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, 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; Upsample and color convert for the case of 2:1 horizontal and 1:1 vertical. 24; 25; GLOBAL(void) 26; jsimd_h2v1_merged_upsample_mmx(JDIMENSION output_width, JSAMPIMAGE input_buf, 27; JDIMENSION in_row_group_ctr, 28; JSAMPARRAY output_buf); 29; 30 31%define output_width(b) (b) + 8 ; JDIMENSION output_width 32%define input_buf(b) (b) + 12 ; JSAMPIMAGE input_buf 33%define in_row_group_ctr(b) (b) + 16 ; JDIMENSION in_row_group_ctr 34%define output_buf(b) (b) + 20 ; JSAMPARRAY output_buf 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 3 39%define gotptr wk(0) - SIZEOF_POINTER ; void * gotptr 40 41 align 32 42 GLOBAL_FUNCTION(jsimd_h2v1_merged_upsample_mmx) 43 44EXTN(jsimd_h2v1_merged_upsample_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 [output_width(eax)] ; col 63 test ecx, ecx 64 jz near .return 65 66 push ecx 67 68 mov edi, JSAMPIMAGE [input_buf(eax)] 69 mov ecx, JDIMENSION [in_row_group_ctr(eax)] 70 mov esi, JSAMPARRAY [edi+0*SIZEOF_JSAMPARRAY] 71 mov ebx, JSAMPARRAY [edi+1*SIZEOF_JSAMPARRAY] 72 mov edx, JSAMPARRAY [edi+2*SIZEOF_JSAMPARRAY] 73 mov edi, JSAMPARRAY [output_buf(eax)] 74 mov esi, JSAMPROW [esi+ecx*SIZEOF_JSAMPROW] ; inptr0 75 mov ebx, JSAMPROW [ebx+ecx*SIZEOF_JSAMPROW] ; inptr1 76 mov edx, JSAMPROW [edx+ecx*SIZEOF_JSAMPROW] ; inptr2 77 mov edi, JSAMPROW [edi] ; outptr 78 79 pop ecx ; col 80 81 alignx 16, 7 82.columnloop: 83 movpic eax, POINTER [gotptr] ; load GOT address (eax) 84 85 movq mm6, MMWORD [ebx] ; mm6=Cb(01234567) 86 movq mm7, MMWORD [edx] ; mm7=Cr(01234567) 87 88 pxor mm1, mm1 ; mm1=(all 0's) 89 pcmpeqw mm3, mm3 90 psllw mm3, 7 ; mm3={0xFF80 0xFF80 0xFF80 0xFF80} 91 92 movq mm4, mm6 93 punpckhbw mm6, mm1 ; mm6=Cb(4567)=CbH 94 punpcklbw mm4, mm1 ; mm4=Cb(0123)=CbL 95 movq mm0, mm7 96 punpckhbw mm7, mm1 ; mm7=Cr(4567)=CrH 97 punpcklbw mm0, mm1 ; mm0=Cr(0123)=CrL 98 99 paddw mm6, mm3 100 paddw mm4, mm3 101 paddw mm7, mm3 102 paddw mm0, mm3 103 104 ; (Original) 105 ; R = Y + 1.40200 * Cr 106 ; G = Y - 0.34414 * Cb - 0.71414 * Cr 107 ; B = Y + 1.77200 * Cb 108 ; 109 ; (This implementation) 110 ; R = Y + 0.40200 * Cr + Cr 111 ; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr 112 ; B = Y - 0.22800 * Cb + Cb + Cb 113 114 movq mm5, mm6 ; mm5=CbH 115 movq mm2, mm4 ; mm2=CbL 116 paddw mm6, mm6 ; mm6=2*CbH 117 paddw mm4, mm4 ; mm4=2*CbL 118 movq mm1, mm7 ; mm1=CrH 119 movq mm3, mm0 ; mm3=CrL 120 paddw mm7, mm7 ; mm7=2*CrH 121 paddw mm0, mm0 ; mm0=2*CrL 122 123 pmulhw mm6, [GOTOFF(eax,PW_MF0228)] ; mm6=(2*CbH * -FIX(0.22800)) 124 pmulhw mm4, [GOTOFF(eax,PW_MF0228)] ; mm4=(2*CbL * -FIX(0.22800)) 125 pmulhw mm7, [GOTOFF(eax,PW_F0402)] ; mm7=(2*CrH * FIX(0.40200)) 126 pmulhw mm0, [GOTOFF(eax,PW_F0402)] ; mm0=(2*CrL * FIX(0.40200)) 127 128 paddw mm6, [GOTOFF(eax,PW_ONE)] 129 paddw mm4, [GOTOFF(eax,PW_ONE)] 130 psraw mm6, 1 ; mm6=(CbH * -FIX(0.22800)) 131 psraw mm4, 1 ; mm4=(CbL * -FIX(0.22800)) 132 paddw mm7, [GOTOFF(eax,PW_ONE)] 133 paddw mm0, [GOTOFF(eax,PW_ONE)] 134 psraw mm7, 1 ; mm7=(CrH * FIX(0.40200)) 135 psraw mm0, 1 ; mm0=(CrL * FIX(0.40200)) 136 137 paddw mm6, mm5 138 paddw mm4, mm2 139 paddw mm6, mm5 ; mm6=(CbH * FIX(1.77200))=(B-Y)H 140 paddw mm4, mm2 ; mm4=(CbL * FIX(1.77200))=(B-Y)L 141 paddw mm7, mm1 ; mm7=(CrH * FIX(1.40200))=(R-Y)H 142 paddw mm0, mm3 ; mm0=(CrL * FIX(1.40200))=(R-Y)L 143 144 movq MMWORD [wk(0)], mm6 ; wk(0)=(B-Y)H 145 movq MMWORD [wk(1)], mm7 ; wk(1)=(R-Y)H 146 147 movq mm6, mm5 148 movq mm7, mm2 149 punpcklwd mm5, mm1 150 punpckhwd mm6, mm1 151 pmaddwd mm5, [GOTOFF(eax,PW_MF0344_F0285)] 152 pmaddwd mm6, [GOTOFF(eax,PW_MF0344_F0285)] 153 punpcklwd mm2, mm3 154 punpckhwd mm7, mm3 155 pmaddwd mm2, [GOTOFF(eax,PW_MF0344_F0285)] 156 pmaddwd mm7, [GOTOFF(eax,PW_MF0344_F0285)] 157 158 paddd mm5, [GOTOFF(eax,PD_ONEHALF)] 159 paddd mm6, [GOTOFF(eax,PD_ONEHALF)] 160 psrad mm5, SCALEBITS 161 psrad mm6, SCALEBITS 162 paddd mm2, [GOTOFF(eax,PD_ONEHALF)] 163 paddd mm7, [GOTOFF(eax,PD_ONEHALF)] 164 psrad mm2, SCALEBITS 165 psrad mm7, SCALEBITS 166 167 packssdw mm5, mm6 ; mm5=CbH*-FIX(0.344)+CrH*FIX(0.285) 168 packssdw mm2, mm7 ; mm2=CbL*-FIX(0.344)+CrL*FIX(0.285) 169 psubw mm5, mm1 ; mm5=CbH*-FIX(0.344)+CrH*-FIX(0.714)=(G-Y)H 170 psubw mm2, mm3 ; mm2=CbL*-FIX(0.344)+CrL*-FIX(0.714)=(G-Y)L 171 172 movq MMWORD [wk(2)], mm5 ; wk(2)=(G-Y)H 173 174 mov al, 2 ; Yctr 175 jmp short .Yloop_1st 176 alignx 16, 7 177 178.Yloop_2nd: 179 movq mm0, MMWORD [wk(1)] ; mm0=(R-Y)H 180 movq mm2, MMWORD [wk(2)] ; mm2=(G-Y)H 181 movq mm4, MMWORD [wk(0)] ; mm4=(B-Y)H 182 alignx 16, 7 183 184.Yloop_1st: 185 movq mm7, MMWORD [esi] ; mm7=Y(01234567) 186 187 pcmpeqw mm6, mm6 188 psrlw mm6, BYTE_BIT ; mm6={0xFF 0x00 0xFF 0x00 ..} 189 pand mm6, mm7 ; mm6=Y(0246)=YE 190 psrlw mm7, BYTE_BIT ; mm7=Y(1357)=YO 191 192 movq mm1, mm0 ; mm1=mm0=(R-Y)(L/H) 193 movq mm3, mm2 ; mm3=mm2=(G-Y)(L/H) 194 movq mm5, mm4 ; mm5=mm4=(B-Y)(L/H) 195 196 paddw mm0, mm6 ; mm0=((R-Y)+YE)=RE=(R0 R2 R4 R6) 197 paddw mm1, mm7 ; mm1=((R-Y)+YO)=RO=(R1 R3 R5 R7) 198 packuswb mm0, mm0 ; mm0=(R0 R2 R4 R6 ** ** ** **) 199 packuswb mm1, mm1 ; mm1=(R1 R3 R5 R7 ** ** ** **) 200 201 paddw mm2, mm6 ; mm2=((G-Y)+YE)=GE=(G0 G2 G4 G6) 202 paddw mm3, mm7 ; mm3=((G-Y)+YO)=GO=(G1 G3 G5 G7) 203 packuswb mm2, mm2 ; mm2=(G0 G2 G4 G6 ** ** ** **) 204 packuswb mm3, mm3 ; mm3=(G1 G3 G5 G7 ** ** ** **) 205 206 paddw mm4, mm6 ; mm4=((B-Y)+YE)=BE=(B0 B2 B4 B6) 207 paddw mm5, mm7 ; mm5=((B-Y)+YO)=BO=(B1 B3 B5 B7) 208 packuswb mm4, mm4 ; mm4=(B0 B2 B4 B6 ** ** ** **) 209 packuswb mm5, mm5 ; mm5=(B1 B3 B5 B7 ** ** ** **) 210 211%if RGB_PIXELSIZE == 3 ; --------------- 212 213 ; mmA=(00 02 04 06 ** ** ** **), mmB=(01 03 05 07 ** ** ** **) 214 ; mmC=(10 12 14 16 ** ** ** **), mmD=(11 13 15 17 ** ** ** **) 215 ; mmE=(20 22 24 26 ** ** ** **), mmF=(21 23 25 27 ** ** ** **) 216 ; mmG=(** ** ** ** ** ** ** **), mmH=(** ** ** ** ** ** ** **) 217 218 punpcklbw mmA, mmC ; mmA=(00 10 02 12 04 14 06 16) 219 punpcklbw mmE, mmB ; mmE=(20 01 22 03 24 05 26 07) 220 punpcklbw mmD, mmF ; mmD=(11 21 13 23 15 25 17 27) 221 222 movq mmG, mmA 223 movq mmH, mmA 224 punpcklwd mmA, mmE ; mmA=(00 10 20 01 02 12 22 03) 225 punpckhwd mmG, mmE ; mmG=(04 14 24 05 06 16 26 07) 226 227 psrlq mmH, 2*BYTE_BIT ; mmH=(02 12 04 14 06 16 -- --) 228 psrlq mmE, 2*BYTE_BIT ; mmE=(22 03 24 05 26 07 -- --) 229 230 movq mmC, mmD 231 movq mmB, mmD 232 punpcklwd mmD, mmH ; mmD=(11 21 02 12 13 23 04 14) 233 punpckhwd mmC, mmH ; mmC=(15 25 06 16 17 27 -- --) 234 235 psrlq mmB, 2*BYTE_BIT ; mmB=(13 23 15 25 17 27 -- --) 236 237 movq mmF, mmE 238 punpcklwd mmE, mmB ; mmE=(22 03 13 23 24 05 15 25) 239 punpckhwd mmF, mmB ; mmF=(26 07 17 27 -- -- -- --) 240 241 punpckldq mmA, mmD ; mmA=(00 10 20 01 11 21 02 12) 242 punpckldq mmE, mmG ; mmE=(22 03 13 23 04 14 24 05) 243 punpckldq mmC, mmF ; mmC=(15 25 06 16 26 07 17 27) 244 245 cmp ecx, byte SIZEOF_MMWORD 246 jb short .column_st16 247 248 movq MMWORD [edi+0*SIZEOF_MMWORD], mmA 249 movq MMWORD [edi+1*SIZEOF_MMWORD], mmE 250 movq MMWORD [edi+2*SIZEOF_MMWORD], mmC 251 252 sub ecx, byte SIZEOF_MMWORD 253 jz near .endcolumn 254 255 add edi, byte RGB_PIXELSIZE*SIZEOF_MMWORD ; outptr 256 add esi, byte SIZEOF_MMWORD ; inptr0 257 dec al ; Yctr 258 jnz near .Yloop_2nd 259 260 add ebx, byte SIZEOF_MMWORD ; inptr1 261 add edx, byte SIZEOF_MMWORD ; inptr2 262 jmp near .columnloop 263 alignx 16, 7 264 265.column_st16: 266 lea ecx, [ecx+ecx*2] ; imul ecx, RGB_PIXELSIZE 267 cmp ecx, byte 2*SIZEOF_MMWORD 268 jb short .column_st8 269 movq MMWORD [edi+0*SIZEOF_MMWORD], mmA 270 movq MMWORD [edi+1*SIZEOF_MMWORD], mmE 271 movq mmA, mmC 272 sub ecx, byte 2*SIZEOF_MMWORD 273 add edi, byte 2*SIZEOF_MMWORD 274 jmp short .column_st4 275.column_st8: 276 cmp ecx, byte SIZEOF_MMWORD 277 jb short .column_st4 278 movq MMWORD [edi+0*SIZEOF_MMWORD], mmA 279 movq mmA, mmE 280 sub ecx, byte SIZEOF_MMWORD 281 add edi, byte SIZEOF_MMWORD 282.column_st4: 283 movd eax, mmA 284 cmp ecx, byte SIZEOF_DWORD 285 jb short .column_st2 286 mov DWORD [edi+0*SIZEOF_DWORD], eax 287 psrlq mmA, DWORD_BIT 288 movd eax, mmA 289 sub ecx, byte SIZEOF_DWORD 290 add edi, byte SIZEOF_DWORD 291.column_st2: 292 cmp ecx, byte SIZEOF_WORD 293 jb short .column_st1 294 mov WORD [edi+0*SIZEOF_WORD], ax 295 shr eax, WORD_BIT 296 sub ecx, byte SIZEOF_WORD 297 add edi, byte SIZEOF_WORD 298.column_st1: 299 cmp ecx, byte SIZEOF_BYTE 300 jb short .endcolumn 301 mov BYTE [edi+0*SIZEOF_BYTE], al 302 303%else ; RGB_PIXELSIZE == 4 ; ----------- 304 305%ifdef RGBX_FILLER_0XFF 306 pcmpeqb mm6, mm6 ; mm6=(X0 X2 X4 X6 ** ** ** **) 307 pcmpeqb mm7, mm7 ; mm7=(X1 X3 X5 X7 ** ** ** **) 308%else 309 pxor mm6, mm6 ; mm6=(X0 X2 X4 X6 ** ** ** **) 310 pxor mm7, mm7 ; mm7=(X1 X3 X5 X7 ** ** ** **) 311%endif 312 ; mmA=(00 02 04 06 ** ** ** **), mmB=(01 03 05 07 ** ** ** **) 313 ; mmC=(10 12 14 16 ** ** ** **), mmD=(11 13 15 17 ** ** ** **) 314 ; mmE=(20 22 24 26 ** ** ** **), mmF=(21 23 25 27 ** ** ** **) 315 ; mmG=(30 32 34 36 ** ** ** **), mmH=(31 33 35 37 ** ** ** **) 316 317 punpcklbw mmA, mmC ; mmA=(00 10 02 12 04 14 06 16) 318 punpcklbw mmE, mmG ; mmE=(20 30 22 32 24 34 26 36) 319 punpcklbw mmB, mmD ; mmB=(01 11 03 13 05 15 07 17) 320 punpcklbw mmF, mmH ; mmF=(21 31 23 33 25 35 27 37) 321 322 movq mmC, mmA 323 punpcklwd mmA, mmE ; mmA=(00 10 20 30 02 12 22 32) 324 punpckhwd mmC, mmE ; mmC=(04 14 24 34 06 16 26 36) 325 movq mmG, mmB 326 punpcklwd mmB, mmF ; mmB=(01 11 21 31 03 13 23 33) 327 punpckhwd mmG, mmF ; mmG=(05 15 25 35 07 17 27 37) 328 329 movq mmD, mmA 330 punpckldq mmA, mmB ; mmA=(00 10 20 30 01 11 21 31) 331 punpckhdq mmD, mmB ; mmD=(02 12 22 32 03 13 23 33) 332 movq mmH, mmC 333 punpckldq mmC, mmG ; mmC=(04 14 24 34 05 15 25 35) 334 punpckhdq mmH, mmG ; mmH=(06 16 26 36 07 17 27 37) 335 336 cmp ecx, byte SIZEOF_MMWORD 337 jb short .column_st16 338 339 movq MMWORD [edi+0*SIZEOF_MMWORD], mmA 340 movq MMWORD [edi+1*SIZEOF_MMWORD], mmD 341 movq MMWORD [edi+2*SIZEOF_MMWORD], mmC 342 movq MMWORD [edi+3*SIZEOF_MMWORD], mmH 343 344 sub ecx, byte SIZEOF_MMWORD 345 jz short .endcolumn 346 347 add edi, byte RGB_PIXELSIZE*SIZEOF_MMWORD ; outptr 348 add esi, byte SIZEOF_MMWORD ; inptr0 349 dec al ; Yctr 350 jnz near .Yloop_2nd 351 352 add ebx, byte SIZEOF_MMWORD ; inptr1 353 add edx, byte SIZEOF_MMWORD ; inptr2 354 jmp near .columnloop 355 alignx 16, 7 356 357.column_st16: 358 cmp ecx, byte SIZEOF_MMWORD/2 359 jb short .column_st8 360 movq MMWORD [edi+0*SIZEOF_MMWORD], mmA 361 movq MMWORD [edi+1*SIZEOF_MMWORD], mmD 362 movq mmA, mmC 363 movq mmD, mmH 364 sub ecx, byte SIZEOF_MMWORD/2 365 add edi, byte 2*SIZEOF_MMWORD 366.column_st8: 367 cmp ecx, byte SIZEOF_MMWORD/4 368 jb short .column_st4 369 movq MMWORD [edi+0*SIZEOF_MMWORD], mmA 370 movq mmA, mmD 371 sub ecx, byte SIZEOF_MMWORD/4 372 add edi, byte 1*SIZEOF_MMWORD 373.column_st4: 374 cmp ecx, byte SIZEOF_MMWORD/8 375 jb short .endcolumn 376 movd DWORD [edi+0*SIZEOF_DWORD], mmA 377 378%endif ; RGB_PIXELSIZE ; --------------- 379 380.endcolumn: 381 emms ; empty MMX state 382 383.return: 384 pop edi 385 pop esi 386; pop edx ; need not be preserved 387; pop ecx ; need not be preserved 388 pop ebx 389 mov esp, ebp ; esp <- aligned ebp 390 pop esp ; esp <- original ebp 391 pop ebp 392 ret 393 394; -------------------------------------------------------------------------- 395; 396; Upsample and color convert for the case of 2:1 horizontal and 2:1 vertical. 397; 398; GLOBAL(void) 399; jsimd_h2v2_merged_upsample_mmx(JDIMENSION output_width, JSAMPIMAGE input_buf, 400; JDIMENSION in_row_group_ctr, 401; JSAMPARRAY output_buf); 402; 403 404%define output_width(b) (b) + 8 ; JDIMENSION output_width 405%define input_buf(b) (b) + 12 ; JSAMPIMAGE input_buf 406%define in_row_group_ctr(b) (b) + 16 ; JDIMENSION in_row_group_ctr 407%define output_buf(b) (b) + 20 ; JSAMPARRAY output_buf 408 409 align 32 410 GLOBAL_FUNCTION(jsimd_h2v2_merged_upsample_mmx) 411 412EXTN(jsimd_h2v2_merged_upsample_mmx): 413 push ebp 414 mov ebp, esp 415 push ebx 416; push ecx ; need not be preserved 417; push edx ; need not be preserved 418 push esi 419 push edi 420 421 mov eax, JDIMENSION [output_width(ebp)] 422 423 mov edi, JSAMPIMAGE [input_buf(ebp)] 424 mov ecx, JDIMENSION [in_row_group_ctr(ebp)] 425 mov esi, JSAMPARRAY [edi+0*SIZEOF_JSAMPARRAY] 426 mov ebx, JSAMPARRAY [edi+1*SIZEOF_JSAMPARRAY] 427 mov edx, JSAMPARRAY [edi+2*SIZEOF_JSAMPARRAY] 428 mov edi, JSAMPARRAY [output_buf(ebp)] 429 lea esi, [esi+ecx*SIZEOF_JSAMPROW] 430 431 push edx ; inptr2 432 push ebx ; inptr1 433 push esi ; inptr00 434 mov ebx, esp 435 436 push edi ; output_buf (outptr0) 437 push ecx ; in_row_group_ctr 438 push ebx ; input_buf 439 push eax ; output_width 440 441 call near EXTN(jsimd_h2v1_merged_upsample_mmx) 442 443 add esi, byte SIZEOF_JSAMPROW ; inptr01 444 add edi, byte SIZEOF_JSAMPROW ; outptr1 445 mov POINTER [ebx+0*SIZEOF_POINTER], esi 446 mov POINTER [ebx-1*SIZEOF_POINTER], edi 447 448 call near EXTN(jsimd_h2v1_merged_upsample_mmx) 449 450 add esp, byte 7*SIZEOF_DWORD 451 452 pop edi 453 pop esi 454; pop edx ; need not be preserved 455; pop ecx ; need not be preserved 456 pop ebx 457 pop ebp 458 ret 459 460; For some reason, the OS X linker does not honor the request to align the 461; segment unless we do this. 462 align 32 463