h264dsp_mmx.c.svn-base

mediastreamer2是开源的网络传输媒体流的库

            if(!(mask_mv & edge)) {
                asm volatile("pxor %%mm0, %%mm0 \n\t":);
                for( l = bidir; l >= 0; l-- ) {
                    asm volatile(
                        "movd %0, %%mm1 \n\t"
                        "punpckldq %1, %%mm1 \n\t"
                        "movq %%mm1, %%mm2 \n\t"
                        "psrlw $7, %%mm2 \n\t"
                        "pand %%mm6, %%mm2 \n\t"
                        "por %%mm2, %%mm1 \n\t" // ref_cache with -2 mapped to -1
                        "punpckldq %%mm1, %%mm2 \n\t"
                        "pcmpeqb %%mm2, %%mm1 \n\t"
                        "paddb %%mm6, %%mm1 \n\t"
                        "punpckhbw %%mm7, %%mm1 \n\t" // ref[b] != ref[bn]
                        "por %%mm1, %%mm0 \n\t"

                        "movq %2, %%mm1 \n\t"
                        "movq %3, %%mm2 \n\t"
                        "psubw %4, %%mm1 \n\t"
                        "psubw %5, %%mm2 \n\t"
                        "packsswb %%mm2, %%mm1 \n\t"
                        "paddb %%mm5, %%mm1 \n\t"
                        "pminub %%mm4, %%mm1 \n\t"
                        "pcmpeqb %%mm4, %%mm1 \n\t" // abs(mv[b] - mv[bn]) >= limit
                        "por %%mm1, %%mm0 \n\t"
                        ::"m"(ref[l][b_idx]),
                          "m"(ref[l][b_idx+d_idx]),
                          "m"(mv[l][b_idx][0]),
                          "m"(mv[l][b_idx+2][0]),
                          "m"(mv[l][b_idx+d_idx][0]),
                          "m"(mv[l][b_idx+d_idx+2][0])
                    );
                }
            }
            asm volatile(
                "movd %0, %%mm1 \n\t"
                "por  %1, %%mm1 \n\t"
                "punpcklbw %%mm7, %%mm1 \n\t"
                "pcmpgtw %%mm7, %%mm1 \n\t" // nnz[b] || nnz[bn]
                ::"m"(nnz[b_idx]),
                  "m"(nnz[b_idx+d_idx])
            );
            asm volatile(
                "pcmpeqw %%mm7, %%mm0 \n\t"
                "pcmpeqw %%mm7, %%mm0 \n\t"
                "psrlw $15, %%mm0 \n\t" // nonzero -> 1
                "psrlw $14, %%mm1 \n\t"
                "movq %%mm0, %%mm2 \n\t"
                "por %%mm1, %%mm2 \n\t"
                "psrlw $1, %%mm1 \n\t"
                "pandn %%mm2, %%mm1 \n\t"
                "movq %%mm1, %0 \n\t"
                :"=m"(*bS[dir][edge])
                ::"memory"
            );
        }
        edges = 4;
        step = 1;
    }
    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"
        TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4)
        "movq %%mm0,   (%0) \n\t"
        "movq %%mm3,  8(%0) \n\t"
        "movq %%mm4, 16(%0) \n\t"
        "movq %%mm2, 24(%0) \n\t"
        ::"r"(bS[0])
        :"memory"
    );
}

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

#define QPEL_H264V_MM(A,B,C,D,E,F,OP,T,Z,d,q)\
        "mov"#q" "#C", "#T"         \n\t"\
        "mov"#d" (%0), "#F"         \n\t"\
        "paddw "#D", "#T"           \n\t"\
        "psllw $2, "#T"             \n\t"\
        "psubw "#B", "#T"           \n\t"\
        "psubw "#E", "#T"           \n\t"\
        "punpcklbw "#Z", "#F"       \n\t"\
        "pmullw %4, "#T"            \n\t"\
        "paddw %5, "#A"             \n\t"\
        "add %2, %0                 \n\t"\
        "paddw "#F", "#A"           \n\t"\
        "paddw "#A", "#T"           \n\t"\
        "psraw $5, "#T"             \n\t"\
        "packuswb "#T", "#T"        \n\t"\
        OP(T, (%1), A, d)\
        "add %3, %1                 \n\t"

#define QPEL_H264HV_MM(A,B,C,D,E,F,OF,T,Z,d,q)\
        "mov"#q" "#C", "#T"         \n\t"\
        "mov"#d" (%0), "#F"         \n\t"\
        "paddw "#D", "#T"           \n\t"\
        "psllw $2, "#T"             \n\t"\
        "paddw %4, "#A"             \n\t"\
        "psubw "#B", "#T"           \n\t"\
        "psubw "#E", "#T"           \n\t"\
        "punpcklbw "#Z", "#F"       \n\t"\
        "pmullw %3, "#T"            \n\t"\
        "paddw "#F", "#A"           \n\t"\
        "add %2, %0                 \n\t"\
        "paddw "#A", "#T"           \n\t"\
        "mov"#q" "#T", "#OF"(%1)    \n\t"

#define QPEL_H264V(A,B,C,D,E,F,OP) QPEL_H264V_MM(A,B,C,D,E,F,OP,%%mm6,%%mm7,d,q)
#define QPEL_H264HV(A,B,C,D,E,F,OF) QPEL_H264HV_MM(A,B,C,D,E,F,OF,%%mm6,%%mm7,d,q)
#define QPEL_H264V_XMM(A,B,C,D,E,F,OP) QPEL_H264V_MM(A,B,C,D,E,F,OP,%%xmm6,%%xmm7,q,dqa)
#define QPEL_H264HV_XMM(A,B,C,D,E,F,OF) QPEL_H264HV_MM(A,B,C,D,E,F,OF,%%xmm6,%%xmm7,q,dqa)


#define QPEL_H264(OPNAME, OP, MMX)\
static av_noinline 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), "+g"(h)\
        : "d"((long)srcStride), "S"((long)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\
        : "memory"\
    );\
}\
static av_noinline 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 av_noinline 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"\
    );\
}\
static av_noinline void OPNAME ## h264_qpel4_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\
    int h=4;\
    int w=3;\
    src -= 2*srcStride+2;\
    while(w--){\
        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_H264HV(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, 0*8*3)\
            QPEL_H264HV(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, 1*8*3)\
            QPEL_H264HV(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, 2*8*3)\
            QPEL_H264HV(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, 3*8*3)\
             \
            : "+a"(src)\
            : "c"(tmp), "S"((long)srcStride), "m"(ff_pw_5), "m"(ff_pw_16)\
            : "memory"\
        );\
        tmp += 4;\
        src += 4 - 9*srcStride;\
    }\
    tmp -= 3*4;\
    asm volatile(\
        "1:                         \n\t"\
        "movq     (%0), %%mm0       \n\t"\
        "paddw  10(%0), %%mm0       \n\t"\
        "movq    2(%0), %%mm1       \n\t"\
        "paddw   8(%0), %%mm1       \n\t"\
        "movq    4(%0), %%mm2       \n\t"\
        "paddw   6(%0), %%mm2       \n\t"\
        "psubw %%mm1, %%mm0         \n\t"/*a-b   (abccba)*/\
        "psraw $2, %%mm0            \n\t"/*(a-b)/4 */\
        "psubw %%mm1, %%mm0         \n\t"/*(a-b)/4-b */\
        "paddsw %%mm2, %%mm0        \n\t"\
        "psraw $2, %%mm0            \n\t"/*((a-b)/4-b+c)/4 */\
        "paddw %%mm2, %%mm0         \n\t"/*(a-5*b+20*c)/16 */\
        "psraw $6, %%mm0            \n\t"\
        "packuswb %%mm0, %%mm0      \n\t"\
        OP(%%mm0, (%1),%%mm7, d)\
        "add $24, %0                \n\t"\
        "add %3, %1                 \n\t"\
        "decl %2                    \n\t"\
        " jnz 1b                    \n\t"\
        : "+a"(tmp), "+c"(dst), "+g"(h)\
        : "S"((long)dstStride)\
        : "memory"\
    );\
}\
\
static av_noinline void OPNAME ## h264_qpel8_h_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
    int h=8;\
    asm volatile(\
        "pxor %%mm7, %%mm7          \n\t"\
        "movq %5, %%mm6             \n\t"\
        "1:                         \n\t"\
        "movq    (%0), %%mm0        \n\t"\
        "movq   1(%0), %%mm2        \n\t"\
        "movq %%mm0, %%mm1          \n\t"\
        "movq %%mm2, %%mm3          \n\t"\
        "punpcklbw %%mm7, %%mm0     \n\t"\
        "punpckhbw %%mm7, %%mm1     \n\t"\
        "punpcklbw %%mm7, %%mm2     \n\t"\
        "punpckhbw %%mm7, %%mm3     \n\t"\
        "paddw %%mm2, %%mm0         \n\t"\
        "paddw %%mm3, %%mm1         \n\t"\
        "psllw $2, %%mm0            \n\t"\
        "psllw $2, %%mm1            \n\t"\
        "movq   -1(%0), %%mm2       \n\t"\
        "movq    2(%0), %%mm4       \n\t"\
        "movq %%mm2, %%mm3          \n\t"\
        "movq %%mm4, %%mm5          \n\t"\
        "punpcklbw %%mm7, %%mm2     \n\t"\
        "punpckhbw %%mm7, %%mm3     \n\t"\
        "punpcklbw %%mm7, %%mm4     \n\t"\
        "punpckhbw %%mm7, %%mm5     \n\t"\
        "paddw %%mm4, %%mm2         \n\t"\
        "paddw %%mm3, %%mm5         \n\t"\
        "psubw %%mm2, %%mm0         \n\t"\
        "psubw %%mm5, %%mm1         \n\t"\
        "pmullw %%mm6, %%mm0        \n\t"\