comparison x86/h264dsp_mmx.c @ 11017:aa10bb3c244c libavcodec

optimize h264_loop_filter_strength_mmx2 244->160 cycles on core2

11016:4aee091df934

static void h264_loop_filter_strength_mmx2( int16_t bS[2][4][4], uint8_t nnz[40], int8_t ref[2][40], int16_t mv[2][40][2],
                                            int bidir, int edges, int step, int mask_mv0, int mask_mv1, int field ) {
    int dir;
    __asm__ volatile(
        "pxor %%mm7, %%mm7 \n\t"
        "movq %0, %%mm6 \n\t"
        "movq %1, %%mm5 \n\t"
        "movq %2, %%mm4 \n\t"
        ::"m"(ff_pb_1), "m"(ff_pb_3), "m"(ff_pb_7)
    );
    if(field)
        __asm__ volatile(
            "movq %0, %%mm5 \n\t"
            "movq %1, %%mm4 \n\t"
            ::"m"(ff_pb_3_1), "m"(ff_pb_7_3)
        );

    // could do a special case for dir==0 && edges==1, but it only reduces the
    // average filter time by 1.2%
    for( dir=1; dir>=0; dir-- ) {
        const x86_reg d_idx = dir ? -8 : -1;
        const int mask_mv = dir ? mask_mv1 : mask_mv0;
        DECLARE_ALIGNED_8(const uint64_t, mask_dir) = dir ? 0 : 0xffffffffffffffffULL;
        int b_idx, edge, l;
        for( b_idx=12, edge=0; edge<edges; edge+=step, b_idx+=8*step ) {
            __asm__ volatile(
                "pand %0, %%mm0 \n\t"
                ::"m"(mask_dir)
            );
            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 (%0,%1), %%mm1 \n\t"
                        "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 8(%2), %%mm2 \n\t"
                        "psubw (%2,%1,4), %%mm1 \n\t"
                        "psubw 8(%2,%1,4), %%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"
                        ::"r"(ref[l]+b_idx),
                          "r"(d_idx),
                          "r"(mv[l]+b_idx)
                    );
                }
                if(bidir==1){
                    __asm__ volatile("pxor %%mm3, %%mm3 \n\t":);
                    for( l = bidir; l >= 0; l-- ) {
                    __asm__ volatile(
                        "movd (%0), %%mm1 \n\t"
                        "punpckldq (%1), %%mm1 \n\t"
                        "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, %%mm3 \n\t"

                        "movq (%2), %%mm1 \n\t"
                        "movq 8(%2), %%mm2 \n\t"
                        "psubw (%3), %%mm1 \n\t"
                        "psubw 8(%3), %%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, %%mm3 \n\t"
                        ::"r"(ref[l]+b_idx),
                          "r"(ref[1-l]+b_idx+d_idx),
                          "r"(mv[l][b_idx]),
                          "r"(mv[1-l][b_idx+d_idx])
                    );
                    }
                    __asm__ volatile(
                        "pcmpeqw %%mm7, %%mm3 \n\t"
                        "psubusw %%mm3, %%mm0 \n\t"
                    :);
                }
            }
            __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;

11017:aa10bb3c244c

static void h264_loop_filter_strength_mmx2( int16_t bS[2][4][4], uint8_t nnz[40], int8_t ref[2][40], int16_t mv[2][40][2],
                                            int bidir, int edges, int step, int mask_mv0, int mask_mv1, int field ) {
    int dir;
    __asm__ volatile(
        "movq %0, %%mm7 \n"
        "movq %1, %%mm6 \n"
        ::"m"(ff_pb_1), "m"(ff_pb_3)
    );
    if(field)
        __asm__ volatile(
            "movq %0, %%mm6 \n"
            ::"m"(ff_pb_3_1)
        );
    __asm__ volatile(
        "movq  %%mm6, %%mm5 \n"
        "paddb %%mm5, %%mm5 \n"
    :);

    // could do a special case for dir==0 && edges==1, but it only reduces the
    // average filter time by 1.2%
    for( dir=1; dir>=0; dir-- ) {
        const x86_reg d_idx = dir ? -8 : -1;
        const int mask_mv = dir ? mask_mv1 : mask_mv0;
        DECLARE_ALIGNED_8(const uint64_t, mask_dir) = dir ? 0 : 0xffffffffffffffffULL;
        int b_idx, edge;
        for( b_idx=12, edge=0; edge<edges; edge+=step, b_idx+=8*step ) {
            __asm__ volatile(
                "pand %0, %%mm0 \n\t"
                ::"m"(mask_dir)
            );
            if(!(mask_mv & edge)) {
                if(bidir) {
                    __asm__ volatile(
                        "movd         (%1,%0), %%mm2 \n"
                        "punpckldq  40(%1,%0), %%mm2 \n" // { ref0[bn], ref1[bn] }
                        "pshufw $0x44,   (%1), %%mm0 \n" // { ref0[b], ref0[b] }
                        "pshufw $0x44, 40(%1), %%mm1 \n" // { ref1[b], ref1[b] }
                        "pshufw $0x4E, %%mm2, %%mm3 \n"
                        "psubb         %%mm2, %%mm0 \n" // { ref0[b]!=ref0[bn], ref0[b]!=ref1[bn] }
                        "psubb         %%mm3, %%mm1 \n" // { ref1[b]!=ref1[bn], ref1[b]!=ref0[bn] }
                        "1: \n"
                        "por           %%mm1, %%mm0 \n"
                        "movq      (%2,%0,4), %%mm1 \n"
                        "movq     8(%2,%0,4), %%mm2 \n"
                        "movq          %%mm1, %%mm3 \n"
                        "movq          %%mm2, %%mm4 \n"
                        "psubw          (%2), %%mm1 \n"
                        "psubw         8(%2), %%mm2 \n"
                        "psubw       160(%2), %%mm3 \n"
                        "psubw       168(%2), %%mm4 \n"
                        "packsswb      %%mm2, %%mm1 \n"
                        "packsswb      %%mm4, %%mm3 \n"
                        "paddb         %%mm6, %%mm1 \n"
                        "paddb         %%mm6, %%mm3 \n"
                        "psubusb       %%mm5, %%mm1 \n" // abs(mv[b] - mv[bn]) >= limit
                        "psubusb       %%mm5, %%mm3 \n"
                        "packsswb      %%mm3, %%mm1 \n"
                        "add $40, %0 \n"
                        "cmp $40, %0 \n"
                        "jl 1b \n"
                        "sub $80, %0 \n"
                        "pshufw $0x4E, %%mm1, %%mm1 \n"
                        "por           %%mm1, %%mm0 \n"
                        "pshufw $0x4E, %%mm0, %%mm1 \n"
                        "pminub        %%mm1, %%mm0 \n"
                        ::"r"(d_idx),
                          "r"(ref[0]+b_idx),
                          "r"(mv[0]+b_idx)
                    );
                } else {
                    __asm__ volatile(
                        "movd        (%1), %%mm0 \n"
                        "psubb    (%1,%0), %%mm0 \n" // ref[b] != ref[bn]
                        "movq        (%2), %%mm1 \n"
                        "movq       8(%2), %%mm2 \n"
                        "psubw  (%2,%0,4), %%mm1 \n"
                        "psubw 8(%2,%0,4), %%mm2 \n"
                        "packsswb   %%mm2, %%mm1 \n"
                        "paddb      %%mm6, %%mm1 \n"
                        "psubusb    %%mm5, %%mm1 \n" // abs(mv[b] - mv[bn]) >= limit
                        "packsswb   %%mm1, %%mm1 \n"
                        "por        %%mm1, %%mm0 \n"
                        ::"r"(d_idx),
                          "r"(ref[0]+b_idx),
                          "r"(mv[0]+b_idx)
                    );
                }
            }
            __asm__ volatile(
                "movd %0, %%mm1 \n"
                "por  %1, %%mm1 \n" // nnz[b] || nnz[bn]
                ::"m"(nnz[b_idx]),
                  "m"(nnz[b_idx+d_idx])
            );
            __asm__ volatile(
                "pminub    %%mm7, %%mm1 \n"
                "pminub    %%mm7, %%mm0 \n"
                "psllw        $1, %%mm1 \n"
                "pxor      %%mm2, %%mm2 \n"
                "pmaxub    %%mm0, %%mm1 \n"
                "punpcklbw %%mm2, %%mm1 \n"
                "movq      %%mm1, %0    \n"
                :"=m"(*bS[dir][edge])
                ::"memory"
            );
        }
        edges = 4;