jidctred-mmx.asm (27106B)
1 ; 2 ; jidctred.asm - reduced-size IDCT (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 ; This file contains inverse-DCT routines that produce reduced-size 14 ; output: either 4x4 or 2x2 pixels from an 8x8 DCT block. 15 ; The following code is based directly on the IJG's original jidctred.c; 16 ; see the jidctred.c for more details. 17 18 %include "jsimdext.inc" 19 %include "jdct.inc" 20 21 ; -------------------------------------------------------------------------- 22 23 %define CONST_BITS 13 24 %define PASS1_BITS 2 25 26 %define DESCALE_P1_4 (CONST_BITS - PASS1_BITS + 1) 27 %define DESCALE_P2_4 (CONST_BITS + PASS1_BITS + 3 + 1) 28 %define DESCALE_P1_2 (CONST_BITS - PASS1_BITS + 2) 29 %define DESCALE_P2_2 (CONST_BITS + PASS1_BITS + 3 + 2) 30 31 %if CONST_BITS == 13 32 F_0_211 equ 1730 ; FIX(0.211164243) 33 F_0_509 equ 4176 ; FIX(0.509795579) 34 F_0_601 equ 4926 ; FIX(0.601344887) 35 F_0_720 equ 5906 ; FIX(0.720959822) 36 F_0_765 equ 6270 ; FIX(0.765366865) 37 F_0_850 equ 6967 ; FIX(0.850430095) 38 F_0_899 equ 7373 ; FIX(0.899976223) 39 F_1_061 equ 8697 ; FIX(1.061594337) 40 F_1_272 equ 10426 ; FIX(1.272758580) 41 F_1_451 equ 11893 ; FIX(1.451774981) 42 F_1_847 equ 15137 ; FIX(1.847759065) 43 F_2_172 equ 17799 ; FIX(2.172734803) 44 F_2_562 equ 20995 ; FIX(2.562915447) 45 F_3_624 equ 29692 ; FIX(3.624509785) 46 %else 47 ; NASM cannot do compile-time arithmetic on floating-point constants. 48 %define DESCALE(x, n) (((x) + (1 << ((n) - 1))) >> (n)) 49 F_0_211 equ DESCALE( 226735879, 30 - CONST_BITS) ; FIX(0.211164243) 50 F_0_509 equ DESCALE( 547388834, 30 - CONST_BITS) ; FIX(0.509795579) 51 F_0_601 equ DESCALE( 645689155, 30 - CONST_BITS) ; FIX(0.601344887) 52 F_0_720 equ DESCALE( 774124714, 30 - CONST_BITS) ; FIX(0.720959822) 53 F_0_765 equ DESCALE( 821806413, 30 - CONST_BITS) ; FIX(0.765366865) 54 F_0_850 equ DESCALE( 913142361, 30 - CONST_BITS) ; FIX(0.850430095) 55 F_0_899 equ DESCALE( 966342111, 30 - CONST_BITS) ; FIX(0.899976223) 56 F_1_061 equ DESCALE(1139878239, 30 - CONST_BITS) ; FIX(1.061594337) 57 F_1_272 equ DESCALE(1366614119, 30 - CONST_BITS) ; FIX(1.272758580) 58 F_1_451 equ DESCALE(1558831516, 30 - CONST_BITS) ; FIX(1.451774981) 59 F_1_847 equ DESCALE(1984016188, 30 - CONST_BITS) ; FIX(1.847759065) 60 F_2_172 equ DESCALE(2332956230, 30 - CONST_BITS) ; FIX(2.172734803) 61 F_2_562 equ DESCALE(2751909506, 30 - CONST_BITS) ; FIX(2.562915447) 62 F_3_624 equ DESCALE(3891787747, 30 - CONST_BITS) ; FIX(3.624509785) 63 %endif 64 65 ; -------------------------------------------------------------------------- 66 SECTION SEG_CONST 67 68 ALIGNZ 32 69 GLOBAL_DATA(jconst_idct_red_mmx) 70 71 EXTN(jconst_idct_red_mmx): 72 73 PW_F184_MF076 times 2 dw F_1_847, -F_0_765 74 PW_F256_F089 times 2 dw F_2_562, F_0_899 75 PW_F106_MF217 times 2 dw F_1_061, -F_2_172 76 PW_MF060_MF050 times 2 dw -F_0_601, -F_0_509 77 PW_F145_MF021 times 2 dw F_1_451, -F_0_211 78 PW_F362_MF127 times 2 dw F_3_624, -F_1_272 79 PW_F085_MF072 times 2 dw F_0_850, -F_0_720 80 PD_DESCALE_P1_4 times 2 dd 1 << (DESCALE_P1_4 - 1) 81 PD_DESCALE_P2_4 times 2 dd 1 << (DESCALE_P2_4 - 1) 82 PD_DESCALE_P1_2 times 2 dd 1 << (DESCALE_P1_2 - 1) 83 PD_DESCALE_P2_2 times 2 dd 1 << (DESCALE_P2_2 - 1) 84 PB_CENTERJSAMP times 8 db CENTERJSAMPLE 85 86 ALIGNZ 32 87 88 ; -------------------------------------------------------------------------- 89 SECTION SEG_TEXT 90 BITS 32 91 ; 92 ; Perform dequantization and inverse DCT on one block of coefficients, 93 ; producing a reduced-size 4x4 output block. 94 ; 95 ; GLOBAL(void) 96 ; jsimd_idct_4x4_mmx(void *dct_table, JCOEFPTR coef_block, 97 ; JSAMPARRAY output_buf, JDIMENSION output_col) 98 ; 99 100 %define dct_table(b) (b) + 8 ; void *dct_table 101 %define coef_block(b) (b) + 12 ; JCOEFPTR coef_block 102 %define output_buf(b) (b) + 16 ; JSAMPARRAY output_buf 103 %define output_col(b) (b) + 20 ; JDIMENSION output_col 104 105 %define original_ebp ebp + 0 106 %define wk(i) ebp - (WK_NUM - (i)) * SIZEOF_MMWORD 107 ; mmword wk[WK_NUM] 108 %define WK_NUM 2 109 %define workspace wk(0) - DCTSIZE2 * SIZEOF_JCOEF 110 ; JCOEF workspace[DCTSIZE2] 111 112 align 32 113 GLOBAL_FUNCTION(jsimd_idct_4x4_mmx) 114 115 EXTN(jsimd_idct_4x4_mmx): 116 push ebp 117 mov eax, esp ; eax = original ebp 118 sub esp, byte 4 119 and esp, byte (-SIZEOF_MMWORD) ; align to 64 bits 120 mov [esp], eax 121 mov ebp, esp ; ebp = aligned ebp 122 lea esp, [workspace] 123 PUSHPIC ebx 124 ; push ecx ; need not be preserved 125 ; push edx ; need not be preserved 126 push esi 127 push edi 128 129 GET_GOT ebx ; get GOT address 130 131 ; ---- Pass 1: process columns from input, store into work array. 132 133 ; mov eax, [original_ebp] 134 mov edx, POINTER [dct_table(eax)] ; quantptr 135 mov esi, JCOEFPTR [coef_block(eax)] ; inptr 136 lea edi, [workspace] ; JCOEF *wsptr 137 mov ecx, DCTSIZE/4 ; ctr 138 ALIGNX 16, 7 139 .columnloop: 140 %ifndef NO_ZERO_COLUMN_TEST_4X4_MMX 141 mov eax, dword [DWBLOCK(1,0,esi,SIZEOF_JCOEF)] 142 or eax, dword [DWBLOCK(2,0,esi,SIZEOF_JCOEF)] 143 jnz short .columnDCT 144 145 movq mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)] 146 movq mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)] 147 por mm0, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)] 148 por mm1, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)] 149 por mm0, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)] 150 por mm1, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)] 151 por mm0, mm1 152 packsswb mm0, mm0 153 movd eax, mm0 154 test eax, eax 155 jnz short .columnDCT 156 157 ; -- AC terms all zero 158 159 movq mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)] 160 pmullw mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)] 161 162 psllw mm0, PASS1_BITS 163 164 movq mm2, mm0 ; mm0=in0=(00 01 02 03) 165 punpcklwd mm0, mm0 ; mm0=(00 00 01 01) 166 punpckhwd mm2, mm2 ; mm2=(02 02 03 03) 167 168 movq mm1, mm0 169 punpckldq mm0, mm0 ; mm0=(00 00 00 00) 170 punpckhdq mm1, mm1 ; mm1=(01 01 01 01) 171 movq mm3, mm2 172 punpckldq mm2, mm2 ; mm2=(02 02 02 02) 173 punpckhdq mm3, mm3 ; mm3=(03 03 03 03) 174 175 movq MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm0 176 movq MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm1 177 movq MMWORD [MMBLOCK(2,0,edi,SIZEOF_JCOEF)], mm2 178 movq MMWORD [MMBLOCK(3,0,edi,SIZEOF_JCOEF)], mm3 179 jmp near .nextcolumn 180 ALIGNX 16, 7 181 %endif 182 .columnDCT: 183 184 ; -- Odd part 185 186 movq mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)] 187 movq mm1, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)] 188 pmullw mm0, MMWORD [MMBLOCK(1,0,edx,SIZEOF_ISLOW_MULT_TYPE)] 189 pmullw mm1, MMWORD [MMBLOCK(3,0,edx,SIZEOF_ISLOW_MULT_TYPE)] 190 movq mm2, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)] 191 movq mm3, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)] 192 pmullw mm2, MMWORD [MMBLOCK(5,0,edx,SIZEOF_ISLOW_MULT_TYPE)] 193 pmullw mm3, MMWORD [MMBLOCK(7,0,edx,SIZEOF_ISLOW_MULT_TYPE)] 194 195 movq mm4, mm0 196 movq mm5, mm0 197 punpcklwd mm4, mm1 198 punpckhwd mm5, mm1 199 movq mm0, mm4 200 movq mm1, mm5 201 pmaddwd mm4, [GOTOFF(ebx,PW_F256_F089)] ; mm4=(tmp2L) 202 pmaddwd mm5, [GOTOFF(ebx,PW_F256_F089)] ; mm5=(tmp2H) 203 pmaddwd mm0, [GOTOFF(ebx,PW_F106_MF217)] ; mm0=(tmp0L) 204 pmaddwd mm1, [GOTOFF(ebx,PW_F106_MF217)] ; mm1=(tmp0H) 205 206 movq mm6, mm2 207 movq mm7, mm2 208 punpcklwd mm6, mm3 209 punpckhwd mm7, mm3 210 movq mm2, mm6 211 movq mm3, mm7 212 pmaddwd mm6, [GOTOFF(ebx,PW_MF060_MF050)] ; mm6=(tmp2L) 213 pmaddwd mm7, [GOTOFF(ebx,PW_MF060_MF050)] ; mm7=(tmp2H) 214 pmaddwd mm2, [GOTOFF(ebx,PW_F145_MF021)] ; mm2=(tmp0L) 215 pmaddwd mm3, [GOTOFF(ebx,PW_F145_MF021)] ; mm3=(tmp0H) 216 217 paddd mm6, mm4 ; mm6=tmp2L 218 paddd mm7, mm5 ; mm7=tmp2H 219 paddd mm2, mm0 ; mm2=tmp0L 220 paddd mm3, mm1 ; mm3=tmp0H 221 222 movq MMWORD [wk(0)], mm2 ; wk(0)=tmp0L 223 movq MMWORD [wk(1)], mm3 ; wk(1)=tmp0H 224 225 ; -- Even part 226 227 movq mm4, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)] 228 movq mm5, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)] 229 movq mm0, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)] 230 pmullw mm4, MMWORD [MMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)] 231 pmullw mm5, MMWORD [MMBLOCK(2,0,edx,SIZEOF_ISLOW_MULT_TYPE)] 232 pmullw mm0, MMWORD [MMBLOCK(6,0,edx,SIZEOF_ISLOW_MULT_TYPE)] 233 234 pxor mm1, mm1 235 pxor mm2, mm2 236 punpcklwd mm1, mm4 ; mm1=tmp0L 237 punpckhwd mm2, mm4 ; mm2=tmp0H 238 psrad mm1, (16-CONST_BITS-1) ; psrad mm1,16 & pslld mm1,CONST_BITS+1 239 psrad mm2, (16-CONST_BITS-1) ; psrad mm2,16 & pslld mm2,CONST_BITS+1 240 241 movq mm3, mm5 ; mm5=in2=z2 242 punpcklwd mm5, mm0 ; mm0=in6=z3 243 punpckhwd mm3, mm0 244 pmaddwd mm5, [GOTOFF(ebx,PW_F184_MF076)] ; mm5=tmp2L 245 pmaddwd mm3, [GOTOFF(ebx,PW_F184_MF076)] ; mm3=tmp2H 246 247 movq mm4, mm1 248 movq mm0, mm2 249 paddd mm1, mm5 ; mm1=tmp10L 250 paddd mm2, mm3 ; mm2=tmp10H 251 psubd mm4, mm5 ; mm4=tmp12L 252 psubd mm0, mm3 ; mm0=tmp12H 253 254 ; -- Final output stage 255 256 movq mm5, mm1 257 movq mm3, mm2 258 paddd mm1, mm6 ; mm1=data0L 259 paddd mm2, mm7 ; mm2=data0H 260 psubd mm5, mm6 ; mm5=data3L 261 psubd mm3, mm7 ; mm3=data3H 262 263 movq mm6, [GOTOFF(ebx,PD_DESCALE_P1_4)] ; mm6=[PD_DESCALE_P1_4] 264 265 paddd mm1, mm6 266 paddd mm2, mm6 267 psrad mm1, DESCALE_P1_4 268 psrad mm2, DESCALE_P1_4 269 paddd mm5, mm6 270 paddd mm3, mm6 271 psrad mm5, DESCALE_P1_4 272 psrad mm3, DESCALE_P1_4 273 274 packssdw mm1, mm2 ; mm1=data0=(00 01 02 03) 275 packssdw mm5, mm3 ; mm5=data3=(30 31 32 33) 276 277 movq mm7, MMWORD [wk(0)] ; mm7=tmp0L 278 movq mm6, MMWORD [wk(1)] ; mm6=tmp0H 279 280 movq mm2, mm4 281 movq mm3, mm0 282 paddd mm4, mm7 ; mm4=data1L 283 paddd mm0, mm6 ; mm0=data1H 284 psubd mm2, mm7 ; mm2=data2L 285 psubd mm3, mm6 ; mm3=data2H 286 287 movq mm7, [GOTOFF(ebx,PD_DESCALE_P1_4)] ; mm7=[PD_DESCALE_P1_4] 288 289 paddd mm4, mm7 290 paddd mm0, mm7 291 psrad mm4, DESCALE_P1_4 292 psrad mm0, DESCALE_P1_4 293 paddd mm2, mm7 294 paddd mm3, mm7 295 psrad mm2, DESCALE_P1_4 296 psrad mm3, DESCALE_P1_4 297 298 packssdw mm4, mm0 ; mm4=data1=(10 11 12 13) 299 packssdw mm2, mm3 ; mm2=data2=(20 21 22 23) 300 301 movq mm6, mm1 ; transpose coefficients(phase 1) 302 punpcklwd mm1, mm4 ; mm1=(00 10 01 11) 303 punpckhwd mm6, mm4 ; mm6=(02 12 03 13) 304 movq mm7, mm2 ; transpose coefficients(phase 1) 305 punpcklwd mm2, mm5 ; mm2=(20 30 21 31) 306 punpckhwd mm7, mm5 ; mm7=(22 32 23 33) 307 308 movq mm0, mm1 ; transpose coefficients(phase 2) 309 punpckldq mm1, mm2 ; mm1=(00 10 20 30) 310 punpckhdq mm0, mm2 ; mm0=(01 11 21 31) 311 movq mm3, mm6 ; transpose coefficients(phase 2) 312 punpckldq mm6, mm7 ; mm6=(02 12 22 32) 313 punpckhdq mm3, mm7 ; mm3=(03 13 23 33) 314 315 movq MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm1 316 movq MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm0 317 movq MMWORD [MMBLOCK(2,0,edi,SIZEOF_JCOEF)], mm6 318 movq MMWORD [MMBLOCK(3,0,edi,SIZEOF_JCOEF)], mm3 319 320 .nextcolumn: 321 add esi, byte 4*SIZEOF_JCOEF ; coef_block 322 add edx, byte 4*SIZEOF_ISLOW_MULT_TYPE ; quantptr 323 add edi, byte 4*DCTSIZE*SIZEOF_JCOEF ; wsptr 324 dec ecx ; ctr 325 jnz near .columnloop 326 327 ; ---- Pass 2: process rows from work array, store into output array. 328 329 mov eax, [original_ebp] 330 lea esi, [workspace] ; JCOEF *wsptr 331 mov edi, JSAMPARRAY [output_buf(eax)] ; (JSAMPROW *) 332 mov eax, JDIMENSION [output_col(eax)] 333 334 ; -- Odd part 335 336 movq mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)] 337 movq mm1, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)] 338 movq mm2, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)] 339 movq mm3, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)] 340 341 movq mm4, mm0 342 movq mm5, mm0 343 punpcklwd mm4, mm1 344 punpckhwd mm5, mm1 345 movq mm0, mm4 346 movq mm1, mm5 347 pmaddwd mm4, [GOTOFF(ebx,PW_F256_F089)] ; mm4=(tmp2L) 348 pmaddwd mm5, [GOTOFF(ebx,PW_F256_F089)] ; mm5=(tmp2H) 349 pmaddwd mm0, [GOTOFF(ebx,PW_F106_MF217)] ; mm0=(tmp0L) 350 pmaddwd mm1, [GOTOFF(ebx,PW_F106_MF217)] ; mm1=(tmp0H) 351 352 movq mm6, mm2 353 movq mm7, mm2 354 punpcklwd mm6, mm3 355 punpckhwd mm7, mm3 356 movq mm2, mm6 357 movq mm3, mm7 358 pmaddwd mm6, [GOTOFF(ebx,PW_MF060_MF050)] ; mm6=(tmp2L) 359 pmaddwd mm7, [GOTOFF(ebx,PW_MF060_MF050)] ; mm7=(tmp2H) 360 pmaddwd mm2, [GOTOFF(ebx,PW_F145_MF021)] ; mm2=(tmp0L) 361 pmaddwd mm3, [GOTOFF(ebx,PW_F145_MF021)] ; mm3=(tmp0H) 362 363 paddd mm6, mm4 ; mm6=tmp2L 364 paddd mm7, mm5 ; mm7=tmp2H 365 paddd mm2, mm0 ; mm2=tmp0L 366 paddd mm3, mm1 ; mm3=tmp0H 367 368 movq MMWORD [wk(0)], mm2 ; wk(0)=tmp0L 369 movq MMWORD [wk(1)], mm3 ; wk(1)=tmp0H 370 371 ; -- Even part 372 373 movq mm4, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)] 374 movq mm5, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)] 375 movq mm0, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)] 376 377 pxor mm1, mm1 378 pxor mm2, mm2 379 punpcklwd mm1, mm4 ; mm1=tmp0L 380 punpckhwd mm2, mm4 ; mm2=tmp0H 381 psrad mm1, (16-CONST_BITS-1) ; psrad mm1,16 & pslld mm1,CONST_BITS+1 382 psrad mm2, (16-CONST_BITS-1) ; psrad mm2,16 & pslld mm2,CONST_BITS+1 383 384 movq mm3, mm5 ; mm5=in2=z2 385 punpcklwd mm5, mm0 ; mm0=in6=z3 386 punpckhwd mm3, mm0 387 pmaddwd mm5, [GOTOFF(ebx,PW_F184_MF076)] ; mm5=tmp2L 388 pmaddwd mm3, [GOTOFF(ebx,PW_F184_MF076)] ; mm3=tmp2H 389 390 movq mm4, mm1 391 movq mm0, mm2 392 paddd mm1, mm5 ; mm1=tmp10L 393 paddd mm2, mm3 ; mm2=tmp10H 394 psubd mm4, mm5 ; mm4=tmp12L 395 psubd mm0, mm3 ; mm0=tmp12H 396 397 ; -- Final output stage 398 399 movq mm5, mm1 400 movq mm3, mm2 401 paddd mm1, mm6 ; mm1=data0L 402 paddd mm2, mm7 ; mm2=data0H 403 psubd mm5, mm6 ; mm5=data3L 404 psubd mm3, mm7 ; mm3=data3H 405 406 movq mm6, [GOTOFF(ebx,PD_DESCALE_P2_4)] ; mm6=[PD_DESCALE_P2_4] 407 408 paddd mm1, mm6 409 paddd mm2, mm6 410 psrad mm1, DESCALE_P2_4 411 psrad mm2, DESCALE_P2_4 412 paddd mm5, mm6 413 paddd mm3, mm6 414 psrad mm5, DESCALE_P2_4 415 psrad mm3, DESCALE_P2_4 416 417 packssdw mm1, mm2 ; mm1=data0=(00 10 20 30) 418 packssdw mm5, mm3 ; mm5=data3=(03 13 23 33) 419 420 movq mm7, MMWORD [wk(0)] ; mm7=tmp0L 421 movq mm6, MMWORD [wk(1)] ; mm6=tmp0H 422 423 movq mm2, mm4 424 movq mm3, mm0 425 paddd mm4, mm7 ; mm4=data1L 426 paddd mm0, mm6 ; mm0=data1H 427 psubd mm2, mm7 ; mm2=data2L 428 psubd mm3, mm6 ; mm3=data2H 429 430 movq mm7, [GOTOFF(ebx,PD_DESCALE_P2_4)] ; mm7=[PD_DESCALE_P2_4] 431 432 paddd mm4, mm7 433 paddd mm0, mm7 434 psrad mm4, DESCALE_P2_4 435 psrad mm0, DESCALE_P2_4 436 paddd mm2, mm7 437 paddd mm3, mm7 438 psrad mm2, DESCALE_P2_4 439 psrad mm3, DESCALE_P2_4 440 441 packssdw mm4, mm0 ; mm4=data1=(01 11 21 31) 442 packssdw mm2, mm3 ; mm2=data2=(02 12 22 32) 443 444 movq mm6, [GOTOFF(ebx,PB_CENTERJSAMP)] ; mm6=[PB_CENTERJSAMP] 445 446 packsswb mm1, mm2 ; mm1=(00 10 20 30 02 12 22 32) 447 packsswb mm4, mm5 ; mm4=(01 11 21 31 03 13 23 33) 448 paddb mm1, mm6 449 paddb mm4, mm6 450 451 movq mm7, mm1 ; transpose coefficients(phase 1) 452 punpcklbw mm1, mm4 ; mm1=(00 01 10 11 20 21 30 31) 453 punpckhbw mm7, mm4 ; mm7=(02 03 12 13 22 23 32 33) 454 455 movq mm0, mm1 ; transpose coefficients(phase 2) 456 punpcklwd mm1, mm7 ; mm1=(00 01 02 03 10 11 12 13) 457 punpckhwd mm0, mm7 ; mm0=(20 21 22 23 30 31 32 33) 458 459 mov edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW] 460 mov esi, JSAMPROW [edi+2*SIZEOF_JSAMPROW] 461 movd dword [edx+eax*SIZEOF_JSAMPLE], mm1 462 movd dword [esi+eax*SIZEOF_JSAMPLE], mm0 463 464 psrlq mm1, 4*BYTE_BIT 465 psrlq mm0, 4*BYTE_BIT 466 467 mov edx, JSAMPROW [edi+1*SIZEOF_JSAMPROW] 468 mov esi, JSAMPROW [edi+3*SIZEOF_JSAMPROW] 469 movd dword [edx+eax*SIZEOF_JSAMPLE], mm1 470 movd dword [esi+eax*SIZEOF_JSAMPLE], mm0 471 472 emms ; empty MMX state 473 474 pop edi 475 pop esi 476 ; pop edx ; need not be preserved 477 ; pop ecx ; need not be preserved 478 POPPIC ebx 479 mov esp, ebp ; esp <- aligned ebp 480 pop esp ; esp <- original ebp 481 pop ebp 482 ret 483 484 ; -------------------------------------------------------------------------- 485 ; 486 ; Perform dequantization and inverse DCT on one block of coefficients, 487 ; producing a reduced-size 2x2 output block. 488 ; 489 ; GLOBAL(void) 490 ; jsimd_idct_2x2_mmx(void *dct_table, JCOEFPTR coef_block, 491 ; JSAMPARRAY output_buf, JDIMENSION output_col) 492 ; 493 494 %define dct_table(b) (b) + 8 ; void *dct_table 495 %define coef_block(b) (b) + 12 ; JCOEFPTR coef_block 496 %define output_buf(b) (b) + 16 ; JSAMPARRAY output_buf 497 %define output_col(b) (b) + 20 ; JDIMENSION output_col 498 499 align 32 500 GLOBAL_FUNCTION(jsimd_idct_2x2_mmx) 501 502 EXTN(jsimd_idct_2x2_mmx): 503 push ebp 504 mov ebp, esp 505 push ebx 506 ; push ecx ; need not be preserved 507 ; push edx ; need not be preserved 508 push esi 509 push edi 510 511 GET_GOT ebx ; get GOT address 512 513 ; ---- Pass 1: process columns from input. 514 515 mov edx, POINTER [dct_table(ebp)] ; quantptr 516 mov esi, JCOEFPTR [coef_block(ebp)] ; inptr 517 518 ; | input: | result: | 519 ; | 00 01 ** 03 ** 05 ** 07 | | 520 ; | 10 11 ** 13 ** 15 ** 17 | | 521 ; | ** ** ** ** ** ** ** ** | | 522 ; | 30 31 ** 33 ** 35 ** 37 | A0 A1 A3 A5 A7 | 523 ; | ** ** ** ** ** ** ** ** | B0 B1 B3 B5 B7 | 524 ; | 50 51 ** 53 ** 55 ** 57 | | 525 ; | ** ** ** ** ** ** ** ** | | 526 ; | 70 71 ** 73 ** 75 ** 77 | | 527 528 ; -- Odd part 529 530 movq mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)] 531 movq mm1, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)] 532 pmullw mm0, MMWORD [MMBLOCK(1,0,edx,SIZEOF_ISLOW_MULT_TYPE)] 533 pmullw mm1, MMWORD [MMBLOCK(3,0,edx,SIZEOF_ISLOW_MULT_TYPE)] 534 movq mm2, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)] 535 movq mm3, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)] 536 pmullw mm2, MMWORD [MMBLOCK(5,0,edx,SIZEOF_ISLOW_MULT_TYPE)] 537 pmullw mm3, MMWORD [MMBLOCK(7,0,edx,SIZEOF_ISLOW_MULT_TYPE)] 538 539 ; mm0=(10 11 ** 13), mm1=(30 31 ** 33) 540 ; mm2=(50 51 ** 53), mm3=(70 71 ** 73) 541 542 pcmpeqd mm7, mm7 543 pslld mm7, WORD_BIT ; mm7={0x0000 0xFFFF 0x0000 0xFFFF} 544 545 movq mm4, mm0 ; mm4=(10 11 ** 13) 546 movq mm5, mm2 ; mm5=(50 51 ** 53) 547 punpcklwd mm4, mm1 ; mm4=(10 30 11 31) 548 punpcklwd mm5, mm3 ; mm5=(50 70 51 71) 549 pmaddwd mm4, [GOTOFF(ebx,PW_F362_MF127)] 550 pmaddwd mm5, [GOTOFF(ebx,PW_F085_MF072)] 551 552 psrld mm0, WORD_BIT ; mm0=(11 -- 13 --) 553 pand mm1, mm7 ; mm1=(-- 31 -- 33) 554 psrld mm2, WORD_BIT ; mm2=(51 -- 53 --) 555 pand mm3, mm7 ; mm3=(-- 71 -- 73) 556 por mm0, mm1 ; mm0=(11 31 13 33) 557 por mm2, mm3 ; mm2=(51 71 53 73) 558 pmaddwd mm0, [GOTOFF(ebx,PW_F362_MF127)] 559 pmaddwd mm2, [GOTOFF(ebx,PW_F085_MF072)] 560 561 paddd mm4, mm5 ; mm4=tmp0[col0 col1] 562 563 movq mm6, MMWORD [MMBLOCK(1,1,esi,SIZEOF_JCOEF)] 564 movq mm1, MMWORD [MMBLOCK(3,1,esi,SIZEOF_JCOEF)] 565 pmullw mm6, MMWORD [MMBLOCK(1,1,edx,SIZEOF_ISLOW_MULT_TYPE)] 566 pmullw mm1, MMWORD [MMBLOCK(3,1,edx,SIZEOF_ISLOW_MULT_TYPE)] 567 movq mm3, MMWORD [MMBLOCK(5,1,esi,SIZEOF_JCOEF)] 568 movq mm5, MMWORD [MMBLOCK(7,1,esi,SIZEOF_JCOEF)] 569 pmullw mm3, MMWORD [MMBLOCK(5,1,edx,SIZEOF_ISLOW_MULT_TYPE)] 570 pmullw mm5, MMWORD [MMBLOCK(7,1,edx,SIZEOF_ISLOW_MULT_TYPE)] 571 572 ; mm6=(** 15 ** 17), mm1=(** 35 ** 37) 573 ; mm3=(** 55 ** 57), mm5=(** 75 ** 77) 574 575 psrld mm6, WORD_BIT ; mm6=(15 -- 17 --) 576 pand mm1, mm7 ; mm1=(-- 35 -- 37) 577 psrld mm3, WORD_BIT ; mm3=(55 -- 57 --) 578 pand mm5, mm7 ; mm5=(-- 75 -- 77) 579 por mm6, mm1 ; mm6=(15 35 17 37) 580 por mm3, mm5 ; mm3=(55 75 57 77) 581 pmaddwd mm6, [GOTOFF(ebx,PW_F362_MF127)] 582 pmaddwd mm3, [GOTOFF(ebx,PW_F085_MF072)] 583 584 paddd mm0, mm2 ; mm0=tmp0[col1 col3] 585 paddd mm6, mm3 ; mm6=tmp0[col5 col7] 586 587 ; -- Even part 588 589 movq mm1, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)] 590 movq mm5, MMWORD [MMBLOCK(0,1,esi,SIZEOF_JCOEF)] 591 pmullw mm1, MMWORD [MMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)] 592 pmullw mm5, MMWORD [MMBLOCK(0,1,edx,SIZEOF_ISLOW_MULT_TYPE)] 593 594 ; mm1=(00 01 ** 03), mm5=(** 05 ** 07) 595 596 movq mm2, mm1 ; mm2=(00 01 ** 03) 597 pslld mm1, WORD_BIT ; mm1=(-- 00 -- **) 598 psrad mm1, (WORD_BIT-CONST_BITS-2) ; mm1=tmp10[col0 ****] 599 600 pand mm2, mm7 ; mm2=(-- 01 -- 03) 601 pand mm5, mm7 ; mm5=(-- 05 -- 07) 602 psrad mm2, (WORD_BIT-CONST_BITS-2) ; mm2=tmp10[col1 col3] 603 psrad mm5, (WORD_BIT-CONST_BITS-2) ; mm5=tmp10[col5 col7] 604 605 ; -- Final output stage 606 607 movq mm3, mm1 608 paddd mm1, mm4 ; mm1=data0[col0 ****]=(A0 **) 609 psubd mm3, mm4 ; mm3=data1[col0 ****]=(B0 **) 610 punpckldq mm1, mm3 ; mm1=(A0 B0) 611 612 movq mm7, [GOTOFF(ebx,PD_DESCALE_P1_2)] ; mm7=[PD_DESCALE_P1_2] 613 614 movq mm4, mm2 615 movq mm3, mm5 616 paddd mm2, mm0 ; mm2=data0[col1 col3]=(A1 A3) 617 paddd mm5, mm6 ; mm5=data0[col5 col7]=(A5 A7) 618 psubd mm4, mm0 ; mm4=data1[col1 col3]=(B1 B3) 619 psubd mm3, mm6 ; mm3=data1[col5 col7]=(B5 B7) 620 621 paddd mm1, mm7 622 psrad mm1, DESCALE_P1_2 623 624 paddd mm2, mm7 625 paddd mm5, mm7 626 psrad mm2, DESCALE_P1_2 627 psrad mm5, DESCALE_P1_2 628 paddd mm4, mm7 629 paddd mm3, mm7 630 psrad mm4, DESCALE_P1_2 631 psrad mm3, DESCALE_P1_2 632 633 ; ---- Pass 2: process rows, store into output array. 634 635 mov edi, JSAMPARRAY [output_buf(ebp)] ; (JSAMPROW *) 636 mov eax, JDIMENSION [output_col(ebp)] 637 638 ; | input:| result:| 639 ; | A0 B0 | | 640 ; | A1 B1 | C0 C1 | 641 ; | A3 B3 | D0 D1 | 642 ; | A5 B5 | | 643 ; | A7 B7 | | 644 645 ; -- Odd part 646 647 packssdw mm2, mm4 ; mm2=(A1 A3 B1 B3) 648 packssdw mm5, mm3 ; mm5=(A5 A7 B5 B7) 649 pmaddwd mm2, [GOTOFF(ebx,PW_F362_MF127)] 650 pmaddwd mm5, [GOTOFF(ebx,PW_F085_MF072)] 651 652 paddd mm2, mm5 ; mm2=tmp0[row0 row1] 653 654 ; -- Even part 655 656 pslld mm1, (CONST_BITS+2) ; mm1=tmp10[row0 row1] 657 658 ; -- Final output stage 659 660 movq mm0, [GOTOFF(ebx,PD_DESCALE_P2_2)] ; mm0=[PD_DESCALE_P2_2] 661 662 movq mm6, mm1 663 paddd mm1, mm2 ; mm1=data0[row0 row1]=(C0 C1) 664 psubd mm6, mm2 ; mm6=data1[row0 row1]=(D0 D1) 665 666 paddd mm1, mm0 667 paddd mm6, mm0 668 psrad mm1, DESCALE_P2_2 669 psrad mm6, DESCALE_P2_2 670 671 movq mm7, mm1 ; transpose coefficients 672 punpckldq mm1, mm6 ; mm1=(C0 D0) 673 punpckhdq mm7, mm6 ; mm7=(C1 D1) 674 675 packssdw mm1, mm7 ; mm1=(C0 D0 C1 D1) 676 packsswb mm1, mm1 ; mm1=(C0 D0 C1 D1 C0 D0 C1 D1) 677 paddb mm1, [GOTOFF(ebx,PB_CENTERJSAMP)] 678 679 movd ecx, mm1 680 movd ebx, mm1 ; ebx=(C0 D0 C1 D1) 681 shr ecx, 2*BYTE_BIT ; ecx=(C1 D1 -- --) 682 683 mov edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW] 684 mov esi, JSAMPROW [edi+1*SIZEOF_JSAMPROW] 685 mov word [edx+eax*SIZEOF_JSAMPLE], bx 686 mov word [esi+eax*SIZEOF_JSAMPLE], cx 687 688 emms ; empty MMX state 689 690 pop edi 691 pop esi 692 ; pop edx ; need not be preserved 693 ; pop ecx ; need not be preserved 694 pop ebx 695 pop ebp 696 ret 697 698 ; For some reason, the OS X linker does not honor the request to align the 699 ; segment unless we do this. 700 align 32