libavcodec/vaapi_hevc.c

   1 /*
   2  * HEVC HW decode acceleration through VA API
   3  *
   4  * Copyright (C) 2015 Timo Rothenpieler <timo@rothenpieler.org>
   5  *
   6  * This file is part of FFmpeg.
   7  *
   8  * FFmpeg is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU Lesser General Public
  10  * License as published by the Free Software Foundation; either
  11  * version 2.1 of the License, or (at your option) any later version.
  12  *
  13  * FFmpeg is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * Lesser General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU Lesser General Public
  19  * License along with FFmpeg; if not, write to the Free Software
  20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  21  */
  22
  23 #include <va/va.h>
  24 #include <va/va_dec_hevc.h>
  25
  26 #include "avcodec.h"
  27 #include "hevcdec.h"
  28 #include "hwaccel_internal.h"
  29 #include "vaapi_decode.h"
  30 #include "vaapi_hevc.h"
  31 #include "h265_profile_level.h"
  32
  33 typedef struct VAAPIDecodePictureHEVC {
  34 #if VA_CHECK_VERSION(1, 2, 0)
  35     VAPictureParameterBufferHEVCExtension pic_param;
  36     VASliceParameterBufferHEVCExtension last_slice_param;
  37 #else
  38     VAPictureParameterBufferHEVC pic_param;
  39     VASliceParameterBufferHEVC last_slice_param;
  40 #endif
  41     const uint8_t *last_buffer;
  42     size_t         last_size;
  43
  44     VAAPIDecodePicture pic;
  45 } VAAPIDecodePictureHEVC;
  46
  47 static void init_vaapi_pic(VAPictureHEVC *va_pic)
  48 {
  49     va_pic->picture_id    = VA_INVALID_ID;
  50     va_pic->flags         = VA_PICTURE_HEVC_INVALID;
  51     va_pic->pic_order_cnt = 0;
  52 }
  53
  54 static void fill_vaapi_pic(VAPictureHEVC *va_pic, const HEVCFrame *pic, int rps_type)
  55 {
  56     va_pic->picture_id    = ff_vaapi_get_surface_id(pic->frame);
  57     va_pic->pic_order_cnt = pic->poc;
  58     va_pic->flags         = rps_type;
  59
  60     if (pic->flags & HEVC_FRAME_FLAG_LONG_REF)
  61         va_pic->flags |= VA_PICTURE_HEVC_LONG_TERM_REFERENCE;
  62
  63     if (pic->frame->flags & AV_FRAME_FLAG_INTERLACED) {
  64         va_pic->flags |= VA_PICTURE_HEVC_FIELD_PIC;
  65
  66         if (!(pic->frame->flags & AV_FRAME_FLAG_TOP_FIELD_FIRST))
  67             va_pic->flags |= VA_PICTURE_HEVC_BOTTOM_FIELD;
  68     }
  69 }
  70
  71 static int find_frame_rps_type(const HEVCContext *h, const HEVCFrame *pic)
  72 {
  73     VASurfaceID pic_surf = ff_vaapi_get_surface_id(pic->frame);
  74     const HEVCFrame *current_picture = h->ref;
  75     int i;
  76
  77     for (i = 0; i < h->rps[ST_CURR_BEF].nb_refs; i++) {
  78         if (pic_surf == ff_vaapi_get_surface_id(h->rps[ST_CURR_BEF].ref[i]->frame))
  79             return VA_PICTURE_HEVC_RPS_ST_CURR_BEFORE;
  80     }
  81
  82     for (i = 0; i < h->rps[ST_CURR_AFT].nb_refs; i++) {
  83         if (pic_surf == ff_vaapi_get_surface_id(h->rps[ST_CURR_AFT].ref[i]->frame))
  84             return VA_PICTURE_HEVC_RPS_ST_CURR_AFTER;
  85     }
  86
  87     for (i = 0; i < h->rps[LT_CURR].nb_refs; i++) {
  88         if (pic_surf == ff_vaapi_get_surface_id(h->rps[LT_CURR].ref[i]->frame))
  89             return VA_PICTURE_HEVC_RPS_LT_CURR;
  90     }
  91
  92     if (h->ps.pps->pps_curr_pic_ref_enabled_flag && current_picture->poc == pic->poc)
  93         return VA_PICTURE_HEVC_LONG_TERM_REFERENCE;
  94
  95     return 0;
  96 }
  97
  98 static void fill_vaapi_reference_frames(const HEVCContext *h, VAPictureParameterBufferHEVC *pp)
  99 {
 100     const HEVCFrame *current_picture = h->ref;
 101     int i, j, rps_type;
 102
 103     for (i = 0, j = 0; i < FF_ARRAY_ELEMS(pp->ReferenceFrames); i++) {
 104         const HEVCFrame *frame = NULL;
 105
 106         while (!frame && j < FF_ARRAY_ELEMS(h->DPB)) {
 107             if ((&h->DPB[j] != current_picture || h->ps.pps->pps_curr_pic_ref_enabled_flag) &&
 108                 (h->DPB[j].flags & (HEVC_FRAME_FLAG_LONG_REF | HEVC_FRAME_FLAG_SHORT_REF)))
 109                 frame = &h->DPB[j];
 110             j++;
 111         }
 112
 113         init_vaapi_pic(&pp->ReferenceFrames[i]);
 114
 115         if (frame) {
 116             rps_type = find_frame_rps_type(h, frame);
 117             fill_vaapi_pic(&pp->ReferenceFrames[i], frame, rps_type);
 118         }
 119     }
 120 }
 121
 122 static int vaapi_hevc_start_frame(AVCodecContext          *avctx,
 123                                   av_unused const uint8_t *buffer,
 124                                   av_unused uint32_t       size)
 125 {
 126     const HEVCContext        *h = avctx->priv_data;
 127     VAAPIDecodePictureHEVC *pic = h->ref->hwaccel_picture_private;
 128     const HEVCSPS          *sps = h->ps.sps;
 129     const HEVCPPS          *pps = h->ps.pps;
 130
 131     const ScalingList *scaling_list = NULL;
 132     int pic_param_size, err, i;
 133
 134 #if VA_CHECK_VERSION(1, 2, 0)
 135     int num_comps, pre_palette_size;
 136 #endif
 137
 138     VAPictureParameterBufferHEVC *pic_param = (VAPictureParameterBufferHEVC *)&pic->pic_param;
 139
 140     pic->pic.output_surface = ff_vaapi_get_surface_id(h->ref->frame);
 141
 142     *pic_param = (VAPictureParameterBufferHEVC) {
 143         .pic_width_in_luma_samples                    = sps->width,
 144         .pic_height_in_luma_samples                   = sps->height,
 145         .log2_min_luma_coding_block_size_minus3       = sps->log2_min_cb_size - 3,
 146         .sps_max_dec_pic_buffering_minus1             = sps->temporal_layer[sps->max_sub_layers - 1].max_dec_pic_buffering - 1,
 147         .log2_diff_max_min_luma_coding_block_size     = sps->log2_diff_max_min_coding_block_size,
 148         .log2_min_transform_block_size_minus2         = sps->log2_min_tb_size - 2,
 149         .log2_diff_max_min_transform_block_size       = sps->log2_max_trafo_size  - sps->log2_min_tb_size,
 150         .max_transform_hierarchy_depth_inter          = sps->max_transform_hierarchy_depth_inter,
 151         .max_transform_hierarchy_depth_intra          = sps->max_transform_hierarchy_depth_intra,
 152         .num_short_term_ref_pic_sets                  = sps->nb_st_rps,
 153         .num_long_term_ref_pic_sps                    = sps->num_long_term_ref_pics_sps,
 154         .num_ref_idx_l0_default_active_minus1         = pps->num_ref_idx_l0_default_active - 1,
 155         .num_ref_idx_l1_default_active_minus1         = pps->num_ref_idx_l1_default_active - 1,
 156         .init_qp_minus26                              = pps->pic_init_qp_minus26,
 157         .pps_cb_qp_offset                             = pps->cb_qp_offset,
 158         .pps_cr_qp_offset                             = pps->cr_qp_offset,
 159         .pcm_sample_bit_depth_luma_minus1             = sps->pcm.bit_depth - 1,
 160         .pcm_sample_bit_depth_chroma_minus1           = sps->pcm.bit_depth_chroma - 1,
 161         .log2_min_pcm_luma_coding_block_size_minus3   = sps->pcm.log2_min_pcm_cb_size - 3,
 162         .log2_diff_max_min_pcm_luma_coding_block_size = sps->pcm.log2_max_pcm_cb_size - sps->pcm.log2_min_pcm_cb_size,
 163         .diff_cu_qp_delta_depth                       = pps->diff_cu_qp_delta_depth,
 164         .pps_beta_offset_div2                         = pps->beta_offset / 2,
 165         .pps_tc_offset_div2                           = pps->tc_offset / 2,
 166         .log2_parallel_merge_level_minus2             = pps->log2_parallel_merge_level - 2,
 167         .bit_depth_luma_minus8                        = sps->bit_depth - 8,
 168         .bit_depth_chroma_minus8                      = sps->bit_depth - 8,
 169         .log2_max_pic_order_cnt_lsb_minus4            = sps->log2_max_poc_lsb - 4,
 170         .num_extra_slice_header_bits                  = pps->num_extra_slice_header_bits,
 171         .pic_fields.bits = {
 172             .chroma_format_idc                          = sps->chroma_format_idc,
 173             .tiles_enabled_flag                         = pps->tiles_enabled_flag,
 174             .separate_colour_plane_flag                 = sps->separate_colour_plane_flag,
 175             .pcm_enabled_flag                           = sps->pcm_enabled_flag,
 176             .scaling_list_enabled_flag                  = sps->scaling_list_enable_flag,
 177             .transform_skip_enabled_flag                = pps->transform_skip_enabled_flag,
 178             .amp_enabled_flag                           = sps->amp_enabled_flag,
 179             .strong_intra_smoothing_enabled_flag        = sps->sps_strong_intra_smoothing_enable_flag,
 180             .sign_data_hiding_enabled_flag              = pps->sign_data_hiding_flag,
 181             .constrained_intra_pred_flag                = pps->constrained_intra_pred_flag,
 182             .cu_qp_delta_enabled_flag                   = pps->cu_qp_delta_enabled_flag,
 183             .weighted_pred_flag                         = pps->weighted_pred_flag,
 184             .weighted_bipred_flag                       = pps->weighted_bipred_flag,
 185             .transquant_bypass_enabled_flag             = pps->transquant_bypass_enable_flag,
 186             .entropy_coding_sync_enabled_flag           = pps->entropy_coding_sync_enabled_flag,
 187             .pps_loop_filter_across_slices_enabled_flag = pps->seq_loop_filter_across_slices_enabled_flag,
 188             .loop_filter_across_tiles_enabled_flag      = pps->loop_filter_across_tiles_enabled_flag,
 189             .pcm_loop_filter_disabled_flag              = sps->pcm.loop_filter_disable_flag,
 190         },
 191         .slice_parsing_fields.bits = {
 192             .lists_modification_present_flag             = pps->lists_modification_present_flag,
 193             .long_term_ref_pics_present_flag             = sps->long_term_ref_pics_present_flag,
 194             .sps_temporal_mvp_enabled_flag               = sps->sps_temporal_mvp_enabled_flag,
 195             .cabac_init_present_flag                     = pps->cabac_init_present_flag,
 196             .output_flag_present_flag                    = pps->output_flag_present_flag,
 197             .dependent_slice_segments_enabled_flag       = pps->dependent_slice_segments_enabled_flag,
 198             .pps_slice_chroma_qp_offsets_present_flag    = pps->pic_slice_level_chroma_qp_offsets_present_flag,
 199             .sample_adaptive_offset_enabled_flag         = sps->sao_enabled,
 200             .deblocking_filter_override_enabled_flag     = pps->deblocking_filter_override_enabled_flag,
 201             .pps_disable_deblocking_filter_flag          = pps->disable_dbf,
 202             .slice_segment_header_extension_present_flag = pps->slice_header_extension_present_flag,
 203             .RapPicFlag                                  = IS_IRAP(h),
 204             .IdrPicFlag                                  = IS_IDR(h),
 205             .IntraPicFlag                                = IS_IRAP(h),
 206         },
 207     };
 208
 209     fill_vaapi_pic(&pic_param->CurrPic, h->ref, 0);
 210     fill_vaapi_reference_frames(h, pic_param);
 211
 212     if (pps->tiles_enabled_flag) {
 213         pic_param->num_tile_columns_minus1 = pps->num_tile_columns - 1;
 214         pic_param->num_tile_rows_minus1    = pps->num_tile_rows - 1;
 215
 216         for (i = 0; i < pps->num_tile_columns; i++)
 217             pic_param->column_width_minus1[i] = pps->column_width[i] - 1;
 218
 219         for (i = 0; i < pps->num_tile_rows; i++)
 220             pic_param->row_height_minus1[i] = pps->row_height[i] - 1;
 221     }
 222
 223     if (h->sh.short_term_ref_pic_set_sps_flag == 0 && h->sh.short_term_rps) {
 224         pic_param->st_rps_bits = h->sh.short_term_ref_pic_set_size;
 225     } else {
 226         pic_param->st_rps_bits = 0;
 227     }
 228
 229 #if VA_CHECK_VERSION(1, 2, 0)
 230     if (avctx->profile == AV_PROFILE_HEVC_REXT ||
 231         avctx->profile == AV_PROFILE_HEVC_SCC) {
 232         pic->pic_param.rext = (VAPictureParameterBufferHEVCRext) {
 233             .range_extension_pic_fields.bits  = {
 234                 .transform_skip_rotation_enabled_flag       = sps->transform_skip_rotation_enabled_flag,
 235                 .transform_skip_context_enabled_flag        = sps->transform_skip_context_enabled_flag,
 236                 .implicit_rdpcm_enabled_flag                = sps->implicit_rdpcm_enabled_flag,
 237                 .explicit_rdpcm_enabled_flag                = sps->explicit_rdpcm_enabled_flag,
 238                 .extended_precision_processing_flag         = sps->extended_precision_processing_flag,
 239                 .intra_smoothing_disabled_flag              = sps->intra_smoothing_disabled_flag,
 240                 .high_precision_offsets_enabled_flag        = sps->high_precision_offsets_enabled_flag,
 241                 .persistent_rice_adaptation_enabled_flag    = sps->persistent_rice_adaptation_enabled_flag,
 242                 .cabac_bypass_alignment_enabled_flag        = sps->cabac_bypass_alignment_enabled_flag,
 243                 .cross_component_prediction_enabled_flag    = pps->cross_component_prediction_enabled_flag,
 244                 .chroma_qp_offset_list_enabled_flag         = pps->chroma_qp_offset_list_enabled_flag,
 245             },
 246             .diff_cu_chroma_qp_offset_depth                 = pps->diff_cu_chroma_qp_offset_depth,
 247             .chroma_qp_offset_list_len_minus1               = pps->chroma_qp_offset_list_len_minus1,
 248             .log2_sao_offset_scale_luma                     = pps->log2_sao_offset_scale_luma,
 249             .log2_sao_offset_scale_chroma                   = pps->log2_sao_offset_scale_chroma,
 250             .log2_max_transform_skip_block_size_minus2      = pps->log2_max_transform_skip_block_size - 2,
 251         };
 252
 253         for (i = 0; i < 6; i++)
 254             pic->pic_param.rext.cb_qp_offset_list[i]        = pps->cb_qp_offset_list[i];
 255         for (i = 0; i < 6; i++)
 256             pic->pic_param.rext.cr_qp_offset_list[i]        = pps->cr_qp_offset_list[i];
 257     }
 258
 259     pre_palette_size = pps->pps_palette_predictor_initializers_present_flag ?
 260                        pps->pps_num_palette_predictor_initializers :
 261                        (sps->sps_palette_predictor_initializers_present_flag ?
 262                        sps->sps_num_palette_predictor_initializers :
 263                        0);
 264
 265     if (avctx->profile == AV_PROFILE_HEVC_SCC) {
 266         pic->pic_param.scc = (VAPictureParameterBufferHEVCScc) {
 267             .screen_content_pic_fields.bits = {
 268                 .pps_curr_pic_ref_enabled_flag              = pps->pps_curr_pic_ref_enabled_flag,
 269                 .palette_mode_enabled_flag                  = sps->palette_mode_enabled_flag,
 270                 .motion_vector_resolution_control_idc       = sps->motion_vector_resolution_control_idc,
 271                 .intra_boundary_filtering_disabled_flag     = sps->intra_boundary_filtering_disabled_flag,
 272                 .residual_adaptive_colour_transform_enabled_flag
 273                                                             = pps->residual_adaptive_colour_transform_enabled_flag,
 274                 .pps_slice_act_qp_offsets_present_flag      = pps->pps_slice_act_qp_offsets_present_flag,
 275             },
 276             .palette_max_size                               = sps->palette_max_size,
 277             .delta_palette_max_predictor_size               = sps->delta_palette_max_predictor_size,
 278             .predictor_palette_size                         = pre_palette_size,
 279             .pps_act_y_qp_offset_plus5                      = pps->residual_adaptive_colour_transform_enabled_flag ?
 280                                                               pps->pps_act_y_qp_offset + 5 : 0,
 281             .pps_act_cb_qp_offset_plus5                     = pps->residual_adaptive_colour_transform_enabled_flag ?
 282                                                               pps->pps_act_cb_qp_offset + 5 : 0,
 283             .pps_act_cr_qp_offset_plus3                     = pps->residual_adaptive_colour_transform_enabled_flag ?
 284                                                               pps->pps_act_cr_qp_offset + 3 : 0,
 285         };
 286
 287         num_comps = pps->monochrome_palette_flag ? 1 : 3;
 288         for (int comp = 0; comp < num_comps; comp++)
 289             for (int j = 0; j < pre_palette_size; j++)
 290                 pic->pic_param.scc.predictor_palette_entries[comp][j] =
 291                     pps->pps_palette_predictor_initializers_present_flag ?
 292                     pps->pps_palette_predictor_initializer[comp][j]:
 293                     sps->sps_palette_predictor_initializer[comp][j];
 294     }
 295
 296 #endif
 297     pic_param_size = avctx->profile >= AV_PROFILE_HEVC_REXT ?
 298                             sizeof(pic->pic_param) : sizeof(VAPictureParameterBufferHEVC);
 299
 300     err = ff_vaapi_decode_make_param_buffer(avctx, &pic->pic,
 301                                             VAPictureParameterBufferType,
 302                                             &pic->pic_param, pic_param_size);
 303     if (err < 0)
 304         goto fail;
 305
 306     if (pps->scaling_list_data_present_flag)
 307         scaling_list = &pps->scaling_list;
 308     else if (sps->scaling_list_enable_flag)
 309         scaling_list = &sps->scaling_list;
 310
 311     if (scaling_list) {
 312         VAIQMatrixBufferHEVC iq_matrix;
 313         int j;
 314
 315         for (i = 0; i < 6; i++) {
 316             for (j = 0; j < 16; j++)
 317                 iq_matrix.ScalingList4x4[i][j] = scaling_list->sl[0][i][j];
 318             for (j = 0; j < 64; j++) {
 319                 iq_matrix.ScalingList8x8[i][j]   = scaling_list->sl[1][i][j];
 320                 iq_matrix.ScalingList16x16[i][j] = scaling_list->sl[2][i][j];
 321                 if (i < 2)
 322                     iq_matrix.ScalingList32x32[i][j] = scaling_list->sl[3][i * 3][j];
 323             }
 324             iq_matrix.ScalingListDC16x16[i] = scaling_list->sl_dc[0][i];
 325             if (i < 2)
 326                 iq_matrix.ScalingListDC32x32[i] = scaling_list->sl_dc[1][i * 3];
 327         }
 328
 329         err = ff_vaapi_decode_make_param_buffer(avctx, &pic->pic,
 330                                                 VAIQMatrixBufferType,
 331                                                 &iq_matrix, sizeof(iq_matrix));
 332         if (err < 0)
 333             goto fail;
 334     }
 335
 336     return 0;
 337
 338 fail:
 339     ff_vaapi_decode_cancel(avctx, &pic->pic);
 340     return err;
 341 }
 342
 343 static int vaapi_hevc_end_frame(AVCodecContext *avctx)
 344 {
 345     const HEVCContext        *h = avctx->priv_data;
 346     VAAPIDecodePictureHEVC *pic = h->ref->hwaccel_picture_private;
 347     VASliceParameterBufferHEVC *last_slice_param = (VASliceParameterBufferHEVC *)&pic->last_slice_param;
 348     int ret;
 349
 350     int slice_param_size = avctx->profile >= AV_PROFILE_HEVC_REXT ?
 351                             sizeof(pic->last_slice_param) : sizeof(VASliceParameterBufferHEVC);
 352
 353     if (pic->last_size) {
 354         last_slice_param->LongSliceFlags.fields.LastSliceOfPic = 1;
 355         ret = ff_vaapi_decode_make_slice_buffer(avctx, &pic->pic,
 356                                                 &pic->last_slice_param, slice_param_size,
 357                                                 pic->last_buffer, pic->last_size);
 358         if (ret < 0)
 359             goto fail;
 360     }
 361
 362
 363     ret = ff_vaapi_decode_issue(avctx, &pic->pic);
 364     if (ret < 0)
 365         goto fail;
 366
 367     return 0;
 368 fail:
 369     ff_vaapi_decode_cancel(avctx, &pic->pic);
 370     return ret;
 371 }
 372
 373 static void fill_pred_weight_table(AVCodecContext *avctx,
 374                                    const HEVCContext *h,
 375                                    const SliceHeader *sh,
 376                                    VASliceParameterBufferHEVC *slice_param)
 377 {
 378     int i;
 379 #if VA_CHECK_VERSION(1, 2, 0)
 380     int is_rext = avctx->profile >= AV_PROFILE_HEVC_REXT;
 381 #else
 382     int is_rext = 0;
 383     if (avctx->profile >= AV_PROFILE_HEVC_REXT)
 384         av_log(avctx, AV_LOG_WARNING, "Please consider to update to VAAPI 1.2.0 "
 385                "or above, which can support REXT related setting correctly.\n");
 386 #endif
 387
 388     memset(slice_param->delta_luma_weight_l0,   0, sizeof(slice_param->delta_luma_weight_l0));
 389     memset(slice_param->delta_luma_weight_l1,   0, sizeof(slice_param->delta_luma_weight_l1));
 390     memset(slice_param->luma_offset_l0,         0, sizeof(slice_param->luma_offset_l0));
 391     memset(slice_param->luma_offset_l1,         0, sizeof(slice_param->luma_offset_l1));
 392     memset(slice_param->delta_chroma_weight_l0, 0, sizeof(slice_param->delta_chroma_weight_l0));
 393     memset(slice_param->delta_chroma_weight_l1, 0, sizeof(slice_param->delta_chroma_weight_l1));
 394     memset(slice_param->ChromaOffsetL0,         0, sizeof(slice_param->ChromaOffsetL0));
 395     memset(slice_param->ChromaOffsetL1,         0, sizeof(slice_param->ChromaOffsetL1));
 396
 397     slice_param->delta_chroma_log2_weight_denom = 0;
 398     slice_param->luma_log2_weight_denom         = 0;
 399
 400     if (sh->slice_type == HEVC_SLICE_I ||
 401         (sh->slice_type == HEVC_SLICE_P && !h->ps.pps->weighted_pred_flag) ||
 402         (sh->slice_type == HEVC_SLICE_B && !h->ps.pps->weighted_bipred_flag))
 403         return;
 404
 405     slice_param->luma_log2_weight_denom = sh->luma_log2_weight_denom;
 406
 407     if (h->ps.sps->chroma_format_idc) {
 408         slice_param->delta_chroma_log2_weight_denom = sh->chroma_log2_weight_denom - sh->luma_log2_weight_denom;
 409     }
 410
 411     for (i = 0; i < 15 && i < sh->nb_refs[L0]; i++) {
 412         slice_param->delta_luma_weight_l0[i] = sh->luma_weight_l0[i] - (1 << sh->luma_log2_weight_denom);
 413         slice_param->delta_chroma_weight_l0[i][0] = sh->chroma_weight_l0[i][0] - (1 << sh->chroma_log2_weight_denom);
 414         slice_param->delta_chroma_weight_l0[i][1] = sh->chroma_weight_l0[i][1] - (1 << sh->chroma_log2_weight_denom);
 415         if (!is_rext) {
 416             slice_param->luma_offset_l0[i] = sh->luma_offset_l0[i];
 417             slice_param->ChromaOffsetL0[i][0] = sh->chroma_offset_l0[i][0];
 418             slice_param->ChromaOffsetL0[i][1] = sh->chroma_offset_l0[i][1];
 419         }
 420     }
 421
 422     if (sh->slice_type == HEVC_SLICE_B) {
 423         for (i = 0; i < 15 && i < sh->nb_refs[L1]; i++) {
 424             slice_param->delta_luma_weight_l1[i] = sh->luma_weight_l1[i] - (1 << sh->luma_log2_weight_denom);
 425             slice_param->delta_chroma_weight_l1[i][0] = sh->chroma_weight_l1[i][0] - (1 << sh->chroma_log2_weight_denom);
 426             slice_param->delta_chroma_weight_l1[i][1] = sh->chroma_weight_l1[i][1] - (1 << sh->chroma_log2_weight_denom);
 427             if (!is_rext) {
 428                 slice_param->luma_offset_l1[i] = sh->luma_offset_l1[i];
 429                 slice_param->ChromaOffsetL1[i][0] = sh->chroma_offset_l1[i][0];
 430                 slice_param->ChromaOffsetL1[i][1] = sh->chroma_offset_l1[i][1];
 431             }
 432         }
 433     }
 434 }
 435
 436 static uint8_t get_ref_pic_index(const HEVCContext *h, const HEVCFrame *frame)
 437 {
 438     VAAPIDecodePictureHEVC *pic = h->ref->hwaccel_picture_private;
 439     VAPictureParameterBufferHEVC *pp = (VAPictureParameterBufferHEVC *)&pic->pic_param;
 440     uint8_t i;
 441
 442     if (!frame)
 443         return 0xff;
 444
 445     for (i = 0; i < FF_ARRAY_ELEMS(pp->ReferenceFrames); i++) {
 446         VASurfaceID pid = pp->ReferenceFrames[i].picture_id;
 447         int poc = pp->ReferenceFrames[i].pic_order_cnt;
 448         if (pid != VA_INVALID_ID && pid == ff_vaapi_get_surface_id(frame->frame) && poc == frame->poc)
 449             return i;
 450     }
 451
 452     return 0xff;
 453 }
 454
 455 static int vaapi_hevc_decode_slice(AVCodecContext *avctx,
 456                                    const uint8_t  *buffer,
 457                                    uint32_t        size)
 458 {
 459     const HEVCContext        *h = avctx->priv_data;
 460     const SliceHeader       *sh = &h->sh;
 461     VAAPIDecodePictureHEVC *pic = h->ref->hwaccel_picture_private;
 462     VASliceParameterBufferHEVC *last_slice_param = (VASliceParameterBufferHEVC *)&pic->last_slice_param;
 463
 464     int slice_param_size = avctx->profile >= AV_PROFILE_HEVC_REXT ?
 465                             sizeof(pic->last_slice_param) : sizeof(VASliceParameterBufferHEVC);
 466
 467     int nb_list = (sh->slice_type == HEVC_SLICE_B) ?
 468                   2 : (sh->slice_type == HEVC_SLICE_I ? 0 : 1);
 469
 470     int err, i, list_idx;
 471
 472     if (!sh->first_slice_in_pic_flag) {
 473         err = ff_vaapi_decode_make_slice_buffer(avctx, &pic->pic,
 474                                                 &pic->last_slice_param, slice_param_size,
 475                                                 pic->last_buffer, pic->last_size);
 476         pic->last_buffer = NULL;
 477         pic->last_size   = 0;
 478         if (err) {
 479             ff_vaapi_decode_cancel(avctx, &pic->pic);
 480             return err;
 481         }
 482     }
 483
 484     *last_slice_param = (VASliceParameterBufferHEVC) {
 485         .slice_data_size               = size,
 486         .slice_data_offset             = 0,
 487         .slice_data_flag               = VA_SLICE_DATA_FLAG_ALL,
 488         /* Add 1 to the bits count here to account for the byte_alignment bit, which
 489          * always is at least one bit and not accounted for otherwise. */
 490         .slice_data_byte_offset        = (get_bits_count(&h->HEVClc->gb) + 1 + 7) / 8,
 491         .slice_segment_address         = sh->slice_segment_addr,
 492         .slice_qp_delta                = sh->slice_qp_delta,
 493         .slice_cb_qp_offset            = sh->slice_cb_qp_offset,
 494         .slice_cr_qp_offset            = sh->slice_cr_qp_offset,
 495         .slice_beta_offset_div2        = sh->beta_offset / 2,
 496         .slice_tc_offset_div2          = sh->tc_offset / 2,
 497         .collocated_ref_idx            = sh->slice_temporal_mvp_enabled_flag ? sh->collocated_ref_idx : 0xFF,
 498         .five_minus_max_num_merge_cand = sh->slice_type == HEVC_SLICE_I ? 0 : 5 - sh->max_num_merge_cand,
 499         .num_ref_idx_l0_active_minus1  = sh->nb_refs[L0] ? sh->nb_refs[L0] - 1 : 0,
 500         .num_ref_idx_l1_active_minus1  = sh->nb_refs[L1] ? sh->nb_refs[L1] - 1 : 0,
 501
 502         .LongSliceFlags.fields = {
 503             .dependent_slice_segment_flag                 = sh->dependent_slice_segment_flag,
 504             .slice_type                                   = sh->slice_type,
 505             .color_plane_id                               = sh->colour_plane_id,
 506             .mvd_l1_zero_flag                             = sh->mvd_l1_zero_flag,
 507             .cabac_init_flag                              = sh->cabac_init_flag,
 508             .slice_temporal_mvp_enabled_flag              = sh->slice_temporal_mvp_enabled_flag,
 509             .slice_deblocking_filter_disabled_flag        = sh->disable_deblocking_filter_flag,
 510             .collocated_from_l0_flag                      = sh->collocated_list == L0 ? 1 : 0,
 511             .slice_loop_filter_across_slices_enabled_flag = sh->slice_loop_filter_across_slices_enabled_flag,
 512             .slice_sao_luma_flag                          = sh->slice_sample_adaptive_offset_flag[0],
 513             .slice_sao_chroma_flag                        = sh->slice_sample_adaptive_offset_flag[1],
 514         },
 515     };
 516
 517     memset(last_slice_param->RefPicList, 0xFF, sizeof(last_slice_param->RefPicList));
 518
 519     for (list_idx = 0; list_idx < nb_list; list_idx++) {
 520         RefPicList *rpl = &h->ref->refPicList[list_idx];
 521
 522         for (i = 0; i < rpl->nb_refs; i++)
 523             last_slice_param->RefPicList[list_idx][i] = get_ref_pic_index(h, rpl->ref[i]);
 524     }
 525
 526     fill_pred_weight_table(avctx, h, sh, last_slice_param);
 527
 528 #if VA_CHECK_VERSION(1, 2, 0)
 529     if (avctx->profile >= AV_PROFILE_HEVC_REXT) {
 530         pic->last_slice_param.rext = (VASliceParameterBufferHEVCRext) {
 531             .slice_ext_flags.bits = {
 532                 .cu_chroma_qp_offset_enabled_flag = sh->cu_chroma_qp_offset_enabled_flag,
 533                 .use_integer_mv_flag = sh->use_integer_mv_flag,
 534             },
 535             .slice_act_y_qp_offset  = sh->slice_act_y_qp_offset,
 536             .slice_act_cb_qp_offset = sh->slice_act_cb_qp_offset,
 537             .slice_act_cr_qp_offset = sh->slice_act_cr_qp_offset,
 538         };
 539         for (i = 0; i < 15 && i < sh->nb_refs[L0]; i++) {
 540             pic->last_slice_param.rext.luma_offset_l0[i] = sh->luma_offset_l0[i];
 541             pic->last_slice_param.rext.ChromaOffsetL0[i][0] = sh->chroma_offset_l0[i][0];
 542             pic->last_slice_param.rext.ChromaOffsetL0[i][1] = sh->chroma_offset_l0[i][1];
 543         }
 544
 545         if (sh->slice_type == HEVC_SLICE_B) {
 546             for (i = 0; i < 15 && i < sh->nb_refs[L1]; i++) {
 547                 pic->last_slice_param.rext.luma_offset_l1[i] = sh->luma_offset_l1[i];
 548                 pic->last_slice_param.rext.ChromaOffsetL1[i][0] = sh->chroma_offset_l1[i][0];
 549                 pic->last_slice_param.rext.ChromaOffsetL1[i][1] = sh->chroma_offset_l1[i][1];
 550             }
 551         }
 552     }
 553 #endif
 554
 555     pic->last_buffer = buffer;
 556     pic->last_size   = size;
 557
 558     return 0;
 559 }
 560
 561 static int ptl_convert(const PTLCommon *general_ptl, H265RawProfileTierLevel *h265_raw_ptl)
 562 {
 563     h265_raw_ptl->general_profile_space = general_ptl->profile_space;
 564     h265_raw_ptl->general_tier_flag     = general_ptl->tier_flag;
 565     h265_raw_ptl->general_profile_idc   = general_ptl->profile_idc;
 566
 567     memcpy(h265_raw_ptl->general_profile_compatibility_flag,
 568                                   general_ptl->profile_compatibility_flag, 32 * sizeof(uint8_t));
 569
 570 #define copy_field(name) h265_raw_ptl->general_ ## name = general_ptl->name
 571     copy_field(progressive_source_flag);
 572     copy_field(interlaced_source_flag);
 573     copy_field(non_packed_constraint_flag);
 574     copy_field(frame_only_constraint_flag);
 575     copy_field(max_12bit_constraint_flag);
 576     copy_field(max_10bit_constraint_flag);
 577     copy_field(max_8bit_constraint_flag);
 578     copy_field(max_422chroma_constraint_flag);
 579     copy_field(max_420chroma_constraint_flag);
 580     copy_field(max_monochrome_constraint_flag);
 581     copy_field(intra_constraint_flag);
 582     copy_field(one_picture_only_constraint_flag);
 583     copy_field(lower_bit_rate_constraint_flag);
 584     copy_field(max_14bit_constraint_flag);
 585     copy_field(inbld_flag);
 586     copy_field(level_idc);
 587 #undef copy_field
 588
 589     return 0;
 590 }
 591
 592 /*
 593  * Find exact va_profile for HEVC Range Extension and Screen Content Coding Extension
 594  */
 595 VAProfile ff_vaapi_parse_hevc_rext_scc_profile(AVCodecContext *avctx)
 596 {
 597     const HEVCContext *h = avctx->priv_data;
 598     const HEVCSPS *sps = h->ps.sps;
 599     const PTL *ptl = &sps->ptl;
 600     const PTLCommon *general_ptl = &ptl->general_ptl;
 601     const H265ProfileDescriptor *profile;
 602     H265RawProfileTierLevel h265_raw_ptl = {0};
 603
 604     /* convert PTLCommon to H265RawProfileTierLevel */
 605     ptl_convert(general_ptl, &h265_raw_ptl);
 606
 607     profile = ff_h265_get_profile(&h265_raw_ptl);
 608     if (!profile) {
 609         av_log(avctx, AV_LOG_WARNING, "HEVC profile is not found.\n");
 610         goto end;
 611     } else {
 612         av_log(avctx, AV_LOG_VERBOSE, "HEVC profile %s is found.\n", profile->name);
 613     }
 614
 615 #if VA_CHECK_VERSION(1, 2, 0)
 616     if (!strcmp(profile->name, "Main 12") ||
 617         !strcmp(profile->name, "Main 12 Intra"))
 618         return VAProfileHEVCMain12;
 619     else if (!strcmp(profile->name, "Main 4:2:2 10") ||
 620         !strcmp(profile->name, "Main 4:2:2 10 Intra"))
 621         return VAProfileHEVCMain422_10;
 622     else if (!strcmp(profile->name, "Main 4:2:2 12") ||
 623         !strcmp(profile->name, "Main 4:2:2 12 Intra"))
 624         return VAProfileHEVCMain422_12;
 625     else if (!strcmp(profile->name, "Main 4:4:4") ||
 626              !strcmp(profile->name, "Main 4:4:4 Intra"))
 627         return VAProfileHEVCMain444;
 628     else if (!strcmp(profile->name, "Main 4:4:4 10") ||
 629              !strcmp(profile->name, "Main 4:4:4 10 Intra"))
 630         return VAProfileHEVCMain444_10;
 631     else if (!strcmp(profile->name, "Main 4:4:4 12") ||
 632              !strcmp(profile->name, "Main 4:4:4 12 Intra"))
 633         return VAProfileHEVCMain444_12;
 634     else if (!strcmp(profile->name, "Screen-Extended Main"))
 635         return VAProfileHEVCSccMain;
 636     else if (!strcmp(profile->name, "Screen-Extended Main 10"))
 637         return VAProfileHEVCSccMain10;
 638     else if (!strcmp(profile->name, "Screen-Extended Main 4:4:4"))
 639         return VAProfileHEVCSccMain444;
 640 #if VA_CHECK_VERSION(1, 8, 0)
 641     else if (!strcmp(profile->name, "Screen-Extended Main 4:4:4 10"))
 642         return VAProfileHEVCSccMain444_10;
 643 #endif
 644 #else
 645     av_log(avctx, AV_LOG_WARNING, "HEVC profile %s is "
 646            "not supported with this VA version.\n", profile->name);
 647 #endif
 648
 649 end:
 650     if (avctx->hwaccel_flags & AV_HWACCEL_FLAG_ALLOW_PROFILE_MISMATCH) {
 651         // Default to selecting Main profile if profile mismatch is allowed
 652         return VAProfileHEVCMain;
 653     } else
 654         return VAProfileNone;
 655 }
 656
 657 const FFHWAccel ff_hevc_vaapi_hwaccel = {
 658     .p.name               = "hevc_vaapi",
 659     .p.type               = AVMEDIA_TYPE_VIDEO,
 660     .p.id                 = AV_CODEC_ID_HEVC,
 661     .p.pix_fmt            = AV_PIX_FMT_VAAPI,
 662     .start_frame          = vaapi_hevc_start_frame,
 663     .end_frame            = vaapi_hevc_end_frame,
 664     .decode_slice         = vaapi_hevc_decode_slice,
 665     .frame_priv_data_size = sizeof(VAAPIDecodePictureHEVC),
 666     .init                 = ff_vaapi_decode_init,
 667     .uninit               = ff_vaapi_decode_uninit,
 668     .frame_params         = ff_vaapi_common_frame_params,
 669     .priv_data_size       = sizeof(VAAPIDecodeContext),
 670     .caps_internal        = HWACCEL_CAP_ASYNC_SAFE,
 671 };