📄 swscale_template.c
字号:
"paddw %%mm1, %%mm4 \n\t"\
"punpckhwd %%mm0, %%mm0 \n\t"\
"punpckhwd %%mm6, %%mm6 \n\t"\
"punpckhwd %%mm3, %%mm3 \n\t"\
"paddw %%mm7, %%mm0 \n\t"\
"paddw %%mm7, %%mm6 \n\t"\
"paddw %%mm7, %%mm3 \n\t"\
/* mm0=B1, mm2=B2, mm3=G2, mm4=G1, mm5=R1, mm6=R2 */\
"packuswb %%mm0, %%mm2 \n\t"\
"packuswb %%mm6, %%mm5 \n\t"\
"packuswb %%mm3, %%mm4 \n\t"\
"pxor %%mm7, %%mm7 \n\t"
#define YSCALEYUV2RGB1(index, c) REAL_YSCALEYUV2RGB1(index, c)
#define REAL_YSCALEYUV2PACKED1b(index, c) \
"xor "#index", "#index" \n\t"\
ASMALIGN(4)\
"1: \n\t"\
"movq (%2, "#index"), %%mm2 \n\t" /* uvbuf0[eax]*/\
"movq (%3, "#index"), %%mm3 \n\t" /* uvbuf1[eax]*/\
"movq 4096(%2, "#index"), %%mm5 \n\t" /* uvbuf0[eax+2048]*/\
"movq 4096(%3, "#index"), %%mm4 \n\t" /* uvbuf1[eax+2048]*/\
"paddw %%mm2, %%mm3 \n\t" /* uvbuf0[eax] + uvbuf1[eax]*/\
"paddw %%mm5, %%mm4 \n\t" /* uvbuf0[eax+2048] + uvbuf1[eax+2048]*/\
"psrlw $8, %%mm3 \n\t" \
"psrlw $8, %%mm4 \n\t" \
"movq (%0, "#index", 2), %%mm1 \n\t" /*buf0[eax]*/\
"movq 8(%0, "#index", 2), %%mm7 \n\t" /*buf0[eax]*/\
"psraw $7, %%mm1 \n\t" \
"psraw $7, %%mm7 \n\t"
#define YSCALEYUV2PACKED1b(index, c) REAL_YSCALEYUV2PACKED1b(index, c)
// do vertical chrominance interpolation
#define REAL_YSCALEYUV2RGB1b(index, c) \
"xor "#index", "#index" \n\t"\
ASMALIGN(4)\
"1: \n\t"\
"movq (%2, "#index"), %%mm2 \n\t" /* uvbuf0[eax]*/\
"movq (%3, "#index"), %%mm3 \n\t" /* uvbuf1[eax]*/\
"movq 4096(%2, "#index"), %%mm5 \n\t" /* uvbuf0[eax+2048]*/\
"movq 4096(%3, "#index"), %%mm4 \n\t" /* uvbuf1[eax+2048]*/\
"paddw %%mm2, %%mm3 \n\t" /* uvbuf0[eax] + uvbuf1[eax]*/\
"paddw %%mm5, %%mm4 \n\t" /* uvbuf0[eax+2048] + uvbuf1[eax+2048]*/\
"psrlw $5, %%mm3 \n\t" /*FIXME might overflow*/\
"psrlw $5, %%mm4 \n\t" /*FIXME might overflow*/\
"psubw "U_OFFSET"("#c"), %%mm3 \n\t" /* (U-128)8*/\
"psubw "V_OFFSET"("#c"), %%mm4 \n\t" /* (V-128)8*/\
"movq %%mm3, %%mm2 \n\t" /* (U-128)8*/\
"movq %%mm4, %%mm5 \n\t" /* (V-128)8*/\
"pmulhw "UG_COEFF"("#c"), %%mm3 \n\t"\
"pmulhw "VG_COEFF"("#c"), %%mm4 \n\t"\
/* mm2=(U-128)8, mm3=ug, mm4=vg mm5=(V-128)8 */\
"movq (%0, "#index", 2), %%mm1 \n\t" /*buf0[eax]*/\
"movq 8(%0, "#index", 2), %%mm7 \n\t" /*buf0[eax]*/\
"psraw $4, %%mm1 \n\t" /* buf0[eax] - buf1[eax] >>4*/\
"psraw $4, %%mm7 \n\t" /* buf0[eax] - buf1[eax] >>4*/\
"pmulhw "UB_COEFF"("#c"), %%mm2 \n\t"\
"pmulhw "VR_COEFF"("#c"), %%mm5 \n\t"\
"psubw "Y_OFFSET"("#c"), %%mm1 \n\t" /* 8(Y-16)*/\
"psubw "Y_OFFSET"("#c"), %%mm7 \n\t" /* 8(Y-16)*/\
"pmulhw "Y_COEFF"("#c"), %%mm1 \n\t"\
"pmulhw "Y_COEFF"("#c"), %%mm7 \n\t"\
/* mm1= Y1, mm2=ub, mm3=ug, mm4=vg mm5=vr, mm7=Y2 */\
"paddw %%mm3, %%mm4 \n\t"\
"movq %%mm2, %%mm0 \n\t"\
"movq %%mm5, %%mm6 \n\t"\
"movq %%mm4, %%mm3 \n\t"\
"punpcklwd %%mm2, %%mm2 \n\t"\
"punpcklwd %%mm5, %%mm5 \n\t"\
"punpcklwd %%mm4, %%mm4 \n\t"\
"paddw %%mm1, %%mm2 \n\t"\
"paddw %%mm1, %%mm5 \n\t"\
"paddw %%mm1, %%mm4 \n\t"\
"punpckhwd %%mm0, %%mm0 \n\t"\
"punpckhwd %%mm6, %%mm6 \n\t"\
"punpckhwd %%mm3, %%mm3 \n\t"\
"paddw %%mm7, %%mm0 \n\t"\
"paddw %%mm7, %%mm6 \n\t"\
"paddw %%mm7, %%mm3 \n\t"\
/* mm0=B1, mm2=B2, mm3=G2, mm4=G1, mm5=R1, mm6=R2 */\
"packuswb %%mm0, %%mm2 \n\t"\
"packuswb %%mm6, %%mm5 \n\t"\
"packuswb %%mm3, %%mm4 \n\t"\
"pxor %%mm7, %%mm7 \n\t"
#define YSCALEYUV2RGB1b(index, c) REAL_YSCALEYUV2RGB1b(index, c)
#define REAL_WRITEBGR32(dst, dstw, index) \
/* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */\
"movq %%mm2, %%mm1 \n\t" /* B */\
"movq %%mm5, %%mm6 \n\t" /* R */\
"punpcklbw %%mm4, %%mm2 \n\t" /* GBGBGBGB 0 */\
"punpcklbw %%mm7, %%mm5 \n\t" /* 0R0R0R0R 0 */\
"punpckhbw %%mm4, %%mm1 \n\t" /* GBGBGBGB 2 */\
"punpckhbw %%mm7, %%mm6 \n\t" /* 0R0R0R0R 2 */\
"movq %%mm2, %%mm0 \n\t" /* GBGBGBGB 0 */\
"movq %%mm1, %%mm3 \n\t" /* GBGBGBGB 2 */\
"punpcklwd %%mm5, %%mm0 \n\t" /* 0RGB0RGB 0 */\
"punpckhwd %%mm5, %%mm2 \n\t" /* 0RGB0RGB 1 */\
"punpcklwd %%mm6, %%mm1 \n\t" /* 0RGB0RGB 2 */\
"punpckhwd %%mm6, %%mm3 \n\t" /* 0RGB0RGB 3 */\
\
MOVNTQ(%%mm0, (dst, index, 4))\
MOVNTQ(%%mm2, 8(dst, index, 4))\
MOVNTQ(%%mm1, 16(dst, index, 4))\
MOVNTQ(%%mm3, 24(dst, index, 4))\
\
"add $8, "#index" \n\t"\
"cmp "#dstw", "#index" \n\t"\
" jb 1b \n\t"
#define WRITEBGR32(dst, dstw, index) REAL_WRITEBGR32(dst, dstw, index)
#define REAL_WRITEBGR16(dst, dstw, index) \
"pand "MANGLE(bF8)", %%mm2 \n\t" /* B */\
"pand "MANGLE(bFC)", %%mm4 \n\t" /* G */\
"pand "MANGLE(bF8)", %%mm5 \n\t" /* R */\
"psrlq $3, %%mm2 \n\t"\
\
"movq %%mm2, %%mm1 \n\t"\
"movq %%mm4, %%mm3 \n\t"\
\
"punpcklbw %%mm7, %%mm3 \n\t"\
"punpcklbw %%mm5, %%mm2 \n\t"\
"punpckhbw %%mm7, %%mm4 \n\t"\
"punpckhbw %%mm5, %%mm1 \n\t"\
\
"psllq $3, %%mm3 \n\t"\
"psllq $3, %%mm4 \n\t"\
\
"por %%mm3, %%mm2 \n\t"\
"por %%mm4, %%mm1 \n\t"\
\
MOVNTQ(%%mm2, (dst, index, 2))\
MOVNTQ(%%mm1, 8(dst, index, 2))\
\
"add $8, "#index" \n\t"\
"cmp "#dstw", "#index" \n\t"\
" jb 1b \n\t"
#define WRITEBGR16(dst, dstw, index) REAL_WRITEBGR16(dst, dstw, index)
#define REAL_WRITEBGR15(dst, dstw, index) \
"pand "MANGLE(bF8)", %%mm2 \n\t" /* B */\
"pand "MANGLE(bF8)", %%mm4 \n\t" /* G */\
"pand "MANGLE(bF8)", %%mm5 \n\t" /* R */\
"psrlq $3, %%mm2 \n\t"\
"psrlq $1, %%mm5 \n\t"\
\
"movq %%mm2, %%mm1 \n\t"\
"movq %%mm4, %%mm3 \n\t"\
\
"punpcklbw %%mm7, %%mm3 \n\t"\
"punpcklbw %%mm5, %%mm2 \n\t"\
"punpckhbw %%mm7, %%mm4 \n\t"\
"punpckhbw %%mm5, %%mm1 \n\t"\
\
"psllq $2, %%mm3 \n\t"\
"psllq $2, %%mm4 \n\t"\
\
"por %%mm3, %%mm2 \n\t"\
"por %%mm4, %%mm1 \n\t"\
\
MOVNTQ(%%mm2, (dst, index, 2))\
MOVNTQ(%%mm1, 8(dst, index, 2))\
\
"add $8, "#index" \n\t"\
"cmp "#dstw", "#index" \n\t"\
" jb 1b \n\t"
#define WRITEBGR15(dst, dstw, index) REAL_WRITEBGR15(dst, dstw, index)
#define WRITEBGR24OLD(dst, dstw, index) \
/* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */\
"movq %%mm2, %%mm1 \n\t" /* B */\
"movq %%mm5, %%mm6 \n\t" /* R */\
"punpcklbw %%mm4, %%mm2 \n\t" /* GBGBGBGB 0 */\
"punpcklbw %%mm7, %%mm5 \n\t" /* 0R0R0R0R 0 */\
"punpckhbw %%mm4, %%mm1 \n\t" /* GBGBGBGB 2 */\
"punpckhbw %%mm7, %%mm6 \n\t" /* 0R0R0R0R 2 */\
"movq %%mm2, %%mm0 \n\t" /* GBGBGBGB 0 */\
"movq %%mm1, %%mm3 \n\t" /* GBGBGBGB 2 */\
"punpcklwd %%mm5, %%mm0 \n\t" /* 0RGB0RGB 0 */\
"punpckhwd %%mm5, %%mm2 \n\t" /* 0RGB0RGB 1 */\
"punpcklwd %%mm6, %%mm1 \n\t" /* 0RGB0RGB 2 */\
"punpckhwd %%mm6, %%mm3 \n\t" /* 0RGB0RGB 3 */\
\
"movq %%mm0, %%mm4 \n\t" /* 0RGB0RGB 0 */\
"psrlq $8, %%mm0 \n\t" /* 00RGB0RG 0 */\
"pand "MANGLE(bm00000111)", %%mm4 \n\t" /* 00000RGB 0 */\
"pand "MANGLE(bm11111000)", %%mm0 \n\t" /* 00RGB000 0.5 */\
"por %%mm4, %%mm0 \n\t" /* 00RGBRGB 0 */\
"movq %%mm2, %%mm4 \n\t" /* 0RGB0RGB 1 */\
"psllq $48, %%mm2 \n\t" /* GB000000 1 */\
"por %%mm2, %%mm0 \n\t" /* GBRGBRGB 0 */\
\
"movq %%mm4, %%mm2 \n\t" /* 0RGB0RGB 1 */\
"psrld $16, %%mm4 \n\t" /* 000R000R 1 */\
"psrlq $24, %%mm2 \n\t" /* 0000RGB0 1.5 */\
"por %%mm4, %%mm2 \n\t" /* 000RRGBR 1 */\
"pand "MANGLE(bm00001111)", %%mm2 \n\t" /* 0000RGBR 1 */\
"movq %%mm1, %%mm4 \n\t" /* 0RGB0RGB 2 */\
"psrlq $8, %%mm1 \n\t" /* 00RGB0RG 2 */\
"pand "MANGLE(bm00000111)", %%mm4 \n\t" /* 00000RGB 2 */\
"pand "MANGLE(bm11111000)", %%mm1 \n\t" /* 00RGB000 2.5 */\
"por %%mm4, %%mm1 \n\t" /* 00RGBRGB 2 */\
"movq %%mm1, %%mm4 \n\t" /* 00RGBRGB 2 */\
"psllq $32, %%mm1 \n\t" /* BRGB0000 2 */\
"por %%mm1, %%mm2 \n\t" /* BRGBRGBR 1 */\
\
"psrlq $32, %%mm4 \n\t" /* 000000RG 2.5 */\
"movq %%mm3, %%mm5 \n\t" /* 0RGB0RGB 3 */\
"psrlq $8, %%mm3 \n\t" /* 00RGB0RG 3 */\
"pand "MANGLE(bm00000111)", %%mm5 \n\t" /* 00000RGB 3 */\
"pand "MANGLE(bm11111000)", %%mm3 \n\t" /* 00RGB000 3.5 */\
"por %%mm5, %%mm3 \n\t" /* 00RGBRGB 3 */\
"psllq $16, %%mm3 \n\t" /* RGBRGB00 3 */\
"por %%mm4, %%mm3 \n\t" /* RGBRGBRG 2.5 */\
\
MOVNTQ(%%mm0, (dst))\
MOVNTQ(%%mm2, 8(dst))\
MOVNTQ(%%mm3, 16(dst))\
"add $24, "#dst" \n\t"\
\
"add $8, "#index" \n\t"\
"cmp "#dstw", "#index" \n\t"\
" jb 1b \n\t"
#define WRITEBGR24MMX(dst, dstw, index) \
/* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */\
"movq %%mm2, %%mm1 \n\t" /* B */\
"movq %%mm5, %%mm6 \n\t" /* R */\
"punpcklbw %%mm4, %%mm2 \n\t" /* GBGBGBGB 0 */\
"punpcklbw %%mm7, %%mm5 \n\t" /* 0R0R0R0R 0 */\
"punpckhbw %%mm4, %%mm1 \n\t" /* GBGBGBGB 2 */\
"punpckhbw %%mm7, %%mm6 \n\t" /* 0R0R0R0R 2 */\
"movq %%mm2, %%mm0 \n\t" /* GBGBGBGB 0 */\
"movq %%mm1, %%mm3 \n\t" /* GBGBGBGB 2 */\
"punpcklwd %%mm5, %%mm0 \n\t" /* 0RGB0RGB 0 */\
"punpckhwd %%mm5, %%mm2 \n\t" /* 0RGB0RGB 1 */\
"punpcklwd %%mm6, %%mm1 \n\t" /* 0RGB0RGB 2 */\
"punpckhwd %%mm6, %%mm3 \n\t" /* 0RGB0RGB 3 */\
\
"movq %%mm0, %%mm4 \n\t" /* 0RGB0RGB 0 */\
"movq %%mm2, %%mm6 \n\t" /* 0RGB0RGB 1 */\
"movq %%mm1, %%mm5 \n\t" /* 0RGB0RGB 2 */\
"movq %%mm3, %%mm7 \n\t" /* 0RGB0RGB 3 */\
\
"psllq $40, %%mm0 \n\t" /* RGB00000 0 */\
"psllq $40, %%mm2 \n\t" /* RGB00000 1 */\
"psllq $40, %%mm1 \n\t" /* RGB00000 2 */\
"psllq $40, %%mm3 \n\t" /* RGB00000 3 */\
\
"punpckhdq %%mm4, %%mm0 \n\t" /* 0RGBRGB0 0 */\
"punpckhdq %%mm6, %%mm2 \n\t" /* 0RGBRGB0 1 */\
"punpckhdq %%mm5, %%mm1 \n\t" /* 0RGBRGB0 2 */\
"punpckhdq %%mm7, %%mm3 \n\t" /* 0RGBRGB0 3 */\
\
"psrlq $8, %%mm0 \n\t" /* 00RGBRGB 0 */\
"movq %%mm2, %%mm6 \n\t" /* 0RGBRGB0 1 */\
"psllq $40, %%mm2 \n\t" /* GB000000 1 */\
"por %%mm2, %%mm0 \n\t" /* GBRGBRGB 0 */\
MOVNTQ(%%mm0, (dst))\
\
"psrlq $24, %%mm6 \n\t" /* 0000RGBR 1 */\
"movq %%mm1, %%mm5 \n\t" /* 0RGBRGB0 2 */\
"psllq $24, %%mm1 \n\t" /* BRGB0000 2 */\
"por %%mm1, %%mm6 \n\t" /* BRGBRGBR 1 */\
MOVNTQ(%%mm6, 8(dst))\
\
"psrlq $40, %%mm5 \n\t" /* 000000RG 2 */\
"psllq $8, %%mm3 \n\t" /* RGBRGB00 3 */\
"por %%mm3, %%mm5 \n\t" /* RGBRGBRG 2 */\
MOVNTQ(%%mm5, 16(dst))\
\
"add $24, "#dst" \n\t"\
\
"add $8, "#index" \n\t"\
"cmp "#dstw", "#index" \n\t"\
" jb 1b \n\t"
#define WRITEBGR24MMX2(dst, dstw, index) \
/* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */\
"movq "MANGLE(M24A)", %%mm0 \n\t"\
"movq "MANGLE(M24C)", %%mm7 \n\t"\
"pshufw $0x50, %%mm2, %%mm1 \n\t" /* B3 B2 B3 B2 B1 B0 B1 B0 */\
"pshufw $0x50, %%mm4, %%mm3 \n\t" /* G3 G2 G3 G2 G1 G0 G1 G0 */\
"pshufw $0x00, %%mm5, %%mm6 \n\t" /* R1 R0 R1 R0 R1 R0 R1 R0 */\
\
"pand %%mm0, %%mm1 \n\t" /* B2 B1 B0 */\
"pand %%mm0, %%mm3 \n\t" /* G2 G1 G0 */\
"pand %%mm7, %%mm6 \n\t" /* R1 R0 */\
\
⌨️ 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 + -