/* PQFx[0][u] = QFx[0][u] - QFc[0][u] * QPc // QPx */
                /* defined in subclause 7.4.3.3 of ISO/IEC 14496-2: 2001(E) */
                if (cur_qp == pred_qp) {
                    /* QPc == QPx */
                    for (i=1; i < SAMPLE_VIDEO_BLOCK_SIZE; i++) {
                        level = coef[i] - coef_bufrow[i];

                        if ((128 <= level) || (-128 >= level)) {
                            sum_err  = MPEG4_PRED_DISABLE;
                            i = SAMPLE_VIDEO_BLOCK_SIZE;
                        } else {
                            sum_err += ABS_MP4(coef[i]) - ABS_MP4(level);
                            pred_buf[blk_indx * 8 + i] = (Ipp16s)level;
                        }
                    }
                } else {
                    /* QPc != QPx */
                    for (i=1; i < SAMPLE_VIDEO_BLOCK_SIZE; i++) {
                        if (0 <= coef_bufrow[i]) {
                            level = coef[i] - (coef_bufrow[i] * pred_qp
                                + (cur_qp >> 1)) / cur_qp;
                        } else {
                            level = coef[i] - (coef_bufrow[i] * pred_qp
                                - (cur_qp >> 1)) / cur_qp;
                        }

                        if ((128 <= level) || (-128 >= level)) {
                            sum_err  = MPEG4_PRED_DISABLE;
                            i = SAMPLE_VIDEO_BLOCK_SIZE;
                        } else {
                            sum_err += ABS_MP4(coef[i]) - ABS_MP4(level);
                            pred_buf[blk_indx * 8 + i] = (Ipp16s)level;
                        }
                    }
                }
            }
        } else if (MPEG4_PRED_DISABLE != sum_err) {

            /* pred_dir == IPP_VIDEO_HORIZONTAL */
            if (0 > coef_bufrow[-8]) {
                pred_buf[blk_indx * 8] = 0;
            } else {
                pred_buf[blk_indx * 8] = 1;
                /* PQFx[v][0] = QFx[v][0] - QFa[v][0] * QPa // QPx */
                /* defined in subclause 7.4.3.3 of ISO/IEC 14496-2: 2001(E) */
                if (cur_qp == pred_qp) {
                    /* QPa == QPx */
                    for (i=1, j=8; i < SAMPLE_VIDEO_BLOCK_SIZE; i++, j+=8) {
                        level = coef[j] - coef_bufcol[i];

                        if ((128 <= level) || (-128 >= level)) {
                            sum_err  = MPEG4_PRED_DISABLE;
                            i = SAMPLE_VIDEO_BLOCK_SIZE;
                        } else {
                            sum_err += ABS_MP4(coef[j]) - ABS_MP4(level);
                            pred_buf[blk_indx * 8 + i] = (Ipp16s)level;
                        }
                    }
                } else {
                    /* QPa != QPx */
                    for (i=1, j=8; i < SAMPLE_VIDEO_BLOCK_SIZE; i++, j+=8) {
                        if (0 <= coef_bufcol[i]) {
                            level = coef[j] - (coef_bufcol[i] * pred_qp
                                + (cur_qp >> 1)) / cur_qp;
                        } else {
                            level = coef[j] - (coef_bufcol[i] * pred_qp
                                - (cur_qp >> 1)) / cur_qp;
                        }

                        if ((128 <= level) || (-128 >= level)) {
                            sum_err  = MPEG4_PRED_DISABLE;
                            i = SAMPLE_VIDEO_BLOCK_SIZE;
                        } else {
                            sum_err += ABS_MP4(coef[j]) - ABS_MP4(level);
                            pred_buf[blk_indx * 8 + i] = (Ipp16s)level;
                        }
                    }
                }
            }
        }

        /* 3. DC coefficient buffer update */
        dc_coef = (Ipp16s)(coef[0] * dc_scaler);
        /* The saturation range after inverse quantisation is listed in 
        // subclause 7.4.4.4 of ISO/IEC 14496-2:2001(E) */
        if (2047 < dc_coef) {
            dc_coef = 2047;
        }

        switch (blk_indx) {
        case Y_BLOCK2:
            tmp_coef         = coef_bufcol[8];
            coef_bufcol[8]   = coef_bufrow[-16];
            coef_bufrow[-16] = tmp_coef;
            
            coef_bufcol[0]   = coef_bufrow[0];
            coef_bufrow[0]   = dc_coef;
            break;
        case Y_BLOCK4:
            coef_bufcol[0]   = dc_coef;
            break;
        default:
            coef_bufcol[0]   = coef_bufrow[0];
            coef_bufrow[0]   = dc_coef;
            break;
        }

        /* 4. AC coefficient buffer update */
        for (i=1, j=8; i < SAMPLE_VIDEO_BLOCK_SIZE; i++, j+= 8) {
            coef_bufrow[i] = coef[i];
            coef_bufcol[i] = coef[j];
        }

        /* 5. DC prediction */
        if (IPP_VIDEO_HORIZONTAL == pred_dir) {
            /* PQFx[0][0] = QFx[0][0] - Fa[0][0] // dc_scaler
            // defined in subclause 7.4.3.2 of ISO/IEC 14496-2:2001(E) */
            coef[0] = (Ipp16s)(coef[0] - (dc_left + (dc_scaler >> 1))
                / dc_scaler);
        } else {
            /* PQFx[0][0] = QFx[0][0] - Fc[0][0] // dc_scaler
            // defined in subclause 7.4.3.2 of ISO/IEC 14496-2:2001(E) */
            coef[0] = (Ipp16s)(coef[0] - (dc_top  + (dc_scaler >> 1))
                / dc_scaler);
        }
    }
    
    /* 6. Determine ac_pred_flag based on the comparison result with the sum
    // of the predicted coefficients absolute values and the sum of original
    // coefficients absolute values */
    if ((MPEG4_PRED_DISABLE != sum_err) && (0 <= sum_err)) {
        
        *ac_pred_flag = 1;
        
        /* AC prediction coefficients update */
        for (blk_indx = Y_BLOCK1; blk_indx <= V_BLOCK; blk_indx++) {
            if (2 <= pred_buf[blk_indx * 8]) {
                pred_dir_buf[blk_indx] = IPP_VIDEO_VERTICAL;
            }
            else {
                pred_dir_buf[blk_indx] = IPP_VIDEO_HORIZONTAL;
            }

            if (3 == pred_buf[blk_indx * 8]) {
                for (i = 1; i < SAMPLE_VIDEO_BLOCK_SIZE; i++) {
                    src_dst_coeff[blk_indx*64 + i] = pred_buf[blk_indx*8 + i];
                }
            }
            else if (1 == pred_buf[blk_indx * 8]) {
                for (i = 1; i < SAMPLE_VIDEO_BLOCK_SIZE; i++) {
                    src_dst_coeff[blk_indx*64 + i*8] = pred_buf[blk_indx*8 + i];
                }
            }
        }
    } else {

        *ac_pred_flag = 0;
        
        for (blk_indx = Y_BLOCK1; blk_indx <= V_BLOCK; blk_indx++) {
            pred_dir_buf[blk_indx] = IPP_VIDEO_NONE;
        }
    }

    return SAMPLE_STATUS_NOERR;
}

/******************************************************************************
//
// Name:            encode_block_inter_mpeg4
//
// Description:     Quantize the DCT coefficients of the inter block and store
//                  them into buffer. Meanwhile, the texture residuals are
//                  reconstructed.
//
// Input Arguments:	
//		src_block	- Pointer to the pixels of current inter block.
//		cur_qp		- Quantization parameter of the macroblock which the
//					  current block belongs to.
//		q_matrix	- If the second inverse quantization method is used,  it is
//                    NULL; If the first inverse quantization method is used,
//                    it points to the quantization weighting coefficients
//                    buffer (for inter MB), whose first 64 elements are the
//                    quantization weighting matrix in Q0, the second 64
//                    elements are their reciprocals in Q21.
//
// Output Arguments:	
//      dst_coeff	- Pointer to the quantized DCT coefficients buffer.
//      rec_block	- Pointer to the reconstructed texture residuals.
//
// Returns:
//      SAMPLE_STATUS_NOERR		- If succeeds.
//      SAMPLE_STATUS_ERR       - If encoding fails.
//	
*********************************************************************************/

sample_status encode_block_inter_mpeg4(Ipp16s    *src_block,
                                       Ipp16s    *rec_block,
                                       Ipp16s    *dst_coeff,                                       
                                       Ipp8u     cur_qp,
                                       const int *q_matrix)
{
    Ipp8u   q_matrix_8u[SAMPLE_VIDEO_BLOCK_SQUARE_SIZE];
    Ipp16s  dct_coef_buf[SAMPLE_VIDEO_BLOCK_SQUARE_SIZE + 8];
    Ipp16s* dct_coef = (Ipp16s*)SAMPLE_ALIGN8(dct_coef_buf);
    int     i;
    sample_status     ret_code;

    /* 1. DCT transform */
    /* src_block, dct_coef must be 8 bytes aligned */
    ret_code = ippiDCT8x8Fwd_Video_16s_C1(src_block, dct_coef);
    if (ippStsNoErr != ret_code) {
        return SAMPLE_STATUS_ERR;
    }

    /* 2. Quantisation */
    ret_code = ippiQuantInter_MPEG4_16s_I(dct_coef, cur_qp, q_matrix);
    if (ippStsNoErr != ret_code) {
        return SAMPLE_STATUS_ERR;
    } else {
        memcpy((void*)dst_coeff, (void*)dct_coef, sizeof(Ipp16s) * 64);
    }

    /* 3. Inverse Quantisation */
    if (q_matrix) {
        for (i = 0; i < SAMPLE_VIDEO_BLOCK_SQUARE_SIZE; i++) {
            q_matrix_8u[i] = (Ipp8u)q_matrix[i];
        }
        ret_code = ippiQuantInvInter_MPEG4_16s_I(dct_coef, cur_qp, q_matrix_8u);
    } else {
        ret_code = ippiQuantInvInter_MPEG4_16s_I(dct_coef, cur_qp, NULL);
    }
    
    if (ippStsNoErr != ret_code) {
        return SAMPLE_STATUS_ERR;
    }

    /* 4. IDCT */
    /* dct_coef, rec_block must be 8 bytes aligned */
    ret_code = ippiDCT8x8Inv_Video_16s_C1(dct_coef, rec_block);
    if (ippStsNoErr != ret_code) {
        return SAMPLE_STATUS_ERR;
    }
    
    return SAMPLE_STATUS_NOERR;
}