h264dsp_mmx.c.svn-base

ffmpeg是一个大项目

        "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);
}

static void h264_h_loop_filter_chroma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0)
{
    //FIXME: could cut some load/stores by merging transpose with filter
    uint8_t trans[8*4];
    transpose4x4(trans, pix-2, 8, stride);
    transpose4x4(trans+4, pix-2+4*stride, 8, stride);
    h264_loop_filter_chroma_mmx2(trans+2*8, 8, alpha-1, beta-1, tc0);
    transpose4x4(pix-2, trans, stride, 8);
    transpose4x4(pix-2+4*stride, trans+4, stride, 8);
}

// p0 = (p0 + q1 + 2*p1 + 2) >> 2
#define H264_FILTER_CHROMA4(p0, p1, q1, one) \
    "movq    "#p0", %%mm4  \n\t"\
    "pxor    "#q1", %%mm4  \n\t"\
    "pand   "#one", %%mm4  \n\t" /* mm4 = (p0^q1)&1 */\
    "pavgb   "#q1", "#p0"  \n\t"\
    "psubusb %%mm4, "#p0"  \n\t"\
    "pavgb   "#p1", "#p0"  \n\t" /* dst = avg(p1, avg(p0,q1) - ((p0^q1)&1)) */\

static inline void h264_loop_filter_chroma_intra_mmx2(uint8_t *pix, int stride, int alpha1, int beta1)
{
    asm volatile(
        "movq    (%0),    %%mm0     \n\t"
        "movq    (%0,%2), %%mm1     \n\t"
        "movq    (%1),    %%mm2     \n\t"
        "movq    (%1,%2), %%mm3     \n\t"
        H264_DEBLOCK_MASK(%3, %4)
        "movq    %%mm1,   %%mm5     \n\t"
        "movq    %%mm2,   %%mm6     \n\t"
        H264_FILTER_CHROMA4(%%mm1, %%mm0, %%mm3, %5) //p0'
        H264_FILTER_CHROMA4(%%mm2, %%mm3, %%mm0, %5) //q0'
        "psubb   %%mm5,   %%mm1     \n\t"
        "psubb   %%mm6,   %%mm2     \n\t"
        "pand    %%mm7,   %%mm1     \n\t"
        "pand    %%mm7,   %%mm2     \n\t"
        "paddb   %%mm5,   %%mm1     \n\t"
        "paddb   %%mm6,   %%mm2     \n\t"
        "movq    %%mm1,   (%0,%2)   \n\t"
        "movq    %%mm2,   (%1)      \n\t"
        :: "r"(pix-2*stride), "r"(pix), "r"((long)stride),
           "m"(alpha1), "m"(beta1), "m"(mm_bone)
    );
}

static void h264_v_loop_filter_chroma_intra_mmx2(uint8_t *pix, int stride, int alpha, int beta)
{
    h264_loop_filter_chroma_intra_mmx2(pix, stride, alpha-1, beta-1);
}

static void h264_h_loop_filter_chroma_intra_mmx2(uint8_t *pix, int stride, int alpha, int beta)
{
    //FIXME: could cut some load/stores by merging transpose with filter
    uint8_t trans[8*4];
    transpose4x4(trans, pix-2, 8, stride);
    transpose4x4(trans+4, pix-2+4*stride, 8, stride);
    h264_loop_filter_chroma_intra_mmx2(trans+2*8, 8, alpha-1, beta-1);
    transpose4x4(pix-2, trans, stride, 8);
    transpose4x4(pix-2+4*stride, trans+4, stride, 8);
}


/***********************************/
/* motion compensation */

#define QPEL_H264V(A,B,C,D,E,F,OP)\
        "movd (%0), "#F"            \n\t"\
        "movq "#C", %%mm6           \n\t"\
        "paddw "#D", %%mm6          \n\t"\
        "psllw $2, %%mm6            \n\t"\
        "psubw "#B", %%mm6          \n\t"\
        "psubw "#E", %%mm6          \n\t"\
        "pmullw %4, %%mm6           \n\t"\
        "add %2, %0                 \n\t"\
        "punpcklbw %%mm7, "#F"      \n\t"\
        "paddw %5, "#A"             \n\t"\
        "paddw "#F", "#A"           \n\t"\
        "paddw "#A", %%mm6          \n\t"\
        "psraw $5, %%mm6            \n\t"\
        "packuswb %%mm6, %%mm6      \n\t"\
        OP(%%mm6, (%1), A, d)\
        "add %3, %1                 \n\t"

#define QPEL_H264HV(A,B,C,D,E,F,OF)\
        "movd (%0), "#F"            \n\t"\
        "movq "#C", %%mm6           \n\t"\
        "paddw "#D", %%mm6          \n\t"\
        "psllw $2, %%mm6            \n\t"\
        "psubw "#B", %%mm6          \n\t"\
        "psubw "#E", %%mm6          \n\t"\
        "pmullw %3, %%mm6           \n\t"\
        "add %2, %0                 \n\t"\
        "punpcklbw %%mm7, "#F"      \n\t"\
        "paddw "#F", "#A"           \n\t"\
        "paddw "#A", %%mm6          \n\t"\
        "movq %%mm6, "#OF"(%1)      \n\t"

#define QPEL_H264(OPNAME, OP, MMX)\
static void OPNAME ## h264_qpel4_h_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
    int h=4;\
\
    asm volatile(\
        "pxor %%mm7, %%mm7          \n\t"\
        "movq %5, %%mm4             \n\t"\
        "movq %6, %%mm5             \n\t"\
        "1:                         \n\t"\
        "movd  -1(%0), %%mm1        \n\t"\
        "movd    (%0), %%mm2        \n\t"\
        "movd   1(%0), %%mm3        \n\t"\
        "movd   2(%0), %%mm0        \n\t"\
        "punpcklbw %%mm7, %%mm1     \n\t"\
        "punpcklbw %%mm7, %%mm2     \n\t"\
        "punpcklbw %%mm7, %%mm3     \n\t"\
        "punpcklbw %%mm7, %%mm0     \n\t"\
        "paddw %%mm0, %%mm1         \n\t"\
        "paddw %%mm3, %%mm2         \n\t"\
        "movd  -2(%0), %%mm0        \n\t"\
        "movd   3(%0), %%mm3        \n\t"\
        "punpcklbw %%mm7, %%mm0     \n\t"\
        "punpcklbw %%mm7, %%mm3     \n\t"\
        "paddw %%mm3, %%mm0         \n\t"\
        "psllw $2, %%mm2            \n\t"\
        "psubw %%mm1, %%mm2         \n\t"\
        "pmullw %%mm4, %%mm2        \n\t"\
        "paddw %%mm5, %%mm0         \n\t"\
        "paddw %%mm2, %%mm0         \n\t"\
        "psraw $5, %%mm0            \n\t"\
        "packuswb %%mm0, %%mm0      \n\t"\
        OP(%%mm0, (%1),%%mm6, d)\
        "add %3, %0                 \n\t"\
        "add %4, %1                 \n\t"\
        "decl %2                    \n\t"\
        " jnz 1b                    \n\t"\
        : "+a"(src), "+c"(dst), "+m"(h)\
        : "d"((long)srcStride), "S"((long)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\
        : "memory"\
    );\
}\
static void OPNAME ## h264_qpel4_h_lowpass_l2_ ## MMX(uint8_t *dst, uint8_t *src, uint8_t *src2, int dstStride, int src2Stride){\
    int h=4;\
    asm volatile(\
        "pxor %%mm7, %%mm7          \n\t"\
        "movq %0, %%mm4             \n\t"\
        "movq %1, %%mm5             \n\t"\
        :: "m"(ff_pw_5), "m"(ff_pw_16)\
    );\
    do{\
    asm volatile(\
        "movd  -1(%0), %%mm1        \n\t"\
        "movd    (%0), %%mm2        \n\t"\
        "movd   1(%0), %%mm3        \n\t"\
        "movd   2(%0), %%mm0        \n\t"\
        "punpcklbw %%mm7, %%mm1     \n\t"\
        "punpcklbw %%mm7, %%mm2     \n\t"\
        "punpcklbw %%mm7, %%mm3     \n\t"\
        "punpcklbw %%mm7, %%mm0     \n\t"\
        "paddw %%mm0, %%mm1         \n\t"\
        "paddw %%mm3, %%mm2         \n\t"\
        "movd  -2(%0), %%mm0        \n\t"\
        "movd   3(%0), %%mm3        \n\t"\
        "punpcklbw %%mm7, %%mm0     \n\t"\
        "punpcklbw %%mm7, %%mm3     \n\t"\
        "paddw %%mm3, %%mm0         \n\t"\
        "psllw $2, %%mm2            \n\t"\
        "psubw %%mm1, %%mm2         \n\t"\
        "pmullw %%mm4, %%mm2        \n\t"\
        "paddw %%mm5, %%mm0         \n\t"\
        "paddw %%mm2, %%mm0         \n\t"\
        "movd   (%2), %%mm3         \n\t"\
        "psraw $5, %%mm0            \n\t"\
        "packuswb %%mm0, %%mm0      \n\t"\
        PAVGB" %%mm3, %%mm0         \n\t"\
        OP(%%mm0, (%1),%%mm6, d)\
        "add %4, %0                 \n\t"\
        "add %4, %1                 \n\t"\
        "add %3, %2                 \n\t"\
        : "+a"(src), "+c"(dst), "+d"(src2)\
        : "D"((long)src2Stride), "S"((long)dstStride)\
        : "memory"\
    );\
    }while(--h);\
}\
static void OPNAME ## h264_qpel4_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
    src -= 2*srcStride;\
    asm volatile(\
        "pxor %%mm7, %%mm7          \n\t"\
        "movd (%0), %%mm0           \n\t"\
        "add %2, %0                 \n\t"\
        "movd (%0), %%mm1           \n\t"\
        "add %2, %0                 \n\t"\
        "movd (%0), %%mm2           \n\t"\
        "add %2, %0                 \n\t"\
        "movd (%0), %%mm3           \n\t"\
        "add %2, %0                 \n\t"\
        "movd (%0), %%mm4           \n\t"\
        "add %2, %0                 \n\t"\
        "punpcklbw %%mm7, %%mm0     \n\t"\
        "punpcklbw %%mm7, %%mm1     \n\t"\
        "punpcklbw %%mm7, %%mm2     \n\t"\
        "punpcklbw %%mm7, %%mm3     \n\t"\
        "punpcklbw %%mm7, %%mm4     \n\t"\
        QPEL_H264V(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP)\
        QPEL_H264V(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP)\
        QPEL_H264V(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP)\
        QPEL_H264V(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP)\
         \
        : "+a"(src), "+c"(dst)\
        : "S"((long)srcStride), "D"((long)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\
        : "memory"\
    );\
}\