jdmrgext-mmx.asm (17796B)
1 ; 2 ; jdmrgext.asm - merged upsampling/color conversion (MMX) 3 ; 4 ; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB 5 ; Copyright (C) 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 ; Upsample and color convert for the case of 2:1 horizontal and 1:1 vertical. 18 ; 19 ; GLOBAL(void) 20 ; jsimd_h2v1_merged_upsample_mmx(JDIMENSION output_width, JSAMPIMAGE input_buf, 21 ; JDIMENSION in_row_group_ctr, 22 ; JSAMPARRAY output_buf); 23 ; 24 25 %define output_width(b) (b) + 8 ; JDIMENSION output_width 26 %define input_buf(b) (b) + 12 ; JSAMPIMAGE input_buf 27 %define in_row_group_ctr(b) (b) + 16 ; JDIMENSION in_row_group_ctr 28 %define output_buf(b) (b) + 20 ; JSAMPARRAY output_buf 29 30 %define original_ebp ebp + 0 31 %define wk(i) ebp - (WK_NUM - (i)) * SIZEOF_MMWORD ; mmword wk[WK_NUM] 32 %define WK_NUM 3 33 %define gotptr wk(0) - SIZEOF_POINTER ; void * gotptr 34 35 align 32 36 GLOBAL_FUNCTION(jsimd_h2v1_merged_upsample_mmx) 37 38 EXTN(jsimd_h2v1_merged_upsample_mmx): 39 push ebp 40 mov eax, esp ; eax = original ebp 41 sub esp, byte 4 42 and esp, byte (-SIZEOF_MMWORD) ; align to 64 bits 43 mov [esp], eax 44 mov ebp, esp ; ebp = aligned ebp 45 lea esp, [wk(0)] 46 PUSHPIC eax ; make a room for GOT address 47 push ebx 48 ; push ecx ; need not be preserved 49 ; push edx ; need not be preserved 50 push esi 51 push edi 52 53 GET_GOT ebx ; get GOT address 54 MOVPIC POINTER [gotptr], ebx ; save GOT address 55 56 mov ecx, JDIMENSION [output_width(eax)] ; col 57 test ecx, ecx 58 jz near .return 59 60 push ecx 61 62 mov edi, JSAMPIMAGE [input_buf(eax)] 63 mov ecx, JDIMENSION [in_row_group_ctr(eax)] 64 mov esi, JSAMPARRAY [edi+0*SIZEOF_JSAMPARRAY] 65 mov ebx, JSAMPARRAY [edi+1*SIZEOF_JSAMPARRAY] 66 mov edx, JSAMPARRAY [edi+2*SIZEOF_JSAMPARRAY] 67 mov edi, JSAMPARRAY [output_buf(eax)] 68 mov esi, JSAMPROW [esi+ecx*SIZEOF_JSAMPROW] ; inptr0 69 mov ebx, JSAMPROW [ebx+ecx*SIZEOF_JSAMPROW] ; inptr1 70 mov edx, JSAMPROW [edx+ecx*SIZEOF_JSAMPROW] ; inptr2 71 mov edi, JSAMPROW [edi] ; outptr 72 73 pop ecx ; col 74 75 ALIGNX 16, 7 76 .columnloop: 77 MOVPIC eax, POINTER [gotptr] ; load GOT address (eax) 78 79 movq mm6, MMWORD [ebx] ; mm6=Cb(01234567) 80 movq mm7, MMWORD [edx] ; mm7=Cr(01234567) 81 82 pxor mm1, mm1 ; mm1=(all 0's) 83 pcmpeqw mm3, mm3 84 psllw mm3, 7 ; mm3={0xFF80 0xFF80 0xFF80 0xFF80} 85 86 movq mm4, mm6 87 punpckhbw mm6, mm1 ; mm6=Cb(4567)=CbH 88 punpcklbw mm4, mm1 ; mm4=Cb(0123)=CbL 89 movq mm0, mm7 90 punpckhbw mm7, mm1 ; mm7=Cr(4567)=CrH 91 punpcklbw mm0, mm1 ; mm0=Cr(0123)=CrL 92 93 paddw mm6, mm3 94 paddw mm4, mm3 95 paddw mm7, mm3 96 paddw mm0, mm3 97 98 ; (Original) 99 ; R = Y + 1.40200 * Cr 100 ; G = Y - 0.34414 * Cb - 0.71414 * Cr 101 ; B = Y + 1.77200 * Cb 102 ; 103 ; (This implementation) 104 ; R = Y + 0.40200 * Cr + Cr 105 ; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr 106 ; B = Y - 0.22800 * Cb + Cb + Cb 107 108 movq mm5, mm6 ; mm5=CbH 109 movq mm2, mm4 ; mm2=CbL 110 paddw mm6, mm6 ; mm6=2*CbH 111 paddw mm4, mm4 ; mm4=2*CbL 112 movq mm1, mm7 ; mm1=CrH 113 movq mm3, mm0 ; mm3=CrL 114 paddw mm7, mm7 ; mm7=2*CrH 115 paddw mm0, mm0 ; mm0=2*CrL 116 117 pmulhw mm6, [GOTOFF(eax,PW_MF0228)] ; mm6=(2*CbH * -FIX(0.22800)) 118 pmulhw mm4, [GOTOFF(eax,PW_MF0228)] ; mm4=(2*CbL * -FIX(0.22800)) 119 pmulhw mm7, [GOTOFF(eax,PW_F0402)] ; mm7=(2*CrH * FIX(0.40200)) 120 pmulhw mm0, [GOTOFF(eax,PW_F0402)] ; mm0=(2*CrL * FIX(0.40200)) 121 122 paddw mm6, [GOTOFF(eax,PW_ONE)] 123 paddw mm4, [GOTOFF(eax,PW_ONE)] 124 psraw mm6, 1 ; mm6=(CbH * -FIX(0.22800)) 125 psraw mm4, 1 ; mm4=(CbL * -FIX(0.22800)) 126 paddw mm7, [GOTOFF(eax,PW_ONE)] 127 paddw mm0, [GOTOFF(eax,PW_ONE)] 128 psraw mm7, 1 ; mm7=(CrH * FIX(0.40200)) 129 psraw mm0, 1 ; mm0=(CrL * FIX(0.40200)) 130 131 paddw mm6, mm5 132 paddw mm4, mm2 133 paddw mm6, mm5 ; mm6=(CbH * FIX(1.77200))=(B-Y)H 134 paddw mm4, mm2 ; mm4=(CbL * FIX(1.77200))=(B-Y)L 135 paddw mm7, mm1 ; mm7=(CrH * FIX(1.40200))=(R-Y)H 136 paddw mm0, mm3 ; mm0=(CrL * FIX(1.40200))=(R-Y)L 137 138 movq MMWORD [wk(0)], mm6 ; wk(0)=(B-Y)H 139 movq MMWORD [wk(1)], mm7 ; wk(1)=(R-Y)H 140 141 movq mm6, mm5 142 movq mm7, mm2 143 punpcklwd mm5, mm1 144 punpckhwd mm6, mm1 145 pmaddwd mm5, [GOTOFF(eax,PW_MF0344_F0285)] 146 pmaddwd mm6, [GOTOFF(eax,PW_MF0344_F0285)] 147 punpcklwd mm2, mm3 148 punpckhwd mm7, mm3 149 pmaddwd mm2, [GOTOFF(eax,PW_MF0344_F0285)] 150 pmaddwd mm7, [GOTOFF(eax,PW_MF0344_F0285)] 151 152 paddd mm5, [GOTOFF(eax,PD_ONEHALF)] 153 paddd mm6, [GOTOFF(eax,PD_ONEHALF)] 154 psrad mm5, SCALEBITS 155 psrad mm6, SCALEBITS 156 paddd mm2, [GOTOFF(eax,PD_ONEHALF)] 157 paddd mm7, [GOTOFF(eax,PD_ONEHALF)] 158 psrad mm2, SCALEBITS 159 psrad mm7, SCALEBITS 160 161 packssdw mm5, mm6 ; mm5=CbH*-FIX(0.344)+CrH*FIX(0.285) 162 packssdw mm2, mm7 ; mm2=CbL*-FIX(0.344)+CrL*FIX(0.285) 163 psubw mm5, mm1 ; mm5=CbH*-FIX(0.344)+CrH*-FIX(0.714)=(G-Y)H 164 psubw mm2, mm3 ; mm2=CbL*-FIX(0.344)+CrL*-FIX(0.714)=(G-Y)L 165 166 movq MMWORD [wk(2)], mm5 ; wk(2)=(G-Y)H 167 168 mov al, 2 ; Yctr 169 jmp short .Yloop_1st 170 ALIGNX 16, 7 171 172 .Yloop_2nd: 173 movq mm0, MMWORD [wk(1)] ; mm0=(R-Y)H 174 movq mm2, MMWORD [wk(2)] ; mm2=(G-Y)H 175 movq mm4, MMWORD [wk(0)] ; mm4=(B-Y)H 176 ALIGNX 16, 7 177 178 .Yloop_1st: 179 movq mm7, MMWORD [esi] ; mm7=Y(01234567) 180 181 pcmpeqw mm6, mm6 182 psrlw mm6, BYTE_BIT ; mm6={0xFF 0x00 0xFF 0x00 ..} 183 pand mm6, mm7 ; mm6=Y(0246)=YE 184 psrlw mm7, BYTE_BIT ; mm7=Y(1357)=YO 185 186 movq mm1, mm0 ; mm1=mm0=(R-Y)(L/H) 187 movq mm3, mm2 ; mm3=mm2=(G-Y)(L/H) 188 movq mm5, mm4 ; mm5=mm4=(B-Y)(L/H) 189 190 paddw mm0, mm6 ; mm0=((R-Y)+YE)=RE=(R0 R2 R4 R6) 191 paddw mm1, mm7 ; mm1=((R-Y)+YO)=RO=(R1 R3 R5 R7) 192 packuswb mm0, mm0 ; mm0=(R0 R2 R4 R6 ** ** ** **) 193 packuswb mm1, mm1 ; mm1=(R1 R3 R5 R7 ** ** ** **) 194 195 paddw mm2, mm6 ; mm2=((G-Y)+YE)=GE=(G0 G2 G4 G6) 196 paddw mm3, mm7 ; mm3=((G-Y)+YO)=GO=(G1 G3 G5 G7) 197 packuswb mm2, mm2 ; mm2=(G0 G2 G4 G6 ** ** ** **) 198 packuswb mm3, mm3 ; mm3=(G1 G3 G5 G7 ** ** ** **) 199 200 paddw mm4, mm6 ; mm4=((B-Y)+YE)=BE=(B0 B2 B4 B6) 201 paddw mm5, mm7 ; mm5=((B-Y)+YO)=BO=(B1 B3 B5 B7) 202 packuswb mm4, mm4 ; mm4=(B0 B2 B4 B6 ** ** ** **) 203 packuswb mm5, mm5 ; mm5=(B1 B3 B5 B7 ** ** ** **) 204 205 %if RGB_PIXELSIZE == 3 ; --------------- 206 207 ; mmA=(00 02 04 06 ** ** ** **), mmB=(01 03 05 07 ** ** ** **) 208 ; mmC=(10 12 14 16 ** ** ** **), mmD=(11 13 15 17 ** ** ** **) 209 ; mmE=(20 22 24 26 ** ** ** **), mmF=(21 23 25 27 ** ** ** **) 210 ; mmG=(** ** ** ** ** ** ** **), mmH=(** ** ** ** ** ** ** **) 211 212 punpcklbw mmA, mmC ; mmA=(00 10 02 12 04 14 06 16) 213 punpcklbw mmE, mmB ; mmE=(20 01 22 03 24 05 26 07) 214 punpcklbw mmD, mmF ; mmD=(11 21 13 23 15 25 17 27) 215 216 movq mmG, mmA 217 movq mmH, mmA 218 punpcklwd mmA, mmE ; mmA=(00 10 20 01 02 12 22 03) 219 punpckhwd mmG, mmE ; mmG=(04 14 24 05 06 16 26 07) 220 221 psrlq mmH, 2*BYTE_BIT ; mmH=(02 12 04 14 06 16 -- --) 222 psrlq mmE, 2*BYTE_BIT ; mmE=(22 03 24 05 26 07 -- --) 223 224 movq mmC, mmD 225 movq mmB, mmD 226 punpcklwd mmD, mmH ; mmD=(11 21 02 12 13 23 04 14) 227 punpckhwd mmC, mmH ; mmC=(15 25 06 16 17 27 -- --) 228 229 psrlq mmB, 2*BYTE_BIT ; mmB=(13 23 15 25 17 27 -- --) 230 231 movq mmF, mmE 232 punpcklwd mmE, mmB ; mmE=(22 03 13 23 24 05 15 25) 233 punpckhwd mmF, mmB ; mmF=(26 07 17 27 -- -- -- --) 234 235 punpckldq mmA, mmD ; mmA=(00 10 20 01 11 21 02 12) 236 punpckldq mmE, mmG ; mmE=(22 03 13 23 04 14 24 05) 237 punpckldq mmC, mmF ; mmC=(15 25 06 16 26 07 17 27) 238 239 cmp ecx, byte SIZEOF_MMWORD 240 jb short .column_st16 241 242 movq MMWORD [edi+0*SIZEOF_MMWORD], mmA 243 movq MMWORD [edi+1*SIZEOF_MMWORD], mmE 244 movq MMWORD [edi+2*SIZEOF_MMWORD], mmC 245 246 sub ecx, byte SIZEOF_MMWORD 247 jz near .endcolumn 248 249 add edi, byte RGB_PIXELSIZE*SIZEOF_MMWORD ; outptr 250 add esi, byte SIZEOF_MMWORD ; inptr0 251 dec al ; Yctr 252 jnz near .Yloop_2nd 253 254 add ebx, byte SIZEOF_MMWORD ; inptr1 255 add edx, byte SIZEOF_MMWORD ; inptr2 256 jmp near .columnloop 257 ALIGNX 16, 7 258 259 .column_st16: 260 lea ecx, [ecx+ecx*2] ; imul ecx, RGB_PIXELSIZE 261 cmp ecx, byte 2*SIZEOF_MMWORD 262 jb short .column_st8 263 movq MMWORD [edi+0*SIZEOF_MMWORD], mmA 264 movq MMWORD [edi+1*SIZEOF_MMWORD], mmE 265 movq mmA, mmC 266 sub ecx, byte 2*SIZEOF_MMWORD 267 add edi, byte 2*SIZEOF_MMWORD 268 jmp short .column_st4 269 .column_st8: 270 cmp ecx, byte SIZEOF_MMWORD 271 jb short .column_st4 272 movq MMWORD [edi+0*SIZEOF_MMWORD], mmA 273 movq mmA, mmE 274 sub ecx, byte SIZEOF_MMWORD 275 add edi, byte SIZEOF_MMWORD 276 .column_st4: 277 movd eax, mmA 278 cmp ecx, byte SIZEOF_DWORD 279 jb short .column_st2 280 mov dword [edi+0*SIZEOF_DWORD], eax 281 psrlq mmA, DWORD_BIT 282 movd eax, mmA 283 sub ecx, byte SIZEOF_DWORD 284 add edi, byte SIZEOF_DWORD 285 .column_st2: 286 cmp ecx, byte SIZEOF_WORD 287 jb short .column_st1 288 mov word [edi+0*SIZEOF_WORD], ax 289 shr eax, WORD_BIT 290 sub ecx, byte SIZEOF_WORD 291 add edi, byte SIZEOF_WORD 292 .column_st1: 293 cmp ecx, byte SIZEOF_BYTE 294 jb short .endcolumn 295 mov byte [edi+0*SIZEOF_BYTE], al 296 297 %else ; RGB_PIXELSIZE == 4 ; ----------- 298 299 %ifdef RGBX_FILLER_0XFF 300 pcmpeqb mm6, mm6 ; mm6=(X0 X2 X4 X6 ** ** ** **) 301 pcmpeqb mm7, mm7 ; mm7=(X1 X3 X5 X7 ** ** ** **) 302 %else 303 pxor mm6, mm6 ; mm6=(X0 X2 X4 X6 ** ** ** **) 304 pxor mm7, mm7 ; mm7=(X1 X3 X5 X7 ** ** ** **) 305 %endif 306 ; mmA=(00 02 04 06 ** ** ** **), mmB=(01 03 05 07 ** ** ** **) 307 ; mmC=(10 12 14 16 ** ** ** **), mmD=(11 13 15 17 ** ** ** **) 308 ; mmE=(20 22 24 26 ** ** ** **), mmF=(21 23 25 27 ** ** ** **) 309 ; mmG=(30 32 34 36 ** ** ** **), mmH=(31 33 35 37 ** ** ** **) 310 311 punpcklbw mmA, mmC ; mmA=(00 10 02 12 04 14 06 16) 312 punpcklbw mmE, mmG ; mmE=(20 30 22 32 24 34 26 36) 313 punpcklbw mmB, mmD ; mmB=(01 11 03 13 05 15 07 17) 314 punpcklbw mmF, mmH ; mmF=(21 31 23 33 25 35 27 37) 315 316 movq mmC, mmA 317 punpcklwd mmA, mmE ; mmA=(00 10 20 30 02 12 22 32) 318 punpckhwd mmC, mmE ; mmC=(04 14 24 34 06 16 26 36) 319 movq mmG, mmB 320 punpcklwd mmB, mmF ; mmB=(01 11 21 31 03 13 23 33) 321 punpckhwd mmG, mmF ; mmG=(05 15 25 35 07 17 27 37) 322 323 movq mmD, mmA 324 punpckldq mmA, mmB ; mmA=(00 10 20 30 01 11 21 31) 325 punpckhdq mmD, mmB ; mmD=(02 12 22 32 03 13 23 33) 326 movq mmH, mmC 327 punpckldq mmC, mmG ; mmC=(04 14 24 34 05 15 25 35) 328 punpckhdq mmH, mmG ; mmH=(06 16 26 36 07 17 27 37) 329 330 cmp ecx, byte SIZEOF_MMWORD 331 jb short .column_st16 332 333 movq MMWORD [edi+0*SIZEOF_MMWORD], mmA 334 movq MMWORD [edi+1*SIZEOF_MMWORD], mmD 335 movq MMWORD [edi+2*SIZEOF_MMWORD], mmC 336 movq MMWORD [edi+3*SIZEOF_MMWORD], mmH 337 338 sub ecx, byte SIZEOF_MMWORD 339 jz short .endcolumn 340 341 add edi, byte RGB_PIXELSIZE*SIZEOF_MMWORD ; outptr 342 add esi, byte SIZEOF_MMWORD ; inptr0 343 dec al ; Yctr 344 jnz near .Yloop_2nd 345 346 add ebx, byte SIZEOF_MMWORD ; inptr1 347 add edx, byte SIZEOF_MMWORD ; inptr2 348 jmp near .columnloop 349 ALIGNX 16, 7 350 351 .column_st16: 352 cmp ecx, byte SIZEOF_MMWORD/2 353 jb short .column_st8 354 movq MMWORD [edi+0*SIZEOF_MMWORD], mmA 355 movq MMWORD [edi+1*SIZEOF_MMWORD], mmD 356 movq mmA, mmC 357 movq mmD, mmH 358 sub ecx, byte SIZEOF_MMWORD/2 359 add edi, byte 2*SIZEOF_MMWORD 360 .column_st8: 361 cmp ecx, byte SIZEOF_MMWORD/4 362 jb short .column_st4 363 movq MMWORD [edi+0*SIZEOF_MMWORD], mmA 364 movq mmA, mmD 365 sub ecx, byte SIZEOF_MMWORD/4 366 add edi, byte 1*SIZEOF_MMWORD 367 .column_st4: 368 cmp ecx, byte SIZEOF_MMWORD/8 369 jb short .endcolumn 370 movd dword [edi+0*SIZEOF_DWORD], mmA 371 372 %endif ; RGB_PIXELSIZE ; --------------- 373 374 .endcolumn: 375 emms ; empty MMX state 376 377 .return: 378 pop edi 379 pop esi 380 ; pop edx ; need not be preserved 381 ; pop ecx ; need not be preserved 382 pop ebx 383 mov esp, ebp ; esp <- aligned ebp 384 pop esp ; esp <- original ebp 385 pop ebp 386 ret 387 388 ; -------------------------------------------------------------------------- 389 ; 390 ; Upsample and color convert for the case of 2:1 horizontal and 2:1 vertical. 391 ; 392 ; GLOBAL(void) 393 ; jsimd_h2v2_merged_upsample_mmx(JDIMENSION output_width, JSAMPIMAGE input_buf, 394 ; JDIMENSION in_row_group_ctr, 395 ; JSAMPARRAY output_buf); 396 ; 397 398 %define output_width(b) (b) + 8 ; JDIMENSION output_width 399 %define input_buf(b) (b) + 12 ; JSAMPIMAGE input_buf 400 %define in_row_group_ctr(b) (b) + 16 ; JDIMENSION in_row_group_ctr 401 %define output_buf(b) (b) + 20 ; JSAMPARRAY output_buf 402 403 align 32 404 GLOBAL_FUNCTION(jsimd_h2v2_merged_upsample_mmx) 405 406 EXTN(jsimd_h2v2_merged_upsample_mmx): 407 push ebp 408 mov ebp, esp 409 push ebx 410 ; push ecx ; need not be preserved 411 ; push edx ; need not be preserved 412 push esi 413 push edi 414 415 mov eax, JDIMENSION [output_width(ebp)] 416 417 mov edi, JSAMPIMAGE [input_buf(ebp)] 418 mov ecx, JDIMENSION [in_row_group_ctr(ebp)] 419 mov esi, JSAMPARRAY [edi+0*SIZEOF_JSAMPARRAY] 420 mov ebx, JSAMPARRAY [edi+1*SIZEOF_JSAMPARRAY] 421 mov edx, JSAMPARRAY [edi+2*SIZEOF_JSAMPARRAY] 422 mov edi, JSAMPARRAY [output_buf(ebp)] 423 lea esi, [esi+ecx*SIZEOF_JSAMPROW] 424 425 push edx ; inptr2 426 push ebx ; inptr1 427 push esi ; inptr00 428 mov ebx, esp 429 430 push edi ; output_buf (outptr0) 431 push ecx ; in_row_group_ctr 432 push ebx ; input_buf 433 push eax ; output_width 434 435 call near EXTN(jsimd_h2v1_merged_upsample_mmx) 436 437 add esi, byte SIZEOF_JSAMPROW ; inptr01 438 add edi, byte SIZEOF_JSAMPROW ; outptr1 439 mov POINTER [ebx+0*SIZEOF_POINTER], esi 440 mov POINTER [ebx-1*SIZEOF_POINTER], edi 441 442 call near EXTN(jsimd_h2v1_merged_upsample_mmx) 443 444 add esp, byte 7*SIZEOF_DWORD 445 446 pop edi 447 pop esi 448 ; pop edx ; need not be preserved 449 ; pop ecx ; need not be preserved 450 pop ebx 451 pop ebp 452 ret 453 454 ; For some reason, the OS X linker does not honor the request to align the 455 ; segment unless we do this. 456 align 32