fdct.c

Motion JPEG编解码器源代码

#undef MERGE_S16
    /* }}} */


/* Some of the initial calculations can be done as vector short before
 * conversion to vector float.  The following code section takes advantage
 * of this.
 */
#if 1
    /* fdct rows {{{ */
    vs16(x0) = vec_add(vs16(b00), vs16(b70));
    vs16(x7) = vec_sub(vs16(b00), vs16(b70));
    vs16(x1) = vec_add(vs16(b10), vs16(b60));
    vs16(x6) = vec_sub(vs16(b10), vs16(b60));
    vs16(x2) = vec_add(vs16(b20), vs16(b50));
    vs16(x5) = vec_sub(vs16(b20), vs16(b50));
    vs16(x3) = vec_add(vs16(b30), vs16(b40));
    vs16(x4) = vec_sub(vs16(b30), vs16(b40));

    vs16(b70) = vec_add(vs16(x0), vs16(x3));
    vs16(b10) = vec_add(vs16(x1), vs16(x2));

    vs16(b00) = vec_add(vs16(b70), vs16(b10));
    vs16(b40) = vec_sub(vs16(b70), vs16(b10));

#define CTF0(n) \
    vs32(b##n##1) = vec_unpackl(vs16(b##n##0)); \
    vs32(b##n##0) = vec_unpackh(vs16(b##n##0)); \
    b##n##1 = vec_ctf(vs32(b##n##1), 0); \
    b##n##0 = vec_ctf(vs32(b##n##0), 0);

    CTF0(0);
    CTF0(4);

    vs16(b20) = vec_sub(vs16(x0), vs16(x3));
    vs16(b60) = vec_sub(vs16(x1), vs16(x2));

    CTF0(2);
    CTF0(6);
#undef CTF0

    x0 = vec_add(b60, b20);
    x1 = vec_add(b61, b21);

    cnst = LD_W2;
    x0 = vec_madd(cnst, x0, mzero);
    x1 = vec_madd(cnst, x1, mzero);
    cnst = LD_W1;
    b20 = vec_madd(cnst, b20, x0);
    b21 = vec_madd(cnst, b21, x1);
    cnst = LD_W0;
    b60 = vec_madd(cnst, b60, x0);
    b61 = vec_madd(cnst, b61, x1);

#define CTFX(x,b) \
    vs32(b##0) = vec_unpackh(vs16(x)); \
    vs32(b##1) = vec_unpackl(vs16(x)); \
    b##0 = vec_ctf(vs32(b##0), 0); \
    b##1 = vec_ctf(vs32(b##1), 0); \

    CTFX(x4, b7);
    CTFX(x5, b5);
    CTFX(x6, b3);
    CTFX(x7, b1);

#undef CTFX


    x0 = vec_add(b70, b10);
    x1 = vec_add(b50, b30);
    x2 = vec_add(b70, b30);
    x3 = vec_add(b50, b10);
    x8 = vec_add(x2, x3);
    cnst = LD_W3;
    x8 = vec_madd(cnst, x8, mzero);

    cnst = LD_W8;
    x0 = vec_madd(cnst, x0, mzero);
    cnst = LD_W9;
    x1 = vec_madd(cnst, x1, mzero);
    cnst = LD_WA;
    x2 = vec_madd(cnst, x2, x8);
    cnst = LD_WB;
    x3 = vec_madd(cnst, x3, x8);

    cnst = LD_W4;
    b70 = vec_madd(cnst, b70, x0);
    cnst = LD_W5;
    b50 = vec_madd(cnst, b50, x1);
    cnst = LD_W6;
    b30 = vec_madd(cnst, b30, x1);
    cnst = LD_W7;
    b10 = vec_madd(cnst, b10, x0);

    b70 = vec_add(b70, x2);
    b50 = vec_add(b50, x3);
    b30 = vec_add(b30, x2);
    b10 = vec_add(b10, x3);


    x0 = vec_add(b71, b11);
    x1 = vec_add(b51, b31);
    x2 = vec_add(b71, b31);
    x3 = vec_add(b51, b11);
    x8 = vec_add(x2, x3);
    cnst = LD_W3;
    x8 = vec_madd(cnst, x8, mzero);

    cnst = LD_W8;
    x0 = vec_madd(cnst, x0, mzero);
    cnst = LD_W9;
    x1 = vec_madd(cnst, x1, mzero);
    cnst = LD_WA;
    x2 = vec_madd(cnst, x2, x8);
    cnst = LD_WB;
    x3 = vec_madd(cnst, x3, x8);

    cnst = LD_W4;
    b71 = vec_madd(cnst, b71, x0);
    cnst = LD_W5;
    b51 = vec_madd(cnst, b51, x1);
    cnst = LD_W6;
    b31 = vec_madd(cnst, b31, x1);
    cnst = LD_W7;
    b11 = vec_madd(cnst, b11, x0);

    b71 = vec_add(b71, x2);
    b51 = vec_add(b51, x3);
    b31 = vec_add(b31, x2);
    b11 = vec_add(b11, x3);
    /* }}} */
#else
    /* convert to float {{{ */
#define CTF(n) \
    vs32(b##n##1) = vec_unpackl(vs16(b##n##0)); \
    vs32(b##n##0) = vec_unpackh(vs16(b##n##0)); \
    b##n##1 = vec_ctf(vs32(b##n##1), 0); \
    b##n##0 = vec_ctf(vs32(b##n##0), 0); \

    CTF(0);
    CTF(1);
    CTF(2);
    CTF(3);
    CTF(4);
    CTF(5);
    CTF(6);
    CTF(7);

#undef CTF
    /* }}} */

    FDCTROW(b00, b10, b20, b30, b40, b50, b60, b70);
    FDCTROW(b01, b11, b21, b31, b41, b51, b61, b71);
#endif


    /* 8x8 matrix transpose (vector float[8][2]) {{{ */
    x0 = vec_mergel(b00, b20);
    x1 = vec_mergeh(b00, b20);
    x2 = vec_mergel(b10, b30);
    x3 = vec_mergeh(b10, b30);

    b00 = vec_mergeh(x1, x3);
    b10 = vec_mergel(x1, x3);
    b20 = vec_mergeh(x0, x2);
    b30 = vec_mergel(x0, x2);

    x4 = vec_mergel(b41, b61);
    x5 = vec_mergeh(b41, b61);
    x6 = vec_mergel(b51, b71);
    x7 = vec_mergeh(b51, b71);

    b41 = vec_mergeh(x5, x7);
    b51 = vec_mergel(x5, x7);
    b61 = vec_mergeh(x4, x6);
    b71 = vec_mergel(x4, x6);

    x0 = vec_mergel(b01, b21);
    x1 = vec_mergeh(b01, b21);
    x2 = vec_mergel(b11, b31);
    x3 = vec_mergeh(b11, b31);

    x4 = vec_mergel(b40, b60);
    x5 = vec_mergeh(b40, b60);
    x6 = vec_mergel(b50, b70);
    x7 = vec_mergeh(b50, b70);

    b40 = vec_mergeh(x1, x3);
    b50 = vec_mergel(x1, x3);
    b60 = vec_mergeh(x0, x2);
    b70 = vec_mergel(x0, x2);

    b01 = vec_mergeh(x5, x7);
    b11 = vec_mergel(x5, x7);
    b21 = vec_mergeh(x4, x6);
    b31 = vec_mergel(x4, x6);
    /* }}} */


    /* divby8 = (vector float)(0.125); {{{ */
    vu8(divby8) = vec_splat_u8(-6);
    vu16(divby8) = vec_sr(vu16(divby8),vu16(divby8));
    vs32(divby8) = vec_unpackh(vs16(divby8));
    divby8 = vec_sld(divby8, divby8, 3);
    /* }}} */
    cnsts0 = vec_madd(cnsts0, divby8, mzero);
    cnsts1 = vec_madd(cnsts1, divby8, mzero);
    cnsts2 = vec_madd(cnsts2, divby8, mzero);


    FDCTCOL(b00, b10, b20, b30, b40, b50, b60, b70);
    FDCTCOL(b01, b11, b21, b31, b41, b51, b61, b71);


    /* round, convert back to short and clip {{{ */

    /* cnsts0 = max = (2047); cnsts2 = min = (-2048); {{{ */
    vu8(cnsts0) = vec_splat_u8(0x7);
    vs16(x8) = vec_splat_s16(-1); /* 0xffff */
    vu8(cnsts0) = vec_mergeh(vu8(cnsts0), vu8(x8)); /* 0x07ff == 2047 */
    vs16(cnsts2) = vec_sub(vs16(x8), vs16(cnsts0));
    /* }}} */

#define CTS(n) \
    b##n##0 = vec_round(b##n##0); \
    b##n##1 = vec_round(b##n##1); \
    vs32(b##n##0) = vec_cts(b##n##0, 0); \
    vs32(b##n##1) = vec_cts(b##n##1, 0); \
    vs16(b##n##0) = vec_pack(vs32(b##n##0), vs32(b##n##1)); \
    vs16(b##n##0) = vec_min(vs16(b##n##0), vs16(cnsts0)); \
    vs16(b##n##0) = vec_max(vs16(b##n##0), vs16(cnsts2)); \
    vec_st(vs16(b##n##0), 0, bp);


    bp = (vector signed short*)block;
    CTS(0); bp++;
    CTS(1); bp++;
    CTS(2); bp++;
    CTS(3); bp++;
    CTS(4); bp++;
    CTS(5); bp++;
    CTS(6); bp++;
    CTS(7);

#undef CTS
    /* }}} */


    AMBER_STOP;
}


#if ALTIVEC_TEST_FUNCTION(fdct) /* {{{ */
#ifdef ALTIVEC_VERIFY
void fdct_altivec_verify(FDCT_PDECL)
{
    int i;

    fdct_altivec(FDCT_ARGS);

    for (i = 0; i < 64; i++) {
	if (block[i] < -2048)
	    mjpeg_warn("fdct: block[%d]=%d < -2048\n", i, block[i]);
	else if (block[i] > 2047)
	    mjpeg_warn("fdct: block[%d]=%d > 2047\n", i, block[i]);
    }
}
#else
ALTIVEC_TEST(fdct, void, (FDCT_PDECL), FDCT_PFMT, FDCT_ARGS);
#endif
#endif /* }}} */

/* vim:set foldmethod=marker foldlevel=0: */