yading@10: /*
yading@10:  * DXVA2 WMV3/VC-1 HW acceleration.
yading@10:  *
yading@10:  * copyright (c) 2010 Laurent Aimar
yading@10:  *
yading@10:  * This file is part of FFmpeg.
yading@10:  *
yading@10:  * FFmpeg is free software; you can redistribute it and/or
yading@10:  * modify it under the terms of the GNU Lesser General Public
yading@10:  * License as published by the Free Software Foundation; either
yading@10:  * version 2.1 of the License, or (at your option) any later version.
yading@10:  *
yading@10:  * FFmpeg is distributed in the hope that it will be useful,
yading@10:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
yading@10:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
yading@10:  * Lesser General Public License for more details.
yading@10:  *
yading@10:  * You should have received a copy of the GNU Lesser General Public
yading@10:  * License along with FFmpeg; if not, write to the Free Software
yading@10:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
yading@10:  */
yading@10: 
yading@10: #include "dxva2_internal.h"
yading@10: #include "vc1.h"
yading@10: #include "vc1data.h"
yading@10: 
yading@10: struct dxva2_picture_context {
yading@10:     DXVA_PictureParameters pp;
yading@10:     DXVA_SliceInfo         si;
yading@10: 
yading@10:     const uint8_t          *bitstream;
yading@10:     unsigned               bitstream_size;
yading@10: };
yading@10: 
yading@10: static void fill_picture_parameters(AVCodecContext *avctx,
yading@10:                                     struct dxva_context *ctx, const VC1Context *v,
yading@10:                                     DXVA_PictureParameters *pp)
yading@10: {
yading@10:     const MpegEncContext *s = &v->s;
yading@10:     const Picture *current_picture = s->current_picture_ptr;
yading@10: 
yading@10:     memset(pp, 0, sizeof(*pp));
yading@10:     pp->wDecodedPictureIndex    =
yading@10:     pp->wDeblockedPictureIndex  = ff_dxva2_get_surface_index(ctx, current_picture);
yading@10:     if (s->pict_type != AV_PICTURE_TYPE_I && !v->bi_type)
yading@10:         pp->wForwardRefPictureIndex = ff_dxva2_get_surface_index(ctx, &s->last_picture);
yading@10:     else
yading@10:         pp->wForwardRefPictureIndex = 0xffff;
yading@10:     if (s->pict_type == AV_PICTURE_TYPE_B && !v->bi_type)
yading@10:         pp->wBackwardRefPictureIndex = ff_dxva2_get_surface_index(ctx, &s->next_picture);
yading@10:     else
yading@10:         pp->wBackwardRefPictureIndex = 0xffff;
yading@10:     if (v->profile == PROFILE_ADVANCED) {
yading@10:         /* It is the cropped width/height -1 of the frame */
yading@10:         pp->wPicWidthInMBminus1 = avctx->width  - 1;
yading@10:         pp->wPicHeightInMBminus1= avctx->height - 1;
yading@10:     } else {
yading@10:         /* It is the coded width/height in macroblock -1 of the frame */
yading@10:         pp->wPicWidthInMBminus1 = s->mb_width  - 1;
yading@10:         pp->wPicHeightInMBminus1= s->mb_height - 1;
yading@10:     }
yading@10:     pp->bMacroblockWidthMinus1  = 15;
yading@10:     pp->bMacroblockHeightMinus1 = 15;
yading@10:     pp->bBlockWidthMinus1       = 7;
yading@10:     pp->bBlockHeightMinus1      = 7;
yading@10:     pp->bBPPminus1              = 7;
yading@10:     if (s->picture_structure & PICT_TOP_FIELD)
yading@10:         pp->bPicStructure      |= 0x01;
yading@10:     if (s->picture_structure & PICT_BOTTOM_FIELD)
yading@10:         pp->bPicStructure      |= 0x02;
yading@10:     pp->bSecondField            = v->interlace && v->fcm == ILACE_FIELD && v->second_field;
yading@10:     pp->bPicIntra               = s->pict_type == AV_PICTURE_TYPE_I || v->bi_type;
yading@10:     pp->bPicBackwardPrediction  = s->pict_type == AV_PICTURE_TYPE_B && !v->bi_type;
yading@10:     pp->bBidirectionalAveragingMode = (1                                           << 7) |
yading@10:                                       ((ctx->cfg->ConfigIntraResidUnsigned != 0)   << 6) |
yading@10:                                       ((ctx->cfg->ConfigResidDiffAccelerator != 0) << 5) |
yading@10:                                       ((v->lumscale != 32 || v->lumshift != 0)     << 4) |
yading@10:                                       ((v->profile == PROFILE_ADVANCED)            << 3);
yading@10:     pp->bMVprecisionAndChromaRelation = ((v->mv_mode == MV_PMODE_1MV_HPEL_BILIN) << 3) |
yading@10:                                         (1                                       << 2) |
yading@10:                                         (0                                       << 1) |
yading@10:                                         (!s->quarter_sample                          );
yading@10:     pp->bChromaFormat           = v->chromaformat;
yading@10:     ctx->report_id++;
yading@10:     if (ctx->report_id >= (1 << 16))
yading@10:         ctx->report_id = 1;
yading@10:     pp->bPicScanFixed           = ctx->report_id >> 8;
yading@10:     pp->bPicScanMethod          = ctx->report_id & 0xff;
yading@10:     pp->bPicReadbackRequests    = 0;
yading@10:     pp->bRcontrol               = v->rnd;
yading@10:     pp->bPicSpatialResid8       = (v->panscanflag  << 7) |
yading@10:                                   (v->refdist_flag << 6) |
yading@10:                                   (s->loop_filter  << 5) |
yading@10:                                   (v->fastuvmc     << 4) |
yading@10:                                   (v->extended_mv  << 3) |
yading@10:                                   (v->dquant       << 1) |
yading@10:                                   (v->vstransform      );
yading@10:     pp->bPicOverflowBlocks      = (v->quantizer_mode << 6) |
yading@10:                                   (v->multires       << 5) |
yading@10:                                   (s->resync_marker  << 4) |
yading@10:                                   (v->rangered       << 3) |
yading@10:                                   (s->max_b_frames       );
yading@10:     pp->bPicExtrapolation       = (!v->interlace || v->fcm == PROGRESSIVE) ? 1 : 2;
yading@10:     pp->bPicDeblocked           = ((!pp->bPicBackwardPrediction && v->overlap)        << 6) |
yading@10:                                   ((v->profile != PROFILE_ADVANCED && v->rangeredfrm) << 5) |
yading@10:                                   (s->loop_filter                                     << 1);
yading@10:     pp->bPicDeblockConfined     = (v->postprocflag             << 7) |
yading@10:                                   (v->broadcast                << 6) |
yading@10:                                   (v->interlace                << 5) |
yading@10:                                   (v->tfcntrflag               << 4) |
yading@10:                                   (v->finterpflag              << 3) |
yading@10:                                   ((s->pict_type != AV_PICTURE_TYPE_B) << 2) |
yading@10:                                   (v->psf                      << 1) |
yading@10:                                   (v->extended_dmv                 );
yading@10:     if (s->pict_type != AV_PICTURE_TYPE_I)
yading@10:         pp->bPic4MVallowed      = v->mv_mode == MV_PMODE_MIXED_MV ||
yading@10:                                   (v->mv_mode == MV_PMODE_INTENSITY_COMP &&
yading@10:                                    v->mv_mode2 == MV_PMODE_MIXED_MV);
yading@10:     if (v->profile == PROFILE_ADVANCED)
yading@10:         pp->bPicOBMC            = (v->range_mapy_flag  << 7) |
yading@10:                                   (v->range_mapy       << 4) |
yading@10:                                   (v->range_mapuv_flag << 3) |
yading@10:                                   (v->range_mapuv          );
yading@10:     pp->bPicBinPB               = 0;
yading@10:     pp->bMV_RPS                 = 0;
yading@10:     pp->bReservedBits           = 0;
yading@10:     if (s->picture_structure == PICT_FRAME) {
yading@10:         pp->wBitstreamFcodes        = v->lumscale;
yading@10:         pp->wBitstreamPCEelements   = v->lumshift;
yading@10:     } else {
yading@10:         /* Syntax: (top_field_param << 8) | bottom_field_param */
yading@10:         pp->wBitstreamFcodes        = (v->lumscale << 8) | v->lumscale;
yading@10:         pp->wBitstreamPCEelements   = (v->lumshift << 8) | v->lumshift;
yading@10:     }
yading@10:     pp->bBitstreamConcealmentNeed   = 0;
yading@10:     pp->bBitstreamConcealmentMethod = 0;
yading@10: }
yading@10: 
yading@10: static void fill_slice(AVCodecContext *avctx, DXVA_SliceInfo *slice,
yading@10:                        unsigned position, unsigned size)
yading@10: {
yading@10:     const VC1Context *v = avctx->priv_data;
yading@10:     const MpegEncContext *s = &v->s;
yading@10: 
yading@10:     memset(slice, 0, sizeof(*slice));
yading@10:     slice->wHorizontalPosition = 0;
yading@10:     slice->wVerticalPosition   = s->mb_y;
yading@10:     slice->dwSliceBitsInBuffer = 8 * size;
yading@10:     slice->dwSliceDataLocation = position;
yading@10:     slice->bStartCodeBitOffset = 0;
yading@10:     slice->bReservedBits       = 0;
yading@10:     slice->wMBbitOffset        = get_bits_count(&s->gb);
yading@10:     slice->wNumberMBsInSlice   = s->mb_width * s->mb_height; /* XXX We assume 1 slice */
yading@10:     slice->wQuantizerScaleCode = v->pq;
yading@10:     slice->wBadSliceChopping   = 0;
yading@10: }
yading@10: 
yading@10: static int commit_bitstream_and_slice_buffer(AVCodecContext *avctx,
yading@10:                                              DXVA2_DecodeBufferDesc *bs,
yading@10:                                              DXVA2_DecodeBufferDesc *sc)
yading@10: {
yading@10:     const VC1Context *v = avctx->priv_data;
yading@10:     struct dxva_context *ctx = avctx->hwaccel_context;
yading@10:     const MpegEncContext *s = &v->s;
yading@10:     struct dxva2_picture_context *ctx_pic = s->current_picture_ptr->hwaccel_picture_private;
yading@10: 
yading@10:     DXVA_SliceInfo *slice = &ctx_pic->si;
yading@10: 
yading@10:     static const uint8_t start_code[] = { 0, 0, 1, 0x0d };
yading@10:     const unsigned start_code_size = avctx->codec_id == AV_CODEC_ID_VC1 ? sizeof(start_code) : 0;
yading@10:     const unsigned slice_size = slice->dwSliceBitsInBuffer / 8;
yading@10:     const unsigned padding = 128 - ((start_code_size + slice_size) & 127);
yading@10:     const unsigned data_size = start_code_size + slice_size + padding;
yading@10: 
yading@10:     uint8_t  *dxva_data;
yading@10:     unsigned dxva_size;
yading@10:     int result;
yading@10: 
yading@10:     if (FAILED(IDirectXVideoDecoder_GetBuffer(ctx->decoder,
yading@10:                                               DXVA2_BitStreamDateBufferType,
yading@10:                                               (void **)&dxva_data, &dxva_size)))
yading@10:         return -1;
yading@10: 
yading@10:     result = data_size <= dxva_size ? 0 : -1;
yading@10:     if (!result) {
yading@10:         if (start_code_size > 0) {
yading@10:             memcpy(dxva_data, start_code, start_code_size);
yading@10:             if (v->second_field)
yading@10:                 dxva_data[3] = 0x0c;
yading@10:         }
yading@10:         memcpy(dxva_data + start_code_size,
yading@10:                ctx_pic->bitstream + slice->dwSliceDataLocation, slice_size);
yading@10:         if (padding > 0)
yading@10:             memset(dxva_data + start_code_size + slice_size, 0, padding);
yading@10:         slice->dwSliceBitsInBuffer = 8 * data_size;
yading@10:     }
yading@10:     if (FAILED(IDirectXVideoDecoder_ReleaseBuffer(ctx->decoder,
yading@10:                                                   DXVA2_BitStreamDateBufferType)))
yading@10:         return -1;
yading@10:     if (result)
yading@10:         return result;
yading@10: 
yading@10:     memset(bs, 0, sizeof(*bs));
yading@10:     bs->CompressedBufferType = DXVA2_BitStreamDateBufferType;
yading@10:     bs->DataSize             = data_size;
yading@10:     bs->NumMBsInBuffer       = s->mb_width * s->mb_height;
yading@10:     assert((bs->DataSize & 127) == 0);
yading@10: 
yading@10:     return ff_dxva2_commit_buffer(avctx, ctx, sc,
yading@10:                                   DXVA2_SliceControlBufferType,
yading@10:                                   slice, sizeof(*slice), bs->NumMBsInBuffer);
yading@10: }
yading@10: 
yading@10: static int dxva2_vc1_start_frame(AVCodecContext *avctx,
yading@10:                                  av_unused const uint8_t *buffer,
yading@10:                                  av_unused uint32_t size)
yading@10: {
yading@10:     const VC1Context *v = avctx->priv_data;
yading@10:     struct dxva_context *ctx = avctx->hwaccel_context;
yading@10:     struct dxva2_picture_context *ctx_pic = v->s.current_picture_ptr->hwaccel_picture_private;
yading@10: 
yading@10:     if (!ctx->decoder || !ctx->cfg || ctx->surface_count <= 0)
yading@10:         return -1;
yading@10:     assert(ctx_pic);
yading@10: 
yading@10:     fill_picture_parameters(avctx, ctx, v, &ctx_pic->pp);
yading@10: 
yading@10:     ctx_pic->bitstream_size = 0;
yading@10:     ctx_pic->bitstream      = NULL;
yading@10:     return 0;
yading@10: }
yading@10: 
yading@10: static int dxva2_vc1_decode_slice(AVCodecContext *avctx,
yading@10:                                   const uint8_t *buffer,
yading@10:                                   uint32_t size)
yading@10: {
yading@10:     const VC1Context *v = avctx->priv_data;
yading@10:     const Picture *current_picture = v->s.current_picture_ptr;
yading@10:     struct dxva2_picture_context *ctx_pic = current_picture->hwaccel_picture_private;
yading@10: 
yading@10:     if (ctx_pic->bitstream_size > 0)
yading@10:         return -1;
yading@10: 
yading@10:     if (avctx->codec_id == AV_CODEC_ID_VC1 &&
yading@10:         size >= 4 && IS_MARKER(AV_RB32(buffer))) {
yading@10:         buffer += 4;
yading@10:         size   -= 4;
yading@10:     }
yading@10: 
yading@10:     ctx_pic->bitstream_size = size;
yading@10:     ctx_pic->bitstream      = buffer;
yading@10: 
yading@10:     fill_slice(avctx, &ctx_pic->si, 0, size);
yading@10:     return 0;
yading@10: }
yading@10: 
yading@10: static int dxva2_vc1_end_frame(AVCodecContext *avctx)
yading@10: {
yading@10:     VC1Context *v = avctx->priv_data;
yading@10:     struct dxva2_picture_context *ctx_pic = v->s.current_picture_ptr->hwaccel_picture_private;
yading@10:     int ret;
yading@10: 
yading@10:     if (ctx_pic->bitstream_size <= 0)
yading@10:         return -1;
yading@10: 
yading@10:     ret = ff_dxva2_common_end_frame(avctx, v->s.current_picture_ptr,
yading@10:                                     &ctx_pic->pp, sizeof(ctx_pic->pp),
yading@10:                                     NULL, 0,
yading@10:                                     commit_bitstream_and_slice_buffer);
yading@10:     if (!ret)
yading@10:         ff_mpeg_draw_horiz_band(&v->s, 0, avctx->height);
yading@10:     return ret;
yading@10: }
yading@10: 
yading@10: #if CONFIG_WMV3_DXVA2_HWACCEL
yading@10: AVHWAccel ff_wmv3_dxva2_hwaccel = {
yading@10:     .name           = "wmv3_dxva2",
yading@10:     .type           = AVMEDIA_TYPE_VIDEO,
yading@10:     .id             = AV_CODEC_ID_WMV3,
yading@10:     .pix_fmt        = AV_PIX_FMT_DXVA2_VLD,
yading@10:     .start_frame    = dxva2_vc1_start_frame,
yading@10:     .decode_slice   = dxva2_vc1_decode_slice,
yading@10:     .end_frame      = dxva2_vc1_end_frame,
yading@10:     .priv_data_size = sizeof(struct dxva2_picture_context),
yading@10: };
yading@10: #endif
yading@10: 
yading@10: AVHWAccel ff_vc1_dxva2_hwaccel = {
yading@10:     .name           = "vc1_dxva2",
yading@10:     .type           = AVMEDIA_TYPE_VIDEO,
yading@10:     .id             = AV_CODEC_ID_VC1,
yading@10:     .pix_fmt        = AV_PIX_FMT_DXVA2_VLD,
yading@10:     .start_frame    = dxva2_vc1_start_frame,
yading@10:     .decode_slice   = dxva2_vc1_decode_slice,
yading@10:     .end_frame      = dxva2_vc1_end_frame,
yading@10:     .priv_data_size = sizeof(struct dxva2_picture_context),
yading@10: };