fdct_mmx.c

ffmpeg的完整源代码和作者自己写的文档。不但有在Linux的工程哦

}};

#define FDCT_COL(cpu, mm, mov)\
static av_always_inline void fdct_col_##cpu(const int16_t *in, int16_t *out, int offset)\
{\
    mov##_m2r(*(in + offset + 1 * 8), mm##0);\
    mov##_m2r(*(in + offset + 6 * 8), mm##1);\
    mov##_r2r(mm##0, mm##2);\
    mov##_m2r(*(in + offset + 2 * 8), mm##3);\
    paddsw_r2r(mm##1, mm##0);\
    mov##_m2r(*(in + offset + 5 * 8), mm##4);\
    psllw_i2r(SHIFT_FRW_COL, mm##0);\
    mov##_m2r(*(in + offset + 0 * 8), mm##5);\
    paddsw_r2r(mm##3, mm##4);\
    paddsw_m2r(*(in + offset + 7 * 8), mm##5);\
    psllw_i2r(SHIFT_FRW_COL, mm##4);\
    mov##_r2r(mm##0, mm##6);\
    psubsw_r2r(mm##1, mm##2);\
    mov##_m2r(*(fdct_tg_all_16 + 8), mm##1);\
    psubsw_r2r(mm##4, mm##0);\
    mov##_m2r(*(in + offset + 3 * 8), mm##7);\
    pmulhw_r2r(mm##0, mm##1);\
    paddsw_m2r(*(in + offset + 4 * 8), mm##7);\
    psllw_i2r(SHIFT_FRW_COL, mm##5);\
    paddsw_r2r(mm##4, mm##6);\
    psllw_i2r(SHIFT_FRW_COL, mm##7);\
    mov##_r2r(mm##5, mm##4);\
    psubsw_r2r(mm##7, mm##5);\
    paddsw_r2r(mm##5, mm##1);\
    paddsw_r2r(mm##7, mm##4);\
    por_m2r(*fdct_one_corr, mm##1);\
    psllw_i2r(SHIFT_FRW_COL + 1, mm##2);\
    pmulhw_m2r(*(fdct_tg_all_16 + 8), mm##5);\
    mov##_r2r(mm##4, mm##7);\
    psubsw_m2r(*(in + offset + 5 * 8), mm##3);\
    psubsw_r2r(mm##6, mm##4);\
    mov##_r2m(mm##1, *(out + offset + 2 * 8));\
    paddsw_r2r(mm##6, mm##7);\
    mov##_m2r(*(in + offset + 3 * 8), mm##1);\
    psllw_i2r(SHIFT_FRW_COL + 1, mm##3);\
    psubsw_m2r(*(in + offset + 4 * 8), mm##1);\
    mov##_r2r(mm##2, mm##6);\
    mov##_r2m(mm##4, *(out + offset + 4 * 8));\
    paddsw_r2r(mm##3, mm##2);\
    pmulhw_m2r(*ocos_4_16, mm##2);\
    psubsw_r2r(mm##3, mm##6);\
    pmulhw_m2r(*ocos_4_16, mm##6);\
    psubsw_r2r(mm##0, mm##5);\
    por_m2r(*fdct_one_corr, mm##5);\
    psllw_i2r(SHIFT_FRW_COL, mm##1);\
    por_m2r(*fdct_one_corr, mm##2);\
    mov##_r2r(mm##1, mm##4);\
    mov##_m2r(*(in + offset + 0 * 8), mm##3);\
    paddsw_r2r(mm##6, mm##1);\
    psubsw_m2r(*(in + offset + 7 * 8), mm##3);\
    psubsw_r2r(mm##6, mm##4);\
    mov##_m2r(*(fdct_tg_all_16 + 0), mm##0);\
    psllw_i2r(SHIFT_FRW_COL, mm##3);\
    mov##_m2r(*(fdct_tg_all_16 + 16), mm##6);\
    pmulhw_r2r(mm##1, mm##0);\
    mov##_r2m(mm##7, *(out + offset + 0 * 8));\
    pmulhw_r2r(mm##4, mm##6);\
    mov##_r2m(mm##5, *(out + offset + 6 * 8));\
    mov##_r2r(mm##3, mm##7);\
    mov##_m2r(*(fdct_tg_all_16 + 16), mm##5);\
    psubsw_r2r(mm##2, mm##7);\
    paddsw_r2r(mm##2, mm##3);\
    pmulhw_r2r(mm##7, mm##5);\
    paddsw_r2r(mm##3, mm##0);\
    paddsw_r2r(mm##4, mm##6);\
    pmulhw_m2r(*(fdct_tg_all_16 + 0), mm##3);\
    por_m2r(*fdct_one_corr, mm##0);\
    paddsw_r2r(mm##7, mm##5);\
    psubsw_r2r(mm##6, mm##7);\
    mov##_r2m(mm##0, *(out + offset + 1 * 8));\
    paddsw_r2r(mm##4, mm##5);\
    mov##_r2m(mm##7, *(out + offset + 3 * 8));\
    psubsw_r2r(mm##1, mm##3);\
    mov##_r2m(mm##5, *(out + offset + 5 * 8));\
    mov##_r2m(mm##3, *(out + offset + 7 * 8));\
}

FDCT_COL(mmx, mm, movq)
FDCT_COL(sse2, xmm, movdqa)

static av_always_inline void fdct_row_sse2(const int16_t *in, int16_t *out)
{
    asm volatile(
#define FDCT_ROW_SSE2_H1(i,t)                    \
        "movq      " #i "(%0), %%xmm2      \n\t" \
        "movq      " #i "+8(%0), %%xmm0    \n\t" \
        "movdqa    " #t "+32(%1), %%xmm3   \n\t" \
        "movdqa    " #t "+48(%1), %%xmm7   \n\t" \
        "movdqa    " #t "(%1), %%xmm4      \n\t" \
        "movdqa    " #t "+16(%1), %%xmm5   \n\t"

#define FDCT_ROW_SSE2_H2(i,t)                    \
        "movq      " #i "(%0), %%xmm2      \n\t" \
        "movq      " #i "+8(%0), %%xmm0    \n\t" \
        "movdqa    " #t "+32(%1), %%xmm3   \n\t" \
        "movdqa    " #t "+48(%1), %%xmm7   \n\t"

#define FDCT_ROW_SSE2(i)                      \
        "movq      %%xmm2, %%xmm1       \n\t" \
        "pshuflw   $27, %%xmm0, %%xmm0  \n\t" \
        "paddsw    %%xmm0, %%xmm1       \n\t" \
        "psubsw    %%xmm0, %%xmm2       \n\t" \
        "punpckldq %%xmm2, %%xmm1       \n\t" \
        "pshufd    $78, %%xmm1, %%xmm2  \n\t" \
        "pmaddwd   %%xmm2, %%xmm3       \n\t" \
        "pmaddwd   %%xmm1, %%xmm7       \n\t" \
        "pmaddwd   %%xmm5, %%xmm2       \n\t" \
        "pmaddwd   %%xmm4, %%xmm1       \n\t" \
        "paddd     %%xmm7, %%xmm3       \n\t" \
        "paddd     %%xmm2, %%xmm1       \n\t" \
        "paddd     %%xmm6, %%xmm3       \n\t" \
        "paddd     %%xmm6, %%xmm1       \n\t" \
        "psrad     %3, %%xmm3           \n\t" \
        "psrad     %3, %%xmm1           \n\t" \
        "packssdw  %%xmm3, %%xmm1       \n\t" \
        "movdqa    %%xmm1, " #i "(%4)   \n\t"

        "movdqa    (%2), %%xmm6         \n\t"
        FDCT_ROW_SSE2_H1(0,0)
        FDCT_ROW_SSE2(0)
        FDCT_ROW_SSE2_H2(64,0)
        FDCT_ROW_SSE2(64)

        FDCT_ROW_SSE2_H1(16,64)
        FDCT_ROW_SSE2(16)
        FDCT_ROW_SSE2_H2(112,64)
        FDCT_ROW_SSE2(112)

        FDCT_ROW_SSE2_H1(32,128)
        FDCT_ROW_SSE2(32)
        FDCT_ROW_SSE2_H2(96,128)
        FDCT_ROW_SSE2(96)

        FDCT_ROW_SSE2_H1(48,192)
        FDCT_ROW_SSE2(48)
        FDCT_ROW_SSE2_H2(80,192)
        FDCT_ROW_SSE2(80)
        :
        : "r" (in), "r" (tab_frw_01234567_sse2.tab_frw_01234567_sse2), "r" (fdct_r_row_sse2.fdct_r_row_sse2), "i" (SHIFT_FRW_ROW), "r" (out)
    );
}

static av_always_inline void fdct_row_mmx2(const int16_t *in, int16_t *out, const int16_t *table)
{
    pshufw_m2r(*(in + 4), mm5, 0x1B);
    movq_m2r(*(in + 0), mm0);
    movq_r2r(mm0, mm1);
    paddsw_r2r(mm5, mm0);
    psubsw_r2r(mm5, mm1);
    movq_r2r(mm0, mm2);
    punpckldq_r2r(mm1, mm0);
    punpckhdq_r2r(mm1, mm2);
    movq_m2r(*(table + 0), mm1);
    movq_m2r(*(table + 4), mm3);
    movq_m2r(*(table + 8), mm4);
    movq_m2r(*(table + 12), mm5);
    movq_m2r(*(table + 16), mm6);
    movq_m2r(*(table + 20), mm7);
    pmaddwd_r2r(mm0, mm1);
    pmaddwd_r2r(mm2, mm3);
    pmaddwd_r2r(mm0, mm4);
    pmaddwd_r2r(mm2, mm5);
    pmaddwd_r2r(mm0, mm6);
    pmaddwd_r2r(mm2, mm7);
    pmaddwd_m2r(*(table + 24), mm0);
    pmaddwd_m2r(*(table + 28), mm2);
    paddd_r2r(mm1, mm3);
    paddd_r2r(mm4, mm5);
    paddd_r2r(mm6, mm7);
    paddd_r2r(mm0, mm2);
    movq_m2r(*fdct_r_row, mm0);
    paddd_r2r(mm0, mm3);
    paddd_r2r(mm0, mm5);
    paddd_r2r(mm0, mm7);
    paddd_r2r(mm0, mm2);
    psrad_i2r(SHIFT_FRW_ROW, mm3);
    psrad_i2r(SHIFT_FRW_ROW, mm5);
    psrad_i2r(SHIFT_FRW_ROW, mm7);
    psrad_i2r(SHIFT_FRW_ROW, mm2);
    packssdw_r2r(mm5, mm3);
    packssdw_r2r(mm2, mm7);
    movq_r2m(mm3, *(out + 0));
    movq_r2m(mm7, *(out + 4));
}

static av_always_inline void fdct_row_mmx(const int16_t *in, int16_t *out, const int16_t *table)
{
//FIXME reorder (I do not have an old MMX-only CPU here to benchmark ...)
    movd_m2r(*(in + 6), mm1);
    punpcklwd_m2r(*(in + 4), mm1);
    movq_r2r(mm1, mm2);
    psrlq_i2r(0x20, mm1);
    movq_m2r(*(in + 0), mm0);
    punpcklwd_r2r(mm2, mm1);
    movq_r2r(mm0, mm5);
    paddsw_r2r(mm1, mm0);
    psubsw_r2r(mm1, mm5);
    movq_r2r(mm0, mm2);
    punpckldq_r2r(mm5, mm0);
    punpckhdq_r2r(mm5, mm2);
    movq_m2r(*(table + 0), mm1);
    movq_m2r(*(table + 4), mm3);
    movq_m2r(*(table + 8), mm4);
    movq_m2r(*(table + 12), mm5);
    movq_m2r(*(table + 16), mm6);
    movq_m2r(*(table + 20), mm7);
    pmaddwd_r2r(mm0, mm1);
    pmaddwd_r2r(mm2, mm3);
    pmaddwd_r2r(mm0, mm4);
    pmaddwd_r2r(mm2, mm5);
    pmaddwd_r2r(mm0, mm6);
    pmaddwd_r2r(mm2, mm7);
    pmaddwd_m2r(*(table + 24), mm0);
    pmaddwd_m2r(*(table + 28), mm2);
    paddd_r2r(mm1, mm3);
    paddd_r2r(mm4, mm5);
    paddd_r2r(mm6, mm7);
    paddd_r2r(mm0, mm2);
    movq_m2r(*fdct_r_row, mm0);
    paddd_r2r(mm0, mm3);
    paddd_r2r(mm0, mm5);
    paddd_r2r(mm0, mm7);
    paddd_r2r(mm0, mm2);
    psrad_i2r(SHIFT_FRW_ROW, mm3);
    psrad_i2r(SHIFT_FRW_ROW, mm5);
    psrad_i2r(SHIFT_FRW_ROW, mm7);
    psrad_i2r(SHIFT_FRW_ROW, mm2);
    packssdw_r2r(mm5, mm3);
    packssdw_r2r(mm2, mm7);
    movq_r2m(mm3, *(out + 0));
    movq_r2m(mm7, *(out + 4));
}

void ff_fdct_mmx(int16_t *block)
{
    int64_t align_tmp[16] ATTR_ALIGN(8);
    int16_t * block1= (int16_t*)align_tmp;
    const int16_t *table= tab_frw_01234567;
    int i;

    fdct_col_mmx(block, block1, 0);
    fdct_col_mmx(block, block1, 4);

    for(i=8;i>0;i--) {
        fdct_row_mmx(block1, block, table);
        block1 += 8;
        table += 32;
        block += 8;
    }
}

void ff_fdct_mmx2(int16_t *block)
{
    int64_t align_tmp[16] ATTR_ALIGN(8);
    int16_t *block1= (int16_t*)align_tmp;
    const int16_t *table= tab_frw_01234567;
    int i;

    fdct_col_mmx(block, block1, 0);
    fdct_col_mmx(block, block1, 4);

    for(i=8;i>0;i--) {
        fdct_row_mmx2(block1, block, table);
        block1 += 8;
        table += 32;
        block += 8;
    }
}

void ff_fdct_sse2(int16_t *block)
{
    int64_t align_tmp[16] ATTR_ALIGN(16);
    int16_t * const block1= (int16_t*)align_tmp;

    fdct_col_sse2(block, block1, 0);
    fdct_row_sse2(block1, block);
}