]> git.sesse.net Git - ffmpeg/blob - libavcodec/dxva2_h264.c
h264/aarch64: optimize neon loop filter
[ffmpeg] / libavcodec / dxva2_h264.c
1 /*
2  * DXVA2 H.264 HW acceleration.
3  *
4  * copyright (c) 2009 Laurent Aimar
5  *
6  * This file is part of Libav.
7  *
8  * Libav is free software; you can redistribute it and/or
9  * modify it under the terms of the GNU Lesser General Public
10  * License as published by the Free Software Foundation; either
11  * version 2.1 of the License, or (at your option) any later version.
12  *
13  * Libav is distributed in the hope that it will be useful,
14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
16  * Lesser General Public License for more details.
17  *
18  * You should have received a copy of the GNU Lesser General Public
19  * License along with Libav; if not, write to the Free Software
20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21  */
22
23 #include "dxva2_internal.h"
24 #include "h264dec.h"
25 #include "h264data.h"
26 #include "h264_ps.h"
27 #include "mpegutils.h"
28
29 struct dxva2_picture_context {
30     DXVA_PicParams_H264   pp;
31     DXVA_Qmatrix_H264     qm;
32     unsigned              slice_count;
33     DXVA_Slice_H264_Short slice_short[MAX_SLICES];
34     DXVA_Slice_H264_Long  slice_long[MAX_SLICES];
35     const uint8_t         *bitstream;
36     unsigned              bitstream_size;
37 };
38
39 static void fill_picture_entry(DXVA_PicEntry_H264 *pic,
40                                unsigned index, unsigned flag)
41 {
42     assert((index&0x7f) == index && (flag&0x01) == flag);
43     pic->bPicEntry = index | (flag << 7);
44 }
45
46 static void fill_picture_parameters(const AVCodecContext *avctx, AVDXVAContext *ctx, const H264Context *h,
47                                     DXVA_PicParams_H264 *pp)
48 {
49     const H264Picture *current_picture = h->cur_pic_ptr;
50     const SPS *sps = h->ps.sps;
51     const PPS *pps = h->ps.pps;
52     int i, j;
53
54     memset(pp, 0, sizeof(*pp));
55     /* Configure current picture */
56     fill_picture_entry(&pp->CurrPic,
57                        ff_dxva2_get_surface_index(avctx, ctx, current_picture->f),
58                        h->picture_structure == PICT_BOTTOM_FIELD);
59     /* Configure the set of references */
60     pp->UsedForReferenceFlags  = 0;
61     pp->NonExistingFrameFlags  = 0;
62     for (i = 0, j = 0; i < FF_ARRAY_ELEMS(pp->RefFrameList); i++) {
63         const H264Picture *r;
64         if (j < h->short_ref_count) {
65             r = h->short_ref[j++];
66         } else {
67             r = NULL;
68             while (!r && j < h->short_ref_count + 16)
69                 r = h->long_ref[j++ - h->short_ref_count];
70         }
71         if (r) {
72             fill_picture_entry(&pp->RefFrameList[i],
73                                ff_dxva2_get_surface_index(avctx, ctx, r->f),
74                                r->long_ref != 0);
75
76             if ((r->reference & PICT_TOP_FIELD) && r->field_poc[0] != INT_MAX)
77                 pp->FieldOrderCntList[i][0] = r->field_poc[0];
78             if ((r->reference & PICT_BOTTOM_FIELD) && r->field_poc[1] != INT_MAX)
79                 pp->FieldOrderCntList[i][1] = r->field_poc[1];
80
81             pp->FrameNumList[i] = r->long_ref ? r->pic_id : r->frame_num;
82             if (r->reference & PICT_TOP_FIELD)
83                 pp->UsedForReferenceFlags |= 1 << (2*i + 0);
84             if (r->reference & PICT_BOTTOM_FIELD)
85                 pp->UsedForReferenceFlags |= 1 << (2*i + 1);
86         } else {
87             pp->RefFrameList[i].bPicEntry = 0xff;
88             pp->FieldOrderCntList[i][0]   = 0;
89             pp->FieldOrderCntList[i][1]   = 0;
90             pp->FrameNumList[i]           = 0;
91         }
92     }
93
94     pp->wFrameWidthInMbsMinus1        = h->mb_width  - 1;
95     pp->wFrameHeightInMbsMinus1       = h->mb_height - 1;
96     pp->num_ref_frames                = sps->ref_frame_count;
97
98     pp->wBitFields                    = ((h->picture_structure != PICT_FRAME) <<  0) |
99                                         ((sps->mb_aff &&
100                                         (h->picture_structure == PICT_FRAME)) <<  1) |
101                                         (sps->residual_color_transform_flag   <<  2) |
102                                         /* sp_for_switch_flag (not implemented by Libav) */
103                                         (0                                    <<  3) |
104                                         (sps->chroma_format_idc               <<  4) |
105                                         ((h->nal_ref_idc != 0)                <<  6) |
106                                         (pps->constrained_intra_pred          <<  7) |
107                                         (pps->weighted_pred                   <<  8) |
108                                         (pps->weighted_bipred_idc             <<  9) |
109                                         /* MbsConsecutiveFlag */
110                                         (1                                    << 11) |
111                                         (sps->frame_mbs_only_flag             << 12) |
112                                         (pps->transform_8x8_mode              << 13) |
113                                         ((sps->level_idc >= 31)               << 14) |
114                                         /* IntraPicFlag (Modified if we detect a non
115                                          * intra slice in dxva2_h264_decode_slice) */
116                                         (1                                    << 15);
117
118     pp->bit_depth_luma_minus8         = sps->bit_depth_luma - 8;
119     pp->bit_depth_chroma_minus8       = sps->bit_depth_chroma - 8;
120     if (DXVA_CONTEXT_WORKAROUND(avctx, ctx) & FF_DXVA2_WORKAROUND_SCALING_LIST_ZIGZAG)
121         pp->Reserved16Bits            = 0;
122     else if (DXVA_CONTEXT_WORKAROUND(avctx, ctx) & FF_DXVA2_WORKAROUND_INTEL_CLEARVIDEO)
123         pp->Reserved16Bits            = 0x34c;
124     else
125         pp->Reserved16Bits            = 3; /* FIXME is there a way to detect the right mode ? */
126     pp->StatusReportFeedbackNumber    = 1 + DXVA_CONTEXT_REPORT_ID(avctx, ctx)++;
127     pp->CurrFieldOrderCnt[0] = 0;
128     if ((h->picture_structure & PICT_TOP_FIELD) &&
129         current_picture->field_poc[0] != INT_MAX)
130         pp->CurrFieldOrderCnt[0] = current_picture->field_poc[0];
131     pp->CurrFieldOrderCnt[1] = 0;
132     if ((h->picture_structure & PICT_BOTTOM_FIELD) &&
133         current_picture->field_poc[1] != INT_MAX)
134         pp->CurrFieldOrderCnt[1] = current_picture->field_poc[1];
135     pp->pic_init_qs_minus26           = pps->init_qs - 26;
136     pp->chroma_qp_index_offset        = pps->chroma_qp_index_offset[0];
137     pp->second_chroma_qp_index_offset = pps->chroma_qp_index_offset[1];
138     pp->ContinuationFlag              = 1;
139     pp->pic_init_qp_minus26           = pps->init_qp - 26;
140     pp->num_ref_idx_l0_active_minus1  = pps->ref_count[0] - 1;
141     pp->num_ref_idx_l1_active_minus1  = pps->ref_count[1] - 1;
142     pp->Reserved8BitsA                = 0;
143     pp->frame_num                     = h->poc.frame_num;
144     pp->log2_max_frame_num_minus4     = sps->log2_max_frame_num - 4;
145     pp->pic_order_cnt_type            = sps->poc_type;
146     if (sps->poc_type == 0)
147         pp->log2_max_pic_order_cnt_lsb_minus4 = sps->log2_max_poc_lsb - 4;
148     else if (sps->poc_type == 1)
149         pp->delta_pic_order_always_zero_flag = sps->delta_pic_order_always_zero_flag;
150     pp->direct_8x8_inference_flag     = sps->direct_8x8_inference_flag;
151     pp->entropy_coding_mode_flag      = pps->cabac;
152     pp->pic_order_present_flag        = pps->pic_order_present;
153     pp->num_slice_groups_minus1       = pps->slice_group_count - 1;
154     pp->slice_group_map_type          = pps->mb_slice_group_map_type;
155     pp->deblocking_filter_control_present_flag = pps->deblocking_filter_parameters_present;
156     pp->redundant_pic_cnt_present_flag= pps->redundant_pic_cnt_present;
157     pp->Reserved8BitsB                = 0;
158     pp->slice_group_change_rate_minus1= 0;  /* XXX not implemented by Libav */
159     //pp->SliceGroupMap[810];               /* XXX not implemented by Libav */
160 }
161
162 static void fill_scaling_lists(const AVCodecContext *avctx, AVDXVAContext *ctx, const H264Context *h, DXVA_Qmatrix_H264 *qm)
163 {
164     unsigned i, j;
165     const PPS *pps = h->ps.pps;
166     memset(qm, 0, sizeof(*qm));
167     if (DXVA_CONTEXT_WORKAROUND(avctx, ctx) & FF_DXVA2_WORKAROUND_SCALING_LIST_ZIGZAG) {
168         for (i = 0; i < 6; i++)
169             for (j = 0; j < 16; j++)
170                 qm->bScalingLists4x4[i][j] = pps->scaling_matrix4[i][j];
171
172         for (i = 0; i < 64; i++) {
173             qm->bScalingLists8x8[0][i] = pps->scaling_matrix8[0][i];
174             qm->bScalingLists8x8[1][i] = pps->scaling_matrix8[3][i];
175         }
176     } else {
177         for (i = 0; i < 6; i++)
178             for (j = 0; j < 16; j++)
179                 qm->bScalingLists4x4[i][j] = pps->scaling_matrix4[i][ff_zigzag_scan[j]];
180
181         for (i = 0; i < 64; i++) {
182             qm->bScalingLists8x8[0][i] = pps->scaling_matrix8[0][ff_zigzag_direct[i]];
183             qm->bScalingLists8x8[1][i] = pps->scaling_matrix8[3][ff_zigzag_direct[i]];
184         }
185     }
186 }
187
188 static int is_slice_short(const AVCodecContext *avctx, AVDXVAContext *ctx)
189 {
190     assert(DXVA_CONTEXT_CFG_BITSTREAM(avctx, ctx) == 1 ||
191            DXVA_CONTEXT_CFG_BITSTREAM(avctx, ctx) == 2);
192     return DXVA_CONTEXT_CFG_BITSTREAM(avctx, ctx) == 2;
193 }
194
195 static void fill_slice_short(DXVA_Slice_H264_Short *slice,
196                              unsigned position, unsigned size)
197 {
198     memset(slice, 0, sizeof(*slice));
199     slice->BSNALunitDataLocation = position;
200     slice->SliceBytesInBuffer    = size;
201     slice->wBadSliceChopping     = 0;
202 }
203
204 static int get_refpic_index(const DXVA_PicParams_H264 *pp, int surface_index)
205 {
206     int i;
207     for (i = 0; i < FF_ARRAY_ELEMS(pp->RefFrameList); i++) {
208         if ((pp->RefFrameList[i].bPicEntry & 0x7f) == surface_index)
209           return i;
210     }
211     return 0x7f;
212 }
213
214 static void fill_slice_long(AVCodecContext *avctx, DXVA_Slice_H264_Long *slice,
215                             const DXVA_PicParams_H264 *pp, unsigned position, unsigned size)
216 {
217     const H264Context *h = avctx->priv_data;
218     H264SliceContext *sl = &h->slice_ctx[0];
219     AVDXVAContext *ctx = DXVA_CONTEXT(avctx);
220     unsigned list;
221
222     memset(slice, 0, sizeof(*slice));
223     slice->BSNALunitDataLocation = position;
224     slice->SliceBytesInBuffer    = size;
225     slice->wBadSliceChopping     = 0;
226
227     slice->first_mb_in_slice     = (sl->mb_y >> FIELD_OR_MBAFF_PICTURE(h)) * h->mb_width + sl->mb_x;
228     slice->NumMbsForSlice        = 0; /* XXX it is set once we have all slices */
229     slice->BitOffsetToSliceData  = get_bits_count(&sl->gb);
230     slice->slice_type            = ff_h264_get_slice_type(sl);
231     if (sl->slice_type_fixed)
232         slice->slice_type += 5;
233     slice->luma_log2_weight_denom       = sl->pwt.luma_log2_weight_denom;
234     slice->chroma_log2_weight_denom     = sl->pwt.chroma_log2_weight_denom;
235     if (sl->list_count > 0)
236         slice->num_ref_idx_l0_active_minus1 = sl->ref_count[0] - 1;
237     if (sl->list_count > 1)
238         slice->num_ref_idx_l1_active_minus1 = sl->ref_count[1] - 1;
239     slice->slice_alpha_c0_offset_div2   = sl->slice_alpha_c0_offset / 2;
240     slice->slice_beta_offset_div2       = sl->slice_beta_offset     / 2;
241     slice->Reserved8Bits                = 0;
242
243     for (list = 0; list < 2; list++) {
244         unsigned i;
245         for (i = 0; i < FF_ARRAY_ELEMS(slice->RefPicList[list]); i++) {
246             if (list < sl->list_count && i < sl->ref_count[list]) {
247                 const H264Picture *r = sl->ref_list[list][i].parent;
248                 unsigned plane;
249                 unsigned index;
250                 if (DXVA_CONTEXT_WORKAROUND(avctx, ctx) & FF_DXVA2_WORKAROUND_INTEL_CLEARVIDEO)
251                     index = ff_dxva2_get_surface_index(avctx, ctx, r->f);
252                 else
253                     index = get_refpic_index(pp, ff_dxva2_get_surface_index(avctx, ctx, r->f));
254                 fill_picture_entry(&slice->RefPicList[list][i], index,
255                                    r->reference == PICT_BOTTOM_FIELD);
256                 for (plane = 0; plane < 3; plane++) {
257                     int w, o;
258                     if (plane == 0 && sl->pwt.luma_weight_flag[list]) {
259                         w = sl->pwt.luma_weight[i][list][0];
260                         o = sl->pwt.luma_weight[i][list][1];
261                     } else if (plane >= 1 && sl->pwt.chroma_weight_flag[list]) {
262                         w = sl->pwt.chroma_weight[i][list][plane-1][0];
263                         o = sl->pwt.chroma_weight[i][list][plane-1][1];
264                     } else {
265                         w = 1 << (plane == 0 ? sl->pwt.luma_log2_weight_denom :
266                                                sl->pwt.chroma_log2_weight_denom);
267                         o = 0;
268                     }
269                     slice->Weights[list][i][plane][0] = w;
270                     slice->Weights[list][i][plane][1] = o;
271                 }
272             } else {
273                 unsigned plane;
274                 slice->RefPicList[list][i].bPicEntry = 0xff;
275                 for (plane = 0; plane < 3; plane++) {
276                     slice->Weights[list][i][plane][0] = 0;
277                     slice->Weights[list][i][plane][1] = 0;
278                 }
279             }
280         }
281     }
282     slice->slice_qs_delta    = 0; /* XXX not implemented by Libav */
283     slice->slice_qp_delta    = sl->qscale - h->ps.pps->init_qp;
284     slice->redundant_pic_cnt = sl->redundant_pic_count;
285     if (sl->slice_type == AV_PICTURE_TYPE_B)
286         slice->direct_spatial_mv_pred_flag = sl->direct_spatial_mv_pred;
287     slice->cabac_init_idc = h->ps.pps->cabac ? sl->cabac_init_idc : 0;
288     if (sl->deblocking_filter < 2)
289         slice->disable_deblocking_filter_idc = 1 - sl->deblocking_filter;
290     else
291         slice->disable_deblocking_filter_idc = sl->deblocking_filter;
292     slice->slice_id = h->current_slice - 1;
293 }
294
295 static int commit_bitstream_and_slice_buffer(AVCodecContext *avctx,
296                                              DECODER_BUFFER_DESC *bs,
297                                              DECODER_BUFFER_DESC *sc)
298 {
299     const H264Context *h = avctx->priv_data;
300     const unsigned mb_count = h->mb_width * h->mb_height;
301     AVDXVAContext *ctx = DXVA_CONTEXT(avctx);
302     const H264Picture *current_picture = h->cur_pic_ptr;
303     struct dxva2_picture_context *ctx_pic = current_picture->hwaccel_picture_private;
304     DXVA_Slice_H264_Short *slice = NULL;
305     void     *dxva_data_ptr;
306     uint8_t  *dxva_data, *current, *end;
307     unsigned dxva_size;
308     void     *slice_data;
309     unsigned slice_size;
310     unsigned padding;
311     unsigned i;
312     unsigned type;
313
314     /* Create an annex B bitstream buffer with only slice NAL and finalize slice */
315 #if CONFIG_D3D11VA
316     if (ff_dxva2_is_d3d11(avctx)) {
317         type = D3D11_VIDEO_DECODER_BUFFER_BITSTREAM;
318         if (FAILED(ID3D11VideoContext_GetDecoderBuffer(D3D11VA_CONTEXT(ctx)->video_context,
319                                                        D3D11VA_CONTEXT(ctx)->decoder,
320                                                        type,
321                                                        &dxva_size, &dxva_data_ptr)))
322             return -1;
323     }
324 #endif
325 #if CONFIG_DXVA2
326     if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD) {
327         type = DXVA2_BitStreamDateBufferType;
328         if (FAILED(IDirectXVideoDecoder_GetBuffer(DXVA2_CONTEXT(ctx)->decoder,
329                                                   type,
330                                                   &dxva_data_ptr, &dxva_size)))
331             return -1;
332     }
333 #endif
334
335     dxva_data = dxva_data_ptr;
336     current = dxva_data;
337     end = dxva_data + dxva_size;
338
339     for (i = 0; i < ctx_pic->slice_count; i++) {
340         static const uint8_t start_code[] = { 0, 0, 1 };
341         static const unsigned start_code_size = sizeof(start_code);
342         unsigned position, size;
343
344         assert(offsetof(DXVA_Slice_H264_Short, BSNALunitDataLocation) ==
345                offsetof(DXVA_Slice_H264_Long,  BSNALunitDataLocation));
346         assert(offsetof(DXVA_Slice_H264_Short, SliceBytesInBuffer) ==
347                offsetof(DXVA_Slice_H264_Long,  SliceBytesInBuffer));
348
349         if (is_slice_short(avctx, ctx))
350             slice = &ctx_pic->slice_short[i];
351         else
352             slice = (DXVA_Slice_H264_Short*)&ctx_pic->slice_long[i];
353
354         position = slice->BSNALunitDataLocation;
355         size     = slice->SliceBytesInBuffer;
356         if (start_code_size + size > end - current) {
357             av_log(avctx, AV_LOG_ERROR, "Failed to build bitstream");
358             break;
359         }
360
361         slice->BSNALunitDataLocation = current - dxva_data;
362         slice->SliceBytesInBuffer    = start_code_size + size;
363
364         if (!is_slice_short(avctx, ctx)) {
365             DXVA_Slice_H264_Long *slice_long = (DXVA_Slice_H264_Long*)slice;
366             if (i < ctx_pic->slice_count - 1)
367                 slice_long->NumMbsForSlice =
368                     slice_long[1].first_mb_in_slice - slice_long[0].first_mb_in_slice;
369             else
370                 slice_long->NumMbsForSlice = mb_count - slice_long->first_mb_in_slice;
371         }
372
373         memcpy(current, start_code, start_code_size);
374         current += start_code_size;
375
376         memcpy(current, &ctx_pic->bitstream[position], size);
377         current += size;
378     }
379     padding = FFMIN(128 - ((current - dxva_data) & 127), end - current);
380     if (slice && padding > 0) {
381         memset(current, 0, padding);
382         current += padding;
383
384         slice->SliceBytesInBuffer += padding;
385     }
386 #if CONFIG_D3D11VA
387     if (ff_dxva2_is_d3d11(avctx))
388         if (FAILED(ID3D11VideoContext_ReleaseDecoderBuffer(D3D11VA_CONTEXT(ctx)->video_context, D3D11VA_CONTEXT(ctx)->decoder, type)))
389             return -1;
390 #endif
391 #if CONFIG_DXVA2
392     if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD)
393         if (FAILED(IDirectXVideoDecoder_ReleaseBuffer(DXVA2_CONTEXT(ctx)->decoder, type)))
394             return -1;
395 #endif
396     if (i < ctx_pic->slice_count)
397         return -1;
398
399 #if CONFIG_D3D11VA
400     if (ff_dxva2_is_d3d11(avctx)) {
401         D3D11_VIDEO_DECODER_BUFFER_DESC *dsc11 = bs;
402         memset(dsc11, 0, sizeof(*dsc11));
403         dsc11->BufferType           = type;
404         dsc11->DataSize             = current - dxva_data;
405         dsc11->NumMBsInBuffer       = mb_count;
406
407         type = D3D11_VIDEO_DECODER_BUFFER_SLICE_CONTROL;
408     }
409 #endif
410 #if CONFIG_DXVA2
411     if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD) {
412         DXVA2_DecodeBufferDesc *dsc2 = bs;
413         memset(dsc2, 0, sizeof(*dsc2));
414         dsc2->CompressedBufferType = type;
415         dsc2->DataSize             = current - dxva_data;
416         dsc2->NumMBsInBuffer       = mb_count;
417
418         type = DXVA2_SliceControlBufferType;
419     }
420 #endif
421
422     if (is_slice_short(avctx, ctx)) {
423         slice_data = ctx_pic->slice_short;
424         slice_size = ctx_pic->slice_count * sizeof(*ctx_pic->slice_short);
425     } else {
426         slice_data = ctx_pic->slice_long;
427         slice_size = ctx_pic->slice_count * sizeof(*ctx_pic->slice_long);
428     }
429     assert((bs->DataSize & 127) == 0);
430     return ff_dxva2_commit_buffer(avctx, ctx, sc,
431                                   type,
432                                   slice_data, slice_size, mb_count);
433 }
434
435
436 static int dxva2_h264_start_frame(AVCodecContext *avctx,
437                                   av_unused const uint8_t *buffer,
438                                   av_unused uint32_t size)
439 {
440     const H264Context *h = avctx->priv_data;
441     AVDXVAContext *ctx = DXVA_CONTEXT(avctx);
442     struct dxva2_picture_context *ctx_pic = h->cur_pic_ptr->hwaccel_picture_private;
443
444     if (!DXVA_CONTEXT_VALID(avctx, ctx))
445         return -1;
446     assert(ctx_pic);
447
448     /* Fill up DXVA_PicParams_H264 */
449     fill_picture_parameters(avctx, ctx, h, &ctx_pic->pp);
450
451     /* Fill up DXVA_Qmatrix_H264 */
452     fill_scaling_lists(avctx, ctx, h, &ctx_pic->qm);
453
454     ctx_pic->slice_count    = 0;
455     ctx_pic->bitstream_size = 0;
456     ctx_pic->bitstream      = NULL;
457     return 0;
458 }
459
460 static int dxva2_h264_decode_slice(AVCodecContext *avctx,
461                                    const uint8_t *buffer,
462                                    uint32_t size)
463 {
464     const H264Context *h = avctx->priv_data;
465     const H264SliceContext *sl = &h->slice_ctx[0];
466     AVDXVAContext *ctx = DXVA_CONTEXT(avctx);
467     const H264Picture *current_picture = h->cur_pic_ptr;
468     struct dxva2_picture_context *ctx_pic = current_picture->hwaccel_picture_private;
469     unsigned position;
470
471     if (ctx_pic->slice_count >= MAX_SLICES)
472         return -1;
473
474     if (!ctx_pic->bitstream)
475         ctx_pic->bitstream = buffer;
476     ctx_pic->bitstream_size += size;
477
478     position = buffer - ctx_pic->bitstream;
479     if (is_slice_short(avctx, ctx))
480         fill_slice_short(&ctx_pic->slice_short[ctx_pic->slice_count],
481                          position, size);
482     else
483         fill_slice_long(avctx, &ctx_pic->slice_long[ctx_pic->slice_count],
484                         &ctx_pic->pp, position, size);
485     ctx_pic->slice_count++;
486
487     if (sl->slice_type != AV_PICTURE_TYPE_I && sl->slice_type != AV_PICTURE_TYPE_SI)
488         ctx_pic->pp.wBitFields &= ~(1 << 15); /* Set IntraPicFlag to 0 */
489     return 0;
490 }
491
492 static int dxva2_h264_end_frame(AVCodecContext *avctx)
493 {
494     H264Context *h = avctx->priv_data;
495     H264SliceContext *sl = &h->slice_ctx[0];
496     struct dxva2_picture_context *ctx_pic =
497         h->cur_pic_ptr->hwaccel_picture_private;
498     int ret;
499
500     if (ctx_pic->slice_count <= 0 || ctx_pic->bitstream_size <= 0)
501         return -1;
502     ret = ff_dxva2_common_end_frame(avctx, h->cur_pic_ptr->f,
503                                     &ctx_pic->pp, sizeof(ctx_pic->pp),
504                                     &ctx_pic->qm, sizeof(ctx_pic->qm),
505                                     commit_bitstream_and_slice_buffer);
506     if (!ret)
507         ff_h264_draw_horiz_band(h, sl, 0, h->avctx->height);
508     return ret;
509 }
510
511 #if CONFIG_H264_DXVA2_HWACCEL
512 const AVHWAccel ff_h264_dxva2_hwaccel = {
513     .name           = "h264_dxva2",
514     .type           = AVMEDIA_TYPE_VIDEO,
515     .id             = AV_CODEC_ID_H264,
516     .pix_fmt        = AV_PIX_FMT_DXVA2_VLD,
517     .init           = ff_dxva2_decode_init,
518     .uninit         = ff_dxva2_decode_uninit,
519     .start_frame    = dxva2_h264_start_frame,
520     .decode_slice   = dxva2_h264_decode_slice,
521     .end_frame      = dxva2_h264_end_frame,
522     .frame_params   = ff_dxva2_common_frame_params,
523     .frame_priv_data_size = sizeof(struct dxva2_picture_context),
524     .priv_data_size = sizeof(FFDXVASharedContext),
525 };
526 #endif
527
528 #if CONFIG_H264_D3D11VA_HWACCEL
529 const AVHWAccel ff_h264_d3d11va_hwaccel = {
530     .name           = "h264_d3d11va",
531     .type           = AVMEDIA_TYPE_VIDEO,
532     .id             = AV_CODEC_ID_H264,
533     .pix_fmt        = AV_PIX_FMT_D3D11VA_VLD,
534     .init           = ff_dxva2_decode_init,
535     .uninit         = ff_dxva2_decode_uninit,
536     .start_frame    = dxva2_h264_start_frame,
537     .decode_slice   = dxva2_h264_decode_slice,
538     .end_frame      = dxva2_h264_end_frame,
539     .frame_params   = ff_dxva2_common_frame_params,
540     .frame_priv_data_size = sizeof(struct dxva2_picture_context),
541     .priv_data_size = sizeof(FFDXVASharedContext),
542 };
543 #endif
544
545 #if CONFIG_H264_D3D11VA2_HWACCEL
546 const AVHWAccel ff_h264_d3d11va2_hwaccel = {
547     .name           = "h264_d3d11va2",
548     .type           = AVMEDIA_TYPE_VIDEO,
549     .id             = AV_CODEC_ID_H264,
550     .pix_fmt        = AV_PIX_FMT_D3D11,
551     .init           = ff_dxva2_decode_init,
552     .uninit         = ff_dxva2_decode_uninit,
553     .start_frame    = dxva2_h264_start_frame,
554     .decode_slice   = dxva2_h264_decode_slice,
555     .end_frame      = dxva2_h264_end_frame,
556     .frame_params   = ff_dxva2_common_frame_params,
557     .frame_priv_data_size = sizeof(struct dxva2_picture_context),
558     .priv_data_size = sizeof(FFDXVASharedContext),
559 };
560 #endif