postprocessing_mmxext.c

来自「VLC媒体播放程序」· C语言 代码 · 共 918 行 · 第 1/3 页

/*****************************************************************************
 * postprocessing_mmxext.c: Post Processing plugin MMXEXT
 *****************************************************************************
 * Copyright (C) 2001 VideoLAN
 * $Id: postprocessing_mmxext.c,v 1.5 2002/12/18 14:17:10 sam Exp $
 *
 * Authors: Laurent Aimar <[e-mail]>
 *
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation; either version 2 of the License, or
 * (at your option) any later version.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with this program; if not, write to the Free Software
 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA  02111, USA.
 *****************************************************************************/

#include <vlc/vlc.h> /* only use uint8_t, uint32_t .... */

#include "postprocessing.h"
#include "postprocessing_common.h"

/*****************************************************************************
 *
 * Internals functions common to pp_Deblock_V and pp_Deblock_H
 *
 *****************************************************************************/

/*****************************************************************************
 * MMX stuff
 *****************************************************************************/


/* XXX PP_THR1 need to be defined as ULL */

/* Use same things as in idct but how it work ? */
#define UNUSED_LONGLONG( foo ) \
    static const unsigned long long foo __asm__ (#foo)  __attribute__((unused))

/* to calculate isDC_mode for mmx */
UNUSED_LONGLONG( mmx_thr1 ) = ( PP_THR1 << 56 )|
                              ( PP_THR1 << 48 )|
                              ( PP_THR1 << 40 )|
                              ( PP_THR1 << 32 )|
                              ( PP_THR1 << 24 )|
                              ( PP_THR1 << 16 )|
                              ( PP_THR1 <<  8 )|
                              ( PP_THR1 );

UNUSED_LONGLONG( mmx_127_thr1 ) = ( ( 127ULL - PP_THR1 ) << 56 )|
                                  ( ( 127ULL - PP_THR1 ) << 48 )|
                                  ( ( 127ULL - PP_THR1 ) << 40 )|
                                  ( ( 127ULL - PP_THR1 ) << 32 )|
                                  ( ( 127ULL - PP_THR1 ) << 24 )|
                                  ( ( 127ULL - PP_THR1 ) << 16 )|
                                  ( ( 127ULL - PP_THR1 ) <<  8 )|
                                  ( ( 127ULL - PP_THR1 ) );

UNUSED_LONGLONG( mmx_127_2xthr1_1 ) = ( ( 127ULL - PP_2xTHR1 -1) << 56 )|
                                    ( ( 127ULL - PP_2xTHR1 -1 ) << 48 )|
                                    ( ( 127ULL - PP_2xTHR1 -1 ) << 40 )|
                                    ( ( 127ULL - PP_2xTHR1 -1 ) << 32 )|
                                    ( ( 127ULL - PP_2xTHR1 -1 ) << 24 )|
                                    ( ( 127ULL - PP_2xTHR1 -1 ) << 16 )|
                                    ( ( 127ULL - PP_2xTHR1 -1 ) <<  8 )|
                                    ( ( 127ULL - PP_2xTHR1 -1 ) );

UNUSED_LONGLONG( mmx_m2_5_m5_2 ) = 0xfffe0005fffb0002ULL;


/* find min bytes from r ans set it in r, t is destroyed */
#define MMXEXT_GET_PMIN( r, t ) \
   "movq      " #r ",     " #t "                                \n" \
   "psrlq       $8,       " #t "                                \n" \
   "pminub    " #t ",     " #r "                                \n" \
   "pshufw $0xf5, " #r ", " #t " #instead of shift with tmp reg \n" \
   "pminub    " #t ",     " #r "                                \n" \
   "pshufw $0xfe, " #r ", " #t "                                \n" \
   "pminub    " #t ",     " #r "                                \n"

 /* find mzx bytes from r ans set it in r, t is destroyed */
#define MMXEXT_GET_PMAX( r, t ) \
   "movq      " #r ",     " #t "                                \n" \
   "psrlq       $8,       " #t "                                \n" \
   "pmaxub    " #t ",     " #r "                                \n" \
   "pshufw $0xf5, " #r ", " #t "                                \n" \
   "pmaxub    " #t ",     " #r "                                \n" \
   "pshufw $0xfe, " #r ", " #t "                                \n" \
   "pmaxub    " #t ",     " #r "                                \n"



#define MMXEXT_GET_LMINMAX( s, m, M, t ) \
    "movq   " #s ",        " #t "   \n" \
    "pminub " #t ",        " #m "   \n" \
    "pmaxub " #t ",        " #M "   \n"

/* Some tips for MMX

    * |a-b| :
        d1 = a - b with unsigned saturate
        d2 = b - a  with ...
        |a-b| = d1 | d2


*/

/****************************************************************************
 * pp_deblock_isDC_mode : Check if we will use DC mode or Default mode
 ****************************************************************************
 * Use constant PP_THR1 and PP_THR2 ( PP_2xTHR1 )
 *
 * Called for for each pixel on a boundary block when doing deblocking
 *  so need to be fast ...
 *
 ****************************************************************************/
static inline int pp_deblock_isDC_mode( uint8_t *p_v )
{
    unsigned int i_eq_cnt;


    /* algo :
       x = v[i] - v[i+1] without signed saturation
        ( XXX see if there is'nt problem, but can't be with signed
        sat because pixel will be saturate :(
       so x within [-128, 127] and we have to test if it fit in [-M, M]
       we add 127-M with wrap around -> good value fit in [ 127-2*M, 127]
       and if x >= 127 - 2 * M ie x > 127 -2*M - 1 value is good
    */
#if 0
    __asm__ __volatile__ (
   "                                #* Do (v0-v1) to (v7-v8)            \n"
   "movq      (%1),         %%mm1   #  load v0->v7                      \n"
   "movq      1(%1),        %%mm2   #  load v1->v8                      \n"
   "psubb    %%mm2,         %%mm1   #  v[i]-v[i+1]                      \n"
   "paddb     mmx_127_thr1, %%mm1   #  + 127-THR1 with wrap             \n"
   "pcmpgtb   mmx_127_2xthr1_1, %%mm1 #  >  127 -2*thr1 - 1             \n"
   "pxor      %%mm0,        %%mm0   # mm0 = 0                           \n"
   "psadbw    %%mm1,        %%mm0                                       \n"
   "movd      %%mm0,        %0      #                                   \n"
   "negl      %0                                                        \n"
   "andl      $255,         %0"

       : "=r"(i_eq_cnt) : "r" (p_v) );
#endif
     __asm__ __volatile__ (
   "                                #* Do (v0-v1) to (v7-v8)            \n"
   "movq      (%1),         %%mm1   #  load v0->v7                      \n"
   "pxor      %%mm0,        %%mm0   # mm0 = 0                           \n"
   "movq      1(%1),        %%mm2   #  load v1->v8                      \n"
   "psubb    %%mm2,         %%mm1   #  v[i]-v[i+1]                      \n"
   "paddb     mmx_127_thr1, %%mm1   #  + 127-THR1 with wrap             \n"
   "pcmpgtb   mmx_127_2xthr1_1, %%mm1 #  >  127 -2*thr1 - 1             \n"
   "psadbw    %%mm1,        %%mm0                                       \n"
   "movd      %%mm0,        %0      #                                   \n"
   "negl      %0"

       : "=r"(i_eq_cnt) : "r" (p_v) );

    /* last test, hey, 9 don't fit in MMX */
    if((  ( p_v[8] - p_v[9] + PP_THR1 )&0xffff )<= PP_2xTHR1 )
    {
         i_eq_cnt++;
    }

#if 0
    /* algo :  if ( | v[i] -v[i+1] | <= PP_THR1 ) { i_eq_cnt++; } */
    i_eq_cnt = 0;

    for( i =0; i < 9; i++ )
    {
        if((  ( p_v[i] - p_v[i+1] + PP_THR1 )&0xffff )<= PP_2xTHR1 )
        {
            i_eq_cnt++;
        }
    }
#endif

    return( (i_eq_cnt >= PP_THR2 ) ? 1 : 0 );
}

static inline int pp_deblock_isMinMaxOk( uint8_t *p_v, int i_QP )
{
    int i_range;

    __asm__ __volatile__ (
   "movq        1(%1),      %%mm0   # 8 bytes                   \n"
   "movq        %%mm0,      %%mm1                               \n"
    MMXEXT_GET_PMIN( %%mm0, %%mm7 )
    MMXEXT_GET_PMAX( %%mm1, %%mm7 )
   "psubd       %%mm0,      %%mm1   # max - min                 \n"
   "movd        %%mm1,      %0                                  \n"
   "andl        $255,       %0" : "=r"(i_range) : "r"(p_v) );

#if 0
    int i_max, i_min;
    int i;

    i_min = i_max = p_v[1];
    for( i = 2; i < 9; i++ )
    {
        if( i_max < p_v[i] ) i_max = p_v[i];
        if( i_min > p_v[i] ) i_min = p_v[i];
    }
    i_range = i_max - i_min;
#endif

    return( i_range< 2*i_QP ? 1 : 0 );
}


static inline void pp_deblock_DefaultMode( uint8_t i_v[10], int i_stride,
                                      int i_QP )
{
    int d, i_delta;
    int a3x0, a3x0_, a3x1, a3x2;
    int b_neg;

    /* d = CLIP( 5(a3x0' - a3x0)//8, 0, (v4-v5)/2 ).d( abs(a3x0) < QP ) */

    /* First calculate a3x0 */
    __asm__ __volatile__ (
   "pxor    %%mm7,  %%mm7           # mm7 = 0          \n"
   "movq    mmx_m2_5_m5_2, %%mm6    # mm6 =(2,-5,5,-2) \n"
   "movd    3(%1),  %%mm0           \n"
   "punpcklbw %%mm7,%%mm0           \n"
   "pmaddwd %%mm6,  %%mm0           \n"
   "pshufw  $0xfe,  %%mm0, %%mm1    \n"
   "paddd   %%mm1,  %%mm0           \n"
   "movd    %%mm0,  %0" : "=r"(a3x0) :"r"(i_v) );
#if 0
    a3x0 = 2 * ( i_v[3] - i_v[6] ) + 5 *( i_v[5] - i_v[4] );
#endif

    if( a3x0 < 0 )
    {
        b_neg = 1;
        a3x0  = -a3x0;
    }
    else
    {
        b_neg = 0;
    }
    /* XXX Now a3x0 is abs( a3x0 ) */
    if( ( a3x0 < 8 * i_QP )&&( a3x0 != 0 ) ) /* |a3x0| < 8*i_QP */
    {
        /* calculate a3x1 et a3x2 */
        __asm__ __volatile__ (
       "                                # mm7 = 0                   \n"
       "                                # mm6 = ( 2, -5, 5, -2 )    \n"
       "movd    1(%2),  %%mm0           \n"
       "movd    5(%2),  %%mm2           \n"
       "punpcklbw %%mm7,%%mm0           \n"
       "punpcklbw %%mm7,%%mm2           \n"
       "pmaddwd %%mm6,  %%mm0           \n"
       "pmaddwd %%mm6,  %%mm2           \n"
       "pshufw  $0xfe,  %%mm0, %%mm1    \n"
       "paddd   %%mm1,  %%mm0           # mm0 = a3x1    \n"
       "movd    %%mm0,  %0              \n"
       "pshufw  $0xfe,  %%mm2, %%mm1    \n"
       "paddd   %%mm1,  %%mm2           # mm2 = a3x2    \n"
       "movd    %%mm2,  %1              \n"
        : "=r"(a3x1), "=r"(a3x2) : "r"(i_v) );
#if 0
        a3x1 = 2 * ( i_v[1] - i_v[4] ) + 5 * ( i_v[3] - i_v[2] );
        a3x2 = 2 * ( i_v[5] - i_v[8] ) + 5 * ( i_v[7] - i_v[6] );
#endif

        if( a3x1 < 0) a3x1 = -a3x1; /* abs( a3x1 ) */
        if( a3x2 < 0) a3x2 = -a3x2; /* abs( a3x2 ) */

        a3x0_ = PP_MIN3( a3x0, a3x1, a3x2 );

        d = 5 *( a3x0 - a3x0_ ) / 8; /* always > 0 */

        i_delta = ( i_v[4] - i_v[5] ) / 2;
        /* clip into [0, i_delta] or [i_delta, 0] */
        if( i_delta < 0 )
        {
            if( !b_neg ) /* since true d has sgn(d) = - sgn( a3x0 ) */
            {
                d = -d;
                if( d < i_delta ) d = i_delta;
                i_v[4] -= d;
                i_v[5] += d;
            }
        }
        else
        {
            if( b_neg )
            {
                if( d > i_delta ) d = i_delta;
                i_v[4] -= d;
                i_v[5] += d;
            }
        }
    }
}