git.sesse.net Git - ffmpeg/blob - libavcodec/dxva2_vp9.c

   1 /*
   2  * DXVA2 VP9 HW acceleration.
   3  *
   4  * copyright (c) 2015 Hendrik Leppkes
   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 "libavutil/avassert.h"
  24 #include "libavutil/pixdesc.h"
  25
  26 #include "dxva2_internal.h"
  27 #include "vp9shared.h"
  28
  29 struct vp9_dxva2_picture_context {
  30     DXVA_PicParams_VP9    pp;
  31     DXVA_Slice_VPx_Short  slice;
  32     const uint8_t         *bitstream;
  33     unsigned              bitstream_size;
  34 };
  35
  36 static void fill_picture_entry(DXVA_PicEntry_VPx *pic,
  37                                unsigned index, unsigned flag)
  38 {
  39     av_assert0((index & 0x7f) == index && (flag & 0x01) == flag);
  40     pic->bPicEntry = index | (flag << 7);
  41 }
  42
  43 static int fill_picture_parameters(const AVCodecContext *avctx, AVDXVAContext *ctx, const VP9SharedContext *h,
  44                                     DXVA_PicParams_VP9 *pp)
  45 {
  46     int i;
  47     const AVPixFmtDescriptor * pixdesc = av_pix_fmt_desc_get(avctx->sw_pix_fmt);
  48
  49     if (!pixdesc)
  50         return -1;
  51
  52     memset(pp, 0, sizeof(*pp));
  53
  54     fill_picture_entry(&pp->CurrPic, ff_dxva2_get_surface_index(avctx, ctx, h->frames[CUR_FRAME].tf.f), 0);
  55
  56     pp->profile = h->h.profile;
  57     pp->wFormatAndPictureInfoFlags = ((h->h.keyframe == 0)   <<  0) |
  58                                      ((h->h.invisible == 0)  <<  1) |
  59                                      (h->h.errorres          <<  2) |
  60                                      (pixdesc->log2_chroma_w <<  3) | /* subsampling_x */
  61                                      (pixdesc->log2_chroma_h <<  4) | /* subsampling_y */
  62                                      (0                      <<  5) | /* extra_plane */
  63                                      (h->h.refreshctx        <<  6) |
  64                                      (h->h.parallelmode      <<  7) |
  65                                      (h->h.intraonly         <<  8) |
  66                                      (h->h.framectxid        <<  9) |
  67                                      (h->h.resetctx          << 11) |
  68                                      ((h->h.keyframe ? 0 : h->h.highprecisionmvs) << 13) |
  69                                      (0                      << 14);  /* ReservedFormatInfo2Bits */
  70
  71     pp->width  = avctx->width;
  72     pp->height = avctx->height;
  73     pp->BitDepthMinus8Luma   = pixdesc->comp[0].depth - 8;
  74     pp->BitDepthMinus8Chroma = pixdesc->comp[1].depth - 8;
  75     /* swap 0/1 to match the reference */
  76     pp->interp_filter = h->h.filtermode ^ (h->h.filtermode <= 1);
  77     pp->Reserved8Bits = 0;
  78
  79     for (i = 0; i < 8; i++) {
  80         if (h->refs[i].f->buf[0]) {
  81             fill_picture_entry(&pp->ref_frame_map[i], ff_dxva2_get_surface_index(avctx, ctx, h->refs[i].f), 0);
  82             pp->ref_frame_coded_width[i]  = h->refs[i].f->width;
  83             pp->ref_frame_coded_height[i] = h->refs[i].f->height;
  84         } else
  85             pp->ref_frame_map[i].bPicEntry = 0xFF;
  86     }
  87
  88     for (i = 0; i < 3; i++) {
  89         uint8_t refidx = h->h.refidx[i];
  90         if (h->refs[refidx].f->buf[0])
  91             fill_picture_entry(&pp->frame_refs[i], ff_dxva2_get_surface_index(avctx, ctx, h->refs[refidx].f), 0);
  92         else
  93             pp->frame_refs[i].bPicEntry = 0xFF;
  94
  95         pp->ref_frame_sign_bias[i + 1] = h->h.signbias[i];
  96     }
  97
  98     pp->filter_level    = h->h.filter.level;
  99     pp->sharpness_level = h->h.filter.sharpness;
 100
 101     pp->wControlInfoFlags = (h->h.lf_delta.enabled   << 0) |
 102                             (h->h.lf_delta.updated   << 1) |
 103                             (h->h.use_last_frame_mvs << 2) |
 104                             (0                       << 3);  /* ReservedControlInfo5Bits */
 105
 106     for (i = 0; i < 4; i++)
 107         pp->ref_deltas[i]  = h->h.lf_delta.ref[i];
 108
 109     for (i = 0; i < 2; i++)
 110         pp->mode_deltas[i]  = h->h.lf_delta.mode[i];
 111
 112     pp->base_qindex   = h->h.yac_qi;
 113     pp->y_dc_delta_q  = h->h.ydc_qdelta;
 114     pp->uv_dc_delta_q = h->h.uvdc_qdelta;
 115     pp->uv_ac_delta_q = h->h.uvac_qdelta;
 116
 117     /* segmentation data */
 118     pp->stVP9Segments.wSegmentInfoFlags = (h->h.segmentation.enabled       << 0) |
 119                                           (h->h.segmentation.update_map    << 1) |
 120                                           (h->h.segmentation.temporal      << 2) |
 121                                           (h->h.segmentation.absolute_vals << 3) |
 122                                           (0                               << 4);  /* ReservedSegmentFlags4Bits */
 123
 124     for (i = 0; i < 7; i++)
 125         pp->stVP9Segments.tree_probs[i] = h->h.segmentation.prob[i];
 126
 127     if (h->h.segmentation.temporal)
 128         for (i = 0; i < 3; i++)
 129             pp->stVP9Segments.pred_probs[i] = h->h.segmentation.pred_prob[i];
 130     else
 131         memset(pp->stVP9Segments.pred_probs, 255, sizeof(pp->stVP9Segments.pred_probs));
 132
 133     for (i = 0; i < 8; i++) {
 134         pp->stVP9Segments.feature_mask[i] = (h->h.segmentation.feat[i].q_enabled    << 0) |
 135                                             (h->h.segmentation.feat[i].lf_enabled   << 1) |
 136                                             (h->h.segmentation.feat[i].ref_enabled  << 2) |
 137                                             (h->h.segmentation.feat[i].skip_enabled << 3);
 138
 139         pp->stVP9Segments.feature_data[i][0] = h->h.segmentation.feat[i].q_val;
 140         pp->stVP9Segments.feature_data[i][1] = h->h.segmentation.feat[i].lf_val;
 141         pp->stVP9Segments.feature_data[i][2] = h->h.segmentation.feat[i].ref_val;
 142         pp->stVP9Segments.feature_data[i][3] = 0; /* no data for skip */
 143     }
 144
 145     pp->log2_tile_cols = h->h.tiling.log2_tile_cols;
 146     pp->log2_tile_rows = h->h.tiling.log2_tile_rows;
 147
 148     pp->uncompressed_header_size_byte_aligned = h->h.uncompressed_header_size;
 149     pp->first_partition_size = h->h.compressed_header_size;
 150
 151     pp->StatusReportFeedbackNumber = 1 + DXVA_CONTEXT_REPORT_ID(avctx, ctx)++;
 152     return 0;
 153 }
 154
 155 static void fill_slice_short(DXVA_Slice_VPx_Short *slice,
 156                              unsigned position, unsigned size)
 157 {
 158     memset(slice, 0, sizeof(*slice));
 159     slice->BSNALunitDataLocation = position;
 160     slice->SliceBytesInBuffer    = size;
 161     slice->wBadSliceChopping     = 0;
 162 }
 163
 164 static int commit_bitstream_and_slice_buffer(AVCodecContext *avctx,
 165                                              DECODER_BUFFER_DESC *bs,
 166                                              DECODER_BUFFER_DESC *sc)
 167 {
 168     const VP9SharedContext *h = avctx->priv_data;
 169     AVDXVAContext *ctx = DXVA_CONTEXT(avctx);
 170     struct vp9_dxva2_picture_context *ctx_pic = h->frames[CUR_FRAME].hwaccel_picture_private;
 171     void     *dxva_data_ptr;
 172     uint8_t  *dxva_data;
 173     unsigned dxva_size;
 174     unsigned padding;
 175     unsigned type;
 176
 177 #if CONFIG_D3D11VA
 178     if (ff_dxva2_is_d3d11(avctx)) {
 179         type = D3D11_VIDEO_DECODER_BUFFER_BITSTREAM;
 180         if (FAILED(ID3D11VideoContext_GetDecoderBuffer(D3D11VA_CONTEXT(ctx)->video_context,
 181                                                        D3D11VA_CONTEXT(ctx)->decoder,
 182                                                        type,
 183                                                        &dxva_size, &dxva_data_ptr)))
 184             return -1;
 185     }
 186 #endif
 187 #if CONFIG_DXVA2
 188     if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD) {
 189         type = DXVA2_BitStreamDateBufferType;
 190         if (FAILED(IDirectXVideoDecoder_GetBuffer(DXVA2_CONTEXT(ctx)->decoder,
 191                                                   type,
 192                                                   &dxva_data_ptr, &dxva_size)))
 193             return -1;
 194     }
 195 #endif
 196
 197     dxva_data = dxva_data_ptr;
 198
 199     if (ctx_pic->slice.SliceBytesInBuffer > dxva_size) {
 200         av_log(avctx, AV_LOG_ERROR, "Failed to build bitstream");
 201         return -1;
 202     }
 203
 204     memcpy(dxva_data, ctx_pic->bitstream, ctx_pic->slice.SliceBytesInBuffer);
 205
 206     padding = FFMIN(128 - ((ctx_pic->slice.SliceBytesInBuffer) & 127), dxva_size - ctx_pic->slice.SliceBytesInBuffer);
 207     if (padding > 0) {
 208         memset(dxva_data + ctx_pic->slice.SliceBytesInBuffer, 0, padding);
 209         ctx_pic->slice.SliceBytesInBuffer += padding;
 210     }
 211
 212 #if CONFIG_D3D11VA
 213     if (ff_dxva2_is_d3d11(avctx))
 214         if (FAILED(ID3D11VideoContext_ReleaseDecoderBuffer(D3D11VA_CONTEXT(ctx)->video_context, D3D11VA_CONTEXT(ctx)->decoder, type)))
 215             return -1;
 216 #endif
 217 #if CONFIG_DXVA2
 218     if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD)
 219         if (FAILED(IDirectXVideoDecoder_ReleaseBuffer(DXVA2_CONTEXT(ctx)->decoder, type)))
 220             return -1;
 221 #endif
 222
 223 #if CONFIG_D3D11VA
 224     if (ff_dxva2_is_d3d11(avctx)) {
 225         D3D11_VIDEO_DECODER_BUFFER_DESC *dsc11 = bs;
 226         memset(dsc11, 0, sizeof(*dsc11));
 227         dsc11->BufferType           = type;
 228         dsc11->DataSize             = ctx_pic->slice.SliceBytesInBuffer;
 229         dsc11->NumMBsInBuffer       = 0;
 230
 231         type = D3D11_VIDEO_DECODER_BUFFER_SLICE_CONTROL;
 232     }
 233 #endif
 234 #if CONFIG_DXVA2
 235     if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD) {
 236         DXVA2_DecodeBufferDesc *dsc2 = bs;
 237         memset(dsc2, 0, sizeof(*dsc2));
 238         dsc2->CompressedBufferType = type;
 239         dsc2->DataSize             = ctx_pic->slice.SliceBytesInBuffer;
 240         dsc2->NumMBsInBuffer       = 0;
 241
 242         type = DXVA2_SliceControlBufferType;
 243     }
 244 #endif
 245
 246     return ff_dxva2_commit_buffer(avctx, ctx, sc,
 247                                   type,
 248                                   &ctx_pic->slice, sizeof(ctx_pic->slice), 0);
 249 }
 250
 251
 252 static int dxva2_vp9_start_frame(AVCodecContext *avctx,
 253                                  av_unused const uint8_t *buffer,
 254                                  av_unused uint32_t size)
 255 {
 256     const VP9SharedContext *h = avctx->priv_data;
 257     AVDXVAContext *ctx = DXVA_CONTEXT(avctx);
 258     struct vp9_dxva2_picture_context *ctx_pic = h->frames[CUR_FRAME].hwaccel_picture_private;
 259
 260     if (!DXVA_CONTEXT_VALID(avctx, ctx))
 261         return -1;
 262     av_assert0(ctx_pic);
 263
 264     /* Fill up DXVA_PicParams_VP9 */
 265     if (fill_picture_parameters(avctx, ctx, h, &ctx_pic->pp) < 0)
 266         return -1;
 267
 268     ctx_pic->bitstream_size = 0;
 269     ctx_pic->bitstream      = NULL;
 270     return 0;
 271 }
 272
 273 static int dxva2_vp9_decode_slice(AVCodecContext *avctx,
 274                                   const uint8_t *buffer,
 275                                   uint32_t size)
 276 {
 277     const VP9SharedContext *h = avctx->priv_data;
 278     struct vp9_dxva2_picture_context *ctx_pic = h->frames[CUR_FRAME].hwaccel_picture_private;
 279     unsigned position;
 280
 281     if (!ctx_pic->bitstream)
 282         ctx_pic->bitstream = buffer;
 283     ctx_pic->bitstream_size += size;
 284
 285     position = buffer - ctx_pic->bitstream;
 286     fill_slice_short(&ctx_pic->slice, position, size);
 287
 288     return 0;
 289 }
 290
 291 static int dxva2_vp9_end_frame(AVCodecContext *avctx)
 292 {
 293     VP9SharedContext *h = avctx->priv_data;
 294     struct vp9_dxva2_picture_context *ctx_pic = h->frames[CUR_FRAME].hwaccel_picture_private;
 295     int ret;
 296
 297     if (ctx_pic->bitstream_size <= 0)
 298         return -1;
 299
 300     ret = ff_dxva2_common_end_frame(avctx, h->frames[CUR_FRAME].tf.f,
 301                                     &ctx_pic->pp, sizeof(ctx_pic->pp),
 302                                     NULL, 0,
 303                                     commit_bitstream_and_slice_buffer);
 304     return ret;
 305 }
 306
 307 #if CONFIG_VP9_DXVA2_HWACCEL
 308 const AVHWAccel ff_vp9_dxva2_hwaccel = {
 309     .name           = "vp9_dxva2",
 310     .type           = AVMEDIA_TYPE_VIDEO,
 311     .id             = AV_CODEC_ID_VP9,
 312     .pix_fmt        = AV_PIX_FMT_DXVA2_VLD,
 313     .init           = ff_dxva2_decode_init,
 314     .uninit         = ff_dxva2_decode_uninit,
 315     .start_frame    = dxva2_vp9_start_frame,
 316     .decode_slice   = dxva2_vp9_decode_slice,
 317     .end_frame      = dxva2_vp9_end_frame,
 318     .frame_params   = ff_dxva2_common_frame_params,
 319     .frame_priv_data_size = sizeof(struct vp9_dxva2_picture_context),
 320     .priv_data_size = sizeof(FFDXVASharedContext),
 321 };
 322 #endif
 323
 324 #if CONFIG_VP9_D3D11VA_HWACCEL
 325 const AVHWAccel ff_vp9_d3d11va_hwaccel = {
 326     .name           = "vp9_d3d11va",
 327     .type           = AVMEDIA_TYPE_VIDEO,
 328     .id             = AV_CODEC_ID_VP9,
 329     .pix_fmt        = AV_PIX_FMT_D3D11VA_VLD,
 330     .init           = ff_dxva2_decode_init,
 331     .uninit         = ff_dxva2_decode_uninit,
 332     .start_frame    = dxva2_vp9_start_frame,
 333     .decode_slice   = dxva2_vp9_decode_slice,
 334     .end_frame      = dxva2_vp9_end_frame,
 335     .frame_params   = ff_dxva2_common_frame_params,
 336     .frame_priv_data_size = sizeof(struct vp9_dxva2_picture_context),
 337     .priv_data_size = sizeof(FFDXVASharedContext),
 338 };
 339 #endif
 340
 341 #if CONFIG_VP9_D3D11VA2_HWACCEL
 342 const AVHWAccel ff_vp9_d3d11va2_hwaccel = {
 343     .name           = "vp9_d3d11va2",
 344     .type           = AVMEDIA_TYPE_VIDEO,
 345     .id             = AV_CODEC_ID_VP9,
 346     .pix_fmt        = AV_PIX_FMT_D3D11,
 347     .init           = ff_dxva2_decode_init,
 348     .uninit         = ff_dxva2_decode_uninit,
 349     .start_frame    = dxva2_vp9_start_frame,
 350     .decode_slice   = dxva2_vp9_decode_slice,
 351     .end_frame      = dxva2_vp9_end_frame,
 352     .frame_params   = ff_dxva2_common_frame_params,
 353     .frame_priv_data_size = sizeof(struct vp9_dxva2_picture_context),
 354     .priv_data_size = sizeof(FFDXVASharedContext),
 355 };
 356 #endif