Add DXVA2 HEVC HWAccel
[ffmpeg.git] / libavcodec / dxva2_hevc.c
1 /*
2  * DXVA2 HEVC HW acceleration.
3  *
4  * copyright (c) 2014 - 2015 Hendrik Leppkes
5  *
6  * This file is part of Libav.
7  *
8  * Libav 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  * Libav 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 Libav; if not, write to the Free Software
20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21  */
22
23 #include "libavutil/avassert.h"
24
25 #include "dxva2_internal.h"
26 #include "hevc.h"
27
28 #define MAX_SLICES 256
29
30 struct hevc_dxva2_picture_context {
31     DXVA_PicParams_HEVC   pp;
32     DXVA_Qmatrix_HEVC     qm;
33     unsigned              slice_count;
34     DXVA_Slice_HEVC_Short slice_short[MAX_SLICES];
35     const uint8_t         *bitstream;
36     unsigned              bitstream_size;
37 };
38
39 static void fill_picture_entry(DXVA_PicEntry_HEVC *pic,
40                                unsigned index, unsigned flag)
41 {
42     av_assert0((index & 0x7f) == index && (flag & 0x01) == flag);
43     pic->bPicEntry = index | (flag << 7);
44 }
45
46 static int get_refpic_index(const DXVA_PicParams_HEVC *pp, int surface_index)
47 {
48     int i;
49     for (i = 0; i < FF_ARRAY_ELEMS(pp->RefPicList); i++) {
50         if ((pp->RefPicList[i].bPicEntry & 0x7f) == surface_index)
51           return i;
52     }
53     return 0xff;
54 }
55
56 static void fill_picture_parameters(struct dxva_context *ctx, const HEVCContext *h,
57                                     DXVA_PicParams_HEVC *pp)
58 {
59     const HEVCFrame *current_picture = h->ref;
60     int i, j, k;
61
62     memset(pp, 0, sizeof(*pp));
63
64     pp->PicWidthInMinCbsY  = h->sps->min_cb_width;
65     pp->PicHeightInMinCbsY = h->sps->min_cb_height;
66
67     pp->wFormatAndSequenceInfoFlags = (h->sps->chroma_format_idc          <<  0) |
68                                       (h->sps->separate_colour_plane_flag <<  2) |
69                                       ((h->sps->bit_depth - 8)            <<  3) |
70                                       ((h->sps->bit_depth - 8)            <<  6) |
71                                       ((h->sps->log2_max_poc_lsb - 4)     <<  9) |
72                                       (0                                  << 13) |
73                                       (0                                  << 14) |
74                                       (0                                  << 15);
75
76     fill_picture_entry(&pp->CurrPic, ff_dxva2_get_surface_index(ctx, current_picture->frame), 0);
77
78     pp->sps_max_dec_pic_buffering_minus1         = h->sps->temporal_layer[h->sps->max_sub_layers - 1].max_dec_pic_buffering - 1;
79     pp->log2_min_luma_coding_block_size_minus3   = h->sps->log2_min_cb_size - 3;
80     pp->log2_diff_max_min_luma_coding_block_size = h->sps->log2_diff_max_min_coding_block_size;
81     pp->log2_min_transform_block_size_minus2     = h->sps->log2_min_tb_size - 2;
82     pp->log2_diff_max_min_transform_block_size   = h->sps->log2_max_trafo_size  - h->sps->log2_min_tb_size;
83     pp->max_transform_hierarchy_depth_inter      = h->sps->max_transform_hierarchy_depth_inter;
84     pp->max_transform_hierarchy_depth_intra      = h->sps->max_transform_hierarchy_depth_intra;
85     pp->num_short_term_ref_pic_sets              = h->sps->nb_st_rps;
86     pp->num_long_term_ref_pics_sps               = h->sps->num_long_term_ref_pics_sps;
87
88     pp->num_ref_idx_l0_default_active_minus1     = h->pps->num_ref_idx_l0_default_active - 1;
89     pp->num_ref_idx_l1_default_active_minus1     = h->pps->num_ref_idx_l1_default_active - 1;
90     pp->init_qp_minus26                          = h->pps->pic_init_qp_minus26;
91
92     if (h->sh.short_term_ref_pic_set_sps_flag == 0 && h->sh.short_term_rps) {
93         pp->ucNumDeltaPocsOfRefRpsIdx            = h->sh.short_term_rps->num_delta_pocs;
94         pp->wNumBitsForShortTermRPSInSlice       = h->sh.short_term_ref_pic_set_size;
95     }
96
97     pp->dwCodingParamToolFlags = (h->sps->scaling_list_enable_flag               <<  0) |
98                                  (h->sps->amp_enabled_flag                       <<  1) |
99                                  (h->sps->sao_enabled                            <<  2) |
100                                  (h->sps->pcm_enabled_flag                       <<  3) |
101                                  ((h->sps->pcm_enabled_flag ? (h->sps->pcm.bit_depth - 1) : 0)            <<  4) |
102                                  ((h->sps->pcm_enabled_flag ? (h->sps->pcm.bit_depth_chroma - 1) : 0)     <<  8) |
103                                  ((h->sps->pcm_enabled_flag ? (h->sps->pcm.log2_min_pcm_cb_size - 3) : 0) << 12) |
104                                  ((h->sps->pcm_enabled_flag ? (h->sps->pcm.log2_max_pcm_cb_size - h->sps->pcm.log2_min_pcm_cb_size) : 0) << 14) |
105                                  (h->sps->pcm.loop_filter_disable_flag           << 16) |
106                                  (h->sps->long_term_ref_pics_present_flag        << 17) |
107                                  (h->sps->sps_temporal_mvp_enabled_flag          << 18) |
108                                  (h->sps->sps_strong_intra_smoothing_enable_flag << 19) |
109                                  (h->pps->dependent_slice_segments_enabled_flag  << 20) |
110                                  (h->pps->output_flag_present_flag               << 21) |
111                                  (h->pps->num_extra_slice_header_bits            << 22) |
112                                  (h->pps->sign_data_hiding_flag                  << 25) |
113                                  (h->pps->cabac_init_present_flag                << 26) |
114                                  (0                                              << 27);
115
116     pp->dwCodingSettingPicturePropertyFlags = (h->pps->constrained_intra_pred_flag                <<  0) |
117                                               (h->pps->transform_skip_enabled_flag                <<  1) |
118                                               (h->pps->cu_qp_delta_enabled_flag                   <<  2) |
119                                               (h->pps->pic_slice_level_chroma_qp_offsets_present_flag <<  3) |
120                                               (h->pps->weighted_pred_flag                         <<  4) |
121                                               (h->pps->weighted_bipred_flag                       <<  5) |
122                                               (h->pps->transquant_bypass_enable_flag              <<  6) |
123                                               (h->pps->tiles_enabled_flag                         <<  7) |
124                                               (h->pps->entropy_coding_sync_enabled_flag           <<  8) |
125                                               (h->pps->uniform_spacing_flag                       <<  9) |
126                                               ((h->pps->tiles_enabled_flag ? h->pps->loop_filter_across_tiles_enabled_flag : 0) << 10) |
127                                               (h->pps->seq_loop_filter_across_slices_enabled_flag << 11) |
128                                               (h->pps->deblocking_filter_override_enabled_flag    << 12) |
129                                               (h->pps->disable_dbf                                << 13) |
130                                               (h->pps->lists_modification_present_flag            << 14) |
131                                               (h->pps->slice_header_extension_present_flag        << 15) |
132                                               (IS_IRAP(h)                                         << 16) |
133                                               (IS_IDR(h)                                          << 17) |
134                                               /* IntraPicFlag */
135                                               (IS_IRAP(h)                                         << 18) |
136                                               (0                                                  << 19);
137     pp->pps_cb_qp_offset            = h->pps->cb_qp_offset;
138     pp->pps_cr_qp_offset            = h->pps->cr_qp_offset;
139     if (h->pps->tiles_enabled_flag) {
140         pp->num_tile_columns_minus1 = h->pps->num_tile_columns - 1;
141         pp->num_tile_rows_minus1    = h->pps->num_tile_rows - 1;
142
143         if (!h->pps->uniform_spacing_flag) {
144             for (i = 0; i < h->pps->num_tile_columns; i++)
145                 pp->column_width_minus1[i] = h->pps->column_width[i] - 1;
146
147             for (i = 0; i < h->pps->num_tile_rows; i++)
148                 pp->row_height_minus1[i] = h->pps->row_height[i] - 1;
149         }
150     }
151
152     pp->diff_cu_qp_delta_depth           = h->pps->diff_cu_qp_delta_depth;
153     pp->pps_beta_offset_div2             = h->pps->beta_offset / 2;
154     pp->pps_tc_offset_div2               = h->pps->tc_offset / 2;
155     pp->log2_parallel_merge_level_minus2 = h->pps->log2_parallel_merge_level - 2;
156     pp->CurrPicOrderCntVal               = h->poc;
157
158     // empty the lists
159     memset(&pp->RefPicList, 0xff, sizeof(pp->RefPicList));
160     memset(&pp->RefPicSetStCurrBefore, 0xff, sizeof(pp->RefPicSetStCurrBefore));
161     memset(&pp->RefPicSetStCurrAfter, 0xff, sizeof(pp->RefPicSetStCurrAfter));
162     memset(&pp->RefPicSetLtCurr, 0xff, sizeof(pp->RefPicSetLtCurr));
163
164     // fill RefPicList from the DPB
165     for (i = 0, j = 0; i < FF_ARRAY_ELEMS(h->DPB); i++) {
166         const HEVCFrame *frame = &h->DPB[i];
167         if (frame != current_picture && (frame->flags & (HEVC_FRAME_FLAG_LONG_REF | HEVC_FRAME_FLAG_SHORT_REF))) {
168             fill_picture_entry(&pp->RefPicList[j], ff_dxva2_get_surface_index(ctx, frame->frame), !!(frame->flags & HEVC_FRAME_FLAG_LONG_REF));
169             pp->PicOrderCntValList[j] = frame->poc;
170             j++;
171         }
172     }
173
174     #define DO_REF_LIST(ref_idx, ref_list) { \
175         const RefPicList *rpl = &h->rps[ref_idx]; \
176         av_assert0(rpl->nb_refs <= FF_ARRAY_ELEMS(pp->ref_list)); \
177         for (j = 0, k = 0; j < rpl->nb_refs; j++) { \
178             if (rpl->ref[j]) { \
179                 pp->ref_list[k] = get_refpic_index(pp, ff_dxva2_get_surface_index(ctx, rpl->ref[j]->frame)); \
180                 k++; \
181             } \
182         } \
183     }
184
185     // Fill short term and long term lists
186     DO_REF_LIST(ST_CURR_BEF, RefPicSetStCurrBefore);
187     DO_REF_LIST(ST_CURR_AFT, RefPicSetStCurrAfter);
188     DO_REF_LIST(LT_CURR, RefPicSetLtCurr);
189
190     pp->StatusReportFeedbackNumber = 1 + ctx->report_id++;
191 }
192
193 static void fill_scaling_lists(struct dxva_context *ctx, const HEVCContext *h, DXVA_Qmatrix_HEVC *qm)
194 {
195     unsigned i, j;
196     memset(qm, 0, sizeof(*qm));
197     for (i = 0; i < 6; i++) {
198         for (j = 0; j < 16; j++)
199             qm->ucScalingLists0[i][j] = h->pps->scaling_list.sl[0][i][j];
200
201         for (j = 0; j < 64; j++) {
202             qm->ucScalingLists1[i][j] = h->pps->scaling_list.sl[1][i][j];
203             qm->ucScalingLists2[i][j] = h->pps->scaling_list.sl[2][i][j];
204
205             if (i < 2)
206                 qm->ucScalingLists3[i][j] = h->pps->scaling_list.sl[3][i][j];
207         }
208
209         qm->ucScalingListDCCoefSizeID2[i] = h->pps->scaling_list.sl_dc[0][i];
210         if (i < 2)
211             qm->ucScalingListDCCoefSizeID3[i] = h->pps->scaling_list.sl_dc[1][i];
212     }
213 }
214
215 static void fill_slice_short(DXVA_Slice_HEVC_Short *slice,
216                              unsigned position, unsigned size)
217 {
218     memset(slice, 0, sizeof(*slice));
219     slice->BSNALunitDataLocation = position;
220     slice->SliceBytesInBuffer    = size;
221     slice->wBadSliceChopping     = 0;
222 }
223
224 static int commit_bitstream_and_slice_buffer(AVCodecContext *avctx,
225                                              DXVA2_DecodeBufferDesc *bs,
226                                              DXVA2_DecodeBufferDesc *sc)
227 {
228     const HEVCContext *h = avctx->priv_data;
229     struct dxva_context *ctx = avctx->hwaccel_context;
230     const HEVCFrame *current_picture = h->ref;
231     struct hevc_dxva2_picture_context *ctx_pic = current_picture->hwaccel_picture_private;
232     DXVA_Slice_HEVC_Short *slice = NULL;
233     void     *dxva_data_ptr;
234     uint8_t  *dxva_data, *current, *end;
235     unsigned dxva_size;
236     void     *slice_data;
237     unsigned slice_size;
238     unsigned padding;
239     unsigned i;
240
241     /* Create an annex B bitstream buffer with only slice NAL and finalize slice */
242     if (FAILED(IDirectXVideoDecoder_GetBuffer(ctx->decoder,
243                                               DXVA2_BitStreamDateBufferType,
244                                               &dxva_data_ptr, &dxva_size)))
245         return -1;
246
247     dxva_data = dxva_data_ptr;
248     current = dxva_data;
249     end = dxva_data + dxva_size;
250
251     for (i = 0; i < ctx_pic->slice_count; i++) {
252         static const uint8_t start_code[] = { 0, 0, 1 };
253         static const unsigned start_code_size = sizeof(start_code);
254         unsigned position, size;
255
256         slice = &ctx_pic->slice_short[i];
257
258         position = slice->BSNALunitDataLocation;
259         size     = slice->SliceBytesInBuffer;
260         if (start_code_size + size > end - current) {
261             av_log(avctx, AV_LOG_ERROR, "Failed to build bitstream");
262             break;
263         }
264
265         slice->BSNALunitDataLocation = current - dxva_data;
266         slice->SliceBytesInBuffer    = start_code_size + size;
267
268         memcpy(current, start_code, start_code_size);
269         current += start_code_size;
270
271         memcpy(current, &ctx_pic->bitstream[position], size);
272         current += size;
273     }
274     padding = FFMIN(128 - ((current - dxva_data) & 127), end - current);
275     if (slice && padding > 0) {
276         memset(current, 0, padding);
277         current += padding;
278
279         slice->SliceBytesInBuffer += padding;
280     }
281     if (FAILED(IDirectXVideoDecoder_ReleaseBuffer(ctx->decoder,
282                                                   DXVA2_BitStreamDateBufferType)))
283         return -1;
284     if (i < ctx_pic->slice_count)
285         return -1;
286
287     memset(bs, 0, sizeof(*bs));
288     bs->CompressedBufferType = DXVA2_BitStreamDateBufferType;
289     bs->DataSize             = current - dxva_data;
290     bs->NumMBsInBuffer       = 0;
291
292     slice_data = ctx_pic->slice_short;
293     slice_size = ctx_pic->slice_count * sizeof(*ctx_pic->slice_short);
294
295     av_assert0((bs->DataSize & 127) == 0);
296     return ff_dxva2_commit_buffer(avctx, ctx, sc,
297                                   DXVA2_SliceControlBufferType,
298                                   slice_data, slice_size, 0);
299 }
300
301
302 static int dxva2_hevc_start_frame(AVCodecContext *avctx,
303                                   av_unused const uint8_t *buffer,
304                                   av_unused uint32_t size)
305 {
306     const HEVCContext *h = avctx->priv_data;
307     struct dxva_context *ctx = avctx->hwaccel_context;
308     struct hevc_dxva2_picture_context *ctx_pic = h->ref->hwaccel_picture_private;
309
310     if (!ctx->decoder || !ctx->cfg || ctx->surface_count <= 0)
311         return -1;
312     av_assert0(ctx_pic);
313
314     /* Fill up DXVA_PicParams_HEVC */
315     fill_picture_parameters(ctx, h, &ctx_pic->pp);
316
317     /* Fill up DXVA_Qmatrix_HEVC */
318     fill_scaling_lists(ctx, h, &ctx_pic->qm);
319
320     ctx_pic->slice_count    = 0;
321     ctx_pic->bitstream_size = 0;
322     ctx_pic->bitstream      = NULL;
323     return 0;
324 }
325
326 static int dxva2_hevc_decode_slice(AVCodecContext *avctx,
327                                    const uint8_t *buffer,
328                                    uint32_t size)
329 {
330     const HEVCContext *h = avctx->priv_data;
331     const HEVCFrame *current_picture = h->ref;
332     struct hevc_dxva2_picture_context *ctx_pic = current_picture->hwaccel_picture_private;
333     unsigned position;
334
335     if (ctx_pic->slice_count >= MAX_SLICES)
336         return -1;
337
338     if (!ctx_pic->bitstream)
339         ctx_pic->bitstream = buffer;
340     ctx_pic->bitstream_size += size;
341
342     position = buffer - ctx_pic->bitstream;
343     fill_slice_short(&ctx_pic->slice_short[ctx_pic->slice_count], position, size);
344     ctx_pic->slice_count++;
345
346     return 0;
347 }
348
349 static int dxva2_hevc_end_frame(AVCodecContext *avctx)
350 {
351     HEVCContext *h = avctx->priv_data;
352     struct hevc_dxva2_picture_context *ctx_pic = h->ref->hwaccel_picture_private;
353     int scale = ctx_pic->pp.dwCodingParamToolFlags & 1;
354     int ret;
355
356     if (ctx_pic->slice_count <= 0 || ctx_pic->bitstream_size <= 0)
357         return -1;
358
359     ret = ff_dxva2_common_end_frame(avctx, h->ref->frame,
360                                     &ctx_pic->pp, sizeof(ctx_pic->pp),
361                                     scale ? &ctx_pic->qm : NULL, scale ? sizeof(ctx_pic->qm) : 0,
362                                     commit_bitstream_and_slice_buffer);
363     return ret;
364 }
365
366 AVHWAccel ff_hevc_dxva2_hwaccel = {
367     .name           = "hevc_dxva2",
368     .type           = AVMEDIA_TYPE_VIDEO,
369     .id             = AV_CODEC_ID_HEVC,
370     .pix_fmt        = AV_PIX_FMT_DXVA2_VLD,
371     .start_frame    = dxva2_hevc_start_frame,
372     .decode_slice   = dxva2_hevc_decode_slice,
373     .end_frame      = dxva2_hevc_end_frame,
374     .frame_priv_data_size = sizeof(struct hevc_dxva2_picture_context),
375 };