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 "mpegutils.h"
  25 #include "vc1.h"
  26 #include "vc1data.h"
  27
  28 struct dxva2_picture_context {
  29     DXVA_PictureParameters pp;
  30     DXVA_SliceInfo         si;
  31
  32     const uint8_t          *bitstream;
  33     unsigned               bitstream_size;
  34 };
  35
  36 static void fill_picture_parameters(AVCodecContext *avctx,
  37                                     AVDXVAContext *ctx, const VC1Context *v,
  38                                     DXVA_PictureParameters *pp)
  39 {
  40     const MpegEncContext *s = &v->s;
  41     const Picture *current_picture = s->current_picture_ptr;
  42     int intcomp = 0;
  43
  44     // determine if intensity compensation is needed
  45     if (s->pict_type == AV_PICTURE_TYPE_P) {
  46       if ((v->fcm == ILACE_FRAME && v->intcomp) || (v->fcm != ILACE_FRAME && v->mv_mode == MV_PMODE_INTENSITY_COMP)) {
  47         if (v->lumscale != 32 || v->lumshift != 0 || (s->picture_structure != PICT_FRAME && (v->lumscale2 != 32 || v->lumshift2 != 0)))
  48           intcomp = 1;
  49       }
  50     }
  51
  52     memset(pp, 0, sizeof(*pp));
  53     pp->wDecodedPictureIndex    =
  54     pp->wDeblockedPictureIndex  = ff_dxva2_get_surface_index(avctx, ctx, current_picture->f);
  55     if (s->pict_type != AV_PICTURE_TYPE_I && !v->bi_type)
  56         pp->wForwardRefPictureIndex = ff_dxva2_get_surface_index(avctx, ctx, s->last_picture.f);
  57     else
  58         pp->wForwardRefPictureIndex = 0xffff;
  59     if (s->pict_type == AV_PICTURE_TYPE_B && !v->bi_type)
  60         pp->wBackwardRefPictureIndex = ff_dxva2_get_surface_index(avctx, ctx, s->next_picture.f);
  61     else
  62         pp->wBackwardRefPictureIndex = 0xffff;
  63     if (v->profile == PROFILE_ADVANCED) {
  64         /* It is the cropped width/height -1 of the frame */
  65         pp->wPicWidthInMBminus1 = avctx->width  - 1;
  66         pp->wPicHeightInMBminus1= avctx->height - 1;
  67     } else {
  68         /* It is the coded width/height in macroblock -1 of the frame */
  69         pp->wPicWidthInMBminus1 = s->mb_width  - 1;
  70         pp->wPicHeightInMBminus1= s->mb_height - 1;
  71     }
  72     pp->bMacroblockWidthMinus1  = 15;
  73     pp->bMacroblockHeightMinus1 = 15;
  74     pp->bBlockWidthMinus1       = 7;
  75     pp->bBlockHeightMinus1      = 7;
  76     pp->bBPPminus1              = 7;
  77     if (s->picture_structure & PICT_TOP_FIELD)
  78         pp->bPicStructure      |= 0x01;
  79     if (s->picture_structure & PICT_BOTTOM_FIELD)
  80         pp->bPicStructure      |= 0x02;
  81     pp->bSecondField            = v->interlace && v->fcm == ILACE_FIELD && v->second_field;
  82     pp->bPicIntra               = s->pict_type == AV_PICTURE_TYPE_I || v->bi_type;
  83     pp->bPicBackwardPrediction  = s->pict_type == AV_PICTURE_TYPE_B && !v->bi_type;
  84     pp->bBidirectionalAveragingMode = (1                                           << 7) |
  85                                       ((DXVA_CONTEXT_CFG_INTRARESID(avctx, ctx) != 0) << 6) |
  86                                       ((DXVA_CONTEXT_CFG_RESIDACCEL(avctx, ctx) != 0) << 5) |
  87                                       (intcomp                                     << 4) |
  88                                       ((v->profile == PROFILE_ADVANCED)            << 3);
  89     pp->bMVprecisionAndChromaRelation = ((v->mv_mode == MV_PMODE_1MV_HPEL_BILIN) << 3) |
  90                                         (1                                       << 2) |
  91                                         (0                                       << 1) |
  92                                         (!s->quarter_sample                          );
  93     pp->bChromaFormat           = v->chromaformat;
  94     DXVA_CONTEXT_REPORT_ID(avctx, ctx)++;
  95     if (DXVA_CONTEXT_REPORT_ID(avctx, ctx) >= (1 << 16))
  96         DXVA_CONTEXT_REPORT_ID(avctx, ctx) = 1;
  97     pp->bPicScanFixed           = DXVA_CONTEXT_REPORT_ID(avctx, ctx) >> 8;
  98     pp->bPicScanMethod          = DXVA_CONTEXT_REPORT_ID(avctx, ctx) & 0xff;
  99     pp->bPicReadbackRequests    = 0;
 100     pp->bRcontrol               = v->rnd;
 101     pp->bPicSpatialResid8       = (v->panscanflag  << 7) |
 102                                   (v->refdist_flag << 6) |
 103                                   (s->loop_filter  << 5) |
 104                                   (v->fastuvmc     << 4) |
 105                                   (v->extended_mv  << 3) |
 106                                   (v->dquant       << 1) |
 107                                   (v->vstransform      );
 108     pp->bPicOverflowBlocks      = (v->quantizer_mode << 6) |
 109                                   (v->multires       << 5) |
 110                                   (v->resync_marker  << 4) |
 111                                   (v->rangered       << 3) |
 112                                   (s->max_b_frames       );
 113     pp->bPicExtrapolation       = (!v->interlace || v->fcm == PROGRESSIVE) ? 1 : 2;
 114     pp->bPicDeblocked           = ((!pp->bPicBackwardPrediction && v->overlap)        << 6) |
 115                                   ((v->profile != PROFILE_ADVANCED && v->rangeredfrm) << 5) |
 116                                   (s->loop_filter                                     << 1);
 117     pp->bPicDeblockConfined     = (v->postprocflag             << 7) |
 118                                   (v->broadcast                << 6) |
 119                                   (v->interlace                << 5) |
 120                                   (v->tfcntrflag               << 4) |
 121                                   (v->finterpflag              << 3) |
 122                                   ((s->pict_type != AV_PICTURE_TYPE_B) << 2) |
 123                                   (v->psf                      << 1) |
 124                                   (v->extended_dmv                 );
 125     if (s->pict_type != AV_PICTURE_TYPE_I)
 126         pp->bPic4MVallowed      = v->mv_mode == MV_PMODE_MIXED_MV ||
 127                                   (v->mv_mode == MV_PMODE_INTENSITY_COMP &&
 128                                    v->mv_mode2 == MV_PMODE_MIXED_MV);
 129     if (v->profile == PROFILE_ADVANCED)
 130         pp->bPicOBMC            = (v->range_mapy_flag  << 7) |
 131                                   (v->range_mapy       << 4) |
 132                                   (v->range_mapuv_flag << 3) |
 133                                   (v->range_mapuv          );
 134     pp->bPicBinPB               = 0;
 135     pp->bMV_RPS                 = (v->fcm == ILACE_FIELD && pp->bPicBackwardPrediction) ? v->refdist + 9 : 0;
 136     pp->bReservedBits           = v->pq;
 137     if (s->picture_structure == PICT_FRAME) {
 138         if (intcomp) {
 139             pp->wBitstreamFcodes      = v->lumscale;
 140             pp->wBitstreamPCEelements = v->lumshift;
 141         } else {
 142             pp->wBitstreamFcodes      = 32;
 143             pp->wBitstreamPCEelements = 0;
 144         }
 145     } else {
 146         /* Syntax: (top_field_param << 8) | bottom_field_param */
 147         if (intcomp) {
 148             pp->wBitstreamFcodes      = (v->lumscale << 8) | v->lumscale2;
 149             pp->wBitstreamPCEelements = (v->lumshift << 8) | v->lumshift2;
 150         } else {
 151             pp->wBitstreamFcodes      = (32 << 8) | 32;
 152             pp->wBitstreamPCEelements = 0;
 153         }
 154     }
 155     pp->bBitstreamConcealmentNeed   = 0;
 156     pp->bBitstreamConcealmentMethod = 0;
 157 }
 158
 159 static void fill_slice(AVCodecContext *avctx, DXVA_SliceInfo *slice,
 160                        unsigned position, unsigned size)
 161 {
 162     const VC1Context *v = avctx->priv_data;
 163     const MpegEncContext *s = &v->s;
 164
 165     memset(slice, 0, sizeof(*slice));
 166     slice->wHorizontalPosition = 0;
 167     slice->wVerticalPosition   = s->mb_y;
 168     slice->dwSliceBitsInBuffer = 8 * size;
 169     slice->dwSliceDataLocation = position;
 170     slice->bStartCodeBitOffset = 0;
 171     slice->bReservedBits       = (s->pict_type == AV_PICTURE_TYPE_B && !v->bi_type) ? v->bfraction_lut_index + 9 : 0;
 172     slice->wMBbitOffset        = v->p_frame_skipped ? 0xffff : get_bits_count(&s->gb) + (avctx->codec_id == AV_CODEC_ID_VC1 ? 32 : 0);
 173     slice->wNumberMBsInSlice   = s->mb_width * s->mb_height; /* XXX We assume 1 slice */
 174     slice->wQuantizerScaleCode = v->pq;
 175     slice->wBadSliceChopping   = 0;
 176 }
 177
 178 static int commit_bitstream_and_slice_buffer(AVCodecContext *avctx,
 179                                              DECODER_BUFFER_DESC *bs,
 180                                              DECODER_BUFFER_DESC *sc)
 181 {
 182     const VC1Context *v = avctx->priv_data;
 183     AVDXVAContext *ctx = avctx->hwaccel_context;
 184     const MpegEncContext *s = &v->s;
 185     struct dxva2_picture_context *ctx_pic = s->current_picture_ptr->hwaccel_picture_private;
 186
 187     DXVA_SliceInfo *slice = &ctx_pic->si;
 188
 189     static const uint8_t start_code[] = { 0, 0, 1, 0x0d };
 190     const unsigned start_code_size = avctx->codec_id == AV_CODEC_ID_VC1 ? sizeof(start_code) : 0;
 191     const unsigned slice_size = slice->dwSliceBitsInBuffer / 8;
 192     const unsigned padding = 128 - ((start_code_size + slice_size) & 127);
 193     const unsigned data_size = start_code_size + slice_size + padding;
 194
 195     void     *dxva_data_ptr;
 196     uint8_t  *dxva_data;
 197     unsigned dxva_size;
 198     int result;
 199     unsigned type;
 200
 201 #if CONFIG_D3D11VA
 202     if (avctx->pix_fmt == AV_PIX_FMT_D3D11VA_VLD) {
 203         type = D3D11_VIDEO_DECODER_BUFFER_BITSTREAM;
 204         if (FAILED(ID3D11VideoContext_GetDecoderBuffer(D3D11VA_CONTEXT(ctx)->video_context,
 205                                                        D3D11VA_CONTEXT(ctx)->decoder,
 206                                                        type,
 207                                                        &dxva_size, &dxva_data_ptr)))
 208             return -1;
 209     }
 210 #endif
 211 #if CONFIG_DXVA2
 212     if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD) {
 213         type = DXVA2_BitStreamDateBufferType;
 214         if (FAILED(IDirectXVideoDecoder_GetBuffer(DXVA2_CONTEXT(ctx)->decoder,
 215                                                   type,
 216                                                   &dxva_data_ptr, &dxva_size)))
 217             return -1;
 218     }
 219 #endif
 220
 221     dxva_data = dxva_data_ptr;
 222     result = data_size <= dxva_size ? 0 : -1;
 223     if (!result) {
 224         if (start_code_size > 0) {
 225             memcpy(dxva_data, start_code, start_code_size);
 226             if (v->second_field)
 227                 dxva_data[3] = 0x0c;
 228         }
 229         memcpy(dxva_data + start_code_size,
 230                ctx_pic->bitstream + slice->dwSliceDataLocation, slice_size);
 231         if (padding > 0)
 232             memset(dxva_data + start_code_size + slice_size, 0, padding);
 233         slice->dwSliceBitsInBuffer = 8 * data_size;
 234     }
 235 #if CONFIG_D3D11VA
 236     if (avctx->pix_fmt == AV_PIX_FMT_D3D11VA_VLD)
 237         if (FAILED(ID3D11VideoContext_ReleaseDecoderBuffer(D3D11VA_CONTEXT(ctx)->video_context, D3D11VA_CONTEXT(ctx)->decoder, type)))
 238             return -1;
 239 #endif
 240 #if CONFIG_DXVA2
 241     if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD)
 242         if (FAILED(IDirectXVideoDecoder_ReleaseBuffer(DXVA2_CONTEXT(ctx)->decoder, type)))
 243             return -1;
 244 #endif
 245     if (result)
 246         return result;
 247
 248 #if CONFIG_D3D11VA
 249     if (avctx->pix_fmt == AV_PIX_FMT_D3D11VA_VLD) {
 250         D3D11_VIDEO_DECODER_BUFFER_DESC *dsc11 = bs;
 251         memset(dsc11, 0, sizeof(*dsc11));
 252         dsc11->BufferType           = type;
 253         dsc11->DataSize             = data_size;
 254         dsc11->NumMBsInBuffer       = s->mb_width * s->mb_height;
 255
 256         type = D3D11_VIDEO_DECODER_BUFFER_SLICE_CONTROL;
 257     }
 258 #endif
 259 #if CONFIG_DXVA2
 260     if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD) {
 261         DXVA2_DecodeBufferDesc *dsc2 = bs;
 262         memset(dsc2, 0, sizeof(*dsc2));
 263         dsc2->CompressedBufferType = type;
 264         dsc2->DataSize             = data_size;
 265         dsc2->NumMBsInBuffer       = s->mb_width * s->mb_height;
 266
 267         type = DXVA2_SliceControlBufferType;
 268     }
 269 #endif
 270     assert((data_size & 127) == 0);
 271
 272     return ff_dxva2_commit_buffer(avctx, ctx, sc,
 273                                   type,
 274                                   slice, sizeof(*slice), s->mb_width * s->mb_height);
 275 }
 276
 277 static int dxva2_vc1_start_frame(AVCodecContext *avctx,
 278                                  av_unused const uint8_t *buffer,
 279                                  av_unused uint32_t size)
 280 {
 281     const VC1Context *v = avctx->priv_data;
 282     AVDXVAContext *ctx = avctx->hwaccel_context;
 283     struct dxva2_picture_context *ctx_pic = v->s.current_picture_ptr->hwaccel_picture_private;
 284
 285     if (DXVA_CONTEXT_DECODER(avctx, ctx) == NULL ||
 286         DXVA_CONTEXT_CFG(avctx, ctx) == NULL ||
 287         DXVA_CONTEXT_COUNT(avctx, ctx) <= 0)
 288         return -1;
 289     assert(ctx_pic);
 290
 291     fill_picture_parameters(avctx, ctx, v, &ctx_pic->pp);
 292
 293     ctx_pic->bitstream_size = 0;
 294     ctx_pic->bitstream      = NULL;
 295     return 0;
 296 }
 297
 298 static int dxva2_vc1_decode_slice(AVCodecContext *avctx,
 299                                   const uint8_t *buffer,
 300                                   uint32_t size)
 301 {
 302     const VC1Context *v = avctx->priv_data;
 303     const Picture *current_picture = v->s.current_picture_ptr;
 304     struct dxva2_picture_context *ctx_pic = current_picture->hwaccel_picture_private;
 305
 306     if (ctx_pic->bitstream_size > 0)
 307         return -1;
 308
 309     if (avctx->codec_id == AV_CODEC_ID_VC1 &&
 310         size >= 4 && IS_MARKER(AV_RB32(buffer))) {
 311         buffer += 4;
 312         size   -= 4;
 313     }
 314
 315     ctx_pic->bitstream_size = size;
 316     ctx_pic->bitstream      = buffer;
 317
 318     fill_slice(avctx, &ctx_pic->si, 0, size);
 319     return 0;
 320 }
 321
 322 static int dxva2_vc1_end_frame(AVCodecContext *avctx)
 323 {
 324     VC1Context *v = avctx->priv_data;
 325     struct dxva2_picture_context *ctx_pic = v->s.current_picture_ptr->hwaccel_picture_private;
 326     int ret;
 327
 328     if (ctx_pic->bitstream_size <= 0)
 329         return -1;
 330
 331     ret = ff_dxva2_common_end_frame(avctx, v->s.current_picture_ptr->f,
 332                                     &ctx_pic->pp, sizeof(ctx_pic->pp),
 333                                     NULL, 0,
 334                                     commit_bitstream_and_slice_buffer);
 335     if (!ret)
 336         ff_mpeg_draw_horiz_band(&v->s, 0, avctx->height);
 337     return ret;
 338 }
 339
 340 #if CONFIG_WMV3_DXVA2_HWACCEL
 341 AVHWAccel ff_wmv3_dxva2_hwaccel = {
 342     .name           = "wmv3_dxva2",
 343     .type           = AVMEDIA_TYPE_VIDEO,
 344     .id             = AV_CODEC_ID_WMV3,
 345     .pix_fmt        = AV_PIX_FMT_DXVA2_VLD,
 346     .start_frame    = dxva2_vc1_start_frame,
 347     .decode_slice   = dxva2_vc1_decode_slice,
 348     .end_frame      = dxva2_vc1_end_frame,
 349     .frame_priv_data_size = sizeof(struct dxva2_picture_context),
 350 };
 351 #endif
 352
 353 #if CONFIG_VC1_DXVA2_HWACCEL
 354 AVHWAccel ff_vc1_dxva2_hwaccel = {
 355     .name           = "vc1_dxva2",
 356     .type           = AVMEDIA_TYPE_VIDEO,
 357     .id             = AV_CODEC_ID_VC1,
 358     .pix_fmt        = AV_PIX_FMT_DXVA2_VLD,
 359     .start_frame    = dxva2_vc1_start_frame,
 360     .decode_slice   = dxva2_vc1_decode_slice,
 361     .end_frame      = dxva2_vc1_end_frame,
 362     .frame_priv_data_size = sizeof(struct dxva2_picture_context),
 363 };
 364 #endif
 365
 366 #if CONFIG_WMV3_D3D11VA_HWACCEL
 367 AVHWAccel ff_wmv3_d3d11va_hwaccel = {
 368     .name           = "wmv3_d3d11va",
 369     .type           = AVMEDIA_TYPE_VIDEO,
 370     .id             = AV_CODEC_ID_WMV3,
 371     .pix_fmt        = AV_PIX_FMT_D3D11VA_VLD,
 372     .start_frame    = dxva2_vc1_start_frame,
 373     .decode_slice   = dxva2_vc1_decode_slice,
 374     .end_frame      = dxva2_vc1_end_frame,
 375     .frame_priv_data_size = sizeof(struct dxva2_picture_context),
 376 };
 377 #endif
 378
 379 #if CONFIG_VC1_D3D11VA_HWACCEL
 380 AVHWAccel ff_vc1_d3d11va_hwaccel = {
 381     .name           = "vc1_d3d11va",
 382     .type           = AVMEDIA_TYPE_VIDEO,
 383     .id             = AV_CODEC_ID_VC1,
 384     .pix_fmt        = AV_PIX_FMT_D3D11VA_VLD,
 385     .start_frame    = dxva2_vc1_start_frame,
 386     .decode_slice   = dxva2_vc1_decode_slice,
 387     .end_frame      = dxva2_vc1_end_frame,
 388     .frame_priv_data_size = sizeof(struct dxva2_picture_context),
 389 };
 390 #endif