 + 1,
981 media::H264SliceHeader::kRefListModSize);
982 pic = dpb_.GetLongRefPicByLongTermPicNum(list_mod->long_term_pic_num);
983 if (!pic) {
984 DVLOG(1) << "Malformed stream, no pic num "
985 << list_mod->long_term_pic_num;
986 return false;
987 }
988 ShiftRightAndInsert(ref_pic_listx, ref_idx_lx,
989 num_ref_idx_lX_active_minus1, pic);
990 ref_idx_lx++;
991
992 for (int src = ref_idx_lx, dst = ref_idx_lx;
993 src <= num_ref_idx_lX_active_minus1 + 1; ++src) {
994 if (LongTermPicNumF((*ref_pic_listx)[src])
995 != static_cast<int>(list_mod->long_term_pic_num))
996 (*ref_pic_listx)[dst++] = (*ref_pic_listx)[src];
997 }
998 break;
999
1000 case 3:
1001 // End of modification list.
1002 done = true;
1003 break;
1004
1005 default:
1006 // May be recoverable.
1007 DVLOG(1) << "Invalid modification_of_pic_nums_idc="
1008 << list_mod->modification_of_pic_nums_idc
1009 << " in position " << i;
1010 break;
1011 }
1012
1013 ++list_mod;
1014 }
1015
1016 // Per NOTE 2 in 8.2.4.3.2, the ref_pic_listx size in the above loop is
1017 // temporarily made one element longer than the required final list.
1018 // Resize the list back to its required size.
1019 ref_pic_listx->resize(num_ref_idx_lX_active_minus1 + 1);
1020
1021 return true;
1022 }
1023
OutputPic(H264Picture * pic)1024 bool VaapiH264Decoder::OutputPic(H264Picture* pic) {
1025 DCHECK(!pic->outputted);
1026 pic->outputted = true;
1027 last_output_poc_ = pic->pic_order_cnt;
1028
1029 DecodeSurface* dec_surface = DecodeSurfaceByPoC(pic->pic_order_cnt);
1030 if (!dec_surface)
1031 return false;
1032
1033 DCHECK_GE(dec_surface->input_id(), 0);
1034 DVLOG(4) << "Posting output task for POC: " << pic->pic_order_cnt
1035 << " input_id: " << dec_surface->input_id();
1036 output_pic_cb_.Run(dec_surface->input_id(), dec_surface->va_surface());
1037
1038 return true;
1039 }
1040
ClearDPB()1041 void VaapiH264Decoder::ClearDPB() {
1042 // Clear DPB contents, marking the pictures as unused first.
1043 for (H264DPB::Pictures::iterator it = dpb_.begin(); it != dpb_.end(); ++it)
1044 UnassignSurfaceFromPoC((*it)->pic_order_cnt);
1045
1046 dpb_.Clear();
1047 last_output_poc_ = std::numeric_limits<int>::min();
1048 }
1049
OutputAllRemainingPics()1050 bool VaapiH264Decoder::OutputAllRemainingPics() {
1051 // Output all pictures that are waiting to be outputted.
1052 FinishPrevFrameIfPresent();
1053 H264Picture::PtrVector to_output;
1054 dpb_.GetNotOutputtedPicsAppending(to_output);
1055 // Sort them by ascending POC to output in order.
1056 std::sort(to_output.begin(), to_output.end(), POCAscCompare());
1057
1058 H264Picture::PtrVector::iterator it;
1059 for (it = to_output.begin(); it != to_output.end(); ++it) {
1060 if (!OutputPic(*it)) {
1061 DVLOG(1) << "Failed to output pic POC: " << (*it)->pic_order_cnt;
1062 return false;
1063 }
1064 }
1065
1066 return true;
1067 }
1068
Flush()1069 bool VaapiH264Decoder::Flush() {
1070 DVLOG(2) << "Decoder flush";
1071
1072 if (!OutputAllRemainingPics())
1073 return false;
1074
1075 ClearDPB();
1076
1077 DCHECK(decode_surfaces_in_use_.empty());
1078 return true;
1079 }
1080
StartNewFrame(media::H264SliceHeader * slice_hdr)1081 bool VaapiH264Decoder::StartNewFrame(media::H264SliceHeader* slice_hdr) {
1082 // TODO posciak: add handling of max_num_ref_frames per spec.
1083
1084 // If the new frame is an IDR, output what's left to output and clear DPB
1085 if (slice_hdr->idr_pic_flag) {
1086 // (unless we are explicitly instructed not to do so).
1087 if (!slice_hdr->no_output_of_prior_pics_flag) {
1088 // Output DPB contents.
1089 if (!Flush())
1090 return false;
1091 }
1092 dpb_.Clear();
1093 last_output_poc_ = std::numeric_limits<int>::min();
1094 }
1095
1096 // curr_pic_ should have either been added to DPB or discarded when finishing
1097 // the last frame. DPB is responsible for releasing that memory once it's
1098 // not needed anymore.
1099 DCHECK(!curr_pic_.get());
1100 curr_pic_.reset(new H264Picture);
1101 CHECK(curr_pic_.get());
1102
1103 if (!InitCurrPicture(slice_hdr))
1104 return false;
1105
1106 DCHECK_GT(max_frame_num_, 0);
1107
1108 UpdatePicNums();
1109
1110 // Send parameter buffers before each new picture, before the first slice.
1111 if (!SendPPS())
1112 return false;
1113
1114 if (!SendIQMatrix())
1115 return false;
1116
1117 if (!QueueSlice(slice_hdr))
1118 return false;
1119
1120 return true;
1121 }
1122
HandleMemoryManagementOps()1123 bool VaapiH264Decoder::HandleMemoryManagementOps() {
1124 // 8.2.5.4
1125 for (unsigned int i = 0; i < arraysize(curr_pic_->ref_pic_marking); ++i) {
1126 // Code below does not support interlaced stream (per-field pictures).
1127 media::H264DecRefPicMarking* ref_pic_marking =
1128 &curr_pic_->ref_pic_marking[i];
1129 H264Picture* to_mark;
1130 int pic_num_x;
1131
1132 switch (ref_pic_marking->memory_mgmnt_control_operation) {
1133 case 0:
1134 // Normal end of operations' specification.
1135 return true;
1136
1137 case 1:
1138 // Mark a short term reference picture as unused so it can be removed
1139 // if outputted.
1140 pic_num_x = curr_pic_->pic_num -
1141 (ref_pic_marking->difference_of_pic_nums_minus1 + 1);
1142 to_mark = dpb_.GetShortRefPicByPicNum(pic_num_x);
1143 if (to_mark) {
1144 to_mark->ref = false;
1145 } else {
1146 DVLOG(1) << "Invalid short ref pic num to unmark";
1147 return false;
1148 }
1149 break;
1150
1151 case 2:
1152 // Mark a long term reference picture as unused so it can be removed
1153 // if outputted.
1154 to_mark = dpb_.GetLongRefPicByLongTermPicNum(
1155 ref_pic_marking->long_term_pic_num);
1156 if (to_mark) {
1157 to_mark->ref = false;
1158 } else {
1159 DVLOG(1) << "Invalid long term ref pic num to unmark";
1160 return false;
1161 }
1162 break;
1163
1164 case 3:
1165 // Mark a short term reference picture as long term reference.
1166 pic_num_x = curr_pic_->pic_num -
1167 (ref_pic_marking->difference_of_pic_nums_minus1 + 1);
1168 to_mark = dpb_.GetShortRefPicByPicNum(pic_num_x);
1169 if (to_mark) {
1170 DCHECK(to_mark->ref && !to_mark->long_term);
1171 to_mark->long_term = true;
1172 to_mark->long_term_frame_idx = ref_pic_marking->long_term_frame_idx;
1173 } else {
1174 DVLOG(1) << "Invalid short term ref pic num to mark as long ref";
1175 return false;
1176 }
1177 break;
1178
1179 case 4: {
1180 // Unmark all reference pictures with long_term_frame_idx over new max.
1181 max_long_term_frame_idx_
1182 = ref_pic_marking->max_long_term_frame_idx_plus1 - 1;
1183 H264Picture::PtrVector long_terms;
1184 dpb_.GetLongTermRefPicsAppending(long_terms);
1185 for (size_t i = 0; i < long_terms.size(); ++i) {
1186 H264Picture* pic = long_terms[i];
1187 DCHECK(pic->ref && pic->long_term);
1188 // Ok to cast, max_long_term_frame_idx is much smaller than 16bit.
1189 if (pic->long_term_frame_idx >
1190 static_cast<int>(max_long_term_frame_idx_))
1191 pic->ref = false;
1192 }
1193 break;
1194 }
1195
1196 case 5:
1197 // Unmark all reference pictures.
1198 dpb_.MarkAllUnusedForRef();
1199 max_long_term_frame_idx_ = -1;
1200 curr_pic_->mem_mgmt_5 = true;
1201 break;
1202
1203 case 6: {
1204 // Replace long term reference pictures with current picture.
1205 // First unmark if any existing with this long_term_frame_idx...
1206 H264Picture::PtrVector long_terms;
1207 dpb_.GetLongTermRefPicsAppending(long_terms);
1208 for (size_t i = 0; i < long_terms.size(); ++i) {
1209 H264Picture* pic = long_terms[i];
1210 DCHECK(pic->ref && pic->long_term);
1211 // Ok to cast, long_term_frame_idx is much smaller than 16bit.
1212 if (pic->long_term_frame_idx ==
1213 static_cast<int>(ref_pic_marking->long_term_frame_idx))
1214 pic->ref = false;
1215 }
1216
1217 // and mark the current one instead.
1218 curr_pic_->ref = true;
1219 curr_pic_->long_term = true;
1220 curr_pic_->long_term_frame_idx = ref_pic_marking->long_term_frame_idx;
1221 break;
1222 }
1223
1224 default:
1225 // Would indicate a bug in parser.
1226 NOTREACHED();
1227 }
1228 }
1229
1230 return true;
1231 }
1232
1233 // This method ensures that DPB does not overflow, either by removing
1234 // reference pictures as specified in the stream, or using a sliding window
1235 // procedure to remove the oldest one.
1236 // It also performs marking and unmarking pictures as reference.
1237 // See spac 8.2.5.1.
ReferencePictureMarking()1238 void VaapiH264Decoder::ReferencePictureMarking() {
1239 if (curr_pic_->idr) {
1240 // If current picture is an IDR, all reference pictures are unmarked.
1241 dpb_.MarkAllUnusedForRef();
1242
1243 if (curr_pic_->long_term_reference_flag) {
1244 curr_pic_->long_term = true;
1245 curr_pic_->long_term_frame_idx = 0;
1246 max_long_term_frame_idx_ = 0;
1247 } else {
1248 curr_pic_->long_term = false;
1249 max_long_term_frame_idx_ = -1;
1250 }
1251 } else {
1252 if (!curr_pic_->adaptive_ref_pic_marking_mode_flag) {
1253 // If non-IDR, and the stream does not indicate what we should do to
1254 // ensure DPB doesn't overflow, discard oldest picture.
1255 // See spec 8.2.5.3.
1256 if (curr_pic_->field == H264Picture::FIELD_NONE) {
1257 DCHECK_LE(dpb_.CountRefPics(),
1258 std::max<int>(parser_.GetSPS(curr_sps_id_)->max_num_ref_frames,
1259 1));
1260 if (dpb_.CountRefPics() ==
1261 std::max<int>(parser_.GetSPS(curr_sps_id_)->max_num_ref_frames,
1262 1)) {
1263 // Max number of reference pics reached,
1264 // need to remove one of the short term ones.
1265 // Find smallest frame_num_wrap short reference picture and mark
1266 // it as unused.
1267 H264Picture* to_unmark = dpb_.GetLowestFrameNumWrapShortRefPic();
1268 if (to_unmark == NULL) {
1269 DVLOG(1) << "Couldn't find a short ref picture to unmark";
1270 return;
1271 }
1272 to_unmark->ref = false;
1273 }
1274 } else {
1275 // Shouldn't get here.
1276 DVLOG(1) << "Interlaced video not supported.";
1277 report_error_to_uma_cb_.Run(INTERLACED_STREAM);
1278 }
1279 } else {
1280 // Stream has instructions how to discard pictures from DPB and how
1281 // to mark/unmark existing reference pictures. Do it.
1282 // Spec 8.2.5.4.
1283 if (curr_pic_->field == H264Picture::FIELD_NONE) {
1284 HandleMemoryManagementOps();
1285 } else {
1286 // Shouldn't get here.
1287 DVLOG(1) << "Interlaced video not supported.";
1288 report_error_to_uma_cb_.Run(INTERLACED_STREAM);
1289 }
1290 }
1291 }
1292 }
1293
FinishPicture()1294 bool VaapiH264Decoder::FinishPicture() {
1295 DCHECK(curr_pic_.get());
1296
1297 // Finish processing previous picture.
1298 // Start by storing previous reference picture data for later use,
1299 // if picture being finished is a reference picture.
1300 if (curr_pic_->ref) {
1301 ReferencePictureMarking();
1302 prev_ref_has_memmgmnt5_ = curr_pic_->mem_mgmt_5;
1303 prev_ref_top_field_order_cnt_ = curr_pic_->top_field_order_cnt;
1304 prev_ref_pic_order_cnt_msb_ = curr_pic_->pic_order_cnt_msb;
1305 prev_ref_pic_order_cnt_lsb_ = curr_pic_->pic_order_cnt_lsb;
1306 prev_ref_field_ = curr_pic_->field;
1307 }
1308 prev_has_memmgmnt5_ = curr_pic_->mem_mgmt_5;
1309 prev_frame_num_offset_ = curr_pic_->frame_num_offset;
1310
1311 // Remove unused (for reference or later output) pictures from DPB, marking
1312 // them as such.
1313 for (H264DPB::Pictures::iterator it = dpb_.begin(); it != dpb_.end(); ++it) {
1314 if ((*it)->outputted && !(*it)->ref)
1315 UnassignSurfaceFromPoC((*it)->pic_order_cnt);
1316 }
1317 dpb_.DeleteUnused();
1318
1319 DVLOG(4) << "Finishing picture, entries in DPB: " << dpb_.size();
1320
1321 // Whatever happens below, curr_pic_ will stop managing the pointer to the
1322 // picture after this function returns. The ownership will either be
1323 // transferred to DPB, if the image is still needed (for output and/or
1324 // reference), or the memory will be released if we manage to output it here
1325 // without having to store it for future reference.
1326 scoped_ptr<H264Picture> pic(curr_pic_.release());
1327
1328 // Get all pictures that haven't been outputted yet.
1329 H264Picture::PtrVector not_outputted;
1330 // TODO(posciak): pass as pointer, not reference (violates coding style).
1331 dpb_.GetNotOutputtedPicsAppending(not_outputted);
1332 // Include the one we've just decoded.
1333 not_outputted.push_back(pic.get());
1334
1335 // Sort in output order.
1336 std::sort(not_outputted.begin(), not_outputted.end(), POCAscCompare());
1337
1338 // Try to output as many pictures as we can. A picture can be output,
1339 // if the number of decoded and not yet outputted pictures that would remain
1340 // in DPB afterwards would at least be equal to max_num_reorder_frames.
1341 // If the outputted picture is not a reference picture, it doesn't have
1342 // to remain in the DPB and can be removed.
1343 H264Picture::PtrVector::iterator output_candidate = not_outputted.begin();
1344 size_t num_remaining = not_outputted.size();
1345 while (num_remaining > max_num_reorder_frames_) {
1346 int poc = (*output_candidate)->pic_order_cnt;
1347 DCHECK_GE(poc, last_output_poc_);
1348 if (!OutputPic(*output_candidate))
1349 return false;
1350
1351 if (!(*output_candidate)->ref) {
1352 // Current picture hasn't been inserted into DPB yet, so don't remove it
1353 // if we managed to output it immediately.
1354 if (*output_candidate != pic)
1355 dpb_.DeleteByPOC(poc);
1356 // Mark as unused.
1357 UnassignSurfaceFromPoC(poc);
1358 }
1359
1360 ++output_candidate;
1361 --num_remaining;
1362 }
1363
1364 // If we haven't managed to output the picture that we just decoded, or if
1365 // it's a reference picture, we have to store it in DPB.
1366 if (!pic->outputted || pic->ref) {
1367 if (dpb_.IsFull()) {
1368 // If we haven't managed to output anything to free up space in DPB
1369 // to store this picture, it's an error in the stream.
1370 DVLOG(1) << "Could not free up space in DPB!";
1371 return false;
1372 }
1373
1374 dpb_.StorePic(pic.release());
1375 }
1376
1377 return true;
1378 }
1379
LevelToMaxDpbMbs(int level)1380 static int LevelToMaxDpbMbs(int level) {
1381 // See table A-1 in spec.
1382 switch (level) {
1383 case 10: return 396;
1384 case 11: return 900;
1385 case 12: // fallthrough
1386 case 13: // fallthrough
1387 case 20: return 2376;
1388 case 21: return 4752;
1389 case 22: // fallthrough
1390 case 30: return 8100;
1391 case 31: return 18000;
1392 case 32: return 20480;
1393 case 40: // fallthrough
1394 case 41: return 32768;
1395 case 42: return 34816;
1396 case 50: return 110400;
1397 case 51: // fallthrough
1398 case 52: return 184320;
1399 default:
1400 DVLOG(1) << "Invalid codec level (" << level << ")";
1401 return 0;
1402 }
1403 }
1404
UpdateMaxNumReorderFrames(const media::H264SPS * sps)1405 bool VaapiH264Decoder::UpdateMaxNumReorderFrames(const media::H264SPS* sps) {
1406 if (sps->vui_parameters_present_flag && sps->bitstream_restriction_flag) {
1407 max_num_reorder_frames_ =
1408 base::checked_cast<size_t>(sps->max_num_reorder_frames);
1409 if (max_num_reorder_frames_ > dpb_.max_num_pics()) {
1410 DVLOG(1)
1411 << "max_num_reorder_frames present, but larger than MaxDpbFrames ("
1412 << max_num_reorder_frames_ << " > " << dpb_.max_num_pics() << ")";
1413 max_num_reorder_frames_ = 0;
1414 return false;
1415 }
1416 return true;
1417 }
1418
1419 // max_num_reorder_frames not present, infer from profile/constraints
1420 // (see VUI semantics in spec).
1421 if (sps->constraint_set3_flag) {
1422 switch (sps->profile_idc) {
1423 case 44:
1424 case 86:
1425 case 100:
1426 case 110:
1427 case 122:
1428 case 244:
1429 max_num_reorder_frames_ = 0;
1430 break;
1431 default:
1432 max_num_reorder_frames_ = dpb_.max_num_pics();
1433 break;
1434 }
1435 } else {
1436 max_num_reorder_frames_ = dpb_.max_num_pics();
1437 }
1438
1439 return true;
1440 }
1441
ProcessSPS(int sps_id,bool * need_new_buffers)1442 bool VaapiH264Decoder::ProcessSPS(int sps_id, bool* need_new_buffers) {
1443 const media::H264SPS* sps = parser_.GetSPS(sps_id);
1444 DCHECK(sps);
1445 DVLOG(4) << "Processing SPS";
1446
1447 *need_new_buffers = false;
1448
1449 if (sps->frame_mbs_only_flag == 0) {
1450 DVLOG(1) << "frame_mbs_only_flag != 1 not supported";
1451 report_error_to_uma_cb_.Run(FRAME_MBS_ONLY_FLAG_NOT_ONE);
1452 return false;
1453 }
1454
1455 if (sps->gaps_in_frame_num_value_allowed_flag) {
1456 DVLOG(1) << "Gaps in frame numbers not supported";
1457 report_error_to_uma_cb_.Run(GAPS_IN_FRAME_NUM);
1458 return false;
1459 }
1460
1461 curr_sps_id_ = sps->seq_parameter_set_id;
1462
1463 // Calculate picture height/width in macroblocks and pixels
1464 // (spec 7.4.2.1.1, 7.4.3).
1465 int width_mb = sps->pic_width_in_mbs_minus1 + 1;
1466 int height_mb = (2 - sps->frame_mbs_only_flag) *
1467 (sps->pic_height_in_map_units_minus1 + 1);
1468
1469 gfx::Size new_pic_size(16 * width_mb, 16 * height_mb);
1470 if (new_pic_size.IsEmpty()) {
1471 DVLOG(1) << "Invalid picture size: " << new_pic_size.ToString();
1472 return false;
1473 }
1474
1475 if (!pic_size_.IsEmpty() && new_pic_size == pic_size_) {
1476 // Already have surfaces and this SPS keeps the same resolution,
1477 // no need to request a new set.
1478 return true;
1479 }
1480
1481 pic_size_ = new_pic_size;
1482 DVLOG(1) << "New picture size: " << pic_size_.ToString();
1483
1484 max_pic_order_cnt_lsb_ = 1 << (sps->log2_max_pic_order_cnt_lsb_minus4 + 4);
1485 max_frame_num_ = 1 << (sps->log2_max_frame_num_minus4 + 4);
1486
1487 int level = sps->level_idc;
1488 int max_dpb_mbs = LevelToMaxDpbMbs(level);
1489 if (max_dpb_mbs == 0)
1490 return false;
1491
1492 size_t max_dpb_size = std::min(max_dpb_mbs / (width_mb * height_mb),
1493 static_cast<int>(H264DPB::kDPBMaxSize));
1494 DVLOG(1) << "Codec level: " << level << ", DPB size: " << max_dpb_size;
1495 if (max_dpb_size == 0) {
1496 DVLOG(1) << "Invalid DPB Size";
1497 return false;
1498 }
1499
1500 dpb_.set_max_num_pics(max_dpb_size);
1501
1502 if (!UpdateMaxNumReorderFrames(sps))
1503 return false;
1504 DVLOG(1) << "max_num_reorder_frames: " << max_num_reorder_frames_;
1505
1506 *need_new_buffers = true;
1507 return true;
1508 }
1509
ProcessPPS(int pps_id)1510 bool VaapiH264Decoder::ProcessPPS(int pps_id) {
1511 const media::H264PPS* pps = parser_.GetPPS(pps_id);
1512 DCHECK(pps);
1513
1514 curr_pps_id_ = pps->pic_parameter_set_id;
1515
1516 return true;
1517 }
1518
FinishPrevFrameIfPresent()1519 bool VaapiH264Decoder::FinishPrevFrameIfPresent() {
1520 // If we already have a frame waiting to be decoded, decode it and finish.
1521 if (curr_pic_ != NULL) {
1522 if (!DecodePicture())
1523 return false;
1524 return FinishPicture();
1525 }
1526
1527 return true;
1528 }
1529
ProcessSlice(media::H264SliceHeader * slice_hdr)1530 bool VaapiH264Decoder::ProcessSlice(media::H264SliceHeader* slice_hdr) {
1531 prev_frame_num_ = frame_num_;
1532 frame_num_ = slice_hdr->frame_num;
1533
1534 if (prev_frame_num_ > 0 && prev_frame_num_ < frame_num_ - 1) {
1535 DVLOG(1) << "Gap in frame_num!";
1536 report_error_to_uma_cb_.Run(GAPS_IN_FRAME_NUM);
1537 return false;
1538 }
1539
1540 if (slice_hdr->field_pic_flag == 0)
1541 max_pic_num_ = max_frame_num_;
1542 else
1543 max_pic_num_ = 2 * max_frame_num_;
1544
1545 // TODO posciak: switch to new picture detection per 7.4.1.2.4.
1546 if (curr_pic_ != NULL && slice_hdr->first_mb_in_slice != 0) {
1547 // This is just some more slice data of the current picture, so
1548 // just queue it and return.
1549 QueueSlice(slice_hdr);
1550 return true;
1551 } else {
1552 // A new frame, so first finish the previous one before processing it...
1553 if (!FinishPrevFrameIfPresent())
1554 return false;
1555
1556 // and then start a new one.
1557 return StartNewFrame(slice_hdr);
1558 }
1559 }
1560
1561 #define SET_ERROR_AND_RETURN() \
1562 do { \
1563 DVLOG(1) << "Error during decode"; \
1564 state_ = kError; \
1565 return VaapiH264Decoder::kDecodeError; \
1566 } while (0)
1567
SetStream(const uint8 * ptr,size_t size,int32 input_id)1568 void VaapiH264Decoder::SetStream(const uint8* ptr,
1569 size_t size,
1570 int32 input_id) {
1571 DCHECK(ptr);
1572 DCHECK(size);
1573
1574 // Got new input stream data from the client.
1575 DVLOG(4) << "New input stream id: " << input_id << " at: " << (void*) ptr
1576 << " size: " << size;
1577 parser_.SetStream(ptr, size);
1578 curr_input_id_ = input_id;
1579 }
1580
Decode()1581 VaapiH264Decoder::DecResult VaapiH264Decoder::Decode() {
1582 media::H264Parser::Result par_res;
1583 media::H264NALU nalu;
1584 DCHECK_NE(state_, kError);
1585
1586 while (1) {
1587 // If we've already decoded some of the stream (after reset, i.e. we are
1588 // not in kNeedStreamMetadata state), we may be able to go back into
1589 // decoding state not only starting at/resuming from an SPS, but also from
1590 // other resume points, such as IDRs. In the latter case we need an output
1591 // surface, because we will end up decoding that IDR in the process.
1592 // Otherwise we just look for an SPS and don't produce any output frames.
1593 if (state_ != kNeedStreamMetadata && available_va_surfaces_.empty()) {
1594 DVLOG(4) << "No output surfaces available";
1595 return kRanOutOfSurfaces;
1596 }
1597
1598 par_res = parser_.AdvanceToNextNALU(&nalu);
1599 if (par_res == media::H264Parser::kEOStream)
1600 return kRanOutOfStreamData;
1601 else if (par_res != media::H264Parser::kOk)
1602 SET_ERROR_AND_RETURN();
1603
1604 DVLOG(4) << "NALU found: " << static_cast<int>(nalu.nal_unit_type);
1605
1606 switch (nalu.nal_unit_type) {
1607 case media::H264NALU::kNonIDRSlice:
1608 // We can't resume from a non-IDR slice.
1609 if (state_ != kDecoding)
1610 break;
1611 // else fallthrough
1612 case media::H264NALU::kIDRSlice: {
1613 // TODO(posciak): the IDR may require an SPS that we don't have
1614 // available. For now we'd fail if that happens, but ideally we'd like
1615 // to keep going until the next SPS in the stream.
1616 if (state_ == kNeedStreamMetadata) {
1617 // We need an SPS, skip this IDR and keep looking.
1618 break;
1619 }
1620
1621 // If after reset, we should be able to recover from an IDR.
1622 media::H264SliceHeader slice_hdr;
1623
1624 par_res = parser_.ParseSliceHeader(nalu, &slice_hdr);
1625 if (par_res != media::H264Parser::kOk)
1626 SET_ERROR_AND_RETURN();
1627
1628 if (!ProcessSlice(&slice_hdr))
1629 SET_ERROR_AND_RETURN();
1630
1631 state_ = kDecoding;
1632 break;
1633 }
1634
1635 case media::H264NALU::kSPS: {
1636 int sps_id;
1637
1638 if (!FinishPrevFrameIfPresent())
1639 SET_ERROR_AND_RETURN();
1640
1641 par_res = parser_.ParseSPS(&sps_id);
1642 if (par_res != media::H264Parser::kOk)
1643 SET_ERROR_AND_RETURN();
1644
1645 bool need_new_buffers = false;
1646 if (!ProcessSPS(sps_id, &need_new_buffers))
1647 SET_ERROR_AND_RETURN();
1648
1649 state_ = kDecoding;
1650
1651 if (need_new_buffers) {
1652 if (!Flush())
1653 return kDecodeError;
1654
1655 available_va_surfaces_.clear();
1656 return kAllocateNewSurfaces;
1657 }
1658 break;
1659 }
1660
1661 case media::H264NALU::kPPS: {
1662 if (state_ != kDecoding)
1663 break;
1664
1665 int pps_id;
1666
1667 if (!FinishPrevFrameIfPresent())
1668 SET_ERROR_AND_RETURN();
1669
1670 par_res = parser_.ParsePPS(&pps_id);
1671 if (par_res != media::H264Parser::kOk)
1672 SET_ERROR_AND_RETURN();
1673
1674 if (!ProcessPPS(pps_id))
1675 SET_ERROR_AND_RETURN();
1676 break;
1677 }
1678
1679 default:
1680 DVLOG(4) << "Skipping NALU type: " << nalu.nal_unit_type;
1681 break;
1682 }
1683 }
1684 }
1685
GetRequiredNumOfPictures()1686 size_t VaapiH264Decoder::GetRequiredNumOfPictures() {
1687 return dpb_.max_num_pics() + kPicsInPipeline;
1688 }
1689
1690 } // namespace content
1691