]> git.sesse.net Git - ffmpeg/blob - libavcodec/vaapi_encode_h265.c
Merge commit '0e4c166cdd6446522a085dd9731967d09ac71f72'
[ffmpeg] / libavcodec / vaapi_encode_h265.c
1 /*
2  * This file is part of FFmpeg.
3  *
4  * FFmpeg is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU Lesser General Public
6  * License as published by the Free Software Foundation; either
7  * version 2.1 of the License, or (at your option) any later version.
8  *
9  * FFmpeg is distributed in the hope that it will be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
12  * Lesser General Public License for more details.
13  *
14  * You should have received a copy of the GNU Lesser General Public
15  * License along with FFmpeg; if not, write to the Free Software
16  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
17  */
18
19 #include <string.h>
20
21 #include <va/va.h>
22 #include <va/va_enc_hevc.h>
23
24 #include "libavutil/avassert.h"
25 #include "libavutil/common.h"
26 #include "libavutil/opt.h"
27
28 #include "avcodec.h"
29 #include "cbs.h"
30 #include "cbs_h265.h"
31 #include "hevc.h"
32 #include "internal.h"
33 #include "put_bits.h"
34 #include "vaapi_encode.h"
35
36
37 typedef struct VAAPIEncodeH265Context {
38     unsigned int ctu_width;
39     unsigned int ctu_height;
40
41     int fixed_qp_idr;
42     int fixed_qp_p;
43     int fixed_qp_b;
44
45     H265RawAUD aud;
46     H265RawVPS vps;
47     H265RawSPS sps;
48     H265RawPPS pps;
49     H265RawSlice slice;
50
51     int64_t last_idr_frame;
52     int pic_order_cnt;
53
54     int slice_nal_unit;
55     int slice_type;
56     int pic_type;
57
58     CodedBitstreamContext *cbc;
59     CodedBitstreamFragment current_access_unit;
60     int aud_needed;
61 } VAAPIEncodeH265Context;
62
63 typedef struct VAAPIEncodeH265Options {
64     int qp;
65     int aud;
66     int profile;
67     int level;
68 } VAAPIEncodeH265Options;
69
70
71 static int vaapi_encode_h265_write_access_unit(AVCodecContext *avctx,
72                                                char *data, size_t *data_len,
73                                                CodedBitstreamFragment *au)
74 {
75     VAAPIEncodeContext      *ctx = avctx->priv_data;
76     VAAPIEncodeH265Context *priv = ctx->priv_data;
77     int err;
78
79     err = ff_cbs_write_fragment_data(priv->cbc, au);
80     if (err < 0) {
81         av_log(avctx, AV_LOG_ERROR, "Failed to write packed header.\n");
82         return err;
83     }
84
85     if (*data_len < 8 * au->data_size - au->data_bit_padding) {
86         av_log(avctx, AV_LOG_ERROR, "Access unit too large: "
87                "%zu < %zu.\n", *data_len,
88                8 * au->data_size - au->data_bit_padding);
89         return AVERROR(ENOSPC);
90     }
91
92     memcpy(data, au->data, au->data_size);
93     *data_len = 8 * au->data_size - au->data_bit_padding;
94
95     return 0;
96 }
97
98 static int vaapi_encode_h265_add_nal(AVCodecContext *avctx,
99                                      CodedBitstreamFragment *au,
100                                      void *nal_unit)
101 {
102     VAAPIEncodeContext      *ctx = avctx->priv_data;
103     VAAPIEncodeH265Context *priv = ctx->priv_data;
104     H265RawNALUnitHeader *header = nal_unit;
105     int err;
106
107     err = ff_cbs_insert_unit_content(priv->cbc, au, -1,
108                                      header->nal_unit_type, nal_unit);
109     if (err < 0) {
110         av_log(avctx, AV_LOG_ERROR, "Failed to add NAL unit: "
111                "type = %d.\n", header->nal_unit_type);
112         return err;
113     }
114
115     return 0;
116 }
117
118 static int vaapi_encode_h265_write_sequence_header(AVCodecContext *avctx,
119                                                    char *data, size_t *data_len)
120 {
121     VAAPIEncodeContext      *ctx = avctx->priv_data;
122     VAAPIEncodeH265Context *priv = ctx->priv_data;
123     CodedBitstreamFragment   *au = &priv->current_access_unit;
124     int err;
125
126     if (priv->aud_needed) {
127         err = vaapi_encode_h265_add_nal(avctx, au, &priv->aud);
128         if (err < 0)
129             goto fail;
130         priv->aud_needed = 0;
131     }
132
133     err = vaapi_encode_h265_add_nal(avctx, au, &priv->vps);
134     if (err < 0)
135         goto fail;
136
137     err = vaapi_encode_h265_add_nal(avctx, au, &priv->sps);
138     if (err < 0)
139         goto fail;
140
141     err = vaapi_encode_h265_add_nal(avctx, au, &priv->pps);
142     if (err < 0)
143         goto fail;
144
145     err = vaapi_encode_h265_write_access_unit(avctx, data, data_len, au);
146 fail:
147     ff_cbs_fragment_uninit(priv->cbc, au);
148     return err;
149 }
150
151 static int vaapi_encode_h265_write_slice_header(AVCodecContext *avctx,
152                                                 VAAPIEncodePicture *pic,
153                                                 VAAPIEncodeSlice *slice,
154                                                 char *data, size_t *data_len)
155 {
156     VAAPIEncodeContext      *ctx = avctx->priv_data;
157     VAAPIEncodeH265Context *priv = ctx->priv_data;
158     CodedBitstreamFragment   *au = &priv->current_access_unit;
159     int err;
160
161     if (priv->aud_needed) {
162         err = vaapi_encode_h265_add_nal(avctx, au, &priv->aud);
163         if (err < 0)
164             goto fail;
165         priv->aud_needed = 0;
166     }
167
168     err = vaapi_encode_h265_add_nal(avctx, au, &priv->slice);
169     if (err < 0)
170         goto fail;
171
172     err = vaapi_encode_h265_write_access_unit(avctx, data, data_len, au);
173 fail:
174     ff_cbs_fragment_uninit(priv->cbc, au);
175     return err;
176 }
177
178 static int vaapi_encode_h265_init_sequence_params(AVCodecContext *avctx)
179 {
180     VAAPIEncodeContext                *ctx = avctx->priv_data;
181     VAAPIEncodeH265Context           *priv = ctx->priv_data;
182     H265RawVPS                        *vps = &priv->vps;
183     H265RawSPS                        *sps = &priv->sps;
184     H265RawPPS                        *pps = &priv->pps;
185     H265RawVUI                        *vui = &sps->vui;
186     VAEncSequenceParameterBufferHEVC *vseq = ctx->codec_sequence_params;
187     VAEncPictureParameterBufferHEVC  *vpic = ctx->codec_picture_params;
188     int i;
189
190     memset(&priv->current_access_unit, 0,
191            sizeof(priv->current_access_unit));
192
193     memset(vps, 0, sizeof(*vps));
194     memset(sps, 0, sizeof(*sps));
195     memset(pps, 0, sizeof(*pps));
196
197
198     // VPS
199
200     vps->nal_unit_header = (H265RawNALUnitHeader) {
201         .nal_unit_type         = HEVC_NAL_VPS,
202         .nuh_layer_id          = 0,
203         .nuh_temporal_id_plus1 = 1,
204     };
205
206     vps->vps_video_parameter_set_id = 0;
207
208     vps->vps_base_layer_internal_flag  = 1;
209     vps->vps_base_layer_available_flag = 1;
210     vps->vps_max_layers_minus1         = 0;
211     vps->vps_max_sub_layers_minus1     = 0;
212     vps->vps_temporal_id_nesting_flag  = 1;
213
214     vps->profile_tier_level = (H265RawProfileTierLevel) {
215         .general_profile_space = 0,
216         .general_profile_idc   = avctx->profile,
217         .general_tier_flag     = 0,
218
219         .general_progressive_source_flag    = 1,
220         .general_interlaced_source_flag     = 0,
221         .general_non_packed_constraint_flag = 1,
222         .general_frame_only_constraint_flag = 1,
223
224         .general_level_idc     = avctx->level,
225     };
226     vps->profile_tier_level.general_profile_compatibility_flag[avctx->profile & 31] = 1;
227
228     vps->vps_sub_layer_ordering_info_present_flag = 0;
229     vps->vps_max_dec_pic_buffering_minus1[0]      = (ctx->b_per_p > 0) + 1;
230     vps->vps_max_num_reorder_pics[0]              = (ctx->b_per_p > 0);
231     vps->vps_max_latency_increase_plus1[0]        = 0;
232
233     vps->vps_max_layer_id             = 0;
234     vps->vps_num_layer_sets_minus1    = 0;
235     vps->layer_id_included_flag[0][0] = 1;
236
237     vps->vps_timing_info_present_flag = 1;
238     if (avctx->framerate.num > 0 && avctx->framerate.den > 0) {
239         vps->vps_num_units_in_tick  = avctx->framerate.den;
240         vps->vps_time_scale         = avctx->framerate.num;
241         vps->vps_poc_proportional_to_timing_flag = 1;
242         vps->vps_num_ticks_poc_diff_one_minus1   = 0;
243     } else {
244         vps->vps_num_units_in_tick  = avctx->time_base.num;
245         vps->vps_time_scale         = avctx->time_base.den;
246         vps->vps_poc_proportional_to_timing_flag = 0;
247     }
248     vps->vps_num_hrd_parameters = 0;
249
250
251     // SPS
252
253     sps->nal_unit_header = (H265RawNALUnitHeader) {
254         .nal_unit_type         = HEVC_NAL_SPS,
255         .nuh_layer_id          = 0,
256         .nuh_temporal_id_plus1 = 1,
257     };
258
259     sps->sps_video_parameter_set_id = vps->vps_video_parameter_set_id;
260
261     sps->sps_max_sub_layers_minus1    = vps->vps_max_sub_layers_minus1;
262     sps->sps_temporal_id_nesting_flag = vps->vps_temporal_id_nesting_flag;
263
264     sps->profile_tier_level = vps->profile_tier_level;
265
266     sps->sps_seq_parameter_set_id = 0;
267
268     sps->chroma_format_idc          = 1; // YUV 4:2:0.
269     sps->separate_colour_plane_flag = 0;
270
271     sps->pic_width_in_luma_samples  = ctx->surface_width;
272     sps->pic_height_in_luma_samples = ctx->surface_height;
273
274     if (avctx->width  != ctx->surface_width ||
275         avctx->height != ctx->surface_height) {
276         sps->conformance_window_flag = 1;
277         sps->conf_win_left_offset   = 0;
278         sps->conf_win_right_offset  =
279             (ctx->surface_width - avctx->width) / 2;
280         sps->conf_win_top_offset    = 0;
281         sps->conf_win_bottom_offset =
282             (ctx->surface_height - avctx->height) / 2;
283     } else {
284         sps->conformance_window_flag = 0;
285     }
286
287     sps->bit_depth_luma_minus8 =
288         avctx->profile == FF_PROFILE_HEVC_MAIN_10 ? 2 : 0;
289     sps->bit_depth_chroma_minus8 = sps->bit_depth_luma_minus8;
290
291     sps->log2_max_pic_order_cnt_lsb_minus4 = 8;
292
293     sps->sps_sub_layer_ordering_info_present_flag =
294         vps->vps_sub_layer_ordering_info_present_flag;
295     for (i = 0; i <= sps->sps_max_sub_layers_minus1; i++) {
296         sps->sps_max_dec_pic_buffering_minus1[i] =
297             vps->vps_max_dec_pic_buffering_minus1[i];
298         sps->sps_max_num_reorder_pics[i] =
299             vps->vps_max_num_reorder_pics[i];
300         sps->sps_max_latency_increase_plus1[i] =
301             vps->vps_max_latency_increase_plus1[i];
302     }
303
304     // These have to come from the capabilities of the encoder.  We have no
305     // way to query them, so just hardcode parameters which work on the Intel
306     // driver.
307     // CTB size from 8x8 to 32x32.
308     sps->log2_min_luma_coding_block_size_minus3   = 0;
309     sps->log2_diff_max_min_luma_coding_block_size = 2;
310     // Transform size from 4x4 to 32x32.
311     sps->log2_min_luma_transform_block_size_minus2   = 0;
312     sps->log2_diff_max_min_luma_transform_block_size = 3;
313     // Full transform hierarchy allowed (2-5).
314     sps->max_transform_hierarchy_depth_inter = 3;
315     sps->max_transform_hierarchy_depth_intra = 3;
316     // AMP works.
317     sps->amp_enabled_flag = 1;
318     // SAO and temporal MVP do not work.
319     sps->sample_adaptive_offset_enabled_flag = 0;
320     sps->sps_temporal_mvp_enabled_flag       = 0;
321
322     sps->pcm_enabled_flag = 0;
323
324     // STRPSs should ideally be here rather than defined individually in
325     // each slice, but the structure isn't completely fixed so for now
326     // don't bother.
327     sps->num_short_term_ref_pic_sets     = 0;
328     sps->long_term_ref_pics_present_flag = 0;
329
330     sps->vui_parameters_present_flag = 1;
331
332     if (avctx->sample_aspect_ratio.num != 0 &&
333         avctx->sample_aspect_ratio.den != 0) {
334         static const AVRational sar_idc[] = {
335             {   0,  0 },
336             {   1,  1 }, {  12, 11 }, {  10, 11 }, {  16, 11 },
337             {  40, 33 }, {  24, 11 }, {  20, 11 }, {  32, 11 },
338             {  80, 33 }, {  18, 11 }, {  15, 11 }, {  64, 33 },
339             { 160, 99 }, {   4,  3 }, {   3,  2 }, {   2,  1 },
340         };
341         int i;
342         for (i = 0; i < FF_ARRAY_ELEMS(sar_idc); i++) {
343             if (avctx->sample_aspect_ratio.num == sar_idc[i].num &&
344                 avctx->sample_aspect_ratio.den == sar_idc[i].den) {
345                 vui->aspect_ratio_idc = i;
346                 break;
347             }
348         }
349         if (i >= FF_ARRAY_ELEMS(sar_idc)) {
350             vui->aspect_ratio_idc = 255;
351             vui->sar_width  = avctx->sample_aspect_ratio.num;
352             vui->sar_height = avctx->sample_aspect_ratio.den;
353         }
354         vui->aspect_ratio_info_present_flag = 1;
355     }
356
357     if (avctx->color_range     != AVCOL_RANGE_UNSPECIFIED ||
358         avctx->color_primaries != AVCOL_PRI_UNSPECIFIED ||
359         avctx->color_trc       != AVCOL_TRC_UNSPECIFIED ||
360         avctx->colorspace      != AVCOL_SPC_UNSPECIFIED) {
361         vui->video_signal_type_present_flag = 1;
362         vui->video_format      = 5; // Unspecified.
363         vui->video_full_range_flag =
364             avctx->color_range == AVCOL_RANGE_JPEG;
365
366         if (avctx->color_primaries != AVCOL_PRI_UNSPECIFIED ||
367             avctx->color_trc       != AVCOL_TRC_UNSPECIFIED ||
368             avctx->colorspace      != AVCOL_SPC_UNSPECIFIED) {
369             vui->colour_description_present_flag = 1;
370             vui->colour_primaries         = avctx->color_primaries;
371             vui->transfer_characteristics = avctx->color_trc;
372             vui->matrix_coefficients      = avctx->colorspace;
373         }
374     } else {
375         vui->video_format             = 5;
376         vui->video_full_range_flag    = 0;
377         vui->colour_primaries         = avctx->color_primaries;
378         vui->transfer_characteristics = avctx->color_trc;
379         vui->matrix_coefficients      = avctx->colorspace;
380     }
381
382     if (avctx->chroma_sample_location != AVCHROMA_LOC_UNSPECIFIED) {
383         vui->chroma_loc_info_present_flag = 1;
384         vui->chroma_sample_loc_type_top_field    =
385         vui->chroma_sample_loc_type_bottom_field =
386             avctx->chroma_sample_location - 1;
387     }
388
389     vui->vui_timing_info_present_flag        = 1;
390     vui->vui_num_units_in_tick               = vps->vps_num_units_in_tick;
391     vui->vui_time_scale                      = vps->vps_time_scale;
392     vui->vui_poc_proportional_to_timing_flag = vps->vps_poc_proportional_to_timing_flag;
393     vui->vui_num_ticks_poc_diff_one_minus1   = vps->vps_num_ticks_poc_diff_one_minus1;
394     vui->vui_hrd_parameters_present_flag     = 0;
395
396     vui->bitstream_restriction_flag    = 1;
397     vui->motion_vectors_over_pic_boundaries_flag = 1;
398     vui->restricted_ref_pic_lists_flag = 1;
399     vui->max_bytes_per_pic_denom       = 0;
400     vui->max_bits_per_min_cu_denom     = 0;
401     vui->log2_max_mv_length_horizontal = 15;
402     vui->log2_max_mv_length_vertical   = 15;
403
404
405     // PPS
406
407     pps->nal_unit_header = (H265RawNALUnitHeader) {
408         .nal_unit_type         = HEVC_NAL_PPS,
409         .nuh_layer_id          = 0,
410         .nuh_temporal_id_plus1 = 1,
411     };
412
413     pps->pps_pic_parameter_set_id = 0;
414     pps->pps_seq_parameter_set_id = sps->sps_seq_parameter_set_id;
415
416     pps->num_ref_idx_l0_default_active_minus1 = 0;
417     pps->num_ref_idx_l1_default_active_minus1 = 0;
418
419     pps->init_qp_minus26 = priv->fixed_qp_idr - 26;
420
421     pps->cu_qp_delta_enabled_flag = (ctx->va_rc_mode != VA_RC_CQP);
422     pps->diff_cu_qp_delta_depth   = 0;
423
424     pps->pps_loop_filter_across_slices_enabled_flag = 1;
425
426
427     // Fill VAAPI parameter buffers.
428
429     *vseq = (VAEncSequenceParameterBufferHEVC) {
430         .general_profile_idc = vps->profile_tier_level.general_profile_idc,
431         .general_level_idc   = vps->profile_tier_level.general_level_idc,
432         .general_tier_flag   = vps->profile_tier_level.general_tier_flag,
433
434         .intra_period     = avctx->gop_size,
435         .intra_idr_period = avctx->gop_size,
436         .ip_period        = ctx->b_per_p + 1,
437         .bits_per_second   = avctx->bit_rate,
438
439         .pic_width_in_luma_samples  = sps->pic_width_in_luma_samples,
440         .pic_height_in_luma_samples = sps->pic_height_in_luma_samples,
441
442         .seq_fields.bits = {
443             .chroma_format_idc             = sps->chroma_format_idc,
444             .separate_colour_plane_flag    = sps->separate_colour_plane_flag,
445             .bit_depth_luma_minus8         = sps->bit_depth_luma_minus8,
446             .bit_depth_chroma_minus8       = sps->bit_depth_chroma_minus8,
447             .scaling_list_enabled_flag     = sps->scaling_list_enabled_flag,
448             .strong_intra_smoothing_enabled_flag =
449                 sps->strong_intra_smoothing_enabled_flag,
450             .amp_enabled_flag              = sps->amp_enabled_flag,
451             .sample_adaptive_offset_enabled_flag =
452                 sps->sample_adaptive_offset_enabled_flag,
453             .pcm_enabled_flag              = sps->pcm_enabled_flag,
454             .pcm_loop_filter_disabled_flag = sps->pcm_loop_filter_disabled_flag,
455             .sps_temporal_mvp_enabled_flag = sps->sps_temporal_mvp_enabled_flag,
456         },
457
458         .log2_min_luma_coding_block_size_minus3 =
459             sps->log2_min_luma_coding_block_size_minus3,
460         .log2_diff_max_min_luma_coding_block_size =
461             sps->log2_diff_max_min_luma_coding_block_size,
462         .log2_min_transform_block_size_minus2 =
463             sps->log2_min_luma_transform_block_size_minus2,
464         .log2_diff_max_min_transform_block_size =
465             sps->log2_diff_max_min_luma_transform_block_size,
466         .max_transform_hierarchy_depth_inter =
467             sps->max_transform_hierarchy_depth_inter,
468         .max_transform_hierarchy_depth_intra =
469             sps->max_transform_hierarchy_depth_intra,
470
471         .pcm_sample_bit_depth_luma_minus1 =
472             sps->pcm_sample_bit_depth_luma_minus1,
473         .pcm_sample_bit_depth_chroma_minus1 =
474             sps->pcm_sample_bit_depth_chroma_minus1,
475         .log2_min_pcm_luma_coding_block_size_minus3 =
476             sps->log2_min_pcm_luma_coding_block_size_minus3,
477         .log2_max_pcm_luma_coding_block_size_minus3 =
478             sps->log2_min_pcm_luma_coding_block_size_minus3 +
479             sps->log2_diff_max_min_pcm_luma_coding_block_size,
480
481         .vui_parameters_present_flag = 0,
482     };
483
484     *vpic = (VAEncPictureParameterBufferHEVC) {
485         .decoded_curr_pic = {
486             .picture_id = VA_INVALID_ID,
487             .flags      = VA_PICTURE_HEVC_INVALID,
488         },
489
490         .coded_buf = VA_INVALID_ID,
491
492         .collocated_ref_pic_index = 0xff,
493
494         .last_picture = 0,
495
496         .pic_init_qp            = pps->init_qp_minus26 + 26,
497         .diff_cu_qp_delta_depth = pps->diff_cu_qp_delta_depth,
498         .pps_cb_qp_offset       = pps->pps_cb_qp_offset,
499         .pps_cr_qp_offset       = pps->pps_cr_qp_offset,
500
501         .num_tile_columns_minus1 = pps->num_tile_columns_minus1,
502         .num_tile_rows_minus1    = pps->num_tile_rows_minus1,
503
504         .log2_parallel_merge_level_minus2 = pps->log2_parallel_merge_level_minus2,
505         .ctu_max_bitsize_allowed          = 0,
506
507         .num_ref_idx_l0_default_active_minus1 =
508             pps->num_ref_idx_l0_default_active_minus1,
509         .num_ref_idx_l1_default_active_minus1 =
510             pps->num_ref_idx_l1_default_active_minus1,
511
512         .slice_pic_parameter_set_id = pps->pps_pic_parameter_set_id,
513
514         .pic_fields.bits = {
515             .sign_data_hiding_enabled_flag  = pps->sign_data_hiding_enabled_flag,
516             .constrained_intra_pred_flag    = pps->constrained_intra_pred_flag,
517             .transform_skip_enabled_flag    = pps->transform_skip_enabled_flag,
518             .cu_qp_delta_enabled_flag       = pps->cu_qp_delta_enabled_flag,
519             .weighted_pred_flag             = pps->weighted_pred_flag,
520             .weighted_bipred_flag           = pps->weighted_bipred_flag,
521             .transquant_bypass_enabled_flag = pps->transquant_bypass_enabled_flag,
522             .tiles_enabled_flag             = pps->tiles_enabled_flag,
523             .entropy_coding_sync_enabled_flag = pps->entropy_coding_sync_enabled_flag,
524             .loop_filter_across_tiles_enabled_flag =
525                 pps->loop_filter_across_tiles_enabled_flag,
526             .scaling_list_data_present_flag = (sps->sps_scaling_list_data_present_flag |
527                                                pps->pps_scaling_list_data_present_flag),
528             .screen_content_flag            = 0,
529             .enable_gpu_weighted_prediction = 0,
530             .no_output_of_prior_pics_flag   = 0,
531         },
532     };
533
534     return 0;
535 }
536
537 static int vaapi_encode_h265_init_picture_params(AVCodecContext *avctx,
538                                                  VAAPIEncodePicture *pic)
539 {
540     VAAPIEncodeContext               *ctx = avctx->priv_data;
541     VAAPIEncodeH265Context          *priv = ctx->priv_data;
542     VAAPIEncodeH265Options           *opt = ctx->codec_options;
543     VAEncPictureParameterBufferHEVC *vpic = pic->codec_picture_params;
544     int i;
545
546     if (pic->type == PICTURE_TYPE_IDR) {
547         av_assert0(pic->display_order == pic->encode_order);
548
549         priv->last_idr_frame = pic->display_order;
550
551         priv->slice_nal_unit = HEVC_NAL_IDR_W_RADL;
552         priv->slice_type     = HEVC_SLICE_I;
553         priv->pic_type       = 0;
554     } else {
555         av_assert0(pic->encode_order > priv->last_idr_frame);
556
557         if (pic->type == PICTURE_TYPE_I) {
558             priv->slice_nal_unit = HEVC_NAL_CRA_NUT;
559             priv->slice_type     = HEVC_SLICE_I;
560             priv->pic_type       = 0;
561         } else if (pic->type == PICTURE_TYPE_P) {
562             av_assert0(pic->refs[0]);
563             priv->slice_nal_unit = HEVC_NAL_TRAIL_R;
564             priv->slice_type     = HEVC_SLICE_P;
565             priv->pic_type       = 1;
566         } else {
567             av_assert0(pic->refs[0] && pic->refs[1]);
568             if (pic->refs[1]->type == PICTURE_TYPE_I)
569                 priv->slice_nal_unit = HEVC_NAL_RASL_N;
570             else
571                 priv->slice_nal_unit = HEVC_NAL_TRAIL_N;
572             priv->slice_type = HEVC_SLICE_B;
573             priv->pic_type   = 2;
574         }
575     }
576     priv->pic_order_cnt = pic->display_order - priv->last_idr_frame;
577
578     if (opt->aud) {
579         priv->aud_needed = 1;
580         priv->aud.nal_unit_header = (H265RawNALUnitHeader) {
581             .nal_unit_type         = HEVC_NAL_AUD,
582             .nuh_layer_id          = 0,
583             .nuh_temporal_id_plus1 = 1,
584         };
585         priv->aud.pic_type = priv->pic_type;
586     } else {
587         priv->aud_needed = 0;
588     }
589
590     vpic->decoded_curr_pic = (VAPictureHEVC) {
591         .picture_id    = pic->recon_surface,
592         .pic_order_cnt = priv->pic_order_cnt,
593         .flags         = 0,
594     };
595
596     for (i = 0; i < pic->nb_refs; i++) {
597         VAAPIEncodePicture *ref = pic->refs[i];
598         av_assert0(ref && ref->encode_order < pic->encode_order);
599
600         vpic->reference_frames[i] = (VAPictureHEVC) {
601             .picture_id    = ref->recon_surface,
602             .pic_order_cnt = ref->display_order - priv->last_idr_frame,
603             .flags = (ref->display_order < pic->display_order ?
604                       VA_PICTURE_HEVC_RPS_ST_CURR_BEFORE : 0) |
605                      (ref->display_order > pic->display_order ?
606                       VA_PICTURE_HEVC_RPS_ST_CURR_AFTER  : 0),
607         };
608     }
609     for (; i < FF_ARRAY_ELEMS(vpic->reference_frames); i++) {
610         vpic->reference_frames[i] = (VAPictureHEVC) {
611             .picture_id = VA_INVALID_ID,
612             .flags      = VA_PICTURE_HEVC_INVALID,
613         };
614     }
615
616     vpic->coded_buf = pic->output_buffer;
617
618     vpic->nal_unit_type = priv->slice_nal_unit;
619
620     switch (pic->type) {
621     case PICTURE_TYPE_IDR:
622         vpic->pic_fields.bits.idr_pic_flag       = 1;
623         vpic->pic_fields.bits.coding_type        = 1;
624         vpic->pic_fields.bits.reference_pic_flag = 1;
625         break;
626     case PICTURE_TYPE_I:
627         vpic->pic_fields.bits.idr_pic_flag       = 0;
628         vpic->pic_fields.bits.coding_type        = 1;
629         vpic->pic_fields.bits.reference_pic_flag = 1;
630         break;
631     case PICTURE_TYPE_P:
632         vpic->pic_fields.bits.idr_pic_flag       = 0;
633         vpic->pic_fields.bits.coding_type        = 2;
634         vpic->pic_fields.bits.reference_pic_flag = 1;
635         break;
636     case PICTURE_TYPE_B:
637         vpic->pic_fields.bits.idr_pic_flag       = 0;
638         vpic->pic_fields.bits.coding_type        = 3;
639         vpic->pic_fields.bits.reference_pic_flag = 0;
640         break;
641     default:
642         av_assert0(0 && "invalid picture type");
643     }
644
645     pic->nb_slices = 1;
646
647     return 0;
648 }
649
650 static int vaapi_encode_h265_init_slice_params(AVCodecContext *avctx,
651                                                VAAPIEncodePicture *pic,
652                                                VAAPIEncodeSlice *slice)
653 {
654     VAAPIEncodeContext                *ctx = avctx->priv_data;
655     VAAPIEncodeH265Context           *priv = ctx->priv_data;
656     const H265RawSPS                  *sps = &priv->sps;
657     const H265RawPPS                  *pps = &priv->pps;
658     H265RawSliceHeader                 *sh = &priv->slice.header;
659     VAEncPictureParameterBufferHEVC  *vpic = pic->codec_picture_params;
660     VAEncSliceParameterBufferHEVC  *vslice = slice->codec_slice_params;
661     int i;
662
663     sh->nal_unit_header = (H265RawNALUnitHeader) {
664         .nal_unit_type         = priv->slice_nal_unit,
665         .nuh_layer_id          = 0,
666         .nuh_temporal_id_plus1 = 1,
667     };
668
669     sh->slice_pic_parameter_set_id      = pps->pps_pic_parameter_set_id;
670
671     // Currently we only support one slice per frame.
672     sh->first_slice_segment_in_pic_flag = 1;
673     sh->slice_segment_address           = 0;
674
675     sh->slice_type = priv->slice_type;
676
677     sh->slice_pic_order_cnt_lsb = priv->pic_order_cnt &
678         (1 << (sps->log2_max_pic_order_cnt_lsb_minus4 + 4)) - 1;
679
680     if (pic->type != PICTURE_TYPE_IDR) {
681         H265RawSTRefPicSet *rps;
682         VAAPIEncodePicture *st;
683         int used;
684
685         sh->short_term_ref_pic_set_sps_flag = 0;
686
687         rps = &sh->short_term_ref_pic_set;
688         memset(rps, 0, sizeof(*rps));
689
690         for (st = ctx->pic_start; st; st = st->next) {
691             if (st->encode_order >= pic->encode_order) {
692                 // Not yet in DPB.
693                 continue;
694             }
695             used = 0;
696             for (i = 0; i < pic->nb_refs; i++) {
697                 if (pic->refs[i] == st)
698                     used = 1;
699             }
700             if (!used) {
701                 // Usually each picture always uses all of the others in the
702                 // DPB as references.  The one case we have to treat here is
703                 // a non-IDR IRAP picture, which may need to hold unused
704                 // references across itself to be used for the decoding of
705                 // following RASL pictures.  This looks for such an RASL
706                 // picture, and keeps the reference if there is one.
707                 VAAPIEncodePicture *rp;
708                 for (rp = ctx->pic_start; rp; rp = rp->next) {
709                     if (rp->encode_order < pic->encode_order)
710                         continue;
711                     if (rp->type != PICTURE_TYPE_B)
712                         continue;
713                     if (rp->refs[0] == st && rp->refs[1] == pic)
714                         break;
715                 }
716                 if (!rp)
717                     continue;
718             }
719             // This only works for one instance of each (delta_poc_sN_minus1
720             // is relative to the previous frame in the list, not relative to
721             // the current frame directly).
722             if (st->display_order < pic->display_order) {
723                 rps->delta_poc_s0_minus1[rps->num_negative_pics] =
724                     pic->display_order - st->display_order - 1;
725                 rps->used_by_curr_pic_s0_flag[rps->num_negative_pics] = used;
726                 ++rps->num_negative_pics;
727             } else {
728                 rps->delta_poc_s1_minus1[rps->num_positive_pics] =
729                     st->display_order - pic->display_order - 1;
730                 rps->used_by_curr_pic_s1_flag[rps->num_positive_pics] = used;
731                 ++rps->num_positive_pics;
732             }
733         }
734
735         sh->num_long_term_sps  = 0;
736         sh->num_long_term_pics = 0;
737
738         sh->slice_temporal_mvp_enabled_flag =
739             sps->sps_temporal_mvp_enabled_flag;
740         if (sh->slice_temporal_mvp_enabled_flag) {
741             sh->collocated_from_l0_flag = sh->slice_type == HEVC_SLICE_B;
742             sh->collocated_ref_idx      = 0;
743         }
744
745         sh->num_ref_idx_active_override_flag = 0;
746         sh->num_ref_idx_l0_active_minus1 = pps->num_ref_idx_l0_default_active_minus1;
747         sh->num_ref_idx_l1_active_minus1 = pps->num_ref_idx_l1_default_active_minus1;
748     }
749
750     sh->slice_sao_luma_flag = sh->slice_sao_chroma_flag =
751         sps->sample_adaptive_offset_enabled_flag;
752
753     if (pic->type == PICTURE_TYPE_B)
754         sh->slice_qp_delta = priv->fixed_qp_b - (pps->init_qp_minus26 + 26);
755     else if (pic->type == PICTURE_TYPE_P)
756         sh->slice_qp_delta = priv->fixed_qp_p - (pps->init_qp_minus26 + 26);
757     else
758         sh->slice_qp_delta = priv->fixed_qp_idr - (pps->init_qp_minus26 + 26);
759
760
761     *vslice = (VAEncSliceParameterBufferHEVC) {
762         .slice_segment_address = sh->slice_segment_address,
763         .num_ctu_in_slice      = priv->ctu_width * priv->ctu_height,
764
765         .slice_type                 = sh->slice_type,
766         .slice_pic_parameter_set_id = sh->slice_pic_parameter_set_id,
767
768         .num_ref_idx_l0_active_minus1 = sh->num_ref_idx_l0_active_minus1,
769         .num_ref_idx_l1_active_minus1 = sh->num_ref_idx_l1_active_minus1,
770
771         .luma_log2_weight_denom         = sh->luma_log2_weight_denom,
772         .delta_chroma_log2_weight_denom = sh->delta_chroma_log2_weight_denom,
773
774         .max_num_merge_cand = 5 - sh->five_minus_max_num_merge_cand,
775
776         .slice_qp_delta     = sh->slice_qp_delta,
777         .slice_cb_qp_offset = sh->slice_cb_qp_offset,
778         .slice_cr_qp_offset = sh->slice_cr_qp_offset,
779
780         .slice_beta_offset_div2 = sh->slice_beta_offset_div2,
781         .slice_tc_offset_div2   = sh->slice_tc_offset_div2,
782
783         .slice_fields.bits = {
784             .last_slice_of_pic_flag       = 1,
785             .dependent_slice_segment_flag = sh->dependent_slice_segment_flag,
786             .colour_plane_id              = sh->colour_plane_id,
787             .slice_temporal_mvp_enabled_flag =
788                 sh->slice_temporal_mvp_enabled_flag,
789             .slice_sao_luma_flag          = sh->slice_sao_luma_flag,
790             .slice_sao_chroma_flag        = sh->slice_sao_chroma_flag,
791             .num_ref_idx_active_override_flag =
792                 sh->num_ref_idx_active_override_flag,
793             .mvd_l1_zero_flag             = sh->mvd_l1_zero_flag,
794             .cabac_init_flag              = sh->cabac_init_flag,
795             .slice_deblocking_filter_disabled_flag =
796                 sh->slice_deblocking_filter_disabled_flag,
797             .slice_loop_filter_across_slices_enabled_flag =
798                 sh->slice_loop_filter_across_slices_enabled_flag,
799             .collocated_from_l0_flag      = sh->collocated_from_l0_flag,
800         },
801     };
802
803     for (i = 0; i < FF_ARRAY_ELEMS(vslice->ref_pic_list0); i++) {
804         vslice->ref_pic_list0[i].picture_id = VA_INVALID_ID;
805         vslice->ref_pic_list0[i].flags      = VA_PICTURE_HEVC_INVALID;
806         vslice->ref_pic_list1[i].picture_id = VA_INVALID_ID;
807         vslice->ref_pic_list1[i].flags      = VA_PICTURE_HEVC_INVALID;
808     }
809
810     av_assert0(pic->nb_refs <= 2);
811     if (pic->nb_refs >= 1) {
812         // Backward reference for P- or B-frame.
813         av_assert0(pic->type == PICTURE_TYPE_P ||
814                    pic->type == PICTURE_TYPE_B);
815         vslice->ref_pic_list0[0] = vpic->reference_frames[0];
816     }
817     if (pic->nb_refs >= 2) {
818         // Forward reference for B-frame.
819         av_assert0(pic->type == PICTURE_TYPE_B);
820         vslice->ref_pic_list1[0] = vpic->reference_frames[1];
821     }
822
823     return 0;
824 }
825
826 static av_cold int vaapi_encode_h265_configure(AVCodecContext *avctx)
827 {
828     VAAPIEncodeContext      *ctx = avctx->priv_data;
829     VAAPIEncodeH265Context *priv = ctx->priv_data;
830     VAAPIEncodeH265Options  *opt = ctx->codec_options;
831     int err;
832
833     err = ff_cbs_init(&priv->cbc, AV_CODEC_ID_HEVC, avctx);
834     if (err < 0)
835         return err;
836
837     priv->ctu_width     = FFALIGN(ctx->surface_width,  32) / 32;
838     priv->ctu_height    = FFALIGN(ctx->surface_height, 32) / 32;
839
840     av_log(avctx, AV_LOG_VERBOSE, "Input %ux%u -> Surface %ux%u -> CTU %ux%u.\n",
841            avctx->width, avctx->height, ctx->surface_width,
842            ctx->surface_height, priv->ctu_width, priv->ctu_height);
843
844     if (ctx->va_rc_mode == VA_RC_CQP) {
845         priv->fixed_qp_p = opt->qp;
846         if (avctx->i_quant_factor > 0.0)
847             priv->fixed_qp_idr = (int)((priv->fixed_qp_p * avctx->i_quant_factor +
848                                         avctx->i_quant_offset) + 0.5);
849         else
850             priv->fixed_qp_idr = priv->fixed_qp_p;
851         if (avctx->b_quant_factor > 0.0)
852             priv->fixed_qp_b = (int)((priv->fixed_qp_p * avctx->b_quant_factor +
853                                       avctx->b_quant_offset) + 0.5);
854         else
855             priv->fixed_qp_b = priv->fixed_qp_p;
856
857         av_log(avctx, AV_LOG_DEBUG, "Using fixed QP = "
858                "%d / %d / %d for IDR- / P- / B-frames.\n",
859                priv->fixed_qp_idr, priv->fixed_qp_p, priv->fixed_qp_b);
860
861     } else if (ctx->va_rc_mode == VA_RC_CBR ||
862                ctx->va_rc_mode == VA_RC_VBR) {
863         // These still need to be  set for pic_init_qp/slice_qp_delta.
864         priv->fixed_qp_idr = 30;
865         priv->fixed_qp_p   = 30;
866         priv->fixed_qp_b   = 30;
867
868         av_log(avctx, AV_LOG_DEBUG, "Using %s-bitrate = %"PRId64" bps.\n",
869                ctx->va_rc_mode == VA_RC_CBR ? "constant" : "variable",
870                avctx->bit_rate);
871
872     } else {
873         av_assert0(0 && "Invalid RC mode.");
874     }
875
876     return 0;
877 }
878
879 static const VAAPIEncodeType vaapi_encode_type_h265 = {
880     .priv_data_size        = sizeof(VAAPIEncodeH265Context),
881
882     .configure             = &vaapi_encode_h265_configure,
883
884     .sequence_params_size  = sizeof(VAEncSequenceParameterBufferHEVC),
885     .init_sequence_params  = &vaapi_encode_h265_init_sequence_params,
886
887     .picture_params_size   = sizeof(VAEncPictureParameterBufferHEVC),
888     .init_picture_params   = &vaapi_encode_h265_init_picture_params,
889
890     .slice_params_size     = sizeof(VAEncSliceParameterBufferHEVC),
891     .init_slice_params     = &vaapi_encode_h265_init_slice_params,
892
893     .sequence_header_type  = VAEncPackedHeaderSequence,
894     .write_sequence_header = &vaapi_encode_h265_write_sequence_header,
895
896     .slice_header_type     = VAEncPackedHeaderHEVC_Slice,
897     .write_slice_header    = &vaapi_encode_h265_write_slice_header,
898 };
899
900 static av_cold int vaapi_encode_h265_init(AVCodecContext *avctx)
901 {
902     VAAPIEncodeContext     *ctx = avctx->priv_data;
903     VAAPIEncodeH265Options *opt =
904         (VAAPIEncodeH265Options*)ctx->codec_options_data;
905
906     ctx->codec = &vaapi_encode_type_h265;
907
908     if (avctx->profile == FF_PROFILE_UNKNOWN)
909         avctx->profile = opt->profile;
910     if (avctx->level == FF_LEVEL_UNKNOWN)
911         avctx->level = opt->level;
912
913     switch (avctx->profile) {
914     case FF_PROFILE_HEVC_MAIN:
915     case FF_PROFILE_UNKNOWN:
916         ctx->va_profile = VAProfileHEVCMain;
917         ctx->va_rt_format = VA_RT_FORMAT_YUV420;
918         break;
919     case FF_PROFILE_HEVC_MAIN_10:
920 #ifdef VA_RT_FORMAT_YUV420_10BPP
921         ctx->va_profile = VAProfileHEVCMain10;
922         ctx->va_rt_format = VA_RT_FORMAT_YUV420_10BPP;
923         break;
924 #else
925         av_log(avctx, AV_LOG_ERROR, "10-bit encoding is not "
926                "supported with this VAAPI version.\n");
927         return AVERROR(ENOSYS);
928 #endif
929     default:
930         av_log(avctx, AV_LOG_ERROR, "Unknown H.265 profile %d.\n",
931                avctx->profile);
932         return AVERROR(EINVAL);
933     }
934     ctx->va_entrypoint = VAEntrypointEncSlice;
935
936     if (avctx->bit_rate > 0) {
937         if (avctx->rc_max_rate == avctx->bit_rate)
938             ctx->va_rc_mode = VA_RC_CBR;
939         else
940             ctx->va_rc_mode = VA_RC_VBR;
941     } else
942         ctx->va_rc_mode = VA_RC_CQP;
943
944     ctx->va_packed_headers =
945         VA_ENC_PACKED_HEADER_SEQUENCE | // VPS, SPS and PPS.
946         VA_ENC_PACKED_HEADER_SLICE;     // Slice headers.
947
948     ctx->surface_width  = FFALIGN(avctx->width,  16);
949     ctx->surface_height = FFALIGN(avctx->height, 16);
950
951     return ff_vaapi_encode_init(avctx);
952 }
953
954 static av_cold int vaapi_encode_h265_close(AVCodecContext *avctx)
955 {
956     VAAPIEncodeContext *ctx = avctx->priv_data;
957     VAAPIEncodeH265Context *priv = ctx->priv_data;
958
959     if (priv)
960         ff_cbs_close(&priv->cbc);
961
962     return ff_vaapi_encode_close(avctx);
963 }
964
965 #define OFFSET(x) (offsetof(VAAPIEncodeContext, codec_options_data) + \
966                    offsetof(VAAPIEncodeH265Options, x))
967 #define FLAGS (AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM)
968 static const AVOption vaapi_encode_h265_options[] = {
969     { "qp", "Constant QP (for P-frames; scaled by qfactor/qoffset for I/B)",
970       OFFSET(qp), AV_OPT_TYPE_INT, { .i64 = 25 }, 0, 52, FLAGS },
971
972     { "aud", "Include AUD",
973       OFFSET(aud), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, FLAGS },
974
975     { "profile", "Set profile (general_profile_idc)",
976       OFFSET(profile), AV_OPT_TYPE_INT,
977       { .i64 = FF_PROFILE_HEVC_MAIN }, 0x00, 0xff, FLAGS, "profile" },
978
979 #define PROFILE(name, value)  name, NULL, 0, AV_OPT_TYPE_CONST, \
980       { .i64 = value }, 0, 0, FLAGS, "profile"
981     { PROFILE("main",               FF_PROFILE_HEVC_MAIN) },
982     { PROFILE("main10",             FF_PROFILE_HEVC_MAIN_10) },
983 #undef PROFILE
984
985     { "level", "Set level (general_level_idc)",
986       OFFSET(level), AV_OPT_TYPE_INT,
987       { .i64 = 153 }, 0x00, 0xff, FLAGS, "level" },
988
989 #define LEVEL(name, value) name, NULL, 0, AV_OPT_TYPE_CONST, \
990       { .i64 = value }, 0, 0, FLAGS, "level"
991     { LEVEL("1",    30) },
992     { LEVEL("2",    60) },
993     { LEVEL("2.1",  63) },
994     { LEVEL("3",    90) },
995     { LEVEL("3.1",  93) },
996     { LEVEL("4",   120) },
997     { LEVEL("4.1", 123) },
998     { LEVEL("5",   150) },
999     { LEVEL("5.1", 153) },
1000     { LEVEL("5.2", 156) },
1001     { LEVEL("6",   180) },
1002     { LEVEL("6.1", 183) },
1003     { LEVEL("6.2", 186) },
1004 #undef LEVEL
1005
1006     { NULL },
1007 };
1008
1009 static const AVCodecDefault vaapi_encode_h265_defaults[] = {
1010     { "b",              "0"   },
1011     { "bf",             "2"   },
1012     { "g",              "120" },
1013     { "i_qfactor",      "1"   },
1014     { "i_qoffset",      "0"   },
1015     { "b_qfactor",      "6/5" },
1016     { "b_qoffset",      "0"   },
1017     { NULL },
1018 };
1019
1020 static const AVClass vaapi_encode_h265_class = {
1021     .class_name = "h265_vaapi",
1022     .item_name  = av_default_item_name,
1023     .option     = vaapi_encode_h265_options,
1024     .version    = LIBAVUTIL_VERSION_INT,
1025 };
1026
1027 AVCodec ff_hevc_vaapi_encoder = {
1028     .name           = "hevc_vaapi",
1029     .long_name      = NULL_IF_CONFIG_SMALL("H.265/HEVC (VAAPI)"),
1030     .type           = AVMEDIA_TYPE_VIDEO,
1031     .id             = AV_CODEC_ID_HEVC,
1032     .priv_data_size = (sizeof(VAAPIEncodeContext) +
1033                        sizeof(VAAPIEncodeH265Options)),
1034     .init           = &vaapi_encode_h265_init,
1035     .encode2        = &ff_vaapi_encode2,
1036     .close          = &vaapi_encode_h265_close,
1037     .priv_class     = &vaapi_encode_h265_class,
1038     .capabilities   = AV_CODEC_CAP_DELAY | AV_CODEC_CAP_HARDWARE,
1039     .defaults       = vaapi_encode_h265_defaults,
1040     .pix_fmts = (const enum AVPixelFormat[]) {
1041         AV_PIX_FMT_VAAPI,
1042         AV_PIX_FMT_NONE,
1043     },
1044     .wrapper_name   = "vaapi",
1045 };