Update to MPlayer SVN rev 30827 and FFmpeg SVN rev 22201.
[vaapi:jass-mplayer.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;
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; i < FF_ARRAY_ELEMS(pp->RefFrameList); i++) {
60         if (i < h->short_ref_count + h->long_ref_count) {
61             const Picture *r;
62             if (i < h->short_ref_count) {
63                 r = h->short_ref[i];
64                 assert(!r->long_ref);
65             } else {
66                 r = h->long_ref[i - h->short_ref_count];
67                 assert(r->long_ref);
68             }
69             fill_picture_entry(&pp->RefFrameList[i],
70                                ff_dxva2_get_surface_index(ctx, r),
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        = 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     pp->Reserved16Bits                = 3; /* FIXME is there a way to detect the right mode ? */
117     pp->StatusReportFeedbackNumber    = 1 + ctx->report_id++;
118     pp->CurrFieldOrderCnt[0] = 0;
119     if ((s->picture_structure & PICT_TOP_FIELD) &&
120         current_picture->field_poc[0] != INT_MAX)
121         pp->CurrFieldOrderCnt[0] = current_picture->field_poc[0];
122     pp->CurrFieldOrderCnt[1] = 0;
123     if ((s->picture_structure & PICT_BOTTOM_FIELD) &&
124         current_picture->field_poc[1] != INT_MAX)
125         pp->CurrFieldOrderCnt[1] = current_picture->field_poc[1];
126     pp->pic_init_qs_minus26           = h->pps.init_qs - 26;
127     pp->chroma_qp_index_offset        = h->pps.chroma_qp_index_offset[0];
128     pp->second_chroma_qp_index_offset = h->pps.chroma_qp_index_offset[1];
129     pp->ContinuationFlag              = 1;
130     pp->pic_init_qp_minus26           = h->pps.init_qp - 26;
131     pp->num_ref_idx_l0_active_minus1  = h->pps.ref_count[0] - 1;
132     pp->num_ref_idx_l1_active_minus1  = h->pps.ref_count[1] - 1;
133     pp->Reserved8BitsA                = 0;
134     pp->frame_num                     = h->frame_num;
135     pp->log2_max_frame_num_minus4     = h->sps.log2_max_frame_num - 4;
136     pp->pic_order_cnt_type            = h->sps.poc_type;
137     if (h->sps.poc_type == 0)
138         pp->log2_max_pic_order_cnt_lsb_minus4 = h->sps.log2_max_poc_lsb - 4;
139     else if (h->sps.poc_type == 1)
140         pp->delta_pic_order_always_zero_flag = h->sps.delta_pic_order_always_zero_flag;
141     pp->direct_8x8_inference_flag     = h->sps.direct_8x8_inference_flag;
142     pp->entropy_coding_mode_flag      = h->pps.cabac;
143     pp->pic_order_present_flag        = h->pps.pic_order_present;
144     pp->num_slice_groups_minus1       = h->pps.slice_group_count - 1;
145     pp->slice_group_map_type          = h->pps.mb_slice_group_map_type;
146     pp->deblocking_filter_control_present_flag = h->pps.deblocking_filter_parameters_present;
147     pp->redundant_pic_cnt_present_flag= h->pps.redundant_pic_cnt_present;
148     pp->Reserved8BitsB                = 0;
149     pp->slice_group_change_rate_minus1= 0;  /* XXX not implemented by FFmpeg */
150     //pp->SliceGroupMap[810];               /* XXX not implemented by FFmpeg */
151 }
152
153 static void fill_scaling_lists(const H264Context *h, DXVA_Qmatrix_H264 *qm)
154 {
155     unsigned i, j;
156     memset(qm, 0, sizeof(*qm));
157     for (i = 0; i < 6; i++)
158         for (j = 0; j < 16; j++)
159             qm->bScalingLists4x4[i][j] = h->pps.scaling_matrix4[i][zigzag_scan[j]];
160
161     for (i = 0; i < 2; i++)
162         for (j = 0; j < 64; j++)
163             qm->bScalingLists8x8[i][j] = h->pps.scaling_matrix8[i][ff_zigzag_direct[j]];
164 }
165
166 static int is_slice_short(struct dxva_context *ctx)
167 {
168     assert(ctx->cfg->ConfigBitstreamRaw == 1 ||
169            ctx->cfg->ConfigBitstreamRaw == 2);
170     return ctx->cfg->ConfigBitstreamRaw == 2;
171 }
172
173 static void fill_slice_short(DXVA_Slice_H264_Short *slice,
174                              unsigned position, unsigned size)
175 {
176     memset(slice, 0, sizeof(*slice));
177     slice->BSNALunitDataLocation = position;
178     slice->SliceBytesInBuffer    = size;
179     slice->wBadSliceChopping     = 0;
180 }
181
182 static void fill_slice_long(AVCodecContext *avctx, DXVA_Slice_H264_Long *slice,
183                             unsigned position, unsigned size)
184 {
185     const H264Context *h = avctx->priv_data;
186     struct dxva_context *ctx = avctx->hwaccel_context;
187     const MpegEncContext *s = &h->s;
188     unsigned list;
189
190     memset(slice, 0, sizeof(*slice));
191     slice->BSNALunitDataLocation = position;
192     slice->SliceBytesInBuffer    = size;
193     slice->wBadSliceChopping     = 0;
194
195     slice->first_mb_in_slice     = (s->mb_y >> FIELD_OR_MBAFF_PICTURE) * s->mb_width + s->mb_x;
196     slice->NumMbsForSlice        = 0; /* XXX it is set once we have all slices */
197     slice->BitOffsetToSliceData  = get_bits_count(&s->gb) + 8;
198     slice->slice_type            = ff_h264_get_slice_type(h);
199     if (h->slice_type_fixed)
200         slice->slice_type += 5;
201     slice->luma_log2_weight_denom       = h->luma_log2_weight_denom;
202     slice->chroma_log2_weight_denom     = h->chroma_log2_weight_denom;
203     if (h->list_count > 0)
204         slice->num_ref_idx_l0_active_minus1 = h->ref_count[0] - 1;
205     if (h->list_count > 1)
206         slice->num_ref_idx_l1_active_minus1 = h->ref_count[1] - 1;
207     slice->slice_alpha_c0_offset_div2   = h->slice_alpha_c0_offset / 2 - 26;
208     slice->slice_beta_offset_div2       = h->slice_beta_offset     / 2 - 26;
209     slice->Reserved8Bits                = 0;
210
211     for (list = 0; list < 2; list++) {
212         unsigned i;
213         for (i = 0; i < FF_ARRAY_ELEMS(slice->RefPicList[list]); i++) {
214             if (list < h->list_count && i < h->ref_count[list]) {
215                 const Picture *r = &h->ref_list[list][i];
216                 unsigned plane;
217                 fill_picture_entry(&slice->RefPicList[list][i],
218                                    ff_dxva2_get_surface_index(ctx, r),
219                                    r->reference == PICT_BOTTOM_FIELD);
220                 for (plane = 0; plane < 3; plane++) {
221                     int w, o;
222                     if (plane == 0 && h->luma_weight_flag[list]) {
223                         w = h->luma_weight[list][i][0];
224                         o = h->luma_weight[list][i][1];
225                     } else if (plane >= 1 && h->chroma_weight_flag[list]) {
226                         w = h->chroma_weight[list][i][plane-1][0];
227                         o = h->chroma_weight[list][i][plane-1][1];
228                     } else {
229                         w = 1 << (plane == 0 ? h->luma_log2_weight_denom :
230                                                h->chroma_log2_weight_denom);
231                         o = 0;
232                     }
233                     slice->Weights[list][i][plane][0] = w;
234                     slice->Weights[list][i][plane][1] = o;
235                 }
236             } else {
237                 unsigned plane;
238                 slice->RefPicList[list][i].bPicEntry = 0xff;
239                 for (plane = 0; plane < 3; plane++) {
240                     slice->Weights[list][i][plane][0] = 0;
241                     slice->Weights[list][i][plane][1] = 0;
242                 }
243             }
244         }
245     }
246     slice->slice_qs_delta    = 0; /* XXX not implemented by FFmpeg */
247     slice->slice_qp_delta    = s->qscale - h->pps.init_qp;
248     slice->redundant_pic_cnt = h->redundant_pic_count;
249     if (h->slice_type == FF_B_TYPE)
250         slice->direct_spatial_mv_pred_flag = h->direct_spatial_mv_pred;
251     slice->cabac_init_idc = h->pps.cabac ? h->cabac_init_idc : 0;
252     if (h->deblocking_filter < 2)
253         slice->disable_deblocking_filter_idc = 1 - h->deblocking_filter;
254     else
255         slice->disable_deblocking_filter_idc = h->deblocking_filter;
256     slice->slice_id = h->current_slice - 1;
257 }
258
259 static int commit_bitstream_and_slice_buffer(AVCodecContext *avctx,
260                                              DXVA2_DecodeBufferDesc *bs,
261                                              DXVA2_DecodeBufferDesc *sc)
262 {
263     const H264Context *h = avctx->priv_data;
264     const MpegEncContext *s = &h->s;
265     const unsigned mb_count = s->mb_width * s->mb_height;
266     struct dxva_context *ctx = avctx->hwaccel_context;
267     const Picture *current_picture = h->s.current_picture_ptr;
268     struct dxva2_picture_context *ctx_pic = current_picture->hwaccel_picture_private;
269     DXVA_Slice_H264_Short *slice = NULL;
270     uint8_t  *dxva_data, *current, *end;
271     unsigned dxva_size;
272     void     *slice_data;
273     unsigned slice_size;
274     unsigned padding;
275     unsigned i;
276
277     /* Create an annex B bitstream buffer with only slice NAL and finalize slice */
278     if (FAILED(IDirectXVideoDecoder_GetBuffer(ctx->decoder,
279                                                DXVA2_BitStreamDateBufferType,
280                                                &dxva_data, &dxva_size)))
281         return -1;
282     current = dxva_data;
283     end = dxva_data + dxva_size;
284
285     for (i = 0; i < ctx_pic->slice_count; i++) {
286         static const uint8_t start_code[] = { 0, 0, 1 };
287         static const unsigned start_code_size = sizeof(start_code);
288         unsigned position, size;
289
290         assert(offsetof(DXVA_Slice_H264_Short, BSNALunitDataLocation) ==
291                offsetof(DXVA_Slice_H264_Long,  BSNALunitDataLocation));
292         assert(offsetof(DXVA_Slice_H264_Short, SliceBytesInBuffer) ==
293                offsetof(DXVA_Slice_H264_Long,  SliceBytesInBuffer));
294
295         if (is_slice_short(ctx))
296             slice = &ctx_pic->slice_short[i];
297         else
298             slice = (DXVA_Slice_H264_Short*)&ctx_pic->slice_long[i];
299
300         position = slice->BSNALunitDataLocation;
301         size     = slice->SliceBytesInBuffer;
302         if (start_code_size + size > end - current) {
303             av_log(avctx, AV_LOG_ERROR, "Failed to build bitstream");
304             break;
305         }
306
307         slice->BSNALunitDataLocation = current - dxva_data;
308         slice->SliceBytesInBuffer    = start_code_size + size;
309
310         if (!is_slice_short(ctx)) {
311             DXVA_Slice_H264_Long *slice_long = (DXVA_Slice_H264_Long*)slice;
312             if (i < ctx_pic->slice_count - 1)
313                 slice_long->NumMbsForSlice =
314                     slice_long[1].first_mb_in_slice - slice_long[0].first_mb_in_slice;
315             else
316                 slice_long->NumMbsForSlice = mb_count - slice_long->first_mb_in_slice;
317         }
318
319         memcpy(current, start_code, start_code_size);
320         current += start_code_size;
321
322         memcpy(current, &ctx_pic->bitstream[position], size);
323         current += size;
324     }
325     padding = FFMIN(128 - ((current - dxva_data) & 127), end - current);
326     if (slice && padding > 0) {
327         memset(current, 0, padding);
328         current += padding;
329
330         slice->SliceBytesInBuffer += padding;
331     }
332     if (FAILED(IDirectXVideoDecoder_ReleaseBuffer(ctx->decoder,
333                                                   DXVA2_BitStreamDateBufferType)))
334         return -1;
335     if (i < ctx_pic->slice_count)
336         return -1;
337
338     memset(bs, 0, sizeof(*bs));
339     bs->CompressedBufferType = DXVA2_BitStreamDateBufferType;
340     bs->DataSize             = current - dxva_data;
341     bs->NumMBsInBuffer       = mb_count;
342
343     if (is_slice_short(ctx)) {
344         slice_data = ctx_pic->slice_short;
345         slice_size = ctx_pic->slice_count * sizeof(*ctx_pic->slice_short);
346     } else {
347         slice_data = ctx_pic->slice_long;
348         slice_size = ctx_pic->slice_count * sizeof(*ctx_pic->slice_long);
349     }
350     assert((bs->DataSize & 127) == 0);
351     return ff_dxva2_commit_buffer(avctx, ctx, sc,
352                                   DXVA2_SliceControlBufferType,
353                                   slice_data, slice_size, mb_count);
354 }
355
356
357 static int start_frame(AVCodecContext *avctx,
358                        av_unused const uint8_t *buffer,
359                        av_unused uint32_t size)
360 {
361     const H264Context *h = avctx->priv_data;
362     struct dxva_context *ctx = avctx->hwaccel_context;
363     struct dxva2_picture_context *ctx_pic = h->s.current_picture_ptr->hwaccel_picture_private;
364
365     if (!ctx->decoder || !ctx->cfg || ctx->surface_count <= 0)
366         return -1;
367     assert(ctx_pic);
368
369     /* Fill up DXVA_PicParams_H264 */
370     fill_picture_parameters(ctx, h, &ctx_pic->pp);
371
372     /* Fill up DXVA_Qmatrix_H264 */
373     fill_scaling_lists(h, &ctx_pic->qm);
374
375     ctx_pic->slice_count    = 0;
376     ctx_pic->bitstream_size = 0;
377     ctx_pic->bitstream      = NULL;
378     return 0;
379 }
380
381 static int decode_slice(AVCodecContext *avctx,
382                         const uint8_t *buffer, uint32_t size)
383 {
384     const H264Context *h = avctx->priv_data;
385     struct dxva_context *ctx = avctx->hwaccel_context;
386     const Picture *current_picture = h->s.current_picture_ptr;
387     struct dxva2_picture_context *ctx_pic = current_picture->hwaccel_picture_private;
388     unsigned position;
389
390     if (ctx_pic->slice_count >= MAX_SLICES)
391         return -1;
392
393     if (!ctx_pic->bitstream)
394         ctx_pic->bitstream = buffer;
395     ctx_pic->bitstream_size += size;
396
397     position = buffer - ctx_pic->bitstream;
398     if (is_slice_short(ctx))
399         fill_slice_short(&ctx_pic->slice_short[ctx_pic->slice_count],
400                          position, size);
401     else
402         fill_slice_long(avctx, &ctx_pic->slice_long[ctx_pic->slice_count],
403                         position, size);
404     ctx_pic->slice_count++;
405
406     if (h->slice_type != FF_I_TYPE && h->slice_type != FF_SI_TYPE)
407         ctx_pic->pp.wBitFields &= ~(1 << 15); /* Set IntraPicFlag to 0 */
408     return 0;
409 }
410
411 static int end_frame(AVCodecContext *avctx)
412 {
413     H264Context *h = avctx->priv_data;
414     MpegEncContext *s = &h->s;
415     struct dxva2_picture_context *ctx_pic =
416         h->s.current_picture_ptr->hwaccel_picture_private;
417
418     if (ctx_pic->slice_count <= 0 || ctx_pic->bitstream_size <= 0)
419         return -1;
420     return ff_dxva2_common_end_frame(avctx, s,
421                                      &ctx_pic->pp, sizeof(ctx_pic->pp),
422                                      &ctx_pic->qm, sizeof(ctx_pic->qm),
423                                      commit_bitstream_and_slice_buffer);
424 }
425
426 AVHWAccel h264_dxva2_hwaccel = {
427     .name           = "h264_dxva2",
428     .type           = CODEC_TYPE_VIDEO,
429     .id             = CODEC_ID_H264,
430     .pix_fmt        = PIX_FMT_DXVA2_VLD,
431     .capabilities   = 0,
432     .start_frame    = start_frame,
433     .decode_slice   = decode_slice,
434     .end_frame      = end_frame,
435     .priv_data_size = sizeof(struct dxva2_picture_context),
436 };
437