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
  27 struct dxva2_picture_context {
  28     DXVA_PicParams_H264   pp;
  29     DXVA_Qmatrix_H264     qm;
  30     unsigned              slice_count;
  31     DXVA_Slice_H264_Short slice_short[MAX_SLICES];
  32     DXVA_Slice_H264_Long  slice_long[MAX_SLICES];
  33     const uint8_t         *bitstream;
  34     unsigned              bitstream_size;
  35 };
  36
  37 static void fill_picture_entry(DXVA_PicEntry_H264 *pic,
  38                                unsigned index, unsigned flag)
  39 {
  40     assert((index&0x7f) == index && (flag&0x01) == flag);
  41     pic->bPicEntry = index | (flag << 7);
  42 }
  43
  44 static void fill_picture_parameters(struct dxva_context *ctx, const H264Context *h,
  45                                     DXVA_PicParams_H264 *pp)
  46 {
  47     const MpegEncContext *s = &h->s;
  48     const Picture *current_picture = s->current_picture_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),
  55                        s->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 Picture *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),
  71                                r->long_ref != 0);
  72
  73             if ((r->f.reference & PICT_TOP_FIELD) && r->field_poc[0] != INT_MAX)
  74                 pp->FieldOrderCntList[i][0] = r->field_poc[0];
  75             if ((r->f.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->f.reference & PICT_TOP_FIELD)
  80                 pp->UsedForReferenceFlags |= 1 << (2*i + 0);
  81             if (r->f.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        = s->mb_width  - 1;
  92     pp->wFrameHeightInMbsMinus1       = s->mb_height - 1;
  93     pp->num_ref_frames                = h->sps.ref_frame_count;
  94
  95     pp->wBitFields                    = ((s->picture_structure != PICT_FRAME) <<  0) |
  96                                         ((h->sps.mb_aff &&
  97                                         (s->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 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 ((s->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 ((s->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 void fill_slice_long(AVCodecContext *avctx, DXVA_Slice_H264_Long *slice,
 199                             unsigned position, unsigned size)
 200 {
 201     const H264Context *h = avctx->priv_data;
 202     struct dxva_context *ctx = avctx->hwaccel_context;
 203     const MpegEncContext *s = &h->s;
 204     unsigned list;
 205
 206     memset(slice, 0, sizeof(*slice));
 207     slice->BSNALunitDataLocation = position;
 208     slice->SliceBytesInBuffer    = size;
 209     slice->wBadSliceChopping     = 0;
 210
 211     slice->first_mb_in_slice     = (s->mb_y >> FIELD_OR_MBAFF_PICTURE) * s->mb_width + s->mb_x;
 212     slice->NumMbsForSlice        = 0; /* XXX it is set once we have all slices */
 213     slice->BitOffsetToSliceData  = get_bits_count(&s->gb);
 214     slice->slice_type            = ff_h264_get_slice_type(h);
 215     if (h->slice_type_fixed)
 216         slice->slice_type += 5;
 217     slice->luma_log2_weight_denom       = h->luma_log2_weight_denom;
 218     slice->chroma_log2_weight_denom     = h->chroma_log2_weight_denom;
 219     if (h->list_count > 0)
 220         slice->num_ref_idx_l0_active_minus1 = h->ref_count[0] - 1;
 221     if (h->list_count > 1)
 222         slice->num_ref_idx_l1_active_minus1 = h->ref_count[1] - 1;
 223     slice->slice_alpha_c0_offset_div2   = h->slice_alpha_c0_offset / 2 - 26;
 224     slice->slice_beta_offset_div2       = h->slice_beta_offset     / 2 - 26;
 225     slice->Reserved8Bits                = 0;
 226
 227     for (list = 0; list < 2; list++) {
 228         unsigned i;
 229         for (i = 0; i < FF_ARRAY_ELEMS(slice->RefPicList[list]); i++) {
 230             if (list < h->list_count && i < h->ref_count[list]) {
 231                 const Picture *r = &h->ref_list[list][i];
 232                 unsigned plane;
 233                 fill_picture_entry(&slice->RefPicList[list][i],
 234                                    ff_dxva2_get_surface_index(ctx, r),
 235                                    r->f.reference == PICT_BOTTOM_FIELD);
 236                 for (plane = 0; plane < 3; plane++) {
 237                     int w, o;
 238                     if (plane == 0 && h->luma_weight_flag[list]) {
 239                         w = h->luma_weight[i][list][0];
 240                         o = h->luma_weight[i][list][1];
 241                     } else if (plane >= 1 && h->chroma_weight_flag[list]) {
 242                         w = h->chroma_weight[i][list][plane-1][0];
 243                         o = h->chroma_weight[i][list][plane-1][1];
 244                     } else {
 245                         w = 1 << (plane == 0 ? h->luma_log2_weight_denom :
 246                                                h->chroma_log2_weight_denom);
 247                         o = 0;
 248                     }
 249                     slice->Weights[list][i][plane][0] = w;
 250                     slice->Weights[list][i][plane][1] = o;
 251                 }
 252             } else {
 253                 unsigned plane;
 254                 slice->RefPicList[list][i].bPicEntry = 0xff;
 255                 for (plane = 0; plane < 3; plane++) {
 256                     slice->Weights[list][i][plane][0] = 0;
 257                     slice->Weights[list][i][plane][1] = 0;
 258                 }
 259             }
 260         }
 261     }
 262     slice->slice_qs_delta    = 0; /* XXX not implemented by Libav */
 263     slice->slice_qp_delta    = s->qscale - h->pps.init_qp;
 264     slice->redundant_pic_cnt = h->redundant_pic_count;
 265     if (h->slice_type == AV_PICTURE_TYPE_B)
 266         slice->direct_spatial_mv_pred_flag = h->direct_spatial_mv_pred;
 267     slice->cabac_init_idc = h->pps.cabac ? h->cabac_init_idc : 0;
 268     if (h->deblocking_filter < 2)
 269         slice->disable_deblocking_filter_idc = 1 - h->deblocking_filter;
 270     else
 271         slice->disable_deblocking_filter_idc = h->deblocking_filter;
 272     slice->slice_id = h->current_slice - 1;
 273 }
 274
 275 static int commit_bitstream_and_slice_buffer(AVCodecContext *avctx,
 276                                              DXVA2_DecodeBufferDesc *bs,
 277                                              DXVA2_DecodeBufferDesc *sc)
 278 {
 279     const H264Context *h = avctx->priv_data;
 280     const MpegEncContext *s = &h->s;
 281     const unsigned mb_count = s->mb_width * s->mb_height;
 282     struct dxva_context *ctx = avctx->hwaccel_context;
 283     const Picture *current_picture = h->s.current_picture_ptr;
 284     struct dxva2_picture_context *ctx_pic = current_picture->f.hwaccel_picture_private;
 285     DXVA_Slice_H264_Short *slice = NULL;
 286     uint8_t  *dxva_data, *current, *end;
 287     unsigned dxva_size;
 288     void     *slice_data;
 289     unsigned slice_size;
 290     unsigned padding;
 291     unsigned i;
 292
 293     /* Create an annex B bitstream buffer with only slice NAL and finalize slice */
 294     if (FAILED(IDirectXVideoDecoder_GetBuffer(ctx->decoder,
 295                                                DXVA2_BitStreamDateBufferType,
 296                                                &dxva_data, &dxva_size)))
 297         return -1;
 298     current = dxva_data;
 299     end = dxva_data + dxva_size;
 300
 301     for (i = 0; i < ctx_pic->slice_count; i++) {
 302         static const uint8_t start_code[] = { 0, 0, 1 };
 303         static const unsigned start_code_size = sizeof(start_code);
 304         unsigned position, size;
 305
 306         assert(offsetof(DXVA_Slice_H264_Short, BSNALunitDataLocation) ==
 307                offsetof(DXVA_Slice_H264_Long,  BSNALunitDataLocation));
 308         assert(offsetof(DXVA_Slice_H264_Short, SliceBytesInBuffer) ==
 309                offsetof(DXVA_Slice_H264_Long,  SliceBytesInBuffer));
 310
 311         if (is_slice_short(ctx))
 312             slice = &ctx_pic->slice_short[i];
 313         else
 314             slice = (DXVA_Slice_H264_Short*)&ctx_pic->slice_long[i];
 315
 316         position = slice->BSNALunitDataLocation;
 317         size     = slice->SliceBytesInBuffer;
 318         if (start_code_size + size > end - current) {
 319             av_log(avctx, AV_LOG_ERROR, "Failed to build bitstream");
 320             break;
 321         }
 322
 323         slice->BSNALunitDataLocation = current - dxva_data;
 324         slice->SliceBytesInBuffer    = start_code_size + size;
 325
 326         if (!is_slice_short(ctx)) {
 327             DXVA_Slice_H264_Long *slice_long = (DXVA_Slice_H264_Long*)slice;
 328             if (i < ctx_pic->slice_count - 1)
 329                 slice_long->NumMbsForSlice =
 330                     slice_long[1].first_mb_in_slice - slice_long[0].first_mb_in_slice;
 331             else
 332                 slice_long->NumMbsForSlice = mb_count - slice_long->first_mb_in_slice;
 333         }
 334
 335         memcpy(current, start_code, start_code_size);
 336         current += start_code_size;
 337
 338         memcpy(current, &ctx_pic->bitstream[position], size);
 339         current += size;
 340     }
 341     padding = FFMIN(128 - ((current - dxva_data) & 127), end - current);
 342     if (slice && padding > 0) {
 343         memset(current, 0, padding);
 344         current += padding;
 345
 346         slice->SliceBytesInBuffer += padding;
 347     }
 348     if (FAILED(IDirectXVideoDecoder_ReleaseBuffer(ctx->decoder,
 349                                                   DXVA2_BitStreamDateBufferType)))
 350         return -1;
 351     if (i < ctx_pic->slice_count)
 352         return -1;
 353
 354     memset(bs, 0, sizeof(*bs));
 355     bs->CompressedBufferType = DXVA2_BitStreamDateBufferType;
 356     bs->DataSize             = current - dxva_data;
 357     bs->NumMBsInBuffer       = mb_count;
 358
 359     if (is_slice_short(ctx)) {
 360         slice_data = ctx_pic->slice_short;
 361         slice_size = ctx_pic->slice_count * sizeof(*ctx_pic->slice_short);
 362     } else {
 363         slice_data = ctx_pic->slice_long;
 364         slice_size = ctx_pic->slice_count * sizeof(*ctx_pic->slice_long);
 365     }
 366     assert((bs->DataSize & 127) == 0);
 367     return ff_dxva2_commit_buffer(avctx, ctx, sc,
 368                                   DXVA2_SliceControlBufferType,
 369                                   slice_data, slice_size, mb_count);
 370 }
 371
 372
 373 static int start_frame(AVCodecContext *avctx,
 374                        av_unused const uint8_t *buffer,
 375                        av_unused uint32_t size)
 376 {
 377     const H264Context *h = avctx->priv_data;
 378     struct dxva_context *ctx = avctx->hwaccel_context;
 379     struct dxva2_picture_context *ctx_pic = h->s.current_picture_ptr->f.hwaccel_picture_private;
 380
 381     if (!ctx->decoder || !ctx->cfg || ctx->surface_count <= 0)
 382         return -1;
 383     assert(ctx_pic);
 384
 385     /* Fill up DXVA_PicParams_H264 */
 386     fill_picture_parameters(ctx, h, &ctx_pic->pp);
 387
 388     /* Fill up DXVA_Qmatrix_H264 */
 389     fill_scaling_lists(ctx, h, &ctx_pic->qm);
 390
 391     ctx_pic->slice_count    = 0;
 392     ctx_pic->bitstream_size = 0;
 393     ctx_pic->bitstream      = NULL;
 394     return 0;
 395 }
 396
 397 static int decode_slice(AVCodecContext *avctx,
 398                         const uint8_t *buffer, uint32_t size)
 399 {
 400     const H264Context *h = avctx->priv_data;
 401     struct dxva_context *ctx = avctx->hwaccel_context;
 402     const Picture *current_picture = h->s.current_picture_ptr;
 403     struct dxva2_picture_context *ctx_pic = current_picture->f.hwaccel_picture_private;
 404     unsigned position;
 405
 406     if (ctx_pic->slice_count >= MAX_SLICES)
 407         return -1;
 408
 409     if (!ctx_pic->bitstream)
 410         ctx_pic->bitstream = buffer;
 411     ctx_pic->bitstream_size += size;
 412
 413     position = buffer - ctx_pic->bitstream;
 414     if (is_slice_short(ctx))
 415         fill_slice_short(&ctx_pic->slice_short[ctx_pic->slice_count],
 416                          position, size);
 417     else
 418         fill_slice_long(avctx, &ctx_pic->slice_long[ctx_pic->slice_count],
 419                         position, size);
 420     ctx_pic->slice_count++;
 421
 422     if (h->slice_type != AV_PICTURE_TYPE_I && h->slice_type != AV_PICTURE_TYPE_SI)
 423         ctx_pic->pp.wBitFields &= ~(1 << 15); /* Set IntraPicFlag to 0 */
 424     return 0;
 425 }
 426
 427 static int end_frame(AVCodecContext *avctx)
 428 {
 429     H264Context *h = avctx->priv_data;
 430     MpegEncContext *s = &h->s;
 431     struct dxva2_picture_context *ctx_pic =
 432         h->s.current_picture_ptr->f.hwaccel_picture_private;
 433
 434     if (ctx_pic->slice_count <= 0 || ctx_pic->bitstream_size <= 0)
 435         return -1;
 436     return ff_dxva2_common_end_frame(avctx, s,
 437                                      &ctx_pic->pp, sizeof(ctx_pic->pp),
 438                                      &ctx_pic->qm, sizeof(ctx_pic->qm),
 439                                      commit_bitstream_and_slice_buffer);
 440 }
 441
 442 AVHWAccel ff_h264_dxva2_hwaccel = {
 443     .name           = "h264_dxva2",
 444     .type           = AVMEDIA_TYPE_VIDEO,
 445     .id             = AV_CODEC_ID_H264,
 446     .pix_fmt        = AV_PIX_FMT_DXVA2_VLD,
 447     .start_frame    = start_frame,
 448     .decode_slice   = decode_slice,
 449     .end_frame      = end_frame,
 450     .priv_data_size = sizeof(struct dxva2_picture_context),
 451 };