deinterlace.c

video linux conference

                {
                    for( ; p_out < p_out_end ; )
                    {
                        Merge( p_out, p_in, p_in + p_pic->p[i_plane].i_pitch,
                               p_pic->p[i_plane].i_pitch );

                        p_out += p_pic->p[i_plane].i_pitch;
                        p_in += p_pic->p[i_plane].i_pitch;
                    }
                }

                else
                {
                    for( ; p_out < p_out_end ; )
                    {
                        Merge( p_out, p_in, p_in + p_pic->p[i_plane].i_pitch,
                               p_pic->p[i_plane].i_pitch );

                        p_out += p_pic->p[i_plane].i_pitch;
                        p_in += 2*p_pic->p[i_plane].i_pitch;
                    }
                }
                break;
        }
    }
    EndMerge();
}

#undef Merge

static void MergeGeneric( void *_p_dest, const void *_p_s1,
                          const void *_p_s2, size_t i_bytes )
{
    uint8_t* p_dest = (uint8_t*)_p_dest;
    const uint8_t *p_s1 = (const uint8_t *)_p_s1;
    const uint8_t *p_s2 = (const uint8_t *)_p_s2;
    uint8_t* p_end = p_dest + i_bytes - 8;

    while( p_dest < p_end )
    {
        *p_dest++ = ( (uint16_t)(*p_s1++) + (uint16_t)(*p_s2++) ) >> 1;
        *p_dest++ = ( (uint16_t)(*p_s1++) + (uint16_t)(*p_s2++) ) >> 1;
        *p_dest++ = ( (uint16_t)(*p_s1++) + (uint16_t)(*p_s2++) ) >> 1;
        *p_dest++ = ( (uint16_t)(*p_s1++) + (uint16_t)(*p_s2++) ) >> 1;
        *p_dest++ = ( (uint16_t)(*p_s1++) + (uint16_t)(*p_s2++) ) >> 1;
        *p_dest++ = ( (uint16_t)(*p_s1++) + (uint16_t)(*p_s2++) ) >> 1;
        *p_dest++ = ( (uint16_t)(*p_s1++) + (uint16_t)(*p_s2++) ) >> 1;
        *p_dest++ = ( (uint16_t)(*p_s1++) + (uint16_t)(*p_s2++) ) >> 1;
    }

    p_end += 8;

    while( p_dest < p_end )
    {
        *p_dest++ = ( (uint16_t)(*p_s1++) + (uint16_t)(*p_s2++) ) >> 1;
    }
}

#if defined(CAN_COMPILE_MMXEXT)
static void MergeMMX( void *_p_dest, const void *_p_s1, const void *_p_s2,
                      size_t i_bytes )
{
    uint8_t* p_dest = (uint8_t*)_p_dest;
    const uint8_t *p_s1 = (const uint8_t *)_p_s1;
    const uint8_t *p_s2 = (const uint8_t *)_p_s2;
    uint8_t* p_end = p_dest + i_bytes - 8;
    while( p_dest < p_end )
    {
        __asm__  __volatile__( "movq %2,%%mm1;"
                               "pavgb %1, %%mm1;"
                               "movq %%mm1, %0" :"=m" (*p_dest):
                                                 "m" (*p_s1),
                                                 "m" (*p_s2) );
        p_dest += 8;
        p_s1 += 8;
        p_s2 += 8;
    }

    p_end += 8;

    while( p_dest < p_end )
    {
        *p_dest++ = ( (uint16_t)(*p_s1++) + (uint16_t)(*p_s2++) ) >> 1;
    }
}
#endif

#if defined(CAN_COMPILE_SSE)
static void MergeSSE2( void *_p_dest, const void *_p_s1, const void *_p_s2,
                       size_t i_bytes )
{
    uint8_t* p_dest = (uint8_t*)_p_dest;
    const uint8_t *p_s1 = (const uint8_t *)_p_s1;
    const uint8_t *p_s2 = (const uint8_t *)_p_s2;
    uint8_t* p_end;
    while( (int)p_s1 % 16 )
    {
        *p_dest++ = ( (uint16_t)(*p_s1++) + (uint16_t)(*p_s2++) ) >> 1;
    }        
    p_end = p_dest + i_bytes - 16;
    while( p_dest < p_end )
    {
        __asm__  __volatile__( "movdqu %2,%%xmm1;"
                               "pavgb %1, %%xmm1;"
                               "movdqu %%xmm1, %0" :"=m" (*p_dest):
                                                 "m" (*p_s1),
                                                 "m" (*p_s2) );
        p_dest += 16;
        p_s1 += 16;
        p_s2 += 16;
    }

    p_end += 16;

    while( p_dest < p_end )
    {
        *p_dest++ = ( (uint16_t)(*p_s1++) + (uint16_t)(*p_s2++) ) >> 1;
    }
}
#endif

#if defined(CAN_COMPILE_MMXEXT) || defined(CAN_COMPILE_SSE)
static void EndMMX( void )
{
    __asm__ __volatile__( "emms" :: );
}
#endif

#ifdef CAN_COMPILE_C_ALTIVEC
static void MergeAltivec( void *_p_dest, const void *_p_s1,
                          const void *_p_s2, size_t i_bytes )
{
    uint8_t *p_dest = (uint8_t *)_p_dest;
    uint8_t *p_s1   = (uint8_t *)_p_s1;
    uint8_t *p_s2   = (uint8_t *)_p_s2;
    uint8_t *p_end  = p_dest + i_bytes - 15;

    /* Use C until the first 16-bytes aligned destination pixel */
    while( (int)p_dest & 0xF )
    {
        *p_dest++ = ( (uint16_t)(*p_s1++) + (uint16_t)(*p_s2++) ) >> 1;
    }

    if( ( (int)p_s1 & 0xF ) | ( (int)p_s2 & 0xF ) )
    {
        /* Unaligned source */
        vector unsigned char s1v, s2v, destv;
        vector unsigned char s1oldv, s2oldv, s1newv, s2newv;
        vector unsigned char perm1v, perm2v;

        perm1v = vec_lvsl( 0, p_s1 );
        perm2v = vec_lvsl( 0, p_s2 );
        s1oldv = vec_ld( 0, p_s1 );
        s2oldv = vec_ld( 0, p_s2 );

        while( p_dest < p_end )
        {
            s1newv = vec_ld( 16, p_s1 );
            s2newv = vec_ld( 16, p_s2 );
            s1v    = vec_perm( s1oldv, s1newv, perm1v );
            s2v    = vec_perm( s2oldv, s2newv, perm2v );
            s1oldv = s1newv;
            s2oldv = s2newv;
            destv  = vec_avg( s1v, s2v );
            vec_st( destv, 0, p_dest );

            p_s1   += 16;
            p_s2   += 16;
            p_dest += 16;
        }
    }
    else
    {
        /* Aligned source */
        vector unsigned char s1v, s2v, destv;

        while( p_dest < p_end )
        {
            s1v   = vec_ld( 0, p_s1 );
            s2v   = vec_ld( 0, p_s2 );
            destv = vec_avg( s1v, s2v );
            vec_st( destv, 0, p_dest );

            p_s1   += 16;
            p_s2   += 16;
            p_dest += 16;
        }
    }

    p_end += 15;

    while( p_dest < p_end )
    {
        *p_dest++ = ( (uint16_t)(*p_s1++) + (uint16_t)(*p_s2++) ) >> 1;
    }
}
#endif

/*****************************************************************************
 * RenderX: This algo works on a 8x8 block basic, it copies the top field
 * and apply a process to recreate the bottom field :
 *  If a 8x8 block is classified as :
 *   - progressive: it applies a small blend (1,6,1)
 *   - interlaced:
 *    * in the MMX version: we do a ME between the 2 fields, if there is a
 *    good match we use MC to recreate the bottom field (with a small
 *    blend (1,6,1) )
 *    * otherwise: it recreates the bottom field by an edge oriented
 *    interpolation.
  *****************************************************************************/

/* XDeint8x8Detect: detect if a 8x8 block is interlaced.
 * XXX: It need to access to 8x10
 * We use more than 8 lines to help with scrolling (text)
 * (and because XDeint8x8Frame use line 9)
 * XXX: smooth/uniform area with noise detection doesn't works well
 * but it's not really a problem because they don't have much details anyway
 */
static inline int ssd( int a ) { return a*a; }
static inline int XDeint8x8DetectC( uint8_t *src, int i_src )
{
    int y, x;
    int ff, fr;
    int fc;

    /* Detect interlacing */
    fc = 0;
    for( y = 0; y < 7; y += 2 )
    {
        ff = fr = 0;
        for( x = 0; x < 8; x++ )
        {
            fr += ssd(src[      x] - src[1*i_src+x]) +
                  ssd(src[i_src+x] - src[2*i_src+x]);
            ff += ssd(src[      x] - src[2*i_src+x]) +
                  ssd(src[i_src+x] - src[3*i_src+x]);
        }
        if( ff < 6*fr/8 && fr > 32 )
            fc++;

        src += 2*i_src;
    }

    return fc < 1 ? VLC_FALSE : VLC_TRUE;
}
#ifdef CAN_COMPILE_MMXEXT
static inline int XDeint8x8DetectMMXEXT( uint8_t *src, int i_src )
{

    int y, x;
    int32_t ff, fr;
    int fc;

    /* Detect interlacing */
    fc = 0;
    pxor_r2r( mm7, mm7 );
    for( y = 0; y < 9; y += 2 )
    {
        ff = fr = 0;
        pxor_r2r( mm5, mm5 );
        pxor_r2r( mm6, mm6 );
        for( x = 0; x < 8; x+=4 )
        {
            movd_m2r( src[        x], mm0 );
            movd_m2r( src[1*i_src+x], mm1 );
            movd_m2r( src[2*i_src+x], mm2 );
            movd_m2r( src[3*i_src+x], mm3 );

            punpcklbw_r2r( mm7, mm0 );
            punpcklbw_r2r( mm7, mm1 );
            punpcklbw_r2r( mm7, mm2 );
            punpcklbw_r2r( mm7, mm3 );

            movq_r2r( mm0, mm4 );

            psubw_r2r( mm1, mm0 );
            psubw_r2r( mm2, mm4 );

            psubw_r2r( mm1, mm2 );
            psubw_r2r( mm1, mm3 );

            pmaddwd_r2r( mm0, mm0 );
            pmaddwd_r2r( mm4, mm4 );
            pmaddwd_r2r( mm2, mm2 );
            pmaddwd_r2r( mm3, mm3 );
            paddd_r2r( mm0, mm2 );
            paddd_r2r( mm4, mm3 );
            paddd_r2r( mm2, mm5 );
            paddd_r2r( mm3, mm6 );
        }

        movq_r2r( mm5, mm0 );
        psrlq_i2r( 32, mm0 );
        paddd_r2r( mm0, mm5 );
        movd_r2m( mm5, fr );

        movq_r2r( mm6, mm0 );
        psrlq_i2r( 32, mm0 );
        paddd_r2r( mm0, mm6 );
        movd_r2m( mm6, ff );

        if( ff < 6*fr/8 && fr > 32 )
            fc++;

        src += 2*i_src;
    }
    return fc;
}
#endif

/* XDeint8x8Frame: apply a small blend between field (1,6,1).
 * This won't destroy details, and help if there is a bit of interlacing.
 * (It helps with paning to avoid flickers)
 * (Use 8x9 pixels)
 */
#if 0
static inline void XDeint8x8FrameC( uint8_t *dst, int i_dst,
                                    uint8_t *src, int i_src )
{
    int y, x;

    /* Progressive */
    for( y = 0; y < 8; y += 2 )
    {
        memcpy( dst, src, 8 );
        dst += i_dst;

        for( x = 0; x < 8; x++ )
            dst[x] = (src[x] + 6*src[1*i_src+x] + src[2*i_src+x] + 4 ) >> 3;
        dst += 1*i_dst;
        src += 2*i_src;
    }
}
#endif
static inline void XDeint8x8MergeC( uint8_t *dst, int i_dst,
                                    uint8_t *src1, int i_src1,
                                    uint8_t *src2, int i_src2 )
{
    int y, x;

    /* Progressive */
    for( y = 0; y < 8; y += 2 )
    {
        memcpy( dst, src1, 8 );
        dst  += i_dst;

        for( x = 0; x < 8; x++ )
            dst[x] = (src1[x] + 6*src2[x] + src1[i_src1+x] + 4 ) >> 3;
        dst += i_dst;

        src1 += i_src1;
        src2 += i_src2;
    }
}

#ifdef CAN_COMPILE_MMXEXT
static inline void XDeint8x8MergeMMXEXT( uint8_t *dst, int i_dst,
                                         uint8_t *src1, int i_src1,
                                         uint8_t *src2, int i_src2 )
{
    static const uint64_t m_4 = I64C(0x0004000400040004);
    int y, x;

    /* Progressive */
    pxor_r2r( mm7, mm7 );
    for( y = 0; y < 8; y += 2 )
    {
        for( x = 0; x < 8; x +=4 )
        {
            movd_m2r( src1[x], mm0 );
            movd_r2m( mm0, dst[x] );

            movd_m2r( src2[x], mm1 );
            movd_m2r( src1[i_src1+x], mm2 );

            punpcklbw_r2r( mm7, mm0 );
            punpcklbw_r2r( mm7, mm1 );
            punpcklbw_r2r( mm7, mm2 );
            paddw_r2r( mm1, mm1 );
            movq_r2r( mm1, mm3 );
            paddw_r2r( mm3, mm3 );
            paddw_r2r( mm2, mm0 );
            paddw_r2r( mm3, mm1 );
            paddw_m2r( m_4, mm1 );
            paddw_r2r( mm1, mm0 );
            psraw_i2r( 3, mm0 );
            packuswb_r2r( mm7, mm0 );
            movd_r2m( mm0, dst[i_dst+x] );
        }
        dst += 2*i_dst;
        src1 += i_src1;
        src2 += i_src2;
    }
}

#endif

/* For debug */
static inline void XDeint8x8Set( uint8_t *dst, int i_dst, uint8_t v )
{
    int y;
    for( y = 0; y < 8; y++ )
        memset( &dst[y*i_dst], v, 8 );
}

/* XDeint8x8FieldE: Stupid deinterlacing (1,0,1) for block that miss a
 * neighbour
 * (Use 8x9 pixels)
 * TODO: a better one for the inner part.
 */
static inline void XDeint8x8FieldEC( uint8_t *dst, int i_dst,
                                     uint8_t *src, int i_src )
{
    int y, x;

    /* Interlaced */
    for( y = 0; y < 8; y += 2 )
    {
        memcpy( dst, src, 8 );
        dst += i_dst;

        for( x = 0; x < 8; x++ )
            dst[x] = (src[x] + src[2*i_src+x] ) >> 1;
        dst += 1*i_dst;
        src += 2*i_src;
    }
}