git.sesse.net Git - ffmpeg/blob - 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 <va/va.h>
  20 #include <va/va_enc_hevc.h>
  21
  22 #include "libavutil/avassert.h"
  23 #include "libavutil/internal.h"
  24 #include "libavutil/opt.h"
  25 #include "libavutil/pixfmt.h"
  26
  27 #include "avcodec.h"
  28 #include "hevc.h"
  29 #include "internal.h"
  30 #include "put_bits.h"
  31 #include "vaapi_encode.h"
  32 #include "vaapi_encode_h26x.h"
  33
  34
  35 #define MAX_ST_REF_PIC_SETS  32
  36 #define MAX_DPB_PICS         16
  37 #define MAX_LAYERS            1
  38
  39
  40 typedef struct VAAPIEncodeH265STRPS {
  41     char inter_ref_pic_set_prediction_flag;
  42
  43     unsigned int num_negative_pics;
  44     unsigned int num_positive_pics;
  45
  46     unsigned int delta_poc_s0_minus1[MAX_DPB_PICS];
  47     char used_by_curr_pic_s0_flag[MAX_DPB_PICS];
  48
  49     unsigned int delta_poc_s1_minus1[MAX_DPB_PICS];
  50     char used_by_curr_pic_s1_flag[MAX_DPB_PICS];
  51 } VAAPIEncodeH265STRPS;
  52
  53 // This structure contains all possibly-useful per-sequence syntax elements
  54 // which are not already contained in the various VAAPI structures.
  55 typedef struct VAAPIEncodeH265MiscSequenceParams {
  56
  57     // Parameter set IDs.
  58     unsigned int video_parameter_set_id;
  59     unsigned int seq_parameter_set_id;
  60
  61     // Layering.
  62     unsigned int vps_max_layers_minus1;
  63     unsigned int vps_max_sub_layers_minus1;
  64     char vps_temporal_id_nesting_flag;
  65     unsigned int vps_max_layer_id;
  66     unsigned int vps_num_layer_sets_minus1;
  67     unsigned int sps_max_sub_layers_minus1;
  68     char sps_temporal_id_nesting_flag;
  69     char layer_id_included_flag[MAX_LAYERS][64];
  70
  71     // Profile/tier/level parameters.
  72     char general_profile_compatibility_flag[32];
  73     char general_progressive_source_flag;
  74     char general_interlaced_source_flag;
  75     char general_non_packed_constraint_flag;
  76     char general_frame_only_constraint_flag;
  77     char general_inbld_flag;
  78
  79     // Decode/display ordering parameters.
  80     unsigned int log2_max_pic_order_cnt_lsb_minus4;
  81     char vps_sub_layer_ordering_info_present_flag;
  82     unsigned int vps_max_dec_pic_buffering_minus1[MAX_LAYERS];
  83     unsigned int vps_max_num_reorder_pics[MAX_LAYERS];
  84     unsigned int vps_max_latency_increase_plus1[MAX_LAYERS];
  85     char sps_sub_layer_ordering_info_present_flag;
  86     unsigned int sps_max_dec_pic_buffering_minus1[MAX_LAYERS];
  87     unsigned int sps_max_num_reorder_pics[MAX_LAYERS];
  88     unsigned int sps_max_latency_increase_plus1[MAX_LAYERS];
  89
  90     // Timing information.
  91     char vps_timing_info_present_flag;
  92     unsigned int vps_num_units_in_tick;
  93     unsigned int vps_time_scale;
  94     char vps_poc_proportional_to_timing_flag;
  95     unsigned int vps_num_ticks_poc_diff_minus1;
  96
  97     // Cropping information.
  98     char conformance_window_flag;
  99     unsigned int conf_win_left_offset;
 100     unsigned int conf_win_right_offset;
 101     unsigned int conf_win_top_offset;
 102     unsigned int conf_win_bottom_offset;
 103
 104     // Short-term reference picture sets.
 105     unsigned int num_short_term_ref_pic_sets;
 106     VAAPIEncodeH265STRPS st_ref_pic_set[MAX_ST_REF_PIC_SETS];
 107
 108     // Long-term reference pictures.
 109     char long_term_ref_pics_present_flag;
 110     unsigned int num_long_term_ref_pics_sps;
 111     struct {
 112         unsigned int lt_ref_pic_poc_lsb_sps;
 113         char used_by_curr_pic_lt_sps_flag;
 114     } lt_ref_pic;
 115
 116     // Deblocking filter control.
 117     char deblocking_filter_control_present_flag;
 118     char deblocking_filter_override_enabled_flag;
 119     char pps_deblocking_filter_disabled_flag;
 120     int pps_beta_offset_div2;
 121     int pps_tc_offset_div2;
 122
 123     // Video Usability Information.
 124     char vui_parameters_present_flag;
 125     char aspect_ratio_info_present_flag;
 126     unsigned int aspect_ratio_idc;
 127     unsigned int sar_width;
 128     unsigned int sar_height;
 129     char video_signal_type_present_flag;
 130     unsigned int video_format;
 131     char video_full_range_flag;
 132     char colour_description_present_flag;
 133     unsigned int colour_primaries;
 134     unsigned int transfer_characteristics;
 135     unsigned int matrix_coeffs;
 136
 137     // Oddments.
 138     char uniform_spacing_flag;
 139     char output_flag_present_flag;
 140     char cabac_init_present_flag;
 141     unsigned int num_extra_slice_header_bits;
 142     char lists_modification_present_flag;
 143     char pps_slice_chroma_qp_offsets_present_flag;
 144     char pps_slice_chroma_offset_list_enabled_flag;
 145 } VAAPIEncodeH265MiscSequenceParams;
 146
 147 // This structure contains all possibly-useful per-slice syntax elements
 148 // which are not already contained in the various VAAPI structures.
 149 typedef struct VAAPIEncodeH265MiscSliceParams {
 150     // Slice segments.
 151     char first_slice_segment_in_pic_flag;
 152     unsigned int slice_segment_address;
 153
 154     // Short-term reference picture sets.
 155     char short_term_ref_pic_set_sps_flag;
 156     unsigned int short_term_ref_pic_idx;
 157     VAAPIEncodeH265STRPS st_ref_pic_set;
 158
 159     // Deblocking filter.
 160     char deblocking_filter_override_flag;
 161
 162     // Oddments.
 163     char slice_reserved_flag[8];
 164     char no_output_of_prior_pics_flag;
 165     char pic_output_flag;
 166 } VAAPIEncodeH265MiscSliceParams;
 167
 168 typedef struct VAAPIEncodeH265Slice {
 169     VAAPIEncodeH265MiscSliceParams misc_slice_params;
 170
 171     int64_t pic_order_cnt;
 172 } VAAPIEncodeH265Slice;
 173
 174 typedef struct VAAPIEncodeH265Context {
 175     VAAPIEncodeH265MiscSequenceParams misc_sequence_params;
 176
 177     unsigned int ctu_width;
 178     unsigned int ctu_height;
 179
 180     int fixed_qp_idr;
 181     int fixed_qp_p;
 182     int fixed_qp_b;
 183
 184     int64_t last_idr_frame;
 185
 186     // Rate control configuration.
 187     struct {
 188         VAEncMiscParameterBuffer misc;
 189         VAEncMiscParameterRateControl rc;
 190     } rc_params;
 191     struct {
 192         VAEncMiscParameterBuffer misc;
 193         VAEncMiscParameterHRD hrd;
 194     } hrd_params;
 195 } VAAPIEncodeH265Context;
 196
 197 typedef struct VAAPIEncodeH265Options {
 198     int qp;
 199 } VAAPIEncodeH265Options;
 200
 201
 202 #define vseq_var(name)     vseq->name, name
 203 #define vseq_field(name)   vseq->seq_fields.bits.name, name
 204 #define vpic_var(name)     vpic->name, name
 205 #define vpic_field(name)   vpic->pic_fields.bits.name, name
 206 #define vslice_var(name)   vslice->name, name
 207 #define vslice_field(name) vslice->slice_fields.bits.name, name
 208 #define mseq_var(name)     mseq->name, name
 209 #define mslice_var(name)   mslice->name, name
 210 #define mstrps_var(name)   mstrps->name, name
 211
 212 static void vaapi_encode_h265_write_nal_unit_header(PutBitContext *pbc,
 213                                                     int nal_unit_type)
 214 {
 215     u(1, 0, forbidden_zero_bit);
 216     u(6, nal_unit_type, nal_unit_type);
 217     u(6, 0, nuh_layer_id);
 218     u(3, 1, nuh_temporal_id_plus1);
 219 }
 220
 221 static void vaapi_encode_h265_write_rbsp_trailing_bits(PutBitContext *pbc)
 222 {
 223     u(1, 1, rbsp_stop_one_bit);
 224     while (put_bits_count(pbc) & 7)
 225         u(1, 0, rbsp_alignment_zero_bit);
 226 }
 227
 228 static void vaapi_encode_h265_write_profile_tier_level(PutBitContext *pbc,
 229                                                        VAAPIEncodeContext *ctx)
 230 {
 231     VAEncSequenceParameterBufferHEVC  *vseq = ctx->codec_sequence_params;
 232     VAAPIEncodeH265Context            *priv = ctx->priv_data;
 233     VAAPIEncodeH265MiscSequenceParams *mseq = &priv->misc_sequence_params;
 234     int j;
 235
 236     if (1) {
 237         u(2, 0, general_profile_space);
 238         u(1, vseq_var(general_tier_flag));
 239         u(5, vseq_var(general_profile_idc));
 240
 241         for (j = 0; j < 32; j++) {
 242             u(1, mseq_var(general_profile_compatibility_flag[j]));
 243         }
 244
 245         u(1, mseq_var(general_progressive_source_flag));
 246         u(1, mseq_var(general_interlaced_source_flag));
 247         u(1, mseq_var(general_non_packed_constraint_flag));
 248         u(1, mseq_var(general_frame_only_constraint_flag));
 249
 250         if (0) {
 251             // Not main profile.
 252             // Lots of extra constraint flags.
 253         } else {
 254             // put_bits only handles up to 31 bits.
 255             u(23, 0, general_reserved_zero_43bits);
 256             u(20, 0, general_reserved_zero_43bits);
 257         }
 258
 259         if (vseq->general_profile_idc >= 1 && vseq->general_profile_idc <= 5) {
 260             u(1, mseq_var(general_inbld_flag));
 261         } else {
 262             u(1, 0, general_reserved_zero_bit);
 263         }
 264     }
 265
 266     u(8, vseq_var(general_level_idc));
 267
 268     // No sublayers.
 269 }
 270
 271 static void vaapi_encode_h265_write_vps(PutBitContext *pbc,
 272                                         VAAPIEncodeContext *ctx)
 273 {
 274     VAAPIEncodeH265Context            *priv = ctx->priv_data;
 275     VAAPIEncodeH265MiscSequenceParams *mseq = &priv->misc_sequence_params;
 276     int i, j;
 277
 278     vaapi_encode_h265_write_nal_unit_header(pbc, NAL_VPS);
 279
 280     u(4, mseq->video_parameter_set_id, vps_video_parameter_set_id);
 281
 282     u(1, 1, vps_base_layer_internal_flag);
 283     u(1, 1, vps_base_layer_available_flag);
 284     u(6, mseq_var(vps_max_layers_minus1));
 285     u(3, mseq_var(vps_max_sub_layers_minus1));
 286     u(1, mseq_var(vps_temporal_id_nesting_flag));
 287
 288     u(16, 0xffff, vps_reserved_0xffff_16bits);
 289
 290     vaapi_encode_h265_write_profile_tier_level(pbc, ctx);
 291
 292     u(1, mseq_var(vps_sub_layer_ordering_info_present_flag));
 293     for (i = (mseq->vps_sub_layer_ordering_info_present_flag ?
 294               0 : mseq->vps_max_sub_layers_minus1);
 295          i <= mseq->vps_max_sub_layers_minus1; i++) {
 296         ue(mseq_var(vps_max_dec_pic_buffering_minus1[i]));
 297         ue(mseq_var(vps_max_num_reorder_pics[i]));
 298         ue(mseq_var(vps_max_latency_increase_plus1[i]));
 299     }
 300
 301     u(6, mseq_var(vps_max_layer_id));
 302     ue(mseq_var(vps_num_layer_sets_minus1));
 303     for (i = 1; i <= mseq->vps_num_layer_sets_minus1; i++) {
 304         for (j = 0; j < mseq->vps_max_layer_id; j++)
 305             u(1, mseq_var(layer_id_included_flag[i][j]));
 306     }
 307
 308     u(1, mseq_var(vps_timing_info_present_flag));
 309     if (mseq->vps_timing_info_present_flag) {
 310         u(1, 0, put_bits_hack_zero_bit);
 311         u(31, mseq_var(vps_num_units_in_tick));
 312         u(1, 0, put_bits_hack_zero_bit);
 313         u(31, mseq_var(vps_time_scale));
 314         u(1, mseq_var(vps_poc_proportional_to_timing_flag));
 315         if (mseq->vps_poc_proportional_to_timing_flag) {
 316             ue(mseq_var(vps_num_ticks_poc_diff_minus1));
 317         }
 318         ue(0, vps_num_hrd_parameters);
 319     }
 320
 321     u(1, 0, vps_extension_flag);
 322
 323     vaapi_encode_h265_write_rbsp_trailing_bits(pbc);
 324 }
 325
 326 static void vaapi_encode_h265_write_st_ref_pic_set(PutBitContext *pbc,
 327                                                    int st_rps_idx,
 328                                                    VAAPIEncodeH265STRPS *mstrps)
 329 {
 330     int i;
 331
 332     if (st_rps_idx != 0)
 333        u(1, mstrps_var(inter_ref_pic_set_prediction_flag));
 334
 335     if (mstrps->inter_ref_pic_set_prediction_flag) {
 336         av_assert0(0 && "inter ref pic set prediction not supported");
 337     } else {
 338         ue(mstrps_var(num_negative_pics));
 339         ue(mstrps_var(num_positive_pics));
 340
 341         for (i = 0; i < mstrps->num_negative_pics; i++) {
 342             ue(mstrps_var(delta_poc_s0_minus1[i]));
 343             u(1, mstrps_var(used_by_curr_pic_s0_flag[i]));
 344         }
 345         for (i = 0; i < mstrps->num_positive_pics; i++) {
 346             ue(mstrps_var(delta_poc_s1_minus1[i]));
 347             u(1, mstrps_var(used_by_curr_pic_s1_flag[i]));
 348         }
 349     }
 350 }
 351
 352 static void vaapi_encode_h265_write_vui_parameters(PutBitContext *pbc,
 353                                                    VAAPIEncodeContext *ctx)
 354 {
 355     VAAPIEncodeH265Context            *priv = ctx->priv_data;
 356     VAAPIEncodeH265MiscSequenceParams *mseq = &priv->misc_sequence_params;
 357
 358     u(1, mseq_var(aspect_ratio_info_present_flag));
 359     if (mseq->aspect_ratio_info_present_flag) {
 360         u(8, mseq_var(aspect_ratio_idc));
 361         if (mseq->aspect_ratio_idc == 255) {
 362             u(16, mseq_var(sar_width));
 363             u(16, mseq_var(sar_height));
 364         }
 365     }
 366
 367     u(1, 0, overscan_info_present_flag);
 368
 369     u(1, mseq_var(video_signal_type_present_flag));
 370     if (mseq->video_signal_type_present_flag) {
 371         u(3, mseq_var(video_format));
 372         u(1, mseq_var(video_full_range_flag));
 373         u(1, mseq_var(colour_description_present_flag));
 374         if (mseq->colour_description_present_flag) {
 375             u(8, mseq_var(colour_primaries));
 376             u(8, mseq_var(transfer_characteristics));
 377             u(8, mseq_var(matrix_coeffs));
 378         }
 379     }
 380
 381     u(1, 0, chroma_loc_info_present_flag);
 382     u(1, 0, neutral_chroma_indication_flag);
 383     u(1, 0, field_seq_flag);
 384     u(1, 0, frame_field_info_present_flag);
 385     u(1, 0, default_display_window_flag);
 386     u(1, 0, vui_timing_info_present_flag);
 387     u(1, 0, bitstream_restriction_flag_flag);
 388 }
 389
 390 static void vaapi_encode_h265_write_sps(PutBitContext *pbc,
 391                                         VAAPIEncodeContext *ctx)
 392 {
 393     VAEncSequenceParameterBufferHEVC  *vseq = ctx->codec_sequence_params;
 394     VAAPIEncodeH265Context            *priv = ctx->priv_data;
 395     VAAPIEncodeH265MiscSequenceParams *mseq = &priv->misc_sequence_params;
 396     int i;
 397
 398     vaapi_encode_h265_write_nal_unit_header(pbc, NAL_SPS);
 399
 400     u(4, mseq->video_parameter_set_id, sps_video_parameter_set_id);
 401
 402     u(3, mseq_var(sps_max_sub_layers_minus1));
 403     u(1, mseq_var(sps_temporal_id_nesting_flag));
 404
 405     vaapi_encode_h265_write_profile_tier_level(pbc, ctx);
 406
 407     ue(mseq->seq_parameter_set_id, sps_seq_parameter_set_id);
 408     ue(vseq_field(chroma_format_idc));
 409     if (vseq->seq_fields.bits.chroma_format_idc == 3)
 410         u(1, 0, separate_colour_plane_flag);
 411
 412     ue(vseq_var(pic_width_in_luma_samples));
 413     ue(vseq_var(pic_height_in_luma_samples));
 414
 415     u(1, mseq_var(conformance_window_flag));
 416     if (mseq->conformance_window_flag) {
 417         ue(mseq_var(conf_win_left_offset));
 418         ue(mseq_var(conf_win_right_offset));
 419         ue(mseq_var(conf_win_top_offset));
 420         ue(mseq_var(conf_win_bottom_offset));
 421     }
 422
 423     ue(vseq_field(bit_depth_luma_minus8));
 424     ue(vseq_field(bit_depth_chroma_minus8));
 425
 426     ue(mseq_var(log2_max_pic_order_cnt_lsb_minus4));
 427
 428     u(1, mseq_var(sps_sub_layer_ordering_info_present_flag));
 429     for (i = (mseq->sps_sub_layer_ordering_info_present_flag ?
 430               0 : mseq->sps_max_sub_layers_minus1);
 431          i <= mseq->sps_max_sub_layers_minus1; i++) {
 432         ue(mseq_var(sps_max_dec_pic_buffering_minus1[i]));
 433         ue(mseq_var(sps_max_num_reorder_pics[i]));
 434         ue(mseq_var(sps_max_latency_increase_plus1[i]));
 435     }
 436
 437     ue(vseq_var(log2_min_luma_coding_block_size_minus3));
 438     ue(vseq_var(log2_diff_max_min_luma_coding_block_size));
 439     ue(vseq_var(log2_min_transform_block_size_minus2));
 440     ue(vseq_var(log2_diff_max_min_transform_block_size));
 441     ue(vseq_var(max_transform_hierarchy_depth_inter));
 442     ue(vseq_var(max_transform_hierarchy_depth_intra));
 443
 444     u(1, vseq_field(scaling_list_enabled_flag));
 445     if (vseq->seq_fields.bits.scaling_list_enabled_flag) {
 446         u(1, 0, sps_scaling_list_data_present_flag);
 447     }
 448
 449     u(1, vseq_field(amp_enabled_flag));
 450     u(1, vseq_field(sample_adaptive_offset_enabled_flag));
 451
 452     u(1, vseq_field(pcm_enabled_flag));
 453     if (vseq->seq_fields.bits.pcm_enabled_flag) {
 454         u(4, vseq_var(pcm_sample_bit_depth_luma_minus1));
 455         u(4, vseq_var(pcm_sample_bit_depth_chroma_minus1));
 456         ue(vseq_var(log2_min_pcm_luma_coding_block_size_minus3));
 457         ue(vseq->log2_max_pcm_luma_coding_block_size_minus3 -
 458            vseq->log2_min_pcm_luma_coding_block_size_minus3,
 459            log2_diff_max_min_pcm_luma_coding_block_size);
 460         u(1, vseq_field(pcm_loop_filter_disabled_flag));
 461     }
 462
 463     ue(mseq_var(num_short_term_ref_pic_sets));
 464     for (i = 0; i < mseq->num_short_term_ref_pic_sets; i++)
 465         vaapi_encode_h265_write_st_ref_pic_set(pbc, i,
 466                                                &mseq->st_ref_pic_set[i]);
 467
 468     u(1, mseq_var(long_term_ref_pics_present_flag));
 469     if (mseq->long_term_ref_pics_present_flag) {
 470         ue(0, num_long_term_ref_pics_sps);
 471     }
 472
 473     u(1, vseq_field(sps_temporal_mvp_enabled_flag));
 474     u(1, vseq_field(strong_intra_smoothing_enabled_flag));
 475
 476     u(1, mseq_var(vui_parameters_present_flag));
 477     if (mseq->vui_parameters_present_flag) {
 478         vaapi_encode_h265_write_vui_parameters(pbc, ctx);
 479     }
 480
 481     u(1, 0, sps_extension_present_flag);
 482
 483     vaapi_encode_h265_write_rbsp_trailing_bits(pbc);
 484 }
 485
 486 static void vaapi_encode_h265_write_pps(PutBitContext *pbc,
 487                                         VAAPIEncodeContext *ctx)
 488 {
 489     VAEncPictureParameterBufferHEVC   *vpic = ctx->codec_picture_params;
 490     VAAPIEncodeH265Context            *priv = ctx->priv_data;
 491     VAAPIEncodeH265MiscSequenceParams *mseq = &priv->misc_sequence_params;
 492     int i;
 493
 494     vaapi_encode_h265_write_nal_unit_header(pbc, NAL_PPS);
 495
 496     ue(vpic->slice_pic_parameter_set_id, pps_pic_parameter_set_id);
 497     ue(mseq->seq_parameter_set_id, pps_seq_parameter_set_id);
 498
 499     u(1, vpic_field(dependent_slice_segments_enabled_flag));
 500     u(1, mseq_var(output_flag_present_flag));
 501     u(3, mseq_var(num_extra_slice_header_bits));
 502     u(1, vpic_field(sign_data_hiding_enabled_flag));
 503     u(1, mseq_var(cabac_init_present_flag));
 504
 505     ue(vpic_var(num_ref_idx_l0_default_active_minus1));
 506     ue(vpic_var(num_ref_idx_l1_default_active_minus1));
 507
 508     se(vpic->pic_init_qp - 26, init_qp_minus26);
 509
 510     u(1, vpic_field(constrained_intra_pred_flag));
 511     u(1, vpic_field(transform_skip_enabled_flag));
 512
 513     u(1, vpic_field(cu_qp_delta_enabled_flag));
 514     if (vpic->pic_fields.bits.cu_qp_delta_enabled_flag)
 515         ue(vpic_var(diff_cu_qp_delta_depth));
 516
 517     se(vpic_var(pps_cb_qp_offset));
 518     se(vpic_var(pps_cr_qp_offset));
 519
 520     u(1, mseq_var(pps_slice_chroma_qp_offsets_present_flag));
 521     u(1, vpic_field(weighted_pred_flag));
 522     u(1, vpic_field(weighted_bipred_flag));
 523     u(1, vpic_field(transquant_bypass_enabled_flag));
 524     u(1, vpic_field(tiles_enabled_flag));
 525     u(1, vpic_field(entropy_coding_sync_enabled_flag));
 526
 527     if (vpic->pic_fields.bits.tiles_enabled_flag) {
 528         ue(vpic_var(num_tile_columns_minus1));
 529         ue(vpic_var(num_tile_rows_minus1));
 530         u(1, mseq_var(uniform_spacing_flag));
 531         if (!mseq->uniform_spacing_flag) {
 532             for (i = 0; i < vpic->num_tile_columns_minus1; i++)
 533                 ue(vpic_var(column_width_minus1[i]));
 534             for (i = 0; i < vpic->num_tile_rows_minus1; i++)
 535                 ue(vpic_var(row_height_minus1[i]));
 536         }
 537         u(1, vpic_field(loop_filter_across_tiles_enabled_flag));
 538     }
 539
 540     u(1, vpic_field(pps_loop_filter_across_slices_enabled_flag));
 541     u(1, mseq_var(deblocking_filter_control_present_flag));
 542     if (mseq->deblocking_filter_control_present_flag) {
 543         u(1, mseq_var(deblocking_filter_override_enabled_flag));
 544         u(1, mseq_var(pps_deblocking_filter_disabled_flag));
 545         if (!mseq->pps_deblocking_filter_disabled_flag) {
 546             se(mseq_var(pps_beta_offset_div2));
 547             se(mseq_var(pps_tc_offset_div2));
 548         }
 549     }
 550
 551     u(1, 0, pps_scaling_list_data_present_flag);
 552     // No scaling list data.
 553
 554     u(1, mseq_var(lists_modification_present_flag));
 555     ue(vpic_var(log2_parallel_merge_level_minus2));
 556     u(1, 0, slice_segment_header_extension_present_flag);
 557     u(1, 0, pps_extension_present_flag);
 558
 559     vaapi_encode_h265_write_rbsp_trailing_bits(pbc);
 560 }
 561
 562 static void vaapi_encode_h265_write_slice_header2(PutBitContext *pbc,
 563                                                   VAAPIEncodeContext *ctx,
 564                                                   VAAPIEncodePicture *pic,
 565                                                   VAAPIEncodeSlice *slice)
 566 {
 567     VAEncSequenceParameterBufferHEVC  *vseq = ctx->codec_sequence_params;
 568     VAEncPictureParameterBufferHEVC   *vpic = pic->codec_picture_params;
 569     VAEncSliceParameterBufferHEVC   *vslice = slice->codec_slice_params;
 570     VAAPIEncodeH265Context            *priv = ctx->priv_data;
 571     VAAPIEncodeH265MiscSequenceParams *mseq = &priv->misc_sequence_params;
 572     VAAPIEncodeH265Slice            *pslice = slice->priv_data;
 573     VAAPIEncodeH265MiscSliceParams  *mslice = &pslice->misc_slice_params;
 574     int i;
 575
 576     vaapi_encode_h265_write_nal_unit_header(pbc, vpic->nal_unit_type);
 577
 578     u(1, mslice_var(first_slice_segment_in_pic_flag));
 579     if (vpic->nal_unit_type >= NAL_BLA_W_LP &&
 580        vpic->nal_unit_type <= 23)
 581         u(1, mslice_var(no_output_of_prior_pics_flag));
 582
 583     ue(vslice_var(slice_pic_parameter_set_id));
 584
 585     if (!mslice->first_slice_segment_in_pic_flag) {
 586         if (vpic->pic_fields.bits.dependent_slice_segments_enabled_flag)
 587             u(1, vslice_field(dependent_slice_segment_flag));
 588         u(av_log2((priv->ctu_width * priv->ctu_height) - 1) + 1,
 589           mslice_var(slice_segment_address));
 590     }
 591     if (!vslice->slice_fields.bits.dependent_slice_segment_flag) {
 592         for (i = 0; i < mseq->num_extra_slice_header_bits; i++)
 593             u(1, mslice_var(slice_reserved_flag[i]));
 594
 595         ue(vslice_var(slice_type));
 596         if (mseq->output_flag_present_flag)
 597             u(1, 1, pic_output_flag);
 598         if (vseq->seq_fields.bits.separate_colour_plane_flag)
 599             u(2, vslice_field(colour_plane_id));
 600         if (vpic->nal_unit_type != NAL_IDR_W_RADL &&
 601            vpic->nal_unit_type != NAL_IDR_N_LP) {
 602             u(4 + mseq->log2_max_pic_order_cnt_lsb_minus4,
 603               (pslice->pic_order_cnt &
 604                ((1 << (mseq->log2_max_pic_order_cnt_lsb_minus4 + 4)) - 1)),
 605               slice_pic_order_cnt_lsb);
 606
 607             u(1, mslice_var(short_term_ref_pic_set_sps_flag));
 608             if (!mslice->short_term_ref_pic_set_sps_flag) {
 609                 vaapi_encode_h265_write_st_ref_pic_set(pbc, mseq->num_short_term_ref_pic_sets,
 610                                                        &mslice->st_ref_pic_set);
 611             } else if (mseq->num_short_term_ref_pic_sets > 1) {
 612                 u(av_log2(mseq->num_short_term_ref_pic_sets - 1) + 1,
 613                   mslice_var(short_term_ref_pic_idx));
 614             }
 615
 616             if (mseq->long_term_ref_pics_present_flag) {
 617                 av_assert0(0);
 618             }
 619
 620             if (vseq->seq_fields.bits.sps_temporal_mvp_enabled_flag) {
 621                 u(1, vslice_field(slice_temporal_mvp_enabled_flag));
 622             }
 623
 624             if (vseq->seq_fields.bits.sample_adaptive_offset_enabled_flag) {
 625                 u(1, vslice_field(slice_sao_luma_flag));
 626                 if (!vseq->seq_fields.bits.separate_colour_plane_flag &&
 627                    vseq->seq_fields.bits.chroma_format_idc != 0) {
 628                     u(1, vslice_field(slice_sao_chroma_flag));
 629                 }
 630             }
 631
 632             if (vslice->slice_type == P_SLICE || vslice->slice_type == B_SLICE) {
 633                 u(1, vslice_field(num_ref_idx_active_override_flag));
 634                 if (vslice->slice_fields.bits.num_ref_idx_active_override_flag) {
 635                     ue(vslice_var(num_ref_idx_l0_active_minus1));
 636                     if (vslice->slice_type == B_SLICE) {
 637                         ue(vslice_var(num_ref_idx_l1_active_minus1));
 638                     }
 639                 }
 640
 641                 if (mseq->lists_modification_present_flag) {
 642                     av_assert0(0);
 643                     // ref_pic_lists_modification()
 644                 }
 645                 if (vslice->slice_type == B_SLICE) {
 646                     u(1, vslice_field(mvd_l1_zero_flag));
 647                 }
 648                 if (mseq->cabac_init_present_flag) {
 649                     u(1, vslice_field(cabac_init_flag));
 650                 }
 651                 if (vslice->slice_fields.bits.slice_temporal_mvp_enabled_flag) {
 652                     if (vslice->slice_type == B_SLICE)
 653                         u(1, vslice_field(collocated_from_l0_flag));
 654                     ue(vpic->collocated_ref_pic_index, collocated_ref_idx);
 655                 }
 656                 if ((vpic->pic_fields.bits.weighted_pred_flag &&
 657                      vslice->slice_type == P_SLICE) ||
 658                     (vpic->pic_fields.bits.weighted_bipred_flag &&
 659                      vslice->slice_type == B_SLICE)) {
 660                     av_assert0(0);
 661                     // pred_weight_table()
 662                 }
 663                 ue(5 - vslice->max_num_merge_cand, five_minus_max_num_merge_cand);
 664             }
 665
 666             se(vslice_var(slice_qp_delta));
 667             if (mseq->pps_slice_chroma_qp_offsets_present_flag) {
 668                 se(vslice_var(slice_cb_qp_offset));
 669                 se(vslice_var(slice_cr_qp_offset));
 670             }
 671             if (mseq->pps_slice_chroma_offset_list_enabled_flag) {
 672                 u(1, 0, cu_chroma_qp_offset_enabled_flag);
 673             }
 674             if (mseq->deblocking_filter_override_enabled_flag) {
 675                 u(1, mslice_var(deblocking_filter_override_flag));
 676             }
 677             if (mslice->deblocking_filter_override_flag) {
 678                 u(1, vslice_field(slice_deblocking_filter_disabled_flag));
 679                 if (!vslice->slice_fields.bits.slice_deblocking_filter_disabled_flag) {
 680                     se(vslice_var(slice_beta_offset_div2));
 681                     se(vslice_var(slice_tc_offset_div2));
 682                 }
 683             }
 684             if (vpic->pic_fields.bits.pps_loop_filter_across_slices_enabled_flag &&
 685                 (vslice->slice_fields.bits.slice_sao_luma_flag ||
 686                  vslice->slice_fields.bits.slice_sao_chroma_flag ||
 687                  vslice->slice_fields.bits.slice_deblocking_filter_disabled_flag)) {
 688                 u(1, vslice_field(slice_loop_filter_across_slices_enabled_flag));
 689             }
 690         }
 691
 692         if (vpic->pic_fields.bits.tiles_enabled_flag ||
 693             vpic->pic_fields.bits.entropy_coding_sync_enabled_flag) {
 694             // num_entry_point_offsets
 695         }
 696
 697         if (0) {
 698             // slice_segment_header_extension_length
 699         }
 700     }
 701
 702     u(1, 1, alignment_bit_equal_to_one);
 703     while (put_bits_count(pbc) & 7)
 704         u(1, 0, alignment_bit_equal_to_zero);
 705 }
 706
 707 static int vaapi_encode_h265_write_sequence_header(AVCodecContext *avctx,
 708                                                    char *data, size_t *data_len)
 709 {
 710     VAAPIEncodeContext *ctx = avctx->priv_data;
 711     PutBitContext pbc;
 712     char tmp[256];
 713     int err;
 714     size_t nal_len, bit_len, bit_pos, next_len;
 715
 716     bit_len = *data_len;
 717     bit_pos = 0;
 718
 719     init_put_bits(&pbc, tmp, sizeof(tmp));
 720     vaapi_encode_h265_write_vps(&pbc, ctx);
 721     nal_len = put_bits_count(&pbc);
 722     flush_put_bits(&pbc);
 723
 724     next_len = bit_len - bit_pos;
 725     err = ff_vaapi_encode_h26x_nal_unit_to_byte_stream(data + bit_pos / 8,
 726                                                        &next_len,
 727                                                        tmp, nal_len);
 728     if (err < 0)
 729         return err;
 730     bit_pos += next_len;
 731
 732     init_put_bits(&pbc, tmp, sizeof(tmp));
 733     vaapi_encode_h265_write_sps(&pbc, ctx);
 734     nal_len = put_bits_count(&pbc);
 735     flush_put_bits(&pbc);
 736
 737     next_len = bit_len - bit_pos;
 738     err = ff_vaapi_encode_h26x_nal_unit_to_byte_stream(data + bit_pos / 8,
 739                                                        &next_len,
 740                                                        tmp, nal_len);
 741     if (err < 0)
 742         return err;
 743     bit_pos += next_len;
 744
 745     init_put_bits(&pbc, tmp, sizeof(tmp));
 746     vaapi_encode_h265_write_pps(&pbc, ctx);
 747     nal_len = put_bits_count(&pbc);
 748     flush_put_bits(&pbc);
 749
 750     next_len = bit_len - bit_pos;
 751     err = ff_vaapi_encode_h26x_nal_unit_to_byte_stream(data + bit_pos / 8,
 752                                                        &next_len,
 753                                                        tmp, nal_len);
 754     if (err < 0)
 755         return err;
 756     bit_pos += next_len;
 757
 758     *data_len = bit_pos;
 759     return 0;
 760 }
 761
 762 static int vaapi_encode_h265_write_slice_header(AVCodecContext *avctx,
 763                                                 VAAPIEncodePicture *pic,
 764                                                 VAAPIEncodeSlice *slice,
 765                                                 char *data, size_t *data_len)
 766 {
 767     VAAPIEncodeContext *ctx = avctx->priv_data;
 768     PutBitContext pbc;
 769     char tmp[256];
 770     size_t header_len;
 771
 772     init_put_bits(&pbc, tmp, sizeof(tmp));
 773     vaapi_encode_h265_write_slice_header2(&pbc, ctx, pic, slice);
 774     header_len = put_bits_count(&pbc);
 775     flush_put_bits(&pbc);
 776
 777     return ff_vaapi_encode_h26x_nal_unit_to_byte_stream(data, data_len,
 778                                                         tmp, header_len);
 779 }
 780
 781 static int vaapi_encode_h265_init_sequence_params(AVCodecContext *avctx)
 782 {
 783     VAAPIEncodeContext                 *ctx = avctx->priv_data;
 784     VAEncSequenceParameterBufferHEVC  *vseq = ctx->codec_sequence_params;
 785     VAEncPictureParameterBufferHEVC   *vpic = ctx->codec_picture_params;
 786     VAAPIEncodeH265Context            *priv = ctx->priv_data;
 787     VAAPIEncodeH265MiscSequenceParams *mseq = &priv->misc_sequence_params;
 788     int i;
 789
 790     {
 791         // general_profile_space == 0.
 792         vseq->general_profile_idc = 1; // Main profile (ctx->codec_profile?)
 793         vseq->general_tier_flag = 0;
 794
 795         vseq->general_level_idc = avctx->level * 3;
 796
 797         vseq->intra_period = 0;
 798         vseq->intra_idr_period = 0;
 799         vseq->ip_period = 0;
 800
 801         vseq->pic_width_in_luma_samples  = ctx->aligned_width;
 802         vseq->pic_height_in_luma_samples = ctx->aligned_height;
 803
 804         vseq->seq_fields.bits.chroma_format_idc = 1; // 4:2:0.
 805         vseq->seq_fields.bits.separate_colour_plane_flag = 0;
 806         vseq->seq_fields.bits.bit_depth_luma_minus8 = 0; // 8-bit luma.
 807         vseq->seq_fields.bits.bit_depth_chroma_minus8 = 0; // 8-bit chroma.
 808         // Other misc flags all zero.
 809
 810         // These have to come from the capabilities of the encoder.  We have
 811         // no way to query it, so just hardcode ones which worked for me...
 812         // CTB size from 8x8 to 32x32.
 813         vseq->log2_min_luma_coding_block_size_minus3 = 0;
 814         vseq->log2_diff_max_min_luma_coding_block_size = 2;
 815         // Transform size from 4x4 to 32x32.
 816         vseq->log2_min_transform_block_size_minus2 = 0;
 817         vseq->log2_diff_max_min_transform_block_size = 3;
 818         // Full transform hierarchy allowed (2-5).
 819         vseq->max_transform_hierarchy_depth_inter = 3;
 820         vseq->max_transform_hierarchy_depth_intra = 3;
 821
 822         vseq->vui_parameters_present_flag = 0;
 823
 824         vseq->bits_per_second = avctx->bit_rate;
 825         if (avctx->framerate.num > 0 && avctx->framerate.den > 0) {
 826             vseq->vui_num_units_in_tick = avctx->framerate.num;
 827             vseq->vui_time_scale        = avctx->framerate.den;
 828         } else {
 829             vseq->vui_num_units_in_tick = avctx->time_base.num;
 830             vseq->vui_time_scale        = avctx->time_base.den;
 831         }
 832
 833         vseq->intra_period     = ctx->p_per_i * (ctx->b_per_p + 1);
 834         vseq->intra_idr_period = vseq->intra_period;
 835         vseq->ip_period        = ctx->b_per_p + 1;
 836     }
 837
 838     {
 839         vpic->decoded_curr_pic.picture_id = VA_INVALID_ID;
 840         vpic->decoded_curr_pic.flags      = VA_PICTURE_HEVC_INVALID;
 841
 842         for (i = 0; i < FF_ARRAY_ELEMS(vpic->reference_frames); i++) {
 843             vpic->reference_frames[i].picture_id = VA_INVALID_ID;
 844             vpic->reference_frames[i].flags      = VA_PICTURE_HEVC_INVALID;
 845         }
 846
 847         vpic->collocated_ref_pic_index = 0xff;
 848
 849         vpic->last_picture = 0;
 850
 851         vpic->pic_init_qp = priv->fixed_qp_idr;
 852
 853         vpic->diff_cu_qp_delta_depth = 0;
 854         vpic->pps_cb_qp_offset = 0;
 855         vpic->pps_cr_qp_offset = 0;
 856
 857         // tiles_enabled_flag == 0, so ignore num_tile_(rows|columns)_minus1.
 858
 859         vpic->log2_parallel_merge_level_minus2 = 0;
 860
 861         // No limit on size.
 862         vpic->ctu_max_bitsize_allowed = 0;
 863
 864         vpic->num_ref_idx_l0_default_active_minus1 = 0;
 865         vpic->num_ref_idx_l1_default_active_minus1 = 0;
 866
 867         vpic->slice_pic_parameter_set_id = 0;
 868
 869         vpic->pic_fields.bits.screen_content_flag = 0;
 870         vpic->pic_fields.bits.enable_gpu_weighted_prediction = 0;
 871         vpic->pic_fields.bits.cu_qp_delta_enabled_flag = 1;
 872     }
 873
 874     {
 875         mseq->video_parameter_set_id = 5;
 876         mseq->seq_parameter_set_id = 5;
 877
 878         mseq->vps_max_layers_minus1 = 0;
 879         mseq->vps_max_sub_layers_minus1 = 0;
 880         mseq->vps_temporal_id_nesting_flag = 1;
 881         mseq->sps_max_sub_layers_minus1 = 0;
 882         mseq->sps_temporal_id_nesting_flag = 1;
 883
 884         for (i = 0; i < 32; i++) {
 885             mseq->general_profile_compatibility_flag[i] =
 886                 (i == vseq->general_profile_idc);
 887         }
 888
 889         mseq->general_progressive_source_flag    = 1;
 890         mseq->general_interlaced_source_flag     = 0;
 891         mseq->general_non_packed_constraint_flag = 0;
 892         mseq->general_frame_only_constraint_flag = 1;
 893         mseq->general_inbld_flag = 0;
 894
 895         mseq->log2_max_pic_order_cnt_lsb_minus4 = 8;
 896         mseq->vps_sub_layer_ordering_info_present_flag = 0;
 897         mseq->vps_max_dec_pic_buffering_minus1[0] = 1;
 898         mseq->vps_max_num_reorder_pics[0]         = ctx->b_per_p;
 899         mseq->vps_max_latency_increase_plus1[0]   = 0;
 900         mseq->sps_sub_layer_ordering_info_present_flag = 0;
 901         mseq->sps_max_dec_pic_buffering_minus1[0] = 1;
 902         mseq->sps_max_num_reorder_pics[0]         = ctx->b_per_p;
 903         mseq->sps_max_latency_increase_plus1[0]   = 0;
 904
 905         mseq->vps_timing_info_present_flag = 1;
 906         mseq->vps_num_units_in_tick = avctx->time_base.num;
 907         mseq->vps_time_scale        = avctx->time_base.den;
 908         mseq->vps_poc_proportional_to_timing_flag = 1;
 909         mseq->vps_num_ticks_poc_diff_minus1 = 0;
 910
 911         if (ctx->input_width  != ctx->aligned_width ||
 912             ctx->input_height != ctx->aligned_height) {
 913             mseq->conformance_window_flag = 1;
 914             mseq->conf_win_left_offset   = 0;
 915             mseq->conf_win_right_offset  =
 916                 (ctx->aligned_width - ctx->input_width) / 2;
 917             mseq->conf_win_top_offset    = 0;
 918             mseq->conf_win_bottom_offset =
 919                 (ctx->aligned_height - ctx->input_height) / 2;
 920         } else {
 921             mseq->conformance_window_flag = 0;
 922         }
 923
 924         mseq->num_short_term_ref_pic_sets = 0;
 925         // STRPSs should ideally be here rather than repeated in each slice.
 926
 927         mseq->vui_parameters_present_flag = 1;
 928         if (avctx->sample_aspect_ratio.num != 0) {
 929             mseq->aspect_ratio_info_present_flag = 1;
 930             if (avctx->sample_aspect_ratio.num ==
 931                 avctx->sample_aspect_ratio.den) {
 932                 mseq->aspect_ratio_idc = 1;
 933             } else {
 934                 mseq->aspect_ratio_idc = 255; // Extended SAR.
 935                 mseq->sar_width  = avctx->sample_aspect_ratio.num;
 936                 mseq->sar_height = avctx->sample_aspect_ratio.den;
 937             }
 938         }
 939         if (1) {
 940             // Should this be conditional on some of these being set?
 941             mseq->video_signal_type_present_flag = 1;
 942             mseq->video_format = 5; // Unspecified.
 943             mseq->video_full_range_flag = 0;
 944             mseq->colour_description_present_flag = 1;
 945             mseq->colour_primaries = avctx->color_primaries;
 946             mseq->transfer_characteristics = avctx->color_trc;
 947             mseq->matrix_coeffs = avctx->colorspace;
 948         }
 949     }
 950
 951     return 0;
 952 }
 953
 954 static int vaapi_encode_h265_init_picture_params(AVCodecContext *avctx,
 955                                                  VAAPIEncodePicture *pic)
 956 {
 957     VAAPIEncodeContext               *ctx = avctx->priv_data;
 958     VAEncPictureParameterBufferHEVC *vpic = pic->codec_picture_params;
 959     VAAPIEncodeH265Context          *priv = ctx->priv_data;
 960     int i;
 961
 962     if (pic->type == PICTURE_TYPE_IDR) {
 963         av_assert0(pic->display_order == pic->encode_order);
 964         priv->last_idr_frame = pic->display_order;
 965     } else {
 966         av_assert0(pic->encode_order > priv->last_idr_frame);
 967         // Display order need not be if we have RA[SD]L pictures, though.
 968     }
 969
 970     vpic->decoded_curr_pic.picture_id    = pic->recon_surface;
 971     vpic->decoded_curr_pic.pic_order_cnt =
 972         pic->display_order - priv->last_idr_frame;
 973     vpic->decoded_curr_pic.flags         = 0;
 974
 975     for (i = 0; i < pic->nb_refs; i++) {
 976         VAAPIEncodePicture *ref = pic->refs[i];
 977         av_assert0(ref);
 978         vpic->reference_frames[i].picture_id    = ref->recon_surface;
 979         vpic->reference_frames[i].pic_order_cnt =
 980             ref->display_order - priv->last_idr_frame;
 981         vpic->reference_frames[i].flags =
 982             (ref->display_order < pic->display_order ?
 983              VA_PICTURE_HEVC_RPS_ST_CURR_BEFORE : 0) |
 984             (ref->display_order > pic->display_order ?
 985              VA_PICTURE_HEVC_RPS_ST_CURR_AFTER  : 0);
 986     }
 987     for (; i < FF_ARRAY_ELEMS(vpic->reference_frames); i++) {
 988         vpic->reference_frames[i].picture_id = VA_INVALID_ID;
 989         vpic->reference_frames[i].flags      = VA_PICTURE_HEVC_INVALID;
 990     }
 991
 992     vpic->coded_buf = pic->output_buffer;
 993
 994     switch (pic->type) {
 995     case PICTURE_TYPE_IDR:
 996         vpic->nal_unit_type = NAL_IDR_W_RADL;
 997         vpic->pic_fields.bits.idr_pic_flag = 1;
 998         vpic->pic_fields.bits.coding_type  = 1;
 999         vpic->pic_fields.bits.reference_pic_flag = 1;
1000         break;
1001     case PICTURE_TYPE_I:
1002         vpic->nal_unit_type = NAL_TRAIL_R;
1003         vpic->pic_fields.bits.idr_pic_flag = 0;
1004         vpic->pic_fields.bits.coding_type  = 1;
1005         vpic->pic_fields.bits.reference_pic_flag = 1;
1006         break;
1007     case PICTURE_TYPE_P:
1008         vpic->nal_unit_type = NAL_TRAIL_R;
1009         vpic->pic_fields.bits.idr_pic_flag = 0;
1010         vpic->pic_fields.bits.coding_type  = 2;
1011         vpic->pic_fields.bits.reference_pic_flag = 1;
1012         break;
1013     case PICTURE_TYPE_B:
1014         vpic->nal_unit_type = NAL_TRAIL_R;
1015         vpic->pic_fields.bits.idr_pic_flag = 0;
1016         vpic->pic_fields.bits.coding_type  = 3;
1017         vpic->pic_fields.bits.reference_pic_flag = 0;
1018         break;
1019     default:
1020         av_assert0(0 && "invalid picture type");
1021     }
1022
1023     pic->nb_slices = 1;
1024
1025     return 0;
1026 }
1027
1028 static int vaapi_encode_h265_init_slice_params(AVCodecContext *avctx,
1029                                                VAAPIEncodePicture *pic,
1030                                                VAAPIEncodeSlice *slice)
1031 {
1032     VAAPIEncodeContext                *ctx = avctx->priv_data;
1033     VAEncPictureParameterBufferHEVC  *vpic = pic->codec_picture_params;
1034     VAEncSliceParameterBufferHEVC  *vslice = slice->codec_slice_params;
1035     VAAPIEncodeH265Context           *priv = ctx->priv_data;
1036     VAAPIEncodeH265Slice           *pslice;
1037     VAAPIEncodeH265MiscSliceParams *mslice;
1038     int i;
1039
1040     slice->priv_data = av_mallocz(sizeof(*pslice));
1041     if (!slice->priv_data)
1042         return AVERROR(ENOMEM);
1043     pslice = slice->priv_data;
1044     mslice = &pslice->misc_slice_params;
1045
1046     // Currently we only support one slice per frame.
1047     vslice->slice_segment_address = 0;
1048     vslice->num_ctu_in_slice = priv->ctu_width * priv->ctu_height;
1049
1050     switch (pic->type) {
1051     case PICTURE_TYPE_IDR:
1052     case PICTURE_TYPE_I:
1053         vslice->slice_type = I_SLICE;
1054         break;
1055     case PICTURE_TYPE_P:
1056         vslice->slice_type = P_SLICE;
1057         break;
1058     case PICTURE_TYPE_B:
1059         vslice->slice_type = B_SLICE;
1060         break;
1061     default:
1062         av_assert0(0 && "invalid picture type");
1063     }
1064
1065     vslice->slice_pic_parameter_set_id = vpic->slice_pic_parameter_set_id;
1066
1067     pslice->pic_order_cnt = pic->display_order - priv->last_idr_frame;
1068
1069     for (i = 0; i < FF_ARRAY_ELEMS(vslice->ref_pic_list0); i++) {
1070         vslice->ref_pic_list0[i].picture_id = VA_INVALID_ID;
1071         vslice->ref_pic_list0[i].flags      = VA_PICTURE_HEVC_INVALID;
1072         vslice->ref_pic_list1[i].picture_id = VA_INVALID_ID;
1073         vslice->ref_pic_list1[i].flags      = VA_PICTURE_HEVC_INVALID;
1074     }
1075
1076     av_assert0(pic->nb_refs <= 2);
1077     if (pic->nb_refs >= 1) {
1078         // Backward reference for P or B frame.
1079         av_assert0(pic->type == PICTURE_TYPE_P ||
1080                    pic->type == PICTURE_TYPE_B);
1081
1082         vslice->num_ref_idx_l0_active_minus1 = 0;
1083         vslice->ref_pic_list0[0] = vpic->reference_frames[0];
1084     }
1085     if (pic->nb_refs >= 2) {
1086         // Forward reference for B frame.
1087         av_assert0(pic->type == PICTURE_TYPE_B);
1088
1089         vslice->num_ref_idx_l1_active_minus1 = 0;
1090         vslice->ref_pic_list1[0] = vpic->reference_frames[1];
1091     }
1092
1093     vslice->max_num_merge_cand = 5;
1094
1095     if (pic->type == PICTURE_TYPE_B)
1096         vslice->slice_qp_delta = priv->fixed_qp_b  - vpic->pic_init_qp;
1097     else if (pic->type == PICTURE_TYPE_P)
1098         vslice->slice_qp_delta = priv->fixed_qp_p - vpic->pic_init_qp;
1099     else
1100         vslice->slice_qp_delta = priv->fixed_qp_idr - vpic->pic_init_qp;
1101
1102     vslice->slice_fields.bits.last_slice_of_pic_flag = 1;
1103
1104     mslice->first_slice_segment_in_pic_flag = 1;
1105
1106     if (pic->type == PICTURE_TYPE_IDR) {
1107         // No reference pictures.
1108     } else if (0) {
1109         mslice->short_term_ref_pic_set_sps_flag = 1;
1110         mslice->short_term_ref_pic_idx = 0;
1111     } else {
1112         VAAPIEncodePicture *st;
1113         int used;
1114
1115         mslice->short_term_ref_pic_set_sps_flag = 0;
1116         mslice->st_ref_pic_set.inter_ref_pic_set_prediction_flag = 0;
1117
1118         for (st = ctx->pic_start; st; st = st->next) {
1119             if (st->encode_order >= pic->encode_order) {
1120                 // Not yet in DPB.
1121                 continue;
1122             }
1123             used = 0;
1124             for (i = 0; i < pic->nb_refs; i++) {
1125                 if (pic->refs[i] == st)
1126                     used = 1;
1127             }
1128             if (!used) {
1129                 // Currently true, but need not be.
1130                 continue;
1131             }
1132             // This only works for one instance of each (delta_poc_sN_minus1
1133             // is relative to the previous frame in the list, not relative to
1134             // the current frame directly).
1135             if (st->display_order < pic->display_order) {
1136                 i = mslice->st_ref_pic_set.num_negative_pics;
1137                 mslice->st_ref_pic_set.delta_poc_s0_minus1[i] =
1138                     pic->display_order - st->display_order - 1;
1139                 mslice->st_ref_pic_set.used_by_curr_pic_s0_flag[i] = used;
1140                 ++mslice->st_ref_pic_set.num_negative_pics;
1141             } else {
1142                 i = mslice->st_ref_pic_set.num_positive_pics;
1143                 mslice->st_ref_pic_set.delta_poc_s1_minus1[i] =
1144                     st->display_order - pic->display_order - 1;
1145                 mslice->st_ref_pic_set.used_by_curr_pic_s1_flag[i] = used;
1146                 ++mslice->st_ref_pic_set.num_positive_pics;
1147             }
1148         }
1149     }
1150
1151     return 0;
1152 }
1153
1154 static av_cold int vaapi_encode_h265_init_constant_bitrate(AVCodecContext *avctx)
1155 {
1156     VAAPIEncodeContext      *ctx = avctx->priv_data;
1157     VAAPIEncodeH265Context *priv = ctx->priv_data;
1158     int hrd_buffer_size;
1159     int hrd_initial_buffer_fullness;
1160
1161     if (avctx->bit_rate > INT32_MAX) {
1162         av_log(avctx, AV_LOG_ERROR, "Target bitrate of 2^31 bps or "
1163                "higher is not supported.\n");
1164         return AVERROR(EINVAL);
1165     }
1166
1167     if (avctx->rc_buffer_size)
1168         hrd_buffer_size = avctx->rc_buffer_size;
1169     else
1170         hrd_buffer_size = avctx->bit_rate;
1171     if (avctx->rc_initial_buffer_occupancy)
1172         hrd_initial_buffer_fullness = avctx->rc_initial_buffer_occupancy;
1173     else
1174         hrd_initial_buffer_fullness = hrd_buffer_size * 3 / 4;
1175
1176     priv->rc_params.misc.type = VAEncMiscParameterTypeRateControl;
1177     priv->rc_params.rc = (VAEncMiscParameterRateControl) {
1178         .bits_per_second   = avctx->bit_rate,
1179         .target_percentage = 66,
1180         .window_size       = 1000,
1181         .initial_qp        = (avctx->qmax >= 0 ? avctx->qmax : 40),
1182         .min_qp            = (avctx->qmin >= 0 ? avctx->qmin : 20),
1183         .basic_unit_size   = 0,
1184     };
1185     ctx->global_params[ctx->nb_global_params] =
1186         &priv->rc_params.misc;
1187     ctx->global_params_size[ctx->nb_global_params++] =
1188         sizeof(priv->rc_params);
1189
1190     priv->hrd_params.misc.type = VAEncMiscParameterTypeHRD;
1191     priv->hrd_params.hrd = (VAEncMiscParameterHRD) {
1192         .initial_buffer_fullness = hrd_initial_buffer_fullness,
1193         .buffer_size             = hrd_buffer_size,
1194     };
1195     ctx->global_params[ctx->nb_global_params] =
1196         &priv->hrd_params.misc;
1197     ctx->global_params_size[ctx->nb_global_params++] =
1198         sizeof(priv->hrd_params);
1199
1200     // These still need to be  set for pic_init_qp/slice_qp_delta.
1201     priv->fixed_qp_idr = 30;
1202     priv->fixed_qp_p   = 30;
1203     priv->fixed_qp_b   = 30;
1204
1205     av_log(avctx, AV_LOG_DEBUG, "Using constant-bitrate = %"PRId64" bps.\n",
1206            avctx->bit_rate);
1207     return 0;
1208 }
1209
1210 static av_cold int vaapi_encode_h265_init_fixed_qp(AVCodecContext *avctx)
1211 {
1212     VAAPIEncodeContext      *ctx = avctx->priv_data;
1213     VAAPIEncodeH265Context *priv = ctx->priv_data;
1214     VAAPIEncodeH265Options  *opt = ctx->codec_options;
1215
1216     priv->fixed_qp_p = opt->qp;
1217     if (avctx->i_quant_factor > 0.0)
1218         priv->fixed_qp_idr = (int)((priv->fixed_qp_p * avctx->i_quant_factor +
1219                                     avctx->i_quant_offset) + 0.5);
1220     else
1221         priv->fixed_qp_idr = priv->fixed_qp_p;
1222     if (avctx->b_quant_factor > 0.0)
1223         priv->fixed_qp_b = (int)((priv->fixed_qp_p * avctx->b_quant_factor +
1224                                   avctx->b_quant_offset) + 0.5);
1225     else
1226         priv->fixed_qp_b = priv->fixed_qp_p;
1227
1228     av_log(avctx, AV_LOG_DEBUG, "Using fixed QP = "
1229            "%d / %d / %d for IDR / P / B frames.\n",
1230            priv->fixed_qp_idr, priv->fixed_qp_p, priv->fixed_qp_b);
1231     return 0;
1232 }
1233
1234 static av_cold int vaapi_encode_h265_init_internal(AVCodecContext *avctx)
1235 {
1236     static const VAConfigAttrib default_config_attributes[] = {
1237         { .type  = VAConfigAttribRTFormat,
1238           .value = VA_RT_FORMAT_YUV420 },
1239         { .type  = VAConfigAttribEncPackedHeaders,
1240           .value = (VA_ENC_PACKED_HEADER_SEQUENCE |
1241                     VA_ENC_PACKED_HEADER_SLICE) },
1242     };
1243
1244     VAAPIEncodeContext      *ctx = avctx->priv_data;
1245     VAAPIEncodeH265Context *priv = ctx->priv_data;
1246     int i, err;
1247
1248     switch (avctx->profile) {
1249     case FF_PROFILE_HEVC_MAIN:
1250     case FF_PROFILE_UNKNOWN:
1251         ctx->va_profile = VAProfileHEVCMain;
1252         break;
1253     case FF_PROFILE_HEVC_MAIN_10:
1254         av_log(avctx, AV_LOG_ERROR, "H.265 main 10-bit profile "
1255                "is not supported.\n");
1256         return AVERROR_PATCHWELCOME;
1257     default:
1258         av_log(avctx, AV_LOG_ERROR, "Unknown H.265 profile %d.\n",
1259                avctx->profile);
1260         return AVERROR(EINVAL);
1261     }
1262     ctx->va_entrypoint = VAEntrypointEncSlice;
1263
1264     ctx->input_width    = avctx->width;
1265     ctx->input_height   = avctx->height;
1266     ctx->aligned_width  = FFALIGN(ctx->input_width,  16);
1267     ctx->aligned_height = FFALIGN(ctx->input_height, 16);
1268     priv->ctu_width     = FFALIGN(ctx->aligned_width,  32) / 32;
1269     priv->ctu_height    = FFALIGN(ctx->aligned_height, 32) / 32;
1270
1271     av_log(avctx, AV_LOG_VERBOSE, "Input %ux%u -> Aligned %ux%u -> CTU %ux%u.\n",
1272            ctx->input_width, ctx->input_height, ctx->aligned_width,
1273            ctx->aligned_height, priv->ctu_width, priv->ctu_height);
1274
1275     for (i = 0; i < FF_ARRAY_ELEMS(default_config_attributes); i++) {
1276         ctx->config_attributes[ctx->nb_config_attributes++] =
1277             default_config_attributes[i];
1278     }
1279
1280     if (avctx->bit_rate > 0) {
1281         ctx->va_rc_mode = VA_RC_CBR;
1282         err = vaapi_encode_h265_init_constant_bitrate(avctx);
1283     } else {
1284         ctx->va_rc_mode = VA_RC_CQP;
1285         err = vaapi_encode_h265_init_fixed_qp(avctx);
1286     }
1287     if (err < 0)
1288         return err;
1289
1290     ctx->config_attributes[ctx->nb_config_attributes++] = (VAConfigAttrib) {
1291         .type  = VAConfigAttribRateControl,
1292         .value = ctx->va_rc_mode,
1293     };
1294
1295     ctx->nb_recon_frames = 20;
1296
1297     return 0;
1298 }
1299
1300 static VAAPIEncodeType vaapi_encode_type_h265 = {
1301     .priv_data_size        = sizeof(VAAPIEncodeH265Context),
1302
1303     .init                  = &vaapi_encode_h265_init_internal,
1304
1305     .sequence_params_size  = sizeof(VAEncSequenceParameterBufferHEVC),
1306     .init_sequence_params  = &vaapi_encode_h265_init_sequence_params,
1307
1308     .picture_params_size   = sizeof(VAEncPictureParameterBufferHEVC),
1309     .init_picture_params   = &vaapi_encode_h265_init_picture_params,
1310
1311     .slice_params_size     = sizeof(VAEncSliceParameterBufferHEVC),
1312     .init_slice_params     = &vaapi_encode_h265_init_slice_params,
1313
1314     .sequence_header_type  = VAEncPackedHeaderSequence,
1315     .write_sequence_header = &vaapi_encode_h265_write_sequence_header,
1316
1317     .slice_header_type     = VAEncPackedHeaderHEVC_Slice,
1318     .write_slice_header    = &vaapi_encode_h265_write_slice_header,
1319 };
1320
1321 static av_cold int vaapi_encode_h265_init(AVCodecContext *avctx)
1322 {
1323     return ff_vaapi_encode_init(avctx, &vaapi_encode_type_h265);
1324 }
1325
1326 #define OFFSET(x) (offsetof(VAAPIEncodeContext, codec_options_data) + \
1327                    offsetof(VAAPIEncodeH265Options, x))
1328 #define FLAGS (AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM)
1329 static const AVOption vaapi_encode_h265_options[] = {
1330     { "qp", "Constant QP (for P frames; scaled by qfactor/qoffset for I/B)",
1331       OFFSET(qp), AV_OPT_TYPE_INT, { .i64 = 25 }, 0, 52, FLAGS },
1332     { NULL },
1333 };
1334
1335 static const AVCodecDefault vaapi_encode_h265_defaults[] = {
1336     { "profile",        "1"   },
1337     { "level",          "51"  },
1338     { "b",              "0"   },
1339     { "bf",             "2"   },
1340     { "g",              "120" },
1341     { "i_qfactor",      "1.0" },
1342     { "i_qoffset",      "0.0" },
1343     { "b_qfactor",      "1.2" },
1344     { "b_qoffset",      "0.0" },
1345     { NULL },
1346 };
1347
1348 static const AVClass vaapi_encode_h265_class = {
1349     .class_name = "h265_vaapi",
1350     .item_name  = av_default_item_name,
1351     .option     = vaapi_encode_h265_options,
1352     .version    = LIBAVUTIL_VERSION_INT,
1353 };
1354
1355 AVCodec ff_hevc_vaapi_encoder = {
1356     .name           = "hevc_vaapi",
1357     .long_name      = NULL_IF_CONFIG_SMALL("H.265/HEVC (VAAPI)"),
1358     .type           = AVMEDIA_TYPE_VIDEO,
1359     .id             = AV_CODEC_ID_HEVC,
1360     .priv_data_size = (sizeof(VAAPIEncodeContext) +
1361                        sizeof(VAAPIEncodeH265Options)),
1362     .init           = &vaapi_encode_h265_init,
1363     .encode2        = &ff_vaapi_encode2,
1364     .close          = &ff_vaapi_encode_close,
1365     .priv_class     = &vaapi_encode_h265_class,
1366     .capabilities   = AV_CODEC_CAP_DELAY,
1367     .defaults       = vaapi_encode_h265_defaults,
1368     .pix_fmts = (const enum AVPixelFormat[]) {
1369         AV_PIX_FMT_VAAPI,
1370         AV_PIX_FMT_NONE,
1371     },
1372 };