1; 2; jccolext.asm - colorspace conversion (MMX) 3; 4; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB 5; 6; Based on the x86 SIMD extension for IJG JPEG library 7; Copyright (C) 1999-2006, MIYASAKA Masaru. 8; For conditions of distribution and use, see copyright notice in jsimdext.inc 9; 10; This file should be assembled with NASM (Netwide Assembler), 11; can *not* be assembled with Microsoft's MASM or any compatible 12; assembler (including Borland's Turbo Assembler). 13; NASM is available from http://nasm.sourceforge.net/ or 14; http://sourceforge.net/project/showfiles.php?group_id=6208 15; 16; [TAB8] 17 18%include "jcolsamp.inc" 19 20; -------------------------------------------------------------------------- 21; 22; Convert some rows of samples to the output colorspace. 23; 24; GLOBAL(void) 25; jsimd_rgb_ycc_convert_mmx (JDIMENSION img_width, 26; JSAMPARRAY input_buf, JSAMPIMAGE output_buf, 27; JDIMENSION output_row, int num_rows); 28; 29 30%define img_width(b) (b)+8 ; JDIMENSION img_width 31%define input_buf(b) (b)+12 ; JSAMPARRAY input_buf 32%define output_buf(b) (b)+16 ; JSAMPIMAGE output_buf 33%define output_row(b) (b)+20 ; JDIMENSION output_row 34%define num_rows(b) (b)+24 ; int num_rows 35 36%define original_ebp ebp+0 37%define wk(i) ebp-(WK_NUM-(i))*SIZEOF_MMWORD ; mmword wk[WK_NUM] 38%define WK_NUM 8 39%define gotptr wk(0)-SIZEOF_POINTER ; void * gotptr 40 41 align 16 42 global EXTN(jsimd_rgb_ycc_convert_mmx) 43 44EXTN(jsimd_rgb_ycc_convert_mmx): 45 push ebp 46 mov eax,esp ; eax = original ebp 47 sub esp, byte 4 48 and esp, byte (-SIZEOF_MMWORD) ; align to 64 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)] ; num_cols 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 mov ebx, JSAMPARRAY [esi+1*SIZEOF_JSAMPARRAY] 72 mov edx, JSAMPARRAY [esi+2*SIZEOF_JSAMPARRAY] 73 lea edi, [edi+ecx*SIZEOF_JSAMPROW] 74 lea ebx, [ebx+ecx*SIZEOF_JSAMPROW] 75 lea edx, [edx+ecx*SIZEOF_JSAMPROW] 76 77 pop ecx 78 79 mov esi, JSAMPARRAY [input_buf(eax)] 80 mov eax, INT [num_rows(eax)] 81 test eax,eax 82 jle near .return 83 alignx 16,7 84.rowloop: 85 pushpic eax 86 push edx 87 push ebx 88 push edi 89 push esi 90 push ecx ; col 91 92 mov esi, JSAMPROW [esi] ; inptr 93 mov edi, JSAMPROW [edi] ; outptr0 94 mov ebx, JSAMPROW [ebx] ; outptr1 95 mov edx, JSAMPROW [edx] ; outptr2 96 movpic eax, POINTER [gotptr] ; load GOT address (eax) 97 98 cmp ecx, byte SIZEOF_MMWORD 99 jae short .columnloop 100 alignx 16,7 101 102%if RGB_PIXELSIZE == 3 ; --------------- 103 104.column_ld1: 105 push eax 106 push edx 107 lea ecx,[ecx+ecx*2] ; imul ecx,RGB_PIXELSIZE 108 test cl, SIZEOF_BYTE 109 jz short .column_ld2 110 sub ecx, byte SIZEOF_BYTE 111 xor eax,eax 112 mov al, BYTE [esi+ecx] 113.column_ld2: 114 test cl, SIZEOF_WORD 115 jz short .column_ld4 116 sub ecx, byte SIZEOF_WORD 117 xor edx,edx 118 mov dx, WORD [esi+ecx] 119 shl eax, WORD_BIT 120 or eax,edx 121.column_ld4: 122 movd mmA,eax 123 pop edx 124 pop eax 125 test cl, SIZEOF_DWORD 126 jz short .column_ld8 127 sub ecx, byte SIZEOF_DWORD 128 movd mmG, DWORD [esi+ecx] 129 psllq mmA, DWORD_BIT 130 por mmA,mmG 131.column_ld8: 132 test cl, SIZEOF_MMWORD 133 jz short .column_ld16 134 movq mmG,mmA 135 movq mmA, MMWORD [esi+0*SIZEOF_MMWORD] 136 mov ecx, SIZEOF_MMWORD 137 jmp short .rgb_ycc_cnv 138.column_ld16: 139 test cl, 2*SIZEOF_MMWORD 140 mov ecx, SIZEOF_MMWORD 141 jz short .rgb_ycc_cnv 142 movq mmF,mmA 143 movq mmA, MMWORD [esi+0*SIZEOF_MMWORD] 144 movq mmG, MMWORD [esi+1*SIZEOF_MMWORD] 145 jmp short .rgb_ycc_cnv 146 alignx 16,7 147 148.columnloop: 149 movq mmA, MMWORD [esi+0*SIZEOF_MMWORD] 150 movq mmG, MMWORD [esi+1*SIZEOF_MMWORD] 151 movq mmF, MMWORD [esi+2*SIZEOF_MMWORD] 152 153.rgb_ycc_cnv: 154 ; mmA=(00 10 20 01 11 21 02 12) 155 ; mmG=(22 03 13 23 04 14 24 05) 156 ; mmF=(15 25 06 16 26 07 17 27) 157 158 movq mmD,mmA 159 psllq mmA,4*BYTE_BIT ; mmA=(-- -- -- -- 00 10 20 01) 160 psrlq mmD,4*BYTE_BIT ; mmD=(11 21 02 12 -- -- -- --) 161 162 punpckhbw mmA,mmG ; mmA=(00 04 10 14 20 24 01 05) 163 psllq mmG,4*BYTE_BIT ; mmG=(-- -- -- -- 22 03 13 23) 164 165 punpcklbw mmD,mmF ; mmD=(11 15 21 25 02 06 12 16) 166 punpckhbw mmG,mmF ; mmG=(22 26 03 07 13 17 23 27) 167 168 movq mmE,mmA 169 psllq mmA,4*BYTE_BIT ; mmA=(-- -- -- -- 00 04 10 14) 170 psrlq mmE,4*BYTE_BIT ; mmE=(20 24 01 05 -- -- -- --) 171 172 punpckhbw mmA,mmD ; mmA=(00 02 04 06 10 12 14 16) 173 psllq mmD,4*BYTE_BIT ; mmD=(-- -- -- -- 11 15 21 25) 174 175 punpcklbw mmE,mmG ; mmE=(20 22 24 26 01 03 05 07) 176 punpckhbw mmD,mmG ; mmD=(11 13 15 17 21 23 25 27) 177 178 pxor mmH,mmH 179 180 movq mmC,mmA 181 punpcklbw mmA,mmH ; mmA=(00 02 04 06) 182 punpckhbw mmC,mmH ; mmC=(10 12 14 16) 183 184 movq mmB,mmE 185 punpcklbw mmE,mmH ; mmE=(20 22 24 26) 186 punpckhbw mmB,mmH ; mmB=(01 03 05 07) 187 188 movq mmF,mmD 189 punpcklbw mmD,mmH ; mmD=(11 13 15 17) 190 punpckhbw mmF,mmH ; mmF=(21 23 25 27) 191 192%else ; RGB_PIXELSIZE == 4 ; ----------- 193 194.column_ld1: 195 test cl, SIZEOF_MMWORD/8 196 jz short .column_ld2 197 sub ecx, byte SIZEOF_MMWORD/8 198 movd mmA, DWORD [esi+ecx*RGB_PIXELSIZE] 199.column_ld2: 200 test cl, SIZEOF_MMWORD/4 201 jz short .column_ld4 202 sub ecx, byte SIZEOF_MMWORD/4 203 movq mmF,mmA 204 movq mmA, MMWORD [esi+ecx*RGB_PIXELSIZE] 205.column_ld4: 206 test cl, SIZEOF_MMWORD/2 207 mov ecx, SIZEOF_MMWORD 208 jz short .rgb_ycc_cnv 209 movq mmD,mmA 210 movq mmC,mmF 211 movq mmA, MMWORD [esi+0*SIZEOF_MMWORD] 212 movq mmF, MMWORD [esi+1*SIZEOF_MMWORD] 213 jmp short .rgb_ycc_cnv 214 alignx 16,7 215 216.columnloop: 217 movq mmA, MMWORD [esi+0*SIZEOF_MMWORD] 218 movq mmF, MMWORD [esi+1*SIZEOF_MMWORD] 219 movq mmD, MMWORD [esi+2*SIZEOF_MMWORD] 220 movq mmC, MMWORD [esi+3*SIZEOF_MMWORD] 221 222.rgb_ycc_cnv: 223 ; mmA=(00 10 20 30 01 11 21 31) 224 ; mmF=(02 12 22 32 03 13 23 33) 225 ; mmD=(04 14 24 34 05 15 25 35) 226 ; mmC=(06 16 26 36 07 17 27 37) 227 228 movq mmB,mmA 229 punpcklbw mmA,mmF ; mmA=(00 02 10 12 20 22 30 32) 230 punpckhbw mmB,mmF ; mmB=(01 03 11 13 21 23 31 33) 231 232 movq mmG,mmD 233 punpcklbw mmD,mmC ; mmD=(04 06 14 16 24 26 34 36) 234 punpckhbw mmG,mmC ; mmG=(05 07 15 17 25 27 35 37) 235 236 movq mmE,mmA 237 punpcklwd mmA,mmD ; mmA=(00 02 04 06 10 12 14 16) 238 punpckhwd mmE,mmD ; mmE=(20 22 24 26 30 32 34 36) 239 240 movq mmH,mmB 241 punpcklwd mmB,mmG ; mmB=(01 03 05 07 11 13 15 17) 242 punpckhwd mmH,mmG ; mmH=(21 23 25 27 31 33 35 37) 243 244 pxor mmF,mmF 245 246 movq mmC,mmA 247 punpcklbw mmA,mmF ; mmA=(00 02 04 06) 248 punpckhbw mmC,mmF ; mmC=(10 12 14 16) 249 250 movq mmD,mmB 251 punpcklbw mmB,mmF ; mmB=(01 03 05 07) 252 punpckhbw mmD,mmF ; mmD=(11 13 15 17) 253 254 movq mmG,mmE 255 punpcklbw mmE,mmF ; mmE=(20 22 24 26) 256 punpckhbw mmG,mmF ; mmG=(30 32 34 36) 257 258 punpcklbw mmF,mmH 259 punpckhbw mmH,mmH 260 psrlw mmF,BYTE_BIT ; mmF=(21 23 25 27) 261 psrlw mmH,BYTE_BIT ; mmH=(31 33 35 37) 262 263%endif ; RGB_PIXELSIZE ; --------------- 264 265 ; mm0=(R0 R2 R4 R6)=RE, mm2=(G0 G2 G4 G6)=GE, mm4=(B0 B2 B4 B6)=BE 266 ; mm1=(R1 R3 R5 R7)=RO, mm3=(G1 G3 G5 G7)=GO, mm5=(B1 B3 B5 B7)=BO 267 268 ; (Original) 269 ; Y = 0.29900 * R + 0.58700 * G + 0.11400 * B 270 ; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE 271 ; Cr = 0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE 272 ; 273 ; (This implementation) 274 ; Y = 0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G 275 ; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE 276 ; Cr = 0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE 277 278 movq MMWORD [wk(0)], mm0 ; wk(0)=RE 279 movq MMWORD [wk(1)], mm1 ; wk(1)=RO 280 movq MMWORD [wk(2)], mm4 ; wk(2)=BE 281 movq MMWORD [wk(3)], mm5 ; wk(3)=BO 282 283 movq mm6,mm1 284 punpcklwd mm1,mm3 285 punpckhwd mm6,mm3 286 movq mm7,mm1 287 movq mm4,mm6 288 pmaddwd mm1,[GOTOFF(eax,PW_F0299_F0337)] ; mm1=ROL*FIX(0.299)+GOL*FIX(0.337) 289 pmaddwd mm6,[GOTOFF(eax,PW_F0299_F0337)] ; mm6=ROH*FIX(0.299)+GOH*FIX(0.337) 290 pmaddwd mm7,[GOTOFF(eax,PW_MF016_MF033)] ; mm7=ROL*-FIX(0.168)+GOL*-FIX(0.331) 291 pmaddwd mm4,[GOTOFF(eax,PW_MF016_MF033)] ; mm4=ROH*-FIX(0.168)+GOH*-FIX(0.331) 292 293 movq MMWORD [wk(4)], mm1 ; wk(4)=ROL*FIX(0.299)+GOL*FIX(0.337) 294 movq MMWORD [wk(5)], mm6 ; wk(5)=ROH*FIX(0.299)+GOH*FIX(0.337) 295 296 pxor mm1,mm1 297 pxor mm6,mm6 298 punpcklwd mm1,mm5 ; mm1=BOL 299 punpckhwd mm6,mm5 ; mm6=BOH 300 psrld mm1,1 ; mm1=BOL*FIX(0.500) 301 psrld mm6,1 ; mm6=BOH*FIX(0.500) 302 303 movq mm5,[GOTOFF(eax,PD_ONEHALFM1_CJ)] ; mm5=[PD_ONEHALFM1_CJ] 304 305 paddd mm7,mm1 306 paddd mm4,mm6 307 paddd mm7,mm5 308 paddd mm4,mm5 309 psrld mm7,SCALEBITS ; mm7=CbOL 310 psrld mm4,SCALEBITS ; mm4=CbOH 311 packssdw mm7,mm4 ; mm7=CbO 312 313 movq mm1, MMWORD [wk(2)] ; mm1=BE 314 315 movq mm6,mm0 316 punpcklwd mm0,mm2 317 punpckhwd mm6,mm2 318 movq mm5,mm0 319 movq mm4,mm6 320 pmaddwd mm0,[GOTOFF(eax,PW_F0299_F0337)] ; mm0=REL*FIX(0.299)+GEL*FIX(0.337) 321 pmaddwd mm6,[GOTOFF(eax,PW_F0299_F0337)] ; mm6=REH*FIX(0.299)+GEH*FIX(0.337) 322 pmaddwd mm5,[GOTOFF(eax,PW_MF016_MF033)] ; mm5=REL*-FIX(0.168)+GEL*-FIX(0.331) 323 pmaddwd mm4,[GOTOFF(eax,PW_MF016_MF033)] ; mm4=REH*-FIX(0.168)+GEH*-FIX(0.331) 324 325 movq MMWORD [wk(6)], mm0 ; wk(6)=REL*FIX(0.299)+GEL*FIX(0.337) 326 movq MMWORD [wk(7)], mm6 ; wk(7)=REH*FIX(0.299)+GEH*FIX(0.337) 327 328 pxor mm0,mm0 329 pxor mm6,mm6 330 punpcklwd mm0,mm1 ; mm0=BEL 331 punpckhwd mm6,mm1 ; mm6=BEH 332 psrld mm0,1 ; mm0=BEL*FIX(0.500) 333 psrld mm6,1 ; mm6=BEH*FIX(0.500) 334 335 movq mm1,[GOTOFF(eax,PD_ONEHALFM1_CJ)] ; mm1=[PD_ONEHALFM1_CJ] 336 337 paddd mm5,mm0 338 paddd mm4,mm6 339 paddd mm5,mm1 340 paddd mm4,mm1 341 psrld mm5,SCALEBITS ; mm5=CbEL 342 psrld mm4,SCALEBITS ; mm4=CbEH 343 packssdw mm5,mm4 ; mm5=CbE 344 345 psllw mm7,BYTE_BIT 346 por mm5,mm7 ; mm5=Cb 347 movq MMWORD [ebx], mm5 ; Save Cb 348 349 movq mm0, MMWORD [wk(3)] ; mm0=BO 350 movq mm6, MMWORD [wk(2)] ; mm6=BE 351 movq mm1, MMWORD [wk(1)] ; mm1=RO 352 353 movq mm4,mm0 354 punpcklwd mm0,mm3 355 punpckhwd mm4,mm3 356 movq mm7,mm0 357 movq mm5,mm4 358 pmaddwd mm0,[GOTOFF(eax,PW_F0114_F0250)] ; mm0=BOL*FIX(0.114)+GOL*FIX(0.250) 359 pmaddwd mm4,[GOTOFF(eax,PW_F0114_F0250)] ; mm4=BOH*FIX(0.114)+GOH*FIX(0.250) 360 pmaddwd mm7,[GOTOFF(eax,PW_MF008_MF041)] ; mm7=BOL*-FIX(0.081)+GOL*-FIX(0.418) 361 pmaddwd mm5,[GOTOFF(eax,PW_MF008_MF041)] ; mm5=BOH*-FIX(0.081)+GOH*-FIX(0.418) 362 363 movq mm3,[GOTOFF(eax,PD_ONEHALF)] ; mm3=[PD_ONEHALF] 364 365 paddd mm0, MMWORD [wk(4)] 366 paddd mm4, MMWORD [wk(5)] 367 paddd mm0,mm3 368 paddd mm4,mm3 369 psrld mm0,SCALEBITS ; mm0=YOL 370 psrld mm4,SCALEBITS ; mm4=YOH 371 packssdw mm0,mm4 ; mm0=YO 372 373 pxor mm3,mm3 374 pxor mm4,mm4 375 punpcklwd mm3,mm1 ; mm3=ROL 376 punpckhwd mm4,mm1 ; mm4=ROH 377 psrld mm3,1 ; mm3=ROL*FIX(0.500) 378 psrld mm4,1 ; mm4=ROH*FIX(0.500) 379 380 movq mm1,[GOTOFF(eax,PD_ONEHALFM1_CJ)] ; mm1=[PD_ONEHALFM1_CJ] 381 382 paddd mm7,mm3 383 paddd mm5,mm4 384 paddd mm7,mm1 385 paddd mm5,mm1 386 psrld mm7,SCALEBITS ; mm7=CrOL 387 psrld mm5,SCALEBITS ; mm5=CrOH 388 packssdw mm7,mm5 ; mm7=CrO 389 390 movq mm3, MMWORD [wk(0)] ; mm3=RE 391 392 movq mm4,mm6 393 punpcklwd mm6,mm2 394 punpckhwd mm4,mm2 395 movq mm1,mm6 396 movq mm5,mm4 397 pmaddwd mm6,[GOTOFF(eax,PW_F0114_F0250)] ; mm6=BEL*FIX(0.114)+GEL*FIX(0.250) 398 pmaddwd mm4,[GOTOFF(eax,PW_F0114_F0250)] ; mm4=BEH*FIX(0.114)+GEH*FIX(0.250) 399 pmaddwd mm1,[GOTOFF(eax,PW_MF008_MF041)] ; mm1=BEL*-FIX(0.081)+GEL*-FIX(0.418) 400 pmaddwd mm5,[GOTOFF(eax,PW_MF008_MF041)] ; mm5=BEH*-FIX(0.081)+GEH*-FIX(0.418) 401 402 movq mm2,[GOTOFF(eax,PD_ONEHALF)] ; mm2=[PD_ONEHALF] 403 404 paddd mm6, MMWORD [wk(6)] 405 paddd mm4, MMWORD [wk(7)] 406 paddd mm6,mm2 407 paddd mm4,mm2 408 psrld mm6,SCALEBITS ; mm6=YEL 409 psrld mm4,SCALEBITS ; mm4=YEH 410 packssdw mm6,mm4 ; mm6=YE 411 412 psllw mm0,BYTE_BIT 413 por mm6,mm0 ; mm6=Y 414 movq MMWORD [edi], mm6 ; Save Y 415 416 pxor mm2,mm2 417 pxor mm4,mm4 418 punpcklwd mm2,mm3 ; mm2=REL 419 punpckhwd mm4,mm3 ; mm4=REH 420 psrld mm2,1 ; mm2=REL*FIX(0.500) 421 psrld mm4,1 ; mm4=REH*FIX(0.500) 422 423 movq mm0,[GOTOFF(eax,PD_ONEHALFM1_CJ)] ; mm0=[PD_ONEHALFM1_CJ] 424 425 paddd mm1,mm2 426 paddd mm5,mm4 427 paddd mm1,mm0 428 paddd mm5,mm0 429 psrld mm1,SCALEBITS ; mm1=CrEL 430 psrld mm5,SCALEBITS ; mm5=CrEH 431 packssdw mm1,mm5 ; mm1=CrE 432 433 psllw mm7,BYTE_BIT 434 por mm1,mm7 ; mm1=Cr 435 movq MMWORD [edx], mm1 ; Save Cr 436 437 sub ecx, byte SIZEOF_MMWORD 438 add esi, byte RGB_PIXELSIZE*SIZEOF_MMWORD ; inptr 439 add edi, byte SIZEOF_MMWORD ; outptr0 440 add ebx, byte SIZEOF_MMWORD ; outptr1 441 add edx, byte SIZEOF_MMWORD ; outptr2 442 cmp ecx, byte SIZEOF_MMWORD 443 jae near .columnloop 444 test ecx,ecx 445 jnz near .column_ld1 446 447 pop ecx ; col 448 pop esi 449 pop edi 450 pop ebx 451 pop edx 452 poppic eax 453 454 add esi, byte SIZEOF_JSAMPROW ; input_buf 455 add edi, byte SIZEOF_JSAMPROW 456 add ebx, byte SIZEOF_JSAMPROW 457 add edx, byte SIZEOF_JSAMPROW 458 dec eax ; num_rows 459 jg near .rowloop 460 461 emms ; empty MMX state 462 463.return: 464 pop edi 465 pop esi 466; pop edx ; need not be preserved 467; pop ecx ; need not be preserved 468 pop ebx 469 mov esp,ebp ; esp <- aligned ebp 470 pop esp ; esp <- original ebp 471 pop ebp 472 ret 473 474; For some reason, the OS X linker does not honor the request to align the 475; segment unless we do this. 476 align 16 477