1; 2; jcgryext.asm - grayscale colorspace conversion (MMX) 3; 4; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB 5; Copyright (C) 2011, D. R. Commander. 6; 7; Based on the x86 SIMD extension for IJG JPEG library 8; Copyright (C) 1999-2006, MIYASAKA Masaru. 9; For conditions of distribution and use, see copyright notice in jsimdext.inc 10; 11; This file should be assembled with NASM (Netwide Assembler), 12; can *not* be assembled with Microsoft's MASM or any compatible 13; assembler (including Borland's Turbo Assembler). 14; NASM is available from http://nasm.sourceforge.net/ or 15; http://sourceforge.net/project/showfiles.php?group_id=6208 16; 17; [TAB8] 18 19%include "jcolsamp.inc" 20 21; -------------------------------------------------------------------------- 22; 23; Convert some rows of samples to the output colorspace. 24; 25; GLOBAL(void) 26; jsimd_rgb_gray_convert_mmx (JDIMENSION img_width, 27; JSAMPARRAY input_buf, JSAMPIMAGE output_buf, 28; JDIMENSION output_row, int num_rows); 29; 30 31%define img_width(b) (b)+8 ; JDIMENSION img_width 32%define input_buf(b) (b)+12 ; JSAMPARRAY input_buf 33%define output_buf(b) (b)+16 ; JSAMPIMAGE output_buf 34%define output_row(b) (b)+20 ; JDIMENSION output_row 35%define num_rows(b) (b)+24 ; int num_rows 36 37%define original_ebp ebp+0 38%define wk(i) ebp-(WK_NUM-(i))*SIZEOF_MMWORD ; mmword wk[WK_NUM] 39%define WK_NUM 2 40%define gotptr wk(0)-SIZEOF_POINTER ; void * gotptr 41 42 align 16 43 global EXTN(jsimd_rgb_gray_convert_mmx) 44 45EXTN(jsimd_rgb_gray_convert_mmx): 46 push ebp 47 mov eax,esp ; eax = original ebp 48 sub esp, byte 4 49 and esp, byte (-SIZEOF_MMWORD) ; align to 64 bits 50 mov [esp],eax 51 mov ebp,esp ; ebp = aligned ebp 52 lea esp, [wk(0)] 53 pushpic eax ; make a room for GOT address 54 push ebx 55; push ecx ; need not be preserved 56; push edx ; need not be preserved 57 push esi 58 push edi 59 60 get_GOT ebx ; get GOT address 61 movpic POINTER [gotptr], ebx ; save GOT address 62 63 mov ecx, JDIMENSION [img_width(eax)] ; num_cols 64 test ecx,ecx 65 jz near .return 66 67 push ecx 68 69 mov esi, JSAMPIMAGE [output_buf(eax)] 70 mov ecx, JDIMENSION [output_row(eax)] 71 mov edi, JSAMPARRAY [esi+0*SIZEOF_JSAMPARRAY] 72 lea edi, [edi+ecx*SIZEOF_JSAMPROW] 73 74 pop ecx 75 76 mov esi, JSAMPARRAY [input_buf(eax)] 77 mov eax, INT [num_rows(eax)] 78 test eax,eax 79 jle near .return 80 alignx 16,7 81.rowloop: 82 pushpic eax 83 push edi 84 push esi 85 push ecx ; col 86 87 mov esi, JSAMPROW [esi] ; inptr 88 mov edi, JSAMPROW [edi] ; outptr0 89 movpic eax, POINTER [gotptr] ; load GOT address (eax) 90 91 cmp ecx, byte SIZEOF_MMWORD 92 jae short .columnloop 93 alignx 16,7 94 95%if RGB_PIXELSIZE == 3 ; --------------- 96 97.column_ld1: 98 push eax 99 push edx 100 lea ecx,[ecx+ecx*2] ; imul ecx,RGB_PIXELSIZE 101 test cl, SIZEOF_BYTE 102 jz short .column_ld2 103 sub ecx, byte SIZEOF_BYTE 104 xor eax,eax 105 mov al, BYTE [esi+ecx] 106.column_ld2: 107 test cl, SIZEOF_WORD 108 jz short .column_ld4 109 sub ecx, byte SIZEOF_WORD 110 xor edx,edx 111 mov dx, WORD [esi+ecx] 112 shl eax, WORD_BIT 113 or eax,edx 114.column_ld4: 115 movd mmA,eax 116 pop edx 117 pop eax 118 test cl, SIZEOF_DWORD 119 jz short .column_ld8 120 sub ecx, byte SIZEOF_DWORD 121 movd mmG, DWORD [esi+ecx] 122 psllq mmA, DWORD_BIT 123 por mmA,mmG 124.column_ld8: 125 test cl, SIZEOF_MMWORD 126 jz short .column_ld16 127 movq mmG,mmA 128 movq mmA, MMWORD [esi+0*SIZEOF_MMWORD] 129 mov ecx, SIZEOF_MMWORD 130 jmp short .rgb_gray_cnv 131.column_ld16: 132 test cl, 2*SIZEOF_MMWORD 133 mov ecx, SIZEOF_MMWORD 134 jz short .rgb_gray_cnv 135 movq mmF,mmA 136 movq mmA, MMWORD [esi+0*SIZEOF_MMWORD] 137 movq mmG, MMWORD [esi+1*SIZEOF_MMWORD] 138 jmp short .rgb_gray_cnv 139 alignx 16,7 140 141.columnloop: 142 movq mmA, MMWORD [esi+0*SIZEOF_MMWORD] 143 movq mmG, MMWORD [esi+1*SIZEOF_MMWORD] 144 movq mmF, MMWORD [esi+2*SIZEOF_MMWORD] 145 146.rgb_gray_cnv: 147 ; mmA=(00 10 20 01 11 21 02 12) 148 ; mmG=(22 03 13 23 04 14 24 05) 149 ; mmF=(15 25 06 16 26 07 17 27) 150 151 movq mmD,mmA 152 psllq mmA,4*BYTE_BIT ; mmA=(-- -- -- -- 00 10 20 01) 153 psrlq mmD,4*BYTE_BIT ; mmD=(11 21 02 12 -- -- -- --) 154 155 punpckhbw mmA,mmG ; mmA=(00 04 10 14 20 24 01 05) 156 psllq mmG,4*BYTE_BIT ; mmG=(-- -- -- -- 22 03 13 23) 157 158 punpcklbw mmD,mmF ; mmD=(11 15 21 25 02 06 12 16) 159 punpckhbw mmG,mmF ; mmG=(22 26 03 07 13 17 23 27) 160 161 movq mmE,mmA 162 psllq mmA,4*BYTE_BIT ; mmA=(-- -- -- -- 00 04 10 14) 163 psrlq mmE,4*BYTE_BIT ; mmE=(20 24 01 05 -- -- -- --) 164 165 punpckhbw mmA,mmD ; mmA=(00 02 04 06 10 12 14 16) 166 psllq mmD,4*BYTE_BIT ; mmD=(-- -- -- -- 11 15 21 25) 167 168 punpcklbw mmE,mmG ; mmE=(20 22 24 26 01 03 05 07) 169 punpckhbw mmD,mmG ; mmD=(11 13 15 17 21 23 25 27) 170 171 pxor mmH,mmH 172 173 movq mmC,mmA 174 punpcklbw mmA,mmH ; mmA=(00 02 04 06) 175 punpckhbw mmC,mmH ; mmC=(10 12 14 16) 176 177 movq mmB,mmE 178 punpcklbw mmE,mmH ; mmE=(20 22 24 26) 179 punpckhbw mmB,mmH ; mmB=(01 03 05 07) 180 181 movq mmF,mmD 182 punpcklbw mmD,mmH ; mmD=(11 13 15 17) 183 punpckhbw mmF,mmH ; mmF=(21 23 25 27) 184 185%else ; RGB_PIXELSIZE == 4 ; ----------- 186 187.column_ld1: 188 test cl, SIZEOF_MMWORD/8 189 jz short .column_ld2 190 sub ecx, byte SIZEOF_MMWORD/8 191 movd mmA, DWORD [esi+ecx*RGB_PIXELSIZE] 192.column_ld2: 193 test cl, SIZEOF_MMWORD/4 194 jz short .column_ld4 195 sub ecx, byte SIZEOF_MMWORD/4 196 movq mmF,mmA 197 movq mmA, MMWORD [esi+ecx*RGB_PIXELSIZE] 198.column_ld4: 199 test cl, SIZEOF_MMWORD/2 200 mov ecx, SIZEOF_MMWORD 201 jz short .rgb_gray_cnv 202 movq mmD,mmA 203 movq mmC,mmF 204 movq mmA, MMWORD [esi+0*SIZEOF_MMWORD] 205 movq mmF, MMWORD [esi+1*SIZEOF_MMWORD] 206 jmp short .rgb_gray_cnv 207 alignx 16,7 208 209.columnloop: 210 movq mmA, MMWORD [esi+0*SIZEOF_MMWORD] 211 movq mmF, MMWORD [esi+1*SIZEOF_MMWORD] 212 movq mmD, MMWORD [esi+2*SIZEOF_MMWORD] 213 movq mmC, MMWORD [esi+3*SIZEOF_MMWORD] 214 215.rgb_gray_cnv: 216 ; mmA=(00 10 20 30 01 11 21 31) 217 ; mmF=(02 12 22 32 03 13 23 33) 218 ; mmD=(04 14 24 34 05 15 25 35) 219 ; mmC=(06 16 26 36 07 17 27 37) 220 221 movq mmB,mmA 222 punpcklbw mmA,mmF ; mmA=(00 02 10 12 20 22 30 32) 223 punpckhbw mmB,mmF ; mmB=(01 03 11 13 21 23 31 33) 224 225 movq mmG,mmD 226 punpcklbw mmD,mmC ; mmD=(04 06 14 16 24 26 34 36) 227 punpckhbw mmG,mmC ; mmG=(05 07 15 17 25 27 35 37) 228 229 movq mmE,mmA 230 punpcklwd mmA,mmD ; mmA=(00 02 04 06 10 12 14 16) 231 punpckhwd mmE,mmD ; mmE=(20 22 24 26 30 32 34 36) 232 233 movq mmH,mmB 234 punpcklwd mmB,mmG ; mmB=(01 03 05 07 11 13 15 17) 235 punpckhwd mmH,mmG ; mmH=(21 23 25 27 31 33 35 37) 236 237 pxor mmF,mmF 238 239 movq mmC,mmA 240 punpcklbw mmA,mmF ; mmA=(00 02 04 06) 241 punpckhbw mmC,mmF ; mmC=(10 12 14 16) 242 243 movq mmD,mmB 244 punpcklbw mmB,mmF ; mmB=(01 03 05 07) 245 punpckhbw mmD,mmF ; mmD=(11 13 15 17) 246 247 movq mmG,mmE 248 punpcklbw mmE,mmF ; mmE=(20 22 24 26) 249 punpckhbw mmG,mmF ; mmG=(30 32 34 36) 250 251 punpcklbw mmF,mmH 252 punpckhbw mmH,mmH 253 psrlw mmF,BYTE_BIT ; mmF=(21 23 25 27) 254 psrlw mmH,BYTE_BIT ; mmH=(31 33 35 37) 255 256%endif ; RGB_PIXELSIZE ; --------------- 257 258 ; mm0=(R0 R2 R4 R6)=RE, mm2=(G0 G2 G4 G6)=GE, mm4=(B0 B2 B4 B6)=BE 259 ; mm1=(R1 R3 R5 R7)=RO, mm3=(G1 G3 G5 G7)=GO, mm5=(B1 B3 B5 B7)=BO 260 261 ; (Original) 262 ; Y = 0.29900 * R + 0.58700 * G + 0.11400 * B 263 ; 264 ; (This implementation) 265 ; Y = 0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G 266 267 movq mm6,mm1 268 punpcklwd mm1,mm3 269 punpckhwd mm6,mm3 270 pmaddwd mm1,[GOTOFF(eax,PW_F0299_F0337)] ; mm1=ROL*FIX(0.299)+GOL*FIX(0.337) 271 pmaddwd mm6,[GOTOFF(eax,PW_F0299_F0337)] ; mm6=ROH*FIX(0.299)+GOH*FIX(0.337) 272 273 movq mm7, mm6 ; mm7=ROH*FIX(0.299)+GOH*FIX(0.337) 274 275 movq mm6,mm0 276 punpcklwd mm0,mm2 277 punpckhwd mm6,mm2 278 pmaddwd mm0,[GOTOFF(eax,PW_F0299_F0337)] ; mm0=REL*FIX(0.299)+GEL*FIX(0.337) 279 pmaddwd mm6,[GOTOFF(eax,PW_F0299_F0337)] ; mm6=REH*FIX(0.299)+GEH*FIX(0.337) 280 281 movq MMWORD [wk(0)], mm0 ; wk(0)=REL*FIX(0.299)+GEL*FIX(0.337) 282 movq MMWORD [wk(1)], mm6 ; wk(1)=REH*FIX(0.299)+GEH*FIX(0.337) 283 284 movq mm0, mm5 ; mm0=BO 285 movq mm6, mm4 ; mm6=BE 286 287 movq mm4,mm0 288 punpcklwd mm0,mm3 289 punpckhwd mm4,mm3 290 pmaddwd mm0,[GOTOFF(eax,PW_F0114_F0250)] ; mm0=BOL*FIX(0.114)+GOL*FIX(0.250) 291 pmaddwd mm4,[GOTOFF(eax,PW_F0114_F0250)] ; mm4=BOH*FIX(0.114)+GOH*FIX(0.250) 292 293 movq mm3,[GOTOFF(eax,PD_ONEHALF)] ; mm3=[PD_ONEHALF] 294 295 paddd mm0, mm1 296 paddd mm4, mm7 297 paddd mm0,mm3 298 paddd mm4,mm3 299 psrld mm0,SCALEBITS ; mm0=YOL 300 psrld mm4,SCALEBITS ; mm4=YOH 301 packssdw mm0,mm4 ; mm0=YO 302 303 movq mm4,mm6 304 punpcklwd mm6,mm2 305 punpckhwd mm4,mm2 306 pmaddwd mm6,[GOTOFF(eax,PW_F0114_F0250)] ; mm6=BEL*FIX(0.114)+GEL*FIX(0.250) 307 pmaddwd mm4,[GOTOFF(eax,PW_F0114_F0250)] ; mm4=BEH*FIX(0.114)+GEH*FIX(0.250) 308 309 movq mm2,[GOTOFF(eax,PD_ONEHALF)] ; mm2=[PD_ONEHALF] 310 311 paddd mm6, MMWORD [wk(0)] 312 paddd mm4, MMWORD [wk(1)] 313 paddd mm6,mm2 314 paddd mm4,mm2 315 psrld mm6,SCALEBITS ; mm6=YEL 316 psrld mm4,SCALEBITS ; mm4=YEH 317 packssdw mm6,mm4 ; mm6=YE 318 319 psllw mm0,BYTE_BIT 320 por mm6,mm0 ; mm6=Y 321 movq MMWORD [edi], mm6 ; Save Y 322 323 sub ecx, byte SIZEOF_MMWORD 324 add esi, byte RGB_PIXELSIZE*SIZEOF_MMWORD ; inptr 325 add edi, byte SIZEOF_MMWORD ; outptr0 326 cmp ecx, byte SIZEOF_MMWORD 327 jae near .columnloop 328 test ecx,ecx 329 jnz near .column_ld1 330 331 pop ecx ; col 332 pop esi 333 pop edi 334 poppic eax 335 336 add esi, byte SIZEOF_JSAMPROW ; input_buf 337 add edi, byte SIZEOF_JSAMPROW 338 dec eax ; num_rows 339 jg near .rowloop 340 341 emms ; empty MMX state 342 343.return: 344 pop edi 345 pop esi 346; pop edx ; need not be preserved 347; pop ecx ; need not be preserved 348 pop ebx 349 mov esp,ebp ; esp <- aligned ebp 350 pop esp ; esp <- original ebp 351 pop ebp 352 ret 353 354; For some reason, the OS X linker does not honor the request to align the 355; segment unless we do this. 356 align 16 357