h264dsp_mmx.c

这是著名的TCPMP播放器在WINDWOWS,和WINCE下编译通过的源程序.笔者对其中的LIBMAD库做了针对ARM MPU的优化. 并增加了词幕功能.

/*
 * Copyright (c) 2004-2005 Michael Niedermayer, Loren Merritt
 *
 * This library is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation; either
 * version 2 of the License, or (at your option) any later version.
 *
 * This library 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
 * Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public
 * License along with this library; if not, write to the Free Software
 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
 */


/***********************************/
/* IDCT */

/* in/out: mma=mma+mmb, mmb=mmb-mma */
#define SUMSUB_BA( a, b ) \
    "paddw "#b", "#a" \n\t"\
    "paddw "#b", "#b" \n\t"\
    "psubw "#a", "#b" \n\t"

#define SUMSUB_BADC( a, b, c, d ) \
    "paddw "#b", "#a" \n\t"\
    "paddw "#d", "#c" \n\t"\
    "paddw "#b", "#b" \n\t"\
    "paddw "#d", "#d" \n\t"\
    "psubw "#a", "#b" \n\t"\
    "psubw "#c", "#d" \n\t"

#define SUMSUBD2_AB( a, b, t ) \
    "movq  "#b", "#t" \n\t"\
    "psraw  $1 , "#b" \n\t"\
    "paddw "#a", "#b" \n\t"\
    "psraw  $1 , "#a" \n\t"\
    "psubw "#t", "#a" \n\t"

#define IDCT4_1D( s02, s13, d02, d13, t ) \
    SUMSUB_BA  ( s02, d02 )\
    SUMSUBD2_AB( s13, d13, t )\
    SUMSUB_BADC( d13, s02, s13, d02 )

#define SBUTTERFLY(a,b,t,n)\
    "movq " #a ", " #t "		\n\t" /* abcd */\
    "punpckl" #n " " #b ", " #a "	\n\t" /* aebf */\
    "punpckh" #n " " #b ", " #t "	\n\t" /* cgdh */\

#define TRANSPOSE4(a,b,c,d,t)\
    SBUTTERFLY(a,b,t,wd) /* a=aebf t=cgdh */\
    SBUTTERFLY(c,d,b,wd) /* c=imjn b=kolp */\
    SBUTTERFLY(a,c,d,dq) /* a=aeim d=bfjn */\
    SBUTTERFLY(t,b,c,dq) /* t=cgko c=dhlp */

#define STORE_DIFF_4P( p, t, z ) \
    "psraw      $6,     "#p" \n\t"\
    "movd       (%0),   "#t" \n\t"\
    "punpcklbw "#z",    "#t" \n\t"\
    "paddsw    "#t",    "#p" \n\t"\
    "packuswb  "#z",    "#p" \n\t"\
    "movd      "#p",    (%0) \n\t"

void ff_h264_idct_add_mmx2(uint8_t *dst, int16_t *block, int stride)
{
    /* Load dct coeffs */
    asm volatile(
        "movq   (%0), %%mm0 \n\t"
        "movq  8(%0), %%mm1 \n\t"
        "movq 16(%0), %%mm2 \n\t"
        "movq 24(%0), %%mm3 \n\t"
    :: "r"(block) );

    asm volatile(
        /* mm1=s02+s13  mm2=s02-s13  mm4=d02+d13  mm0=d02-d13 */
        IDCT4_1D( %%mm2, %%mm1, %%mm0, %%mm3, %%mm4 )

        "movq      %0,    %%mm6 \n\t"
        /* in: 1,4,0,2  out: 1,2,3,0 */
        TRANSPOSE4( %%mm3, %%mm1, %%mm0, %%mm2, %%mm4 )

        "paddw     %%mm6, %%mm3 \n\t"

        /* mm2=s02+s13  mm3=s02-s13  mm4=d02+d13  mm1=d02-d13 */
        IDCT4_1D( %%mm4, %%mm2, %%mm3, %%mm0, %%mm1 )

        "pxor %%mm7, %%mm7    \n\t"
    :: "m"(ff_pw_32));

    asm volatile(
    STORE_DIFF_4P( %%mm0, %%mm1, %%mm7)
        "add %1, %0             \n\t"
    STORE_DIFF_4P( %%mm2, %%mm1, %%mm7)
        "add %1, %0             \n\t"
    STORE_DIFF_4P( %%mm3, %%mm1, %%mm7)
        "add %1, %0             \n\t"
    STORE_DIFF_4P( %%mm4, %%mm1, %%mm7)
        : "+r"(dst)
        : "r" ((long)stride)
    );
}


/***********************************/
/* deblocking */

// out: o = |x-y|>a
// clobbers: t
#define DIFF_GT_MMX(x,y,a,o,t)\
    "movq     "#y", "#t"  \n\t"\
    "movq     "#x", "#o"  \n\t"\
    "psubusb  "#x", "#t"  \n\t"\
    "psubusb  "#y", "#o"  \n\t"\
    "por      "#t", "#o"  \n\t"\
    "psubusb  "#a", "#o"  \n\t"

// in: mm0=p1 mm1=p0 mm2=q0 mm3=q1
// out: mm5=beta-1, mm7=mask
// clobbers: mm4,mm6
#define H264_DEBLOCK_MASK(alpha1, beta1) \
    "pshufw $0, "#alpha1", %%mm4 \n\t"\
    "pshufw $0, "#beta1 ", %%mm5 \n\t"\
    "packuswb  %%mm4, %%mm4      \n\t"\
    "packuswb  %%mm5, %%mm5      \n\t"\
    DIFF_GT_MMX(%%mm1, %%mm2, %%mm4, %%mm7, %%mm6) /* |p0-q0| > alpha-1 */\
    DIFF_GT_MMX(%%mm0, %%mm1, %%mm5, %%mm4, %%mm6) /* |p1-p0| > beta-1 */\
    "por       %%mm4, %%mm7      \n\t"\
    DIFF_GT_MMX(%%mm3, %%mm2, %%mm5, %%mm4, %%mm6) /* |q1-q0| > beta-1 */\
    "por       %%mm4, %%mm7      \n\t"\
    "pxor      %%mm6, %%mm6      \n\t"\
    "pcmpeqb   %%mm6, %%mm7      \n\t"

// in: mm0=p1 mm1=p0 mm2=q0 mm3=q1 mm7=(tc&mask)
// out: mm1=p0' mm2=q0'
// clobbers: mm0,3-6
#define H264_DEBLOCK_P0_Q0(pb_01, pb_3f)\
        /* a = q0^p0^((p1-q1)>>2) */\
        "movq    %%mm0, %%mm4  \n\t"\
        "psubb   %%mm3, %%mm4  \n\t"\
        "psrlw   $2,    %%mm4  \n\t"\
        "pxor    %%mm1, %%mm4  \n\t"\
        "pxor    %%mm2, %%mm4  \n\t"\
        /* b = p0^(q1>>2) */\
        "psrlw   $2,    %%mm3  \n\t"\
        "pand "#pb_3f", %%mm3  \n\t"\
        "movq    %%mm1, %%mm5  \n\t"\
        "pxor    %%mm3, %%mm5  \n\t"\
        /* c = q0^(p1>>2) */\
        "psrlw   $2,    %%mm0  \n\t"\
        "pand "#pb_3f", %%mm0  \n\t"\
        "movq    %%mm2, %%mm6  \n\t"\
        "pxor    %%mm0, %%mm6  \n\t"\
        /* d = (c^b) & ~(b^a) & 1 */\
        "pxor    %%mm5, %%mm6  \n\t"\
        "pxor    %%mm4, %%mm5  \n\t"\
        "pandn   %%mm6, %%mm5  \n\t"\
        "pand "#pb_01", %%mm5  \n\t"\
        /* delta = (avg(q0, p1>>2) + (d&a))
         *       - (avg(p0, q1>>2) + (d&~a)) */\
        "pavgb   %%mm2, %%mm0  \n\t"\
        "pand    %%mm5, %%mm4  \n\t"\
        "paddusb %%mm4, %%mm0  \n\t"\
        "pavgb   %%mm1, %%mm3  \n\t"\
        "pxor    %%mm5, %%mm4  \n\t"\
        "paddusb %%mm4, %%mm3  \n\t"\
        /* p0 += clip(delta, -tc0, tc0)
         * q0 -= clip(delta, -tc0, tc0) */\
        "movq    %%mm0, %%mm4  \n\t"\
        "psubusb %%mm3, %%mm0  \n\t"\
        "psubusb %%mm4, %%mm3  \n\t"\
        "pminub  %%mm7, %%mm0  \n\t"\
        "pminub  %%mm7, %%mm3  \n\t"\
        "paddusb %%mm0, %%mm1  \n\t"\
        "paddusb %%mm3, %%mm2  \n\t"\
        "psubusb %%mm3, %%mm1  \n\t"\
        "psubusb %%mm0, %%mm2  \n\t"

// in: mm0=p1 mm1=p0 mm2=q0 mm3=q1 mm7=(tc&mask) %8=mm_bone
// out: (q1addr) = clip( (q2+((p0+q0+1)>>1))>>1, q1-tc0, q1+tc0 )
// clobbers: q2, tmp, tc0
#define H264_DEBLOCK_Q1(p1, q2, q2addr, q1addr, tc0, tmp)\
        "movq     %%mm1,  "#tmp"   \n\t"\
        "pavgb    %%mm2,  "#tmp"   \n\t"\
        "pavgb    "#tmp", "#q2"    \n\t" /* avg(p2,avg(p0,q0)) */\
        "pxor   "q2addr", "#tmp"   \n\t"\
        "pand     %8,     "#tmp"   \n\t" /* (p2^avg(p0,q0))&1 */\
        "psubusb  "#tmp", "#q2"    \n\t" /* (p2+((p0+q0+1)>>1))>>1 */\
        "movq     "#p1",  "#tmp"   \n\t"\
        "psubusb  "#tc0", "#tmp"   \n\t"\
        "paddusb  "#p1",  "#tc0"   \n\t"\
        "pmaxub   "#tmp", "#q2"    \n\t"\
        "pminub   "#tc0", "#q2"    \n\t"\
        "movq     "#q2",  "q1addr" \n\t"

static inline void h264_loop_filter_luma_mmx2(uint8_t *pix, int stride, int alpha1, int beta1, int8_t *tc0)
{
    uint64_t tmp0;
    uint64_t tc = (uint8_t)tc0[1]*0x01010000 | (uint8_t)tc0[0]*0x0101;
    // with luma, tc0=0 doesn't mean no filtering, so we need a separate input mask
    uint32_t mask[2] = { (tc0[0]>=0)*0xffffffff, (tc0[1]>=0)*0xffffffff };

    asm volatile(
        "movq    (%1,%3), %%mm0    \n\t" //p1
        "movq    (%1,%3,2), %%mm1  \n\t" //p0
        "movq    (%2),    %%mm2    \n\t" //q0
        "movq    (%2,%3), %%mm3    \n\t" //q1
        H264_DEBLOCK_MASK(%6, %7)
        "pand     %5,     %%mm7    \n\t"
        "movq     %%mm7,  %0       \n\t"

        /* filter p1 */
        "movq     (%1),   %%mm3    \n\t" //p2
        DIFF_GT_MMX(%%mm1, %%mm3, %%mm5, %%mm6, %%mm4) // |p2-p0|>beta-1
        "pandn    %%mm7,  %%mm6    \n\t"
        "pcmpeqb  %%mm7,  %%mm6    \n\t"
        "pand     %%mm7,  %%mm6    \n\t" // mask & |p2-p0|<beta
        "pshufw  $80, %4, %%mm4    \n\t"
        "pand     %%mm7,  %%mm4    \n\t" // mask & tc0
        "movq     %8,     %%mm7    \n\t"
        "pand     %%mm6,  %%mm7    \n\t" // mask & |p2-p0|<beta & 1
        "pand     %%mm4,  %%mm6    \n\t" // mask & |p2-p0|<beta & tc0
        "paddb    %%mm4,  %%mm7    \n\t" // tc++
        H264_DEBLOCK_Q1(%%mm0, %%mm3, "(%1)", "(%1,%3)", %%mm6, %%mm4)

        /* filter q1 */
        "movq    (%2,%3,2), %%mm4  \n\t" //q2
        DIFF_GT_MMX(%%mm2, %%mm4, %%mm5, %%mm6, %%mm3) // |q2-q0|>beta-1
        "pandn    %0,     %%mm6    \n\t"
        "pcmpeqb  %0,     %%mm6    \n\t"
        "pand     %0,     %%mm6    \n\t"
        "pshufw  $80, %4, %%mm5    \n\t"
        "pand     %%mm6,  %%mm5    \n\t"
        "pand     %8,     %%mm6    \n\t"
        "paddb    %%mm6,  %%mm7    \n\t"
        "movq    (%2,%3), %%mm3    \n\t"
        H264_DEBLOCK_Q1(%%mm3, %%mm4, "(%2,%3,2)", "(%2,%3)", %%mm5, %%mm6)

        /* filter p0, q0 */
        H264_DEBLOCK_P0_Q0(%8, %9)
        "movq      %%mm1, (%1,%3,2) \n\t"
        "movq      %%mm2, (%2)      \n\t"

        : "=m"(tmp0)
        : "r"(pix-3*stride), "r"(pix), "r"((long)stride),
          "m"(tc), "m"(*(uint64_t*)mask), "m"(alpha1), "m"(beta1),
          "m"(mm_bone), "m"(ff_pb_3F)
    );
}

static void h264_v_loop_filter_luma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0)
{
    if((tc0[0] & tc0[1]) >= 0)
        h264_loop_filter_luma_mmx2(pix, stride, alpha-1, beta-1, tc0);
    if((tc0[2] & tc0[3]) >= 0)
        h264_loop_filter_luma_mmx2(pix+8, stride, alpha-1, beta-1, tc0+2);
}
static void h264_h_loop_filter_luma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0)
{
    //FIXME: could cut some load/stores by merging transpose with filter
    // also, it only needs to transpose 6x8
    uint8_t trans[8*8];
    int i;
    for(i=0; i<2; i++, pix+=8*stride, tc0+=2) {
        if((tc0[0] & tc0[1]) < 0)
            continue;
        transpose4x4(trans,       pix-4,          8, stride);
        transpose4x4(trans  +4*8, pix,            8, stride);
        transpose4x4(trans+4,     pix-4+4*stride, 8, stride);
        transpose4x4(trans+4+4*8, pix  +4*stride, 8, stride);
        h264_loop_filter_luma_mmx2(trans+4*8, 8, alpha-1, beta-1, tc0);
        transpose4x4(pix-2,          trans  +2*8, stride, 8);
        transpose4x4(pix-2+4*stride, trans+4+2*8, stride, 8);
    }
}

static inline void h264_loop_filter_chroma_mmx2(uint8_t *pix, int stride, int alpha1, int beta1, int8_t *tc0)
{
    asm volatile(
        "movq    (%0),    %%mm0     \n\t" //p1
        "movq    (%0,%2), %%mm1     \n\t" //p0
        "movq    (%1),    %%mm2     \n\t" //q0
        "movq    (%1,%2), %%mm3     \n\t" //q1
        H264_DEBLOCK_MASK(%4, %5)
        "movd      %3,    %%mm6     \n\t"
        "punpcklbw %%mm6, %%mm6     \n\t"
        "pand      %%mm6, %%mm7     \n\t" // mm7 = tc&mask
        H264_DEBLOCK_P0_Q0(%6, %7)
        "movq      %%mm1, (%0,%2)   \n\t"
        "movq      %%mm2, (%1)      \n\t"

        :: "r"(pix-2*stride), "r"(pix), "r"((long)stride),
           "r"(*(uint32_t*)tc0),
           "m"(alpha1), "m"(beta1), "m"(mm_bone), "m"(ff_pb_3F)
    );
}

static void h264_v_loop_filter_chroma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0)
{
    h264_loop_filter_chroma_mmx2(pix, stride, alpha-1, beta-1, tc0);
}