umc_mpeg2_dec_mc.cpp

这是在PCA下的基于IPP库示例代码例子,在网上下了IPP的库之后,设置相关参数就可以编译该代码.

            }
            break;
        case IPPVC_MC_16X8:
            for(r = 0; r < pMacro->motion_vector_count; r++)
            {
                pMacro->vector_luma[4*r + 3] = (pMacro->vector[4*r + 3] >> 1) << 1;
                pMacro->vector_luma[4*r + 2] = pMacro->vector[4*r + 2] >> 1;


                pMacro->vector_chroma[4*r + 3] = (pMacro->vector[4*r + 3] >> 1) << 1;//down
                pMacro->vector_chroma[4*r + 2] = pMacro->vector[4*r + 2] >> 2;//right


                offs = (pMacro->vector_chroma[4*r + 3] + pMacro->row_c + pMacro->motion_vertical_field_select[2*r+1])*pitch_uv +
                            pMacro->vector_chroma[4*r + 2] + pMacro->col_c;

                flag1 =  ((pMacro->vector[4*r + 2]  & 1) << 3) | ((pMacro->vector[4*r + 3]  & 1) << 2);
                flag2 =  ((pMacro->vector[4*r + 2]  & 2) << 2) | ((pMacro->vector[4*r + 3]  & 1) << 2);

            i = ippiMC16x8_8u_C1(    ref_Y_data +
                                    (pMacro->vector_luma[4*r + 3]  + pMacro->row_l  + 16*r + pMacro->motion_vertical_field_select[2*r+1])*pitch_y +
                                    pMacro->vector_luma[4*r + 2] + pMacro->col_l,
                                    pitch_y_2,
                                    diff+128*r,32,
                                    video->frame_buffer.frame_p_c_n[video->frame_buffer.curr_index].Y_comp_data + pMacro->offset_l  + pMacro->shift_y[r] + pMacro->odd_pitch_y,
                                    pitch_y_2, flag1, 0);
            VM_ASSERT(i == ippStsOk);
            i = ippiMC8x8_8u_C1(  ref_U_data + offs  + 8 * r * pitch_uv,
                                    pitch_uv_2,
                                    diff+256+64*r,16,
                                    video->frame_buffer.frame_p_c_n[video->frame_buffer.curr_index].U_comp_data + pMacro->offset_c + pMacro->shift_uv[r] + pMacro->odd_pitch_uv,
                                    pitch_uv_2, flag2, 0);
            VM_ASSERT(i == ippStsOk);
            i = ippiMC8x8_8u_C1(  ref_V_data + offs  + 8 * r * pitch_uv,
                                    pitch_uv_2,
                                    diff+384+64*r,16,
                                    video->frame_buffer.frame_p_c_n[video->frame_buffer.curr_index].V_comp_data + pMacro->offset_c + pMacro->shift_uv[r] + pMacro->odd_pitch_uv,
                                    pitch_uv_2, flag2, 0);
            VM_ASSERT(i == ippStsOk);
        }//for(r)
        }
    }

}//mc_mp2_422

void MPEG2VideoDecoderBase::mc_mp2_444(short* /*diff*/, int /*threadID*/)
{

}//mc_mp2_444

void MPEG2VideoDecoderBase::mc_mp2_420b_frame_skip(short* diff, int threadID)
{
  IppVideoContext*   video = &Video[threadID];
  sMacroblock       *pMacro = &Video[threadID].macroblock;
  sSlice            *pSlice = &Video[threadID].slice;
  Ipp8u *ref_Y_data1, *ref_U_data1;
  Ipp8u *ref_Y_data2, *ref_U_data2;
  Ipp8u *ref_Y_data3, *ref_U_data3;
  Ipp8u *ref_Y_data4, *ref_U_data4;
  Ipp8u *cur_Y_data, *cur_U_data;
  // Used when U & V not merged
  Ipp8u *ref_V_data1 = 0, *ref_V_data2 = 0, *ref_V_data3 = 0, *ref_V_data4 = 0, *cur_V_data = 0;

  unsigned int        flag1, flag2, flag3, flag4;
  int                 offs_uv1, offs_uv2, offs_y1, offs_y2;
  unsigned int        pitch_Y, pitch_UV;
  unsigned int        pitch_Y_2, pitch_UV_2;
  int i;

  ref_Y_data1 = video->frame_buffer.frame_p_c_n[video->frame_buffer.prev_index].Y_comp_data;
  ref_U_data1 = video->frame_buffer.frame_p_c_n[video->frame_buffer.prev_index].U_comp_data;
  ref_Y_data2 = video->frame_buffer.frame_p_c_n[video->frame_buffer.next_index].Y_comp_data;
  ref_U_data2 = video->frame_buffer.frame_p_c_n[video->frame_buffer.next_index].U_comp_data;
  ref_Y_data3 = video->frame_buffer.frame_p_c_n[video->frame_buffer.prev_index].Y_comp_data;
  ref_U_data3 = video->frame_buffer.frame_p_c_n[video->frame_buffer.prev_index].U_comp_data;
  ref_Y_data4 = video->frame_buffer.frame_p_c_n[video->frame_buffer.next_index].Y_comp_data;
  ref_U_data4 = video->frame_buffer.frame_p_c_n[video->frame_buffer.next_index].U_comp_data;
  cur_Y_data = video->frame_buffer.frame_p_c_n[video->frame_buffer.curr_index].Y_comp_data;
  cur_U_data = video->frame_buffer.frame_p_c_n[video->frame_buffer.curr_index].U_comp_data;

  cur_V_data = video->frame_buffer.frame_p_c_n[video->frame_buffer.curr_index].V_comp_data;
  ref_V_data1 = video->frame_buffer.frame_p_c_n[video->frame_buffer.prev_index].V_comp_data;
  ref_V_data2 = video->frame_buffer.frame_p_c_n[video->frame_buffer.next_index].V_comp_data;
  ref_V_data3 = video->frame_buffer.frame_p_c_n[video->frame_buffer.prev_index].V_comp_data;
  ref_V_data4 = video->frame_buffer.frame_p_c_n[video->frame_buffer.next_index].V_comp_data;

  pitch_Y   =  video->frame_buffer.Y_comp_pitch;
  pitch_UV  =  video->frame_buffer.U_comp_pitch;
  pitch_Y_2   =  video->frame_buffer.Y_comp_pitch << 1;
  pitch_UV_2  =  video->frame_buffer.U_comp_pitch << 1;

  cur_Y_data += pMacro->offset_l;
  cur_U_data += pMacro->offset_c;
  cur_V_data += pMacro->offset_c;

  if(pMacro->prediction_type == IPPVC_MC_FRAME)
  {
    CALC_OFFSETS_FRAME(offs_y1, offs_uv1, flag1, flag2, pMacro->vector[0], pMacro->vector[1], HP_FLAG_MC)
    CALC_OFFSETS_FRAME(offs_y2, offs_uv2, flag3, flag4, pMacro->vector[2], pMacro->vector[3], HP_FLAG_MC)

    for (i = 0; i < pSlice->mb_address_increment - 1; i++) {
      FUNC_MCB(16, 16, Frame, ref_Y_data1 + offs_y1, pitch_Y, flag1,
        ref_Y_data2 + offs_y2, pitch_Y, flag3,
        diff, 32,
        cur_Y_data,
        pitch_Y, 0);

      FUNC_MCB(8, 8, Frame, ref_U_data1 + offs_uv1, pitch_UV, flag2,
          ref_U_data2 + offs_uv2, pitch_UV, flag4,
          diff + 256, 16,
          cur_U_data,
          pitch_UV, 0);

      FUNC_MCB(8, 8, Frame, ref_V_data1 + offs_uv1, pitch_UV, flag2,
          ref_V_data2 + offs_uv2, pitch_UV, flag4,
          diff + 320, 16,
          cur_V_data,
          pitch_UV, 0);
      cur_Y_data += 16;
      cur_U_data += 8;
      cur_V_data += 8;
      offs_y1 += 16;
      offs_uv1 += 8;
      offs_y2 += 16;
      offs_uv2 += 8;
    }
  }
  else
  {
    CALC_OFFSETS_FIELD(offs_y1, offs_uv1, flag1, flag2, pMacro->vector[0], pMacro->vector[1], pMacro->motion_vertical_field_select[0], HP_FLAG_MC)
    CALC_OFFSETS_FIELD(offs_y2, offs_uv2, flag3, flag4, pMacro->vector[2], pMacro->vector[3], pMacro->motion_vertical_field_select[1], HP_FLAG_MC)

    for (i = 0; i < pSlice->mb_address_increment - 1; i++) {
      FUNC_MCB(16, 8, FieldDCT, ref_Y_data1 + offs_y1, pitch_Y_2, flag1,
        ref_Y_data2 + offs_y2, pitch_Y_2, flag3,
        diff, 64,
        cur_Y_data, pitch_Y_2,  0);

      FUNC_MCB(8, 4, FieldDCT, ref_U_data1 + offs_uv1, pitch_UV_2, flag2,
          ref_U_data2 + offs_uv2, pitch_UV_2, flag4,
          diff + 256, 32,
          cur_U_data, pitch_UV_2, 0);

      FUNC_MCB(8, 4, FieldDCT, ref_V_data1 + offs_uv1, pitch_UV_2, flag2,
          ref_V_data2 + offs_uv2, pitch_UV_2, flag4,
          diff + 320, 32,
          cur_V_data, pitch_UV_2,  0);
      cur_Y_data += 16;
      cur_U_data += 8;
      cur_V_data += 8;
      offs_y1 += 16;
      offs_uv1 += 8;
      offs_y2 += 16;
      offs_uv2 += 8;
    }

    if( pMacro->motion_vector_count == 2)
    {
      CALC_OFFSETS_FIELD(offs_y1, offs_uv1, flag1, flag2, pMacro->vector[4], pMacro->vector[5], pMacro->motion_vertical_field_select[2], HP_FLAG_MC)
      CALC_OFFSETS_FIELD(offs_y2, offs_uv2, flag3, flag4, pMacro->vector[6], pMacro->vector[7], pMacro->motion_vertical_field_select[3], HP_FLAG_MC)

      for (i = 0; i < pSlice->mb_address_increment - 1; i++) {
        FUNC_MCB(16, 8, FieldDCT, ref_Y_data1 + offs_y1, pitch_Y_2, flag1,
          ref_Y_data2 + offs_y2, pitch_Y_2, flag3,
          diff + 16, 64,
          cur_Y_data  + pitch_Y, pitch_Y_2,  0);

        FUNC_MCB(8, 4, FieldDCT, ref_U_data1 + offs_uv1, pitch_UV_2, flag2,
            ref_U_data2 + offs_uv2, pitch_UV_2, flag4,
            diff + 256 + 8, 32,
            cur_U_data  + pitch_UV , pitch_UV_2, 0);

        FUNC_MCB(8, 4, FieldDCT, ref_V_data1 + offs_uv1, pitch_UV_2, flag2,
            ref_V_data2 + offs_uv2, pitch_UV_2, flag4,
            diff + 320 + 8, 32,
            cur_V_data  + pitch_UV, pitch_UV_2,  0);
        cur_Y_data += 16;
        cur_U_data += 8;
        cur_V_data += 8;
        offs_y1 += 16;
        offs_uv1 += 8;
        offs_y2 += 16;
        offs_uv2 += 8;
      }
    }
  }
}

void MPEG2VideoDecoderBase::mc_mp2_420_frame_skip(int threadID)
{
  IppVideoContext *video = &Video[threadID];
  sMacroblock *pMacro = &Video[threadID].macroblock;
  int ref_index = (pMacro->macroblock_motion_backward) ? video->frame_buffer.next_index : video->frame_buffer.prev_index;
  Ipp8u *ref_Y_data = video->frame_buffer.frame_p_c_n[ref_index].Y_comp_data;
  Ipp8u *ref_U_data = video->frame_buffer.frame_p_c_n[ref_index].U_comp_data;
  Ipp8u *ref_V_data = video->frame_buffer.frame_p_c_n[ref_index].V_comp_data;
  Ipp8u *cur_Y_data = video->frame_buffer.frame_p_c_n[video->frame_buffer.curr_index].Y_comp_data;
  Ipp8u *cur_U_data = video->frame_buffer.frame_p_c_n[video->frame_buffer.curr_index].U_comp_data;
  Ipp8u *cur_V_data = video->frame_buffer.frame_p_c_n[video->frame_buffer.curr_index].V_comp_data;
  unsigned int pitch_Y  = video->frame_buffer.Y_comp_pitch;
  unsigned int pitch_UV = video->frame_buffer.U_comp_pitch;
  unsigned int pitch_Y_2  = video->frame_buffer.Y_comp_pitch << 1;
  unsigned int pitch_UV_2 = video->frame_buffer.U_comp_pitch << 1;
  unsigned int flag1, flag2;
  int offs_uv1, offs_y1;
  short *vector = pMacro->vector;
  sSlice *pSlice = &Video[threadID].slice;
  IppiSize roi = {16*(pSlice->mb_address_increment - 1), 16};
  int i;

  cur_Y_data += pMacro->offset_l;
  cur_U_data += pMacro->offset_c;
  cur_V_data += pMacro->offset_c;

  if (pMacro->macroblock_motion_backward) vector += 2;

  if (PictureHeader.picture_coding_type == P_FRAME) {
    if(PictureHeader.picture_structure != FRAME_PICTURE) {
      if(PictureHeader.picture_structure == BOTTOM_FIELD) {
        cur_Y_data += pitch_Y;
        cur_U_data += pitch_UV;
        cur_V_data += pitch_UV;
      }
      pitch_Y <<= 1;
      pitch_UV <<= 1;
    }
    ippiCopy_8u_C1R( ref_Y_data, pitch_Y, cur_Y_data, pitch_Y, roi );
    roi.height >>= 1;
    roi.width  >>= 1;
    ippiCopy_8u_C1R(ref_U_data, pitch_UV, cur_U_data, pitch_UV, roi);
    ippiCopy_8u_C1R(ref_V_data, pitch_UV, cur_V_data, pitch_UV, roi);
    return;
  }

  if(pMacro->prediction_type == IPPVC_MC_FRAME) {
    CALC_OFFSETS_FRAME(offs_y1, offs_uv1, flag1, flag2, vector[0], vector[1], HP_FLAG_CP)
    for (i = 0; i < pSlice->mb_address_increment - 1; i++) {
      MCZERO_FRAME(ref_Y_data + offs_y1, ref_U_data + offs_uv1, ref_V_data + offs_uv1, flag1, flag2)
      cur_Y_data += 16;
      cur_U_data += 8;
      cur_V_data += 8;
      offs_y1 += 16;
      offs_uv1 += 8;
    }
  }
  else
  {
    int *motion_vertical_field_select = pMacro->motion_vertical_field_select;
    Ipp8u* cur_Y_data_save = cur_Y_data;
    Ipp8u* cur_U_data_save = cur_U_data;
    Ipp8u* cur_V_data_save = cur_V_data;
    if (pMacro->macroblock_motion_backward) motion_vertical_field_select++;

    CALC_OFFSETS_FIELD(offs_y1, offs_uv1, flag1, flag2, vector[0], vector[1], motion_vertical_field_select[0], HP_FLAG_CP)
    for (i = 0; i < pSlice->mb_address_increment - 1; i++) {
      MCZERO_FIELD0(ref_Y_data + offs_y1, ref_U_data + offs_uv1, ref_V_data + offs_uv1, flag1, flag2)
      cur_Y_data += 16;
      cur_U_data += 8;
      cur_V_data += 8;
      offs_y1 += 16;
      offs_uv1 += 8;
    }

    if(pMacro->motion_vector_count == 2) {
      cur_Y_data = cur_Y_data_save;
      cur_U_data = cur_U_data_save;
      cur_V_data = cur_V_data_save;
      CALC_OFFSETS_FIELD(offs_y1, offs_uv1, flag1, flag2, vector[4], vector[5], motion_vertical_field_select[2], HP_FLAG_CP)
      for (i = 0; i < pSlice->mb_address_increment - 1; i++) {
        MCZERO_FIELD1(ref_Y_data + offs_y1, ref_U_data + offs_uv1, ref_V_data + offs_uv1, flag1, flag2)
        cur_Y_data += 16;
        cur_U_data += 8;
        cur_V_data += 8;
        offs_y1 += 16;
        offs_uv1 += 8;
      }
    }
  }
}