idct_mmx_xvid.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

  "paddd %%mm6,%%mm5           \n\t"/* 6 ; b3=sum(odd3) b2=sum(odd2)*/\
  "movq %%mm0,%%mm4            \n\t"/* 4 ; a3 a2*/\
  "paddd %%mm5,%%mm0           \n\t"/* a3+b3 a2+b2*/\
  "psubd %%mm5,%%mm4           \n\t"/* 5 ; a3-b3 a2-b2*/\
  "psrad $11,%%mm0 \n\t"/* y3=a3+b3 y2=a2+b2*/\
  "psrad $11,%%mm4 \n\t"/* y4=a3-b3 y5=a2-b2*/\
  "packssdw %%mm0,%%mm1        \n\t"/* 0 ; y3 y2 y1 y0*/\
  "packssdw %%mm3,%%mm4        \n\t"/* 3 ; y6 y7 y4 y5*/\
  "movq %%mm4,%%mm7            \n\t"/* 7 ; y6 y7 y4 y5*/\
  "psrld $16,%%mm4            \n\t"/* 0 y6 0 y4*/\
  "pslld $16,%%mm7            \n\t"/* y7 0 y5 0*/\
  "movq %%mm1," #A2 "              \n\t"/* 1 ; save y3 y2 y1 y0*/\
  "por %%mm4,%%mm7             \n\t"/* 4 ; y7 y6 y5 y4*/\
  "movq %%mm7,8            +" #A2 "\n\t"/* 7 ; save y7 y6 y5 y4*/\


//-----------------------------------------------------------------------------
// DCT_8_INV_ROW_XMM( INP, OUT, TABLE, ROUNDER
//-----------------------------------------------------------------------------

#define DCT_8_INV_ROW_XMM(A1,A2,A3,A4)\
  "movq " #A1 ",%%mm0                  \n\t"/* 0     ; x3 x2 x1 x0*/\
  "movq 8+" #A1 ",%%mm1                \n\t"/* 1     ; x7 x6 x5 x4*/\
  "movq %%mm0,%%mm2                \n\t"/* 2     ; x3 x2 x1 x0*/\
  "movq " #A3 ",%%mm3                  \n\t"/* 3     ; w05 w04 w01 w00*/\
  "pshufw $0x88,%%mm0,%%mm0        \n\t"/* x2 x0 x2 x0*/\
  "movq 8+" #A3 ",%%mm4                \n\t"/* 4     ; w07 w06 w03 w02*/\
  "movq %%mm1,%%mm5                \n\t"/* 5     ; x7 x6 x5 x4*/\
  "pmaddwd %%mm0,%%mm3             \n\t"/* x2*w05+x0*w04 x2*w01+x0*w00*/\
  "movq 32+" #A3 ",%%mm6               \n\t"/* 6     ; w21 w20 w17 w16*/\
  "pshufw $0x88,%%mm1,%%mm1        \n\t"/* x6 x4 x6 x4*/\
  "pmaddwd %%mm1,%%mm4             \n\t"/* x6*w07+x4*w06 x6*w03+x4*w02*/\
  "movq 40+" #A3 ",%%mm7               \n\t"/* 7    ; w23 w22 w19 w18*/\
  "pshufw $0xdd,%%mm2,%%mm2        \n\t"/* x3 x1 x3 x1*/\
  "pmaddwd %%mm2,%%mm6             \n\t"/* x3*w21+x1*w20 x3*w17+x1*w16*/\
  "pshufw $0xdd,%%mm5,%%mm5        \n\t"/* x7 x5 x7 x5*/\
  "pmaddwd %%mm5,%%mm7             \n\t"/* x7*w23+x5*w22 x7*w19+x5*w18*/\
  "paddd " #A4 ",%%mm3                 \n\t"/* +%4*/\
  "pmaddwd 16+" #A3 ",%%mm0            \n\t"/* x2*w13+x0*w12 x2*w09+x0*w08*/\
  "paddd %%mm4,%%mm3               \n\t"/* 4     ; a1=sum(even1) a0=sum(even0)*/\
  "pmaddwd 24+" #A3 ",%%mm1            \n\t"/* x6*w15+x4*w14 x6*w11+x4*w10*/\
  "movq %%mm3,%%mm4                \n\t"/* 4     ; a1 a0*/\
  "pmaddwd 48+" #A3 ",%%mm2            \n\t"/* x3*w29+x1*w28 x3*w25+x1*w24*/\
  "paddd %%mm7,%%mm6               \n\t"/* 7     ; b1=sum(odd1) b0=sum(odd0)*/\
  "pmaddwd 56+" #A3 ",%%mm5            \n\t"/* x7*w31+x5*w30 x7*w27+x5*w26*/\
  "paddd %%mm6,%%mm3               \n\t"/* a1+b1 a0+b0*/\
  "paddd " #A4 ",%%mm0                 \n\t"/* +%4*/\
  "psrad $11,%%mm3     \n\t"/* y1=a1+b1 y0=a0+b0*/\
  "paddd %%mm1,%%mm0               \n\t"/* 1     ; a3=sum(even3) a2=sum(even2)*/\
  "psubd %%mm6,%%mm4               \n\t"/* 6     ; a1-b1 a0-b0*/\
  "movq %%mm0,%%mm7                \n\t"/* 7     ; a3 a2*/\
  "paddd %%mm5,%%mm2               \n\t"/* 5     ; b3=sum(odd3) b2=sum(odd2)*/\
  "paddd %%mm2,%%mm0               \n\t"/* a3+b3 a2+b2*/\
  "psrad $11,%%mm4     \n\t"/* y6=a1-b1 y7=a0-b0*/\
  "psubd %%mm2,%%mm7               \n\t"/* 2     ; a3-b3 a2-b2*/\
  "psrad $11,%%mm0     \n\t"/* y3=a3+b3 y2=a2+b2*/\
  "psrad $11,%%mm7     \n\t"/* y4=a3-b3 y5=a2-b2*/\
  "packssdw %%mm0,%%mm3            \n\t"/* 0     ; y3 y2 y1 y0*/\
  "packssdw %%mm4,%%mm7            \n\t"/* 4     ; y6 y7 y4 y5*/\
  "movq %%mm3, " #A2 "                  \n\t"/* 3     ; save y3 y2 y1 y0*/\
  "pshufw $0xb1,%%mm7,%%mm7        \n\t"/* y7 y6 y5 y4*/\
  "movq %%mm7,8                +" #A2 "\n\t"/* 7     ; save y7 y6 y5 y4*/\


//-----------------------------------------------------------------------------
//
// The first stage DCT 8x8 - forward DCTs of columns
//
// The %2puts are multiplied
// for rows 0,4 - on cos_4_16,
// for rows 1,7 - on cos_1_16,
// for rows 2,6 - on cos_2_16,
// for rows 3,5 - on cos_3_16
// and are shifted to the left for rise of accuracy
//
//-----------------------------------------------------------------------------
//
// The 8-point scaled forward DCT algorithm (26a8m)
//
//-----------------------------------------------------------------------------
//
// #define DCT_8_FRW_COL(x, y)
//{
// short t0, t1, t2, t3, t4, t5, t6, t7;
// short tp03, tm03, tp12, tm12, tp65, tm65;
// short tp465, tm465, tp765, tm765;
//
// t0 = LEFT_SHIFT ( x[0] + x[7] );
// t1 = LEFT_SHIFT ( x[1] + x[6] );
// t2 = LEFT_SHIFT ( x[2] + x[5] );
// t3 = LEFT_SHIFT ( x[3] + x[4] );
// t4 = LEFT_SHIFT ( x[3] - x[4] );
// t5 = LEFT_SHIFT ( x[2] - x[5] );
// t6 = LEFT_SHIFT ( x[1] - x[6] );
// t7 = LEFT_SHIFT ( x[0] - x[7] );
//
// tp03 = t0 + t3;
// tm03 = t0 - t3;
// tp12 = t1 + t2;
// tm12 = t1 - t2;
//
// y[0] = tp03 + tp12;
// y[4] = tp03 - tp12;
//
// y[2] = tm03 + tm12 * tg_2_16;
// y[6] = tm03 * tg_2_16 - tm12;
//
// tp65 =(t6 +t5 )*cos_4_16;
// tm65 =(t6 -t5 )*cos_4_16;
//
// tp765 = t7 + tp65;
// tm765 = t7 - tp65;
// tp465 = t4 + tm65;
// tm465 = t4 - tm65;
//
// y[1] = tp765 + tp465 * tg_1_16;
// y[7] = tp765 * tg_1_16 - tp465;
// y[5] = tm765 * tg_3_16 + tm465;
// y[3] = tm765 - tm465 * tg_3_16;
//}
//
//-----------------------------------------------------------------------------

//-----------------------------------------------------------------------------
// DCT_8_INV_COL_4  INP,OUT
//-----------------------------------------------------------------------------

#define DCT_8_INV_COL(A1,A2)\
  "movq 2*8(%3),%%mm0\n\t"\
  "movq 16*3+" #A1 ",%%mm3\n\t"\
  "movq %%mm0,%%mm1            \n\t"/* tg_3_16*/\
  "movq 16*5+" #A1 ",%%mm5\n\t"\
  "pmulhw %%mm3,%%mm0          \n\t"/* x3*(tg_3_16-1)*/\
  "movq (%3),%%mm4\n\t"\
  "pmulhw %%mm5,%%mm1          \n\t"/* x5*(tg_3_16-1)*/\
  "movq 16*7+" #A1 ",%%mm7\n\t"\
  "movq %%mm4,%%mm2            \n\t"/* tg_1_16*/\
  "movq 16*1+" #A1 ",%%mm6\n\t"\
  "pmulhw %%mm7,%%mm4          \n\t"/* x7*tg_1_16*/\
  "paddsw %%mm3,%%mm0          \n\t"/* x3*tg_3_16*/\
  "pmulhw %%mm6,%%mm2          \n\t"/* x1*tg_1_16*/\
  "paddsw %%mm3,%%mm1          \n\t"/* x3+x5*(tg_3_16-1)*/\
  "psubsw %%mm5,%%mm0          \n\t"/* x3*tg_3_16-x5 = tm35*/\
  "movq 3*8(%3),%%mm3\n\t"\
  "paddsw %%mm5,%%mm1          \n\t"/* x3+x5*tg_3_16 = tp35*/\
  "paddsw %%mm6,%%mm4          \n\t"/* x1+tg_1_16*x7 = tp17*/\
  "psubsw %%mm7,%%mm2          \n\t"/* x1*tg_1_16-x7 = tm17*/\
  "movq %%mm4,%%mm5            \n\t"/* tp17*/\
  "movq %%mm2,%%mm6            \n\t"/* tm17*/\
  "paddsw %%mm1,%%mm5          \n\t"/* tp17+tp35 = b0*/\
  "psubsw %%mm0,%%mm6          \n\t"/* tm17-tm35 = b3*/\
  "psubsw %%mm1,%%mm4          \n\t"/* tp17-tp35 = t1*/\
  "paddsw %%mm0,%%mm2          \n\t"/* tm17+tm35 = t2*/\
  "movq 1*8(%3),%%mm7\n\t"\
  "movq %%mm4,%%mm1            \n\t"/* t1*/\
  "movq %%mm5,3*16         +" #A2 "\n\t"/* save b0*/\
  "paddsw %%mm2,%%mm1          \n\t"/* t1+t2*/\
  "movq %%mm6,5*16         +" #A2 "\n\t"/* save b3*/\
  "psubsw %%mm2,%%mm4          \n\t"/* t1-t2*/\
  "movq 2*16+" #A1 ",%%mm5\n\t"\
  "movq %%mm7,%%mm0            \n\t"/* tg_2_16*/\
  "movq 6*16+" #A1 ",%%mm6\n\t"\
  "pmulhw %%mm5,%%mm0          \n\t"/* x2*tg_2_16*/\
  "pmulhw %%mm6,%%mm7          \n\t"/* x6*tg_2_16*/\
  "pmulhw %%mm3,%%mm1          \n\t"/* ocos_4_16*(t1+t2) = b1/2*/\
  "movq 0*16+" #A1 ",%%mm2\n\t"\
  "pmulhw %%mm3,%%mm4          \n\t"/* ocos_4_16*(t1-t2) = b2/2*/\
  "psubsw %%mm6,%%mm0          \n\t"/* t2*tg_2_16-x6 = tm26*/\
  "movq %%mm2,%%mm3            \n\t"/* x0*/\
  "movq 4*16+" #A1 ",%%mm6\n\t"\
  "paddsw %%mm5,%%mm7          \n\t"/* x2+x6*tg_2_16 = tp26*/\
  "paddsw %%mm6,%%mm2          \n\t"/* x0+x4 = tp04*/\
  "psubsw %%mm6,%%mm3          \n\t"/* x0-x4 = tm04*/\
  "movq %%mm2,%%mm5            \n\t"/* tp04*/\
  "movq %%mm3,%%mm6            \n\t"/* tm04*/\
  "psubsw %%mm7,%%mm2          \n\t"/* tp04-tp26 = a3*/\
  "paddsw %%mm0,%%mm3          \n\t"/* tm04+tm26 = a1*/\
  "paddsw %%mm1,%%mm1          \n\t"/* b1*/\
  "paddsw %%mm4,%%mm4          \n\t"/* b2*/\
  "paddsw %%mm7,%%mm5          \n\t"/* tp04+tp26 = a0*/\
  "psubsw %%mm0,%%mm6          \n\t"/* tm04-tm26 = a2*/\
  "movq %%mm3,%%mm7            \n\t"/* a1*/\
  "movq %%mm6,%%mm0            \n\t"/* a2*/\
  "paddsw %%mm1,%%mm3          \n\t"/* a1+b1*/\
  "paddsw %%mm4,%%mm6          \n\t"/* a2+b2*/\
  "psraw $6,%%mm3 \n\t"/* dst1*/\
  "psubsw %%mm1,%%mm7          \n\t"/* a1-b1*/\
  "psraw $6,%%mm6 \n\t"/* dst2*/\
  "psubsw %%mm4,%%mm0          \n\t"/* a2-b2*/\
  "movq 3*16+" #A2 ",%%mm1         \n\t"/* load b0*/\
  "psraw $6,%%mm7 \n\t"/* dst6*/\
  "movq %%mm5,%%mm4            \n\t"/* a0*/\
  "psraw $6,%%mm0 \n\t"/* dst5*/\
  "movq %%mm3,1*16+" #A2 "\n\t"\
  "paddsw %%mm1,%%mm5          \n\t"/* a0+b0*/\
  "movq %%mm6,2*16+" #A2 "\n\t"\
  "psubsw %%mm1,%%mm4          \n\t"/* a0-b0*/\
  "movq 5*16+" #A2 ",%%mm3         \n\t"/* load b3*/\
  "psraw $6,%%mm5 \n\t"/* dst0*/\
  "movq %%mm2,%%mm6            \n\t"/* a3*/\
  "psraw $6,%%mm4 \n\t"/* dst7*/\
  "movq %%mm0,5*16+" #A2 "\n\t"\
  "paddsw %%mm3,%%mm2          \n\t"/* a3+b3*/\
  "movq %%mm7,6*16+" #A2 "\n\t"\
  "psubsw %%mm3,%%mm6          \n\t"/* a3-b3*/\
  "movq %%mm5,0*16+" #A2 "\n\t"\
  "psraw $6,%%mm2 \n\t"/* dst3*/\
  "movq %%mm4,7*16+" #A2 "\n\t"\
  "psraw $6,%%mm6 \n\t"/* dst4*/\
  "movq %%mm2,3*16+" #A2 "\n\t"\
  "movq %%mm6,4*16+" #A2 "\n\t"

//=============================================================================
// Code
//=============================================================================

//-----------------------------------------------------------------------------
// void idct_mmx(uint16_t block[64]);
//-----------------------------------------------------------------------------


void ff_idct_xvid_mmx(short *block){
asm volatile(
            //# Process each row
    DCT_8_INV_ROW_MMX(0*16(%0), 0*16(%0), 64*0(%2), 8*0(%1))
    DCT_8_INV_ROW_MMX(1*16(%0), 1*16(%0), 64*1(%2), 8*1(%1))
    DCT_8_INV_ROW_MMX(2*16(%0), 2*16(%0), 64*2(%2), 8*2(%1))
    DCT_8_INV_ROW_MMX(3*16(%0), 3*16(%0), 64*3(%2), 8*3(%1))
    DCT_8_INV_ROW_MMX(4*16(%0), 4*16(%0), 64*0(%2), 8*4(%1))
    DCT_8_INV_ROW_MMX(5*16(%0), 5*16(%0), 64*3(%2), 8*5(%1))
    DCT_8_INV_ROW_MMX(6*16(%0), 6*16(%0), 64*2(%2), 8*6(%1))
    DCT_8_INV_ROW_MMX(7*16(%0), 7*16(%0), 64*1(%2), 8*7(%1))

            //# Process the columns (4 at a time)
    DCT_8_INV_COL(0(%0), 0(%0))
    DCT_8_INV_COL(8(%0), 8(%0))
    :: "r"(block), "r"(rounder_0), "r"(tab_i_04_mmx), "r"(tg_1_16));
}

//-----------------------------------------------------------------------------
// void idct_xmm(uint16_t block[64]);
//-----------------------------------------------------------------------------


void ff_idct_xvid_mmx2(short *block){
asm volatile(
            //# Process each row
    DCT_8_INV_ROW_XMM(0*16(%0), 0*16(%0), 64*0(%2), 8*0(%1))
    DCT_8_INV_ROW_XMM(1*16(%0), 1*16(%0), 64*1(%2), 8*1(%1))
    DCT_8_INV_ROW_XMM(2*16(%0), 2*16(%0), 64*2(%2), 8*2(%1))
    DCT_8_INV_ROW_XMM(3*16(%0), 3*16(%0), 64*3(%2), 8*3(%1))
    DCT_8_INV_ROW_XMM(4*16(%0), 4*16(%0), 64*0(%2), 8*4(%1))
    DCT_8_INV_ROW_XMM(5*16(%0), 5*16(%0), 64*3(%2), 8*5(%1))
    DCT_8_INV_ROW_XMM(6*16(%0), 6*16(%0), 64*2(%2), 8*6(%1))
    DCT_8_INV_ROW_XMM(7*16(%0), 7*16(%0), 64*1(%2), 8*7(%1))

            //# Process the columns (4 at a time)
    DCT_8_INV_COL(0(%0), 0(%0))
    DCT_8_INV_COL(8(%0), 8(%0))
    :: "r"(block), "r"(rounder_0), "r"(tab_i_04_xmm), "r"(tg_1_16));
}