Merge commit 'a0f2946068c62e18cb05ac25c0df3d86077251a6'
[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     H264SliceContext *sl = &h->slice_ctx[0];
215     struct dxva_context *ctx = avctx->hwaccel_context;
216     unsigned list;
217
218     memset(slice, 0, sizeof(*slice));
219     slice->BSNALunitDataLocation = position;
220     slice->SliceBytesInBuffer    = size;
221     slice->wBadSliceChopping     = 0;
222
223     slice->first_mb_in_slice     = (sl->mb_y >> FIELD_OR_MBAFF_PICTURE(h)) * h->mb_width + sl->mb_x;
224     slice->NumMbsForSlice        = 0; /* XXX it is set once we have all slices */
225     slice->BitOffsetToSliceData  = get_bits_count(&sl->gb);
226     slice->slice_type            = ff_h264_get_slice_type(sl);
227     if (sl->slice_type_fixed)
228         slice->slice_type += 5;
229     slice->luma_log2_weight_denom       = sl->luma_log2_weight_denom;
230     slice->chroma_log2_weight_denom     = sl->chroma_log2_weight_denom;
231     if (sl->list_count > 0)
232         slice->num_ref_idx_l0_active_minus1 = sl->ref_count[0] - 1;
233     if (sl->list_count > 1)
234         slice->num_ref_idx_l1_active_minus1 = sl->ref_count[1] - 1;
235     slice->slice_alpha_c0_offset_div2   = sl->slice_alpha_c0_offset / 2;
236     slice->slice_beta_offset_div2       = sl->slice_beta_offset     / 2;
237     slice->Reserved8Bits                = 0;
238
239     for (list = 0; list < 2; list++) {
240         unsigned i;
241         for (i = 0; i < FF_ARRAY_ELEMS(slice->RefPicList[list]); i++) {
242             if (list < sl->list_count && i < sl->ref_count[list]) {
243                 const H264Picture *r = sl->ref_list[list][i].parent;
244                 unsigned plane;
245                 unsigned index;
246                 if (ctx->workaround & FF_DXVA2_WORKAROUND_INTEL_CLEARVIDEO)
247                     index = ff_dxva2_get_surface_index(ctx, r->f);
248                 else
249                     index = get_refpic_index(pp, ff_dxva2_get_surface_index(ctx, r->f));
250                 fill_picture_entry(&slice->RefPicList[list][i], index,
251                                    r->reference == PICT_BOTTOM_FIELD);
252                 for (plane = 0; plane < 3; plane++) {
253                     int w, o;
254                     if (plane == 0 && sl->luma_weight_flag[list]) {
255                         w = sl->luma_weight[i][list][0];
256                         o = sl->luma_weight[i][list][1];
257                     } else if (plane >= 1 && sl->chroma_weight_flag[list]) {
258                         w = sl->chroma_weight[i][list][plane-1][0];
259                         o = sl->chroma_weight[i][list][plane-1][1];
260                     } else {
261                         w = 1 << (plane == 0 ? sl->luma_log2_weight_denom :
262                                                sl->chroma_log2_weight_denom);
263                         o = 0;
264                     }
265                     slice->Weights[list][i][plane][0] = w;
266                     slice->Weights[list][i][plane][1] = o;
267                 }
268             } else {
269                 unsigned plane;
270                 slice->RefPicList[list][i].bPicEntry = 0xff;
271                 for (plane = 0; plane < 3; plane++) {
272                     slice->Weights[list][i][plane][0] = 0;
273                     slice->Weights[list][i][plane][1] = 0;
274                 }
275             }
276         }
277     }
278     slice->slice_qs_delta    = 0; /* XXX not implemented by FFmpeg */
279     slice->slice_qp_delta    = sl->qscale - h->pps.init_qp;
280     slice->redundant_pic_cnt = sl->redundant_pic_count;
281     if (sl->slice_type == AV_PICTURE_TYPE_B)
282         slice->direct_spatial_mv_pred_flag = sl->direct_spatial_mv_pred;
283     slice->cabac_init_idc = h->pps.cabac ? sl->cabac_init_idc : 0;
284     if (sl->deblocking_filter < 2)
285         slice->disable_deblocking_filter_idc = 1 - sl->deblocking_filter;
286     else
287         slice->disable_deblocking_filter_idc = sl->deblocking_filter;
288     slice->slice_id = h->current_slice - 1;
289 }
290
291 static int commit_bitstream_and_slice_buffer(AVCodecContext *avctx,
292                                              DXVA2_DecodeBufferDesc *bs,
293                                              DXVA2_DecodeBufferDesc *sc)
294 {
295     const H264Context *h = avctx->priv_data;
296     const unsigned mb_count = h->mb_width * h->mb_height;
297     struct dxva_context *ctx = avctx->hwaccel_context;
298     const H264Picture *current_picture = h->cur_pic_ptr;
299     struct dxva2_picture_context *ctx_pic = current_picture->hwaccel_picture_private;
300     DXVA_Slice_H264_Short *slice = NULL;
301     void     *dxva_data_ptr;
302     uint8_t  *dxva_data, *current, *end;
303     unsigned dxva_size;
304     void     *slice_data;
305     unsigned slice_size;
306     unsigned padding;
307     unsigned i;
308
309     /* Create an annex B bitstream buffer with only slice NAL and finalize slice */
310     if (FAILED(IDirectXVideoDecoder_GetBuffer(ctx->decoder,
311                                               DXVA2_BitStreamDateBufferType,
312                                               &dxva_data_ptr, &dxva_size)))
313         return -1;
314
315     dxva_data = dxva_data_ptr;
316     current = dxva_data;
317     end = dxva_data + dxva_size;
318
319     for (i = 0; i < ctx_pic->slice_count; i++) {
320         static const uint8_t start_code[] = { 0, 0, 1 };
321         static const unsigned start_code_size = sizeof(start_code);
322         unsigned position, size;
323
324         assert(offsetof(DXVA_Slice_H264_Short, BSNALunitDataLocation) ==
325                offsetof(DXVA_Slice_H264_Long,  BSNALunitDataLocation));
326         assert(offsetof(DXVA_Slice_H264_Short, SliceBytesInBuffer) ==
327                offsetof(DXVA_Slice_H264_Long,  SliceBytesInBuffer));
328
329         if (is_slice_short(ctx))
330             slice = &ctx_pic->slice_short[i];
331         else
332             slice = (DXVA_Slice_H264_Short*)&ctx_pic->slice_long[i];
333
334         position = slice->BSNALunitDataLocation;
335         size     = slice->SliceBytesInBuffer;
336         if (start_code_size + size > end - current) {
337             av_log(avctx, AV_LOG_ERROR, "Failed to build bitstream");
338             break;
339         }
340
341         slice->BSNALunitDataLocation = current - dxva_data;
342         slice->SliceBytesInBuffer    = start_code_size + size;
343
344         if (!is_slice_short(ctx)) {
345             DXVA_Slice_H264_Long *slice_long = (DXVA_Slice_H264_Long*)slice;
346             if (i < ctx_pic->slice_count - 1)
347                 slice_long->NumMbsForSlice =
348                     slice_long[1].first_mb_in_slice - slice_long[0].first_mb_in_slice;
349             else
350                 slice_long->NumMbsForSlice = mb_count - slice_long->first_mb_in_slice;
351         }
352
353         memcpy(current, start_code, start_code_size);
354         current += start_code_size;
355
356         memcpy(current, &ctx_pic->bitstream[position], size);
357         current += size;
358     }
359     padding = FFMIN(128 - ((current - dxva_data) & 127), end - current);
360     if (slice && padding > 0) {
361         memset(current, 0, padding);
362         current += padding;
363
364         slice->SliceBytesInBuffer += padding;
365     }
366     if (FAILED(IDirectXVideoDecoder_ReleaseBuffer(ctx->decoder,
367                                                   DXVA2_BitStreamDateBufferType)))
368         return -1;
369     if (i < ctx_pic->slice_count)
370         return -1;
371
372     memset(bs, 0, sizeof(*bs));
373     bs->CompressedBufferType = DXVA2_BitStreamDateBufferType;
374     bs->DataSize             = current - dxva_data;
375     bs->NumMBsInBuffer       = mb_count;
376
377     if (is_slice_short(ctx)) {
378         slice_data = ctx_pic->slice_short;
379         slice_size = ctx_pic->slice_count * sizeof(*ctx_pic->slice_short);
380     } else {
381         slice_data = ctx_pic->slice_long;
382         slice_size = ctx_pic->slice_count * sizeof(*ctx_pic->slice_long);
383     }
384     assert((bs->DataSize & 127) == 0);
385     return ff_dxva2_commit_buffer(avctx, ctx, sc,
386                                   DXVA2_SliceControlBufferType,
387                                   slice_data, slice_size, mb_count);
388 }
389
390
391 static int dxva2_h264_start_frame(AVCodecContext *avctx,
392                                   av_unused const uint8_t *buffer,
393                                   av_unused uint32_t size)
394 {
395     const H264Context *h = avctx->priv_data;
396     struct dxva_context *ctx = avctx->hwaccel_context;
397     struct dxva2_picture_context *ctx_pic = h->cur_pic_ptr->hwaccel_picture_private;
398
399     if (!ctx->decoder || !ctx->cfg || ctx->surface_count <= 0)
400         return -1;
401     assert(ctx_pic);
402
403     /* Fill up DXVA_PicParams_H264 */
404     fill_picture_parameters(ctx, h, &ctx_pic->pp);
405
406     /* Fill up DXVA_Qmatrix_H264 */
407     fill_scaling_lists(ctx, h, &ctx_pic->qm);
408
409     ctx_pic->slice_count    = 0;
410     ctx_pic->bitstream_size = 0;
411     ctx_pic->bitstream      = NULL;
412     return 0;
413 }
414
415 static int dxva2_h264_decode_slice(AVCodecContext *avctx,
416                                    const uint8_t *buffer,
417                                    uint32_t size)
418 {
419     const H264Context *h = avctx->priv_data;
420     const H264SliceContext *sl = &h->slice_ctx[0];
421     struct dxva_context *ctx = avctx->hwaccel_context;
422     const H264Picture *current_picture = h->cur_pic_ptr;
423     struct dxva2_picture_context *ctx_pic = current_picture->hwaccel_picture_private;
424     unsigned position;
425
426     if (ctx_pic->slice_count >= MAX_SLICES)
427         return -1;
428
429     if (!ctx_pic->bitstream)
430         ctx_pic->bitstream = buffer;
431     ctx_pic->bitstream_size += size;
432
433     position = buffer - ctx_pic->bitstream;
434     if (is_slice_short(ctx))
435         fill_slice_short(&ctx_pic->slice_short[ctx_pic->slice_count],
436                          position, size);
437     else
438         fill_slice_long(avctx, &ctx_pic->slice_long[ctx_pic->slice_count],
439                         &ctx_pic->pp, position, size);
440     ctx_pic->slice_count++;
441
442     if (sl->slice_type != AV_PICTURE_TYPE_I && sl->slice_type != AV_PICTURE_TYPE_SI)
443         ctx_pic->pp.wBitFields &= ~(1 << 15); /* Set IntraPicFlag to 0 */
444     return 0;
445 }
446
447 static int dxva2_h264_end_frame(AVCodecContext *avctx)
448 {
449     H264Context *h = avctx->priv_data;
450     H264SliceContext *sl = &h->slice_ctx[0];
451     struct dxva2_picture_context *ctx_pic =
452         h->cur_pic_ptr->hwaccel_picture_private;
453     int ret;
454
455     if (ctx_pic->slice_count <= 0 || ctx_pic->bitstream_size <= 0)
456         return -1;
457     ret = ff_dxva2_common_end_frame(avctx, h->cur_pic_ptr->f,
458                                     &ctx_pic->pp, sizeof(ctx_pic->pp),
459                                     &ctx_pic->qm, sizeof(ctx_pic->qm),
460                                     commit_bitstream_and_slice_buffer);
461     if (!ret)
462         ff_h264_draw_horiz_band(h, sl, 0, h->avctx->height);
463     return ret;
464 }
465
466 AVHWAccel ff_h264_dxva2_hwaccel = {
467     .name           = "h264_dxva2",
468     .type           = AVMEDIA_TYPE_VIDEO,
469     .id             = AV_CODEC_ID_H264,
470     .pix_fmt        = AV_PIX_FMT_DXVA2_VLD,
471     .start_frame    = dxva2_h264_start_frame,
472     .decode_slice   = dxva2_h264_decode_slice,
473     .end_frame      = dxva2_h264_end_frame,
474     .frame_priv_data_size = sizeof(struct dxva2_picture_context),
475 };