dct.c

linux下编译已经通过

#undef DST
}

static void sub16x16_dct8( int16_t dct[4][8][8], uint8_t *pix1, uint8_t *pix2 )
{
    sub8x8_dct8( dct[0], &pix1[0],               &pix2[0] );
    sub8x8_dct8( dct[1], &pix1[8],               &pix2[8] );
    sub8x8_dct8( dct[2], &pix1[8*FENC_STRIDE+0], &pix2[8*FDEC_STRIDE+0] );
    sub8x8_dct8( dct[3], &pix1[8*FENC_STRIDE+8], &pix2[8*FDEC_STRIDE+8] );
}

#define IDCT8_1D {\
    const int a0 =  SRC(0) + SRC(4);\
    const int a2 =  SRC(0) - SRC(4);\
    const int a4 = (SRC(2)>>1) - SRC(6);\
    const int a6 = (SRC(6)>>1) + SRC(2);\
    const int b0 = a0 + a6;\
    const int b2 = a2 + a4;\
    const int b4 = a2 - a4;\
    const int b6 = a0 - a6;\
    const int a1 = -SRC(3) + SRC(5) - SRC(7) - (SRC(7)>>1);\
    const int a3 =  SRC(1) + SRC(7) - SRC(3) - (SRC(3)>>1);\
    const int a5 = -SRC(1) + SRC(7) + SRC(5) + (SRC(5)>>1);\
    const int a7 =  SRC(3) + SRC(5) + SRC(1) + (SRC(1)>>1);\
    const int b1 = (a7>>2) + a1;\
    const int b3 =  a3 + (a5>>2);\
    const int b5 = (a3>>2) - a5;\
    const int b7 =  a7 - (a1>>2);\
    DST(0, b0 + b7);\
    DST(1, b2 + b5);\
    DST(2, b4 + b3);\
    DST(3, b6 + b1);\
    DST(4, b6 - b1);\
    DST(5, b4 - b3);\
    DST(6, b2 - b5);\
    DST(7, b0 - b7);\
}

static void add8x8_idct8( uint8_t *dst, int16_t dct[8][8] )
{
    int i;

    dct[0][0] += 32; // rounding for the >>6 at the end

#define SRC(x)     dct[x][i]
#define DST(x,rhs) dct[x][i] = (rhs)
    for( i = 0; i < 8; i++ )
        IDCT8_1D
#undef SRC
#undef DST

#define SRC(x)     dct[i][x]
#define DST(x,rhs) dst[i + x*FDEC_STRIDE] = clip_uint8( dst[i + x*FDEC_STRIDE] + ((rhs) >> 6) );
    for( i = 0; i < 8; i++ )
        IDCT8_1D
#undef SRC
#undef DST
}

static void add16x16_idct8( uint8_t *dst, int16_t dct[4][8][8] )
{
    add8x8_idct8( &dst[0],               dct[0] );
    add8x8_idct8( &dst[8],               dct[1] );
    add8x8_idct8( &dst[8*FDEC_STRIDE+0], dct[2] );
    add8x8_idct8( &dst[8*FDEC_STRIDE+8], dct[3] );
}


/****************************************************************************
 * x264_dct_init:
 ****************************************************************************/
void x264_dct_init( int cpu, x264_dct_function_t *dctf )
{
    dctf->sub4x4_dct    = sub4x4_dct;
    dctf->add4x4_idct   = add4x4_idct;

    dctf->sub8x8_dct    = sub8x8_dct;
    dctf->add8x8_idct   = add8x8_idct;

    dctf->sub16x16_dct  = sub16x16_dct;
    dctf->add16x16_idct = add16x16_idct;

    dctf->sub8x8_dct8   = sub8x8_dct8;
    dctf->add8x8_idct8  = add8x8_idct8;

    dctf->sub16x16_dct8  = sub16x16_dct8;
    dctf->add16x16_idct8 = add16x16_idct8;

    dctf->dct4x4dc  = dct4x4dc;
    dctf->idct4x4dc = idct4x4dc;

    dctf->dct2x2dc  = dct2x2dc;
    dctf->idct2x2dc = dct2x2dc;

#ifdef HAVE_MMX
    if( cpu&X264_CPU_MMX )
    {
        dctf->sub4x4_dct    = x264_sub4x4_dct_mmx;
        dctf->sub8x8_dct    = x264_sub8x8_dct_mmx;
        dctf->sub16x16_dct  = x264_sub16x16_dct_mmx;

        dctf->add4x4_idct   = x264_add4x4_idct_mmx;
        dctf->add8x8_idct   = x264_add8x8_idct_mmx;
        dctf->add16x16_idct = x264_add16x16_idct_mmx;

        dctf->dct4x4dc      = x264_dct4x4dc_mmx;
        dctf->idct4x4dc     = x264_idct4x4dc_mmx;

#ifndef ARCH_X86_64
        dctf->sub8x8_dct8   = x264_sub8x8_dct8_mmx;
        dctf->sub16x16_dct8 = x264_sub16x16_dct8_mmx;

        dctf->add8x8_idct8  = x264_add8x8_idct8_mmx;
        dctf->add16x16_idct8= x264_add16x16_idct8_mmx;
#endif
    }

#ifdef ARCH_X86_64
    if( cpu&X264_CPU_SSE2 )
    {
        dctf->sub8x8_dct8   = x264_sub8x8_dct8_sse2;
        dctf->sub16x16_dct8 = x264_sub16x16_dct8_sse2;

        dctf->add8x8_idct8  = x264_add8x8_idct8_sse2;
        dctf->add16x16_idct8= x264_add16x16_idct8_sse2;
    }
#endif
#endif //HAVE_MMX

#ifdef ARCH_PPC
    if( cpu&X264_CPU_ALTIVEC )
    {
        dctf->sub4x4_dct    = x264_sub4x4_dct_altivec;
        dctf->sub8x8_dct    = x264_sub8x8_dct_altivec;
        dctf->sub16x16_dct  = x264_sub16x16_dct_altivec;

        dctf->sub8x8_dct8   = x264_sub8x8_dct8_altivec;
        dctf->sub16x16_dct8 = x264_sub16x16_dct8_altivec;

        dctf->add8x8_idct8  = x264_add8x8_idct8_altivec;
        dctf->add16x16_idct8= x264_add16x16_idct8_altivec;
    }
#endif
}

void x264_dct_init_weights( void )
{
    int i, j;
    for( j=0; j<2; j++ )
    {
        for( i=0; i<16; i++ )
            x264_dct4_weight2_zigzag[j][i] = x264_dct4_weight2_tab[ x264_zigzag_scan4[j][i] ];
        for( i=0; i<64; i++ )
            x264_dct8_weight2_zigzag[j][i] = x264_dct8_weight2_tab[ x264_zigzag_scan8[j][i] ];
    }
}


#define ZIG(i,y,x) level[i] = dct[x][y];

static void zigzag_scan_8x8_frame( int level[64], int16_t dct[8][8] )
{
    ZIG( 0,0,0) ZIG( 1,0,1) ZIG( 2,1,0) ZIG( 3,2,0)
    ZIG( 4,1,1) ZIG( 5,0,2) ZIG( 6,0,3) ZIG( 7,1,2)
    ZIG( 8,2,1) ZIG( 9,3,0) ZIG(10,4,0) ZIG(11,3,1)
    ZIG(12,2,2) ZIG(13,1,3) ZIG(14,0,4) ZIG(15,0,5)
    ZIG(16,1,4) ZIG(17,2,3) ZIG(18,3,2) ZIG(19,4,1)
    ZIG(20,5,0) ZIG(21,6,0) ZIG(22,5,1) ZIG(23,4,2)
    ZIG(24,3,3) ZIG(25,2,4) ZIG(26,1,5) ZIG(27,0,6)
    ZIG(28,0,7) ZIG(29,1,6) ZIG(30,2,5) ZIG(31,3,4)
    ZIG(32,4,3) ZIG(33,5,2) ZIG(34,6,1) ZIG(35,7,0)
    ZIG(36,7,1) ZIG(37,6,2) ZIG(38,5,3) ZIG(39,4,4)
    ZIG(40,3,5) ZIG(41,2,6) ZIG(42,1,7) ZIG(43,2,7)
    ZIG(44,3,6) ZIG(45,4,5) ZIG(46,5,4) ZIG(47,6,3)
    ZIG(48,7,2) ZIG(49,7,3) ZIG(50,6,4) ZIG(51,5,5)
    ZIG(52,4,6) ZIG(53,3,7) ZIG(54,4,7) ZIG(55,5,6)
    ZIG(56,6,5) ZIG(57,7,4) ZIG(58,7,5) ZIG(59,6,6)
    ZIG(60,5,7) ZIG(61,6,7) ZIG(62,7,6) ZIG(63,7,7)
}

static void zigzag_scan_8x8_field( int level[64], int16_t dct[8][8] )
{
    ZIG( 0,0,0) ZIG( 1,1,0) ZIG( 2,2,0) ZIG( 3,0,1)
    ZIG( 4,1,1) ZIG( 5,3,0) ZIG( 6,4,0) ZIG( 7,2,1)
    ZIG( 8,0,2) ZIG( 9,3,1) ZIG(10,5,0) ZIG(11,6,0)
    ZIG(12,7,0) ZIG(13,4,1) ZIG(14,1,2) ZIG(15,0,3)
    ZIG(16,2,2) ZIG(17,5,1) ZIG(18,6,1) ZIG(19,7,1)
    ZIG(20,3,2) ZIG(21,1,3) ZIG(22,0,4) ZIG(23,2,3)
    ZIG(24,4,2) ZIG(25,5,2) ZIG(26,6,2) ZIG(27,7,2)
    ZIG(28,3,3) ZIG(29,1,4) ZIG(30,0,5) ZIG(31,2,4)
    ZIG(32,4,3) ZIG(33,5,3) ZIG(34,6,3) ZIG(35,7,3)
    ZIG(36,3,4) ZIG(37,1,5) ZIG(38,0,6) ZIG(39,2,5)
    ZIG(40,4,4) ZIG(41,5,4) ZIG(42,6,4) ZIG(43,7,4)
    ZIG(44,3,5) ZIG(45,1,6) ZIG(46,2,6) ZIG(47,4,5)
    ZIG(48,5,5) ZIG(49,6,5) ZIG(50,7,5) ZIG(51,3,6)
    ZIG(52,0,7) ZIG(53,1,7) ZIG(54,4,6) ZIG(55,5,6)
    ZIG(56,6,6) ZIG(57,7,6) ZIG(58,2,7) ZIG(59,3,7)
    ZIG(60,4,7) ZIG(61,5,7) ZIG(62,6,7) ZIG(63,7,7)
}

static void zigzag_scan_4x4_frame( int level[16], int16_t dct[4][4] )
{
    ZIG( 0,0,0) ZIG( 1,0,1) ZIG( 2,1,0) ZIG( 3,2,0)
    ZIG( 4,1,1) ZIG( 5,0,2) ZIG( 6,0,3) ZIG( 7,1,2)
    ZIG( 8,2,1) ZIG( 9,3,0) ZIG(10,3,1) ZIG(11,2,2)
    ZIG(12,1,3) ZIG(13,2,3) ZIG(14,3,2) ZIG(15,3,3)
}

static void zigzag_scan_4x4_field( int level[16], int16_t dct[4][4] )
{
    ZIG( 0,0,0) ZIG( 1,1,0) ZIG( 2,0,1) ZIG( 3,2,0)
    ZIG( 4,3,0) ZIG( 5,1,1) ZIG( 6,2,1) ZIG( 7,3,1)
    ZIG( 8,0,2) ZIG( 9,1,2) ZIG(10,2,2) ZIG(11,3,2)
    ZIG(12,0,3) ZIG(13,1,3) ZIG(14,2,3) ZIG(15,3,3)
}

static void zigzag_scan_4x4ac_frame( int level[15], int16_t dct[4][4] )
{
                ZIG( 0,0,1) ZIG( 1,1,0) ZIG( 2,2,0)
    ZIG( 3,1,1) ZIG( 4,0,2) ZIG( 5,0,3) ZIG( 6,1,2)
    ZIG( 7,2,1) ZIG( 8,3,0) ZIG( 9,3,1) ZIG(10,2,2)
    ZIG(11,1,3) ZIG(12,2,3) ZIG(13,3,2) ZIG(14,3,3)
}

static void zigzag_scan_4x4ac_field( int level[15], int16_t dct[4][4] )
{
                ZIG( 0,1,0) ZIG( 1,0,1) ZIG( 2,2,0)
    ZIG( 3,3,0) ZIG( 4,1,1) ZIG( 5,2,1) ZIG( 6,3,1)
    ZIG( 7,0,2) ZIG( 8,1,2) ZIG( 9,2,2) ZIG(10,3,2)
    ZIG(11,0,3) ZIG(12,1,3) ZIG(13,2,3) ZIG(14,3,3)
}

#undef ZIG

#define ZIG(i,y,x) {\
    int oe = x+y*FENC_STRIDE;\
    int od = x+y*FDEC_STRIDE;\
    level[i] = p_src[oe] - p_dst[od];\
    p_dst[od] = p_src[oe];\
}

static void zigzag_sub_4x4_frame( int level[16], const uint8_t *p_src, uint8_t *p_dst )
{
    ZIG( 0,0,0) ZIG( 1,0,1) ZIG( 2,1,0) ZIG( 3,2,0)
    ZIG( 4,1,1) ZIG( 5,0,2) ZIG( 6,0,3) ZIG( 7,1,2)
    ZIG( 8,2,1) ZIG( 9,3,0) ZIG(10,3,1) ZIG(11,2,2)
    ZIG(12,1,3) ZIG(13,2,3) ZIG(14,3,2) ZIG(15,3,3)
}

static void zigzag_sub_4x4_field( int level[16], const uint8_t *p_src, uint8_t *p_dst )
{
    ZIG( 0,0,0) ZIG( 1,1,0) ZIG( 2,0,1) ZIG( 3,2,0)
    ZIG( 4,3,0) ZIG( 5,1,1) ZIG( 6,2,1) ZIG( 7,3,1)
    ZIG( 8,0,2) ZIG( 9,1,2) ZIG(10,2,2) ZIG(11,3,2)
    ZIG(12,0,3) ZIG(13,1,3) ZIG(14,2,3) ZIG(15,3,3)
}

static void zigzag_sub_4x4ac_frame( int level[15], const uint8_t *p_src, uint8_t *p_dst )
{
                ZIG( 0,0,1) ZIG( 1,1,0) ZIG( 2,2,0)
    ZIG( 3,1,1) ZIG( 4,0,2) ZIG( 5,0,3) ZIG( 6,1,2)
    ZIG( 7,2,1) ZIG( 8,3,0) ZIG( 9,3,1) ZIG(10,2,2)
    ZIG(11,1,3) ZIG(12,2,3) ZIG(13,3,2) ZIG(14,3,3)
}

static void zigzag_sub_4x4ac_field( int level[15], const uint8_t *p_src, uint8_t *p_dst )
{
                ZIG( 0,1,0) ZIG( 1,0,1) ZIG( 2,2,0)
    ZIG( 3,3,0) ZIG( 4,1,1) ZIG( 5,2,1) ZIG( 6,3,1)
    ZIG( 7,0,2) ZIG( 8,1,2) ZIG( 9,2,2) ZIG(10,3,2)
    ZIG(11,0,3) ZIG(12,1,3) ZIG(13,2,3) ZIG(14,3,3)
}

#undef ZIG

void x264_zigzag_init( int cpu, x264_zigzag_function_t *pf, int b_interlaced )
{
    if( b_interlaced )
    {
        pf->scan_8x8   = zigzag_scan_8x8_field;
        pf->scan_4x4   = zigzag_scan_4x4_field;
        pf->scan_4x4ac = zigzag_scan_4x4ac_field;
        pf->sub_4x4    = zigzag_sub_4x4_field;
        pf->sub_4x4ac  = zigzag_sub_4x4ac_field;
#ifdef HAVE_MMX
#ifdef ARCH_X86
        if( cpu&X264_CPU_MMX )
            pf->scan_4x4 = x264_zigzag_scan_4x4_field_mmx;
#endif
#ifdef ARCH_X86_64
        if( cpu&X264_CPU_SSE2 )
            pf->scan_4x4 = x264_zigzag_scan_4x4_field_sse2;
#endif
#endif
    }
    else
    {
        pf->scan_8x8   = zigzag_scan_8x8_frame;
        pf->scan_4x4   = zigzag_scan_4x4_frame;
        pf->scan_4x4ac = zigzag_scan_4x4ac_frame;
        pf->sub_4x4    = zigzag_sub_4x4_frame;
        pf->sub_4x4ac  = zigzag_sub_4x4ac_frame;
    }
}