jcgryext-mmx.asm (12109B)
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, 2024, 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) or Yasm. 12 13 %include "jcolsamp.inc" 14 15 ; -------------------------------------------------------------------------- 16 ; 17 ; Convert some rows of samples to the output colorspace. 18 ; 19 ; GLOBAL(void) 20 ; jsimd_rgb_gray_convert_mmx(JDIMENSION img_width, JSAMPARRAY input_buf, 21 ; JSAMPIMAGE output_buf, JDIMENSION output_row, 22 ; int num_rows); 23 ; 24 25 %define img_width(b) (b) + 8 ; JDIMENSION img_width 26 %define input_buf(b) (b) + 12 ; JSAMPARRAY input_buf 27 %define output_buf(b) (b) + 16 ; JSAMPIMAGE output_buf 28 %define output_row(b) (b) + 20 ; JDIMENSION output_row 29 %define num_rows(b) (b) + 24 ; int num_rows 30 31 %define original_ebp ebp + 0 32 %define wk(i) ebp - (WK_NUM - (i)) * SIZEOF_MMWORD 33 ; mmword wk[WK_NUM] 34 %define WK_NUM 2 35 %define gotptr wk(0) - SIZEOF_POINTER ; void * gotptr 36 37 align 32 38 GLOBAL_FUNCTION(jsimd_rgb_gray_convert_mmx) 39 40 EXTN(jsimd_rgb_gray_convert_mmx): 41 push ebp 42 mov eax, esp ; eax = original ebp 43 sub esp, byte 4 44 and esp, byte (-SIZEOF_MMWORD) ; align to 64 bits 45 mov [esp], eax 46 mov ebp, esp ; ebp = aligned ebp 47 lea esp, [wk(0)] 48 PUSHPIC eax ; make a room for GOT address 49 push ebx 50 ; push ecx ; need not be preserved 51 ; push edx ; need not be preserved 52 push esi 53 push edi 54 55 GET_GOT ebx ; get GOT address 56 MOVPIC POINTER [gotptr], ebx ; save GOT address 57 58 mov ecx, JDIMENSION [img_width(eax)] ; num_cols 59 test ecx, ecx 60 jz near .return 61 62 push ecx 63 64 mov esi, JSAMPIMAGE [output_buf(eax)] 65 mov ecx, JDIMENSION [output_row(eax)] 66 mov edi, JSAMPARRAY [esi+0*SIZEOF_JSAMPARRAY] 67 lea edi, [edi+ecx*SIZEOF_JSAMPROW] 68 69 pop ecx 70 71 mov esi, JSAMPARRAY [input_buf(eax)] 72 mov eax, INT [num_rows(eax)] 73 test eax, eax 74 jle near .return 75 ALIGNX 16, 7 76 .rowloop: 77 PUSHPIC eax 78 push edi 79 push esi 80 push ecx ; col 81 82 mov esi, JSAMPROW [esi] ; inptr 83 mov edi, JSAMPROW [edi] ; outptr0 84 MOVPIC eax, POINTER [gotptr] ; load GOT address (eax) 85 86 cmp ecx, byte SIZEOF_MMWORD 87 jae short .columnloop 88 ALIGNX 16, 7 89 90 %if RGB_PIXELSIZE == 3 ; --------------- 91 92 .column_ld1: 93 push eax 94 push edx 95 lea ecx, [ecx+ecx*2] ; imul ecx,RGB_PIXELSIZE 96 test cl, SIZEOF_BYTE 97 jz short .column_ld2 98 sub ecx, byte SIZEOF_BYTE 99 xor eax, eax 100 mov al, byte [esi+ecx] 101 .column_ld2: 102 test cl, SIZEOF_WORD 103 jz short .column_ld4 104 sub ecx, byte SIZEOF_WORD 105 xor edx, edx 106 mov dx, word [esi+ecx] 107 shl eax, WORD_BIT 108 or eax, edx 109 .column_ld4: 110 movd mmA, eax 111 pop edx 112 pop eax 113 test cl, SIZEOF_DWORD 114 jz short .column_ld8 115 sub ecx, byte SIZEOF_DWORD 116 movd mmG, dword [esi+ecx] 117 psllq mmA, DWORD_BIT 118 por mmA, mmG 119 .column_ld8: 120 test cl, SIZEOF_MMWORD 121 jz short .column_ld16 122 movq mmG, mmA 123 movq mmA, MMWORD [esi+0*SIZEOF_MMWORD] 124 mov ecx, SIZEOF_MMWORD 125 jmp short .rgb_gray_cnv 126 .column_ld16: 127 test cl, 2*SIZEOF_MMWORD 128 mov ecx, SIZEOF_MMWORD 129 jz short .rgb_gray_cnv 130 movq mmF, mmA 131 movq mmA, MMWORD [esi+0*SIZEOF_MMWORD] 132 movq mmG, MMWORD [esi+1*SIZEOF_MMWORD] 133 jmp short .rgb_gray_cnv 134 ALIGNX 16, 7 135 136 .columnloop: 137 movq mmA, MMWORD [esi+0*SIZEOF_MMWORD] 138 movq mmG, MMWORD [esi+1*SIZEOF_MMWORD] 139 movq mmF, MMWORD [esi+2*SIZEOF_MMWORD] 140 141 .rgb_gray_cnv: 142 ; mmA=(00 10 20 01 11 21 02 12) 143 ; mmG=(22 03 13 23 04 14 24 05) 144 ; mmF=(15 25 06 16 26 07 17 27) 145 146 movq mmD, mmA 147 psllq mmA, 4*BYTE_BIT ; mmA=(-- -- -- -- 00 10 20 01) 148 psrlq mmD, 4*BYTE_BIT ; mmD=(11 21 02 12 -- -- -- --) 149 150 punpckhbw mmA, mmG ; mmA=(00 04 10 14 20 24 01 05) 151 psllq mmG, 4*BYTE_BIT ; mmG=(-- -- -- -- 22 03 13 23) 152 153 punpcklbw mmD, mmF ; mmD=(11 15 21 25 02 06 12 16) 154 punpckhbw mmG, mmF ; mmG=(22 26 03 07 13 17 23 27) 155 156 movq mmE, mmA 157 psllq mmA, 4*BYTE_BIT ; mmA=(-- -- -- -- 00 04 10 14) 158 psrlq mmE, 4*BYTE_BIT ; mmE=(20 24 01 05 -- -- -- --) 159 160 punpckhbw mmA, mmD ; mmA=(00 02 04 06 10 12 14 16) 161 psllq mmD, 4*BYTE_BIT ; mmD=(-- -- -- -- 11 15 21 25) 162 163 punpcklbw mmE, mmG ; mmE=(20 22 24 26 01 03 05 07) 164 punpckhbw mmD, mmG ; mmD=(11 13 15 17 21 23 25 27) 165 166 pxor mmH, mmH 167 168 movq mmC, mmA 169 punpcklbw mmA, mmH ; mmA=(00 02 04 06) 170 punpckhbw mmC, mmH ; mmC=(10 12 14 16) 171 172 movq mmB, mmE 173 punpcklbw mmE, mmH ; mmE=(20 22 24 26) 174 punpckhbw mmB, mmH ; mmB=(01 03 05 07) 175 176 movq mmF, mmD 177 punpcklbw mmD, mmH ; mmD=(11 13 15 17) 178 punpckhbw mmF, mmH ; mmF=(21 23 25 27) 179 180 %else ; RGB_PIXELSIZE == 4 ; ----------- 181 182 .column_ld1: 183 test cl, SIZEOF_MMWORD/8 184 jz short .column_ld2 185 sub ecx, byte SIZEOF_MMWORD/8 186 movd mmA, dword [esi+ecx*RGB_PIXELSIZE] 187 .column_ld2: 188 test cl, SIZEOF_MMWORD/4 189 jz short .column_ld4 190 sub ecx, byte SIZEOF_MMWORD/4 191 movq mmF, mmA 192 movq mmA, MMWORD [esi+ecx*RGB_PIXELSIZE] 193 .column_ld4: 194 test cl, SIZEOF_MMWORD/2 195 mov ecx, SIZEOF_MMWORD 196 jz short .rgb_gray_cnv 197 movq mmD, mmA 198 movq mmC, mmF 199 movq mmA, MMWORD [esi+0*SIZEOF_MMWORD] 200 movq mmF, MMWORD [esi+1*SIZEOF_MMWORD] 201 jmp short .rgb_gray_cnv 202 ALIGNX 16, 7 203 204 .columnloop: 205 movq mmA, MMWORD [esi+0*SIZEOF_MMWORD] 206 movq mmF, MMWORD [esi+1*SIZEOF_MMWORD] 207 movq mmD, MMWORD [esi+2*SIZEOF_MMWORD] 208 movq mmC, MMWORD [esi+3*SIZEOF_MMWORD] 209 210 .rgb_gray_cnv: 211 ; mmA=(00 10 20 30 01 11 21 31) 212 ; mmF=(02 12 22 32 03 13 23 33) 213 ; mmD=(04 14 24 34 05 15 25 35) 214 ; mmC=(06 16 26 36 07 17 27 37) 215 216 movq mmB, mmA 217 punpcklbw mmA, mmF ; mmA=(00 02 10 12 20 22 30 32) 218 punpckhbw mmB, mmF ; mmB=(01 03 11 13 21 23 31 33) 219 220 movq mmG, mmD 221 punpcklbw mmD, mmC ; mmD=(04 06 14 16 24 26 34 36) 222 punpckhbw mmG, mmC ; mmG=(05 07 15 17 25 27 35 37) 223 224 movq mmE, mmA 225 punpcklwd mmA, mmD ; mmA=(00 02 04 06 10 12 14 16) 226 punpckhwd mmE, mmD ; mmE=(20 22 24 26 30 32 34 36) 227 228 movq mmH, mmB 229 punpcklwd mmB, mmG ; mmB=(01 03 05 07 11 13 15 17) 230 punpckhwd mmH, mmG ; mmH=(21 23 25 27 31 33 35 37) 231 232 pxor mmF, mmF 233 234 movq mmC, mmA 235 punpcklbw mmA, mmF ; mmA=(00 02 04 06) 236 punpckhbw mmC, mmF ; mmC=(10 12 14 16) 237 238 movq mmD, mmB 239 punpcklbw mmB, mmF ; mmB=(01 03 05 07) 240 punpckhbw mmD, mmF ; mmD=(11 13 15 17) 241 242 movq mmG, mmE 243 punpcklbw mmE, mmF ; mmE=(20 22 24 26) 244 punpckhbw mmG, mmF ; mmG=(30 32 34 36) 245 246 punpcklbw mmF, mmH 247 punpckhbw mmH, mmH 248 psrlw mmF, BYTE_BIT ; mmF=(21 23 25 27) 249 psrlw mmH, BYTE_BIT ; mmH=(31 33 35 37) 250 251 %endif ; RGB_PIXELSIZE ; --------------- 252 253 ; mm0=(R0 R2 R4 R6)=RE, mm2=(G0 G2 G4 G6)=GE, mm4=(B0 B2 B4 B6)=BE 254 ; mm1=(R1 R3 R5 R7)=RO, mm3=(G1 G3 G5 G7)=GO, mm5=(B1 B3 B5 B7)=BO 255 256 ; (Original) 257 ; Y = 0.29900 * R + 0.58700 * G + 0.11400 * B 258 ; 259 ; (This implementation) 260 ; Y = 0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G 261 262 movq mm6, mm1 263 punpcklwd mm1, mm3 264 punpckhwd mm6, mm3 265 pmaddwd mm1, [GOTOFF(eax,PW_F0299_F0337)] ; mm1=ROL*FIX(0.299)+GOL*FIX(0.337) 266 pmaddwd mm6, [GOTOFF(eax,PW_F0299_F0337)] ; mm6=ROH*FIX(0.299)+GOH*FIX(0.337) 267 268 movq mm7, mm6 ; mm7=ROH*FIX(0.299)+GOH*FIX(0.337) 269 270 movq mm6, mm0 271 punpcklwd mm0, mm2 272 punpckhwd mm6, mm2 273 pmaddwd mm0, [GOTOFF(eax,PW_F0299_F0337)] ; mm0=REL*FIX(0.299)+GEL*FIX(0.337) 274 pmaddwd mm6, [GOTOFF(eax,PW_F0299_F0337)] ; mm6=REH*FIX(0.299)+GEH*FIX(0.337) 275 276 movq MMWORD [wk(0)], mm0 ; wk(0)=REL*FIX(0.299)+GEL*FIX(0.337) 277 movq MMWORD [wk(1)], mm6 ; wk(1)=REH*FIX(0.299)+GEH*FIX(0.337) 278 279 movq mm0, mm5 ; mm0=BO 280 movq mm6, mm4 ; mm6=BE 281 282 movq mm4, mm0 283 punpcklwd mm0, mm3 284 punpckhwd mm4, mm3 285 pmaddwd mm0, [GOTOFF(eax,PW_F0114_F0250)] ; mm0=BOL*FIX(0.114)+GOL*FIX(0.250) 286 pmaddwd mm4, [GOTOFF(eax,PW_F0114_F0250)] ; mm4=BOH*FIX(0.114)+GOH*FIX(0.250) 287 288 movq mm3, [GOTOFF(eax,PD_ONEHALF)] ; mm3=[PD_ONEHALF] 289 290 paddd mm0, mm1 291 paddd mm4, mm7 292 paddd mm0, mm3 293 paddd mm4, mm3 294 psrld mm0, SCALEBITS ; mm0=YOL 295 psrld mm4, SCALEBITS ; mm4=YOH 296 packssdw mm0, mm4 ; mm0=YO 297 298 movq mm4, mm6 299 punpcklwd mm6, mm2 300 punpckhwd mm4, mm2 301 pmaddwd mm6, [GOTOFF(eax,PW_F0114_F0250)] ; mm6=BEL*FIX(0.114)+GEL*FIX(0.250) 302 pmaddwd mm4, [GOTOFF(eax,PW_F0114_F0250)] ; mm4=BEH*FIX(0.114)+GEH*FIX(0.250) 303 304 movq mm2, [GOTOFF(eax,PD_ONEHALF)] ; mm2=[PD_ONEHALF] 305 306 paddd mm6, MMWORD [wk(0)] 307 paddd mm4, MMWORD [wk(1)] 308 paddd mm6, mm2 309 paddd mm4, mm2 310 psrld mm6, SCALEBITS ; mm6=YEL 311 psrld mm4, SCALEBITS ; mm4=YEH 312 packssdw mm6, mm4 ; mm6=YE 313 314 psllw mm0, BYTE_BIT 315 por mm6, mm0 ; mm6=Y 316 movq MMWORD [edi], mm6 ; Save Y 317 318 sub ecx, byte SIZEOF_MMWORD 319 add esi, byte RGB_PIXELSIZE*SIZEOF_MMWORD ; inptr 320 add edi, byte SIZEOF_MMWORD ; outptr0 321 cmp ecx, byte SIZEOF_MMWORD 322 jae near .columnloop 323 test ecx, ecx 324 jnz near .column_ld1 325 326 pop ecx ; col 327 pop esi 328 pop edi 329 POPPIC eax 330 331 add esi, byte SIZEOF_JSAMPROW ; input_buf 332 add edi, byte SIZEOF_JSAMPROW 333 dec eax ; num_rows 334 jg near .rowloop 335 336 emms ; empty MMX state 337 338 .return: 339 pop edi 340 pop esi 341 ; pop edx ; need not be preserved 342 ; pop ecx ; need not be preserved 343 pop ebx 344 mov esp, ebp ; esp <- aligned ebp 345 pop esp ; esp <- original ebp 346 pop ebp 347 ret 348 349 ; For some reason, the OS X linker does not honor the request to align the 350 ; segment unless we do this. 351 align 32