fdct_mmx.c

这是著名的TCPMP播放器在WINDWOWS,和WINCE下编译通过的源程序.笔者对其中的LIBMAD库做了针对ARM MPU的优化. 并增加了词幕功能.

    psllw_i2r(SHIFT_FRW_COL, mm4);
    movq_r2r(mm0, mm6);
    psubsw_r2r(mm1, mm2);
    movq_m2r(*(fdct_tg_all_16 + 4), mm1);
    psubsw_r2r(mm4, mm0);
    movq_m2r(*(in + offset + 3 * 8), mm7);
    pmulhw_r2r(mm0, mm1);
    paddsw_m2r(*(in + offset + 4 * 8), mm7);
    psllw_i2r(SHIFT_FRW_COL, mm5);
    paddsw_r2r(mm4, mm6);
    psllw_i2r(SHIFT_FRW_COL, mm7);
    movq_r2r(mm5, mm4);
    psubsw_r2r(mm7, mm5);
    paddsw_r2r(mm5, mm1);
    paddsw_r2r(mm7, mm4);
    por_m2r(fdct_one_corr, mm1);
    psllw_i2r(SHIFT_FRW_COL + 1, mm2);
    pmulhw_m2r(*(fdct_tg_all_16 + 4), mm5);
    movq_r2r(mm4, mm7);
    psubsw_m2r(*(in + offset + 5 * 8), mm3);
    psubsw_r2r(mm6, mm4);
    movq_r2m(mm1, *(out + offset + 2 * 8));
    paddsw_r2r(mm6, mm7);
    movq_m2r(*(in + offset + 3 * 8), mm1);
    psllw_i2r(SHIFT_FRW_COL + 1, mm3);
    psubsw_m2r(*(in + offset + 4 * 8), mm1);
    movq_r2r(mm2, mm6);
    movq_r2m(mm4, *(out + offset + 4 * 8));
    paddsw_r2r(mm3, mm2);
    pmulhw_m2r(*ocos_4_16, mm2);
    psubsw_r2r(mm3, mm6);
    pmulhw_m2r(*ocos_4_16, mm6);
    psubsw_r2r(mm0, mm5);
    por_m2r(fdct_one_corr, mm5);
    psllw_i2r(SHIFT_FRW_COL, mm1);
    por_m2r(fdct_one_corr, mm2);
    movq_r2r(mm1, mm4);
    movq_m2r(*(in + offset + 0 * 8), mm3);
    paddsw_r2r(mm6, mm1);
    psubsw_m2r(*(in + offset + 7 * 8), mm3);
    psubsw_r2r(mm6, mm4);
    movq_m2r(*(fdct_tg_all_16 + 0), mm0);
    psllw_i2r(SHIFT_FRW_COL, mm3);
    movq_m2r(*(fdct_tg_all_16 + 8), mm6);
    pmulhw_r2r(mm1, mm0);
    movq_r2m(mm7, *(out + offset + 0 * 8));
    pmulhw_r2r(mm4, mm6);
    movq_r2m(mm5, *(out + offset + 6 * 8));
    movq_r2r(mm3, mm7);
    movq_m2r(*(fdct_tg_all_16 + 8), mm5);
    psubsw_r2r(mm2, mm7);
    paddsw_r2r(mm2, mm3);
    pmulhw_r2r(mm7, mm5);
    paddsw_r2r(mm3, mm0);
    paddsw_r2r(mm4, mm6);
    pmulhw_m2r(*(fdct_tg_all_16 + 0), mm3);
    por_m2r(fdct_one_corr, mm0);
    paddsw_r2r(mm7, mm5);
    psubsw_r2r(mm6, mm7);
    movq_r2m(mm0, *(out + offset + 1 * 8));
    paddsw_r2r(mm4, mm5);
    movq_r2m(mm7, *(out + offset + 3 * 8));
    psubsw_r2r(mm1, mm3);
    movq_r2m(mm5, *(out + offset + 5 * 8));
    movq_r2m(mm3, *(out + offset + 7 * 8));
}


static always_inline void fdct_row_sse2(const int16_t *in, int16_t *out)
{
    asm volatile(
        ".macro FDCT_ROW_SSE2_H1 i t   \n\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"	
	".endm                         \n\t"
        ".macro FDCT_ROW_SSE2_H2 i t   \n\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"	
	".endm                         \n\t"
	".macro FDCT_ROW_SSE2 i        \n\t"	
	"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"
	".endm                         \n\t"	
	"movdqa    (%2), %%xmm6        \n\t"		
	"FDCT_ROW_SSE2_H1 0 0 \n\t"
	"FDCT_ROW_SSE2 0 \n\t"
	"FDCT_ROW_SSE2_H2 64 0 \n\t"
	"FDCT_ROW_SSE2 64 \n\t"

	"FDCT_ROW_SSE2_H1 16 64 \n\t"
	"FDCT_ROW_SSE2 16 \n\t"
	"FDCT_ROW_SSE2_H2 112 64 \n\t"
	"FDCT_ROW_SSE2 112 \n\t"

	"FDCT_ROW_SSE2_H1 32 128 \n\t"
	"FDCT_ROW_SSE2 32 \n\t"
	"FDCT_ROW_SSE2_H2 96 128 \n\t"
	"FDCT_ROW_SSE2 96 \n\t"

	"FDCT_ROW_SSE2_H1 48 192 \n\t"
	"FDCT_ROW_SSE2 48 \n\t"
	"FDCT_ROW_SSE2_H2 80 192 \n\t"
	"FDCT_ROW_SSE2 80 \n\t"
	:
	: "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 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 always_inline void fdct_row_mmx(const int16_t *in, int16_t *out, const int16_t *table)
{ 
//FIXME reorder (i dont have a 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 * const block_tmp= (int16_t*)align_tmp;
    int16_t *block1, *out;
    const int16_t *table;
    int i;

    block1 = block_tmp;
    fdct_col(block, block1, 0);
    fdct_col(block, block1, 4);

    block1 = block_tmp;
    table = tab_frw_01234567;
    out = block;
    for(i=8;i>0;i--) {
        fdct_row_mmx(block1, out, table);
        block1 += 8;
        table += 32;
        out += 8;
    }
}

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

    block1 = block_tmp;
    fdct_col(block, block1, 0);
    fdct_col(block, block1, 4);

    block1 = block_tmp;
    table = tab_frw_01234567;
    out = block;
    for(i=8;i>0;i--) {
        fdct_row_mmx2(block1, out, table);
        block1 += 8;
        table += 32;
        out += 8;
    }
}

void ff_fdct_sse2(int16_t *block) 
{
    int64_t align_tmp[16] ATTR_ALIGN(8);
    int16_t * const block_tmp= (int16_t*)align_tmp;
    int16_t *block1;

    block1 = block_tmp;
    fdct_col(block, block1, 0);
    fdct_col(block, block1, 4);

    fdct_row_sse2(block1, block);
}