mpegvideo_altivec.c

mediastreamer2是开源的网络传输媒体流的库

            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);
        signed char lastNonZeroChar;

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


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

    return lastNonZero;
}
#undef FOUROF

/*
  AltiVec version of dct_unquantize_h263
  this code assumes `block' is 16 bytes-aligned
*/
void dct_unquantize_h263_altivec(MpegEncContext *s,
                                 DCTELEM *block, int n, int qscale)
{
POWERPC_PERF_DECLARE(altivec_dct_unquantize_h263_num, 1);
    int i, level, qmul, qadd;
    int nCoeffs;

    assert(s->block_last_index[n]>=0);

POWERPC_PERF_START_COUNT(altivec_dct_unquantize_h263_num, 1);

    qadd = (qscale - 1) | 1;
    qmul = qscale << 1;

    if (s->mb_intra) {
        if (!s->h263_aic) {
            if (n < 4)
                block[0] = block[0] * s->y_dc_scale;
            else
                block[0] = block[0] * s->c_dc_scale;
        }else
            qadd = 0;
        i = 1;
        nCoeffs= 63; //does not always use zigzag table
    } else {
        i = 0;
        nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
    }

    {
      register const vector signed short vczero = (const vector signed short)vec_splat_s16(0);
      DECLARE_ALIGNED_16(short, qmul8[]) =
          {
            qmul, qmul, qmul, qmul,
            qmul, qmul, qmul, qmul
          };
      DECLARE_ALIGNED_16(short, qadd8[]) =
          {
            qadd, qadd, qadd, qadd,
            qadd, qadd, qadd, qadd
          };
      DECLARE_ALIGNED_16(short, nqadd8[]) =
          {
            -qadd, -qadd, -qadd, -qadd,
            -qadd, -qadd, -qadd, -qadd
          };
      register vector signed short blockv, qmulv, qaddv, nqaddv, temp1;
      register vector bool short blockv_null, blockv_neg;
      register short backup_0 = block[0];
      register int j = 0;

      qmulv = vec_ld(0, qmul8);
      qaddv = vec_ld(0, qadd8);
      nqaddv = vec_ld(0, nqadd8);

#if 0 // block *is* 16 bytes-aligned, it seems.
      // first make sure block[j] is 16 bytes-aligned
      for(j = 0; (j <= nCoeffs) && ((((unsigned long)block) + (j << 1)) & 0x0000000F) ; j++) {
        level = block[j];
        if (level) {
          if (level < 0) {
                level = level * qmul - qadd;
            } else {
                level = level * qmul + qadd;
            }
            block[j] = level;
        }
      }
#endif

      // vectorize all the 16 bytes-aligned blocks
      // of 8 elements
      for(; (j + 7) <= nCoeffs ; j+=8)
      {
        blockv = vec_ld(j << 1, block);
        blockv_neg = vec_cmplt(blockv, vczero);
        blockv_null = vec_cmpeq(blockv, vczero);
        // choose between +qadd or -qadd as the third operand
        temp1 = vec_sel(qaddv, nqaddv, blockv_neg);
        // multiply & add (block{i,i+7} * qmul [+-] qadd)
        temp1 = vec_mladd(blockv, qmulv, temp1);
        // put 0 where block[{i,i+7} used to have 0
        blockv = vec_sel(temp1, blockv, blockv_null);
        vec_st(blockv, j << 1, block);
      }

      // if nCoeffs isn't a multiple of 8, finish the job
      // using good old scalar units.
      // (we could do it using a truncated vector,
      // but I'm not sure it's worth the hassle)
      for(; j <= nCoeffs ; j++) {
        level = block[j];
        if (level) {
          if (level < 0) {
                level = level * qmul - qadd;
            } else {
                level = level * qmul + qadd;
            }
            block[j] = level;
        }
      }

      if (i == 1)
      { // cheat. this avoid special-casing the first iteration
        block[0] = backup_0;
      }
    }
POWERPC_PERF_STOP_COUNT(altivec_dct_unquantize_h263_num, nCoeffs == 63);
}


extern void idct_put_altivec(uint8_t *dest, int line_size, int16_t *block);
extern void idct_add_altivec(uint8_t *dest, int line_size, int16_t *block);

void MPV_common_init_altivec(MpegEncContext *s)
{
    if (s->avctx->lowres==0)
    {
        if ((s->avctx->idct_algo == FF_IDCT_AUTO) ||
                (s->avctx->idct_algo == FF_IDCT_ALTIVEC))
        {
            s->dsp.idct_put = idct_put_altivec;
            s->dsp.idct_add = idct_add_altivec;
            s->dsp.idct_permutation_type = FF_TRANSPOSE_IDCT_PERM;
        }
    }

    // Test to make sure that the dct required alignments are met.
    if ((((long)(s->q_intra_matrix) & 0x0f) != 0) ||
        (((long)(s->q_inter_matrix) & 0x0f) != 0))
    {
        av_log(s->avctx, AV_LOG_INFO, "Internal Error: q-matrix blocks must be 16-byte aligned "
                "to use AltiVec DCT. Reverting to non-AltiVec version.\n");
        return;
    }

    if (((long)(s->intra_scantable.inverse) & 0x0f) != 0)
    {
        av_log(s->avctx, AV_LOG_INFO, "Internal Error: scan table blocks must be 16-byte aligned "
                "to use AltiVec DCT. Reverting to non-AltiVec version.\n");
        return;
    }


    if ((s->avctx->dct_algo == FF_DCT_AUTO) ||
            (s->avctx->dct_algo == FF_DCT_ALTIVEC))
    {
#if 0 /* seems to cause trouble under some circumstances */
        s->dct_quantize = dct_quantize_altivec;
#endif
        s->dct_unquantize_h263_intra = dct_unquantize_h263_altivec;
        s->dct_unquantize_h263_inter = dct_unquantize_h263_altivec;
    }
}