git.sesse.net Git - ffmpeg/blob - libavcodec/dxva2_vc1.c

   1 /*
   2  * DXVA2 WMV3/VC-1 HW acceleration.
   3  *
   4  * copyright (c) 2010 Laurent Aimar
   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 "dxva2_internal.h"
  24 #include "vc1.h"
  25 #include "vc1data.h"
  26
  27 struct dxva2_picture_context {
  28     DXVA_PictureParameters pp;
  29     DXVA_SliceInfo         si;
  30
  31     const uint8_t          *bitstream;
  32     unsigned               bitstream_size;
  33 };
  34
  35 static void fill_picture_parameters(AVCodecContext *avctx,
  36                                     struct dxva_context *ctx, const VC1Context *v,
  37                                     DXVA_PictureParameters *pp)
  38 {
  39     const MpegEncContext *s = &v->s;
  40     const Picture *current_picture = s->current_picture_ptr;
  41     int intcomp = 0;
  42
  43     // determine if intensity compensation is needed
  44     if (s->pict_type == AV_PICTURE_TYPE_P) {
  45       if ((v->fcm == ILACE_FRAME && v->intcomp) || (v->fcm != ILACE_FRAME && v->mv_mode == MV_PMODE_INTENSITY_COMP)) {
  46         if (v->lumscale != 32 || v->lumshift != 0 || (s->picture_structure != PICT_FRAME && (v->lumscale2 != 32 && v->lumshift2 != 0)))
  47           intcomp = 1;
  48       }
  49     }
  50
  51     memset(pp, 0, sizeof(*pp));
  52     pp->wDecodedPictureIndex    =
  53     pp->wDeblockedPictureIndex  = ff_dxva2_get_surface_index(ctx, current_picture);
  54     if (s->pict_type != AV_PICTURE_TYPE_I && !v->bi_type)
  55         pp->wForwardRefPictureIndex = ff_dxva2_get_surface_index(ctx, &s->last_picture);
  56     else
  57         pp->wForwardRefPictureIndex = 0xffff;
  58     if (s->pict_type == AV_PICTURE_TYPE_B && !v->bi_type)
  59         pp->wBackwardRefPictureIndex = ff_dxva2_get_surface_index(ctx, &s->next_picture);
  60     else
  61         pp->wBackwardRefPictureIndex = 0xffff;
  62     if (v->profile == PROFILE_ADVANCED) {
  63         /* It is the cropped width/height -1 of the frame */
  64         pp->wPicWidthInMBminus1 = avctx->width  - 1;
  65         pp->wPicHeightInMBminus1= avctx->height - 1;
  66     } else {
  67         /* It is the coded width/height in macroblock -1 of the frame */
  68         pp->wPicWidthInMBminus1 = s->mb_width  - 1;
  69         pp->wPicHeightInMBminus1= s->mb_height - 1;
  70     }
  71     pp->bMacroblockWidthMinus1  = 15;
  72     pp->bMacroblockHeightMinus1 = 15;
  73     pp->bBlockWidthMinus1       = 7;
  74     pp->bBlockHeightMinus1      = 7;
  75     pp->bBPPminus1              = 7;
  76     if (s->picture_structure & PICT_TOP_FIELD)
  77         pp->bPicStructure      |= 0x01;
  78     if (s->picture_structure & PICT_BOTTOM_FIELD)
  79         pp->bPicStructure      |= 0x02;
  80     pp->bSecondField            = v->interlace && v->fcm == ILACE_FIELD && v->second_field;
  81     pp->bPicIntra               = s->pict_type == AV_PICTURE_TYPE_I || v->bi_type;
  82     pp->bPicBackwardPrediction  = s->pict_type == AV_PICTURE_TYPE_B && !v->bi_type;
  83     pp->bBidirectionalAveragingMode = (1                                           << 7) |
  84                                       ((ctx->cfg->ConfigIntraResidUnsigned != 0)   << 6) |
  85                                       ((ctx->cfg->ConfigResidDiffAccelerator != 0) << 5) |
  86                                       (intcomp                                     << 4) |
  87                                       ((v->profile == PROFILE_ADVANCED)            << 3);
  88     pp->bMVprecisionAndChromaRelation = ((v->mv_mode == MV_PMODE_1MV_HPEL_BILIN) << 3) |
  89                                         (1                                       << 2) |
  90                                         (0                                       << 1) |
  91                                         (!s->quarter_sample                          );
  92     pp->bChromaFormat           = v->chromaformat;
  93     ctx->report_id++;
  94     if (ctx->report_id >= (1 << 16))
  95         ctx->report_id = 1;
  96     pp->bPicScanFixed           = ctx->report_id >> 8;
  97     pp->bPicScanMethod          = ctx->report_id & 0xff;
  98     pp->bPicReadbackRequests    = 0;
  99     pp->bRcontrol               = v->rnd;
 100     pp->bPicSpatialResid8       = (v->panscanflag  << 7) |
 101                                   (v->refdist_flag << 6) |
 102                                   (s->loop_filter  << 5) |
 103                                   (v->fastuvmc     << 4) |
 104                                   (v->extended_mv  << 3) |
 105                                   (v->dquant       << 1) |
 106                                   (v->vstransform      );
 107     pp->bPicOverflowBlocks      = (v->quantizer_mode << 6) |
 108                                   (v->multires       << 5) |
 109                                   (v->resync_marker  << 4) |
 110                                   (v->rangered       << 3) |
 111                                   (s->max_b_frames       );
 112     pp->bPicExtrapolation       = (!v->interlace || v->fcm == PROGRESSIVE) ? 1 : 2;
 113     pp->bPicDeblocked           = ((!pp->bPicBackwardPrediction && v->overlap)        << 6) |
 114                                   ((v->profile != PROFILE_ADVANCED && v->rangeredfrm) << 5) |
 115                                   (s->loop_filter                                     << 1);
 116     pp->bPicDeblockConfined     = (v->postprocflag             << 7) |
 117                                   (v->broadcast                << 6) |
 118                                   (v->interlace                << 5) |
 119                                   (v->tfcntrflag               << 4) |
 120                                   (v->finterpflag              << 3) |
 121                                   ((s->pict_type != AV_PICTURE_TYPE_B) << 2) |
 122                                   (v->psf                      << 1) |
 123                                   (v->extended_dmv                 );
 124     if (s->pict_type != AV_PICTURE_TYPE_I)
 125         pp->bPic4MVallowed      = v->mv_mode == MV_PMODE_MIXED_MV ||
 126                                   (v->mv_mode == MV_PMODE_INTENSITY_COMP &&
 127                                    v->mv_mode2 == MV_PMODE_MIXED_MV);
 128     if (v->profile == PROFILE_ADVANCED)
 129         pp->bPicOBMC            = (v->range_mapy_flag  << 7) |
 130                                   (v->range_mapy       << 4) |
 131                                   (v->range_mapuv_flag << 3) |
 132                                   (v->range_mapuv          );
 133     pp->bPicBinPB               = 0;
 134     pp->bMV_RPS                 = (v->fcm == ILACE_FIELD && pp->bPicBackwardPrediction) ? v->refdist + 9 : 0;
 135     pp->bReservedBits           = v->pq;
 136     if (s->picture_structure == PICT_FRAME) {
 137         if (intcomp) {
 138             pp->wBitstreamFcodes      = v->lumscale;
 139             pp->wBitstreamPCEelements = v->lumshift;
 140         } else {
 141             pp->wBitstreamFcodes      = 32;
 142             pp->wBitstreamPCEelements = 0;
 143         }
 144     } else {
 145         /* Syntax: (top_field_param << 8) | bottom_field_param */
 146         if (intcomp) {
 147             pp->wBitstreamFcodes      = (v->lumscale << 8) | v->lumscale2;
 148             pp->wBitstreamPCEelements = (v->lumshift << 8) | v->lumshift2;
 149         } else {
 150             pp->wBitstreamFcodes      = (32 << 8) | 32;
 151             pp->wBitstreamPCEelements = 0;
 152         }
 153     }
 154     pp->bBitstreamConcealmentNeed   = 0;
 155     pp->bBitstreamConcealmentMethod = 0;
 156 }
 157
 158 static void fill_slice(AVCodecContext *avctx, DXVA_SliceInfo *slice,
 159                        unsigned position, unsigned size)
 160 {
 161     const VC1Context *v = avctx->priv_data;
 162     const MpegEncContext *s = &v->s;
 163
 164     memset(slice, 0, sizeof(*slice));
 165     slice->wHorizontalPosition = 0;
 166     slice->wVerticalPosition   = s->mb_y;
 167     slice->dwSliceBitsInBuffer = 8 * size;
 168     slice->dwSliceDataLocation = position;
 169     slice->bStartCodeBitOffset = 0;
 170     slice->bReservedBits       = (s->pict_type == AV_PICTURE_TYPE_B && !v->bi_type) ? v->bfraction_lut_index + 9 : 0;
 171     slice->wMBbitOffset        = v->p_frame_skipped ? 0xffff : get_bits_count(&s->gb);
 172     slice->wNumberMBsInSlice   = s->mb_width * s->mb_height; /* XXX We assume 1 slice */
 173     slice->wQuantizerScaleCode = v->pq;
 174     slice->wBadSliceChopping   = 0;
 175 }
 176
 177 static int commit_bitstream_and_slice_buffer(AVCodecContext *avctx,
 178                                              DXVA2_DecodeBufferDesc *bs,
 179                                              DXVA2_DecodeBufferDesc *sc)
 180 {
 181     const VC1Context *v = avctx->priv_data;
 182     struct dxva_context *ctx = avctx->hwaccel_context;
 183     const MpegEncContext *s = &v->s;
 184     struct dxva2_picture_context *ctx_pic = s->current_picture_ptr->hwaccel_picture_private;
 185
 186     DXVA_SliceInfo *slice = &ctx_pic->si;
 187
 188     static const uint8_t start_code[] = { 0, 0, 1, 0x0d };
 189     const unsigned start_code_size = avctx->codec_id == AV_CODEC_ID_VC1 ? sizeof(start_code) : 0;
 190     const unsigned slice_size = slice->dwSliceBitsInBuffer / 8;
 191     const unsigned padding = 128 - ((start_code_size + slice_size) & 127);
 192     const unsigned data_size = start_code_size + slice_size + padding;
 193
 194     uint8_t  *dxva_data;
 195     unsigned dxva_size;
 196     int result;
 197
 198     if (FAILED(IDirectXVideoDecoder_GetBuffer(ctx->decoder,
 199                                               DXVA2_BitStreamDateBufferType,
 200                                               (void **)&dxva_data, &dxva_size)))
 201         return -1;
 202
 203     result = data_size <= dxva_size ? 0 : -1;
 204     if (!result) {
 205         if (start_code_size > 0) {
 206             memcpy(dxva_data, start_code, start_code_size);
 207             if (v->second_field)
 208                 dxva_data[3] = 0x0c;
 209         }
 210         memcpy(dxva_data + start_code_size,
 211                ctx_pic->bitstream + slice->dwSliceDataLocation, slice_size);
 212         if (padding > 0)
 213             memset(dxva_data + start_code_size + slice_size, 0, padding);
 214         slice->dwSliceBitsInBuffer = 8 * data_size;
 215     }
 216     if (FAILED(IDirectXVideoDecoder_ReleaseBuffer(ctx->decoder,
 217                                                   DXVA2_BitStreamDateBufferType)))
 218         return -1;
 219     if (result)
 220         return result;
 221
 222     memset(bs, 0, sizeof(*bs));
 223     bs->CompressedBufferType = DXVA2_BitStreamDateBufferType;
 224     bs->DataSize             = data_size;
 225     bs->NumMBsInBuffer       = s->mb_width * s->mb_height;
 226     assert((bs->DataSize & 127) == 0);
 227
 228     return ff_dxva2_commit_buffer(avctx, ctx, sc,
 229                                   DXVA2_SliceControlBufferType,
 230                                   slice, sizeof(*slice), bs->NumMBsInBuffer);
 231 }
 232
 233 static int dxva2_vc1_start_frame(AVCodecContext *avctx,
 234                                  av_unused const uint8_t *buffer,
 235                                  av_unused uint32_t size)
 236 {
 237     const VC1Context *v = avctx->priv_data;
 238     struct dxva_context *ctx = avctx->hwaccel_context;
 239     struct dxva2_picture_context *ctx_pic = v->s.current_picture_ptr->hwaccel_picture_private;
 240
 241     if (!ctx->decoder || !ctx->cfg || ctx->surface_count <= 0)
 242         return -1;
 243     assert(ctx_pic);
 244
 245     fill_picture_parameters(avctx, ctx, v, &ctx_pic->pp);
 246
 247     ctx_pic->bitstream_size = 0;
 248     ctx_pic->bitstream      = NULL;
 249     return 0;
 250 }
 251
 252 static int dxva2_vc1_decode_slice(AVCodecContext *avctx,
 253                                   const uint8_t *buffer,
 254                                   uint32_t size)
 255 {
 256     const VC1Context *v = avctx->priv_data;
 257     const Picture *current_picture = v->s.current_picture_ptr;
 258     struct dxva2_picture_context *ctx_pic = current_picture->hwaccel_picture_private;
 259
 260     if (ctx_pic->bitstream_size > 0)
 261         return -1;
 262
 263     if (avctx->codec_id == AV_CODEC_ID_VC1 &&
 264         size >= 4 && IS_MARKER(AV_RB32(buffer))) {
 265         buffer += 4;
 266         size   -= 4;
 267     }
 268
 269     ctx_pic->bitstream_size = size;
 270     ctx_pic->bitstream      = buffer;
 271
 272     fill_slice(avctx, &ctx_pic->si, 0, size);
 273     return 0;
 274 }
 275
 276 static int dxva2_vc1_end_frame(AVCodecContext *avctx)
 277 {
 278     VC1Context *v = avctx->priv_data;
 279     struct dxva2_picture_context *ctx_pic = v->s.current_picture_ptr->hwaccel_picture_private;
 280     int ret;
 281
 282     if (ctx_pic->bitstream_size <= 0)
 283         return -1;
 284
 285     ret = ff_dxva2_common_end_frame(avctx, v->s.current_picture_ptr,
 286                                     &ctx_pic->pp, sizeof(ctx_pic->pp),
 287                                     NULL, 0,
 288                                     commit_bitstream_and_slice_buffer);
 289     if (!ret)
 290         ff_mpeg_draw_horiz_band(&v->s, 0, avctx->height);
 291     return ret;
 292 }
 293
 294 #if CONFIG_WMV3_DXVA2_HWACCEL
 295 AVHWAccel ff_wmv3_dxva2_hwaccel = {
 296     .name           = "wmv3_dxva2",
 297     .type           = AVMEDIA_TYPE_VIDEO,
 298     .id             = AV_CODEC_ID_WMV3,
 299     .pix_fmt        = AV_PIX_FMT_DXVA2_VLD,
 300     .start_frame    = dxva2_vc1_start_frame,
 301     .decode_slice   = dxva2_vc1_decode_slice,
 302     .end_frame      = dxva2_vc1_end_frame,
 303     .priv_data_size = sizeof(struct dxva2_picture_context),
 304 };
 305 #endif
 306
 307 AVHWAccel ff_vc1_dxva2_hwaccel = {
 308     .name           = "vc1_dxva2",
 309     .type           = AVMEDIA_TYPE_VIDEO,
 310     .id             = AV_CODEC_ID_VC1,
 311     .pix_fmt        = AV_PIX_FMT_DXVA2_VLD,
 312     .start_frame    = dxva2_vc1_start_frame,
 313     .decode_slice   = dxva2_vc1_decode_slice,
 314     .end_frame      = dxva2_vc1_end_frame,
 315     .priv_data_size = sizeof(struct dxva2_picture_context),
 316 };