video_decoder.c

vlc stand 0.1.99 ist sehr einfach

            "movq       %%mm1,%%mm2\n\t"
            "punpckhbw  %%mm7,%%mm1\n\t"
            "punpcklbw  %%mm7,%%mm2\n\t"
            "paddw      32(%1),%%mm2\n\t"
            "paddw      40(%1),%%mm1\n\t"
            "packuswb   %%mm1,%%mm2\n\t"
            "movq       %%mm2,(%0)\n\t"
            "addl       %2,%0\n\t"

            "movq       (%0),%%mm1\n\t"
            "movq       %%mm1,%%mm2\n\t"
            "punpckhbw  %%mm7,%%mm1\n\t"
            "punpcklbw  %%mm7,%%mm2\n\t"
            "paddw      48(%1),%%mm2\n\t"
            "paddw      56(%1),%%mm1\n\t"
            "packuswb   %%mm1,%%mm2\n\t"
            "movq       %%mm2,(%0)\n\t"
            "addl       %2,%0\n\t"

            "movq       (%0),%%mm1\n\t"
            "movq       %%mm1,%%mm2\n\t"
            "punpckhbw  %%mm7,%%mm1\n\t"
            "punpcklbw  %%mm7,%%mm2\n\t"
            "paddw      64(%1),%%mm2\n\t"
            "paddw      72(%1),%%mm1\n\t"
            "packuswb   %%mm1,%%mm2\n\t"
            "movq       %%mm2,(%0)\n\t"
            "addl       %2,%0\n\t"

            "movq       (%0),%%mm1\n\t"
            "movq       %%mm1,%%mm2\n\t"
            "punpckhbw  %%mm7,%%mm1\n\t"
            "punpcklbw  %%mm7,%%mm2\n\t"
            "paddw      80(%1),%%mm2\n\t"
            "paddw      88(%1),%%mm1\n\t"
            "packuswb   %%mm1,%%mm2\n\t"
            "movq       %%mm2,(%0)\n\t"
            "addl       %2,%0\n\t"

            "movq       (%0),%%mm1\n\t"
            "movq       %%mm1,%%mm2\n\t"
            "punpckhbw  %%mm7,%%mm1\n\t"
            "punpcklbw  %%mm7,%%mm2\n\t"
            "paddw      96(%1),%%mm2\n\t"
            "paddw      104(%1),%%mm1\n\t"
            "packuswb   %%mm1,%%mm2\n\t"
            "movq       %%mm2,(%0)\n\t"
            "addl       %2,%0\n\t"

            "movq       (%0),%%mm1\n\t"
            "movq       %%mm1,%%mm2\n\t"
            "punpckhbw  %%mm7,%%mm1\n\t"
            "punpcklbw  %%mm7,%%mm2\n\t"
            "paddw      112(%1),%%mm2\n\t"
            "paddw      120(%1),%%mm1\n\t"
            "packuswb   %%mm1,%%mm2\n\t"
            "movq       %%mm2,(%0)\n\t"

            "emms"
             :"+r" (p_data): "r" (p_block),"r" (i_incr+8));
}
#endif


/*****************************************************************************
 * CopyBlock : copy a block
 *****************************************************************************/
#ifndef HAVE_MMX
static __inline__ void CopyBlock( vdec_thread_t * p_vdec, dctelem_t * p_block,
                                  yuv_data_t * p_data, int i_incr )
{
    int i_x, i_y;

    for( i_y = 0; i_y < 8; i_y++ )
    {
        for( i_x = 0; i_x < 8; i_x++ )
        {
            *p_data++ = p_vdec->pi_crop[*p_block++];
        }
        p_data += i_incr;
    }
}
#else
static  __inline__ void CopyBlock( vdec_thread_t * p_vdec, dctelem_t * p_block,
                                          yuv_data_t * p_data, int i_incr )
{
    asm __volatile__ (
            "movq         (%1),%%mm0\n\t"
            "packuswb   8(%1),%%mm0\n\t"
            "movq        %%mm0,(%0)\n\t"
            "addl           %2,%0\n\t"

            "movq        16(%1),%%mm0\n\t"
            "packuswb   24(%1),%%mm0\n\t"
            "movq        %%mm0,(%0)\n\t"
            "addl           %2,%0\n\t"

            "movq        32(%1),%%mm0\n\t"
            "packuswb   40(%1),%%mm0\n\t"
            "movq        %%mm0,(%0)\n\t"
            "addl           %2,%0\n\t"

            "movq        48(%1),%%mm0\n\t"
            "packuswb   56(%1),%%mm0\n\t"
            "movq        %%mm0,(%0)\n\t"
            "addl           %2,%0\n\t"

            "movq        64(%1),%%mm0\n\t"
            "packuswb   72(%1),%%mm0\n\t"
            "movq        %%mm0,(%0)\n\t"
            "addl           %2,%0\n\t"

            "movq        80(%1),%%mm0\n\t"
            "packuswb   88(%1),%%mm0\n\t"
            "movq        %%mm0,(%0)\n\t"
            "addl           %2,%0\n\t"

            "movq        96(%1),%%mm0\n\t"
            "packuswb   104(%1),%%mm0\n\t"
            "movq        %%mm0,(%0)\n\t"
            "addl           %2,%0\n\t"

            "movq        112(%1),%%mm0\n\t"
            "packuswb   120(%1),%%mm0\n\t"
            "movq        %%mm0,(%0)\n\t"
            "emms"
            :"+r" (p_data): "r" (p_block),"r" (i_incr+8));
}
#endif


/*****************************************************************************
 * vdec_DecodeMacroblock : decode a macroblock of a picture
 *****************************************************************************/
#define DECODEBLOCKS( OPBLOCK )                                         \
{                                                                       \
    int             i_b, i_mask;                                        \
                                                                        \
    i_mask = 1 << (3 + p_mb->i_chroma_nb_blocks);                       \
                                                                        \
    /* luminance */                                                     \
    for( i_b = 0; i_b < 4; i_b++, i_mask >>= 1 )                        \
    {                                                                   \
        if( p_mb->i_coded_block_pattern & i_mask )                      \
        {                                                               \
            /*                                                          \
             * Inverse DCT (ISO/IEC 13818-2 section Annex A)            \
             */                                                         \
            (p_mb->pf_idct[i_b])( p_vdec, p_mb->ppi_blocks[i_b],        \
                                  p_mb->pi_sparse_pos[i_b] );           \
                                                                        \
            /*                                                          \
             * Adding prediction and coefficient data (ISO/IEC 13818-2  \
             * section 7.6.8)                                           \
             */                                                         \
            OPBLOCK( p_vdec, p_mb->ppi_blocks[i_b],                     \
                     p_mb->p_data[i_b], p_mb->i_addb_l_stride );        \
        }                                                               \
    }                                                                   \
                                                                        \
    /* chrominance */                                                   \
    for( i_b = 4; i_b < 4 + p_mb->i_chroma_nb_blocks;                   \
         i_b++, i_mask >>= 1 )                                          \
    {                                                                   \
        if( p_mb->i_coded_block_pattern & i_mask )                      \
        {                                                               \
            /*                                                          \
             * Inverse DCT (ISO/IEC 13818-2 section Annex A)            \
             */                                                         \
            (p_mb->pf_idct[i_b])( p_vdec, p_mb->ppi_blocks[i_b],        \
                                  p_mb->pi_sparse_pos[i_b] );           \
                                                                        \
            /*                                                          \
             * Adding prediction and coefficient data (ISO/IEC 13818-2  \
             * section 7.6.8)                                           \
             */                                                         \
            OPBLOCK( p_vdec, p_mb->ppi_blocks[i_b],                     \
                     p_mb->p_data[i_b], p_mb->i_addb_c_stride );        \
        }                                                               \
    }                                                                   \
}

#ifdef VDEC_SMP
static __inline__ void vdec_DecodeMacroblock( vdec_thread_t *p_vdec, macroblock_t * p_mb )
#else
void vdec_DecodeMacroblock( vdec_thread_t *p_vdec, macroblock_t * p_mb )
#endif
{
    if( !(p_mb->i_mb_type & MB_INTRA) )
    {
        /*
         * Motion Compensation (ISO/IEC 13818-2 section 7.6)
         */
        if( p_mb->pf_motion == 0 )
        {
            intf_ErrMsg( "vdec error: pf_motion set to NULL\n" );
        }
        else
        {
            p_mb->pf_motion( p_mb );
        }

        DECODEBLOCKS( AddBlock )
    }
    else
    {
        DECODEBLOCKS( CopyBlock )
    }

    /*
     * Decoding is finished, release the macroblock and free
     * unneeded memory.
     */
    vpar_ReleaseMacroblock( &p_vdec->p_vpar->vfifo, p_mb );
}


/*****************************************************************************
 * RunThread: video decoder thread
 *****************************************************************************
 * Video decoder thread. This function does only return when the thread is
 * terminated.
 *****************************************************************************/
static void RunThread( vdec_thread_t *p_vdec )
{
    intf_DbgMsg("vdec debug: running video decoder thread (%p) (pid == %i)\n",
                p_vdec, getpid());

    /*
     * Initialize thread and free configuration
     */
    p_vdec->b_error = vdec_InitThread( p_vdec );
    if( p_vdec->b_error )
    {
        return;
    }
    p_vdec->b_run = 1;

    /*
     * Main loop - it is not executed if an error occured during
     * initialization
     */
    while( (!p_vdec->b_die) && (!p_vdec->b_error) )
    {
        macroblock_t *          p_mb;

        if( (p_mb = vpar_GetMacroblock( &p_vdec->p_vpar->vfifo )) != NULL )
        {
            vdec_DecodeMacroblock( p_vdec, p_mb );
        }
    }

    /*
     * Error loop
     */
    if( p_vdec->b_error )
    {
        ErrorThread( p_vdec );
    }

    /* End of thread */
    EndThread( p_vdec );
    p_vdec->b_run = 0;
}