git.sesse.net Git - ffmpeg/blob - libavcodec/vaapi_encode_h264.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_h264.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 "h264.h"
  29 #include "h264_sei.h"
  30 #include "internal.h"
  31 #include "vaapi_encode.h"
  32 #include "vaapi_encode_h26x.h"
  33
  34 enum {
  35     SLICE_TYPE_P  = 0,
  36     SLICE_TYPE_B  = 1,
  37     SLICE_TYPE_I  = 2,
  38     SLICE_TYPE_SP = 3,
  39     SLICE_TYPE_SI = 4,
  40 };
  41
  42 // This structure contains all possibly-useful per-sequence syntax elements
  43 // which are not already contained in the various VAAPI structures.
  44 typedef struct VAAPIEncodeH264MiscSequenceParams {
  45     unsigned int profile_idc;
  46     char constraint_set0_flag;
  47     char constraint_set1_flag;
  48     char constraint_set2_flag;
  49     char constraint_set3_flag;
  50     char constraint_set4_flag;
  51     char constraint_set5_flag;
  52
  53     char separate_colour_plane_flag;
  54     char qpprime_y_zero_transform_bypass_flag;
  55
  56     char gaps_in_frame_num_allowed_flag;
  57     char delta_pic_order_always_zero_flag;
  58     char bottom_field_pic_order_in_frame_present_flag;
  59
  60     unsigned int num_slice_groups_minus1;
  61     unsigned int slice_group_map_type;
  62
  63     int pic_init_qs_minus26;
  64
  65     char overscan_info_present_flag;
  66     char overscan_appropriate_flag;
  67
  68     char video_signal_type_present_flag;
  69     unsigned int video_format;
  70     char video_full_range_flag;
  71     char colour_description_present_flag;
  72     unsigned int colour_primaries;
  73     unsigned int transfer_characteristics;
  74     unsigned int matrix_coefficients;
  75
  76     char chroma_loc_info_present_flag;
  77     unsigned int chroma_sample_loc_type_top_field;
  78     unsigned int chroma_sample_loc_type_bottom_field;
  79
  80     // Some timing elements are in VAEncSequenceParameterBufferH264.
  81     char fixed_frame_rate_flag;
  82
  83     char nal_hrd_parameters_present_flag;
  84     char vcl_hrd_parameters_present_flag;
  85     char low_delay_hrd_flag;
  86     char pic_struct_present_flag;
  87
  88     char motion_vectors_over_pic_boundaries_flag;
  89     unsigned int max_bytes_per_pic_denom;
  90     unsigned int max_bits_per_mb_denom;
  91     unsigned int max_num_reorder_frames;
  92     unsigned int max_dec_pic_buffering;
  93
  94     unsigned int cpb_cnt_minus1;
  95     unsigned int bit_rate_scale;
  96     unsigned int cpb_size_scale;
  97     unsigned int bit_rate_value_minus1[32];
  98     unsigned int cpb_size_value_minus1[32];
  99     char cbr_flag[32];
 100     unsigned int initial_cpb_removal_delay_length_minus1;
 101     unsigned int cpb_removal_delay_length_minus1;
 102     unsigned int dpb_output_delay_length_minus1;
 103     unsigned int time_offset_length;
 104
 105     unsigned int initial_cpb_removal_delay;
 106     unsigned int initial_cpb_removal_delay_offset;
 107
 108     unsigned int pic_struct;
 109 } VAAPIEncodeH264MiscSequenceParams;
 110
 111 // This structure contains all possibly-useful per-slice syntax elements
 112 // which are not already contained in the various VAAPI structures.
 113 typedef struct VAAPIEncodeH264MiscSliceParams {
 114     unsigned int nal_unit_type;
 115     unsigned int nal_ref_idc;
 116
 117     unsigned int colour_plane_id;
 118     char field_pic_flag;
 119     char bottom_field_flag;
 120
 121     unsigned int redundant_pic_cnt;
 122
 123     char sp_for_switch_flag;
 124     int slice_qs_delta;
 125
 126     char ref_pic_list_modification_flag_l0;
 127     char ref_pic_list_modification_flag_l1;
 128
 129     char no_output_of_prior_pics_flag;
 130     char long_term_reference_flag;
 131     char adaptive_ref_pic_marking_mode_flag;
 132 } VAAPIEncodeH264MiscSliceParams;
 133
 134 typedef struct VAAPIEncodeH264Slice {
 135     VAAPIEncodeH264MiscSliceParams misc_slice_params;
 136 } VAAPIEncodeH264Slice;
 137
 138 typedef struct VAAPIEncodeH264Context {
 139     VAAPIEncodeH264MiscSequenceParams misc_sequence_params;
 140
 141     int mb_width;
 142     int mb_height;
 143
 144     int fixed_qp_idr;
 145     int fixed_qp_p;
 146     int fixed_qp_b;
 147
 148     int next_frame_num;
 149     int64_t last_idr_frame;
 150     int64_t idr_pic_count;
 151
 152     int cpb_delay;
 153     int dpb_delay;
 154
 155     // Rate control configuration.
 156     int send_timing_sei;
 157 } VAAPIEncodeH264Context;
 158
 159 typedef struct VAAPIEncodeH264Options {
 160     int qp;
 161     int quality;
 162     int low_power;
 163     // Entropy encoder type.
 164     int coder;
 165 } VAAPIEncodeH264Options;
 166
 167
 168 #define vseq_var(name)     vseq->name, name
 169 #define vseq_field(name)   vseq->seq_fields.bits.name, name
 170 #define vvui_field(name)   vseq->vui_fields.bits.name, name
 171 #define vpic_var(name)     vpic->name, name
 172 #define vpic_field(name)   vpic->pic_fields.bits.name, name
 173 #define vslice_var(name)   vslice->name, name
 174 #define vslice_field(name) vslice->slice_fields.bits.name, name
 175 #define mseq_var(name)     mseq->name, name
 176 #define mslice_var(name)   mslice->name, name
 177
 178 static void vaapi_encode_h264_write_nal_header(PutBitContext *pbc,
 179                                                int nal_unit_type, int nal_ref_idc)
 180 {
 181     u(1, 0, forbidden_zero_bit);
 182     u(2, nal_ref_idc, nal_ref_idc);
 183     u(5, nal_unit_type, nal_unit_type);
 184 }
 185
 186 static void vaapi_encode_h264_write_trailing_rbsp(PutBitContext *pbc)
 187 {
 188     u(1, 1, rbsp_stop_one_bit);
 189     while (put_bits_count(pbc) & 7)
 190         u(1, 0, rbsp_alignment_zero_bit);
 191 }
 192
 193 static void vaapi_encode_h264_write_vui(PutBitContext *pbc,
 194                                         VAAPIEncodeContext *ctx)
 195 {
 196     VAEncSequenceParameterBufferH264  *vseq = ctx->codec_sequence_params;
 197     VAAPIEncodeH264Context            *priv = ctx->priv_data;
 198     VAAPIEncodeH264MiscSequenceParams *mseq = &priv->misc_sequence_params;
 199     int i;
 200
 201     u(1, vvui_field(aspect_ratio_info_present_flag));
 202     if (vseq->vui_fields.bits.aspect_ratio_info_present_flag) {
 203         u(8, vseq_var(aspect_ratio_idc));
 204         if (vseq->aspect_ratio_idc == 255) {
 205             u(16, vseq_var(sar_width));
 206             u(16, vseq_var(sar_height));
 207         }
 208     }
 209
 210     u(1, mseq_var(overscan_info_present_flag));
 211     if (mseq->overscan_info_present_flag)
 212         u(1, mseq_var(overscan_appropriate_flag));
 213
 214     u(1, mseq_var(video_signal_type_present_flag));
 215     if (mseq->video_signal_type_present_flag) {
 216         u(3, mseq_var(video_format));
 217         u(1, mseq_var(video_full_range_flag));
 218         u(1, mseq_var(colour_description_present_flag));
 219         if (mseq->colour_description_present_flag) {
 220             u(8, mseq_var(colour_primaries));
 221             u(8, mseq_var(transfer_characteristics));
 222             u(8, mseq_var(matrix_coefficients));
 223         }
 224     }
 225
 226     u(1, mseq_var(chroma_loc_info_present_flag));
 227     if (mseq->chroma_loc_info_present_flag) {
 228         ue(mseq_var(chroma_sample_loc_type_top_field));
 229         ue(mseq_var(chroma_sample_loc_type_bottom_field));
 230     }
 231
 232     u(1, vvui_field(timing_info_present_flag));
 233     if (vseq->vui_fields.bits.timing_info_present_flag) {
 234         u(32, vseq_var(num_units_in_tick));
 235         u(32, vseq_var(time_scale));
 236         u(1, mseq_var(fixed_frame_rate_flag));
 237     }
 238
 239     u(1, mseq_var(nal_hrd_parameters_present_flag));
 240     if (mseq->nal_hrd_parameters_present_flag) {
 241         ue(mseq_var(cpb_cnt_minus1));
 242         u(4, mseq_var(bit_rate_scale));
 243         u(4, mseq_var(cpb_size_scale));
 244         for (i = 0; i <= mseq->cpb_cnt_minus1; i++) {
 245             ue(mseq_var(bit_rate_value_minus1[i]));
 246             ue(mseq_var(cpb_size_value_minus1[i]));
 247             u(1, mseq_var(cbr_flag[i]));
 248         }
 249         u(5, mseq_var(initial_cpb_removal_delay_length_minus1));
 250         u(5, mseq_var(cpb_removal_delay_length_minus1));
 251         u(5, mseq_var(dpb_output_delay_length_minus1));
 252         u(5, mseq_var(time_offset_length));
 253     }
 254     u(1, mseq_var(vcl_hrd_parameters_present_flag));
 255     if (mseq->vcl_hrd_parameters_present_flag) {
 256         av_assert0(0 && "vcl hrd parameters not supported");
 257     }
 258
 259     if (mseq->nal_hrd_parameters_present_flag ||
 260         mseq->vcl_hrd_parameters_present_flag)
 261         u(1, mseq_var(low_delay_hrd_flag));
 262     u(1, mseq_var(pic_struct_present_flag));
 263
 264     u(1, vvui_field(bitstream_restriction_flag));
 265     if (vseq->vui_fields.bits.bitstream_restriction_flag) {
 266         u(1, mseq_var(motion_vectors_over_pic_boundaries_flag));
 267         ue(mseq_var(max_bytes_per_pic_denom));
 268         ue(mseq_var(max_bits_per_mb_denom));
 269         ue(vvui_field(log2_max_mv_length_horizontal));
 270         ue(vvui_field(log2_max_mv_length_vertical));
 271         ue(mseq_var(max_num_reorder_frames));
 272         ue(mseq_var(max_dec_pic_buffering));
 273     }
 274 }
 275
 276 static void vaapi_encode_h264_write_sps(PutBitContext *pbc,
 277                                         VAAPIEncodeContext *ctx)
 278 {
 279     VAEncSequenceParameterBufferH264  *vseq = ctx->codec_sequence_params;
 280     VAAPIEncodeH264Context            *priv = ctx->priv_data;
 281     VAAPIEncodeH264MiscSequenceParams *mseq = &priv->misc_sequence_params;
 282     int i;
 283
 284     vaapi_encode_h264_write_nal_header(pbc, H264_NAL_SPS, 3);
 285
 286     u(8, mseq_var(profile_idc));
 287     u(1, mseq_var(constraint_set0_flag));
 288     u(1, mseq_var(constraint_set1_flag));
 289     u(1, mseq_var(constraint_set2_flag));
 290     u(1, mseq_var(constraint_set3_flag));
 291     u(1, mseq_var(constraint_set4_flag));
 292     u(1, mseq_var(constraint_set5_flag));
 293     u(2, 0, reserved_zero_2bits);
 294
 295     u(8, vseq_var(level_idc));
 296
 297     ue(vseq_var(seq_parameter_set_id));
 298
 299     if (mseq->profile_idc == 100 || mseq->profile_idc == 110 ||
 300         mseq->profile_idc == 122 || mseq->profile_idc == 244 ||
 301         mseq->profile_idc ==  44 || mseq->profile_idc ==  83 ||
 302         mseq->profile_idc ==  86 || mseq->profile_idc == 118 ||
 303         mseq->profile_idc == 128 || mseq->profile_idc == 138) {
 304         ue(vseq_field(chroma_format_idc));
 305
 306         if (vseq->seq_fields.bits.chroma_format_idc == 3)
 307             u(1, mseq_var(separate_colour_plane_flag));
 308
 309         ue(vseq_var(bit_depth_luma_minus8));
 310         ue(vseq_var(bit_depth_chroma_minus8));
 311
 312         u(1, mseq_var(qpprime_y_zero_transform_bypass_flag));
 313
 314         u(1, vseq_field(seq_scaling_matrix_present_flag));
 315         if (vseq->seq_fields.bits.seq_scaling_matrix_present_flag) {
 316             av_assert0(0 && "scaling matrices not supported");
 317         }
 318     }
 319
 320     ue(vseq_field(log2_max_frame_num_minus4));
 321     ue(vseq_field(pic_order_cnt_type));
 322
 323     if (vseq->seq_fields.bits.pic_order_cnt_type == 0) {
 324         ue(vseq_field(log2_max_pic_order_cnt_lsb_minus4));
 325     } else if (vseq->seq_fields.bits.pic_order_cnt_type == 1) {
 326         u(1, mseq_var(delta_pic_order_always_zero_flag));
 327         se(vseq_var(offset_for_non_ref_pic));
 328         se(vseq_var(offset_for_top_to_bottom_field));
 329         ue(vseq_var(num_ref_frames_in_pic_order_cnt_cycle));
 330
 331         for (i = 0; i < vseq->num_ref_frames_in_pic_order_cnt_cycle; i++)
 332             se(vseq_var(offset_for_ref_frame[i]));
 333     }
 334
 335     ue(vseq_var(max_num_ref_frames));
 336     u(1, mseq_var(gaps_in_frame_num_allowed_flag));
 337
 338     ue(vseq->picture_width_in_mbs  - 1, pic_width_in_mbs_minus1);
 339     ue(vseq->picture_height_in_mbs - 1, pic_height_in_mbs_minus1);
 340
 341     u(1, vseq_field(frame_mbs_only_flag));
 342     if (!vseq->seq_fields.bits.frame_mbs_only_flag)
 343         u(1, vseq_field(mb_adaptive_frame_field_flag));
 344
 345     u(1, vseq_field(direct_8x8_inference_flag));
 346
 347     u(1, vseq_var(frame_cropping_flag));
 348     if (vseq->frame_cropping_flag) {
 349         ue(vseq_var(frame_crop_left_offset));
 350         ue(vseq_var(frame_crop_right_offset));
 351         ue(vseq_var(frame_crop_top_offset));
 352         ue(vseq_var(frame_crop_bottom_offset));
 353     }
 354
 355     u(1, vseq_var(vui_parameters_present_flag));
 356     if (vseq->vui_parameters_present_flag)
 357         vaapi_encode_h264_write_vui(pbc, ctx);
 358
 359     vaapi_encode_h264_write_trailing_rbsp(pbc);
 360 }
 361
 362 static void vaapi_encode_h264_write_pps(PutBitContext *pbc,
 363                                         VAAPIEncodeContext *ctx)
 364 {
 365     VAEncPictureParameterBufferH264   *vpic = ctx->codec_picture_params;
 366     VAAPIEncodeH264Context            *priv = ctx->priv_data;
 367     VAAPIEncodeH264MiscSequenceParams *mseq = &priv->misc_sequence_params;
 368
 369     vaapi_encode_h264_write_nal_header(pbc, H264_NAL_PPS, 3);
 370
 371     ue(vpic_var(pic_parameter_set_id));
 372     ue(vpic_var(seq_parameter_set_id));
 373
 374     u(1, vpic_field(entropy_coding_mode_flag));
 375     u(1, mseq_var(bottom_field_pic_order_in_frame_present_flag));
 376
 377     ue(mseq_var(num_slice_groups_minus1));
 378     if (mseq->num_slice_groups_minus1 > 0) {
 379         ue(mseq_var(slice_group_map_type));
 380         av_assert0(0 && "slice groups not supported");
 381     }
 382
 383     ue(vpic_var(num_ref_idx_l0_active_minus1));
 384     ue(vpic_var(num_ref_idx_l1_active_minus1));
 385
 386     u(1, vpic_field(weighted_pred_flag));
 387     u(2, vpic_field(weighted_bipred_idc));
 388
 389     se(vpic->pic_init_qp - 26, pic_init_qp_minus26);
 390     se(mseq_var(pic_init_qs_minus26));
 391     se(vpic_var(chroma_qp_index_offset));
 392
 393     u(1, vpic_field(deblocking_filter_control_present_flag));
 394     u(1, vpic_field(constrained_intra_pred_flag));
 395     u(1, vpic_field(redundant_pic_cnt_present_flag));
 396     u(1, vpic_field(transform_8x8_mode_flag));
 397
 398     u(1, vpic_field(pic_scaling_matrix_present_flag));
 399     if (vpic->pic_fields.bits.pic_scaling_matrix_present_flag) {
 400         av_assert0(0 && "scaling matrices not supported");
 401     }
 402
 403     se(vpic_var(second_chroma_qp_index_offset));
 404
 405     vaapi_encode_h264_write_trailing_rbsp(pbc);
 406 }
 407
 408 static void vaapi_encode_h264_write_slice_header2(PutBitContext *pbc,
 409                                                   VAAPIEncodeContext *ctx,
 410                                                   VAAPIEncodePicture *pic,
 411                                                   VAAPIEncodeSlice *slice)
 412 {
 413     VAEncSequenceParameterBufferH264  *vseq = ctx->codec_sequence_params;
 414     VAEncPictureParameterBufferH264   *vpic = pic->codec_picture_params;
 415     VAEncSliceParameterBufferH264   *vslice = slice->codec_slice_params;
 416     VAAPIEncodeH264Context            *priv = ctx->priv_data;
 417     VAAPIEncodeH264MiscSequenceParams *mseq = &priv->misc_sequence_params;
 418     VAAPIEncodeH264Slice            *pslice = slice->priv_data;
 419     VAAPIEncodeH264MiscSliceParams  *mslice = &pslice->misc_slice_params;
 420
 421     vaapi_encode_h264_write_nal_header(pbc, mslice->nal_unit_type,
 422                                        mslice->nal_ref_idc);
 423
 424     ue(vslice->macroblock_address, first_mb_in_slice);
 425     ue(vslice_var(slice_type));
 426     ue(vpic_var(pic_parameter_set_id));
 427
 428     if (mseq->separate_colour_plane_flag) {
 429         u(2, mslice_var(colour_plane_id));
 430     }
 431
 432     u(4 + vseq->seq_fields.bits.log2_max_frame_num_minus4,
 433       (vpic->frame_num &
 434        ((1 << (4 + vseq->seq_fields.bits.log2_max_frame_num_minus4)) - 1)),
 435       frame_num);
 436
 437     if (!vseq->seq_fields.bits.frame_mbs_only_flag) {
 438         u(1, mslice_var(field_pic_flag));
 439         if (mslice->field_pic_flag)
 440             u(1, mslice_var(bottom_field_flag));
 441     }
 442
 443     if (vpic->pic_fields.bits.idr_pic_flag) {
 444         ue(vslice_var(idr_pic_id));
 445     }
 446
 447     if (vseq->seq_fields.bits.pic_order_cnt_type == 0) {
 448         u(4 + vseq->seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4,
 449           vslice_var(pic_order_cnt_lsb));
 450         if (mseq->bottom_field_pic_order_in_frame_present_flag &&
 451             !mslice->field_pic_flag) {
 452             se(vslice_var(delta_pic_order_cnt_bottom));
 453         }
 454     }
 455
 456     if (vseq->seq_fields.bits.pic_order_cnt_type == 1 &&
 457         !vseq->seq_fields.bits.delta_pic_order_always_zero_flag) {
 458         se(vslice_var(delta_pic_order_cnt[0]));
 459         if (mseq->bottom_field_pic_order_in_frame_present_flag &&
 460             !mslice->field_pic_flag) {
 461             se(vslice_var(delta_pic_order_cnt[1]));
 462         }
 463     }
 464
 465     if (vpic->pic_fields.bits.redundant_pic_cnt_present_flag) {
 466         ue(mslice_var(redundant_pic_cnt));
 467     }
 468
 469     if (vslice->slice_type == SLICE_TYPE_B) {
 470         u(1, vslice_var(direct_spatial_mv_pred_flag));
 471     }
 472
 473     if (vslice->slice_type == SLICE_TYPE_P ||
 474         vslice->slice_type == SLICE_TYPE_SP ||
 475         vslice->slice_type == SLICE_TYPE_B) {
 476         u(1, vslice_var(num_ref_idx_active_override_flag));
 477         if (vslice->num_ref_idx_active_override_flag) {
 478             ue(vslice_var(num_ref_idx_l0_active_minus1));
 479             if (vslice->slice_type == SLICE_TYPE_B)
 480                 ue(vslice_var(num_ref_idx_l1_active_minus1));
 481         }
 482     }
 483
 484     if (mslice->nal_unit_type == 20 || mslice->nal_unit_type == 21) {
 485         av_assert0(0 && "no MVC support");
 486     } else {
 487         if (vslice->slice_type % 5 != 2 && vslice->slice_type % 5 != 4) {
 488             u(1, mslice_var(ref_pic_list_modification_flag_l0));
 489             if (mslice->ref_pic_list_modification_flag_l0) {
 490                 av_assert0(0 && "ref pic list modification");
 491             }
 492         }
 493         if (vslice->slice_type % 5 == 1) {
 494             u(1, mslice_var(ref_pic_list_modification_flag_l1));
 495             if (mslice->ref_pic_list_modification_flag_l1) {
 496                 av_assert0(0 && "ref pic list modification");
 497             }
 498         }
 499     }
 500
 501     if ((vpic->pic_fields.bits.weighted_pred_flag &&
 502          (vslice->slice_type == SLICE_TYPE_P ||
 503           vslice->slice_type == SLICE_TYPE_SP)) ||
 504         (vpic->pic_fields.bits.weighted_bipred_idc == 1 &&
 505          vslice->slice_type == SLICE_TYPE_B)) {
 506         av_assert0(0 && "prediction weights not supported");
 507     }
 508
 509     av_assert0(mslice->nal_ref_idc > 0 ==
 510                vpic->pic_fields.bits.reference_pic_flag);
 511     if (mslice->nal_ref_idc != 0) {
 512         if (vpic->pic_fields.bits.idr_pic_flag) {
 513             u(1, mslice_var(no_output_of_prior_pics_flag));
 514             u(1, mslice_var(long_term_reference_flag));
 515         } else {
 516             u(1, mslice_var(adaptive_ref_pic_marking_mode_flag));
 517             if (mslice->adaptive_ref_pic_marking_mode_flag) {
 518                 av_assert0(0 && "MMCOs not supported");
 519             }
 520         }
 521     }
 522
 523     if (vpic->pic_fields.bits.entropy_coding_mode_flag &&
 524         vslice->slice_type != SLICE_TYPE_I &&
 525         vslice->slice_type != SLICE_TYPE_SI) {
 526         ue(vslice_var(cabac_init_idc));
 527     }
 528
 529     se(vslice_var(slice_qp_delta));
 530     if (vslice->slice_type == SLICE_TYPE_SP ||
 531         vslice->slice_type == SLICE_TYPE_SI) {
 532         if (vslice->slice_type == SLICE_TYPE_SP)
 533             u(1, mslice_var(sp_for_switch_flag));
 534         se(mslice_var(slice_qs_delta));
 535     }
 536
 537     if (vpic->pic_fields.bits.deblocking_filter_control_present_flag) {
 538         ue(vslice_var(disable_deblocking_filter_idc));
 539         if (vslice->disable_deblocking_filter_idc != 1) {
 540             se(vslice_var(slice_alpha_c0_offset_div2));
 541             se(vslice_var(slice_beta_offset_div2));
 542         }
 543     }
 544
 545     if (mseq->num_slice_groups_minus1 > 0 &&
 546         mseq->slice_group_map_type >= 3 && mseq->slice_group_map_type <= 5) {
 547         av_assert0(0 && "slice groups not supported");
 548     }
 549
 550     // No alignment - this need not be a byte boundary.
 551 }
 552
 553 static void vaapi_encode_h264_write_buffering_period(PutBitContext *pbc,
 554                                                      VAAPIEncodeContext *ctx,
 555                                                      VAAPIEncodePicture *pic)
 556 {
 557     VAAPIEncodeH264Context            *priv = ctx->priv_data;
 558     VAAPIEncodeH264MiscSequenceParams *mseq = &priv->misc_sequence_params;
 559     VAEncPictureParameterBufferH264   *vpic = pic->codec_picture_params;
 560     int i;
 561
 562     ue(vpic_var(seq_parameter_set_id));
 563
 564     if (mseq->nal_hrd_parameters_present_flag) {
 565         for (i = 0; i <= mseq->cpb_cnt_minus1; i++) {
 566             u(mseq->initial_cpb_removal_delay_length_minus1 + 1,
 567               mseq_var(initial_cpb_removal_delay));
 568             u(mseq->initial_cpb_removal_delay_length_minus1 + 1,
 569               mseq_var(initial_cpb_removal_delay_offset));
 570         }
 571     }
 572     if (mseq->vcl_hrd_parameters_present_flag) {
 573         av_assert0(0 && "vcl hrd parameters not supported");
 574     }
 575 }
 576
 577 static void vaapi_encode_h264_write_pic_timing(PutBitContext *pbc,
 578                                                VAAPIEncodeContext *ctx,
 579                                                VAAPIEncodePicture *pic)
 580 {
 581     VAEncSequenceParameterBufferH264  *vseq = ctx->codec_sequence_params;
 582     VAAPIEncodeH264Context            *priv = ctx->priv_data;
 583     VAAPIEncodeH264MiscSequenceParams *mseq = &priv->misc_sequence_params;
 584     int i, num_clock_ts;
 585
 586     if (mseq->nal_hrd_parameters_present_flag ||
 587         mseq->vcl_hrd_parameters_present_flag) {
 588         u(mseq->cpb_removal_delay_length_minus1 + 1,
 589           2 * vseq->num_units_in_tick * priv->cpb_delay,
 590           cpb_removal_delay);
 591         u(mseq->dpb_output_delay_length_minus1 + 1,
 592           2 * vseq->num_units_in_tick * priv->dpb_delay,
 593           dpb_output_delay);
 594     }
 595     if (mseq->pic_struct_present_flag) {
 596         u(4, mseq_var(pic_struct));
 597         num_clock_ts = (mseq->pic_struct <= 2 ? 1 :
 598                         mseq->pic_struct <= 4 ? 2 :
 599                         mseq->pic_struct <= 8 ? 3 : 0);
 600         for (i = 0; i < num_clock_ts; i++) {
 601             u(1, 0, clock_timestamp_flag[i]);
 602             // No full timestamp information.
 603         }
 604     }
 605 }
 606
 607 static void vaapi_encode_h264_write_identifier(PutBitContext *pbc,
 608                                                VAAPIEncodeContext *ctx,
 609                                                VAAPIEncodePicture *pic)
 610 {
 611     const char *lavc   = LIBAVCODEC_IDENT;
 612     const char *vaapi  = VA_VERSION_S;
 613     const char *driver = vaQueryVendorString(ctx->hwctx->display);
 614     char tmp[256];
 615     int i;
 616
 617     // Random (version 4) ISO 11578 UUID.
 618     uint8_t uuid[16] = {
 619         0x59, 0x94, 0x8b, 0x28, 0x11, 0xec, 0x45, 0xaf,
 620         0x96, 0x75, 0x19, 0xd4, 0x1f, 0xea, 0xa9, 0x4d,
 621     };
 622
 623     for (i = 0; i < 16; i++)
 624         u(8, uuid[i], uuid_iso_iec_11578);
 625
 626     snprintf(tmp, sizeof(tmp), "%s / VAAPI %s / %s", lavc, vaapi, driver);
 627     for (i = 0; i < sizeof(tmp) && tmp[i]; i++)
 628         u(8, tmp[i], user_data_payload_byte);
 629 }
 630
 631 static void vaapi_encode_h264_write_sei(PutBitContext *pbc,
 632                                         VAAPIEncodeContext *ctx,
 633                                         VAAPIEncodePicture *pic)
 634 {
 635     VAAPIEncodeH264Context *priv = ctx->priv_data;
 636     PutBitContext payload_bits;
 637     char payload[256];
 638     int payload_type, payload_size, i;
 639     void (*write_payload)(PutBitContext *pbc,
 640                           VAAPIEncodeContext *ctx,
 641                           VAAPIEncodePicture *pic) = NULL;
 642
 643     vaapi_encode_h264_write_nal_header(pbc, H264_NAL_SEI, 0);
 644
 645     for (payload_type = 0; payload_type < 64; payload_type++) {
 646         switch (payload_type) {
 647         case H264_SEI_TYPE_BUFFERING_PERIOD:
 648             if (!priv->send_timing_sei ||
 649                 pic->type != PICTURE_TYPE_IDR)
 650                 continue;
 651             write_payload = &vaapi_encode_h264_write_buffering_period;
 652             break;
 653         case H264_SEI_TYPE_PIC_TIMING:
 654             if (!priv->send_timing_sei)
 655                 continue;
 656             write_payload = &vaapi_encode_h264_write_pic_timing;
 657             break;
 658         case H264_SEI_TYPE_USER_DATA_UNREGISTERED:
 659             if (pic->encode_order != 0)
 660                 continue;
 661             write_payload = &vaapi_encode_h264_write_identifier;
 662             break;
 663         default:
 664             continue;
 665         }
 666
 667         init_put_bits(&payload_bits, payload, sizeof(payload));
 668         write_payload(&payload_bits, ctx, pic);
 669         if (put_bits_count(&payload_bits) & 7) {
 670             write_u(&payload_bits, 1, 1, bit_equal_to_one);
 671             while (put_bits_count(&payload_bits) & 7)
 672                 write_u(&payload_bits, 1, 0, bit_equal_to_zero);
 673         }
 674         payload_size = put_bits_count(&payload_bits) / 8;
 675         flush_put_bits(&payload_bits);
 676
 677         u(8, payload_type, last_payload_type_byte);
 678         u(8, payload_size, last_payload_size_byte);
 679         for (i = 0; i < payload_size; i++)
 680             u(8, payload[i] & 0xff, sei_payload);
 681     }
 682
 683     vaapi_encode_h264_write_trailing_rbsp(pbc);
 684 }
 685
 686 static int vaapi_encode_h264_write_sequence_header(AVCodecContext *avctx,
 687                                                    char *data, size_t *data_len)
 688 {
 689     VAAPIEncodeContext *ctx = avctx->priv_data;
 690     PutBitContext pbc;
 691     char tmp[256];
 692     int err;
 693     size_t nal_len, bit_len, bit_pos, next_len;
 694
 695     bit_len = *data_len;
 696     bit_pos = 0;
 697
 698     init_put_bits(&pbc, tmp, sizeof(tmp));
 699     vaapi_encode_h264_write_sps(&pbc, ctx);
 700     nal_len = put_bits_count(&pbc);
 701     flush_put_bits(&pbc);
 702
 703     next_len = bit_len - bit_pos;
 704     err = ff_vaapi_encode_h26x_nal_unit_to_byte_stream(data + bit_pos / 8,
 705                                                        &next_len,
 706                                                        tmp, nal_len);
 707     if (err < 0)
 708         return err;
 709     bit_pos += next_len;
 710
 711     init_put_bits(&pbc, tmp, sizeof(tmp));
 712     vaapi_encode_h264_write_pps(&pbc, ctx);
 713     nal_len = put_bits_count(&pbc);
 714     flush_put_bits(&pbc);
 715
 716     next_len = bit_len - bit_pos;
 717     err = ff_vaapi_encode_h26x_nal_unit_to_byte_stream(data + bit_pos / 8,
 718                                                        &next_len,
 719                                                        tmp, nal_len);
 720     if (err < 0)
 721         return err;
 722     bit_pos += next_len;
 723
 724     *data_len = bit_pos;
 725     return 0;
 726 }
 727
 728 static int vaapi_encode_h264_write_slice_header(AVCodecContext *avctx,
 729                                                 VAAPIEncodePicture *pic,
 730                                                 VAAPIEncodeSlice *slice,
 731                                                 char *data, size_t *data_len)
 732 {
 733     VAAPIEncodeContext *ctx = avctx->priv_data;
 734     PutBitContext pbc;
 735     char tmp[256];
 736     size_t header_len;
 737
 738     init_put_bits(&pbc, tmp, sizeof(tmp));
 739     vaapi_encode_h264_write_slice_header2(&pbc, ctx, pic, slice);
 740     header_len = put_bits_count(&pbc);
 741     flush_put_bits(&pbc);
 742
 743     return ff_vaapi_encode_h26x_nal_unit_to_byte_stream(data, data_len,
 744                                                         tmp, header_len);
 745 }
 746
 747 static int vaapi_encode_h264_write_extra_header(AVCodecContext *avctx,
 748                                                 VAAPIEncodePicture *pic,
 749                                                 int index, int *type,
 750                                                 char *data, size_t *data_len)
 751 {
 752     VAAPIEncodeContext *ctx = avctx->priv_data;
 753     PutBitContext pbc;
 754     char tmp[256];
 755     size_t header_len;
 756
 757     if (index == 0 && ctx->va_rc_mode == VA_RC_CBR) {
 758         *type = VAEncPackedHeaderH264_SEI;
 759
 760         init_put_bits(&pbc, tmp, sizeof(tmp));
 761         vaapi_encode_h264_write_sei(&pbc, ctx, pic);
 762         header_len = put_bits_count(&pbc);
 763         flush_put_bits(&pbc);
 764
 765         return ff_vaapi_encode_h26x_nal_unit_to_byte_stream(data, data_len,
 766                                                             tmp, header_len);
 767
 768     } else {
 769         return AVERROR_EOF;
 770     }
 771 }
 772
 773 static int vaapi_encode_h264_init_sequence_params(AVCodecContext *avctx)
 774 {
 775     VAAPIEncodeContext                 *ctx = avctx->priv_data;
 776     VAEncSequenceParameterBufferH264  *vseq = ctx->codec_sequence_params;
 777     VAEncPictureParameterBufferH264   *vpic = ctx->codec_picture_params;
 778     VAAPIEncodeH264Context            *priv = ctx->priv_data;
 779     VAAPIEncodeH264MiscSequenceParams *mseq = &priv->misc_sequence_params;
 780     VAAPIEncodeH264Options             *opt =
 781         (VAAPIEncodeH264Options*)ctx->codec_options_data;
 782     int i;
 783
 784     {
 785         vseq->seq_parameter_set_id = 0;
 786
 787         vseq->level_idc = avctx->level;
 788
 789         vseq->max_num_ref_frames = 1 + (avctx->max_b_frames > 0);
 790
 791         vseq->picture_width_in_mbs  = priv->mb_width;
 792         vseq->picture_height_in_mbs = priv->mb_height;
 793
 794         vseq->seq_fields.bits.chroma_format_idc = 1;
 795         vseq->seq_fields.bits.frame_mbs_only_flag = 1;
 796         vseq->seq_fields.bits.direct_8x8_inference_flag = 1;
 797         vseq->seq_fields.bits.log2_max_frame_num_minus4 = 4;
 798         vseq->seq_fields.bits.pic_order_cnt_type = 0;
 799         vseq->seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 =
 800             av_clip(av_log2(avctx->max_b_frames + 1) - 2, 0, 12);
 801
 802         if (avctx->width  != ctx->surface_width ||
 803             avctx->height != ctx->surface_height) {
 804             vseq->frame_cropping_flag = 1;
 805
 806             vseq->frame_crop_left_offset   = 0;
 807             vseq->frame_crop_right_offset  =
 808                 (ctx->surface_width - avctx->width) / 2;
 809             vseq->frame_crop_top_offset    = 0;
 810             vseq->frame_crop_bottom_offset =
 811                 (ctx->surface_height - avctx->height) / 2;
 812         } else {
 813             vseq->frame_cropping_flag = 0;
 814         }
 815
 816         vseq->vui_parameters_present_flag = 1;
 817         if (avctx->sample_aspect_ratio.num != 0) {
 818             vseq->vui_fields.bits.aspect_ratio_info_present_flag = 1;
 819             // There is a large enum of these which we could support
 820             // individually rather than using the generic X/Y form?
 821             if (avctx->sample_aspect_ratio.num ==
 822                 avctx->sample_aspect_ratio.den) {
 823                 vseq->aspect_ratio_idc = 1;
 824             } else {
 825                 vseq->aspect_ratio_idc = 255; // Extended SAR.
 826                 vseq->sar_width  = avctx->sample_aspect_ratio.num;
 827                 vseq->sar_height = avctx->sample_aspect_ratio.den;
 828             }
 829         }
 830         if (avctx->color_primaries != AVCOL_PRI_UNSPECIFIED ||
 831             avctx->color_trc       != AVCOL_TRC_UNSPECIFIED ||
 832             avctx->colorspace      != AVCOL_SPC_UNSPECIFIED) {
 833             mseq->video_signal_type_present_flag = 1;
 834             mseq->video_format             = 5; // Unspecified.
 835             mseq->video_full_range_flag    = 0;
 836             mseq->colour_description_present_flag = 1;
 837             // These enums are derived from the standard and hence
 838             // we can just use the values directly.
 839             mseq->colour_primaries         = avctx->color_primaries;
 840             mseq->transfer_characteristics = avctx->color_trc;
 841             mseq->matrix_coefficients      = avctx->colorspace;
 842         }
 843
 844         vseq->vui_fields.bits.bitstream_restriction_flag = 1;
 845         mseq->motion_vectors_over_pic_boundaries_flag = 1;
 846         mseq->max_bytes_per_pic_denom = 0;
 847         mseq->max_bits_per_mb_denom   = 0;
 848         vseq->vui_fields.bits.log2_max_mv_length_horizontal = 16;
 849         vseq->vui_fields.bits.log2_max_mv_length_vertical   = 16;
 850
 851         mseq->max_num_reorder_frames = (avctx->max_b_frames > 0);
 852         mseq->max_dec_pic_buffering  = vseq->max_num_ref_frames;
 853
 854         vseq->bits_per_second = avctx->bit_rate;
 855
 856         vseq->vui_fields.bits.timing_info_present_flag = 1;
 857         if (avctx->framerate.num > 0 && avctx->framerate.den > 0) {
 858             vseq->num_units_in_tick = avctx->framerate.den;
 859             vseq->time_scale        = 2 * avctx->framerate.num;
 860             mseq->fixed_frame_rate_flag = 1;
 861         } else {
 862             vseq->num_units_in_tick = avctx->time_base.num;
 863             vseq->time_scale        = 2 * avctx->time_base.den;
 864             mseq->fixed_frame_rate_flag = 0;
 865         }
 866
 867         if (ctx->va_rc_mode == VA_RC_CBR) {
 868             priv->send_timing_sei = 1;
 869             mseq->nal_hrd_parameters_present_flag = 1;
 870
 871             mseq->cpb_cnt_minus1 = 0;
 872
 873             // Try to scale these to a sensible range so that the
 874             // golomb encode of the value is not overlong.
 875             mseq->bit_rate_scale =
 876                 av_clip_uintp2(av_log2(avctx->bit_rate) - 15 - 6, 4);
 877             mseq->bit_rate_value_minus1[0] =
 878                 (avctx->bit_rate >> mseq->bit_rate_scale + 6) - 1;
 879
 880             mseq->cpb_size_scale =
 881                 av_clip_uintp2(av_log2(ctx->hrd_params.hrd.buffer_size) - 15 - 4, 4);
 882             mseq->cpb_size_value_minus1[0] =
 883                 (ctx->hrd_params.hrd.buffer_size >> mseq->cpb_size_scale + 4) - 1;
 884
 885             // CBR mode isn't actually available here, despite naming.
 886             mseq->cbr_flag[0] = 0;
 887
 888             mseq->initial_cpb_removal_delay_length_minus1 = 23;
 889             mseq->cpb_removal_delay_length_minus1         = 23;
 890             mseq->dpb_output_delay_length_minus1          = 7;
 891             mseq->time_offset_length = 0;
 892
 893             // This calculation can easily overflow 32 bits.
 894             mseq->initial_cpb_removal_delay = 90000 *
 895                 (uint64_t)ctx->hrd_params.hrd.initial_buffer_fullness /
 896                 ctx->hrd_params.hrd.buffer_size;
 897
 898             mseq->initial_cpb_removal_delay_offset = 0;
 899         } else {
 900             priv->send_timing_sei = 0;
 901             mseq->nal_hrd_parameters_present_flag = 0;
 902         }
 903
 904         vseq->intra_period     = avctx->gop_size;
 905         vseq->intra_idr_period = avctx->gop_size;
 906         vseq->ip_period        = ctx->b_per_p + 1;
 907     }
 908
 909     {
 910         vpic->CurrPic.picture_id = VA_INVALID_ID;
 911         vpic->CurrPic.flags      = VA_PICTURE_H264_INVALID;
 912
 913         for (i = 0; i < FF_ARRAY_ELEMS(vpic->ReferenceFrames); i++) {
 914             vpic->ReferenceFrames[i].picture_id = VA_INVALID_ID;
 915             vpic->ReferenceFrames[i].flags      = VA_PICTURE_H264_INVALID;
 916         }
 917
 918         vpic->coded_buf = VA_INVALID_ID;
 919
 920         vpic->pic_parameter_set_id = 0;
 921         vpic->seq_parameter_set_id = 0;
 922
 923         vpic->num_ref_idx_l0_active_minus1 = 0;
 924         vpic->num_ref_idx_l1_active_minus1 = 0;
 925
 926         vpic->pic_fields.bits.entropy_coding_mode_flag =
 927             opt->coder ? ((avctx->profile & 0xff) != 66) : 0;
 928         vpic->pic_fields.bits.weighted_pred_flag = 0;
 929         vpic->pic_fields.bits.weighted_bipred_idc = 0;
 930         vpic->pic_fields.bits.transform_8x8_mode_flag =
 931             ((avctx->profile & 0xff) >= 100);
 932
 933         vpic->pic_init_qp = priv->fixed_qp_idr;
 934     }
 935
 936     {
 937         mseq->profile_idc = avctx->profile & 0xff;
 938
 939         if (avctx->profile & FF_PROFILE_H264_CONSTRAINED)
 940             mseq->constraint_set1_flag = 1;
 941         if (avctx->profile & FF_PROFILE_H264_INTRA)
 942             mseq->constraint_set3_flag = 1;
 943     }
 944
 945     return 0;
 946 }
 947
 948 static int vaapi_encode_h264_init_picture_params(AVCodecContext *avctx,
 949                                                  VAAPIEncodePicture *pic)
 950 {
 951     VAAPIEncodeContext                *ctx = avctx->priv_data;
 952     VAEncSequenceParameterBufferH264 *vseq = ctx->codec_sequence_params;
 953     VAEncPictureParameterBufferH264  *vpic = pic->codec_picture_params;
 954     VAAPIEncodeH264Context           *priv = ctx->priv_data;
 955     int i;
 956
 957     if (pic->type == PICTURE_TYPE_IDR) {
 958         av_assert0(pic->display_order == pic->encode_order);
 959         vpic->frame_num = 0;
 960         priv->next_frame_num = 1;
 961         priv->cpb_delay = 0;
 962         priv->last_idr_frame = pic->display_order;
 963     } else {
 964         vpic->frame_num = priv->next_frame_num;
 965         if (pic->type != PICTURE_TYPE_B) {
 966             // nal_ref_idc != 0
 967             ++priv->next_frame_num;
 968         }
 969         ++priv->cpb_delay;
 970     }
 971     priv->dpb_delay = pic->display_order - pic->encode_order + 1;
 972
 973     vpic->frame_num = vpic->frame_num &
 974         ((1 << (4 + vseq->seq_fields.bits.log2_max_frame_num_minus4)) - 1);
 975
 976     vpic->CurrPic.picture_id          = pic->recon_surface;
 977     vpic->CurrPic.frame_idx           = vpic->frame_num;
 978     vpic->CurrPic.flags               = 0;
 979     vpic->CurrPic.TopFieldOrderCnt    = pic->display_order - priv->last_idr_frame;
 980     vpic->CurrPic.BottomFieldOrderCnt = pic->display_order - priv->last_idr_frame;
 981
 982     for (i = 0; i < pic->nb_refs; i++) {
 983         VAAPIEncodePicture *ref = pic->refs[i];
 984         av_assert0(ref && ref->encode_order < pic->encode_order);
 985         vpic->ReferenceFrames[i].picture_id = ref->recon_surface;
 986         vpic->ReferenceFrames[i].frame_idx  = ref->encode_order;
 987         vpic->ReferenceFrames[i].flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
 988         vpic->ReferenceFrames[i].TopFieldOrderCnt    = ref->display_order - priv->last_idr_frame;
 989         vpic->ReferenceFrames[i].BottomFieldOrderCnt = ref->display_order - priv->last_idr_frame;
 990     }
 991     for (; i < FF_ARRAY_ELEMS(vpic->ReferenceFrames); i++) {
 992         vpic->ReferenceFrames[i].picture_id = VA_INVALID_ID;
 993         vpic->ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
 994     }
 995
 996     vpic->coded_buf = pic->output_buffer;
 997
 998     vpic->pic_fields.bits.idr_pic_flag = (pic->type == PICTURE_TYPE_IDR);
 999     vpic->pic_fields.bits.reference_pic_flag = (pic->type != PICTURE_TYPE_B);
1000
1001     pic->nb_slices = 1;
1002
1003     return 0;
1004 }
1005
1006 static int vaapi_encode_h264_init_slice_params(AVCodecContext *avctx,
1007                                                VAAPIEncodePicture *pic,
1008                                                VAAPIEncodeSlice *slice)
1009 {
1010     VAAPIEncodeContext                 *ctx = avctx->priv_data;
1011     VAEncSequenceParameterBufferH264  *vseq = ctx->codec_sequence_params;
1012     VAEncPictureParameterBufferH264   *vpic = pic->codec_picture_params;
1013     VAEncSliceParameterBufferH264   *vslice = slice->codec_slice_params;
1014     VAAPIEncodeH264Context            *priv = ctx->priv_data;
1015     VAAPIEncodeH264Slice            *pslice;
1016     VAAPIEncodeH264MiscSliceParams  *mslice;
1017     int i;
1018
1019     slice->priv_data = av_mallocz(sizeof(*pslice));
1020     if (!slice->priv_data)
1021         return AVERROR(ENOMEM);
1022     pslice = slice->priv_data;
1023     mslice = &pslice->misc_slice_params;
1024
1025     if (pic->type == PICTURE_TYPE_IDR)
1026         mslice->nal_unit_type = H264_NAL_IDR_SLICE;
1027     else
1028         mslice->nal_unit_type = H264_NAL_SLICE;
1029
1030     switch (pic->type) {
1031     case PICTURE_TYPE_IDR:
1032         vslice->slice_type  = SLICE_TYPE_I;
1033         mslice->nal_ref_idc = 3;
1034         break;
1035     case PICTURE_TYPE_I:
1036         vslice->slice_type  = SLICE_TYPE_I;
1037         mslice->nal_ref_idc = 2;
1038         break;
1039     case PICTURE_TYPE_P:
1040         vslice->slice_type  = SLICE_TYPE_P;
1041         mslice->nal_ref_idc = 1;
1042         break;
1043     case PICTURE_TYPE_B:
1044         vslice->slice_type  = SLICE_TYPE_B;
1045         mslice->nal_ref_idc = 0;
1046         break;
1047     default:
1048         av_assert0(0 && "invalid picture type");
1049     }
1050
1051     // Only one slice per frame.
1052     vslice->macroblock_address = 0;
1053     vslice->num_macroblocks = priv->mb_width * priv->mb_height;
1054
1055     vslice->macroblock_info = VA_INVALID_ID;
1056
1057     vslice->pic_parameter_set_id = vpic->pic_parameter_set_id;
1058     vslice->idr_pic_id = priv->idr_pic_count++;
1059
1060     vslice->pic_order_cnt_lsb = (pic->display_order - priv->last_idr_frame) &
1061         ((1 << (4 + vseq->seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4)) - 1);
1062
1063     for (i = 0; i < FF_ARRAY_ELEMS(vslice->RefPicList0); i++) {
1064         vslice->RefPicList0[i].picture_id = VA_INVALID_ID;
1065         vslice->RefPicList0[i].flags      = VA_PICTURE_H264_INVALID;
1066         vslice->RefPicList1[i].picture_id = VA_INVALID_ID;
1067         vslice->RefPicList1[i].flags      = VA_PICTURE_H264_INVALID;
1068     }
1069
1070     av_assert0(pic->nb_refs <= 2);
1071     if (pic->nb_refs >= 1) {
1072         // Backward reference for P- or B-frame.
1073         av_assert0(pic->type == PICTURE_TYPE_P ||
1074                    pic->type == PICTURE_TYPE_B);
1075
1076         vslice->num_ref_idx_l0_active_minus1 = 0;
1077         vslice->RefPicList0[0] = vpic->ReferenceFrames[0];
1078     }
1079     if (pic->nb_refs >= 2) {
1080         // Forward reference for B-frame.
1081         av_assert0(pic->type == PICTURE_TYPE_B);
1082
1083         vslice->num_ref_idx_l1_active_minus1 = 0;
1084         vslice->RefPicList1[0] = vpic->ReferenceFrames[1];
1085     }
1086
1087     if (pic->type == PICTURE_TYPE_B)
1088         vslice->slice_qp_delta = priv->fixed_qp_b - vpic->pic_init_qp;
1089     else if (pic->type == PICTURE_TYPE_P)
1090         vslice->slice_qp_delta = priv->fixed_qp_p - vpic->pic_init_qp;
1091     else
1092         vslice->slice_qp_delta = priv->fixed_qp_idr - vpic->pic_init_qp;
1093
1094     vslice->direct_spatial_mv_pred_flag = 1;
1095
1096     return 0;
1097 }
1098
1099 static av_cold int vaapi_encode_h264_configure(AVCodecContext *avctx)
1100 {
1101     VAAPIEncodeContext      *ctx = avctx->priv_data;
1102     VAAPIEncodeH264Context *priv = ctx->priv_data;
1103     VAAPIEncodeH264Options  *opt = ctx->codec_options;
1104
1105     priv->mb_width  = FFALIGN(avctx->width,  16) / 16;
1106     priv->mb_height = FFALIGN(avctx->height, 16) / 16;
1107
1108     if (ctx->va_rc_mode == VA_RC_CQP) {
1109         priv->fixed_qp_p = opt->qp;
1110         if (avctx->i_quant_factor > 0.0)
1111             priv->fixed_qp_idr = (int)((priv->fixed_qp_p * avctx->i_quant_factor +
1112                                         avctx->i_quant_offset) + 0.5);
1113         else
1114             priv->fixed_qp_idr = priv->fixed_qp_p;
1115         if (avctx->b_quant_factor > 0.0)
1116             priv->fixed_qp_b = (int)((priv->fixed_qp_p * avctx->b_quant_factor +
1117                                       avctx->b_quant_offset) + 0.5);
1118         else
1119             priv->fixed_qp_b = priv->fixed_qp_p;
1120
1121         av_log(avctx, AV_LOG_DEBUG, "Using fixed QP = "
1122                "%d / %d / %d for IDR- / P- / B-frames.\n",
1123                priv->fixed_qp_idr, priv->fixed_qp_p, priv->fixed_qp_b);
1124
1125     } else if (ctx->va_rc_mode == VA_RC_CBR ||
1126                ctx->va_rc_mode == VA_RC_VBR) {
1127         // These still need to be  set for pic_init_qp/slice_qp_delta.
1128         priv->fixed_qp_idr = 26;
1129         priv->fixed_qp_p   = 26;
1130         priv->fixed_qp_b   = 26;
1131
1132         av_log(avctx, AV_LOG_DEBUG, "Using %s-bitrate = %"PRId64" bps.\n",
1133                ctx->va_rc_mode == VA_RC_CBR ? "constant" : "variable",
1134                avctx->bit_rate);
1135
1136     } else {
1137         av_assert0(0 && "Invalid RC mode.");
1138     }
1139
1140     if (avctx->compression_level == FF_COMPRESSION_DEFAULT)
1141         avctx->compression_level = opt->quality;
1142
1143     return 0;
1144 }
1145
1146 static const VAAPIEncodeType vaapi_encode_type_h264 = {
1147     .priv_data_size        = sizeof(VAAPIEncodeH264Context),
1148
1149     .configure             = &vaapi_encode_h264_configure,
1150
1151     .sequence_params_size  = sizeof(VAEncSequenceParameterBufferH264),
1152     .init_sequence_params  = &vaapi_encode_h264_init_sequence_params,
1153
1154     .picture_params_size   = sizeof(VAEncPictureParameterBufferH264),
1155     .init_picture_params   = &vaapi_encode_h264_init_picture_params,
1156
1157     .slice_params_size     = sizeof(VAEncSliceParameterBufferH264),
1158     .init_slice_params     = &vaapi_encode_h264_init_slice_params,
1159
1160     .sequence_header_type  = VAEncPackedHeaderSequence,
1161     .write_sequence_header = &vaapi_encode_h264_write_sequence_header,
1162
1163     .slice_header_type     = VAEncPackedHeaderH264_Slice,
1164     .write_slice_header    = &vaapi_encode_h264_write_slice_header,
1165
1166     .write_extra_header    = &vaapi_encode_h264_write_extra_header,
1167 };
1168
1169 static av_cold int vaapi_encode_h264_init(AVCodecContext *avctx)
1170 {
1171     VAAPIEncodeContext     *ctx = avctx->priv_data;
1172     VAAPIEncodeH264Options *opt =
1173         (VAAPIEncodeH264Options*)ctx->codec_options_data;
1174
1175     ctx->codec = &vaapi_encode_type_h264;
1176
1177     switch (avctx->profile) {
1178     case FF_PROFILE_H264_BASELINE:
1179         av_log(avctx, AV_LOG_WARNING, "H.264 baseline profile is not "
1180                "supported, using constrained baseline profile instead.\n");
1181         avctx->profile = FF_PROFILE_H264_CONSTRAINED_BASELINE;
1182     case FF_PROFILE_H264_CONSTRAINED_BASELINE:
1183         ctx->va_profile = VAProfileH264ConstrainedBaseline;
1184         if (avctx->max_b_frames != 0) {
1185             avctx->max_b_frames = 0;
1186             av_log(avctx, AV_LOG_WARNING, "H.264 constrained baseline profile "
1187                    "doesn't support encoding with B frames, disabling them.\n");
1188         }
1189         break;
1190     case FF_PROFILE_H264_MAIN:
1191         ctx->va_profile = VAProfileH264Main;
1192         break;
1193     case FF_PROFILE_H264_EXTENDED:
1194         av_log(avctx, AV_LOG_ERROR, "H.264 extended profile "
1195                "is not supported.\n");
1196         return AVERROR_PATCHWELCOME;
1197     case FF_PROFILE_UNKNOWN:
1198     case FF_PROFILE_H264_HIGH:
1199         ctx->va_profile = VAProfileH264High;
1200         break;
1201     case FF_PROFILE_H264_HIGH_10:
1202     case FF_PROFILE_H264_HIGH_10_INTRA:
1203         av_log(avctx, AV_LOG_ERROR, "H.264 10-bit profiles "
1204                "are not supported.\n");
1205         return AVERROR_PATCHWELCOME;
1206     case FF_PROFILE_H264_HIGH_422:
1207     case FF_PROFILE_H264_HIGH_422_INTRA:
1208     case FF_PROFILE_H264_HIGH_444:
1209     case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
1210     case FF_PROFILE_H264_HIGH_444_INTRA:
1211     case FF_PROFILE_H264_CAVLC_444:
1212         av_log(avctx, AV_LOG_ERROR, "H.264 non-4:2:0 profiles "
1213                "are not supported.\n");
1214         return AVERROR_PATCHWELCOME;
1215     default:
1216         av_log(avctx, AV_LOG_ERROR, "Unknown H.264 profile %d.\n",
1217                avctx->profile);
1218         return AVERROR(EINVAL);
1219     }
1220     if (opt->low_power) {
1221 #if VA_CHECK_VERSION(0, 39, 2)
1222         ctx->va_entrypoint = VAEntrypointEncSliceLP;
1223 #else
1224         av_log(avctx, AV_LOG_ERROR, "Low-power encoding is not "
1225                "supported with this VAAPI version.\n");
1226         return AVERROR(EINVAL);
1227 #endif
1228     } else {
1229         ctx->va_entrypoint = VAEntrypointEncSlice;
1230     }
1231
1232     // Only 8-bit encode is supported.
1233     ctx->va_rt_format = VA_RT_FORMAT_YUV420;
1234
1235     if (avctx->bit_rate > 0) {
1236         if (avctx->rc_max_rate == avctx->bit_rate)
1237             ctx->va_rc_mode = VA_RC_CBR;
1238         else
1239             ctx->va_rc_mode = VA_RC_VBR;
1240     } else
1241         ctx->va_rc_mode = VA_RC_CQP;
1242
1243     ctx->va_packed_headers =
1244         VA_ENC_PACKED_HEADER_SEQUENCE | // SPS and PPS.
1245         VA_ENC_PACKED_HEADER_SLICE    | // Slice headers.
1246         VA_ENC_PACKED_HEADER_MISC;      // SEI.
1247
1248     ctx->surface_width  = FFALIGN(avctx->width,  16);
1249     ctx->surface_height = FFALIGN(avctx->height, 16);
1250
1251     return ff_vaapi_encode_init(avctx);
1252 }
1253
1254 #define OFFSET(x) (offsetof(VAAPIEncodeContext, codec_options_data) + \
1255                    offsetof(VAAPIEncodeH264Options, x))
1256 #define FLAGS (AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM)
1257 static const AVOption vaapi_encode_h264_options[] = {
1258     { "qp", "Constant QP (for P-frames; scaled by qfactor/qoffset for I/B)",
1259       OFFSET(qp), AV_OPT_TYPE_INT, { .i64 = 20 }, 0, 52, FLAGS },
1260     { "quality", "Set encode quality (trades off against speed, higher is faster)",
1261       OFFSET(quality), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 8, FLAGS },
1262     { "low_power", "Use low-power encoding mode (experimental: only supported "
1263       "on some platforms, does not support all features)",
1264       OFFSET(low_power), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 1, FLAGS },
1265     { "coder", "Entropy coder type",
1266       OFFSET(coder), AV_OPT_TYPE_INT, { .i64 = 1 }, 0, 1, FLAGS, "coder" },
1267         { "cavlc", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 0 }, INT_MIN, INT_MAX, FLAGS, "coder" },
1268         { "cabac", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 1 }, INT_MIN, INT_MAX, FLAGS, "coder" },
1269         { "vlc",   NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 0 }, INT_MIN, INT_MAX, FLAGS, "coder" },
1270         { "ac",    NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 1 }, INT_MIN, INT_MAX, FLAGS, "coder" },
1271     { NULL },
1272 };
1273
1274 static const AVCodecDefault vaapi_encode_h264_defaults[] = {
1275     { "profile",        "100" },
1276     { "level",          "51"  },
1277     { "b",              "0"   },
1278     { "bf",             "2"   },
1279     { "g",              "120" },
1280     { "i_qfactor",      "1"   },
1281     { "i_qoffset",      "0"   },
1282     { "b_qfactor",      "6/5" },
1283     { "b_qoffset",      "0"   },
1284     { "qmin",           "0"   },
1285     { NULL },
1286 };
1287
1288 static const AVClass vaapi_encode_h264_class = {
1289     .class_name = "h264_vaapi",
1290     .item_name  = av_default_item_name,
1291     .option     = vaapi_encode_h264_options,
1292     .version    = LIBAVUTIL_VERSION_INT,
1293 };
1294
1295 AVCodec ff_h264_vaapi_encoder = {
1296     .name           = "h264_vaapi",
1297     .long_name      = NULL_IF_CONFIG_SMALL("H.264/AVC (VAAPI)"),
1298     .type           = AVMEDIA_TYPE_VIDEO,
1299     .id             = AV_CODEC_ID_H264,
1300     .priv_data_size = (sizeof(VAAPIEncodeContext) +
1301                        sizeof(VAAPIEncodeH264Options)),
1302     .init           = &vaapi_encode_h264_init,
1303     .encode2        = &ff_vaapi_encode2,
1304     .close          = &ff_vaapi_encode_close,
1305     .priv_class     = &vaapi_encode_h264_class,
1306     .capabilities   = AV_CODEC_CAP_DELAY,
1307     .defaults       = vaapi_encode_h264_defaults,
1308     .pix_fmts = (const enum AVPixelFormat[]) {
1309         AV_PIX_FMT_VAAPI,
1310         AV_PIX_FMT_NONE,
1311     },
1312 };