1; 2; jcgryext.asm - grayscale colorspace conversion (MMX) 3; 4; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB 5; Copyright (C) 2011, 2016, 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, JSAMPARRAY input_buf, 27; JSAMPIMAGE output_buf, JDIMENSION output_row, 28; 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 39 ; mmword wk[WK_NUM] 40%define WK_NUM 2 41%define gotptr wk(0) - SIZEOF_POINTER ; void * gotptr 42 43 align 32 44 GLOBAL_FUNCTION(jsimd_rgb_gray_convert_mmx) 45 46EXTN(jsimd_rgb_gray_convert_mmx): 47 push ebp 48 mov eax, esp ; eax = original ebp 49 sub esp, byte 4 50 and esp, byte (-SIZEOF_MMWORD) ; align to 64 bits 51 mov [esp], eax 52 mov ebp, esp ; ebp = aligned ebp 53 lea esp, [wk(0)] 54 pushpic eax ; make a room for GOT address 55 push ebx 56; push ecx ; need not be preserved 57; push edx ; need not be preserved 58 push esi 59 push edi 60 61 get_GOT ebx ; get GOT address 62 movpic POINTER [gotptr], ebx ; save GOT address 63 64 mov ecx, JDIMENSION [img_width(eax)] ; num_cols 65 test ecx, ecx 66 jz near .return 67 68 push ecx 69 70 mov esi, JSAMPIMAGE [output_buf(eax)] 71 mov ecx, JDIMENSION [output_row(eax)] 72 mov edi, JSAMPARRAY [esi+0*SIZEOF_JSAMPARRAY] 73 lea edi, [edi+ecx*SIZEOF_JSAMPROW] 74 75 pop ecx 76 77 mov esi, JSAMPARRAY [input_buf(eax)] 78 mov eax, INT [num_rows(eax)] 79 test eax, eax 80 jle near .return 81 alignx 16, 7 82.rowloop: 83 pushpic eax 84 push edi 85 push esi 86 push ecx ; col 87 88 mov esi, JSAMPROW [esi] ; inptr 89 mov edi, JSAMPROW [edi] ; outptr0 90 movpic eax, POINTER [gotptr] ; load GOT address (eax) 91 92 cmp ecx, byte SIZEOF_MMWORD 93 jae short .columnloop 94 alignx 16, 7 95 96%if RGB_PIXELSIZE == 3 ; --------------- 97 98.column_ld1: 99 push eax 100 push edx 101 lea ecx, [ecx+ecx*2] ; imul ecx,RGB_PIXELSIZE 102 test cl, SIZEOF_BYTE 103 jz short .column_ld2 104 sub ecx, byte SIZEOF_BYTE 105 xor eax, eax 106 mov al, BYTE [esi+ecx] 107.column_ld2: 108 test cl, SIZEOF_WORD 109 jz short .column_ld4 110 sub ecx, byte SIZEOF_WORD 111 xor edx, edx 112 mov dx, WORD [esi+ecx] 113 shl eax, WORD_BIT 114 or eax, edx 115.column_ld4: 116 movd mmA, eax 117 pop edx 118 pop eax 119 test cl, SIZEOF_DWORD 120 jz short .column_ld8 121 sub ecx, byte SIZEOF_DWORD 122 movd mmG, DWORD [esi+ecx] 123 psllq mmA, DWORD_BIT 124 por mmA, mmG 125.column_ld8: 126 test cl, SIZEOF_MMWORD 127 jz short .column_ld16 128 movq mmG, mmA 129 movq mmA, MMWORD [esi+0*SIZEOF_MMWORD] 130 mov ecx, SIZEOF_MMWORD 131 jmp short .rgb_gray_cnv 132.column_ld16: 133 test cl, 2*SIZEOF_MMWORD 134 mov ecx, SIZEOF_MMWORD 135 jz short .rgb_gray_cnv 136 movq mmF, mmA 137 movq mmA, MMWORD [esi+0*SIZEOF_MMWORD] 138 movq mmG, MMWORD [esi+1*SIZEOF_MMWORD] 139 jmp short .rgb_gray_cnv 140 alignx 16, 7 141 142.columnloop: 143 movq mmA, MMWORD [esi+0*SIZEOF_MMWORD] 144 movq mmG, MMWORD [esi+1*SIZEOF_MMWORD] 145 movq mmF, MMWORD [esi+2*SIZEOF_MMWORD] 146 147.rgb_gray_cnv: 148 ; mmA=(00 10 20 01 11 21 02 12) 149 ; mmG=(22 03 13 23 04 14 24 05) 150 ; mmF=(15 25 06 16 26 07 17 27) 151 152 movq mmD, mmA 153 psllq mmA, 4*BYTE_BIT ; mmA=(-- -- -- -- 00 10 20 01) 154 psrlq mmD, 4*BYTE_BIT ; mmD=(11 21 02 12 -- -- -- --) 155 156 punpckhbw mmA, mmG ; mmA=(00 04 10 14 20 24 01 05) 157 psllq mmG, 4*BYTE_BIT ; mmG=(-- -- -- -- 22 03 13 23) 158 159 punpcklbw mmD, mmF ; mmD=(11 15 21 25 02 06 12 16) 160 punpckhbw mmG, mmF ; mmG=(22 26 03 07 13 17 23 27) 161 162 movq mmE, mmA 163 psllq mmA, 4*BYTE_BIT ; mmA=(-- -- -- -- 00 04 10 14) 164 psrlq mmE, 4*BYTE_BIT ; mmE=(20 24 01 05 -- -- -- --) 165 166 punpckhbw mmA, mmD ; mmA=(00 02 04 06 10 12 14 16) 167 psllq mmD, 4*BYTE_BIT ; mmD=(-- -- -- -- 11 15 21 25) 168 169 punpcklbw mmE, mmG ; mmE=(20 22 24 26 01 03 05 07) 170 punpckhbw mmD, mmG ; mmD=(11 13 15 17 21 23 25 27) 171 172 pxor mmH, mmH 173 174 movq mmC, mmA 175 punpcklbw mmA, mmH ; mmA=(00 02 04 06) 176 punpckhbw mmC, mmH ; mmC=(10 12 14 16) 177 178 movq mmB, mmE 179 punpcklbw mmE, mmH ; mmE=(20 22 24 26) 180 punpckhbw mmB, mmH ; mmB=(01 03 05 07) 181 182 movq mmF, mmD 183 punpcklbw mmD, mmH ; mmD=(11 13 15 17) 184 punpckhbw mmF, mmH ; mmF=(21 23 25 27) 185 186%else ; RGB_PIXELSIZE == 4 ; ----------- 187 188.column_ld1: 189 test cl, SIZEOF_MMWORD/8 190 jz short .column_ld2 191 sub ecx, byte SIZEOF_MMWORD/8 192 movd mmA, DWORD [esi+ecx*RGB_PIXELSIZE] 193.column_ld2: 194 test cl, SIZEOF_MMWORD/4 195 jz short .column_ld4 196 sub ecx, byte SIZEOF_MMWORD/4 197 movq mmF, mmA 198 movq mmA, MMWORD [esi+ecx*RGB_PIXELSIZE] 199.column_ld4: 200 test cl, SIZEOF_MMWORD/2 201 mov ecx, SIZEOF_MMWORD 202 jz short .rgb_gray_cnv 203 movq mmD, mmA 204 movq mmC, mmF 205 movq mmA, MMWORD [esi+0*SIZEOF_MMWORD] 206 movq mmF, MMWORD [esi+1*SIZEOF_MMWORD] 207 jmp short .rgb_gray_cnv 208 alignx 16, 7 209 210.columnloop: 211 movq mmA, MMWORD [esi+0*SIZEOF_MMWORD] 212 movq mmF, MMWORD [esi+1*SIZEOF_MMWORD] 213 movq mmD, MMWORD [esi+2*SIZEOF_MMWORD] 214 movq mmC, MMWORD [esi+3*SIZEOF_MMWORD] 215 216.rgb_gray_cnv: 217 ; mmA=(00 10 20 30 01 11 21 31) 218 ; mmF=(02 12 22 32 03 13 23 33) 219 ; mmD=(04 14 24 34 05 15 25 35) 220 ; mmC=(06 16 26 36 07 17 27 37) 221 222 movq mmB, mmA 223 punpcklbw mmA, mmF ; mmA=(00 02 10 12 20 22 30 32) 224 punpckhbw mmB, mmF ; mmB=(01 03 11 13 21 23 31 33) 225 226 movq mmG, mmD 227 punpcklbw mmD, mmC ; mmD=(04 06 14 16 24 26 34 36) 228 punpckhbw mmG, mmC ; mmG=(05 07 15 17 25 27 35 37) 229 230 movq mmE, mmA 231 punpcklwd mmA, mmD ; mmA=(00 02 04 06 10 12 14 16) 232 punpckhwd mmE, mmD ; mmE=(20 22 24 26 30 32 34 36) 233 234 movq mmH, mmB 235 punpcklwd mmB, mmG ; mmB=(01 03 05 07 11 13 15 17) 236 punpckhwd mmH, mmG ; mmH=(21 23 25 27 31 33 35 37) 237 238 pxor mmF, mmF 239 240 movq mmC, mmA 241 punpcklbw mmA, mmF ; mmA=(00 02 04 06) 242 punpckhbw mmC, mmF ; mmC=(10 12 14 16) 243 244 movq mmD, mmB 245 punpcklbw mmB, mmF ; mmB=(01 03 05 07) 246 punpckhbw mmD, mmF ; mmD=(11 13 15 17) 247 248 movq mmG, mmE 249 punpcklbw mmE, mmF ; mmE=(20 22 24 26) 250 punpckhbw mmG, mmF ; mmG=(30 32 34 36) 251 252 punpcklbw mmF, mmH 253 punpckhbw mmH, mmH 254 psrlw mmF, BYTE_BIT ; mmF=(21 23 25 27) 255 psrlw mmH, BYTE_BIT ; mmH=(31 33 35 37) 256 257%endif ; RGB_PIXELSIZE ; --------------- 258 259 ; mm0=(R0 R2 R4 R6)=RE, mm2=(G0 G2 G4 G6)=GE, mm4=(B0 B2 B4 B6)=BE 260 ; mm1=(R1 R3 R5 R7)=RO, mm3=(G1 G3 G5 G7)=GO, mm5=(B1 B3 B5 B7)=BO 261 262 ; (Original) 263 ; Y = 0.29900 * R + 0.58700 * G + 0.11400 * B 264 ; 265 ; (This implementation) 266 ; Y = 0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G 267 268 movq mm6, mm1 269 punpcklwd mm1, mm3 270 punpckhwd mm6, mm3 271 pmaddwd mm1, [GOTOFF(eax,PW_F0299_F0337)] ; mm1=ROL*FIX(0.299)+GOL*FIX(0.337) 272 pmaddwd mm6, [GOTOFF(eax,PW_F0299_F0337)] ; mm6=ROH*FIX(0.299)+GOH*FIX(0.337) 273 274 movq mm7, mm6 ; mm7=ROH*FIX(0.299)+GOH*FIX(0.337) 275 276 movq mm6, mm0 277 punpcklwd mm0, mm2 278 punpckhwd mm6, mm2 279 pmaddwd mm0, [GOTOFF(eax,PW_F0299_F0337)] ; mm0=REL*FIX(0.299)+GEL*FIX(0.337) 280 pmaddwd mm6, [GOTOFF(eax,PW_F0299_F0337)] ; mm6=REH*FIX(0.299)+GEH*FIX(0.337) 281 282 movq MMWORD [wk(0)], mm0 ; wk(0)=REL*FIX(0.299)+GEL*FIX(0.337) 283 movq MMWORD [wk(1)], mm6 ; wk(1)=REH*FIX(0.299)+GEH*FIX(0.337) 284 285 movq mm0, mm5 ; mm0=BO 286 movq mm6, mm4 ; mm6=BE 287 288 movq mm4, mm0 289 punpcklwd mm0, mm3 290 punpckhwd mm4, mm3 291 pmaddwd mm0, [GOTOFF(eax,PW_F0114_F0250)] ; mm0=BOL*FIX(0.114)+GOL*FIX(0.250) 292 pmaddwd mm4, [GOTOFF(eax,PW_F0114_F0250)] ; mm4=BOH*FIX(0.114)+GOH*FIX(0.250) 293 294 movq mm3, [GOTOFF(eax,PD_ONEHALF)] ; mm3=[PD_ONEHALF] 295 296 paddd mm0, mm1 297 paddd mm4, mm7 298 paddd mm0, mm3 299 paddd mm4, mm3 300 psrld mm0, SCALEBITS ; mm0=YOL 301 psrld mm4, SCALEBITS ; mm4=YOH 302 packssdw mm0, mm4 ; mm0=YO 303 304 movq mm4, mm6 305 punpcklwd mm6, mm2 306 punpckhwd mm4, mm2 307 pmaddwd mm6, [GOTOFF(eax,PW_F0114_F0250)] ; mm6=BEL*FIX(0.114)+GEL*FIX(0.250) 308 pmaddwd mm4, [GOTOFF(eax,PW_F0114_F0250)] ; mm4=BEH*FIX(0.114)+GEH*FIX(0.250) 309 310 movq mm2, [GOTOFF(eax,PD_ONEHALF)] ; mm2=[PD_ONEHALF] 311 312 paddd mm6, MMWORD [wk(0)] 313 paddd mm4, MMWORD [wk(1)] 314 paddd mm6, mm2 315 paddd mm4, mm2 316 psrld mm6, SCALEBITS ; mm6=YEL 317 psrld mm4, SCALEBITS ; mm4=YEH 318 packssdw mm6, mm4 ; mm6=YE 319 320 psllw mm0, BYTE_BIT 321 por mm6, mm0 ; mm6=Y 322 movq MMWORD [edi], mm6 ; Save Y 323 324 sub ecx, byte SIZEOF_MMWORD 325 add esi, byte RGB_PIXELSIZE*SIZEOF_MMWORD ; inptr 326 add edi, byte SIZEOF_MMWORD ; outptr0 327 cmp ecx, byte SIZEOF_MMWORD 328 jae near .columnloop 329 test ecx, ecx 330 jnz near .column_ld1 331 332 pop ecx ; col 333 pop esi 334 pop edi 335 poppic eax 336 337 add esi, byte SIZEOF_JSAMPROW ; input_buf 338 add edi, byte SIZEOF_JSAMPROW 339 dec eax ; num_rows 340 jg near .rowloop 341 342 emms ; empty MMX state 343 344.return: 345 pop edi 346 pop esi 347; pop edx ; need not be preserved 348; pop ecx ; need not be preserved 349 pop ebx 350 mov esp, ebp ; esp <- aligned ebp 351 pop esp ; esp <- original ebp 352 pop ebp 353 ret 354 355; For some reason, the OS X linker does not honor the request to align the 356; segment unless we do this. 357 align 32 358