dxva2_h264: set the correct ref frame index in the long slice struct
[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 Picture *current_picture = h->cur_pic_ptr;
48     int i, j;
49
50     memset(pp, 0, sizeof(*pp));
51     /* Configure current picture */
52     fill_picture_entry(&pp->CurrPic,
53                        ff_dxva2_get_surface_index(ctx, current_picture),
54                        h->picture_structure == PICT_BOTTOM_FIELD);
55     /* Configure the set of references */
56     pp->UsedForReferenceFlags  = 0;
57     pp->NonExistingFrameFlags  = 0;
58     for (i = 0, j = 0; i < FF_ARRAY_ELEMS(pp->RefFrameList); i++) {
59         const Picture *r;
60         if (j < h->short_ref_count) {
61             r = h->short_ref[j++];
62         } else {
63             r = NULL;
64             while (!r && j < h->short_ref_count + 16)
65                 r = h->long_ref[j++ - h->short_ref_count];
66         }
67         if (r) {
68             fill_picture_entry(&pp->RefFrameList[i],
69                                ff_dxva2_get_surface_index(ctx, r),
70                                r->long_ref != 0);
71
72             if ((r->reference & PICT_TOP_FIELD) && r->field_poc[0] != INT_MAX)
73                 pp->FieldOrderCntList[i][0] = r->field_poc[0];
74             if ((r->reference & PICT_BOTTOM_FIELD) && r->field_poc[1] != INT_MAX)
75                 pp->FieldOrderCntList[i][1] = r->field_poc[1];
76
77             pp->FrameNumList[i] = r->long_ref ? r->pic_id : r->frame_num;
78             if (r->reference & PICT_TOP_FIELD)
79                 pp->UsedForReferenceFlags |= 1 << (2*i + 0);
80             if (r->reference & PICT_BOTTOM_FIELD)
81                 pp->UsedForReferenceFlags |= 1 << (2*i + 1);
82         } else {
83             pp->RefFrameList[i].bPicEntry = 0xff;
84             pp->FieldOrderCntList[i][0]   = 0;
85             pp->FieldOrderCntList[i][1]   = 0;
86             pp->FrameNumList[i]           = 0;
87         }
88     }
89
90     pp->wFrameWidthInMbsMinus1        = h->mb_width  - 1;
91     pp->wFrameHeightInMbsMinus1       = h->mb_height - 1;
92     pp->num_ref_frames                = h->sps.ref_frame_count;
93
94     pp->wBitFields                    = ((h->picture_structure != PICT_FRAME) <<  0) |
95                                         ((h->sps.mb_aff &&
96                                         (h->picture_structure == PICT_FRAME)) <<  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 dxva2_h264_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 ((h->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 ((h->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 (i = 0; i < 64; i++) {
166             qm->bScalingLists8x8[0][i] = h->pps.scaling_matrix8[0][i];
167             qm->bScalingLists8x8[1][i] = h->pps.scaling_matrix8[3][i];
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 (i = 0; i < 64; i++) {
175             qm->bScalingLists8x8[0][i] = h->pps.scaling_matrix8[0][ff_zigzag_direct[i]];
176             qm->bScalingLists8x8[1][i] = h->pps.scaling_matrix8[3][ff_zigzag_direct[i]];
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 int get_refpic_index(const DXVA_PicParams_H264 *pp, int surface_index)
198 {
199     int i;
200     for (i = 0; i < FF_ARRAY_ELEMS(pp->RefFrameList); i++) {
201         if ((pp->RefFrameList[i].bPicEntry & 0x7f) == surface_index)
202           return i;
203     }
204     return 0x7f;
205 }
206
207 static void fill_slice_long(AVCodecContext *avctx, DXVA_Slice_H264_Long *slice,
208                             const DXVA_PicParams_H264 *pp, unsigned position, unsigned size)
209 {
210     const H264Context *h = avctx->priv_data;
211     struct dxva_context *ctx = avctx->hwaccel_context;
212     unsigned list;
213
214     memset(slice, 0, sizeof(*slice));
215     slice->BSNALunitDataLocation = position;
216     slice->SliceBytesInBuffer    = size;
217     slice->wBadSliceChopping     = 0;
218
219     slice->first_mb_in_slice     = (h->mb_y >> FIELD_OR_MBAFF_PICTURE(h)) * h->mb_width + h->mb_x;
220     slice->NumMbsForSlice        = 0; /* XXX it is set once we have all slices */
221     slice->BitOffsetToSliceData  = get_bits_count(&h->gb);
222     slice->slice_type            = ff_h264_get_slice_type(h);
223     if (h->slice_type_fixed)
224         slice->slice_type += 5;
225     slice->luma_log2_weight_denom       = h->luma_log2_weight_denom;
226     slice->chroma_log2_weight_denom     = h->chroma_log2_weight_denom;
227     if (h->list_count > 0)
228         slice->num_ref_idx_l0_active_minus1 = h->ref_count[0] - 1;
229     if (h->list_count > 1)
230         slice->num_ref_idx_l1_active_minus1 = h->ref_count[1] - 1;
231     slice->slice_alpha_c0_offset_div2   = h->slice_alpha_c0_offset / 2;
232     slice->slice_beta_offset_div2       = h->slice_beta_offset     / 2;
233     slice->Reserved8Bits                = 0;
234
235     for (list = 0; list < 2; list++) {
236         unsigned i;
237         for (i = 0; i < FF_ARRAY_ELEMS(slice->RefPicList[list]); i++) {
238             if (list < h->list_count && i < h->ref_count[list]) {
239                 const Picture *r = &h->ref_list[list][i];
240                 unsigned plane;
241                 unsigned index = get_refpic_index(pp, ff_dxva2_get_surface_index(ctx, r));
242                 fill_picture_entry(&slice->RefPicList[list][i], index,
243                                    r->reference == PICT_BOTTOM_FIELD);
244                 for (plane = 0; plane < 3; plane++) {
245                     int w, o;
246                     if (plane == 0 && h->luma_weight_flag[list]) {
247                         w = h->luma_weight[i][list][0];
248                         o = h->luma_weight[i][list][1];
249                     } else if (plane >= 1 && h->chroma_weight_flag[list]) {
250                         w = h->chroma_weight[i][list][plane-1][0];
251                         o = h->chroma_weight[i][list][plane-1][1];
252                     } else {
253                         w = 1 << (plane == 0 ? h->luma_log2_weight_denom :
254                                                h->chroma_log2_weight_denom);
255                         o = 0;
256                     }
257                     slice->Weights[list][i][plane][0] = w;
258                     slice->Weights[list][i][plane][1] = o;
259                 }
260             } else {
261                 unsigned plane;
262                 slice->RefPicList[list][i].bPicEntry = 0xff;
263                 for (plane = 0; plane < 3; plane++) {
264                     slice->Weights[list][i][plane][0] = 0;
265                     slice->Weights[list][i][plane][1] = 0;
266                 }
267             }
268         }
269     }
270     slice->slice_qs_delta    = 0; /* XXX not implemented by FFmpeg */
271     slice->slice_qp_delta    = h->qscale - h->pps.init_qp;
272     slice->redundant_pic_cnt = h->redundant_pic_count;
273     if (h->slice_type == AV_PICTURE_TYPE_B)
274         slice->direct_spatial_mv_pred_flag = h->direct_spatial_mv_pred;
275     slice->cabac_init_idc = h->pps.cabac ? h->cabac_init_idc : 0;
276     if (h->deblocking_filter < 2)
277         slice->disable_deblocking_filter_idc = 1 - h->deblocking_filter;
278     else
279         slice->disable_deblocking_filter_idc = h->deblocking_filter;
280     slice->slice_id = h->current_slice - 1;
281 }
282
283 static int commit_bitstream_and_slice_buffer(AVCodecContext *avctx,
284                                              DXVA2_DecodeBufferDesc *bs,
285                                              DXVA2_DecodeBufferDesc *sc)
286 {
287     const H264Context *h = avctx->priv_data;
288     const unsigned mb_count = h->mb_width * h->mb_height;
289     struct dxva_context *ctx = avctx->hwaccel_context;
290     const Picture *current_picture = h->cur_pic_ptr;
291     struct dxva2_picture_context *ctx_pic = current_picture->hwaccel_picture_private;
292     DXVA_Slice_H264_Short *slice = NULL;
293     uint8_t  *dxva_data, *current, *end;
294     unsigned dxva_size;
295     void     *slice_data;
296     unsigned slice_size;
297     unsigned padding;
298     unsigned i;
299
300     /* Create an annex B bitstream buffer with only slice NAL and finalize slice */
301     if (FAILED(IDirectXVideoDecoder_GetBuffer(ctx->decoder,
302                                                DXVA2_BitStreamDateBufferType,
303                                                (void **)&dxva_data, &dxva_size)))
304         return -1;
305     current = dxva_data;
306     end = dxva_data + dxva_size;
307
308     for (i = 0; i < ctx_pic->slice_count; i++) {
309         static const uint8_t start_code[] = { 0, 0, 1 };
310         static const unsigned start_code_size = sizeof(start_code);
311         unsigned position, size;
312
313         assert(offsetof(DXVA_Slice_H264_Short, BSNALunitDataLocation) ==
314                offsetof(DXVA_Slice_H264_Long,  BSNALunitDataLocation));
315         assert(offsetof(DXVA_Slice_H264_Short, SliceBytesInBuffer) ==
316                offsetof(DXVA_Slice_H264_Long,  SliceBytesInBuffer));
317
318         if (is_slice_short(ctx))
319             slice = &ctx_pic->slice_short[i];
320         else
321             slice = (DXVA_Slice_H264_Short*)&ctx_pic->slice_long[i];
322
323         position = slice->BSNALunitDataLocation;
324         size     = slice->SliceBytesInBuffer;
325         if (start_code_size + size > end - current) {
326             av_log(avctx, AV_LOG_ERROR, "Failed to build bitstream");
327             break;
328         }
329
330         slice->BSNALunitDataLocation = current - dxva_data;
331         slice->SliceBytesInBuffer    = start_code_size + size;
332
333         if (!is_slice_short(ctx)) {
334             DXVA_Slice_H264_Long *slice_long = (DXVA_Slice_H264_Long*)slice;
335             if (i < ctx_pic->slice_count - 1)
336                 slice_long->NumMbsForSlice =
337                     slice_long[1].first_mb_in_slice - slice_long[0].first_mb_in_slice;
338             else
339                 slice_long->NumMbsForSlice = mb_count - slice_long->first_mb_in_slice;
340         }
341
342         memcpy(current, start_code, start_code_size);
343         current += start_code_size;
344
345         memcpy(current, &ctx_pic->bitstream[position], size);
346         current += size;
347     }
348     padding = FFMIN(128 - ((current - dxva_data) & 127), end - current);
349     if (slice && padding > 0) {
350         memset(current, 0, padding);
351         current += padding;
352
353         slice->SliceBytesInBuffer += padding;
354     }
355     if (FAILED(IDirectXVideoDecoder_ReleaseBuffer(ctx->decoder,
356                                                   DXVA2_BitStreamDateBufferType)))
357         return -1;
358     if (i < ctx_pic->slice_count)
359         return -1;
360
361     memset(bs, 0, sizeof(*bs));
362     bs->CompressedBufferType = DXVA2_BitStreamDateBufferType;
363     bs->DataSize             = current - dxva_data;
364     bs->NumMBsInBuffer       = mb_count;
365
366     if (is_slice_short(ctx)) {
367         slice_data = ctx_pic->slice_short;
368         slice_size = ctx_pic->slice_count * sizeof(*ctx_pic->slice_short);
369     } else {
370         slice_data = ctx_pic->slice_long;
371         slice_size = ctx_pic->slice_count * sizeof(*ctx_pic->slice_long);
372     }
373     assert((bs->DataSize & 127) == 0);
374     return ff_dxva2_commit_buffer(avctx, ctx, sc,
375                                   DXVA2_SliceControlBufferType,
376                                   slice_data, slice_size, mb_count);
377 }
378
379
380 static int dxva2_h264_start_frame(AVCodecContext *avctx,
381                                   av_unused const uint8_t *buffer,
382                                   av_unused uint32_t size)
383 {
384     const H264Context *h = avctx->priv_data;
385     struct dxva_context *ctx = avctx->hwaccel_context;
386     struct dxva2_picture_context *ctx_pic = h->cur_pic_ptr->hwaccel_picture_private;
387
388     if (!ctx->decoder || !ctx->cfg || ctx->surface_count <= 0)
389         return -1;
390     assert(ctx_pic);
391
392     /* Fill up DXVA_PicParams_H264 */
393     fill_picture_parameters(ctx, h, &ctx_pic->pp);
394
395     /* Fill up DXVA_Qmatrix_H264 */
396     fill_scaling_lists(ctx, h, &ctx_pic->qm);
397
398     ctx_pic->slice_count    = 0;
399     ctx_pic->bitstream_size = 0;
400     ctx_pic->bitstream      = NULL;
401     return 0;
402 }
403
404 static int dxva2_h264_decode_slice(AVCodecContext *avctx,
405                                    const uint8_t *buffer,
406                                    uint32_t size)
407 {
408     const H264Context *h = avctx->priv_data;
409     struct dxva_context *ctx = avctx->hwaccel_context;
410     const Picture *current_picture = h->cur_pic_ptr;
411     struct dxva2_picture_context *ctx_pic = current_picture->hwaccel_picture_private;
412     unsigned position;
413
414     if (ctx_pic->slice_count >= MAX_SLICES)
415         return -1;
416
417     if (!ctx_pic->bitstream)
418         ctx_pic->bitstream = buffer;
419     ctx_pic->bitstream_size += size;
420
421     position = buffer - ctx_pic->bitstream;
422     if (is_slice_short(ctx))
423         fill_slice_short(&ctx_pic->slice_short[ctx_pic->slice_count],
424                          position, size);
425     else
426         fill_slice_long(avctx, &ctx_pic->slice_long[ctx_pic->slice_count],
427                         &ctx_pic->pp, position, size);
428     ctx_pic->slice_count++;
429
430     if (h->slice_type != AV_PICTURE_TYPE_I && h->slice_type != AV_PICTURE_TYPE_SI)
431         ctx_pic->pp.wBitFields &= ~(1 << 15); /* Set IntraPicFlag to 0 */
432     return 0;
433 }
434
435 static int dxva2_h264_end_frame(AVCodecContext *avctx)
436 {
437     H264Context *h = avctx->priv_data;
438     struct dxva2_picture_context *ctx_pic =
439         h->cur_pic_ptr->hwaccel_picture_private;
440     int ret;
441
442     if (ctx_pic->slice_count <= 0 || ctx_pic->bitstream_size <= 0)
443         return -1;
444     ret = ff_dxva2_common_end_frame(avctx, h->cur_pic_ptr,
445                                     &ctx_pic->pp, sizeof(ctx_pic->pp),
446                                     &ctx_pic->qm, sizeof(ctx_pic->qm),
447                                     commit_bitstream_and_slice_buffer);
448     if (!ret)
449         ff_h264_draw_horiz_band(h, 0, h->avctx->height);
450     return ret;
451 }
452
453 AVHWAccel ff_h264_dxva2_hwaccel = {
454     .name           = "h264_dxva2",
455     .type           = AVMEDIA_TYPE_VIDEO,
456     .id             = AV_CODEC_ID_H264,
457     .pix_fmt        = AV_PIX_FMT_DXVA2_VLD,
458     .start_frame    = dxva2_h264_start_frame,
459     .decode_slice   = dxva2_h264_decode_slice,
460     .end_frame      = dxva2_h264_end_frame,
461     .priv_data_size = sizeof(struct dxva2_picture_context),
462 };