Home | History | Annotate | Download | only in simd
      1 ;
      2 ; jcgryss2.asm - grayscale colorspace conversion (SSE2)
      3 ;
      4 ; x86 SIMD extension for IJG JPEG library
      5 ; Copyright (C) 1999-2006, MIYASAKA Masaru.
      6 ; Copyright (C) 2011, D. R. Commander.
      7 ; For conditions of distribution and use, see copyright notice in jsimdext.inc
      8 ;
      9 ; This file should be assembled with NASM (Netwide Assembler),
     10 ; can *not* be assembled with Microsoft's MASM or any compatible
     11 ; assembler (including Borland's Turbo Assembler).
     12 ; NASM is available from http://nasm.sourceforge.net/ or
     13 ; http://sourceforge.net/project/showfiles.php?group_id=6208
     14 ;
     15 ; [TAB8]
     16 
     17 %include "jcolsamp.inc"
     18 
     19 ; --------------------------------------------------------------------------
     20 ;
     21 ; Convert some rows of samples to the output colorspace.
     22 ;
     23 ; GLOBAL(void)
     24 ; jsimd_rgb_gray_convert_sse2 (JDIMENSION img_width,
     25 ;                              JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
     26 ;                              JDIMENSION output_row, int num_rows);
     27 ;
     28 
     29 %define img_width(b)	(b)+8			; JDIMENSION img_width
     30 %define input_buf(b)	(b)+12		; JSAMPARRAY input_buf
     31 %define output_buf(b)	(b)+16		; JSAMPIMAGE output_buf
     32 %define output_row(b)	(b)+20		; JDIMENSION output_row
     33 %define num_rows(b)	(b)+24		; int num_rows
     34 
     35 %define original_ebp	ebp+0
     36 %define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
     37 %define WK_NUM		2
     38 %define gotptr		wk(0)-SIZEOF_POINTER	; void * gotptr
     39 
     40 	align	16
     41 
     42 	global	EXTN(jsimd_rgb_gray_convert_sse2) PRIVATE
     43 
     44 EXTN(jsimd_rgb_gray_convert_sse2):
     45 	push	ebp
     46 	mov	eax,esp				; eax = original ebp
     47 	sub	esp, byte 4
     48 	and	esp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
     49 	mov	[esp],eax
     50 	mov	ebp,esp				; ebp = aligned ebp
     51 	lea	esp, [wk(0)]
     52 	pushpic	eax		; make a room for GOT address
     53 	push	ebx
     54 ;	push	ecx		; need not be preserved
     55 ;	push	edx		; need not be preserved
     56 	push	esi
     57 	push	edi
     58 
     59 	get_GOT	ebx			; get GOT address
     60 	movpic	POINTER [gotptr], ebx	; save GOT address
     61 
     62 	mov	ecx, JDIMENSION [img_width(eax)]
     63 	test	ecx,ecx
     64 	jz	near .return
     65 
     66 	push	ecx
     67 
     68 	mov	esi, JSAMPIMAGE [output_buf(eax)]
     69 	mov	ecx, JDIMENSION [output_row(eax)]
     70 	mov	edi, JSAMPARRAY [esi+0*SIZEOF_JSAMPARRAY]
     71 	lea	edi, [edi+ecx*SIZEOF_JSAMPROW]
     72 
     73 	pop	ecx
     74 
     75 	mov	esi, JSAMPARRAY [input_buf(eax)]
     76 	mov	eax, INT [num_rows(eax)]
     77 	test	eax,eax
     78 	jle	near .return
     79 	alignx	16,7
     80 .rowloop:
     81 	pushpic	eax
     82 	push	edi
     83 	push	esi
     84 	push	ecx			; col
     85 
     86 	mov	esi, JSAMPROW [esi]	; inptr
     87 	mov	edi, JSAMPROW [edi]	; outptr0
     88 	movpic	eax, POINTER [gotptr]	; load GOT address (eax)
     89 
     90 	cmp	ecx, byte SIZEOF_XMMWORD
     91 	jae	near .columnloop
     92 	alignx	16,7
     93 
     94 %if RGB_PIXELSIZE == 3 ; ---------------
     95 
     96 .column_ld1:
     97 	push	eax
     98 	push	edx
     99 	lea	ecx,[ecx+ecx*2]		; imul ecx,RGB_PIXELSIZE
    100 	test	cl, SIZEOF_BYTE
    101 	jz	short .column_ld2
    102 	sub	ecx, byte SIZEOF_BYTE
    103 	movzx	eax, BYTE [esi+ecx]
    104 .column_ld2:
    105 	test	cl, SIZEOF_WORD
    106 	jz	short .column_ld4
    107 	sub	ecx, byte SIZEOF_WORD
    108 	movzx	edx, WORD [esi+ecx]
    109 	shl	eax, WORD_BIT
    110 	or	eax,edx
    111 .column_ld4:
    112 	movd	xmmA,eax
    113 	pop	edx
    114 	pop	eax
    115 	test	cl, SIZEOF_DWORD
    116 	jz	short .column_ld8
    117 	sub	ecx, byte SIZEOF_DWORD
    118 	movd	xmmF, XMM_DWORD [esi+ecx]
    119 	pslldq	xmmA, SIZEOF_DWORD
    120 	por	xmmA,xmmF
    121 .column_ld8:
    122 	test	cl, SIZEOF_MMWORD
    123 	jz	short .column_ld16
    124 	sub	ecx, byte SIZEOF_MMWORD
    125 	movq	xmmB, XMM_MMWORD [esi+ecx]
    126 	pslldq	xmmA, SIZEOF_MMWORD
    127 	por	xmmA,xmmB
    128 .column_ld16:
    129 	test	cl, SIZEOF_XMMWORD
    130 	jz	short .column_ld32
    131 	movdqa	xmmF,xmmA
    132 	movdqu	xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
    133 	mov	ecx, SIZEOF_XMMWORD
    134 	jmp	short .rgb_gray_cnv
    135 .column_ld32:
    136 	test	cl, 2*SIZEOF_XMMWORD
    137 	mov	ecx, SIZEOF_XMMWORD
    138 	jz	short .rgb_gray_cnv
    139 	movdqa	xmmB,xmmA
    140 	movdqu	xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
    141 	movdqu	xmmF, XMMWORD [esi+1*SIZEOF_XMMWORD]
    142 	jmp	short .rgb_gray_cnv
    143 	alignx	16,7
    144 
    145 .columnloop:
    146 	movdqu	xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
    147 	movdqu	xmmF, XMMWORD [esi+1*SIZEOF_XMMWORD]
    148 	movdqu	xmmB, XMMWORD [esi+2*SIZEOF_XMMWORD]
    149 
    150 .rgb_gray_cnv:
    151 	; xmmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05)
    152 	; xmmF=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
    153 	; xmmB=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F)
    154 
    155 	movdqa    xmmG,xmmA
    156 	pslldq    xmmA,8	; xmmA=(-- -- -- -- -- -- -- -- 00 10 20 01 11 21 02 12)
    157 	psrldq    xmmG,8	; xmmG=(22 03 13 23 04 14 24 05 -- -- -- -- -- -- -- --)
    158 
    159 	punpckhbw xmmA,xmmF	; xmmA=(00 08 10 18 20 28 01 09 11 19 21 29 02 0A 12 1A)
    160 	pslldq    xmmF,8	; xmmF=(-- -- -- -- -- -- -- -- 15 25 06 16 26 07 17 27)
    161 
    162 	punpcklbw xmmG,xmmB	; xmmG=(22 2A 03 0B 13 1B 23 2B 04 0C 14 1C 24 2C 05 0D)
    163 	punpckhbw xmmF,xmmB	; xmmF=(15 1D 25 2D 06 0E 16 1E 26 2E 07 0F 17 1F 27 2F)
    164 
    165 	movdqa    xmmD,xmmA
    166 	pslldq    xmmA,8	; xmmA=(-- -- -- -- -- -- -- -- 00 08 10 18 20 28 01 09)
    167 	psrldq    xmmD,8	; xmmD=(11 19 21 29 02 0A 12 1A -- -- -- -- -- -- -- --)
    168 
    169 	punpckhbw xmmA,xmmG	; xmmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 01 05 09 0D)
    170 	pslldq    xmmG,8	; xmmG=(-- -- -- -- -- -- -- -- 22 2A 03 0B 13 1B 23 2B)
    171 
    172 	punpcklbw xmmD,xmmF	; xmmD=(11 15 19 1D 21 25 29 2D 02 06 0A 0E 12 16 1A 1E)
    173 	punpckhbw xmmG,xmmF	; xmmG=(22 26 2A 2E 03 07 0B 0F 13 17 1B 1F 23 27 2B 2F)
    174 
    175 	movdqa    xmmE,xmmA
    176 	pslldq    xmmA,8	; xmmA=(-- -- -- -- -- -- -- -- 00 04 08 0C 10 14 18 1C)
    177 	psrldq    xmmE,8	; xmmE=(20 24 28 2C 01 05 09 0D -- -- -- -- -- -- -- --)
    178 
    179 	punpckhbw xmmA,xmmD	; xmmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E)
    180 	pslldq    xmmD,8	; xmmD=(-- -- -- -- -- -- -- -- 11 15 19 1D 21 25 29 2D)
    181 
    182 	punpcklbw xmmE,xmmG	; xmmE=(20 22 24 26 28 2A 2C 2E 01 03 05 07 09 0B 0D 0F)
    183 	punpckhbw xmmD,xmmG	; xmmD=(11 13 15 17 19 1B 1D 1F 21 23 25 27 29 2B 2D 2F)
    184 
    185 	pxor      xmmH,xmmH
    186 
    187 	movdqa    xmmC,xmmA
    188 	punpcklbw xmmA,xmmH	; xmmA=(00 02 04 06 08 0A 0C 0E)
    189 	punpckhbw xmmC,xmmH	; xmmC=(10 12 14 16 18 1A 1C 1E)
    190 
    191 	movdqa    xmmB,xmmE
    192 	punpcklbw xmmE,xmmH	; xmmE=(20 22 24 26 28 2A 2C 2E)
    193 	punpckhbw xmmB,xmmH	; xmmB=(01 03 05 07 09 0B 0D 0F)
    194 
    195 	movdqa    xmmF,xmmD
    196 	punpcklbw xmmD,xmmH	; xmmD=(11 13 15 17 19 1B 1D 1F)
    197 	punpckhbw xmmF,xmmH	; xmmF=(21 23 25 27 29 2B 2D 2F)
    198 
    199 %else ; RGB_PIXELSIZE == 4 ; -----------
    200 
    201 .column_ld1:
    202 	test	cl, SIZEOF_XMMWORD/16
    203 	jz	short .column_ld2
    204 	sub	ecx, byte SIZEOF_XMMWORD/16
    205 	movd	xmmA, XMM_DWORD [esi+ecx*RGB_PIXELSIZE]
    206 .column_ld2:
    207 	test	cl, SIZEOF_XMMWORD/8
    208 	jz	short .column_ld4
    209 	sub	ecx, byte SIZEOF_XMMWORD/8
    210 	movq	xmmE, XMM_MMWORD [esi+ecx*RGB_PIXELSIZE]
    211 	pslldq	xmmA, SIZEOF_MMWORD
    212 	por	xmmA,xmmE
    213 .column_ld4:
    214 	test	cl, SIZEOF_XMMWORD/4
    215 	jz	short .column_ld8
    216 	sub	ecx, byte SIZEOF_XMMWORD/4
    217 	movdqa	xmmE,xmmA
    218 	movdqu	xmmA, XMMWORD [esi+ecx*RGB_PIXELSIZE]
    219 .column_ld8:
    220 	test	cl, SIZEOF_XMMWORD/2
    221 	mov	ecx, SIZEOF_XMMWORD
    222 	jz	short .rgb_gray_cnv
    223 	movdqa	xmmF,xmmA
    224 	movdqa	xmmH,xmmE
    225 	movdqu	xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
    226 	movdqu	xmmE, XMMWORD [esi+1*SIZEOF_XMMWORD]
    227 	jmp	short .rgb_gray_cnv
    228 	alignx	16,7
    229 
    230 .columnloop:
    231 	movdqu	xmmA, XMMWORD [esi+0*SIZEOF_XMMWORD]
    232 	movdqu	xmmE, XMMWORD [esi+1*SIZEOF_XMMWORD]
    233 	movdqu	xmmF, XMMWORD [esi+2*SIZEOF_XMMWORD]
    234 	movdqu	xmmH, XMMWORD [esi+3*SIZEOF_XMMWORD]
    235 
    236 .rgb_gray_cnv:
    237 	; xmmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33)
    238 	; xmmE=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
    239 	; xmmF=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B)
    240 	; xmmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
    241 
    242 	movdqa    xmmD,xmmA
    243 	punpcklbw xmmA,xmmE	; xmmA=(00 04 10 14 20 24 30 34 01 05 11 15 21 25 31 35)
    244 	punpckhbw xmmD,xmmE	; xmmD=(02 06 12 16 22 26 32 36 03 07 13 17 23 27 33 37)
    245 
    246 	movdqa    xmmC,xmmF
    247 	punpcklbw xmmF,xmmH	; xmmF=(08 0C 18 1C 28 2C 38 3C 09 0D 19 1D 29 2D 39 3D)
    248 	punpckhbw xmmC,xmmH	; xmmC=(0A 0E 1A 1E 2A 2E 3A 3E 0B 0F 1B 1F 2B 2F 3B 3F)
    249 
    250 	movdqa    xmmB,xmmA
    251 	punpcklwd xmmA,xmmF	; xmmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 30 34 38 3C)
    252 	punpckhwd xmmB,xmmF	; xmmB=(01 05 09 0D 11 15 19 1D 21 25 29 2D 31 35 39 3D)
    253 
    254 	movdqa    xmmG,xmmD
    255 	punpcklwd xmmD,xmmC	; xmmD=(02 06 0A 0E 12 16 1A 1E 22 26 2A 2E 32 36 3A 3E)
    256 	punpckhwd xmmG,xmmC	; xmmG=(03 07 0B 0F 13 17 1B 1F 23 27 2B 2F 33 37 3B 3F)
    257 
    258 	movdqa    xmmE,xmmA
    259 	punpcklbw xmmA,xmmD	; xmmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E)
    260 	punpckhbw xmmE,xmmD	; xmmE=(20 22 24 26 28 2A 2C 2E 30 32 34 36 38 3A 3C 3E)
    261 
    262 	movdqa    xmmH,xmmB
    263 	punpcklbw xmmB,xmmG	; xmmB=(01 03 05 07 09 0B 0D 0F 11 13 15 17 19 1B 1D 1F)
    264 	punpckhbw xmmH,xmmG	; xmmH=(21 23 25 27 29 2B 2D 2F 31 33 35 37 39 3B 3D 3F)
    265 
    266 	pxor      xmmF,xmmF
    267 
    268 	movdqa    xmmC,xmmA
    269 	punpcklbw xmmA,xmmF	; xmmA=(00 02 04 06 08 0A 0C 0E)
    270 	punpckhbw xmmC,xmmF	; xmmC=(10 12 14 16 18 1A 1C 1E)
    271 
    272 	movdqa    xmmD,xmmB
    273 	punpcklbw xmmB,xmmF	; xmmB=(01 03 05 07 09 0B 0D 0F)
    274 	punpckhbw xmmD,xmmF	; xmmD=(11 13 15 17 19 1B 1D 1F)
    275 
    276 	movdqa    xmmG,xmmE
    277 	punpcklbw xmmE,xmmF	; xmmE=(20 22 24 26 28 2A 2C 2E)
    278 	punpckhbw xmmG,xmmF	; xmmG=(30 32 34 36 38 3A 3C 3E)
    279 
    280 	punpcklbw xmmF,xmmH
    281 	punpckhbw xmmH,xmmH
    282 	psrlw     xmmF,BYTE_BIT	; xmmF=(21 23 25 27 29 2B 2D 2F)
    283 	psrlw     xmmH,BYTE_BIT	; xmmH=(31 33 35 37 39 3B 3D 3F)
    284 
    285 %endif ; RGB_PIXELSIZE ; ---------------
    286 
    287 	; xmm0=R(02468ACE)=RE, xmm2=G(02468ACE)=GE, xmm4=B(02468ACE)=BE
    288 	; xmm1=R(13579BDF)=RO, xmm3=G(13579BDF)=GO, xmm5=B(13579BDF)=BO
    289 
    290 	; (Original)
    291 	; Y  =  0.29900 * R + 0.58700 * G + 0.11400 * B
    292 	;
    293 	; (This implementation)
    294 	; Y  =  0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G
    295 
    296 	movdqa    xmm6,xmm1
    297 	punpcklwd xmm1,xmm3
    298 	punpckhwd xmm6,xmm3
    299 	pmaddwd   xmm1,[GOTOFF(eax,PW_F0299_F0337)] ; xmm1=ROL*FIX(0.299)+GOL*FIX(0.337)
    300 	pmaddwd   xmm6,[GOTOFF(eax,PW_F0299_F0337)] ; xmm6=ROH*FIX(0.299)+GOH*FIX(0.337)
    301 
    302 	movdqa    xmm7, xmm6	; xmm7=ROH*FIX(0.299)+GOH*FIX(0.337)
    303 
    304 	movdqa    xmm6,xmm0
    305 	punpcklwd xmm0,xmm2
    306 	punpckhwd xmm6,xmm2
    307 	pmaddwd   xmm0,[GOTOFF(eax,PW_F0299_F0337)] ; xmm0=REL*FIX(0.299)+GEL*FIX(0.337)
    308 	pmaddwd   xmm6,[GOTOFF(eax,PW_F0299_F0337)] ; xmm6=REH*FIX(0.299)+GEH*FIX(0.337)
    309 
    310 	movdqa    XMMWORD [wk(0)], xmm0	; wk(0)=REL*FIX(0.299)+GEL*FIX(0.337)
    311 	movdqa    XMMWORD [wk(1)], xmm6	; wk(1)=REH*FIX(0.299)+GEH*FIX(0.337)
    312 
    313 	movdqa    xmm0, xmm5	; xmm0=BO
    314 	movdqa    xmm6, xmm4	; xmm6=BE
    315 
    316 	movdqa    xmm4,xmm0
    317 	punpcklwd xmm0,xmm3
    318 	punpckhwd xmm4,xmm3
    319 	pmaddwd   xmm0,[GOTOFF(eax,PW_F0114_F0250)] ; xmm0=BOL*FIX(0.114)+GOL*FIX(0.250)
    320 	pmaddwd   xmm4,[GOTOFF(eax,PW_F0114_F0250)] ; xmm4=BOH*FIX(0.114)+GOH*FIX(0.250)
    321 
    322 	movdqa    xmm3,[GOTOFF(eax,PD_ONEHALF)]	; xmm3=[PD_ONEHALF]
    323 
    324 	paddd     xmm0, xmm1
    325 	paddd     xmm4, xmm7
    326 	paddd     xmm0,xmm3
    327 	paddd     xmm4,xmm3
    328 	psrld     xmm0,SCALEBITS	; xmm0=YOL
    329 	psrld     xmm4,SCALEBITS	; xmm4=YOH
    330 	packssdw  xmm0,xmm4		; xmm0=YO
    331 
    332 	movdqa    xmm4,xmm6
    333 	punpcklwd xmm6,xmm2
    334 	punpckhwd xmm4,xmm2
    335 	pmaddwd   xmm6,[GOTOFF(eax,PW_F0114_F0250)] ; xmm6=BEL*FIX(0.114)+GEL*FIX(0.250)
    336 	pmaddwd   xmm4,[GOTOFF(eax,PW_F0114_F0250)] ; xmm4=BEH*FIX(0.114)+GEH*FIX(0.250)
    337 
    338 	movdqa    xmm2,[GOTOFF(eax,PD_ONEHALF)]	; xmm2=[PD_ONEHALF]
    339 
    340 	paddd     xmm6, XMMWORD [wk(0)]
    341 	paddd     xmm4, XMMWORD [wk(1)]
    342 	paddd     xmm6,xmm2
    343 	paddd     xmm4,xmm2
    344 	psrld     xmm6,SCALEBITS	; xmm6=YEL
    345 	psrld     xmm4,SCALEBITS	; xmm4=YEH
    346 	packssdw  xmm6,xmm4		; xmm6=YE
    347 
    348 	psllw     xmm0,BYTE_BIT
    349 	por       xmm6,xmm0		; xmm6=Y
    350 	movdqa    XMMWORD [edi], xmm6	; Save Y
    351 
    352 	sub	ecx, byte SIZEOF_XMMWORD
    353 	add	esi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD	; inptr
    354 	add	edi, byte SIZEOF_XMMWORD		; outptr0
    355 	cmp	ecx, byte SIZEOF_XMMWORD
    356 	jae	near .columnloop
    357 	test	ecx,ecx
    358 	jnz	near .column_ld1
    359 
    360 	pop	ecx			; col
    361 	pop	esi
    362 	pop	edi
    363 	poppic	eax
    364 
    365 	add	esi, byte SIZEOF_JSAMPROW	; input_buf
    366 	add	edi, byte SIZEOF_JSAMPROW
    367 	dec	eax				; num_rows
    368 	jg	near .rowloop
    369 
    370 .return:
    371 	pop	edi
    372 	pop	esi
    373 ;	pop	edx		; need not be preserved
    374 ;	pop	ecx		; need not be preserved
    375 	pop	ebx
    376 	mov	esp,ebp		; esp <- aligned ebp
    377 	pop	esp		; esp <- original ebp
    378 	pop	ebp
    379 	ret
    380 
    381 ; For some reason, the OS X linker does not honor the request to align the
    382 ; segment unless we do this.
    383 	align	16
    384