postprocess_template.c

ffmpeg移植到symbian的全部源代码

        PAVGB(%%mm7, %%mm6)                           //      1 12        /4
        PAVGB(%%mm2, %%mm6)                           //   1122424        /4
        "movq %%mm6, (%%"REG_c")                \n\t" //       X
        // mm0=8 mm3=6(11) mm4=5 mm5=4(11) mm7=9
        PAVGB(%%mm7, %%mm5)                           //    11   2        /4
        PAVGB(%%mm7, %%mm5)                           //    11   6        /8

        PAVGB(%%mm3, %%mm0)                           //      112        /4
        PAVGB(%%mm0, %%mm5)                           //    112246        /16
        "movq %%mm5, (%%"REG_a", %1, 4)         \n\t" //        X
        "sub %1, %0                             \n\t"

        :
        : "r" (src), "r" ((long)stride), "m" (c->pQPb)
        : "%"REG_a, "%"REG_c
    );
#else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
    const int l1= stride;
    const int l2= stride + l1;
    const int l3= stride + l2;
    const int l4= stride + l3;
    const int l5= stride + l4;
    const int l6= stride + l5;
    const int l7= stride + l6;
    const int l8= stride + l7;
    const int l9= stride + l8;
    int x;
    src+= stride*3;
    for(x=0; x<BLOCK_SIZE; x++){
        const int first= FFABS(src[0] - src[l1]) < c->QP ? src[0] : src[l1];
        const int last= FFABS(src[l8] - src[l9]) < c->QP ? src[l9] : src[l8];

        int sums[10];
        sums[0] = 4*first + src[l1] + src[l2] + src[l3] + 4;
        sums[1] = sums[0] - first  + src[l4];
        sums[2] = sums[1] - first  + src[l5];
        sums[3] = sums[2] - first  + src[l6];
        sums[4] = sums[3] - first  + src[l7];
        sums[5] = sums[4] - src[l1] + src[l8];
        sums[6] = sums[5] - src[l2] + last;
        sums[7] = sums[6] - src[l3] + last;
        sums[8] = sums[7] - src[l4] + last;
        sums[9] = sums[8] - src[l5] + last;

        src[l1]= (sums[0] + sums[2] + 2*src[l1])>>4;
        src[l2]= (sums[1] + sums[3] + 2*src[l2])>>4;
        src[l3]= (sums[2] + sums[4] + 2*src[l3])>>4;
        src[l4]= (sums[3] + sums[5] + 2*src[l4])>>4;
        src[l5]= (sums[4] + sums[6] + 2*src[l5])>>4;
        src[l6]= (sums[5] + sums[7] + 2*src[l6])>>4;
        src[l7]= (sums[6] + sums[8] + 2*src[l7])>>4;
        src[l8]= (sums[7] + sums[9] + 2*src[l8])>>4;

        src++;
    }
#endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
}
#endif //HAVE_ALTIVEC

#if 0
/**
 * Experimental implementation of the filter (Algorithm 1) described in a paper from Ramkishor & Karandikar
 * values are correctly clipped (MMX2)
 * values are wraparound (C)
 * Conclusion: It is fast, but introduces ugly horizontal patterns
 * if there is a continuous gradient.
        0 8 16 24
        x = 8
        x/2 = 4
        x/8 = 1
        1 12 12 23
 */
static inline void RENAME(vertRK1Filter)(uint8_t *src, int stride, int QP)
{
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
    src+= stride*3;
// FIXME rounding
    asm volatile(
        "pxor %%mm7, %%mm7                      \n\t" // 0
        "movq "MANGLE(b80)", %%mm6              \n\t" // MIN_SIGNED_BYTE
        "leal (%0, %1), %%"REG_a"               \n\t"
        "leal (%%"REG_a", %1, 4), %%"REG_c"     \n\t"
//      0       1       2       3       4       5       6       7       8       9
//      %0      eax     eax+%1  eax+2%1 %0+4%1  ecx     ecx+%1  ecx+2%1 %0+8%1  ecx+4%1
        "movq "MANGLE(pQPb)", %%mm0             \n\t" // QP,..., QP
        "movq %%mm0, %%mm1                      \n\t" // QP,..., QP
        "paddusb "MANGLE(b02)", %%mm0           \n\t"
        "psrlw $2, %%mm0                        \n\t"
        "pand "MANGLE(b3F)", %%mm0              \n\t" // QP/4,..., QP/4
        "paddusb %%mm1, %%mm0                   \n\t" // QP*1.25 ...
        "movq (%0, %1, 4), %%mm2                \n\t" // line 4
        "movq (%%"REG_c"), %%mm3                \n\t" // line 5
        "movq %%mm2, %%mm4                      \n\t" // line 4
        "pcmpeqb %%mm5, %%mm5                   \n\t" // -1
        "pxor %%mm2, %%mm5                      \n\t" // -line 4 - 1
        PAVGB(%%mm3, %%mm5)
        "paddb %%mm6, %%mm5                     \n\t" // (l5-l4)/2
        "psubusb %%mm3, %%mm4                   \n\t"
        "psubusb %%mm2, %%mm3                   \n\t"
        "por %%mm3, %%mm4                       \n\t" // |l4 - l5|
        "psubusb %%mm0, %%mm4                   \n\t"
        "pcmpeqb %%mm7, %%mm4                   \n\t"
        "pand %%mm4, %%mm5                      \n\t" // d/2

//        "paddb %%mm6, %%mm2                     \n\t" // line 4 + 0x80
        "paddb %%mm5, %%mm2                     \n\t"
//        "psubb %%mm6, %%mm2                     \n\t"
        "movq %%mm2, (%0,%1, 4)                 \n\t"

        "movq (%%"REG_c"), %%mm2                \n\t"
//        "paddb %%mm6, %%mm2                     \n\t" // line 5 + 0x80
        "psubb %%mm5, %%mm2                     \n\t"
//        "psubb %%mm6, %%mm2                     \n\t"
        "movq %%mm2, (%%"REG_c")                \n\t"

        "paddb %%mm6, %%mm5                     \n\t"
        "psrlw $2, %%mm5                        \n\t"
        "pand "MANGLE(b3F)", %%mm5              \n\t"
        "psubb "MANGLE(b20)", %%mm5             \n\t" // (l5-l4)/8

        "movq (%%"REG_a", %1, 2), %%mm2         \n\t"
        "paddb %%mm6, %%mm2                     \n\t" // line 3 + 0x80
        "paddsb %%mm5, %%mm2                    \n\t"
        "psubb %%mm6, %%mm2                     \n\t"
        "movq %%mm2, (%%"REG_a", %1, 2)         \n\t"

        "movq (%%"REG_c", %1), %%mm2            \n\t"
        "paddb %%mm6, %%mm2                     \n\t" // line 6 + 0x80
        "psubsb %%mm5, %%mm2                    \n\t"
        "psubb %%mm6, %%mm2                     \n\t"
        "movq %%mm2, (%%"REG_c", %1)            \n\t"

        :
        : "r" (src), "r" ((long)stride)
        : "%"REG_a, "%"REG_c
    );
#else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
    const int l1= stride;
    const int l2= stride + l1;
    const int l3= stride + l2;
    const int l4= stride + l3;
    const int l5= stride + l4;
    const int l6= stride + l5;
//    const int l7= stride + l6;
//    const int l8= stride + l7;
//    const int l9= stride + l8;
    int x;
    const int QP15= QP + (QP>>2);
    src+= stride*3;
    for(x=0; x<BLOCK_SIZE; x++){
        const int v = (src[x+l5] - src[x+l4]);
        if(FFABS(v) < QP15){
            src[x+l3] +=v>>3;
            src[x+l4] +=v>>1;
            src[x+l5] -=v>>1;
            src[x+l6] -=v>>3;
        }
    }

#endif //defined (HAVE_MMX2) || defined (HAVE_3DNOW)
}
#endif //0

/**
 * Experimental Filter 1
 * will not damage linear gradients
 * Flat blocks should look like they were passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter
 * can only smooth blocks at the expected locations (it cannot smooth them if they did move)
 * MMX2 version does correct clipping C version does not
 */
static inline void RENAME(vertX1Filter)(uint8_t *src, int stride, PPContext *co)
{
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
    src+= stride*3;

    asm volatile(
        "pxor %%mm7, %%mm7                      \n\t" // 0
        "lea (%0, %1), %%"REG_a"                \n\t"
        "lea (%%"REG_a", %1, 4), %%"REG_c"      \n\t"
//      0       1       2       3       4       5       6       7       8       9
//      %0      eax     eax+%1  eax+2%1 %0+4%1  ecx     ecx+%1  ecx+2%1 %0+8%1  ecx+4%1
        "movq (%%"REG_a", %1, 2), %%mm0         \n\t" // line 3
        "movq (%0, %1, 4), %%mm1                \n\t" // line 4
        "movq %%mm1, %%mm2                      \n\t" // line 4
        "psubusb %%mm0, %%mm1                   \n\t"
        "psubusb %%mm2, %%mm0                   \n\t"
        "por %%mm1, %%mm0                       \n\t" // |l2 - l3|
        "movq (%%"REG_c"), %%mm3                \n\t" // line 5
        "movq (%%"REG_c", %1), %%mm4            \n\t" // line 6
        "movq %%mm3, %%mm5                      \n\t" // line 5
        "psubusb %%mm4, %%mm3                   \n\t"
        "psubusb %%mm5, %%mm4                   \n\t"
        "por %%mm4, %%mm3                       \n\t" // |l5 - l6|
        PAVGB(%%mm3, %%mm0)                           // (|l2 - l3| + |l5 - l6|)/2
        "movq %%mm2, %%mm1                      \n\t" // line 4
        "psubusb %%mm5, %%mm2                   \n\t"
        "movq %%mm2, %%mm4                      \n\t"
        "pcmpeqb %%mm7, %%mm2                   \n\t" // (l4 - l5) <= 0 ? -1 : 0
        "psubusb %%mm1, %%mm5                   \n\t"
        "por %%mm5, %%mm4                       \n\t" // |l4 - l5|
        "psubusb %%mm0, %%mm4                   \n\t" //d = MAX(0, |l4-l5| - (|l2-l3| + |l5-l6|)/2)
        "movq %%mm4, %%mm3                      \n\t" // d
        "movq %2, %%mm0                         \n\t"
        "paddusb %%mm0, %%mm0                   \n\t"
        "psubusb %%mm0, %%mm4                   \n\t"
        "pcmpeqb %%mm7, %%mm4                   \n\t" // d <= QP ? -1 : 0
        "psubusb "MANGLE(b01)", %%mm3           \n\t"
        "pand %%mm4, %%mm3                      \n\t" // d <= QP ? d : 0

        PAVGB(%%mm7, %%mm3)                           // d/2
        "movq %%mm3, %%mm1                      \n\t" // d/2
        PAVGB(%%mm7, %%mm3)                           // d/4
        PAVGB(%%mm1, %%mm3)                           // 3*d/8

        "movq (%0, %1, 4), %%mm0                \n\t" // line 4
        "pxor %%mm2, %%mm0                      \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4
        "psubusb %%mm3, %%mm0                   \n\t"
        "pxor %%mm2, %%mm0                      \n\t"
        "movq %%mm0, (%0, %1, 4)                \n\t" // line 4

        "movq (%%"REG_c"), %%mm0                \n\t" // line 5
        "pxor %%mm2, %%mm0                      \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5
        "paddusb %%mm3, %%mm0                   \n\t"
        "pxor %%mm2, %%mm0                      \n\t"
        "movq %%mm0, (%%"REG_c")                \n\t" // line 5

        PAVGB(%%mm7, %%mm1)                           // d/4

        "movq (%%"REG_a", %1, 2), %%mm0         \n\t" // line 3
        "pxor %%mm2, %%mm0                      \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4
        "psubusb %%mm1, %%mm0                   \n\t"
        "pxor %%mm2, %%mm0                      \n\t"
        "movq %%mm0, (%%"REG_a", %1, 2)         \n\t" // line 3

        "movq (%%"REG_c", %1), %%mm0            \n\t" // line 6
        "pxor %%mm2, %%mm0                      \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5
        "paddusb %%mm1, %%mm0                   \n\t"
        "pxor %%mm2, %%mm0                      \n\t"
        "movq %%mm0, (%%"REG_c", %1)            \n\t" // line 6

        PAVGB(%%mm7, %%mm1)                           // d/8

        "movq (%%"REG_a", %1), %%mm0            \n\t" // line 2
        "pxor %%mm2, %%mm0                      \n\t" //(l4 - l5) <= 0 ? -l2-1 : l2
        "psubusb %%mm1, %%mm0                   \n\t"
        "pxor %%mm2, %%mm0                      \n\t"
        "movq %%mm0, (%%"REG_a", %1)            \n\t" // line 2

        "movq (%%"REG_c", %1, 2), %%mm0         \n\t" // line 7
        "pxor %%mm2, %%mm0                      \n\t" //(l4 - l5) <= 0 ? -l7-1 : l7
        "paddusb %%mm1, %%mm0                   \n\t"
        "pxor %%mm2, %%mm0                      \n\t"
        "movq %%mm0, (%%"REG_c", %1, 2)         \n\t" // line 7

        :
        : "r" (src), "r" ((long)stride), "m" (co->pQPb)
        : "%"REG_a, "%"REG_c
    );
#else //defined (HAVE_MMX2) || defined (HAVE_3DNOW)

    const int l1= stride;
    const int l2= stride + l1;
    const int l3= stride + l2;
    const int l4= stride + l3;
    const int l5= stride + l4;
    const int l6= stride + l5;
    const int l7= stride + l6;
//    const int l8= stride + l7;
//    const int l9= stride + l8;
    int x;

    src+= stride*3;
    for(x=0; x<BLOCK_SIZE; x++){
        int a= src[l3] - src[l4];
        int b= src[l4] - src[l5];
        int c= src[l5] - src[l6];

        int d= FFABS(b) - ((FFABS(a) + FFABS(c))>>1);
        d= FFMAX(d, 0);

        if(d < co->QP*2){
            int v = d * FFSIGN(-b);

            src[l2] +=v>>3;
            src[l3] +=v>>2;
            src[l4] +=(3*v)>>3;
            src[l5] -=(3*v)>>3;
            src[l6] -=v>>2;
            src[l7] -=v>>3;
        }
        src++;
    }