mpegvideo_altivec.c

Trolltech公司发布的图形界面操作系统。可在qt-embedded-2.3.7平台上编译为嵌入式图形界面操作系统。

                SWAP(row5, alt5);
                SWAP(row6, alt6);
                SWAP(row7, alt7);
            }

            if (whichPass == 1)
            {
                // transpose the data for the second pass
                 
                // First, block transpose the upper right with lower left.
                SWAP(row4, alt0);
                SWAP(row5, alt1);
                SWAP(row6, alt2);
                SWAP(row7, alt3);

                // Now, transpose each block of four
                TRANSPOSE4(row0, row1, row2, row3);
                TRANSPOSE4(row4, row5, row6, row7);
                TRANSPOSE4(alt0, alt1, alt2, alt3);
                TRANSPOSE4(alt4, alt5, alt6, alt7);
            }
        }
    }

    // used after quantise step
    int oldBaseValue = 0;

    // perform the quantise step, using the floating point data
    // still in the row/alt registers
    {
        const int* biasAddr;
        const vector signed int* qmat;
        vector float bias, negBias;

        if (s->mb_intra)
        {
            vector signed int baseVector;

            // We must cache element 0 in the intra case
            // (it needs special handling).
            baseVector = vec_cts(vec_splat(row0, 0), 0);
            vec_ste(baseVector, 0, &oldBaseValue);

            qmat = (vector signed int*)s->q_intra_matrix[qscale];
            biasAddr = &(s->intra_quant_bias);
        }
        else
        {
            qmat = (vector signed int*)s->q_inter_matrix[qscale];
            biasAddr = &(s->inter_quant_bias);
        }

        // Load the bias vector (We add 0.5 to the bias so that we're
				// rounding when we convert to int, instead of flooring.)
        {
            vector signed int biasInt;
            const vector float negOneFloat = (vector float)(FOUR_INSTANCES(-1.0f));
            LOAD4(biasInt, biasAddr);
            bias = vec_ctf(biasInt, QUANT_BIAS_SHIFT);
            negBias = vec_madd(bias, negOneFloat, zero);
        }

        {
            vector float q0, q1, q2, q3, q4, q5, q6, q7;

            q0 = vec_ctf(qmat[0], QMAT_SHIFT);
            q1 = vec_ctf(qmat[2], QMAT_SHIFT);
            q2 = vec_ctf(qmat[4], QMAT_SHIFT);
            q3 = vec_ctf(qmat[6], QMAT_SHIFT);
            q4 = vec_ctf(qmat[8], QMAT_SHIFT);
            q5 = vec_ctf(qmat[10], QMAT_SHIFT);
            q6 = vec_ctf(qmat[12], QMAT_SHIFT);
            q7 = vec_ctf(qmat[14], QMAT_SHIFT);

            row0 = vec_sel(vec_madd(row0, q0, negBias), vec_madd(row0, q0, bias),
                    vec_cmpgt(row0, zero));
            row1 = vec_sel(vec_madd(row1, q1, negBias), vec_madd(row1, q1, bias),
                    vec_cmpgt(row1, zero));
            row2 = vec_sel(vec_madd(row2, q2, negBias), vec_madd(row2, q2, bias),
                    vec_cmpgt(row2, zero));
            row3 = vec_sel(vec_madd(row3, q3, negBias), vec_madd(row3, q3, bias),
                    vec_cmpgt(row3, zero));
            row4 = vec_sel(vec_madd(row4, q4, negBias), vec_madd(row4, q4, bias),
                    vec_cmpgt(row4, zero));
            row5 = vec_sel(vec_madd(row5, q5, negBias), vec_madd(row5, q5, bias),
                    vec_cmpgt(row5, zero));
            row6 = vec_sel(vec_madd(row6, q6, negBias), vec_madd(row6, q6, bias),
                    vec_cmpgt(row6, zero));
            row7 = vec_sel(vec_madd(row7, q7, negBias), vec_madd(row7, q7, bias),
                    vec_cmpgt(row7, zero));

            q0 = vec_ctf(qmat[1], QMAT_SHIFT);
            q1 = vec_ctf(qmat[3], QMAT_SHIFT);
            q2 = vec_ctf(qmat[5], QMAT_SHIFT);
            q3 = vec_ctf(qmat[7], QMAT_SHIFT);
            q4 = vec_ctf(qmat[9], QMAT_SHIFT);
            q5 = vec_ctf(qmat[11], QMAT_SHIFT);
            q6 = vec_ctf(qmat[13], QMAT_SHIFT);
            q7 = vec_ctf(qmat[15], QMAT_SHIFT);

            alt0 = vec_sel(vec_madd(alt0, q0, negBias), vec_madd(alt0, q0, bias),
                    vec_cmpgt(alt0, zero));
            alt1 = vec_sel(vec_madd(alt1, q1, negBias), vec_madd(alt1, q1, bias),
                    vec_cmpgt(alt1, zero));
            alt2 = vec_sel(vec_madd(alt2, q2, negBias), vec_madd(alt2, q2, bias),
                    vec_cmpgt(alt2, zero));
            alt3 = vec_sel(vec_madd(alt3, q3, negBias), vec_madd(alt3, q3, bias),
                    vec_cmpgt(alt3, zero));
            alt4 = vec_sel(vec_madd(alt4, q4, negBias), vec_madd(alt4, q4, bias),
                    vec_cmpgt(alt4, zero));
            alt5 = vec_sel(vec_madd(alt5, q5, negBias), vec_madd(alt5, q5, bias),
                    vec_cmpgt(alt5, zero));
            alt6 = vec_sel(vec_madd(alt6, q6, negBias), vec_madd(alt6, q6, bias),
                    vec_cmpgt(alt6, zero));
            alt7 = vec_sel(vec_madd(alt7, q7, negBias), vec_madd(alt7, q7, bias),
                    vec_cmpgt(alt7, zero));
        }

 
    }

    // Store the data back into the original block
    {
        vector signed short data0, data1, data2, data3, data4, data5, data6, data7;

        data0 = vec_pack(vec_cts(row0, 0), vec_cts(alt0, 0));
        data1 = vec_pack(vec_cts(row1, 0), vec_cts(alt1, 0));
        data2 = vec_pack(vec_cts(row2, 0), vec_cts(alt2, 0));
        data3 = vec_pack(vec_cts(row3, 0), vec_cts(alt3, 0));
        data4 = vec_pack(vec_cts(row4, 0), vec_cts(alt4, 0));
        data5 = vec_pack(vec_cts(row5, 0), vec_cts(alt5, 0));
        data6 = vec_pack(vec_cts(row6, 0), vec_cts(alt6, 0));
        data7 = vec_pack(vec_cts(row7, 0), vec_cts(alt7, 0));

        {
            // Clamp for overflow
            vector signed int max_q_int, min_q_int;
            vector signed short max_q, min_q;

            LOAD4(max_q_int, &(s->max_qcoeff));
            LOAD4(min_q_int, &(s->min_qcoeff));

            max_q = vec_pack(max_q_int, max_q_int);
            min_q = vec_pack(min_q_int, min_q_int);

            data0 = vec_max(vec_min(data0, max_q), min_q);
            data1 = vec_max(vec_min(data1, max_q), min_q);
            data2 = vec_max(vec_min(data2, max_q), min_q);
            data4 = vec_max(vec_min(data4, max_q), min_q);
            data5 = vec_max(vec_min(data5, max_q), min_q);
            data6 = vec_max(vec_min(data6, max_q), min_q);
            data7 = vec_max(vec_min(data7, max_q), min_q);
        }

        vector bool char zero_01, zero_23, zero_45, zero_67;
        vector signed char scanIndices_01, scanIndices_23, scanIndices_45, scanIndices_67;
        vector signed char negOne = vec_splat_s8(-1);
        vector signed char* scanPtr =
                (vector signed char*)(s->intra_scantable.inverse);

        // Determine the largest non-zero index.
        zero_01 = vec_pack(vec_cmpeq(data0, (vector short)zero),
                vec_cmpeq(data1, (vector short)zero));
        zero_23 = vec_pack(vec_cmpeq(data2, (vector short)zero),
                vec_cmpeq(data3, (vector short)zero));
        zero_45 = vec_pack(vec_cmpeq(data4, (vector short)zero),
                vec_cmpeq(data5, (vector short)zero));
        zero_67 = vec_pack(vec_cmpeq(data6, (vector short)zero),
                vec_cmpeq(data7, (vector short)zero));

        // 64 biggest values
        scanIndices_01 = vec_sel(scanPtr[0], negOne, zero_01);
        scanIndices_23 = vec_sel(scanPtr[1], negOne, zero_23);
        scanIndices_45 = vec_sel(scanPtr[2], negOne, zero_45);
        scanIndices_67 = vec_sel(scanPtr[3], negOne, zero_67);

        // 32 largest values
        scanIndices_01 = vec_max(scanIndices_01, scanIndices_23);
        scanIndices_45 = vec_max(scanIndices_45, scanIndices_67);

        // 16 largest values
        scanIndices_01 = vec_max(scanIndices_01, scanIndices_45);

        // 8 largest values
        scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
                vec_mergel(scanIndices_01, negOne));

        // 4 largest values
        scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
                vec_mergel(scanIndices_01, negOne));

        // 2 largest values
        scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
                vec_mergel(scanIndices_01, negOne));

        // largest value
        scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
                vec_mergel(scanIndices_01, negOne));

        scanIndices_01 = vec_splat(scanIndices_01, 0);

        signed char lastNonZeroChar;

        vec_ste(scanIndices_01, 0, &lastNonZeroChar);

        lastNonZero = lastNonZeroChar;
        
        // While the data is still in vectors we check for the transpose IDCT permute
        // and handle it using the vector unit if we can.  This is the permute used
        // by the altivec idct, so it is common when using the altivec dct.

        if ((lastNonZero > 0) && (s->idct_permutation_type == FF_TRANSPOSE_IDCT_PERM))
        {
            TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
        }

        vec_st(data0, 0, data);
        vec_st(data1, 16, data);
        vec_st(data2, 32, data);
        vec_st(data3, 48, data);
        vec_st(data4, 64, data);
        vec_st(data5, 80, data);
        vec_st(data6, 96, data);
        vec_st(data7, 112, data);
    }

    // special handling of block[0]
    if (s->mb_intra)
    {
        if (!s->h263_aic)
        {
            if (n < 4)
                oldBaseValue /= s->y_dc_scale;
            else
                oldBaseValue /= s->c_dc_scale;
        }

        // Divide by 8, rounding the result
        data[0] = (oldBaseValue + 4) >> 3;
    }

    // We handled the tranpose permutation above and we don't
    // need to permute the "no" permutation case.
    if ((lastNonZero > 0) &&
        (s->idct_permutation_type != FF_TRANSPOSE_IDCT_PERM) &&
        (s->idct_permutation_type != FF_NO_IDCT_PERM))
    {
        ff_block_permute(data, s->idct_permutation,
                s->intra_scantable.scantable, lastNonZero);
    }

    return lastNonZero;
}