Home | History | Annotate | Download | only in simd 
      1 ;
      2 ; jcclrss2.asm - colorspace conversion (SSE2)
      3 ;
      4 ; x86 SIMD extension for IJG JPEG library
      5 ; Copyright (C) 1999-2006, MIYASAKA Masaru.
      6 ; For conditions of distribution and use, see copyright notice in jsimdext.inc
      7 ;
      8 ; This file should be assembled with NASM (Netwide Assembler),
      9 ; can *not* be assembled with Microsoft's MASM or any compatible
     10 ; assembler (including Borland's Turbo Assembler).
     11 ; NASM is available from http://nasm.sourceforge.net/ or
     12 ; http://sourceforge.net/project/showfiles.php?group_id=6208
     13 ;
     14 ; [TAB8]
     15 
     16 %include "jcolsamp.inc"
     17 
     18 ; --------------------------------------------------------------------------
     19 ;
     20 ; Convert some rows of samples to the output colorspace.
     21 ;
     22 ; GLOBAL(void)
     23 ; jsimd_rgb_ycc_convert_sse2 (JDIMENSION img_width,
     24 ;                             JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
     25 ;                             JDIMENSION output_row, int num_rows);
     26 ;
     27 
     28 %define img_width(b)	(b)+8			; JDIMENSION img_width
     29 %define input_buf(b)	(b)+12		; JSAMPARRAY input_buf
     30 %define output_buf(b)	(b)+16		; JSAMPIMAGE output_buf
     31 %define output_row(b)	(b)+20		; JDIMENSION output_row
     32 %define num_rows(b)	(b)+24		; int num_rows
     33 
     34 %define original_ebp	ebp+0
     35 %define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
     36 %define WK_NUM		8
     37 %define gotptr		wk(0)-SIZEOF_POINTER	; void * gotptr
     38 
     39 	align	16
     40 
     41 	global	EXTN(jsimd_rgb_ycc_convert_sse2) PRIVATE
     42 
     43 EXTN(jsimd_rgb_ycc_convert_sse2):
     44 	push	ebp
     45 	mov	eax,esp				; eax = original ebp
     46 	sub	esp, byte 4
     47 	and	esp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
     48 	mov	[esp],eax
     49 	mov	ebp,esp				; ebp = aligned ebp
     50 	lea	esp, [wk(0)]
     51 	pushpic	eax		; make a room for GOT address
     52 	push	ebx
     53 ;	push	ecx		; need not be preserved
     54 ;	push	edx		; need not be preserved
     55 	push	esi
     56 	push	edi
     57 
     58 	get_GOT	ebx			; get GOT address
     59 	movpic	POINTER [gotptr], ebx	; save GOT address
     60 
     61 	mov	ecx, JDIMENSION [img_width(eax)]
     62 	test	ecx,ecx
     63 	jz	near .return
     64 
     65 	push	ecx
     66 
     67 	mov	esi, JSAMPIMAGE [output_buf(eax)]
     68 	mov	ecx, JDIMENSION [output_row(eax)]
     69 	mov	edi, JSAMPARRAY [esi+0*SIZEOF_JSAMPARRAY]
     70 	mov	ebx, JSAMPARRAY [esi+1*SIZEOF_JSAMPARRAY]
     71 	mov	edx, JSAMPARRAY [esi+2*SIZEOF_JSAMPARRAY]
     72 	lea	edi, [edi+ecx*SIZEOF_JSAMPROW]
     73 	lea	ebx, [ebx+ecx*SIZEOF_JSAMPROW]
     74 	lea	edx, [edx+ecx*SIZEOF_JSAMPROW]
     75 
     76 	pop	ecx
     77 
     78 	mov	esi, JSAMPARRAY [input_buf(eax)]
     79 	mov	eax, INT [num_rows(eax)]
     80 	test	eax,eax
     81 	jle	near .return
     82 	alignx	16,7
     83 .rowloop:
     84 	pushpic	eax
     85 	push	edx
     86 	push	ebx
     87 	push	edi
     88 	push	esi
     89 	push	ecx			; col
     90 
     91 	mov	esi, JSAMPROW [esi]	; inptr
     92 	mov	edi, JSAMPROW [edi]	; outptr0
     93 	mov	ebx, JSAMPROW [ebx]	; outptr1
     94 	mov	edx, JSAMPROW [edx]	; outptr2
     95 	movpic	eax, POINTER [gotptr]	; load GOT address (eax)
     96 
     97 	cmp	ecx, byte SIZEOF_XMMWORD
     98 	jae	near .columnloop
     99 	alignx	16,7
    100 
    101 %if RGB_PIXELSIZE == 3 ; ---------------
    102 
    103 .column_ld1:
    104 	push	eax
    105 	push	edx
    106 	lea	ecx,[ecx+ecx*2]		; imul ecx,RGB_PIXELSIZE
    107 	test	cl, SIZEOF_BYTE
    108 	jz	short .column_ld2
    109 	sub	ecx, byte SIZEOF_BYTE
    110 	movzx	eax, BYTE [esi+ecx]
    111 .column_ld2:
    112 	test	cl, SIZEOF_WORD
    113 	jz	short .column_ld4
    114 	sub	ecx, byte SIZEOF_WORD
    115 	movzx	edx, WORD [esi+ecx]
    116 	shl	eax, WORD_BIT
    117 	or	eax,edx
    118 .column_ld4:
    119 	movd	xmmA,eax
    120 	pop	edx
    121 	pop	eax
    122 	test	cl, SIZEOF_DWORD
    123 	jz	short .column_ld8
    124 	sub	ecx, byte SIZEOF_DWORD
    125 	movd	xmmF, XMM_DWORD [esi+ecx]
    126 	pslldq	xmmA, SIZEOF_DWORD
    127 	por	xmmA,xmmF
    128 .column_ld8:
    129 	test	cl, SIZEOF_MMWORD
    130 	jz	short .column_ld16
    131 	sub	ecx, byte SIZEOF_MMWORD
    132 	movq	xmmB, XMM_MMWORD [esi+ecx]
    133 	pslldq	xmmA, SIZEOF_MMWORD
    134 	por	xmmA,xmmB
    135 .column_ld16:
    136 	test	cl, SIZEOF_XMMWORD
    137 	jz	short .column_ld32
    138 	movdqa	xmmF,xmmA
    139 	movdqu	xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
    140 	mov	ecx, SIZEOF_XMMWORD
    141 	jmp	short .rgb_ycc_cnv
    142 .column_ld32:
    143 	test	cl, 2*SIZEOF_XMMWORD
    144 	mov	ecx, SIZEOF_XMMWORD
    145 	jz	short .rgb_ycc_cnv
    146 	movdqa	xmmB,xmmA
    147 	movdqu	xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
    148 	movdqu	xmmF, XMMWORD [esi+1*SIZEOF_XMMWORD]
    149 	jmp	short .rgb_ycc_cnv
    150 	alignx	16,7
    151 
    152 .columnloop:
    153 	movdqu	xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
    154 	movdqu	xmmF, XMMWORD [esi+1*SIZEOF_XMMWORD]
    155 	movdqu	xmmB, XMMWORD [esi+2*SIZEOF_XMMWORD]
    156 
    157 .rgb_ycc_cnv:
    158 	; xmmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05)
    159 	; xmmF=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
    160 	; xmmB=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F)
    161 
    162 	movdqa    xmmG,xmmA
    163 	pslldq    xmmA,8	; xmmA=(-- -- -- -- -- -- -- -- 00 10 20 01 11 21 02 12)
    164 	psrldq    xmmG,8	; xmmG=(22 03 13 23 04 14 24 05 -- -- -- -- -- -- -- --)
    165 
    166 	punpckhbw xmmA,xmmF	; xmmA=(00 08 10 18 20 28 01 09 11 19 21 29 02 0A 12 1A)
    167 	pslldq    xmmF,8	; xmmF=(-- -- -- -- -- -- -- -- 15 25 06 16 26 07 17 27)
    168 
    169 	punpcklbw xmmG,xmmB	; xmmG=(22 2A 03 0B 13 1B 23 2B 04 0C 14 1C 24 2C 05 0D)
    170 	punpckhbw xmmF,xmmB	; xmmF=(15 1D 25 2D 06 0E 16 1E 26 2E 07 0F 17 1F 27 2F)
    171 
    172 	movdqa    xmmD,xmmA
    173 	pslldq    xmmA,8	; xmmA=(-- -- -- -- -- -- -- -- 00 08 10 18 20 28 01 09)
    174 	psrldq    xmmD,8	; xmmD=(11 19 21 29 02 0A 12 1A -- -- -- -- -- -- -- --)
    175 
    176 	punpckhbw xmmA,xmmG	; xmmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 01 05 09 0D)
    177 	pslldq    xmmG,8	; xmmG=(-- -- -- -- -- -- -- -- 22 2A 03 0B 13 1B 23 2B)
    178 
    179 	punpcklbw xmmD,xmmF	; xmmD=(11 15 19 1D 21 25 29 2D 02 06 0A 0E 12 16 1A 1E)
    180 	punpckhbw xmmG,xmmF	; xmmG=(22 26 2A 2E 03 07 0B 0F 13 17 1B 1F 23 27 2B 2F)
    181 
    182 	movdqa    xmmE,xmmA
    183 	pslldq    xmmA,8	; xmmA=(-- -- -- -- -- -- -- -- 00 04 08 0C 10 14 18 1C)
    184 	psrldq    xmmE,8	; xmmE=(20 24 28 2C 01 05 09 0D -- -- -- -- -- -- -- --)
    185 
    186 	punpckhbw xmmA,xmmD	; xmmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E)
    187 	pslldq    xmmD,8	; xmmD=(-- -- -- -- -- -- -- -- 11 15 19 1D 21 25 29 2D)
    188 
    189 	punpcklbw xmmE,xmmG	; xmmE=(20 22 24 26 28 2A 2C 2E 01 03 05 07 09 0B 0D 0F)
    190 	punpckhbw xmmD,xmmG	; xmmD=(11 13 15 17 19 1B 1D 1F 21 23 25 27 29 2B 2D 2F)
    191 
    192 	pxor      xmmH,xmmH
    193 
    194 	movdqa    xmmC,xmmA
    195 	punpcklbw xmmA,xmmH	; xmmA=(00 02 04 06 08 0A 0C 0E)
    196 	punpckhbw xmmC,xmmH	; xmmC=(10 12 14 16 18 1A 1C 1E)
    197 
    198 	movdqa    xmmB,xmmE
    199 	punpcklbw xmmE,xmmH	; xmmE=(20 22 24 26 28 2A 2C 2E)
    200 	punpckhbw xmmB,xmmH	; xmmB=(01 03 05 07 09 0B 0D 0F)
    201 
    202 	movdqa    xmmF,xmmD
    203 	punpcklbw xmmD,xmmH	; xmmD=(11 13 15 17 19 1B 1D 1F)
    204 	punpckhbw xmmF,xmmH	; xmmF=(21 23 25 27 29 2B 2D 2F)
    205 
    206 %else ; RGB_PIXELSIZE == 4 ; -----------
    207 
    208 .column_ld1:
    209 	test	cl, SIZEOF_XMMWORD/16
    210 	jz	short .column_ld2
    211 	sub	ecx, byte SIZEOF_XMMWORD/16
    212 	movd	xmmA, XMM_DWORD [esi+ecx*RGB_PIXELSIZE]
    213 .column_ld2:
    214 	test	cl, SIZEOF_XMMWORD/8
    215 	jz	short .column_ld4
    216 	sub	ecx, byte SIZEOF_XMMWORD/8
    217 	movq	xmmE, XMM_MMWORD [esi+ecx*RGB_PIXELSIZE]
    218 	pslldq	xmmA, SIZEOF_MMWORD
    219 	por	xmmA,xmmE
    220 .column_ld4:
    221 	test	cl, SIZEOF_XMMWORD/4
    222 	jz	short .column_ld8
    223 	sub	ecx, byte SIZEOF_XMMWORD/4
    224 	movdqa	xmmE,xmmA
    225 	movdqu	xmmA, XMMWORD [esi+ecx*RGB_PIXELSIZE]
    226 .column_ld8:
    227 	test	cl, SIZEOF_XMMWORD/2
    228 	mov	ecx, SIZEOF_XMMWORD
    229 	jz	short .rgb_ycc_cnv
    230 	movdqa	xmmF,xmmA
    231 	movdqa	xmmH,xmmE
    232 	movdqu	xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
    233 	movdqu	xmmE, XMMWORD [esi+1*SIZEOF_XMMWORD]
    234 	jmp	short .rgb_ycc_cnv
    235 	alignx	16,7
    236 
    237 .columnloop:
    238 	movdqu	xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
    239 	movdqu	xmmE, XMMWORD [esi+1*SIZEOF_XMMWORD]
    240 	movdqu	xmmF, XMMWORD [esi+2*SIZEOF_XMMWORD]
    241 	movdqu	xmmH, XMMWORD [esi+3*SIZEOF_XMMWORD]
    242 
    243 .rgb_ycc_cnv:
    244 	; xmmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33)
    245 	; xmmE=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
    246 	; xmmF=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B)
    247 	; xmmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
    248 
    249 	movdqa    xmmD,xmmA
    250 	punpcklbw xmmA,xmmE	; xmmA=(00 04 10 14 20 24 30 34 01 05 11 15 21 25 31 35)
    251 	punpckhbw xmmD,xmmE	; xmmD=(02 06 12 16 22 26 32 36 03 07 13 17 23 27 33 37)
    252 
    253 	movdqa    xmmC,xmmF
    254 	punpcklbw xmmF,xmmH	; xmmF=(08 0C 18 1C 28 2C 38 3C 09 0D 19 1D 29 2D 39 3D)
    255 	punpckhbw xmmC,xmmH	; xmmC=(0A 0E 1A 1E 2A 2E 3A 3E 0B 0F 1B 1F 2B 2F 3B 3F)
    256 
    257 	movdqa    xmmB,xmmA
    258 	punpcklwd xmmA,xmmF	; xmmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 30 34 38 3C)
    259 	punpckhwd xmmB,xmmF	; xmmB=(01 05 09 0D 11 15 19 1D 21 25 29 2D 31 35 39 3D)
    260 
    261 	movdqa    xmmG,xmmD
    262 	punpcklwd xmmD,xmmC	; xmmD=(02 06 0A 0E 12 16 1A 1E 22 26 2A 2E 32 36 3A 3E)
    263 	punpckhwd xmmG,xmmC	; xmmG=(03 07 0B 0F 13 17 1B 1F 23 27 2B 2F 33 37 3B 3F)
    264 
    265 	movdqa    xmmE,xmmA
    266 	punpcklbw xmmA,xmmD	; xmmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E)
    267 	punpckhbw xmmE,xmmD	; xmmE=(20 22 24 26 28 2A 2C 2E 30 32 34 36 38 3A 3C 3E)
    268 
    269 	movdqa    xmmH,xmmB
    270 	punpcklbw xmmB,xmmG	; xmmB=(01 03 05 07 09 0B 0D 0F 11 13 15 17 19 1B 1D 1F)
    271 	punpckhbw xmmH,xmmG	; xmmH=(21 23 25 27 29 2B 2D 2F 31 33 35 37 39 3B 3D 3F)
    272 
    273 	pxor      xmmF,xmmF
    274 
    275 	movdqa    xmmC,xmmA
    276 	punpcklbw xmmA,xmmF	; xmmA=(00 02 04 06 08 0A 0C 0E)
    277 	punpckhbw xmmC,xmmF	; xmmC=(10 12 14 16 18 1A 1C 1E)
    278 
    279 	movdqa    xmmD,xmmB
    280 	punpcklbw xmmB,xmmF	; xmmB=(01 03 05 07 09 0B 0D 0F)
    281 	punpckhbw xmmD,xmmF	; xmmD=(11 13 15 17 19 1B 1D 1F)
    282 
    283 	movdqa    xmmG,xmmE
    284 	punpcklbw xmmE,xmmF	; xmmE=(20 22 24 26 28 2A 2C 2E)
    285 	punpckhbw xmmG,xmmF	; xmmG=(30 32 34 36 38 3A 3C 3E)
    286 
    287 	punpcklbw xmmF,xmmH
    288 	punpckhbw xmmH,xmmH
    289 	psrlw     xmmF,BYTE_BIT	; xmmF=(21 23 25 27 29 2B 2D 2F)
    290 	psrlw     xmmH,BYTE_BIT	; xmmH=(31 33 35 37 39 3B 3D 3F)
    291 
    292 %endif ; RGB_PIXELSIZE ; ---------------
    293 
    294 	; xmm0=R(02468ACE)=RE, xmm2=G(02468ACE)=GE, xmm4=B(02468ACE)=BE
    295 	; xmm1=R(13579BDF)=RO, xmm3=G(13579BDF)=GO, xmm5=B(13579BDF)=BO
    296 
    297 	; (Original)
    298 	; Y  =  0.29900 * R + 0.58700 * G + 0.11400 * B
    299 	; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE
    300 	; Cr =  0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE
    301 	;
    302 	; (This implementation)
    303 	; Y  =  0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G
    304 	; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE
    305 	; Cr =  0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE
    306 
    307 	movdqa    XMMWORD [wk(0)], xmm0	; wk(0)=RE
    308 	movdqa    XMMWORD [wk(1)], xmm1	; wk(1)=RO
    309 	movdqa    XMMWORD [wk(2)], xmm4	; wk(2)=BE
    310 	movdqa    XMMWORD [wk(3)], xmm5	; wk(3)=BO
    311 
    312 	movdqa    xmm6,xmm1
    313 	punpcklwd xmm1,xmm3
    314 	punpckhwd xmm6,xmm3
    315 	movdqa    xmm7,xmm1
    316 	movdqa    xmm4,xmm6
    317 	pmaddwd   xmm1,[GOTOFF(eax,PW_F0299_F0337)] ; xmm1=ROL*FIX(0.299)+GOL*FIX(0.337)
    318 	pmaddwd   xmm6,[GOTOFF(eax,PW_F0299_F0337)] ; xmm6=ROH*FIX(0.299)+GOH*FIX(0.337)
    319 	pmaddwd   xmm7,[GOTOFF(eax,PW_MF016_MF033)] ; xmm7=ROL*-FIX(0.168)+GOL*-FIX(0.331)
    320 	pmaddwd   xmm4,[GOTOFF(eax,PW_MF016_MF033)] ; xmm4=ROH*-FIX(0.168)+GOH*-FIX(0.331)
    321 
    322 	movdqa    XMMWORD [wk(4)], xmm1	; wk(4)=ROL*FIX(0.299)+GOL*FIX(0.337)
    323 	movdqa    XMMWORD [wk(5)], xmm6	; wk(5)=ROH*FIX(0.299)+GOH*FIX(0.337)
    324 
    325 	pxor      xmm1,xmm1
    326 	pxor      xmm6,xmm6
    327 	punpcklwd xmm1,xmm5		; xmm1=BOL
    328 	punpckhwd xmm6,xmm5		; xmm6=BOH
    329 	psrld     xmm1,1		; xmm1=BOL*FIX(0.500)
    330 	psrld     xmm6,1		; xmm6=BOH*FIX(0.500)
    331 
    332 	movdqa    xmm5,[GOTOFF(eax,PD_ONEHALFM1_CJ)] ; xmm5=[PD_ONEHALFM1_CJ]
    333 
    334 	paddd     xmm7,xmm1
    335 	paddd     xmm4,xmm6
    336 	paddd     xmm7,xmm5
    337 	paddd     xmm4,xmm5
    338 	psrld     xmm7,SCALEBITS	; xmm7=CbOL
    339 	psrld     xmm4,SCALEBITS	; xmm4=CbOH
    340 	packssdw  xmm7,xmm4		; xmm7=CbO
    341 
    342 	movdqa    xmm1, XMMWORD [wk(2)]	; xmm1=BE
    343 
    344 	movdqa    xmm6,xmm0
    345 	punpcklwd xmm0,xmm2
    346 	punpckhwd xmm6,xmm2
    347 	movdqa    xmm5,xmm0
    348 	movdqa    xmm4,xmm6
    349 	pmaddwd   xmm0,[GOTOFF(eax,PW_F0299_F0337)] ; xmm0=REL*FIX(0.299)+GEL*FIX(0.337)
    350 	pmaddwd   xmm6,[GOTOFF(eax,PW_F0299_F0337)] ; xmm6=REH*FIX(0.299)+GEH*FIX(0.337)
    351 	pmaddwd   xmm5,[GOTOFF(eax,PW_MF016_MF033)] ; xmm5=REL*-FIX(0.168)+GEL*-FIX(0.331)
    352 	pmaddwd   xmm4,[GOTOFF(eax,PW_MF016_MF033)] ; xmm4=REH*-FIX(0.168)+GEH*-FIX(0.331)
    353 
    354 	movdqa    XMMWORD [wk(6)], xmm0	; wk(6)=REL*FIX(0.299)+GEL*FIX(0.337)
    355 	movdqa    XMMWORD [wk(7)], xmm6	; wk(7)=REH*FIX(0.299)+GEH*FIX(0.337)
    356 
    357 	pxor      xmm0,xmm0
    358 	pxor      xmm6,xmm6
    359 	punpcklwd xmm0,xmm1		; xmm0=BEL
    360 	punpckhwd xmm6,xmm1		; xmm6=BEH
    361 	psrld     xmm0,1		; xmm0=BEL*FIX(0.500)
    362 	psrld     xmm6,1		; xmm6=BEH*FIX(0.500)
    363 
    364 	movdqa    xmm1,[GOTOFF(eax,PD_ONEHALFM1_CJ)] ; xmm1=[PD_ONEHALFM1_CJ]
    365 
    366 	paddd     xmm5,xmm0
    367 	paddd     xmm4,xmm6
    368 	paddd     xmm5,xmm1
    369 	paddd     xmm4,xmm1
    370 	psrld     xmm5,SCALEBITS	; xmm5=CbEL
    371 	psrld     xmm4,SCALEBITS	; xmm4=CbEH
    372 	packssdw  xmm5,xmm4		; xmm5=CbE
    373 
    374 	psllw     xmm7,BYTE_BIT
    375 	por       xmm5,xmm7		; xmm5=Cb
    376 	movdqa    XMMWORD [ebx], xmm5	; Save Cb
    377 
    378 	movdqa    xmm0, XMMWORD [wk(3)]	; xmm0=BO
    379 	movdqa    xmm6, XMMWORD [wk(2)]	; xmm6=BE
    380 	movdqa    xmm1, XMMWORD [wk(1)]	; xmm1=RO
    381 
    382 	movdqa    xmm4,xmm0
    383 	punpcklwd xmm0,xmm3
    384 	punpckhwd xmm4,xmm3
    385 	movdqa    xmm7,xmm0
    386 	movdqa    xmm5,xmm4
    387 	pmaddwd   xmm0,[GOTOFF(eax,PW_F0114_F0250)] ; xmm0=BOL*FIX(0.114)+GOL*FIX(0.250)
    388 	pmaddwd   xmm4,[GOTOFF(eax,PW_F0114_F0250)] ; xmm4=BOH*FIX(0.114)+GOH*FIX(0.250)
    389 	pmaddwd   xmm7,[GOTOFF(eax,PW_MF008_MF041)] ; xmm7=BOL*-FIX(0.081)+GOL*-FIX(0.418)
    390 	pmaddwd   xmm5,[GOTOFF(eax,PW_MF008_MF041)] ; xmm5=BOH*-FIX(0.081)+GOH*-FIX(0.418)
    391 
    392 	movdqa    xmm3,[GOTOFF(eax,PD_ONEHALF)]	; xmm3=[PD_ONEHALF]
    393 
    394 	paddd     xmm0, XMMWORD [wk(4)]
    395 	paddd     xmm4, XMMWORD [wk(5)]
    396 	paddd     xmm0,xmm3
    397 	paddd     xmm4,xmm3
    398 	psrld     xmm0,SCALEBITS	; xmm0=YOL
    399 	psrld     xmm4,SCALEBITS	; xmm4=YOH
    400 	packssdw  xmm0,xmm4		; xmm0=YO
    401 
    402 	pxor      xmm3,xmm3
    403 	pxor      xmm4,xmm4
    404 	punpcklwd xmm3,xmm1		; xmm3=ROL
    405 	punpckhwd xmm4,xmm1		; xmm4=ROH
    406 	psrld     xmm3,1		; xmm3=ROL*FIX(0.500)
    407 	psrld     xmm4,1		; xmm4=ROH*FIX(0.500)
    408 
    409 	movdqa    xmm1,[GOTOFF(eax,PD_ONEHALFM1_CJ)] ; xmm1=[PD_ONEHALFM1_CJ]
    410 
    411 	paddd     xmm7,xmm3
    412 	paddd     xmm5,xmm4
    413 	paddd     xmm7,xmm1
    414 	paddd     xmm5,xmm1
    415 	psrld     xmm7,SCALEBITS	; xmm7=CrOL
    416 	psrld     xmm5,SCALEBITS	; xmm5=CrOH
    417 	packssdw  xmm7,xmm5		; xmm7=CrO
    418 
    419 	movdqa    xmm3, XMMWORD [wk(0)]	; xmm3=RE
    420 
    421 	movdqa    xmm4,xmm6
    422 	punpcklwd xmm6,xmm2
    423 	punpckhwd xmm4,xmm2
    424 	movdqa    xmm1,xmm6
    425 	movdqa    xmm5,xmm4
    426 	pmaddwd   xmm6,[GOTOFF(eax,PW_F0114_F0250)] ; xmm6=BEL*FIX(0.114)+GEL*FIX(0.250)
    427 	pmaddwd   xmm4,[GOTOFF(eax,PW_F0114_F0250)] ; xmm4=BEH*FIX(0.114)+GEH*FIX(0.250)
    428 	pmaddwd   xmm1,[GOTOFF(eax,PW_MF008_MF041)] ; xmm1=BEL*-FIX(0.081)+GEL*-FIX(0.418)
    429 	pmaddwd   xmm5,[GOTOFF(eax,PW_MF008_MF041)] ; xmm5=BEH*-FIX(0.081)+GEH*-FIX(0.418)
    430 
    431 	movdqa    xmm2,[GOTOFF(eax,PD_ONEHALF)]	; xmm2=[PD_ONEHALF]
    432 
    433 	paddd     xmm6, XMMWORD [wk(6)]
    434 	paddd     xmm4, XMMWORD [wk(7)]
    435 	paddd     xmm6,xmm2
    436 	paddd     xmm4,xmm2
    437 	psrld     xmm6,SCALEBITS	; xmm6=YEL
    438 	psrld     xmm4,SCALEBITS	; xmm4=YEH
    439 	packssdw  xmm6,xmm4		; xmm6=YE
    440 
    441 	psllw     xmm0,BYTE_BIT
    442 	por       xmm6,xmm0		; xmm6=Y
    443 	movdqa    XMMWORD [edi], xmm6	; Save Y
    444 
    445 	pxor      xmm2,xmm2
    446 	pxor      xmm4,xmm4
    447 	punpcklwd xmm2,xmm3		; xmm2=REL
    448 	punpckhwd xmm4,xmm3		; xmm4=REH
    449 	psrld     xmm2,1		; xmm2=REL*FIX(0.500)
    450 	psrld     xmm4,1		; xmm4=REH*FIX(0.500)
    451 
    452 	movdqa    xmm0,[GOTOFF(eax,PD_ONEHALFM1_CJ)] ; xmm0=[PD_ONEHALFM1_CJ]
    453 
    454 	paddd     xmm1,xmm2
    455 	paddd     xmm5,xmm4
    456 	paddd     xmm1,xmm0
    457 	paddd     xmm5,xmm0
    458 	psrld     xmm1,SCALEBITS	; xmm1=CrEL
    459 	psrld     xmm5,SCALEBITS	; xmm5=CrEH
    460 	packssdw  xmm1,xmm5		; xmm1=CrE
    461 
    462 	psllw     xmm7,BYTE_BIT
    463 	por       xmm1,xmm7		; xmm1=Cr
    464 	movdqa    XMMWORD [edx], xmm1	; Save Cr
    465 
    466 	sub	ecx, byte SIZEOF_XMMWORD
    467 	add	esi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD	; inptr
    468 	add	edi, byte SIZEOF_XMMWORD		; outptr0
    469 	add	ebx, byte SIZEOF_XMMWORD		; outptr1
    470 	add	edx, byte SIZEOF_XMMWORD		; outptr2
    471 	cmp	ecx, byte SIZEOF_XMMWORD
    472 	jae	near .columnloop
    473 	test	ecx,ecx
    474 	jnz	near .column_ld1
    475 
    476 	pop	ecx			; col
    477 	pop	esi
    478 	pop	edi
    479 	pop	ebx
    480 	pop	edx
    481 	poppic	eax
    482 
    483 	add	esi, byte SIZEOF_JSAMPROW	; input_buf
    484 	add	edi, byte SIZEOF_JSAMPROW
    485 	add	ebx, byte SIZEOF_JSAMPROW
    486 	add	edx, byte SIZEOF_JSAMPROW
    487 	dec	eax				; num_rows
    488 	jg	near .rowloop
    489 
    490 .return:
    491 	pop	edi
    492 	pop	esi
    493 ;	pop	edx		; need not be preserved
    494 ;	pop	ecx		; need not be preserved
    495 	pop	ebx
    496 	mov	esp,ebp		; esp <- aligned ebp
    497 	pop	esp		; esp <- original ebp
    498 	pop	ebp
    499 	ret
    500 
    501 ; For some reason, the OS X linker does not honor the request to align the
    502 ; segment unless we do this.
    503 	align	16
    504