git.sesse.net Git - ffmpeg/blob - libavcodec/dxva2_hevc.c

   1 /*
   2  * DXVA2 HEVC HW acceleration.
   3  *
   4  * copyright (c) 2014 - 2015 Hendrik Leppkes
   5  *
   6  * This file is part of Libav.
   7  *
   8  * Libav is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU Lesser General Public
  10  * License as published by the Free Software Foundation; either
  11  * version 2.1 of the License, or (at your option) any later version.
  12  *
  13  * Libav is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * Lesser General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU Lesser General Public
  19  * License along with Libav; if not, write to the Free Software
  20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  21  */
  22
  23 #include "libavutil/avassert.h"
  24
  25 #include "dxva2_internal.h"
  26 #include "hevc.h"
  27
  28 #define MAX_SLICES 256
  29
  30 struct hevc_dxva2_picture_context {
  31     DXVA_PicParams_HEVC   pp;
  32     DXVA_Qmatrix_HEVC     qm;
  33     unsigned              slice_count;
  34     DXVA_Slice_HEVC_Short slice_short[MAX_SLICES];
  35     const uint8_t         *bitstream;
  36     unsigned              bitstream_size;
  37 };
  38
  39 static void fill_picture_entry(DXVA_PicEntry_HEVC *pic,
  40                                unsigned index, unsigned flag)
  41 {
  42     av_assert0((index & 0x7f) == index && (flag & 0x01) == flag);
  43     pic->bPicEntry = index | (flag << 7);
  44 }
  45
  46 static int get_refpic_index(const DXVA_PicParams_HEVC *pp, int surface_index)
  47 {
  48     int i;
  49     for (i = 0; i < FF_ARRAY_ELEMS(pp->RefPicList); i++) {
  50         if ((pp->RefPicList[i].bPicEntry & 0x7f) == surface_index)
  51           return i;
  52     }
  53     return 0xff;
  54 }
  55
  56 static void fill_picture_parameters(const AVCodecContext *avctx, AVDXVAContext *ctx, const HEVCContext *h,
  57                                     DXVA_PicParams_HEVC *pp)
  58 {
  59     const HEVCFrame *current_picture = h->ref;
  60     int i, j;
  61
  62     memset(pp, 0, sizeof(*pp));
  63
  64     pp->PicWidthInMinCbsY  = h->sps->min_cb_width;
  65     pp->PicHeightInMinCbsY = h->sps->min_cb_height;
  66
  67     pp->wFormatAndSequenceInfoFlags = (h->sps->chroma_format_idc          <<  0) |
  68                                       (h->sps->separate_colour_plane_flag <<  2) |
  69                                       ((h->sps->bit_depth - 8)            <<  3) |
  70                                       ((h->sps->bit_depth - 8)            <<  6) |
  71                                       ((h->sps->log2_max_poc_lsb - 4)     <<  9) |
  72                                       (0                                  << 13) |
  73                                       (0                                  << 14) |
  74                                       (0                                  << 15);
  75
  76     fill_picture_entry(&pp->CurrPic, ff_dxva2_get_surface_index(avctx, ctx, current_picture->frame), 0);
  77
  78     pp->sps_max_dec_pic_buffering_minus1         = h->sps->temporal_layer[h->sps->max_sub_layers - 1].max_dec_pic_buffering - 1;
  79     pp->log2_min_luma_coding_block_size_minus3   = h->sps->log2_min_cb_size - 3;
  80     pp->log2_diff_max_min_luma_coding_block_size = h->sps->log2_diff_max_min_coding_block_size;
  81     pp->log2_min_transform_block_size_minus2     = h->sps->log2_min_tb_size - 2;
  82     pp->log2_diff_max_min_transform_block_size   = h->sps->log2_max_trafo_size  - h->sps->log2_min_tb_size;
  83     pp->max_transform_hierarchy_depth_inter      = h->sps->max_transform_hierarchy_depth_inter;
  84     pp->max_transform_hierarchy_depth_intra      = h->sps->max_transform_hierarchy_depth_intra;
  85     pp->num_short_term_ref_pic_sets              = h->sps->nb_st_rps;
  86     pp->num_long_term_ref_pics_sps               = h->sps->num_long_term_ref_pics_sps;
  87
  88     pp->num_ref_idx_l0_default_active_minus1     = h->pps->num_ref_idx_l0_default_active - 1;
  89     pp->num_ref_idx_l1_default_active_minus1     = h->pps->num_ref_idx_l1_default_active - 1;
  90     pp->init_qp_minus26                          = h->pps->pic_init_qp_minus26;
  91
  92     if (h->sh.short_term_ref_pic_set_sps_flag == 0 && h->sh.short_term_rps) {
  93         pp->ucNumDeltaPocsOfRefRpsIdx            = h->sh.short_term_rps->num_delta_pocs;
  94         pp->wNumBitsForShortTermRPSInSlice       = h->sh.short_term_ref_pic_set_size;
  95     }
  96
  97     pp->dwCodingParamToolFlags = (h->sps->scaling_list_enable_flag               <<  0) |
  98                                  (h->sps->amp_enabled_flag                       <<  1) |
  99                                  (h->sps->sao_enabled                            <<  2) |
 100                                  (h->sps->pcm_enabled_flag                       <<  3) |
 101                                  ((h->sps->pcm_enabled_flag ? (h->sps->pcm.bit_depth - 1) : 0)            <<  4) |
 102                                  ((h->sps->pcm_enabled_flag ? (h->sps->pcm.bit_depth_chroma - 1) : 0)     <<  8) |
 103                                  ((h->sps->pcm_enabled_flag ? (h->sps->pcm.log2_min_pcm_cb_size - 3) : 0) << 12) |
 104                                  ((h->sps->pcm_enabled_flag ? (h->sps->pcm.log2_max_pcm_cb_size - h->sps->pcm.log2_min_pcm_cb_size) : 0) << 14) |
 105                                  (h->sps->pcm.loop_filter_disable_flag           << 16) |
 106                                  (h->sps->long_term_ref_pics_present_flag        << 17) |
 107                                  (h->sps->sps_temporal_mvp_enabled_flag          << 18) |
 108                                  (h->sps->sps_strong_intra_smoothing_enable_flag << 19) |
 109                                  (h->pps->dependent_slice_segments_enabled_flag  << 20) |
 110                                  (h->pps->output_flag_present_flag               << 21) |
 111                                  (h->pps->num_extra_slice_header_bits            << 22) |
 112                                  (h->pps->sign_data_hiding_flag                  << 25) |
 113                                  (h->pps->cabac_init_present_flag                << 26) |
 114                                  (0                                              << 27);
 115
 116     pp->dwCodingSettingPicturePropertyFlags = (h->pps->constrained_intra_pred_flag                <<  0) |
 117                                               (h->pps->transform_skip_enabled_flag                <<  1) |
 118                                               (h->pps->cu_qp_delta_enabled_flag                   <<  2) |
 119                                               (h->pps->pic_slice_level_chroma_qp_offsets_present_flag <<  3) |
 120                                               (h->pps->weighted_pred_flag                         <<  4) |
 121                                               (h->pps->weighted_bipred_flag                       <<  5) |
 122                                               (h->pps->transquant_bypass_enable_flag              <<  6) |
 123                                               (h->pps->tiles_enabled_flag                         <<  7) |
 124                                               (h->pps->entropy_coding_sync_enabled_flag           <<  8) |
 125                                               (h->pps->uniform_spacing_flag                       <<  9) |
 126                                               ((h->pps->tiles_enabled_flag ? h->pps->loop_filter_across_tiles_enabled_flag : 0) << 10) |
 127                                               (h->pps->seq_loop_filter_across_slices_enabled_flag << 11) |
 128                                               (h->pps->deblocking_filter_override_enabled_flag    << 12) |
 129                                               (h->pps->disable_dbf                                << 13) |
 130                                               (h->pps->lists_modification_present_flag            << 14) |
 131                                               (h->pps->slice_header_extension_present_flag        << 15) |
 132                                               (IS_IRAP(h)                                         << 16) |
 133                                               (IS_IDR(h)                                          << 17) |
 134                                               /* IntraPicFlag */
 135                                               (IS_IRAP(h)                                         << 18) |
 136                                               (0                                                  << 19);
 137     pp->pps_cb_qp_offset            = h->pps->cb_qp_offset;
 138     pp->pps_cr_qp_offset            = h->pps->cr_qp_offset;
 139     if (h->pps->tiles_enabled_flag) {
 140         pp->num_tile_columns_minus1 = h->pps->num_tile_columns - 1;
 141         pp->num_tile_rows_minus1    = h->pps->num_tile_rows - 1;
 142
 143         if (!h->pps->uniform_spacing_flag) {
 144             for (i = 0; i < h->pps->num_tile_columns; i++)
 145                 pp->column_width_minus1[i] = h->pps->column_width[i] - 1;
 146
 147             for (i = 0; i < h->pps->num_tile_rows; i++)
 148                 pp->row_height_minus1[i] = h->pps->row_height[i] - 1;
 149         }
 150     }
 151
 152     pp->diff_cu_qp_delta_depth           = h->pps->diff_cu_qp_delta_depth;
 153     pp->pps_beta_offset_div2             = h->pps->beta_offset / 2;
 154     pp->pps_tc_offset_div2               = h->pps->tc_offset / 2;
 155     pp->log2_parallel_merge_level_minus2 = h->pps->log2_parallel_merge_level - 2;
 156     pp->CurrPicOrderCntVal               = h->poc;
 157
 158     // fill RefPicList from the DPB
 159     for (i = 0, j = 0; i < FF_ARRAY_ELEMS(pp->RefPicList); i++) {
 160         const HEVCFrame *frame = NULL;
 161         while (!frame && j < FF_ARRAY_ELEMS(h->DPB)) {
 162             if (&h->DPB[j] != current_picture && (h->DPB[j].flags & (HEVC_FRAME_FLAG_LONG_REF | HEVC_FRAME_FLAG_SHORT_REF)))
 163                 frame = &h->DPB[j];
 164             j++;
 165         }
 166
 167         if (frame) {
 168             fill_picture_entry(&pp->RefPicList[i], ff_dxva2_get_surface_index(avctx, ctx, frame->frame), !!(frame->flags & HEVC_FRAME_FLAG_LONG_REF));
 169             pp->PicOrderCntValList[i] = frame->poc;
 170         } else {
 171             pp->RefPicList[i].bPicEntry = 0xff;
 172             pp->PicOrderCntValList[i]   = 0;
 173         }
 174     }
 175
 176     #define DO_REF_LIST(ref_idx, ref_list) { \
 177         const RefPicList *rpl = &h->rps[ref_idx]; \
 178         for (i = 0, j = 0; i < FF_ARRAY_ELEMS(pp->ref_list); i++) { \
 179             const HEVCFrame *frame = NULL; \
 180             while (!frame && j < rpl->nb_refs) \
 181                 frame = rpl->ref[j++]; \
 182             if (frame) \
 183                 pp->ref_list[i] = get_refpic_index(pp, ff_dxva2_get_surface_index(avctx, ctx, frame->frame)); \
 184             else \
 185                 pp->ref_list[i] = 0xff; \
 186         } \
 187     }
 188
 189     // Fill short term and long term lists
 190     DO_REF_LIST(ST_CURR_BEF, RefPicSetStCurrBefore);
 191     DO_REF_LIST(ST_CURR_AFT, RefPicSetStCurrAfter);
 192     DO_REF_LIST(LT_CURR, RefPicSetLtCurr);
 193
 194     pp->StatusReportFeedbackNumber = 1 + DXVA_CONTEXT_REPORT_ID(avctx, ctx)++;
 195 }
 196
 197 static void fill_scaling_lists(AVDXVAContext *ctx, const HEVCContext *h, DXVA_Qmatrix_HEVC *qm)
 198 {
 199     unsigned i, j, pos;
 200     const ScalingList *sl = h->pps->scaling_list_data_present_flag ?
 201                             &h->pps->scaling_list : &h->sps->scaling_list;
 202
 203     memset(qm, 0, sizeof(*qm));
 204     for (i = 0; i < 6; i++) {
 205         for (j = 0; j < 16; j++) {
 206             pos = 4 * ff_hevc_diag_scan4x4_y[j] + ff_hevc_diag_scan4x4_x[j];
 207             qm->ucScalingLists0[i][j] = sl->sl[0][i][pos];
 208         }
 209
 210         for (j = 0; j < 64; j++) {
 211             pos = 8 * ff_hevc_diag_scan8x8_y[j] + ff_hevc_diag_scan8x8_x[j];
 212             qm->ucScalingLists1[i][j] = sl->sl[1][i][pos];
 213             qm->ucScalingLists2[i][j] = sl->sl[2][i][pos];
 214
 215             if (i < 2)
 216                 qm->ucScalingLists3[i][j] = sl->sl[3][i * 3][pos];
 217         }
 218
 219         qm->ucScalingListDCCoefSizeID2[i] = sl->sl_dc[0][i];
 220         if (i < 2)
 221             qm->ucScalingListDCCoefSizeID3[i] = sl->sl_dc[1][i * 3];
 222     }
 223 }
 224
 225 static void fill_slice_short(DXVA_Slice_HEVC_Short *slice,
 226                              unsigned position, unsigned size)
 227 {
 228     memset(slice, 0, sizeof(*slice));
 229     slice->BSNALunitDataLocation = position;
 230     slice->SliceBytesInBuffer    = size;
 231     slice->wBadSliceChopping     = 0;
 232 }
 233
 234 static int commit_bitstream_and_slice_buffer(AVCodecContext *avctx,
 235                                              DECODER_BUFFER_DESC *bs,
 236                                              DECODER_BUFFER_DESC *sc)
 237 {
 238     const HEVCContext *h = avctx->priv_data;
 239     AVDXVAContext *ctx = avctx->hwaccel_context;
 240     const HEVCFrame *current_picture = h->ref;
 241     struct hevc_dxva2_picture_context *ctx_pic = current_picture->hwaccel_picture_private;
 242     DXVA_Slice_HEVC_Short *slice = NULL;
 243     void     *dxva_data_ptr;
 244     uint8_t  *dxva_data, *current, *end;
 245     unsigned dxva_size;
 246     void     *slice_data;
 247     unsigned slice_size;
 248     unsigned padding;
 249     unsigned i;
 250     unsigned type;
 251
 252     /* Create an annex B bitstream buffer with only slice NAL and finalize slice */
 253 #if CONFIG_D3D11VA
 254     if (avctx->pix_fmt == AV_PIX_FMT_D3D11VA_VLD) {
 255         type = D3D11_VIDEO_DECODER_BUFFER_BITSTREAM;
 256         if (FAILED(ID3D11VideoContext_GetDecoderBuffer(D3D11VA_CONTEXT(ctx)->video_context,
 257                                                        D3D11VA_CONTEXT(ctx)->decoder,
 258                                                        type,
 259                                                        &dxva_size, &dxva_data_ptr)))
 260             return -1;
 261     }
 262 #endif
 263 #if CONFIG_DXVA2
 264     if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD) {
 265         type = DXVA2_BitStreamDateBufferType;
 266         if (FAILED(IDirectXVideoDecoder_GetBuffer(DXVA2_CONTEXT(ctx)->decoder,
 267                                                   type,
 268                                                   &dxva_data_ptr, &dxva_size)))
 269             return -1;
 270     }
 271 #endif
 272
 273     dxva_data = dxva_data_ptr;
 274     current = dxva_data;
 275     end = dxva_data + dxva_size;
 276
 277     for (i = 0; i < ctx_pic->slice_count; i++) {
 278         static const uint8_t start_code[] = { 0, 0, 1 };
 279         static const unsigned start_code_size = sizeof(start_code);
 280         unsigned position, size;
 281
 282         slice = &ctx_pic->slice_short[i];
 283
 284         position = slice->BSNALunitDataLocation;
 285         size     = slice->SliceBytesInBuffer;
 286         if (start_code_size + size > end - current) {
 287             av_log(avctx, AV_LOG_ERROR, "Failed to build bitstream");
 288             break;
 289         }
 290
 291         slice->BSNALunitDataLocation = current - dxva_data;
 292         slice->SliceBytesInBuffer    = start_code_size + size;
 293
 294         memcpy(current, start_code, start_code_size);
 295         current += start_code_size;
 296
 297         memcpy(current, &ctx_pic->bitstream[position], size);
 298         current += size;
 299     }
 300     padding = FFMIN(128 - ((current - dxva_data) & 127), end - current);
 301     if (slice && padding > 0) {
 302         memset(current, 0, padding);
 303         current += padding;
 304
 305         slice->SliceBytesInBuffer += padding;
 306     }
 307 #if CONFIG_D3D11VA
 308     if (avctx->pix_fmt == AV_PIX_FMT_D3D11VA_VLD)
 309         if (FAILED(ID3D11VideoContext_ReleaseDecoderBuffer(D3D11VA_CONTEXT(ctx)->video_context, D3D11VA_CONTEXT(ctx)->decoder, type)))
 310             return -1;
 311 #endif
 312 #if CONFIG_DXVA2
 313     if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD)
 314         if (FAILED(IDirectXVideoDecoder_ReleaseBuffer(DXVA2_CONTEXT(ctx)->decoder, type)))
 315             return -1;
 316 #endif
 317     if (i < ctx_pic->slice_count)
 318         return -1;
 319
 320 #if CONFIG_D3D11VA
 321     if (avctx->pix_fmt == AV_PIX_FMT_D3D11VA_VLD) {
 322         D3D11_VIDEO_DECODER_BUFFER_DESC *dsc11 = bs;
 323         memset(dsc11, 0, sizeof(*dsc11));
 324         dsc11->BufferType           = type;
 325         dsc11->DataSize             = current - dxva_data;
 326         dsc11->NumMBsInBuffer       = 0;
 327
 328         type = D3D11_VIDEO_DECODER_BUFFER_SLICE_CONTROL;
 329     }
 330 #endif
 331 #if CONFIG_DXVA2
 332     if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD) {
 333         DXVA2_DecodeBufferDesc *dsc2 = bs;
 334         memset(dsc2, 0, sizeof(*dsc2));
 335         dsc2->CompressedBufferType = type;
 336         dsc2->DataSize             = current - dxva_data;
 337         dsc2->NumMBsInBuffer       = 0;
 338
 339         type = DXVA2_SliceControlBufferType;
 340     }
 341 #endif
 342
 343     slice_data = ctx_pic->slice_short;
 344     slice_size = ctx_pic->slice_count * sizeof(*ctx_pic->slice_short);
 345
 346     av_assert0(((current - dxva_data) & 127) == 0);
 347     return ff_dxva2_commit_buffer(avctx, ctx, sc,
 348                                   type,
 349                                   slice_data, slice_size, 0);
 350 }
 351
 352
 353 static int dxva2_hevc_start_frame(AVCodecContext *avctx,
 354                                   av_unused const uint8_t *buffer,
 355                                   av_unused uint32_t size)
 356 {
 357     const HEVCContext *h = avctx->priv_data;
 358     AVDXVAContext *ctx = avctx->hwaccel_context;
 359     struct hevc_dxva2_picture_context *ctx_pic = h->ref->hwaccel_picture_private;
 360
 361     if (DXVA_CONTEXT_DECODER(avctx, ctx) == NULL ||
 362         DXVA_CONTEXT_CFG(avctx, ctx) == NULL ||
 363         DXVA_CONTEXT_COUNT(avctx, ctx) <= 0)
 364         return -1;
 365     av_assert0(ctx_pic);
 366
 367     /* Fill up DXVA_PicParams_HEVC */
 368     fill_picture_parameters(avctx, ctx, h, &ctx_pic->pp);
 369
 370     /* Fill up DXVA_Qmatrix_HEVC */
 371     fill_scaling_lists(ctx, h, &ctx_pic->qm);
 372
 373     ctx_pic->slice_count    = 0;
 374     ctx_pic->bitstream_size = 0;
 375     ctx_pic->bitstream      = NULL;
 376     return 0;
 377 }
 378
 379 static int dxva2_hevc_decode_slice(AVCodecContext *avctx,
 380                                    const uint8_t *buffer,
 381                                    uint32_t size)
 382 {
 383     const HEVCContext *h = avctx->priv_data;
 384     const HEVCFrame *current_picture = h->ref;
 385     struct hevc_dxva2_picture_context *ctx_pic = current_picture->hwaccel_picture_private;
 386     unsigned position;
 387
 388     if (ctx_pic->slice_count >= MAX_SLICES)
 389         return -1;
 390
 391     if (!ctx_pic->bitstream)
 392         ctx_pic->bitstream = buffer;
 393     ctx_pic->bitstream_size += size;
 394
 395     position = buffer - ctx_pic->bitstream;
 396     fill_slice_short(&ctx_pic->slice_short[ctx_pic->slice_count], position, size);
 397     ctx_pic->slice_count++;
 398
 399     return 0;
 400 }
 401
 402 static int dxva2_hevc_end_frame(AVCodecContext *avctx)
 403 {
 404     HEVCContext *h = avctx->priv_data;
 405     struct hevc_dxva2_picture_context *ctx_pic = h->ref->hwaccel_picture_private;
 406     int scale = ctx_pic->pp.dwCodingParamToolFlags & 1;
 407     int ret;
 408
 409     if (ctx_pic->slice_count <= 0 || ctx_pic->bitstream_size <= 0)
 410         return -1;
 411
 412     ret = ff_dxva2_common_end_frame(avctx, h->ref->frame,
 413                                     &ctx_pic->pp, sizeof(ctx_pic->pp),
 414                                     scale ? &ctx_pic->qm : NULL, scale ? sizeof(ctx_pic->qm) : 0,
 415                                     commit_bitstream_and_slice_buffer);
 416     return ret;
 417 }
 418
 419 #if CONFIG_HEVC_DXVA2_HWACCEL
 420 AVHWAccel ff_hevc_dxva2_hwaccel = {
 421     .name           = "hevc_dxva2",
 422     .type           = AVMEDIA_TYPE_VIDEO,
 423     .id             = AV_CODEC_ID_HEVC,
 424     .pix_fmt        = AV_PIX_FMT_DXVA2_VLD,
 425     .start_frame    = dxva2_hevc_start_frame,
 426     .decode_slice   = dxva2_hevc_decode_slice,
 427     .end_frame      = dxva2_hevc_end_frame,
 428     .frame_priv_data_size = sizeof(struct hevc_dxva2_picture_context),
 429 };
 430 #endif
 431
 432 #if CONFIG_HEVC_D3D11VA_HWACCEL
 433 AVHWAccel ff_hevc_d3d11va_hwaccel = {
 434     .name           = "hevc_d3d11va",
 435     .type           = AVMEDIA_TYPE_VIDEO,
 436     .id             = AV_CODEC_ID_HEVC,
 437     .pix_fmt        = AV_PIX_FMT_D3D11VA_VLD,
 438     .start_frame    = dxva2_hevc_start_frame,
 439     .decode_slice   = dxva2_hevc_decode_slice,
 440     .end_frame      = dxva2_hevc_end_frame,
 441     .frame_priv_data_size = sizeof(struct hevc_dxva2_picture_context),
 442 };
 443 #endif