Fix dxva2 decoding for some H264 samples.
[ffmpeg.git] / libavcodec / dxva2_h264.c
1 /*
2  * DXVA2 H264 HW acceleration.
3  *
4  * copyright (c) 2009 Laurent Aimar
5  *
6  * This file is part of FFmpeg.
7  *
8  * FFmpeg 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  * FFmpeg 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 FFmpeg; 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 "h264.h"
25 #include "h264data.h"
26
27 struct dxva2_picture_context {
28     DXVA_PicParams_H264   pp;
29     DXVA_Qmatrix_H264     qm;
30     unsigned              slice_count;
31     DXVA_Slice_H264_Short slice_short[MAX_SLICES];
32     DXVA_Slice_H264_Long  slice_long[MAX_SLICES];
33     const uint8_t         *bitstream;
34     unsigned              bitstream_size;
35 };
36
37 static void fill_picture_entry(DXVA_PicEntry_H264 *pic,
38                                unsigned index, unsigned flag)
39 {
40     assert((index&0x7f) == index && (flag&0x01) == flag);
41     pic->bPicEntry = index | (flag << 7);
42 }
43
44 static void fill_picture_parameters(struct dxva_context *ctx, const H264Context *h,
45                                     DXVA_PicParams_H264 *pp)
46 {
47     const MpegEncContext *s = &h->s;
48     const Picture *current_picture = s->current_picture_ptr;
49     int i, j;
50
51     memset(pp, 0, sizeof(*pp));
52     /* Configure current picture */
53     fill_picture_entry(&pp->CurrPic,
54                        ff_dxva2_get_surface_index(ctx, current_picture),
55                        s->picture_structure == PICT_BOTTOM_FIELD);
56     /* Configure the set of references */
57     pp->UsedForReferenceFlags  = 0;
58     pp->NonExistingFrameFlags  = 0;
59     for (i = 0, j = 0; i < FF_ARRAY_ELEMS(pp->RefFrameList); i++) {
60         const Picture *r;
61         if (j < h->short_ref_count) {
62             r = h->short_ref[j++];
63         } else {
64             r = NULL;
65             while (!r && j < h->short_ref_count + 16)
66                 r = h->long_ref[j++ - h->short_ref_count];
67         }
68         if (r) {
69             fill_picture_entry(&pp->RefFrameList[i],
70                                ff_dxva2_get_surface_index(ctx, r),
71                                r->long_ref != 0);
72
73             if ((r->f.reference & PICT_TOP_FIELD) && r->field_poc[0] != INT_MAX)
74                 pp->FieldOrderCntList[i][0] = r->field_poc[0];
75             if ((r->f.reference & PICT_BOTTOM_FIELD) && r->field_poc[1] != INT_MAX)
76                 pp->FieldOrderCntList[i][1] = r->field_poc[1];
77
78             pp->FrameNumList[i] = r->long_ref ? r->pic_id : r->frame_num;
79             if (r->f.reference & PICT_TOP_FIELD)
80                 pp->UsedForReferenceFlags |= 1 << (2*i + 0);
81             if (r->f.reference & PICT_BOTTOM_FIELD)
82                 pp->UsedForReferenceFlags |= 1 << (2*i + 1);
83         } else {
84             pp->RefFrameList[i].bPicEntry = 0xff;
85             pp->FieldOrderCntList[i][0]   = 0;
86             pp->FieldOrderCntList[i][1]   = 0;
87             pp->FrameNumList[i]           = 0;
88         }
89     }
90
91     pp->wFrameWidthInMbsMinus1        = s->mb_width  - 1;
92     pp->wFrameHeightInMbsMinus1       = s->mb_height - 1;
93     pp->num_ref_frames                = h->sps.ref_frame_count;
94
95     pp->wBitFields                    = ((s->picture_structure != PICT_FRAME) <<  0) |
96                                         (h->sps.mb_aff                        <<  1) |
97                                         (h->sps.residual_color_transform_flag <<  2) |
98                                         /* sp_for_switch_flag (not implemented by FFmpeg) */
99                                         (0                                    <<  3) |
100                                         (h->sps.chroma_format_idc             <<  4) |
101                                         ((h->nal_ref_idc != 0)                <<  6) |
102                                         (h->pps.constrained_intra_pred        <<  7) |
103                                         (h->pps.weighted_pred                 <<  8) |
104                                         (h->pps.weighted_bipred_idc           <<  9) |
105                                         /* MbsConsecutiveFlag */
106                                         (1                                    << 11) |
107                                         (h->sps.frame_mbs_only_flag           << 12) |
108                                         (h->pps.transform_8x8_mode            << 13) |
109                                         ((h->sps.level_idc >= 31)             << 14) |
110                                         /* IntraPicFlag (Modified if we detect a non
111                                          * intra slice in decode_slice) */
112                                         (1                                    << 15);
113
114     pp->bit_depth_luma_minus8         = h->sps.bit_depth_luma - 8;
115     pp->bit_depth_chroma_minus8       = h->sps.bit_depth_chroma - 8;
116     if (ctx->workaround & FF_DXVA2_WORKAROUND_SCALING_LIST_ZIGZAG)
117         pp->Reserved16Bits            = 0;
118     else
119         pp->Reserved16Bits            = 3; /* FIXME is there a way to detect the right mode ? */
120     pp->StatusReportFeedbackNumber    = 1 + ctx->report_id++;
121     pp->CurrFieldOrderCnt[0] = 0;
122     if ((s->picture_structure & PICT_TOP_FIELD) &&
123         current_picture->field_poc[0] != INT_MAX)
124         pp->CurrFieldOrderCnt[0] = current_picture->field_poc[0];
125     pp->CurrFieldOrderCnt[1] = 0;
126     if ((s->picture_structure & PICT_BOTTOM_FIELD) &&
127         current_picture->field_poc[1] != INT_MAX)
128         pp->CurrFieldOrderCnt[1] = current_picture->field_poc[1];
129     pp->pic_init_qs_minus26           = h->pps.init_qs - 26;
130     pp->chroma_qp_index_offset        = h->pps.chroma_qp_index_offset[0];
131     pp->second_chroma_qp_index_offset = h->pps.chroma_qp_index_offset[1];
132     pp->ContinuationFlag              = 1;
133     pp->pic_init_qp_minus26           = h->pps.init_qp - 26;
134     pp->num_ref_idx_l0_active_minus1  = h->pps.ref_count[0] - 1;
135     pp->num_ref_idx_l1_active_minus1  = h->pps.ref_count[1] - 1;
136     pp->Reserved8BitsA                = 0;
137     pp->frame_num                     = h->frame_num;
138     pp->log2_max_frame_num_minus4     = h->sps.log2_max_frame_num - 4;
139     pp->pic_order_cnt_type            = h->sps.poc_type;
140     if (h->sps.poc_type == 0)
141         pp->log2_max_pic_order_cnt_lsb_minus4 = h->sps.log2_max_poc_lsb - 4;
142     else if (h->sps.poc_type == 1)
143         pp->delta_pic_order_always_zero_flag = h->sps.delta_pic_order_always_zero_flag;
144     pp->direct_8x8_inference_flag     = h->sps.direct_8x8_inference_flag;
145     pp->entropy_coding_mode_flag      = h->pps.cabac;
146     pp->pic_order_present_flag        = h->pps.pic_order_present;
147     pp->num_slice_groups_minus1       = h->pps.slice_group_count - 1;
148     pp->slice_group_map_type          = h->pps.mb_slice_group_map_type;
149     pp->deblocking_filter_control_present_flag = h->pps.deblocking_filter_parameters_present;
150     pp->redundant_pic_cnt_present_flag= h->pps.redundant_pic_cnt_present;
151     pp->Reserved8BitsB                = 0;
152     pp->slice_group_change_rate_minus1= 0;  /* XXX not implemented by FFmpeg */
153     //pp->SliceGroupMap[810];               /* XXX not implemented by FFmpeg */
154 }
155
156 static void fill_scaling_lists(struct dxva_context *ctx, const H264Context *h, DXVA_Qmatrix_H264 *qm)
157 {
158     unsigned i, j;
159     memset(qm, 0, sizeof(*qm));
160     if (ctx->workaround & FF_DXVA2_WORKAROUND_SCALING_LIST_ZIGZAG) {
161         for (i = 0; i < 6; i++)
162             for (j = 0; j < 16; j++)
163                 qm->bScalingLists4x4[i][j] = h->pps.scaling_matrix4[i][j];
164
165         for (j = 0; j < 64; j++) {
166             qm->bScalingLists8x8[0][j] = h->pps.scaling_matrix8[0][j];
167             qm->bScalingLists8x8[1][j] = h->pps.scaling_matrix8[3][j];
168         }
169     } else {
170         for (i = 0; i < 6; i++)
171             for (j = 0; j < 16; j++)
172                 qm->bScalingLists4x4[i][j] = h->pps.scaling_matrix4[i][zigzag_scan[j]];
173
174         for (j = 0; j < 64; j++) {
175             qm->bScalingLists8x8[0][j] = h->pps.scaling_matrix8[0][ff_zigzag_direct[j]];
176             qm->bScalingLists8x8[1][j] = h->pps.scaling_matrix8[3][ff_zigzag_direct[j]];
177         }
178     }
179 }
180
181 static int is_slice_short(struct dxva_context *ctx)
182 {
183     assert(ctx->cfg->ConfigBitstreamRaw == 1 ||
184            ctx->cfg->ConfigBitstreamRaw == 2);
185     return ctx->cfg->ConfigBitstreamRaw == 2;
186 }
187
188 static void fill_slice_short(DXVA_Slice_H264_Short *slice,
189                              unsigned position, unsigned size)
190 {
191     memset(slice, 0, sizeof(*slice));
192     slice->BSNALunitDataLocation = position;
193     slice->SliceBytesInBuffer    = size;
194     slice->wBadSliceChopping     = 0;
195 }
196
197 static void fill_slice_long(AVCodecContext *avctx, DXVA_Slice_H264_Long *slice,
198                             unsigned position, unsigned size)
199 {
200     const H264Context *h = avctx->priv_data;
201     struct dxva_context *ctx = avctx->hwaccel_context;
202     const MpegEncContext *s = &h->s;
203     unsigned list;
204
205     memset(slice, 0, sizeof(*slice));
206     slice->BSNALunitDataLocation = position;
207     slice->SliceBytesInBuffer    = size;
208     slice->wBadSliceChopping     = 0;
209
210     slice->first_mb_in_slice     = (s->mb_y >> FIELD_OR_MBAFF_PICTURE) * s->mb_width + s->mb_x;
211     slice->NumMbsForSlice        = 0; /* XXX it is set once we have all slices */
212     slice->BitOffsetToSliceData  = get_bits_count(&s->gb);
213     slice->slice_type            = ff_h264_get_slice_type(h);
214     if (h->slice_type_fixed)
215         slice->slice_type += 5;
216     slice->luma_log2_weight_denom       = h->luma_log2_weight_denom;
217     slice->chroma_log2_weight_denom     = h->chroma_log2_weight_denom;
218     if (h->list_count > 0)
219         slice->num_ref_idx_l0_active_minus1 = h->ref_count[0] - 1;
220     if (h->list_count > 1)
221         slice->num_ref_idx_l1_active_minus1 = h->ref_count[1] - 1;
222     slice->slice_alpha_c0_offset_div2   = h->slice_alpha_c0_offset / 2 - 26;
223     slice->slice_beta_offset_div2       = h->slice_beta_offset     / 2 - 26;
224     slice->Reserved8Bits                = 0;
225
226     for (list = 0; list < 2; list++) {
227         unsigned i;
228         for (i = 0; i < FF_ARRAY_ELEMS(slice->RefPicList[list]); i++) {
229             if (list < h->list_count && i < h->ref_count[list]) {
230                 const Picture *r = &h->ref_list[list][i];
231                 unsigned plane;
232                 fill_picture_entry(&slice->RefPicList[list][i],
233                                    ff_dxva2_get_surface_index(ctx, r),
234                                    r->f.reference == PICT_BOTTOM_FIELD);
235                 for (plane = 0; plane < 3; plane++) {
236                     int w, o;
237                     if (plane == 0 && h->luma_weight_flag[list]) {
238                         w = h->luma_weight[i][list][0];
239                         o = h->luma_weight[i][list][1];
240                     } else if (plane >= 1 && h->chroma_weight_flag[list]) {
241                         w = h->chroma_weight[i][list][plane-1][0];
242                         o = h->chroma_weight[i][list][plane-1][1];
243                     } else {
244                         w = 1 << (plane == 0 ? h->luma_log2_weight_denom :
245                                                h->chroma_log2_weight_denom);
246                         o = 0;
247                     }
248                     slice->Weights[list][i][plane][0] = w;
249                     slice->Weights[list][i][plane][1] = o;
250                 }
251             } else {
252                 unsigned plane;
253                 slice->RefPicList[list][i].bPicEntry = 0xff;
254                 for (plane = 0; plane < 3; plane++) {
255                     slice->Weights[list][i][plane][0] = 0;
256                     slice->Weights[list][i][plane][1] = 0;
257                 }
258             }
259         }
260     }
261     slice->slice_qs_delta    = 0; /* XXX not implemented by FFmpeg */
262     slice->slice_qp_delta    = s->qscale - h->pps.init_qp;
263     slice->redundant_pic_cnt = h->redundant_pic_count;
264     if (h->slice_type == AV_PICTURE_TYPE_B)
265         slice->direct_spatial_mv_pred_flag = h->direct_spatial_mv_pred;
266     slice->cabac_init_idc = h->pps.cabac ? h->cabac_init_idc : 0;
267     if (h->deblocking_filter < 2)
268         slice->disable_deblocking_filter_idc = 1 - h->deblocking_filter;
269     else
270         slice->disable_deblocking_filter_idc = h->deblocking_filter;
271     slice->slice_id = h->current_slice - 1;
272 }
273
274 static int commit_bitstream_and_slice_buffer(AVCodecContext *avctx,
275                                              DXVA2_DecodeBufferDesc *bs,
276                                              DXVA2_DecodeBufferDesc *sc)
277 {
278     const H264Context *h = avctx->priv_data;
279     const MpegEncContext *s = &h->s;
280     const unsigned mb_count = s->mb_width * s->mb_height;
281     struct dxva_context *ctx = avctx->hwaccel_context;
282     const Picture *current_picture = h->s.current_picture_ptr;
283     struct dxva2_picture_context *ctx_pic = current_picture->f.hwaccel_picture_private;
284     DXVA_Slice_H264_Short *slice = NULL;
285     uint8_t  *dxva_data, *current, *end;
286     unsigned dxva_size;
287     void     *slice_data;
288     unsigned slice_size;
289     unsigned padding;
290     unsigned i;
291
292     /* Create an annex B bitstream buffer with only slice NAL and finalize slice */
293     if (FAILED(IDirectXVideoDecoder_GetBuffer(ctx->decoder,
294                                                DXVA2_BitStreamDateBufferType,
295                                                &dxva_data, &dxva_size)))
296         return -1;
297     current = dxva_data;
298     end = dxva_data + dxva_size;
299
300     for (i = 0; i < ctx_pic->slice_count; i++) {
301         static const uint8_t start_code[] = { 0, 0, 1 };
302         static const unsigned start_code_size = sizeof(start_code);
303         unsigned position, size;
304
305         assert(offsetof(DXVA_Slice_H264_Short, BSNALunitDataLocation) ==
306                offsetof(DXVA_Slice_H264_Long,  BSNALunitDataLocation));
307         assert(offsetof(DXVA_Slice_H264_Short, SliceBytesInBuffer) ==
308                offsetof(DXVA_Slice_H264_Long,  SliceBytesInBuffer));
309
310         if (is_slice_short(ctx))
311             slice = &ctx_pic->slice_short[i];
312         else
313             slice = (DXVA_Slice_H264_Short*)&ctx_pic->slice_long[i];
314
315         position = slice->BSNALunitDataLocation;
316         size     = slice->SliceBytesInBuffer;
317         if (start_code_size + size > end - current) {
318             av_log(avctx, AV_LOG_ERROR, "Failed to build bitstream");
319             break;
320         }
321
322         slice->BSNALunitDataLocation = current - dxva_data;
323         slice->SliceBytesInBuffer    = start_code_size + size;
324
325         if (!is_slice_short(ctx)) {
326             DXVA_Slice_H264_Long *slice_long = (DXVA_Slice_H264_Long*)slice;
327             if (i < ctx_pic->slice_count - 1)
328                 slice_long->NumMbsForSlice =
329                     slice_long[1].first_mb_in_slice - slice_long[0].first_mb_in_slice;
330             else
331                 slice_long->NumMbsForSlice = mb_count - slice_long->first_mb_in_slice;
332         }
333
334         memcpy(current, start_code, start_code_size);
335         current += start_code_size;
336
337         memcpy(current, &ctx_pic->bitstream[position], size);
338         current += size;
339     }
340     padding = FFMIN(128 - ((current - dxva_data) & 127), end - current);
341     if (slice && padding > 0) {
342         memset(current, 0, padding);
343         current += padding;
344
345         slice->SliceBytesInBuffer += padding;
346     }
347     if (FAILED(IDirectXVideoDecoder_ReleaseBuffer(ctx->decoder,
348                                                   DXVA2_BitStreamDateBufferType)))
349         return -1;
350     if (i < ctx_pic->slice_count)
351         return -1;
352
353     memset(bs, 0, sizeof(*bs));
354     bs->CompressedBufferType = DXVA2_BitStreamDateBufferType;
355     bs->DataSize             = current - dxva_data;
356     bs->NumMBsInBuffer       = mb_count;
357
358     if (is_slice_short(ctx)) {
359         slice_data = ctx_pic->slice_short;
360         slice_size = ctx_pic->slice_count * sizeof(*ctx_pic->slice_short);
361     } else {
362         slice_data = ctx_pic->slice_long;
363         slice_size = ctx_pic->slice_count * sizeof(*ctx_pic->slice_long);
364     }
365     assert((bs->DataSize & 127) == 0);
366     return ff_dxva2_commit_buffer(avctx, ctx, sc,
367                                   DXVA2_SliceControlBufferType,
368                                   slice_data, slice_size, mb_count);
369 }
370
371
372 static int start_frame(AVCodecContext *avctx,
373                        av_unused const uint8_t *buffer,
374                        av_unused uint32_t size)
375 {
376     const H264Context *h = avctx->priv_data;
377     struct dxva_context *ctx = avctx->hwaccel_context;
378     struct dxva2_picture_context *ctx_pic = h->s.current_picture_ptr->f.hwaccel_picture_private;
379
380     if (!ctx->decoder || !ctx->cfg || ctx->surface_count <= 0)
381         return -1;
382     assert(ctx_pic);
383
384     /* Fill up DXVA_PicParams_H264 */
385     fill_picture_parameters(ctx, h, &ctx_pic->pp);
386
387     /* Fill up DXVA_Qmatrix_H264 */
388     fill_scaling_lists(ctx, h, &ctx_pic->qm);
389
390     ctx_pic->slice_count    = 0;
391     ctx_pic->bitstream_size = 0;
392     ctx_pic->bitstream      = NULL;
393     return 0;
394 }
395
396 static int decode_slice(AVCodecContext *avctx,
397                         const uint8_t *buffer, uint32_t size)
398 {
399     const H264Context *h = avctx->priv_data;
400     struct dxva_context *ctx = avctx->hwaccel_context;
401     const Picture *current_picture = h->s.current_picture_ptr;
402     struct dxva2_picture_context *ctx_pic = current_picture->f.hwaccel_picture_private;
403     unsigned position;
404
405     if (ctx_pic->slice_count >= MAX_SLICES)
406         return -1;
407
408     if (!ctx_pic->bitstream)
409         ctx_pic->bitstream = buffer;
410     ctx_pic->bitstream_size += size;
411
412     position = buffer - ctx_pic->bitstream;
413     if (is_slice_short(ctx))
414         fill_slice_short(&ctx_pic->slice_short[ctx_pic->slice_count],
415                          position, size);
416     else
417         fill_slice_long(avctx, &ctx_pic->slice_long[ctx_pic->slice_count],
418                         position, size);
419     ctx_pic->slice_count++;
420
421     if (h->slice_type != AV_PICTURE_TYPE_I && h->slice_type != AV_PICTURE_TYPE_SI)
422         ctx_pic->pp.wBitFields &= ~(1 << 15); /* Set IntraPicFlag to 0 */
423     return 0;
424 }
425
426 static int end_frame(AVCodecContext *avctx)
427 {
428     H264Context *h = avctx->priv_data;
429     MpegEncContext *s = &h->s;
430     struct dxva2_picture_context *ctx_pic =
431         h->s.current_picture_ptr->f.hwaccel_picture_private;
432
433     if (ctx_pic->slice_count <= 0 || ctx_pic->bitstream_size <= 0)
434         return -1;
435     return ff_dxva2_common_end_frame(avctx, s,
436                                      &ctx_pic->pp, sizeof(ctx_pic->pp),
437                                      &ctx_pic->qm, sizeof(ctx_pic->qm),
438                                      commit_bitstream_and_slice_buffer);
439 }
440
441 AVHWAccel ff_h264_dxva2_hwaccel = {
442     .name           = "h264_dxva2",
443     .type           = AVMEDIA_TYPE_VIDEO,
444     .id             = CODEC_ID_H264,
445     .pix_fmt        = PIX_FMT_DXVA2_VLD,
446     .capabilities   = 0,
447     .start_frame    = start_frame,
448     .decode_slice   = decode_slice,
449     .end_frame      = end_frame,
450     .priv_data_size = sizeof(struct dxva2_picture_context),
451 };
452