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.

    int pitch_y        = video->frame_buffer.Y_comp_pitch;
    int pitch_uv    = video->frame_buffer.U_comp_pitch;
    int pitch_y_2    = pitch_y*2;
    int pitch_uv_2    = pitch_uv*2;
    int ptc_uv, ptc_y;
    int flag1, flag2, flag3, flag4;
    int offs;
    int offs1;
    int offs2;

    ptc_uv = pitch_uv  + pitch_uv*(!pMacro->prediction_type);
    ptc_y  = pitch_y  + pitch_y*(!pMacro->prediction_type);

    //bi_direction
    if(pMacro->macroblock_motion_forward && pMacro->macroblock_motion_backward)
    {
        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_V_data1 = video->frame_buffer.frame_p_c_n[video->frame_buffer.prev_index].V_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_V_data2 = video->frame_buffer.frame_p_c_n[video->frame_buffer.next_index].V_comp_data;

        switch(pMacro->prediction_type)
        {
        case IPPVC_MC_FRAME:
            pMacro->vector_luma[1] = pMacro->vector[1] >> 1;
            pMacro->vector_luma[0] = pMacro->vector[0] >> 1;
            pMacro->vector_luma[3] = pMacro->vector[3] >> 1;
            pMacro->vector_luma[2] = pMacro->vector[2] >> 1;

            pMacro->vector_chroma[1] = pMacro->vector[1] >> 1;//down
            pMacro->vector_chroma[0] = pMacro->vector[0] >> 2;//right
            pMacro->vector_chroma[3] = pMacro->vector[3] >> 1;//down
            pMacro->vector_chroma[2] = pMacro->vector[2] >> 2;//right

            offs1 = (pMacro->vector_chroma[1]  + pMacro->row_c + pMacro->motion_vertical_field_select[0])*pitch_uv +
                    pMacro->vector_chroma[0] + pMacro->col_c;
            offs2 = (pMacro->vector_chroma[3]  + pMacro->row_c + pMacro->motion_vertical_field_select[1])*pitch_uv +
                        pMacro->vector_chroma[2] + pMacro->col_c;

            flag1 =  ((pMacro->vector[0]  & 1) << 3) | ((pMacro->vector[1]  & 1) << 2);
            flag2 =  ((pMacro->vector[0]  & 2) << 2) | ((pMacro->vector[1]  & 1) << 2);
            flag3 =  ((pMacro->vector[2]  & 1) << 3) | ((pMacro->vector[3]  & 1) << 2);
            flag4 =  ((pMacro->vector[2]  & 2) << 2) | ((pMacro->vector[3]  & 1) << 2);

            i = ippiMC16x16B_8u_C1(  ref_Y_data1 +
                                    (pMacro->vector_luma[1]  + pMacro->row_l)*pitch_y +
                                    pMacro->vector_luma[0] + pMacro->col_l,
                                    ptc_y,
                                    flag1,
                                    ref_Y_data2 +
                                    (pMacro->vector_luma[3]  + pMacro->row_l)*pitch_y +
                                    pMacro->vector_luma[2] + pMacro->col_l,
                                    ptc_y,
                                    flag3,
                                    diff,32,
                                    video->frame_buffer.frame_p_c_n[video->frame_buffer.curr_index].Y_comp_data + pMacro->offset_l,
                                    ptc_y, 0);
            VM_ASSERT(i == ippStsOk);
            i = ippiMC8x16B_8u_C1 (    ref_U_data1 + offs1, ptc_uv, flag2,
                                    ref_U_data2 + offs2, ptc_uv, flag4,
                                    diff+256,16,
                                    video->frame_buffer.frame_p_c_n[video->frame_buffer.curr_index].U_comp_data + pMacro->offset_c,
                                    ptc_uv, 0);
            VM_ASSERT(i == ippStsOk);
            i = ippiMC8x16B_8u_C1 (    ref_V_data1 + offs1, ptc_uv,flag2,
                                    ref_V_data2 + offs2, ptc_uv,flag4,
                                    diff+384,16,
                                    video->frame_buffer.frame_p_c_n[video->frame_buffer.curr_index].V_comp_data + pMacro->offset_c,
                                    ptc_uv, 0);
            VM_ASSERT(i == ippStsOk);
           break;
       case IPPVC_MC_FIELD:
           if(PictureHeader.picture_structure == FRAME_PICTURE)
           {
            for(r = 0; r < pMacro->motion_vector_count; r++)
            {
                pMacro->vector_luma[4*r + 1] = (pMacro->vector[4*r + 1] >> 1) << 1;
                pMacro->vector_luma[4*r] = pMacro->vector[4*r] >> 1;
                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 + 1] = pMacro->vector_luma[4*r + 1];//down
                pMacro->vector_chroma[4*r] = pMacro->vector[4*r] >> 2;//right
                 pMacro->vector_chroma[4*r + 3] = pMacro->vector_luma[4*r + 3];//down
                pMacro->vector_chroma[4*r + 2] = pMacro->vector[4*r + 2] >> 2;//right

                offs1 = (pMacro->vector_chroma[4*r + 1]  + pMacro->row_c + pMacro->motion_vertical_field_select[2*r])*pitch_uv +
                        pMacro->vector_chroma[4*r] + pMacro->col_c;
                offs2 = (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]  & 1) << 3) | ((pMacro->vector[4*r + 1]  & 1) << 2);
                flag2 =  ((pMacro->vector[4*r]  & 2) << 2) | ((pMacro->vector[4*r + 1]  & 1) << 2);
                flag3 =  ((pMacro->vector[4*r + 2]  & 1) << 3) | ((pMacro->vector[4*r + 3]  & 1) << 2);
                flag4 =  ((pMacro->vector[4*r + 2]  & 2) << 2) | ((pMacro->vector[4*r + 3]  & 1) << 2);

                i = ippiMC16x8B_8u_C1(  ref_Y_data1 +
                                        (pMacro->vector_luma[4*r + 1]  + pMacro->row_l + pMacro->motion_vertical_field_select[2*r])*pitch_y +
                                        pMacro->vector_luma[4*r] + pMacro->col_l,
                                        ptc_y,
                                        flag1,
                                        ref_Y_data2 +
                                        (pMacro->vector_luma[4*r + 3]  + pMacro->row_l + pMacro->motion_vertical_field_select[2*r+1])*pitch_y +
                                        pMacro->vector_luma[4*r + 2] + pMacro->col_l,
                                        ptc_y,
                                        flag3,
                                        diff+16*r,64,
                                        video->frame_buffer.frame_p_c_n[video->frame_buffer.curr_index].Y_comp_data + pMacro->offset_l  + pMacro->shift_y[r],
                                        ptc_y, 0);
                VM_ASSERT(i == ippStsOk);
                i = ippiMC8x8B_8u_C1(    ref_U_data1 + offs1, ptc_uv, flag2,
                                        ref_U_data2 + offs2, ptc_uv, flag4,
                                        diff+256+8*r,32,
                                        video->frame_buffer.frame_p_c_n[video->frame_buffer.curr_index].U_comp_data + pMacro->offset_c  + pMacro->shift_uv[r],
                                        ptc_uv, 0);
                VM_ASSERT(i == ippStsOk);
                i = ippiMC8x8B_8u_C1(    ref_V_data1 + offs1, ptc_uv,flag2,
                                        ref_V_data2 + offs2, ptc_uv,flag4,
                                        diff+384+8*r,32,
                                        video->frame_buffer.frame_p_c_n[video->frame_buffer.curr_index].V_comp_data + pMacro->offset_c  + pMacro->shift_uv[r],
                                        ptc_uv, 0);
                VM_ASSERT(i == ippStsOk);
            }//for(r)
           }
           else
           {
                pMacro->vector_luma[1] = (pMacro->vector[1] >> 1) << 1;
                pMacro->vector_luma[0] = pMacro->vector[0] >> 1;
                pMacro->vector_luma[3] = (pMacro->vector[3] >> 1) << 1;
                pMacro->vector_luma[2] = pMacro->vector[2] >> 1;

                pMacro->vector_chroma[1] = (pMacro->vector[1] >> 1) << 1;//down
                pMacro->vector_chroma[0] = pMacro->vector[0] >> 2;//right
                pMacro->vector_chroma[3] = (pMacro->vector[3] >> 1) << 1;//down
                pMacro->vector_chroma[2] = pMacro->vector[2] >> 2;//right

                offs1 = (pMacro->vector_chroma[1]  + pMacro->row_c + pMacro->motion_vertical_field_select[0])*pitch_uv +
                        pMacro->vector_chroma[0] + pMacro->col_c;
                offs2 = (pMacro->vector_chroma[3]  + pMacro->row_c + pMacro->motion_vertical_field_select[1])*pitch_uv +
                            pMacro->vector_chroma[2] + pMacro->col_c;

                flag1 =  ((pMacro->vector[0]  & 1) << 3) | ((pMacro->vector[1]  & 1) << 2);
                flag2 =  ((pMacro->vector[0]  & 2) << 2) | ((pMacro->vector[1]  & 1) << 2);
                flag3 =  ((pMacro->vector[2]  & 1) << 3) | ((pMacro->vector[3]  & 1) << 2);
                flag4 =  ((pMacro->vector[2]  & 2) << 2) | ((pMacro->vector[3]  & 1) << 2);

                i = ippiMC16x16B_8u_C1(  ref_Y_data1 +
                                        (pMacro->vector_luma[1]  + pMacro->row_l + pMacro->motion_vertical_field_select[0])*pitch_y +
                                        pMacro->vector_luma[0] + pMacro->col_l,
                                        ptc_y,
                                        flag1,
                                        ref_Y_data2 +
                                        (pMacro->vector_luma[3]  + pMacro->row_l + pMacro->motion_vertical_field_select[1])*pitch_y +
                                        pMacro->vector_luma[2] + pMacro->col_l,
                                        ptc_y,
                                        flag3,
                                        diff,32,
                                        video->frame_buffer.frame_p_c_n[video->frame_buffer.curr_index].Y_comp_data + pMacro->offset_l
                                         + pMacro->odd_pitch_y,
                                        ptc_y, 0);
                VM_ASSERT(i == ippStsOk);
                i = ippiMC8x16B_8u_C1(    ref_U_data1 + offs1, ptc_uv, flag2,
                                        ref_U_data2 + offs2, ptc_uv, flag4,
                                        diff+256,16,
                                        video->frame_buffer.frame_p_c_n[video->frame_buffer.curr_index].U_comp_data + pMacro->offset_c
                                         + pMacro->odd_pitch_uv,
                                        ptc_uv, 0);
                VM_ASSERT(i == ippStsOk);
                i = ippiMC8x16B_8u_C1(    ref_V_data1 + offs1, ptc_uv,flag2,
                                        ref_V_data2 + offs2, ptc_uv,flag4,
                                        diff+384,16,
                                        video->frame_buffer.frame_p_c_n[video->frame_buffer.curr_index].V_comp_data + pMacro->offset_c
                                         + pMacro->odd_pitch_uv,
                                        ptc_uv, 0);
                VM_ASSERT(i == ippStsOk);
           }

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

                offs1 = (pMacro->vector_chroma[4*r + 1]  + pMacro->row_c + pMacro->motion_vertical_field_select[2*r])*pitch_uv +
                        pMacro->vector_chroma[4*r] + pMacro->col_c;
                offs2 = (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]  & 1) << 3) | ((pMacro->vector[4*r + 1]  & 1) << 2);
                flag2 =  ((pMacro->vector[4*r]  & 2) << 2) | ((pMacro->vector[4*r + 1]  & 1) << 2);
                flag3 =  ((pMacro->vector[4*r + 2]  & 1) << 3) | ((pMacro->vector[4*r + 3]  & 1) << 2);
                flag4 =  ((pMacro->vector[4*r + 2]  & 2) << 2) | ((pMacro->vector[4*r + 3]  & 1) << 2);

                i = ippiMC16x8B_8u_C1(  ref_Y_data1 +
                                        (pMacro->vector_luma[4*r + 1]  + pMacro->row_l + 16 * r + pMacro->motion_vertical_field_select[2*r])*pitch_y +
                                        pMacro->vector_luma[4*r] + pMacro->col_l,
                                        pitch_y_2,
                                        flag1,
                                        ref_Y_data2 +
                                        (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,
                                        flag3,
                                        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, 0);
                VM_ASSERT(i == ippStsOk);
                i = ippiMC8x8B_8u_C1(    ref_U_data1 + offs1  + 8 * r * pitch_uv, pitch_uv_2, flag2,
                                        ref_U_data2 + offs2  + 8 * r * pitch_uv, pitch_uv_2, flag4,
                                        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, 0);
                VM_ASSERT(i == ippStsOk);
                i = ippiMC8x8B_8u_C1(    ref_V_data1 + offs1  + 8 * r * pitch_uv, pitch_uv_2,flag2,
                                        ref_V_data2 + offs2  + 8 * r * pitch_uv, pitch_uv_2,flag4,
                                        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, 0);
                VM_ASSERT(i == ippStsOk);
            }//for(r)
           break;
        }
    }
    else   if(pMacro->macroblock_motion_forward
              || PictureHeader.picture_coding_type == P_FRAME)
    {
      switch(pMacro->prediction_type)
      {
      case IPPVC_MC_FRAME:
            ref_Y_data = video->frame_buffer.frame_p_c_n[video->frame_buffer.prev_index].Y_comp_data;
            ref_U_data = video->frame_buffer.frame_p_c_n[video->frame_buffer.prev_index].U_comp_data;
            ref_V_data = video->frame_buffer.frame_p_c_n[video->frame_buffer.prev_index].V_comp_data;

            pMacro->vector_luma[1] = pMacro->vector[1] >> 1;
            pMacro->vector_luma[0] = pMacro->vector[0] >> 1;

            pMacro->vector_chroma[1] = pMacro->vector[1] >> 1;//down
            pMacro->vector_chroma[0] = pMacro->vector[0] >> 2;//right

            offs = (pMacro->vector_chroma[1]  + pMacro->row_c + pMacro->motion_vertical_field_select[0])* pitch_uv +
                        pMacro->vector_chroma[0] + pMacro->col_c;

            flag1 =  ((pMacro->vector[0]  & 1) << 3) | ((pMacro->vector[1]  & 1) << 2);
            flag2 =  ((pMacro->vector[0]  & 2) << 2) | ((pMacro->vector[1]  & 1) << 2);

            i = ippiMC16x16_8u_C1(  ref_Y_data +
                                    (pMacro->vector_luma[1]  + pMacro->row_l) * pitch_y +
                                    pMacro->vector_luma[0] + pMacro->col_l,
                                    ptc_y,
                                    diff,32,
                                    video->frame_buffer.frame_p_c_n[video->frame_buffer.curr_index].Y_comp_data + pMacro->offset_l,
                                    ptc_y, flag1, 0);
            VM_ASSERT(i == ippStsOk);

            i = ippiMC8x16_8u_C1(  ref_U_data + offs, ptc_uv,
                                    diff+256,16,
                                    video->frame_buffer.frame_p_c_n[video->frame_buffer.curr_index].U_comp_data + pMacro->offset_c,
                                    ptc_uv, flag2, 0);
            VM_ASSERT(i == ippStsOk);
            i = ippiMC8x16_8u_C1(  ref_V_data + offs, ptc_uv,
                                    diff+384,16,
                                    video->frame_buffer.frame_p_c_n[video->frame_buffer.curr_index].V_comp_data + pMacro->offset_c,
                                    ptc_uv, flag2, 0);
            VM_ASSERT(i == ippStsOk);
            break;
      case IPPVC_MC_FIELD:
            if(PictureHeader.picture_structure == FRAME_PICTURE)
            {
                ref_Y_data = video->frame_buffer.frame_p_c_n[video->frame_buffer.prev_index].Y_comp_data;