comparison ppc/mpegvideo_altivec.c @ 1033:b4172ff70d27 libavcodec

Altivec on non darwin systems patch by Romain Dolbeau

1032:8f440ca8e0b0

    vec = vec_ld(0, _load_addr); \
    vec = vec_perm(vec, vec, _perm_vec); \
    vec = vec_splat(vec, 0); \
}

int dct_quantize_altivec(MpegEncContext* s, 
                        DCTELEM* data, int n,
                        int qscale, int* overflow)
{
    int lastNonZero;
    vector float row0, row1, row2, row3, row4, row5, row6, row7;
    vector float alt0, alt1, alt2, alt3, alt4, alt5, alt6, alt7;
    const vector float zero = (const vector float)(0.0f);

    // Load the data into the row/alt vectors
    {
        vector signed short data0, data1, data2, data3, data4, data5, data6, data7;

    // The following block could exist as a separate an altivec dct
		// function.  However, if we put it inline, the DCT data can remain
		// in the vector local variables, as floats, which we'll use during the
		// quantize step...
    {
        const vector float vec_0_298631336 = (vector float)(0.298631336f);
        const vector float vec_0_390180644 = (vector float)(-0.390180644f);
        const vector float vec_0_541196100 = (vector float)(0.541196100f);
        const vector float vec_0_765366865 = (vector float)(0.765366865f);
        const vector float vec_0_899976223 = (vector float)(-0.899976223f);
        const vector float vec_1_175875602 = (vector float)(1.175875602f);
        const vector float vec_1_501321110 = (vector float)(1.501321110f);
        const vector float vec_1_847759065 = (vector float)(-1.847759065f);
        const vector float vec_1_961570560 = (vector float)(-1.961570560f);
        const vector float vec_2_053119869 = (vector float)(2.053119869f);
        const vector float vec_2_562915447 = (vector float)(-2.562915447f);
        const vector float vec_3_072711026 = (vector float)(3.072711026f);


        int whichPass, whichHalf;

        for(whichPass = 1; whichPass<=2; whichPass++)

        // 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)(-1.0f);
            LOAD4(biasInt, biasAddr);
            bias = vec_ctf(biasInt, QUANT_BIAS_SHIFT);
            negBias = vec_madd(bias, negOneFloat, zero);
        }

                s->intra_scantable.scantable, lastNonZero);
    }

    return lastNonZero;
}

/*
  AltiVec version of dct_unquantize_h263
  this code assumes `block' is 16 bytes-aligned
*/

            block[i] = level;
        }
    }
#else /* ALTIVEC_USE_REFERENCE_C_CODE */
    {
      register const vector short vczero = (const vector short)(0);
      short __attribute__ ((aligned(16))) qmul8[] =
          {
            qmul, qmul, qmul, qmul,
            qmul, qmul, qmul, qmul
          };

1033:b4172ff70d27

    vec = vec_ld(0, _load_addr); \
    vec = vec_perm(vec, vec, _perm_vec); \
    vec = vec_splat(vec, 0); \
}


#ifdef CONFIG_DARWIN
#define FOUROF(a) (a)
#else
// slower, for dumb non-apple GCC
#define FOUROF(a) {a,a,a,a}
#endif
int dct_quantize_altivec(MpegEncContext* s, 
                        DCTELEM* data, int n,
                        int qscale, int* overflow)
{
    int lastNonZero;
    vector float row0, row1, row2, row3, row4, row5, row6, row7;
    vector float alt0, alt1, alt2, alt3, alt4, alt5, alt6, alt7;
    const vector float zero = (const vector float)FOUROF(0.);

    // Load the data into the row/alt vectors
    {
        vector signed short data0, data1, data2, data3, data4, data5, data6, data7;

    // The following block could exist as a separate an altivec dct
		// function.  However, if we put it inline, the DCT data can remain
		// in the vector local variables, as floats, which we'll use during the
		// quantize step...
    {
        const vector float vec_0_298631336 = (vector float)FOUROF(0.298631336f);
        const vector float vec_0_390180644 = (vector float)FOUROF(-0.390180644f);
        const vector float vec_0_541196100 = (vector float)FOUROF(0.541196100f);
        const vector float vec_0_765366865 = (vector float)FOUROF(0.765366865f);
        const vector float vec_0_899976223 = (vector float)FOUROF(-0.899976223f);
        const vector float vec_1_175875602 = (vector float)FOUROF(1.175875602f);
        const vector float vec_1_501321110 = (vector float)FOUROF(1.501321110f);
        const vector float vec_1_847759065 = (vector float)FOUROF(-1.847759065f);
        const vector float vec_1_961570560 = (vector float)FOUROF(-1.961570560f);
        const vector float vec_2_053119869 = (vector float)FOUROF(2.053119869f);
        const vector float vec_2_562915447 = (vector float)FOUROF(-2.562915447f);
        const vector float vec_3_072711026 = (vector float)FOUROF(3.072711026f);


        int whichPass, whichHalf;

        for(whichPass = 1; whichPass<=2; whichPass++)

        // 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)FOUROF(-1.0f);
            LOAD4(biasInt, biasAddr);
            bias = vec_ctf(biasInt, QUANT_BIAS_SHIFT);
            negBias = vec_madd(bias, negOneFloat, zero);
        }

                s->intra_scantable.scantable, lastNonZero);
    }

    return lastNonZero;
}
#undef FOUROF

/*
  AltiVec version of dct_unquantize_h263
  this code assumes `block' is 16 bytes-aligned
*/

            block[i] = level;
        }
    }
#else /* ALTIVEC_USE_REFERENCE_C_CODE */
    {
      register const vector short vczero = (const vector short)vec_splat_s16(0);
      short __attribute__ ((aligned(16))) qmul8[] =
          {
            qmul, qmul, qmul, qmul,
            qmul, qmul, qmul, qmul
          };