Merge commit '1173320249745eab01c901a39054fc0fced33c87'
[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 #include "mpegutils.h"
27
28 struct dxva2_picture_context {
29     DXVA_PicParams_H264   pp;
30     DXVA_Qmatrix_H264     qm;
31     unsigned              slice_count;
32     DXVA_Slice_H264_Short slice_short[MAX_SLICES];
33     DXVA_Slice_H264_Long  slice_long[MAX_SLICES];
34     const uint8_t         *bitstream;
35     unsigned              bitstream_size;
36 };
37
38 static void fill_picture_entry(DXVA_PicEntry_H264 *pic,
39                                unsigned index, unsigned flag)
40 {
41     assert((index&0x7f) == index && (flag&0x01) == flag);
42     pic->bPicEntry = index | (flag << 7);
43 }
44
45 static void fill_picture_parameters(struct dxva_context *ctx, const H264Context *h,
46                                     DXVA_PicParams_H264 *pp)
47 {
48     const H264Picture *current_picture = h->cur_pic_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->f),
55                        h->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 H264Picture *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->f),
71                                r->long_ref != 0);
72
73             if ((r->reference & PICT_TOP_FIELD) && r->field_poc[0] != INT_MAX)
74                 pp->FieldOrderCntList[i][0] = r->field_poc[0];
75             if ((r->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->reference & PICT_TOP_FIELD)
80                 pp->UsedForReferenceFlags |= 1 << (2*i + 0);
81             if (r->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        = h->mb_width  - 1;
92     pp->wFrameHeightInMbsMinus1       = h->mb_height - 1;
93     pp->num_ref_frames                = h->sps.ref_frame_count;
94
95     pp->wBitFields                    = ((h->picture_structure != PICT_FRAME) <<  0) |
96                                         ((h->sps.mb_aff &&
97                                         (h->picture_structure == PICT_FRAME)) <<  1) |
98                                         (h->sps.residual_color_transform_flag <<  2) |
99                                         /* sp_for_switch_flag (not implemented by FFmpeg) */
100                                         (0                                    <<  3) |
101                                         (h->sps.chroma_format_idc             <<  4) |
102                                         ((h->nal_ref_idc != 0)                <<  6) |
103                                         (h->pps.constrained_intra_pred        <<  7) |
104                                         (h->pps.weighted_pred                 <<  8) |
105                                         (h->pps.weighted_bipred_idc           <<  9) |
106                                         /* MbsConsecutiveFlag */
107                                         (1                                    << 11) |
108                                         (h->sps.frame_mbs_only_flag           << 12) |
109                                         (h->pps.transform_8x8_mode            << 13) |
110                                         ((h->sps.level_idc >= 31)             << 14) |
111                                         /* IntraPicFlag (Modified if we detect a non
112                                          * intra slice in dxva2_h264_decode_slice) */
113                                         (1                                    << 15);
114
115     pp->bit_depth_luma_minus8         = h->sps.bit_depth_luma - 8;
116     pp->bit_depth_chroma_minus8       = h->sps.bit_depth_chroma - 8;
117     if (ctx->workaround & FF_DXVA2_WORKAROUND_SCALING_LIST_ZIGZAG)
118         pp->Reserved16Bits            = 0;
119     else if (ctx->workaround & FF_DXVA2_WORKAROUND_INTEL_CLEARVIDEO)
120         pp->Reserved16Bits            = 0x34c;
121     else
122         pp->Reserved16Bits            = 3; /* FIXME is there a way to detect the right mode ? */
123     pp->StatusReportFeedbackNumber    = 1 + ctx->report_id++;
124     pp->CurrFieldOrderCnt[0] = 0;
125     if ((h->picture_structure & PICT_TOP_FIELD) &&
126         current_picture->field_poc[0] != INT_MAX)
127         pp->CurrFieldOrderCnt[0] = current_picture->field_poc[0];
128     pp->CurrFieldOrderCnt[1] = 0;
129     if ((h->picture_structure & PICT_BOTTOM_FIELD) &&
130         current_picture->field_poc[1] != INT_MAX)
131         pp->CurrFieldOrderCnt[1] = current_picture->field_poc[1];
132     pp->pic_init_qs_minus26           = h->pps.init_qs - 26;
133     pp->chroma_qp_index_offset        = h->pps.chroma_qp_index_offset[0];
134     pp->second_chroma_qp_index_offset = h->pps.chroma_qp_index_offset[1];
135     pp->ContinuationFlag              = 1;
136     pp->pic_init_qp_minus26           = h->pps.init_qp - 26;
137     pp->num_ref_idx_l0_active_minus1  = h->pps.ref_count[0] - 1;
138     pp->num_ref_idx_l1_active_minus1  = h->pps.ref_count[1] - 1;
139     pp->Reserved8BitsA                = 0;
140     pp->frame_num                     = h->frame_num;
141     pp->log2_max_frame_num_minus4     = h->sps.log2_max_frame_num - 4;
142     pp->pic_order_cnt_type            = h->sps.poc_type;
143     if (h->sps.poc_type == 0)
144         pp->log2_max_pic_order_cnt_lsb_minus4 = h->sps.log2_max_poc_lsb - 4;
145     else if (h->sps.poc_type == 1)
146         pp->delta_pic_order_always_zero_flag = h->sps.delta_pic_order_always_zero_flag;
147     pp->direct_8x8_inference_flag     = h->sps.direct_8x8_inference_flag;
148     pp->entropy_coding_mode_flag      = h->pps.cabac;
149     pp->pic_order_present_flag        = h->pps.pic_order_present;
150     pp->num_slice_groups_minus1       = h->pps.slice_group_count - 1;
151     pp->slice_group_map_type          = h->pps.mb_slice_group_map_type;
152     pp->deblocking_filter_control_present_flag = h->pps.deblocking_filter_parameters_present;
153     pp->redundant_pic_cnt_present_flag= h->pps.redundant_pic_cnt_present;
154     pp->Reserved8BitsB                = 0;
155     pp->slice_group_change_rate_minus1= 0;  /* XXX not implemented by FFmpeg */
156     //pp->SliceGroupMap[810];               /* XXX not implemented by FFmpeg */
157 }
158
159 static void fill_scaling_lists(struct dxva_context *ctx, const H264Context *h, DXVA_Qmatrix_H264 *qm)
160 {
161     unsigned i, j;
162     memset(qm, 0, sizeof(*qm));
163     if (ctx->workaround & FF_DXVA2_WORKAROUND_SCALING_LIST_ZIGZAG) {
164         for (i = 0; i < 6; i++)
165             for (j = 0; j < 16; j++)
166                 qm->bScalingLists4x4[i][j] = h->pps.scaling_matrix4[i][j];
167
168         for (i = 0; i < 64; i++) {
169             qm->bScalingLists8x8[0][i] = h->pps.scaling_matrix8[0][i];
170             qm->bScalingLists8x8[1][i] = h->pps.scaling_matrix8[3][i];
171         }
172     } else {
173         for (i = 0; i < 6; i++)
174             for (j = 0; j < 16; j++)
175                 qm->bScalingLists4x4[i][j] = h->pps.scaling_matrix4[i][zigzag_scan[j]];
176
177         for (i = 0; i < 64; i++) {
178             qm->bScalingLists8x8[0][i] = h->pps.scaling_matrix8[0][ff_zigzag_direct[i]];
179             qm->bScalingLists8x8[1][i] = h->pps.scaling_matrix8[3][ff_zigzag_direct[i]];
180         }
181     }
182 }
183
184 static int is_slice_short(struct dxva_context *ctx)
185 {
186     assert(ctx->cfg->ConfigBitstreamRaw == 1 ||
187            ctx->cfg->ConfigBitstreamRaw == 2);
188     return ctx->cfg->ConfigBitstreamRaw == 2;
189 }
190
191 static void fill_slice_short(DXVA_Slice_H264_Short *slice,
192                              unsigned position, unsigned size)
193 {
194     memset(slice, 0, sizeof(*slice));
195     slice->BSNALunitDataLocation = position;
196     slice->SliceBytesInBuffer    = size;
197     slice->wBadSliceChopping     = 0;
198 }
199
200 static int get_refpic_index(const DXVA_PicParams_H264 *pp, int surface_index)
201 {
202     int i;
203     for (i = 0; i < FF_ARRAY_ELEMS(pp->RefFrameList); i++) {
204         if ((pp->RefFrameList[i].bPicEntry & 0x7f) == surface_index)
205           return i;
206     }
207     return 0x7f;
208 }
209
210 static void fill_slice_long(AVCodecContext *avctx, DXVA_Slice_H264_Long *slice,
211                             const DXVA_PicParams_H264 *pp, unsigned position, unsigned size)
212 {
213     const H264Context *h = avctx->priv_data;
214     struct dxva_context *ctx = avctx->hwaccel_context;
215     unsigned list;
216
217     memset(slice, 0, sizeof(*slice));
218     slice->BSNALunitDataLocation = position;
219     slice->SliceBytesInBuffer    = size;
220     slice->wBadSliceChopping     = 0;
221
222     slice->first_mb_in_slice     = (h->mb_y >> FIELD_OR_MBAFF_PICTURE(h)) * h->mb_width + h->mb_x;
223     slice->NumMbsForSlice        = 0; /* XXX it is set once we have all slices */
224     slice->BitOffsetToSliceData  = get_bits_count(&h->gb);
225     slice->slice_type            = ff_h264_get_slice_type(h);
226     if (h->slice_type_fixed)
227         slice->slice_type += 5;
228     slice->luma_log2_weight_denom       = h->luma_log2_weight_denom;
229     slice->chroma_log2_weight_denom     = h->chroma_log2_weight_denom;
230     if (h->list_count > 0)
231         slice->num_ref_idx_l0_active_minus1 = h->ref_count[0] - 1;
232     if (h->list_count > 1)
233         slice->num_ref_idx_l1_active_minus1 = h->ref_count[1] - 1;
234     slice->slice_alpha_c0_offset_div2   = h->slice_alpha_c0_offset / 2;
235     slice->slice_beta_offset_div2       = h->slice_beta_offset     / 2;
236     slice->Reserved8Bits                = 0;
237
238     for (list = 0; list < 2; list++) {
239         unsigned i;
240         for (i = 0; i < FF_ARRAY_ELEMS(slice->RefPicList[list]); i++) {
241             if (list < h->list_count && i < h->ref_count[list]) {
242                 const H264Picture *r = &h->ref_list[list][i];
243                 unsigned plane;
244                 unsigned index;
245                 if (ctx->workaround & FF_DXVA2_WORKAROUND_INTEL_CLEARVIDEO)
246                     index = ff_dxva2_get_surface_index(ctx, &r->f);
247                 else
248                     index = get_refpic_index(pp, ff_dxva2_get_surface_index(ctx, &r->f));
249                 fill_picture_entry(&slice->RefPicList[list][i], index,
250                                    r->reference == PICT_BOTTOM_FIELD);
251                 for (plane = 0; plane < 3; plane++) {
252                     int w, o;
253                     if (plane == 0 && h->luma_weight_flag[list]) {
254                         w = h->luma_weight[i][list][0];
255                         o = h->luma_weight[i][list][1];
256                     } else if (plane >= 1 && h->chroma_weight_flag[list]) {
257                         w = h->chroma_weight[i][list][plane-1][0];
258                         o = h->chroma_weight[i][list][plane-1][1];
259                     } else {
260                         w = 1 << (plane == 0 ? h->luma_log2_weight_denom :
261                                                h->chroma_log2_weight_denom);
262                         o = 0;
263                     }
264                     slice->Weights[list][i][plane][0] = w;
265                     slice->Weights[list][i][plane][1] = o;
266                 }
267             } else {
268                 unsigned plane;
269                 slice->RefPicList[list][i].bPicEntry = 0xff;
270                 for (plane = 0; plane < 3; plane++) {
271                     slice->Weights[list][i][plane][0] = 0;
272                     slice->Weights[list][i][plane][1] = 0;
273                 }
274             }
275         }
276     }
277     slice->slice_qs_delta    = 0; /* XXX not implemented by FFmpeg */
278     slice->slice_qp_delta    = h->qscale - h->pps.init_qp;
279     slice->redundant_pic_cnt = h->redundant_pic_count;
280     if (h->slice_type == AV_PICTURE_TYPE_B)
281         slice->direct_spatial_mv_pred_flag = h->direct_spatial_mv_pred;
282     slice->cabac_init_idc = h->pps.cabac ? h->cabac_init_idc : 0;
283     if (h->deblocking_filter < 2)
284         slice->disable_deblocking_filter_idc = 1 - h->deblocking_filter;
285     else
286         slice->disable_deblocking_filter_idc = h->deblocking_filter;
287     slice->slice_id = h->current_slice - 1;
288 }
289
290 static int commit_bitstream_and_slice_buffer(AVCodecContext *avctx,
291                                              DXVA2_DecodeBufferDesc *bs,
292                                              DXVA2_DecodeBufferDesc *sc)
293 {
294     const H264Context *h = avctx->priv_data;
295     const unsigned mb_count = h->mb_width * h->mb_height;
296     struct dxva_context *ctx = avctx->hwaccel_context;
297     const H264Picture *current_picture = h->cur_pic_ptr;
298     struct dxva2_picture_context *ctx_pic = current_picture->hwaccel_picture_private;
299     DXVA_Slice_H264_Short *slice = NULL;
300     uint8_t  *dxva_data, *current, *end;
301     unsigned dxva_size;
302     void     *slice_data;
303     unsigned slice_size;
304     unsigned padding;
305     unsigned i;
306
307     /* Create an annex B bitstream buffer with only slice NAL and finalize slice */
308     if (FAILED(IDirectXVideoDecoder_GetBuffer(ctx->decoder,
309                                                DXVA2_BitStreamDateBufferType,
310                                                (void **)&dxva_data, &dxva_size)))
311         return -1;
312     current = dxva_data;
313     end = dxva_data + dxva_size;
314
315     for (i = 0; i < ctx_pic->slice_count; i++) {
316         static const uint8_t start_code[] = { 0, 0, 1 };
317         static const unsigned start_code_size = sizeof(start_code);
318         unsigned position, size;
319
320         assert(offsetof(DXVA_Slice_H264_Short, BSNALunitDataLocation) ==
321                offsetof(DXVA_Slice_H264_Long,  BSNALunitDataLocation));
322         assert(offsetof(DXVA_Slice_H264_Short, SliceBytesInBuffer) ==
323                offsetof(DXVA_Slice_H264_Long,  SliceBytesInBuffer));
324
325         if (is_slice_short(ctx))
326             slice = &ctx_pic->slice_short[i];
327         else
328             slice = (DXVA_Slice_H264_Short*)&ctx_pic->slice_long[i];
329
330         position = slice->BSNALunitDataLocation;
331         size     = slice->SliceBytesInBuffer;
332         if (start_code_size + size > end - current) {
333             av_log(avctx, AV_LOG_ERROR, "Failed to build bitstream");
334             break;
335         }
336
337         slice->BSNALunitDataLocation = current - dxva_data;
338         slice->SliceBytesInBuffer    = start_code_size + size;
339
340         if (!is_slice_short(ctx)) {
341             DXVA_Slice_H264_Long *slice_long = (DXVA_Slice_H264_Long*)slice;
342             if (i < ctx_pic->slice_count - 1)
343                 slice_long->NumMbsForSlice =
344                     slice_long[1].first_mb_in_slice - slice_long[0].first_mb_in_slice;
345             else
346                 slice_long->NumMbsForSlice = mb_count - slice_long->first_mb_in_slice;
347         }
348
349         memcpy(current, start_code, start_code_size);
350         current += start_code_size;
351
352         memcpy(current, &ctx_pic->bitstream[position], size);
353         current += size;
354     }
355     padding = FFMIN(128 - ((current - dxva_data) & 127), end - current);
356     if (slice && padding > 0) {
357         memset(current, 0, padding);
358         current += padding;
359
360         slice->SliceBytesInBuffer += padding;
361     }
362     if (FAILED(IDirectXVideoDecoder_ReleaseBuffer(ctx->decoder,
363                                                   DXVA2_BitStreamDateBufferType)))
364         return -1;
365     if (i < ctx_pic->slice_count)
366         return -1;
367
368     memset(bs, 0, sizeof(*bs));
369     bs->CompressedBufferType = DXVA2_BitStreamDateBufferType;
370     bs->DataSize             = current - dxva_data;
371     bs->NumMBsInBuffer       = mb_count;
372
373     if (is_slice_short(ctx)) {
374         slice_data = ctx_pic->slice_short;
375         slice_size = ctx_pic->slice_count * sizeof(*ctx_pic->slice_short);
376     } else {
377         slice_data = ctx_pic->slice_long;
378         slice_size = ctx_pic->slice_count * sizeof(*ctx_pic->slice_long);
379     }
380     assert((bs->DataSize & 127) == 0);
381     return ff_dxva2_commit_buffer(avctx, ctx, sc,
382                                   DXVA2_SliceControlBufferType,
383                                   slice_data, slice_size, mb_count);
384 }
385
386
387 static int dxva2_h264_start_frame(AVCodecContext *avctx,
388                                   av_unused const uint8_t *buffer,
389                                   av_unused uint32_t size)
390 {
391     const H264Context *h = avctx->priv_data;
392     struct dxva_context *ctx = avctx->hwaccel_context;
393     struct dxva2_picture_context *ctx_pic = h->cur_pic_ptr->hwaccel_picture_private;
394
395     if (!ctx->decoder || !ctx->cfg || ctx->surface_count <= 0)
396         return -1;
397     assert(ctx_pic);
398
399     /* Fill up DXVA_PicParams_H264 */
400     fill_picture_parameters(ctx, h, &ctx_pic->pp);
401
402     /* Fill up DXVA_Qmatrix_H264 */
403     fill_scaling_lists(ctx, h, &ctx_pic->qm);
404
405     ctx_pic->slice_count    = 0;
406     ctx_pic->bitstream_size = 0;
407     ctx_pic->bitstream      = NULL;
408     return 0;
409 }
410
411 static int dxva2_h264_decode_slice(AVCodecContext *avctx,
412                                    const uint8_t *buffer,
413                                    uint32_t size)
414 {
415     const H264Context *h = avctx->priv_data;
416     struct dxva_context *ctx = avctx->hwaccel_context;
417     const H264Picture *current_picture = h->cur_pic_ptr;
418     struct dxva2_picture_context *ctx_pic = current_picture->hwaccel_picture_private;
419     unsigned position;
420
421     if (ctx_pic->slice_count >= MAX_SLICES)
422         return -1;
423
424     if (!ctx_pic->bitstream)
425         ctx_pic->bitstream = buffer;
426     ctx_pic->bitstream_size += size;
427
428     position = buffer - ctx_pic->bitstream;
429     if (is_slice_short(ctx))
430         fill_slice_short(&ctx_pic->slice_short[ctx_pic->slice_count],
431                          position, size);
432     else
433         fill_slice_long(avctx, &ctx_pic->slice_long[ctx_pic->slice_count],
434                         &ctx_pic->pp, position, size);
435     ctx_pic->slice_count++;
436
437     if (h->slice_type != AV_PICTURE_TYPE_I && h->slice_type != AV_PICTURE_TYPE_SI)
438         ctx_pic->pp.wBitFields &= ~(1 << 15); /* Set IntraPicFlag to 0 */
439     return 0;
440 }
441
442 static int dxva2_h264_end_frame(AVCodecContext *avctx)
443 {
444     H264Context *h = avctx->priv_data;
445     struct dxva2_picture_context *ctx_pic =
446         h->cur_pic_ptr->hwaccel_picture_private;
447     int ret;
448
449     if (ctx_pic->slice_count <= 0 || ctx_pic->bitstream_size <= 0)
450         return -1;
451     ret = ff_dxva2_common_end_frame(avctx, &h->cur_pic_ptr->f,
452                                     &ctx_pic->pp, sizeof(ctx_pic->pp),
453                                     &ctx_pic->qm, sizeof(ctx_pic->qm),
454                                     commit_bitstream_and_slice_buffer);
455     if (!ret)
456         ff_h264_draw_horiz_band(h, 0, h->avctx->height);
457     return ret;
458 }
459
460 AVHWAccel ff_h264_dxva2_hwaccel = {
461     .name           = "h264_dxva2",
462     .type           = AVMEDIA_TYPE_VIDEO,
463     .id             = AV_CODEC_ID_H264,
464     .pix_fmt        = AV_PIX_FMT_DXVA2_VLD,
465     .start_frame    = dxva2_h264_start_frame,
466     .decode_slice   = dxva2_h264_decode_slice,
467     .end_frame      = dxva2_h264_end_frame,
468     .frame_priv_data_size = sizeof(struct dxva2_picture_context),
469 };