git.sesse.net Git - ffmpeg/blob - libavcodec/dxva2_h264.c

   1 /*
   2  * DXVA2 H264 HW acceleration.
   3  *
   4  * copyright (c) 2009 Laurent Aimar
   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 "dxva2_internal.h"
  24 #include "h264.h"
  25 #include "h264data.h"
  26 #include "mpegutils.h"
  27
  28 struct dxva2_picture_context {
  29     DXVA_PicParams_H264   pp;
  30     DXVA_Qmatrix_H264     qm;
  31     unsigned              slice_count;
  32     DXVA_Slice_H264_Short slice_short[MAX_SLICES];
  33     DXVA_Slice_H264_Long  slice_long[MAX_SLICES];
  34     const uint8_t         *bitstream;
  35     unsigned              bitstream_size;
  36 };
  37
  38 static void fill_picture_entry(DXVA_PicEntry_H264 *pic,
  39                                unsigned index, unsigned flag)
  40 {
  41     assert((index&0x7f) == index && (flag&0x01) == flag);
  42     pic->bPicEntry = index | (flag << 7);
  43 }
  44
  45 static void fill_picture_parameters(struct dxva_context *ctx, const H264Context *h,
  46                                     DXVA_PicParams_H264 *pp)
  47 {
  48     const H264Picture *current_picture = h->cur_pic_ptr;
  49     int i, j;
  50
  51     memset(pp, 0, sizeof(*pp));
  52     /* Configure current picture */
  53     fill_picture_entry(&pp->CurrPic,
  54                        ff_dxva2_get_surface_index(ctx, &current_picture->f),
  55                        h->picture_structure == PICT_BOTTOM_FIELD);
  56     /* Configure the set of references */
  57     pp->UsedForReferenceFlags  = 0;
  58     pp->NonExistingFrameFlags  = 0;
  59     for (i = 0, j = 0; i < FF_ARRAY_ELEMS(pp->RefFrameList); i++) {
  60         const H264Picture *r;
  61         if (j < h->short_ref_count) {
  62             r = h->short_ref[j++];
  63         } else {
  64             r = NULL;
  65             while (!r && j < h->short_ref_count + 16)
  66                 r = h->long_ref[j++ - h->short_ref_count];
  67         }
  68         if (r) {
  69             fill_picture_entry(&pp->RefFrameList[i],
  70                                ff_dxva2_get_surface_index(ctx, &r->f),
  71                                r->long_ref != 0);
  72
  73             if ((r->reference & PICT_TOP_FIELD) && r->field_poc[0] != INT_MAX)
  74                 pp->FieldOrderCntList[i][0] = r->field_poc[0];
  75             if ((r->reference & PICT_BOTTOM_FIELD) && r->field_poc[1] != INT_MAX)
  76                 pp->FieldOrderCntList[i][1] = r->field_poc[1];
  77
  78             pp->FrameNumList[i] = r->long_ref ? r->pic_id : r->frame_num;
  79             if (r->reference & PICT_TOP_FIELD)
  80                 pp->UsedForReferenceFlags |= 1 << (2*i + 0);
  81             if (r->reference & PICT_BOTTOM_FIELD)
  82                 pp->UsedForReferenceFlags |= 1 << (2*i + 1);
  83         } else {
  84             pp->RefFrameList[i].bPicEntry = 0xff;
  85             pp->FieldOrderCntList[i][0]   = 0;
  86             pp->FieldOrderCntList[i][1]   = 0;
  87             pp->FrameNumList[i]           = 0;
  88         }
  89     }
  90
  91     pp->wFrameWidthInMbsMinus1        = h->mb_width  - 1;
  92     pp->wFrameHeightInMbsMinus1       = h->mb_height - 1;
  93     pp->num_ref_frames                = h->sps.ref_frame_count;
  94
  95     pp->wBitFields                    = ((h->picture_structure != PICT_FRAME) <<  0) |
  96                                         ((h->sps.mb_aff &&
  97                                         (h->picture_structure == PICT_FRAME)) <<  1) |
  98                                         (h->sps.residual_color_transform_flag <<  2) |
  99                                         /* sp_for_switch_flag (not implemented by Libav) */
 100                                         (0                                    <<  3) |
 101                                         (h->sps.chroma_format_idc             <<  4) |
 102                                         ((h->nal_ref_idc != 0)                <<  6) |
 103                                         (h->pps.constrained_intra_pred        <<  7) |
 104                                         (h->pps.weighted_pred                 <<  8) |
 105                                         (h->pps.weighted_bipred_idc           <<  9) |
 106                                         /* MbsConsecutiveFlag */
 107                                         (1                                    << 11) |
 108                                         (h->sps.frame_mbs_only_flag           << 12) |
 109                                         (h->pps.transform_8x8_mode            << 13) |
 110                                         ((h->sps.level_idc >= 31)             << 14) |
 111                                         /* IntraPicFlag (Modified if we detect a non
 112                                          * intra slice in dxva2_h264_decode_slice) */
 113                                         (1                                    << 15);
 114
 115     pp->bit_depth_luma_minus8         = h->sps.bit_depth_luma - 8;
 116     pp->bit_depth_chroma_minus8       = h->sps.bit_depth_chroma - 8;
 117     if (ctx->workaround & FF_DXVA2_WORKAROUND_SCALING_LIST_ZIGZAG)
 118         pp->Reserved16Bits            = 0;
 119     else
 120         pp->Reserved16Bits            = 3; /* FIXME is there a way to detect the right mode ? */
 121     pp->StatusReportFeedbackNumber    = 1 + ctx->report_id++;
 122     pp->CurrFieldOrderCnt[0] = 0;
 123     if ((h->picture_structure & PICT_TOP_FIELD) &&
 124         current_picture->field_poc[0] != INT_MAX)
 125         pp->CurrFieldOrderCnt[0] = current_picture->field_poc[0];
 126     pp->CurrFieldOrderCnt[1] = 0;
 127     if ((h->picture_structure & PICT_BOTTOM_FIELD) &&
 128         current_picture->field_poc[1] != INT_MAX)
 129         pp->CurrFieldOrderCnt[1] = current_picture->field_poc[1];
 130     pp->pic_init_qs_minus26           = h->pps.init_qs - 26;
 131     pp->chroma_qp_index_offset        = h->pps.chroma_qp_index_offset[0];
 132     pp->second_chroma_qp_index_offset = h->pps.chroma_qp_index_offset[1];
 133     pp->ContinuationFlag              = 1;
 134     pp->pic_init_qp_minus26           = h->pps.init_qp - 26;
 135     pp->num_ref_idx_l0_active_minus1  = h->pps.ref_count[0] - 1;
 136     pp->num_ref_idx_l1_active_minus1  = h->pps.ref_count[1] - 1;
 137     pp->Reserved8BitsA                = 0;
 138     pp->frame_num                     = h->frame_num;
 139     pp->log2_max_frame_num_minus4     = h->sps.log2_max_frame_num - 4;
 140     pp->pic_order_cnt_type            = h->sps.poc_type;
 141     if (h->sps.poc_type == 0)
 142         pp->log2_max_pic_order_cnt_lsb_minus4 = h->sps.log2_max_poc_lsb - 4;
 143     else if (h->sps.poc_type == 1)
 144         pp->delta_pic_order_always_zero_flag = h->sps.delta_pic_order_always_zero_flag;
 145     pp->direct_8x8_inference_flag     = h->sps.direct_8x8_inference_flag;
 146     pp->entropy_coding_mode_flag      = h->pps.cabac;
 147     pp->pic_order_present_flag        = h->pps.pic_order_present;
 148     pp->num_slice_groups_minus1       = h->pps.slice_group_count - 1;
 149     pp->slice_group_map_type          = h->pps.mb_slice_group_map_type;
 150     pp->deblocking_filter_control_present_flag = h->pps.deblocking_filter_parameters_present;
 151     pp->redundant_pic_cnt_present_flag= h->pps.redundant_pic_cnt_present;
 152     pp->Reserved8BitsB                = 0;
 153     pp->slice_group_change_rate_minus1= 0;  /* XXX not implemented by Libav */
 154     //pp->SliceGroupMap[810];               /* XXX not implemented by Libav */
 155 }
 156
 157 static void fill_scaling_lists(struct dxva_context *ctx, const H264Context *h, DXVA_Qmatrix_H264 *qm)
 158 {
 159     unsigned i, j;
 160     memset(qm, 0, sizeof(*qm));
 161     if (ctx->workaround & FF_DXVA2_WORKAROUND_SCALING_LIST_ZIGZAG) {
 162         for (i = 0; i < 6; i++)
 163             for (j = 0; j < 16; j++)
 164                 qm->bScalingLists4x4[i][j] = h->pps.scaling_matrix4[i][j];
 165
 166         for (i = 0; i < 64; i++) {
 167             qm->bScalingLists8x8[0][i] = h->pps.scaling_matrix8[0][i];
 168             qm->bScalingLists8x8[1][i] = h->pps.scaling_matrix8[3][i];
 169         }
 170     } else {
 171         for (i = 0; i < 6; i++)
 172             for (j = 0; j < 16; j++)
 173                 qm->bScalingLists4x4[i][j] = h->pps.scaling_matrix4[i][zigzag_scan[j]];
 174
 175         for (i = 0; i < 64; i++) {
 176             qm->bScalingLists8x8[0][i] = h->pps.scaling_matrix8[0][ff_zigzag_direct[i]];
 177             qm->bScalingLists8x8[1][i] = h->pps.scaling_matrix8[3][ff_zigzag_direct[i]];
 178         }
 179     }
 180 }
 181
 182 static int is_slice_short(struct dxva_context *ctx)
 183 {
 184     assert(ctx->cfg->ConfigBitstreamRaw == 1 ||
 185            ctx->cfg->ConfigBitstreamRaw == 2);
 186     return ctx->cfg->ConfigBitstreamRaw == 2;
 187 }
 188
 189 static void fill_slice_short(DXVA_Slice_H264_Short *slice,
 190                              unsigned position, unsigned size)
 191 {
 192     memset(slice, 0, sizeof(*slice));
 193     slice->BSNALunitDataLocation = position;
 194     slice->SliceBytesInBuffer    = size;
 195     slice->wBadSliceChopping     = 0;
 196 }
 197
 198 static int get_refpic_index(const DXVA_PicParams_H264 *pp, int surface_index)
 199 {
 200     int i;
 201     for (i = 0; i < FF_ARRAY_ELEMS(pp->RefFrameList); i++) {
 202         if ((pp->RefFrameList[i].bPicEntry & 0x7f) == surface_index)
 203           return i;
 204     }
 205     return 0x7f;
 206 }
 207
 208 static void fill_slice_long(AVCodecContext *avctx, DXVA_Slice_H264_Long *slice,
 209                             const DXVA_PicParams_H264 *pp, unsigned position, unsigned size)
 210 {
 211     const H264Context *h = avctx->priv_data;
 212     struct dxva_context *ctx = avctx->hwaccel_context;
 213     unsigned list;
 214
 215     memset(slice, 0, sizeof(*slice));
 216     slice->BSNALunitDataLocation = position;
 217     slice->SliceBytesInBuffer    = size;
 218     slice->wBadSliceChopping     = 0;
 219
 220     slice->first_mb_in_slice     = (h->mb_y >> FIELD_OR_MBAFF_PICTURE(h)) * h->mb_width + h->mb_x;
 221     slice->NumMbsForSlice        = 0; /* XXX it is set once we have all slices */
 222     slice->BitOffsetToSliceData  = get_bits_count(&h->gb);
 223     slice->slice_type            = ff_h264_get_slice_type(h);
 224     if (h->slice_type_fixed)
 225         slice->slice_type += 5;
 226     slice->luma_log2_weight_denom       = h->luma_log2_weight_denom;
 227     slice->chroma_log2_weight_denom     = h->chroma_log2_weight_denom;
 228     if (h->list_count > 0)
 229         slice->num_ref_idx_l0_active_minus1 = h->ref_count[0] - 1;
 230     if (h->list_count > 1)
 231         slice->num_ref_idx_l1_active_minus1 = h->ref_count[1] - 1;
 232     slice->slice_alpha_c0_offset_div2   = h->slice_alpha_c0_offset / 2 - 26;
 233     slice->slice_beta_offset_div2       = h->slice_beta_offset     / 2 - 26;
 234     slice->Reserved8Bits                = 0;
 235
 236     for (list = 0; list < 2; list++) {
 237         unsigned i;
 238         for (i = 0; i < FF_ARRAY_ELEMS(slice->RefPicList[list]); i++) {
 239             if (list < h->list_count && i < h->ref_count[list]) {
 240                 const H264Picture *r = &h->ref_list[list][i];
 241                 unsigned plane;
 242                 unsigned index = get_refpic_index(pp, ff_dxva2_get_surface_index(ctx, &r->f));
 243                 fill_picture_entry(&slice->RefPicList[list][i], index,
 244                                    r->reference == PICT_BOTTOM_FIELD);
 245                 for (plane = 0; plane < 3; plane++) {
 246                     int w, o;
 247                     if (plane == 0 && h->luma_weight_flag[list]) {
 248                         w = h->luma_weight[i][list][0];
 249                         o = h->luma_weight[i][list][1];
 250                     } else if (plane >= 1 && h->chroma_weight_flag[list]) {
 251                         w = h->chroma_weight[i][list][plane-1][0];
 252                         o = h->chroma_weight[i][list][plane-1][1];
 253                     } else {
 254                         w = 1 << (plane == 0 ? h->luma_log2_weight_denom :
 255                                                h->chroma_log2_weight_denom);
 256                         o = 0;
 257                     }
 258                     slice->Weights[list][i][plane][0] = w;
 259                     slice->Weights[list][i][plane][1] = o;
 260                 }
 261             } else {
 262                 unsigned plane;
 263                 slice->RefPicList[list][i].bPicEntry = 0xff;
 264                 for (plane = 0; plane < 3; plane++) {
 265                     slice->Weights[list][i][plane][0] = 0;
 266                     slice->Weights[list][i][plane][1] = 0;
 267                 }
 268             }
 269         }
 270     }
 271     slice->slice_qs_delta    = 0; /* XXX not implemented by Libav */
 272     slice->slice_qp_delta    = h->qscale - h->pps.init_qp;
 273     slice->redundant_pic_cnt = h->redundant_pic_count;
 274     if (h->slice_type == AV_PICTURE_TYPE_B)
 275         slice->direct_spatial_mv_pred_flag = h->direct_spatial_mv_pred;
 276     slice->cabac_init_idc = h->pps.cabac ? h->cabac_init_idc : 0;
 277     if (h->deblocking_filter < 2)
 278         slice->disable_deblocking_filter_idc = 1 - h->deblocking_filter;
 279     else
 280         slice->disable_deblocking_filter_idc = h->deblocking_filter;
 281     slice->slice_id = h->current_slice - 1;
 282 }
 283
 284 static int commit_bitstream_and_slice_buffer(AVCodecContext *avctx,
 285                                              DXVA2_DecodeBufferDesc *bs,
 286                                              DXVA2_DecodeBufferDesc *sc)
 287 {
 288     const H264Context *h = avctx->priv_data;
 289     const unsigned mb_count = h->mb_width * h->mb_height;
 290     struct dxva_context *ctx = avctx->hwaccel_context;
 291     const H264Picture *current_picture = h->cur_pic_ptr;
 292     struct dxva2_picture_context *ctx_pic = current_picture->hwaccel_picture_private;
 293     DXVA_Slice_H264_Short *slice = NULL;
 294     uint8_t  *dxva_data, *current, *end;
 295     unsigned dxva_size;
 296     void     *slice_data;
 297     unsigned slice_size;
 298     unsigned padding;
 299     unsigned i;
 300
 301     /* Create an annex B bitstream buffer with only slice NAL and finalize slice */
 302     if (FAILED(IDirectXVideoDecoder_GetBuffer(ctx->decoder,
 303                                                DXVA2_BitStreamDateBufferType,
 304                                                &dxva_data, &dxva_size)))
 305         return -1;
 306     current = dxva_data;
 307     end = dxva_data + dxva_size;
 308
 309     for (i = 0; i < ctx_pic->slice_count; i++) {
 310         static const uint8_t start_code[] = { 0, 0, 1 };
 311         static const unsigned start_code_size = sizeof(start_code);
 312         unsigned position, size;
 313
 314         assert(offsetof(DXVA_Slice_H264_Short, BSNALunitDataLocation) ==
 315                offsetof(DXVA_Slice_H264_Long,  BSNALunitDataLocation));
 316         assert(offsetof(DXVA_Slice_H264_Short, SliceBytesInBuffer) ==
 317                offsetof(DXVA_Slice_H264_Long,  SliceBytesInBuffer));
 318
 319         if (is_slice_short(ctx))
 320             slice = &ctx_pic->slice_short[i];
 321         else
 322             slice = (DXVA_Slice_H264_Short*)&ctx_pic->slice_long[i];
 323
 324         position = slice->BSNALunitDataLocation;
 325         size     = slice->SliceBytesInBuffer;
 326         if (start_code_size + size > end - current) {
 327             av_log(avctx, AV_LOG_ERROR, "Failed to build bitstream");
 328             break;
 329         }
 330
 331         slice->BSNALunitDataLocation = current - dxva_data;
 332         slice->SliceBytesInBuffer    = start_code_size + size;
 333
 334         if (!is_slice_short(ctx)) {
 335             DXVA_Slice_H264_Long *slice_long = (DXVA_Slice_H264_Long*)slice;
 336             if (i < ctx_pic->slice_count - 1)
 337                 slice_long->NumMbsForSlice =
 338                     slice_long[1].first_mb_in_slice - slice_long[0].first_mb_in_slice;
 339             else
 340                 slice_long->NumMbsForSlice = mb_count - slice_long->first_mb_in_slice;
 341         }
 342
 343         memcpy(current, start_code, start_code_size);
 344         current += start_code_size;
 345
 346         memcpy(current, &ctx_pic->bitstream[position], size);
 347         current += size;
 348     }
 349     padding = FFMIN(128 - ((current - dxva_data) & 127), end - current);
 350     if (slice && padding > 0) {
 351         memset(current, 0, padding);
 352         current += padding;
 353
 354         slice->SliceBytesInBuffer += padding;
 355     }
 356     if (FAILED(IDirectXVideoDecoder_ReleaseBuffer(ctx->decoder,
 357                                                   DXVA2_BitStreamDateBufferType)))
 358         return -1;
 359     if (i < ctx_pic->slice_count)
 360         return -1;
 361
 362     memset(bs, 0, sizeof(*bs));
 363     bs->CompressedBufferType = DXVA2_BitStreamDateBufferType;
 364     bs->DataSize             = current - dxva_data;
 365     bs->NumMBsInBuffer       = mb_count;
 366
 367     if (is_slice_short(ctx)) {
 368         slice_data = ctx_pic->slice_short;
 369         slice_size = ctx_pic->slice_count * sizeof(*ctx_pic->slice_short);
 370     } else {
 371         slice_data = ctx_pic->slice_long;
 372         slice_size = ctx_pic->slice_count * sizeof(*ctx_pic->slice_long);
 373     }
 374     assert((bs->DataSize & 127) == 0);
 375     return ff_dxva2_commit_buffer(avctx, ctx, sc,
 376                                   DXVA2_SliceControlBufferType,
 377                                   slice_data, slice_size, mb_count);
 378 }
 379
 380
 381 static int dxva2_h264_start_frame(AVCodecContext *avctx,
 382                                   av_unused const uint8_t *buffer,
 383                                   av_unused uint32_t size)
 384 {
 385     const H264Context *h = avctx->priv_data;
 386     struct dxva_context *ctx = avctx->hwaccel_context;
 387     struct dxva2_picture_context *ctx_pic = h->cur_pic_ptr->hwaccel_picture_private;
 388
 389     if (!ctx->decoder || !ctx->cfg || ctx->surface_count <= 0)
 390         return -1;
 391     assert(ctx_pic);
 392
 393     /* Fill up DXVA_PicParams_H264 */
 394     fill_picture_parameters(ctx, h, &ctx_pic->pp);
 395
 396     /* Fill up DXVA_Qmatrix_H264 */
 397     fill_scaling_lists(ctx, h, &ctx_pic->qm);
 398
 399     ctx_pic->slice_count    = 0;
 400     ctx_pic->bitstream_size = 0;
 401     ctx_pic->bitstream      = NULL;
 402     return 0;
 403 }
 404
 405 static int dxva2_h264_decode_slice(AVCodecContext *avctx,
 406                                    const uint8_t *buffer,
 407                                    uint32_t size)
 408 {
 409     const H264Context *h = avctx->priv_data;
 410     struct dxva_context *ctx = avctx->hwaccel_context;
 411     const H264Picture *current_picture = h->cur_pic_ptr;
 412     struct dxva2_picture_context *ctx_pic = current_picture->hwaccel_picture_private;
 413     unsigned position;
 414
 415     if (ctx_pic->slice_count >= MAX_SLICES)
 416         return -1;
 417
 418     if (!ctx_pic->bitstream)
 419         ctx_pic->bitstream = buffer;
 420     ctx_pic->bitstream_size += size;
 421
 422     position = buffer - ctx_pic->bitstream;
 423     if (is_slice_short(ctx))
 424         fill_slice_short(&ctx_pic->slice_short[ctx_pic->slice_count],
 425                          position, size);
 426     else
 427         fill_slice_long(avctx, &ctx_pic->slice_long[ctx_pic->slice_count],
 428                         &ctx_pic->pp, position, size);
 429     ctx_pic->slice_count++;
 430
 431     if (h->slice_type != AV_PICTURE_TYPE_I && h->slice_type != AV_PICTURE_TYPE_SI)
 432         ctx_pic->pp.wBitFields &= ~(1 << 15); /* Set IntraPicFlag to 0 */
 433     return 0;
 434 }
 435
 436 static int dxva2_h264_end_frame(AVCodecContext *avctx)
 437 {
 438     H264Context *h = avctx->priv_data;
 439     struct dxva2_picture_context *ctx_pic =
 440         h->cur_pic_ptr->hwaccel_picture_private;
 441     int ret;
 442
 443     if (ctx_pic->slice_count <= 0 || ctx_pic->bitstream_size <= 0)
 444         return -1;
 445     ret = ff_dxva2_common_end_frame(avctx, &h->cur_pic_ptr->f,
 446                                     &ctx_pic->pp, sizeof(ctx_pic->pp),
 447                                     &ctx_pic->qm, sizeof(ctx_pic->qm),
 448                                     commit_bitstream_and_slice_buffer);
 449     if (!ret)
 450         ff_h264_draw_horiz_band(h, 0, h->avctx->height);
 451     return ret;
 452 }
 453
 454 AVHWAccel ff_h264_dxva2_hwaccel = {
 455     .name           = "h264_dxva2",
 456     .type           = AVMEDIA_TYPE_VIDEO,
 457     .id             = AV_CODEC_ID_H264,
 458     .pix_fmt        = AV_PIX_FMT_DXVA2_VLD,
 459     .start_frame    = dxva2_h264_start_frame,
 460     .decode_slice   = dxva2_h264_decode_slice,
 461     .end_frame      = dxva2_h264_end_frame,
 462     .priv_data_size = sizeof(struct dxva2_picture_context),
 463 };