📄 fdct_mmx.asm
字号:
paddd mm3, mm4 ; 4 ; a1=sum(even1) a0=sum(even0)
pmaddwd mm5, [TABLE_3+48] ; x5*w30+x1*w28 x5*w26+x1*w24
pmaddwd mm6, [TABLE_3+56] ; x7*w31+x3*w29 x7*w27+x3*w25
paddd mm1, mm7 ; 7 ; b1=sum(odd1) b0=sum(odd0)
paddd mm0, [round_frw_row] ; +rounder (y6,y4)
psrad mm3, SHIFT_FRW_ROW_CLIP1 ; (y2, y0)
paddd mm1, [round_frw_row] ; +rounder (y3,y1)
paddd mm0, mm2 ; 2 ; a3=sum(even3) a2=sum(even2)
paddd mm5, [round_frw_row] ; +rounder (y7,y5)
psrad mm1, SHIFT_FRW_ROW_CLIP1 ; y1=a1+b1 y0=a0+b0
paddd mm5, mm6 ; 6 ; b3=sum(odd3) b2=sum(odd2)
psrad mm0, SHIFT_FRW_ROW_CLIP1 ; y3=a3+b3 y2=a2+b2
psrad mm5, SHIFT_FRW_ROW_CLIP1 ; y4=a3-b3 y5=a2-b2
packssdw mm3, mm0 ; 0 ; y6 y4 y2 y0, saturate {-32768,+32767}
packssdw mm1, mm5 ; 3 ; y7 y5 y3 y1, saturate {-32768,+32767}
movq mm6, mm3 ; mm0 = y6 y4 y2 y0
punpcklwd mm3, mm1 ; y3 y2 y1 y0
punpckhwd mm6, mm1 ; y7 y6 y5 y4
psraw mm3, SHIFT_FRW_ROW_CLIP2 ; descale [y3 y2 y1 y0] to {-2048,+2047}
psraw mm6, SHIFT_FRW_ROW_CLIP2 ; descale [y7 y6 y5 y4] to {-2048,+2047}
movq [OUT_4-16], mm3 ; 1 ; save y3 y2 y1 y0
movq [OUT_4-8], mm6 ; 7 ; save y7 y6 y5 y4
movd mm5, [INP_4+12] ; mm5 = 7 6
punpcklwd mm5, [INP_4+8]
movq mm2, mm5 ; mm2 = 5 7 4 6
psrlq mm5, 32 ; mm5 = _ _ 5 7
movq mm0, [INP_4] ; mm0 = 3 2 1 0
punpcklwd mm5, mm2 ; mm5 = 4 5 6 7
movq mm1, mm0 ; mm1 = 3 2 1 0
paddsw mm0, mm5 ; mm0 = [3+4, 2+5, 1+6, 0+7] (xt3, xt2, xt1, xt0)
psubsw mm1, mm5 ; mm1 = [3-4, 2-5, 1-6, 0-7] (xt7, xt6, xt5, xt4)
movq mm2, mm0 ; mm2 = [ xt3 xt2 xt1 xt0 ]
punpcklwd mm0, mm1 ; mm0 = [ xt5 xt1 xt4 xt0 ]
punpckhwd mm2, mm1 ; mm2 = [ xt7 xt3 xt6 xt2 ]
movq mm1, mm2 ; mm1
movq mm2, mm0 ; 2 ; x3 x2 x1 x0
movq mm3, [TABLE_4] ; 3 ; w06 w04 w02 w00
punpcklwd mm0, mm1 ; x5 x1 x4 x0
movq mm5, mm0 ; 5 ; x5 x1 x4 x0
punpckldq mm0, mm0 ; x4 x0 x4 x0 [ xt2 xt0 xt2 xt0 ]
movq mm4, [TABLE_4+8] ; 4 ; w07 w05 w03 w01
punpckhwd mm2, mm1 ; 1 ; x7 x3 x6 x2
pmaddwd mm3, mm0 ; x4*w06+x0*w04 x4*w02+x0*w00
movq mm6, mm2 ; 6 ; x7 x3 x6 x2
movq mm1, [TABLE_4+32] ; 1 ; w22 w20 w18 w16
punpckldq mm2, mm2 ; x6 x2 x6 x2 [ xt3 xt1 xt3 xt1 ]
pmaddwd mm4, mm2 ; x6*w07+x2*w05 x6*w03+x2*w01
punpckhdq mm5, mm5 ; x5 x1 x5 x1 [ xt6 xt4 xt6 xt4 ]
pmaddwd mm0, [TABLE_4+16] ; x4*w14+x0*w12 x4*w10+x0*w08
punpckhdq mm6, mm6 ; x7 x3 x7 x3 [ xt7 xt5 xt7 xt5 ]
movq mm7, [TABLE_4+40] ; 7 ; w23 w21 w19 w17
pmaddwd mm1, mm5 ; x5*w22+x1*w20 x5*w18+x1*w16
paddd mm3, [round_frw_row] ; +rounder (y2,y0)
pmaddwd mm7, mm6 ; x7*w23+x3*w21 x7*w19+x3*w17
pmaddwd mm2, [TABLE_4+24] ; x6*w15+x2*w13 x6*w11+x2*w09
paddd mm3, mm4 ; 4 ; a1=sum(even1) a0=sum(even0)
pmaddwd mm5, [TABLE_4+48] ; x5*w30+x1*w28 x5*w26+x1*w24
pmaddwd mm6, [TABLE_4+56] ; x7*w31+x3*w29 x7*w27+x3*w25
paddd mm1, mm7 ; 7 ; b1=sum(odd1) b0=sum(odd0)
paddd mm0, [round_frw_row] ; +rounder (y6,y4)
psrad mm3, SHIFT_FRW_ROW_CLIP1 ; (y2, y0)
paddd mm1, [round_frw_row] ; +rounder (y3,y1)
paddd mm0, mm2 ; 2 ; a3=sum(even3) a2=sum(even2)
paddd mm5, [round_frw_row] ; +rounder (y7,y5)
psrad mm1, SHIFT_FRW_ROW_CLIP1 ; y1=a1+b1 y0=a0+b0
paddd mm5, mm6 ; 6 ; b3=sum(odd3) b2=sum(odd2)
psrad mm0, SHIFT_FRW_ROW_CLIP1 ; y3=a3+b3 y2=a2+b2
psrad mm5, SHIFT_FRW_ROW_CLIP1 ; y4=a3-b3 y5=a2-b2
packssdw mm3, mm0 ; 0 ; y6 y4 y2 y0, saturate {-32768,+32767}
packssdw mm1, mm5 ; 3 ; y7 y5 y3 y1, saturate {-32768,+32767}
movq mm6, mm3 ; mm0 = y6 y4 y2 y0
punpcklwd mm3, mm1 ; y3 y2 y1 y0
punpckhwd mm6, mm1 ; y7 y6 y5 y4
psraw mm3, SHIFT_FRW_ROW_CLIP2 ; descale [y3 y2 y1 y0] to {-2048,+2047}
psraw mm6, SHIFT_FRW_ROW_CLIP2 ; descale [y7 y6 y5 y4] to {-2048,+2047}
movq [OUT_5-16], mm3 ; 1 ; save y3 y2 y1 y0
movq [OUT_5-8], mm6 ; 7 ; save y7 y6 y5 y4
movd mm5, [INP_5+12] ; mm5 = 7 6
punpcklwd mm5, [INP_5+8]
movq mm2, mm5 ; mm2 = 5 7 4 6
psrlq mm5, 32 ; mm5 = _ _ 5 7
movq mm0, [INP_5] ; mm0 = 3 2 1 0
punpcklwd mm5, mm2 ; mm5 = 4 5 6 7
movq mm1, mm0 ; mm1 = 3 2 1 0
paddsw mm0, mm5 ; mm0 = [3+4, 2+5, 1+6, 0+7] (xt3, xt2, xt1, xt0)
psubsw mm1, mm5 ; mm1 = [3-4, 2-5, 1-6, 0-7] (xt7, xt6, xt5, xt4)
movq mm2, mm0 ; mm2 = [ xt3 xt2 xt1 xt0 ]
punpcklwd mm0, mm1 ; mm0 = [ xt5 xt1 xt4 xt0 ]
punpckhwd mm2, mm1 ; mm2 = [ xt7 xt3 xt6 xt2 ]
movq mm1, mm2 ; mm1
movq mm2, mm0 ; 2 ; x3 x2 x1 x0
movq mm3, [TABLE_5] ; 3 ; w06 w04 w02 w00
punpcklwd mm0, mm1 ; x5 x1 x4 x0
movq mm5, mm0 ; 5 ; x5 x1 x4 x0
punpckldq mm0, mm0 ; x4 x0 x4 x0 [ xt2 xt0 xt2 xt0 ]
movq mm4, [TABLE_5+8] ; 4 ; w07 w05 w03 w01
punpckhwd mm2, mm1 ; 1 ; x7 x3 x6 x2
pmaddwd mm3, mm0 ; x4*w06+x0*w04 x4*w02+x0*w00
movq mm6, mm2 ; 6 ; x7 x3 x6 x2
movq mm1, [TABLE_5+32] ; 1 ; w22 w20 w18 w16
punpckldq mm2, mm2 ; x6 x2 x6 x2 [ xt3 xt1 xt3 xt1 ]
pmaddwd mm4, mm2 ; x6*w07+x2*w05 x6*w03+x2*w01
punpckhdq mm5, mm5 ; x5 x1 x5 x1 [ xt6 xt4 xt6 xt4 ]
pmaddwd mm0, [TABLE_5+16] ; x4*w14+x0*w12 x4*w10+x0*w08
punpckhdq mm6, mm6 ; x7 x3 x7 x3 [ xt7 xt5 xt7 xt5 ]
movq mm7, [TABLE_5+40] ; 7 ; w23 w21 w19 w17
pmaddwd mm1, mm5 ; x5*w22+x1*w20 x5*w18+x1*w16
paddd mm3, [round_frw_row] ; +rounder (y2,y0)
pmaddwd mm7, mm6 ; x7*w23+x3*w21 x7*w19+x3*w17
pmaddwd mm2, [TABLE_5+24] ; x6*w15+x2*w13 x6*w11+x2*w09
paddd mm3, mm4 ; 4 ; a1=sum(even1) a0=sum(even0)
pmaddwd mm5, [TABLE_5+48] ; x5*w30+x1*w28 x5*w26+x1*w24
pmaddwd mm6, [TABLE_5+56] ; x7*w31+x3*w29 x7*w27+x3*w25
paddd mm1, mm7 ; 7 ; b1=sum(odd1) b0=sum(odd0)
paddd mm0, [round_frw_row] ; +rounder (y6,y4)
psrad mm3, SHIFT_FRW_ROW_CLIP1 ; (y2, y0)
paddd mm1, [round_frw_row] ; +rounder (y3,y1)
paddd mm0, mm2 ; 2 ; a3=sum(even3) a2=sum(even2)
paddd mm5, [round_frw_row] ; +rounder (y7,y5)
psrad mm1, SHIFT_FRW_ROW_CLIP1 ; y1=a1+b1 y0=a0+b0
paddd mm5, mm6 ; 6 ; b3=sum(odd3) b2=sum(odd2)
psrad mm0, SHIFT_FRW_ROW_CLIP1 ; y3=a3+b3 y2=a2+b2
psrad mm5, SHIFT_FRW_ROW_CLIP1 ; y4=a3-b3 y5=a2-b2
packssdw mm3, mm0 ; 0 ; y6 y4 y2 y0, saturate {-32768,+32767}
packssdw mm1, mm5 ; 3 ; y7 y5 y3 y1, saturate {-32768,+32767}
movq mm6, mm3 ; mm0 = y6 y4 y2 y0
punpcklwd mm3, mm1 ; y3 y2 y1 y0
punpckhwd mm6, mm1 ; y7 y6 y5 y4
psraw mm3, SHIFT_FRW_ROW_CLIP2 ; descale [y3 y2 y1 y0] to {-2048,+2047}
psraw mm6, SHIFT_FRW_ROW_CLIP2 ; descale [y7 y6 y5 y4] to {-2048,+2047}
movq [OUT_6-16], mm3 ; 1 ; save y3 y2 y1 y0
movq [OUT_6-8], mm6 ; 7 ; save y7 y6 y5 y4
movd mm5, [INP_6+12] ; mm5 = 7 6
punpcklwd mm5, [INP_6+8]
movq mm2, mm5 ; mm2 = 5 7 4 6
psrlq mm5, 32 ; mm5 = _ _ 5 7
movq mm0, [INP_6] ; mm0 = 3 2 1 0
punpcklwd mm5, mm2 ; mm5 = 4 5 6 7
movq mm1, mm0 ; mm1 = 3 2 1 0
paddsw mm0, mm5 ; mm0 = [3+4, 2+5, 1+6, 0+7] (xt3, xt2, xt1, xt0)
psubsw mm1, mm5 ; mm1 = [3-4, 2-5, 1-6, 0-7] (xt7, xt6, xt5, xt4)
movq mm2, mm0 ; mm2 = [ xt3 xt2 xt1 xt0 ]
punpcklwd mm0, mm1 ; mm0 = [ xt5 xt1 xt4 xt0 ]
punpckhwd mm2, mm1 ; mm2 = [ xt7 xt3 xt6 xt2 ]
movq mm1, mm2 ; mm1
movq mm2, mm0 ; 2 ; x3 x2 x1 x0
movq mm3, [TABLE_6] ; 3 ; w06 w04 w02 w00
punpcklwd mm0, mm1 ; x5 x1 x4 x0
movq mm5, mm0 ; 5 ; x5 x1 x4 x0
punpckldq mm0, mm0 ; x4 x0 x4 x0 [ xt2 xt0 xt2 xt0 ]
movq mm4, [TABLE_6+8] ; 4 ; w07 w05 w03 w01
punpckhwd mm2, mm1 ; 1 ; x7 x3 x6 x2
pmaddwd mm3, mm0 ; x4*w06+x0*w04 x4*w02+x0*w00
movq mm6, mm2 ; 6 ; x7 x3 x6 x2
movq mm1, [TABLE_6+32] ; 1 ; w22 w20 w18 w16
punpckldq mm2, mm2 ; x6 x2 x6 x2 [ xt3 xt1 xt3 xt1 ]
pmaddwd mm4, mm2 ; x6*w07+x2*w05 x6*w03+x2*w01
punpckhdq mm5, mm5 ; x5 x1 x5 x1 [ xt6 xt4 xt6 xt4 ]
pmaddwd mm0, [TABLE_6+16] ; x4*w14+x0*w12 x4*w10+x0*w08
punpckhdq mm6, mm6 ; x7 x3 x7 x3 [ xt7 xt5 xt7 xt5 ]
movq mm7, [TABLE_6+40] ; 7 ; w23 w21 w19 w17
pmaddwd mm1, mm5 ; x5*w22+x1*w20 x5*w18+x1*w16
paddd mm3, [round_frw_row] ; +rounder (y2,y0)
pmaddwd mm7, mm6 ; x7*w23+x3*w21 x7*w19+x3*w17
pmaddwd mm2, [TABLE_6+24] ; x6*w15+x2*w13 x6*w11+x2*w09
paddd mm3, mm4 ; 4 ; a1=sum(even1) a0=sum(even0)
pmaddwd mm5, [TABLE_6+48] ; x5*w30+x1*w28 x5*w26+x1*w24
pmaddwd mm6, [TABLE_6+56] ; x7*w31+x3*w29 x7*w27+x3*w25
paddd mm1, mm7 ; 7 ; b1=sum(odd1) b0=sum(odd0)
paddd mm0, [round_frw_row] ; +rounder (y6,y4)
psrad mm3, SHIFT_FRW_ROW_CLIP1 ; (y2, y0)
paddd mm1, [round_frw_row] ; +rounder (y3,y1)
paddd mm0, mm2 ; 2 ; a3=sum(even3) a2=sum(even2)
paddd mm5, [round_frw_row] ; +rounder (y7,y5)
psrad mm1, SHIFT_FRW_ROW_CLIP1 ; y1=a1+b1 y0=a0+b0
paddd mm5, mm6 ; 6 ; b3=sum(odd3) b2=sum(odd2)
psrad mm0, SHIFT_FRW_ROW_CLIP1 ; y3=a3+b3 y2=a2+b2
psrad mm5, SHIFT_FRW_ROW_CLIP1 ; y4=a3-b3 y5=a2-b2
packssdw mm3, mm0 ; 0 ; y6 y4 y2 y0, saturate {-32768,+32767}
packssdw mm1, mm5 ; 3 ; y7 y5 y3 y1, saturate {-32768,+32767}
movq mm6, mm3 ; mm0 = y6 y4 y2 y0
punpcklwd mm3, mm1 ; y3 y2 y1 y0
punpckhwd mm6, mm1 ; y7 y6 y5 y4
psraw mm3, SHIFT_FRW_ROW_CLIP2 ; descale [y3 y2 y1 y0] to {-2048,+2047}
psraw mm6, SHIFT_FRW_ROW_CLIP2 ; descale [y7 y6 y5 y4] to {-2048,+2047}
movq [OUT_7-16], mm3 ; 1 ; save y3 y2 y1 y0
movq [OUT_7-8], mm6 ; 7 ; save y7 y6 y5 y4
movd mm5, [INP_7+12] ; mm5 = 7 6
punpcklwd mm5, [INP_7+8]
movq mm2, mm5 ; mm2 = 5 7 4 6
psrlq mm5, 32 ; mm5 = _ _ 5 7
movq mm0, [INP_7] ; mm0 = 3 2 1 0
punpcklwd mm5, mm2 ; mm5 = 4 5 6 7
movq mm1, mm0 ; mm1 = 3 2 1 0
paddsw mm0, mm5 ; mm0 = [3+4, 2+5, 1+6, 0+7] (xt3, xt2, xt1, xt0)
psubsw mm1, mm5 ; mm1 = [3-4, 2-5, 1-6, 0-7] (xt7, xt6, xt5, xt4)
movq mm2, mm0 ; mm2 = [ xt3 xt2 xt1 xt0 ]
punpcklwd mm0, mm1 ; mm0 = [ xt5 xt1 xt4 xt0 ]
punpckhwd mm2, mm1 ; mm2 = [ xt7 xt3 xt6 xt2 ]
movq mm1, mm2 ; mm1
movq mm2, mm0 ; 2 ; x3 x2 x1 x0
movq mm3, [TABLE_7] ; 3 ; w06 w04 w02 w00
punpcklwd mm0, mm1 ; x5 x1 x4 x0
movq mm5, mm0 ; 5 ; x5 x1 x4 x0
punpckldq mm0, mm0 ; x4 x0 x4 x0 [ xt2 xt0 xt2 xt0 ]
movq mm4, [TABLE_7+8] ; 4 ; w07 w05 w03 w01
punpckhwd mm2, mm1 ; 1 ; x7 x3 x6 x2
pmaddwd mm3, mm0 ; x4*w06+x0*w04 x4*w02+x0*w00
movq mm6, mm2 ; 6 ; x7 x3 x6 x2
movq mm1, [TABLE_7+32] ; 1 ; w22 w20 w18 w16
punpckldq mm2, mm2 ; x6 x2 x6 x2 [ xt3 xt1 xt3 xt1 ]
pmaddwd mm4, mm2 ; x6*w07+x2*w05 x6*w03+x2*w01
punpckhdq mm5, mm5 ; x5 x1 x5 x1 [ xt6 xt4 xt6 xt4 ]
pmaddwd mm0, [TABLE_7+16] ; x4*w14+x0*w12 x4*w10+x0*w08
punpckhdq mm6, mm6 ; x7 x3 x7 x3 [ xt7 xt5 xt7 xt5 ]
movq mm7, [TABLE_7+40] ; 7 ; w23 w21 w19 w17
pmaddwd mm1, mm5 ; x5*w22+x1*w20 x5*w18+x1*w16
paddd mm3, [round_frw_row] ; +rounder (y2,y0)
pmaddwd mm7, mm6 ; x7*w23+x3*w21 x7*w19+x3*w17
pmaddwd mm2, [TABLE_7+24] ; x6*w15+x2*w13 x6*w11+x2*w09
paddd mm3, mm4 ; 4 ; a1=sum(even1) a0=sum(even0)
pmaddwd mm5, [TABLE_7+48] ; x5*w30+x1*w28 x5*w26+x1*w24
pmaddwd mm6, [TABLE_7+56] ; x7*w31+x3*w29 x7*w27+x3*w25
paddd mm1, mm7 ; 7 ; b1=sum(odd1) b0=sum(odd0)
paddd mm0, [round_frw_row] ; +rounder (y6,y4)
psrad mm3, SHIFT_FRW_ROW_CLIP1 ; (y2, y0)
paddd mm1, [round_frw_row] ; +rounder (y3,y1)
paddd mm0, mm2 ; 2 ; a3=sum(even3) a2=sum(even2)
paddd mm5, [round_frw_row] ; +rounder (y7,y5)
psrad mm1, SHIFT_FRW_ROW_CLIP1 ; y1=a1+b1 y0=a0+b0
paddd mm5, mm6 ; 6 ; b3=sum(odd3) b2=sum(odd2)
psrad mm0, SHIFT_FRW_ROW_CLIP1 ; y3=a3+b3 y2=a2+b2
psrad mm5, SHIFT_FRW_ROW_CLIP1 ; y4=a3-b3 y5=a2-b2
packssdw mm3, mm0 ; 0 ; y6 y4 y2 y0, saturate {-32768,+32767}
packssdw mm1, mm5 ; 3 ; y7 y5 y3 y1, saturate {-32768,+32767}
movq mm6, mm3 ; mm0 = y6 y4 y2 y0
punpcklwd mm3, mm1 ; y3 y2 y1 y0
punpckhwd mm6, mm1 ; y7 y6 y5 y4
psraw mm3, SHIFT_FRW_ROW_CLIP2 ; descale [y3 y2 y1 y0] to {-2048,+2047}
psraw mm6, SHIFT_FRW_ROW_CLIP2 ; descale [y7 y6 y5 y4] to {-2048,+2047}
movq [OUT_8-16], mm3 ; 1 ; save y3 y2 y1 y0
movq [OUT_8-8], mm6 ; 7 ; save y7 y6 y5 y4
pop ebx
emms
ret
⌨️ 快捷键说明
复制代码
Ctrl + C
搜索代码
Ctrl + F
全屏模式
F11
切换主题
Ctrl + Shift + D
显示快捷键
?
增大字号
Ctrl + =
减小字号
Ctrl + -