jidctfst-sse2.asm (21348B)
1 ; 2 ; jidctfst.asm - fast integer IDCT (SSE2) 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 a fast, not so accurate integer implementation of 14 ; the inverse DCT (Discrete Cosine Transform). The following code is 15 ; based directly on the IJG's original jidctfst.c; see the jidctfst.c 16 ; for more details. 17 18 %include "jsimdext.inc" 19 %include "jdct.inc" 20 21 ; -------------------------------------------------------------------------- 22 23 %define CONST_BITS 8 ; 14 is also OK. 24 %define PASS1_BITS 2 25 26 %if IFAST_SCALE_BITS != PASS1_BITS 27 %error "'IFAST_SCALE_BITS' must be equal to 'PASS1_BITS'." 28 %endif 29 30 %if CONST_BITS == 8 31 F_1_082 equ 277 ; FIX(1.082392200) 32 F_1_414 equ 362 ; FIX(1.414213562) 33 F_1_847 equ 473 ; FIX(1.847759065) 34 F_2_613 equ 669 ; FIX(2.613125930) 35 F_1_613 equ (F_2_613 - 256) ; FIX(2.613125930) - FIX(1) 36 %else 37 ; NASM cannot do compile-time arithmetic on floating-point constants. 38 %define DESCALE(x, n) (((x) + (1 << ((n) - 1))) >> (n)) 39 F_1_082 equ DESCALE(1162209775, 30 - CONST_BITS) ; FIX(1.082392200) 40 F_1_414 equ DESCALE(1518500249, 30 - CONST_BITS) ; FIX(1.414213562) 41 F_1_847 equ DESCALE(1984016188, 30 - CONST_BITS) ; FIX(1.847759065) 42 F_2_613 equ DESCALE(2805822602, 30 - CONST_BITS) ; FIX(2.613125930) 43 F_1_613 equ (F_2_613 - (1 << CONST_BITS)) ; FIX(2.613125930) - FIX(1) 44 %endif 45 46 ; -------------------------------------------------------------------------- 47 SECTION SEG_CONST 48 49 ; PRE_MULTIPLY_SCALE_BITS <= 2 (to avoid overflow) 50 ; CONST_BITS + CONST_SHIFT + PRE_MULTIPLY_SCALE_BITS == 16 (for pmulhw) 51 52 %define PRE_MULTIPLY_SCALE_BITS 2 53 %define CONST_SHIFT (16 - PRE_MULTIPLY_SCALE_BITS - CONST_BITS) 54 55 ALIGNZ 32 56 GLOBAL_DATA(jconst_idct_ifast_sse2) 57 58 EXTN(jconst_idct_ifast_sse2): 59 60 PW_F1414 times 8 dw F_1_414 << CONST_SHIFT 61 PW_F1847 times 8 dw F_1_847 << CONST_SHIFT 62 PW_MF1613 times 8 dw -F_1_613 << CONST_SHIFT 63 PW_F1082 times 8 dw F_1_082 << CONST_SHIFT 64 PB_CENTERJSAMP times 16 db CENTERJSAMPLE 65 66 ALIGNZ 32 67 68 ; -------------------------------------------------------------------------- 69 SECTION SEG_TEXT 70 BITS 32 71 ; 72 ; Perform dequantization and inverse DCT on one block of coefficients. 73 ; 74 ; GLOBAL(void) 75 ; jsimd_idct_ifast_sse2(void *dct_table, JCOEFPTR coef_block, 76 ; JSAMPARRAY output_buf, JDIMENSION output_col) 77 ; 78 79 %define dct_table(b) (b) + 8 ; jpeg_component_info *compptr 80 %define coef_block(b) (b) + 12 ; JCOEFPTR coef_block 81 %define output_buf(b) (b) + 16 ; JSAMPARRAY output_buf 82 %define output_col(b) (b) + 20 ; JDIMENSION output_col 83 84 %define original_ebp ebp + 0 85 %define wk(i) ebp - (WK_NUM - (i)) * SIZEOF_XMMWORD 86 ; xmmword wk[WK_NUM] 87 %define WK_NUM 2 88 89 align 32 90 GLOBAL_FUNCTION(jsimd_idct_ifast_sse2) 91 92 EXTN(jsimd_idct_ifast_sse2): 93 push ebp 94 mov eax, esp ; eax = original ebp 95 sub esp, byte 4 96 and esp, byte (-SIZEOF_XMMWORD) ; align to 128 bits 97 mov [esp], eax 98 mov ebp, esp ; ebp = aligned ebp 99 lea esp, [wk(0)] 100 PUSHPIC ebx 101 ; push ecx ; unused 102 ; push edx ; need not be preserved 103 push esi 104 push edi 105 106 GET_GOT ebx ; get GOT address 107 108 ; ---- Pass 1: process columns from input. 109 110 ; mov eax, [original_ebp] 111 mov edx, POINTER [dct_table(eax)] ; quantptr 112 mov esi, JCOEFPTR [coef_block(eax)] ; inptr 113 114 %ifndef NO_ZERO_COLUMN_TEST_IFAST_SSE2 115 mov eax, dword [DWBLOCK(1,0,esi,SIZEOF_JCOEF)] 116 or eax, dword [DWBLOCK(2,0,esi,SIZEOF_JCOEF)] 117 jnz near .columnDCT 118 119 movdqa xmm0, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_JCOEF)] 120 movdqa xmm1, XMMWORD [XMMBLOCK(2,0,esi,SIZEOF_JCOEF)] 121 por xmm0, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_JCOEF)] 122 por xmm1, XMMWORD [XMMBLOCK(4,0,esi,SIZEOF_JCOEF)] 123 por xmm0, XMMWORD [XMMBLOCK(5,0,esi,SIZEOF_JCOEF)] 124 por xmm1, XMMWORD [XMMBLOCK(6,0,esi,SIZEOF_JCOEF)] 125 por xmm0, XMMWORD [XMMBLOCK(7,0,esi,SIZEOF_JCOEF)] 126 por xmm1, xmm0 127 packsswb xmm1, xmm1 128 packsswb xmm1, xmm1 129 movd eax, xmm1 130 test eax, eax 131 jnz short .columnDCT 132 133 ; -- AC terms all zero 134 135 movdqa xmm0, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_JCOEF)] 136 pmullw xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)] 137 138 movdqa xmm7, xmm0 ; xmm0=in0=(00 01 02 03 04 05 06 07) 139 punpcklwd xmm0, xmm0 ; xmm0=(00 00 01 01 02 02 03 03) 140 punpckhwd xmm7, xmm7 ; xmm7=(04 04 05 05 06 06 07 07) 141 142 pshufd xmm6, xmm0, 0x00 ; xmm6=col0=(00 00 00 00 00 00 00 00) 143 pshufd xmm2, xmm0, 0x55 ; xmm2=col1=(01 01 01 01 01 01 01 01) 144 pshufd xmm5, xmm0, 0xAA ; xmm5=col2=(02 02 02 02 02 02 02 02) 145 pshufd xmm0, xmm0, 0xFF ; xmm0=col3=(03 03 03 03 03 03 03 03) 146 pshufd xmm1, xmm7, 0x00 ; xmm1=col4=(04 04 04 04 04 04 04 04) 147 pshufd xmm4, xmm7, 0x55 ; xmm4=col5=(05 05 05 05 05 05 05 05) 148 pshufd xmm3, xmm7, 0xAA ; xmm3=col6=(06 06 06 06 06 06 06 06) 149 pshufd xmm7, xmm7, 0xFF ; xmm7=col7=(07 07 07 07 07 07 07 07) 150 151 movdqa XMMWORD [wk(0)], xmm2 ; wk(0)=col1 152 movdqa XMMWORD [wk(1)], xmm0 ; wk(1)=col3 153 jmp near .column_end 154 ALIGNX 16, 7 155 %endif 156 .columnDCT: 157 158 ; -- Even part 159 160 movdqa xmm0, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_JCOEF)] 161 movdqa xmm1, XMMWORD [XMMBLOCK(2,0,esi,SIZEOF_JCOEF)] 162 pmullw xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_IFAST_MULT_TYPE)] 163 pmullw xmm1, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_IFAST_MULT_TYPE)] 164 movdqa xmm2, XMMWORD [XMMBLOCK(4,0,esi,SIZEOF_JCOEF)] 165 movdqa xmm3, XMMWORD [XMMBLOCK(6,0,esi,SIZEOF_JCOEF)] 166 pmullw xmm2, XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_IFAST_MULT_TYPE)] 167 pmullw xmm3, XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_IFAST_MULT_TYPE)] 168 169 movdqa xmm4, xmm0 170 movdqa xmm5, xmm1 171 psubw xmm0, xmm2 ; xmm0=tmp11 172 psubw xmm1, xmm3 173 paddw xmm4, xmm2 ; xmm4=tmp10 174 paddw xmm5, xmm3 ; xmm5=tmp13 175 176 psllw xmm1, PRE_MULTIPLY_SCALE_BITS 177 pmulhw xmm1, [GOTOFF(ebx,PW_F1414)] 178 psubw xmm1, xmm5 ; xmm1=tmp12 179 180 movdqa xmm6, xmm4 181 movdqa xmm7, xmm0 182 psubw xmm4, xmm5 ; xmm4=tmp3 183 psubw xmm0, xmm1 ; xmm0=tmp2 184 paddw xmm6, xmm5 ; xmm6=tmp0 185 paddw xmm7, xmm1 ; xmm7=tmp1 186 187 movdqa XMMWORD [wk(1)], xmm4 ; wk(1)=tmp3 188 movdqa XMMWORD [wk(0)], xmm0 ; wk(0)=tmp2 189 190 ; -- Odd part 191 192 movdqa xmm2, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_JCOEF)] 193 movdqa xmm3, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_JCOEF)] 194 pmullw xmm2, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_IFAST_MULT_TYPE)] 195 pmullw xmm3, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_IFAST_MULT_TYPE)] 196 movdqa xmm5, XMMWORD [XMMBLOCK(5,0,esi,SIZEOF_JCOEF)] 197 movdqa xmm1, XMMWORD [XMMBLOCK(7,0,esi,SIZEOF_JCOEF)] 198 pmullw xmm5, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_IFAST_MULT_TYPE)] 199 pmullw xmm1, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_IFAST_MULT_TYPE)] 200 201 movdqa xmm4, xmm2 202 movdqa xmm0, xmm5 203 psubw xmm2, xmm1 ; xmm2=z12 204 psubw xmm5, xmm3 ; xmm5=z10 205 paddw xmm4, xmm1 ; xmm4=z11 206 paddw xmm0, xmm3 ; xmm0=z13 207 208 movdqa xmm1, xmm5 ; xmm1=z10(unscaled) 209 psllw xmm2, PRE_MULTIPLY_SCALE_BITS 210 psllw xmm5, PRE_MULTIPLY_SCALE_BITS 211 212 movdqa xmm3, xmm4 213 psubw xmm4, xmm0 214 paddw xmm3, xmm0 ; xmm3=tmp7 215 216 psllw xmm4, PRE_MULTIPLY_SCALE_BITS 217 pmulhw xmm4, [GOTOFF(ebx,PW_F1414)] ; xmm4=tmp11 218 219 ; To avoid overflow... 220 ; 221 ; (Original) 222 ; tmp12 = -2.613125930 * z10 + z5; 223 ; 224 ; (This implementation) 225 ; tmp12 = (-1.613125930 - 1) * z10 + z5; 226 ; = -1.613125930 * z10 - z10 + z5; 227 228 movdqa xmm0, xmm5 229 paddw xmm5, xmm2 230 pmulhw xmm5, [GOTOFF(ebx,PW_F1847)] ; xmm5=z5 231 pmulhw xmm0, [GOTOFF(ebx,PW_MF1613)] 232 pmulhw xmm2, [GOTOFF(ebx,PW_F1082)] 233 psubw xmm0, xmm1 234 psubw xmm2, xmm5 ; xmm2=tmp10 235 paddw xmm0, xmm5 ; xmm0=tmp12 236 237 ; -- Final output stage 238 239 psubw xmm0, xmm3 ; xmm0=tmp6 240 movdqa xmm1, xmm6 241 movdqa xmm5, xmm7 242 paddw xmm6, xmm3 ; xmm6=data0=(00 01 02 03 04 05 06 07) 243 paddw xmm7, xmm0 ; xmm7=data1=(10 11 12 13 14 15 16 17) 244 psubw xmm1, xmm3 ; xmm1=data7=(70 71 72 73 74 75 76 77) 245 psubw xmm5, xmm0 ; xmm5=data6=(60 61 62 63 64 65 66 67) 246 psubw xmm4, xmm0 ; xmm4=tmp5 247 248 movdqa xmm3, xmm6 ; transpose coefficients(phase 1) 249 punpcklwd xmm6, xmm7 ; xmm6=(00 10 01 11 02 12 03 13) 250 punpckhwd xmm3, xmm7 ; xmm3=(04 14 05 15 06 16 07 17) 251 movdqa xmm0, xmm5 ; transpose coefficients(phase 1) 252 punpcklwd xmm5, xmm1 ; xmm5=(60 70 61 71 62 72 63 73) 253 punpckhwd xmm0, xmm1 ; xmm0=(64 74 65 75 66 76 67 77) 254 255 movdqa xmm7, XMMWORD [wk(0)] ; xmm7=tmp2 256 movdqa xmm1, XMMWORD [wk(1)] ; xmm1=tmp3 257 258 movdqa XMMWORD [wk(0)], xmm5 ; wk(0)=(60 70 61 71 62 72 63 73) 259 movdqa XMMWORD [wk(1)], xmm0 ; wk(1)=(64 74 65 75 66 76 67 77) 260 261 paddw xmm2, xmm4 ; xmm2=tmp4 262 movdqa xmm5, xmm7 263 movdqa xmm0, xmm1 264 paddw xmm7, xmm4 ; xmm7=data2=(20 21 22 23 24 25 26 27) 265 paddw xmm1, xmm2 ; xmm1=data4=(40 41 42 43 44 45 46 47) 266 psubw xmm5, xmm4 ; xmm5=data5=(50 51 52 53 54 55 56 57) 267 psubw xmm0, xmm2 ; xmm0=data3=(30 31 32 33 34 35 36 37) 268 269 movdqa xmm4, xmm7 ; transpose coefficients(phase 1) 270 punpcklwd xmm7, xmm0 ; xmm7=(20 30 21 31 22 32 23 33) 271 punpckhwd xmm4, xmm0 ; xmm4=(24 34 25 35 26 36 27 37) 272 movdqa xmm2, xmm1 ; transpose coefficients(phase 1) 273 punpcklwd xmm1, xmm5 ; xmm1=(40 50 41 51 42 52 43 53) 274 punpckhwd xmm2, xmm5 ; xmm2=(44 54 45 55 46 56 47 57) 275 276 movdqa xmm0, xmm3 ; transpose coefficients(phase 2) 277 punpckldq xmm3, xmm4 ; xmm3=(04 14 24 34 05 15 25 35) 278 punpckhdq xmm0, xmm4 ; xmm0=(06 16 26 36 07 17 27 37) 279 movdqa xmm5, xmm6 ; transpose coefficients(phase 2) 280 punpckldq xmm6, xmm7 ; xmm6=(00 10 20 30 01 11 21 31) 281 punpckhdq xmm5, xmm7 ; xmm5=(02 12 22 32 03 13 23 33) 282 283 movdqa xmm4, XMMWORD [wk(0)] ; xmm4=(60 70 61 71 62 72 63 73) 284 movdqa xmm7, XMMWORD [wk(1)] ; xmm7=(64 74 65 75 66 76 67 77) 285 286 movdqa XMMWORD [wk(0)], xmm3 ; wk(0)=(04 14 24 34 05 15 25 35) 287 movdqa XMMWORD [wk(1)], xmm0 ; wk(1)=(06 16 26 36 07 17 27 37) 288 289 movdqa xmm3, xmm1 ; transpose coefficients(phase 2) 290 punpckldq xmm1, xmm4 ; xmm1=(40 50 60 70 41 51 61 71) 291 punpckhdq xmm3, xmm4 ; xmm3=(42 52 62 72 43 53 63 73) 292 movdqa xmm0, xmm2 ; transpose coefficients(phase 2) 293 punpckldq xmm2, xmm7 ; xmm2=(44 54 64 74 45 55 65 75) 294 punpckhdq xmm0, xmm7 ; xmm0=(46 56 66 76 47 57 67 77) 295 296 movdqa xmm4, xmm6 ; transpose coefficients(phase 3) 297 punpcklqdq xmm6, xmm1 ; xmm6=col0=(00 10 20 30 40 50 60 70) 298 punpckhqdq xmm4, xmm1 ; xmm4=col1=(01 11 21 31 41 51 61 71) 299 movdqa xmm7, xmm5 ; transpose coefficients(phase 3) 300 punpcklqdq xmm5, xmm3 ; xmm5=col2=(02 12 22 32 42 52 62 72) 301 punpckhqdq xmm7, xmm3 ; xmm7=col3=(03 13 23 33 43 53 63 73) 302 303 movdqa xmm1, XMMWORD [wk(0)] ; xmm1=(04 14 24 34 05 15 25 35) 304 movdqa xmm3, XMMWORD [wk(1)] ; xmm3=(06 16 26 36 07 17 27 37) 305 306 movdqa XMMWORD [wk(0)], xmm4 ; wk(0)=col1 307 movdqa XMMWORD [wk(1)], xmm7 ; wk(1)=col3 308 309 movdqa xmm4, xmm1 ; transpose coefficients(phase 3) 310 punpcklqdq xmm1, xmm2 ; xmm1=col4=(04 14 24 34 44 54 64 74) 311 punpckhqdq xmm4, xmm2 ; xmm4=col5=(05 15 25 35 45 55 65 75) 312 movdqa xmm7, xmm3 ; transpose coefficients(phase 3) 313 punpcklqdq xmm3, xmm0 ; xmm3=col6=(06 16 26 36 46 56 66 76) 314 punpckhqdq xmm7, xmm0 ; xmm7=col7=(07 17 27 37 47 57 67 77) 315 .column_end: 316 317 ; -- Prefetch the next coefficient block 318 319 prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 0*32] 320 prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 1*32] 321 prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 2*32] 322 prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 3*32] 323 324 ; ---- Pass 2: process rows from work array, store into output array. 325 326 mov eax, [original_ebp] 327 mov edi, JSAMPARRAY [output_buf(eax)] ; (JSAMPROW *) 328 mov eax, JDIMENSION [output_col(eax)] 329 330 ; -- Even part 331 332 ; xmm6=col0, xmm5=col2, xmm1=col4, xmm3=col6 333 334 movdqa xmm2, xmm6 335 movdqa xmm0, xmm5 336 psubw xmm6, xmm1 ; xmm6=tmp11 337 psubw xmm5, xmm3 338 paddw xmm2, xmm1 ; xmm2=tmp10 339 paddw xmm0, xmm3 ; xmm0=tmp13 340 341 psllw xmm5, PRE_MULTIPLY_SCALE_BITS 342 pmulhw xmm5, [GOTOFF(ebx,PW_F1414)] 343 psubw xmm5, xmm0 ; xmm5=tmp12 344 345 movdqa xmm1, xmm2 346 movdqa xmm3, xmm6 347 psubw xmm2, xmm0 ; xmm2=tmp3 348 psubw xmm6, xmm5 ; xmm6=tmp2 349 paddw xmm1, xmm0 ; xmm1=tmp0 350 paddw xmm3, xmm5 ; xmm3=tmp1 351 352 movdqa xmm0, XMMWORD [wk(0)] ; xmm0=col1 353 movdqa xmm5, XMMWORD [wk(1)] ; xmm5=col3 354 355 movdqa XMMWORD [wk(0)], xmm2 ; wk(0)=tmp3 356 movdqa XMMWORD [wk(1)], xmm6 ; wk(1)=tmp2 357 358 ; -- Odd part 359 360 ; xmm0=col1, xmm5=col3, xmm4=col5, xmm7=col7 361 362 movdqa xmm2, xmm0 363 movdqa xmm6, xmm4 364 psubw xmm0, xmm7 ; xmm0=z12 365 psubw xmm4, xmm5 ; xmm4=z10 366 paddw xmm2, xmm7 ; xmm2=z11 367 paddw xmm6, xmm5 ; xmm6=z13 368 369 movdqa xmm7, xmm4 ; xmm7=z10(unscaled) 370 psllw xmm0, PRE_MULTIPLY_SCALE_BITS 371 psllw xmm4, PRE_MULTIPLY_SCALE_BITS 372 373 movdqa xmm5, xmm2 374 psubw xmm2, xmm6 375 paddw xmm5, xmm6 ; xmm5=tmp7 376 377 psllw xmm2, PRE_MULTIPLY_SCALE_BITS 378 pmulhw xmm2, [GOTOFF(ebx,PW_F1414)] ; xmm2=tmp11 379 380 ; To avoid overflow... 381 ; 382 ; (Original) 383 ; tmp12 = -2.613125930 * z10 + z5; 384 ; 385 ; (This implementation) 386 ; tmp12 = (-1.613125930 - 1) * z10 + z5; 387 ; = -1.613125930 * z10 - z10 + z5; 388 389 movdqa xmm6, xmm4 390 paddw xmm4, xmm0 391 pmulhw xmm4, [GOTOFF(ebx,PW_F1847)] ; xmm4=z5 392 pmulhw xmm6, [GOTOFF(ebx,PW_MF1613)] 393 pmulhw xmm0, [GOTOFF(ebx,PW_F1082)] 394 psubw xmm6, xmm7 395 psubw xmm0, xmm4 ; xmm0=tmp10 396 paddw xmm6, xmm4 ; xmm6=tmp12 397 398 ; -- Final output stage 399 400 psubw xmm6, xmm5 ; xmm6=tmp6 401 movdqa xmm7, xmm1 402 movdqa xmm4, xmm3 403 paddw xmm1, xmm5 ; xmm1=data0=(00 10 20 30 40 50 60 70) 404 paddw xmm3, xmm6 ; xmm3=data1=(01 11 21 31 41 51 61 71) 405 psraw xmm1, (PASS1_BITS+3) ; descale 406 psraw xmm3, (PASS1_BITS+3) ; descale 407 psubw xmm7, xmm5 ; xmm7=data7=(07 17 27 37 47 57 67 77) 408 psubw xmm4, xmm6 ; xmm4=data6=(06 16 26 36 46 56 66 76) 409 psraw xmm7, (PASS1_BITS+3) ; descale 410 psraw xmm4, (PASS1_BITS+3) ; descale 411 psubw xmm2, xmm6 ; xmm2=tmp5 412 413 packsswb xmm1, xmm4 ; xmm1=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76) 414 packsswb xmm3, xmm7 ; xmm3=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77) 415 416 movdqa xmm5, XMMWORD [wk(1)] ; xmm5=tmp2 417 movdqa xmm6, XMMWORD [wk(0)] ; xmm6=tmp3 418 419 paddw xmm0, xmm2 ; xmm0=tmp4 420 movdqa xmm4, xmm5 421 movdqa xmm7, xmm6 422 paddw xmm5, xmm2 ; xmm5=data2=(02 12 22 32 42 52 62 72) 423 paddw xmm6, xmm0 ; xmm6=data4=(04 14 24 34 44 54 64 74) 424 psraw xmm5, (PASS1_BITS+3) ; descale 425 psraw xmm6, (PASS1_BITS+3) ; descale 426 psubw xmm4, xmm2 ; xmm4=data5=(05 15 25 35 45 55 65 75) 427 psubw xmm7, xmm0 ; xmm7=data3=(03 13 23 33 43 53 63 73) 428 psraw xmm4, (PASS1_BITS+3) ; descale 429 psraw xmm7, (PASS1_BITS+3) ; descale 430 431 movdqa xmm2, [GOTOFF(ebx,PB_CENTERJSAMP)] ; xmm2=[PB_CENTERJSAMP] 432 433 packsswb xmm5, xmm6 ; xmm5=(02 12 22 32 42 52 62 72 04 14 24 34 44 54 64 74) 434 packsswb xmm7, xmm4 ; xmm7=(03 13 23 33 43 53 63 73 05 15 25 35 45 55 65 75) 435 436 paddb xmm1, xmm2 437 paddb xmm3, xmm2 438 paddb xmm5, xmm2 439 paddb xmm7, xmm2 440 441 movdqa xmm0, xmm1 ; transpose coefficients(phase 1) 442 punpcklbw xmm1, xmm3 ; xmm1=(00 01 10 11 20 21 30 31 40 41 50 51 60 61 70 71) 443 punpckhbw xmm0, xmm3 ; xmm0=(06 07 16 17 26 27 36 37 46 47 56 57 66 67 76 77) 444 movdqa xmm6, xmm5 ; transpose coefficients(phase 1) 445 punpcklbw xmm5, xmm7 ; xmm5=(02 03 12 13 22 23 32 33 42 43 52 53 62 63 72 73) 446 punpckhbw xmm6, xmm7 ; xmm6=(04 05 14 15 24 25 34 35 44 45 54 55 64 65 74 75) 447 448 movdqa xmm4, xmm1 ; transpose coefficients(phase 2) 449 punpcklwd xmm1, xmm5 ; xmm1=(00 01 02 03 10 11 12 13 20 21 22 23 30 31 32 33) 450 punpckhwd xmm4, xmm5 ; xmm4=(40 41 42 43 50 51 52 53 60 61 62 63 70 71 72 73) 451 movdqa xmm2, xmm6 ; transpose coefficients(phase 2) 452 punpcklwd xmm6, xmm0 ; xmm6=(04 05 06 07 14 15 16 17 24 25 26 27 34 35 36 37) 453 punpckhwd xmm2, xmm0 ; xmm2=(44 45 46 47 54 55 56 57 64 65 66 67 74 75 76 77) 454 455 movdqa xmm3, xmm1 ; transpose coefficients(phase 3) 456 punpckldq xmm1, xmm6 ; xmm1=(00 01 02 03 04 05 06 07 10 11 12 13 14 15 16 17) 457 punpckhdq xmm3, xmm6 ; xmm3=(20 21 22 23 24 25 26 27 30 31 32 33 34 35 36 37) 458 movdqa xmm7, xmm4 ; transpose coefficients(phase 3) 459 punpckldq xmm4, xmm2 ; xmm4=(40 41 42 43 44 45 46 47 50 51 52 53 54 55 56 57) 460 punpckhdq xmm7, xmm2 ; xmm7=(60 61 62 63 64 65 66 67 70 71 72 73 74 75 76 77) 461 462 pshufd xmm5, xmm1, 0x4E ; xmm5=(10 11 12 13 14 15 16 17 00 01 02 03 04 05 06 07) 463 pshufd xmm0, xmm3, 0x4E ; xmm0=(30 31 32 33 34 35 36 37 20 21 22 23 24 25 26 27) 464 pshufd xmm6, xmm4, 0x4E ; xmm6=(50 51 52 53 54 55 56 57 40 41 42 43 44 45 46 47) 465 pshufd xmm2, xmm7, 0x4E ; xmm2=(70 71 72 73 74 75 76 77 60 61 62 63 64 65 66 67) 466 467 mov edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW] 468 mov esi, JSAMPROW [edi+2*SIZEOF_JSAMPROW] 469 movq XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE], xmm1 470 movq XMM_MMWORD [esi+eax*SIZEOF_JSAMPLE], xmm3 471 mov edx, JSAMPROW [edi+4*SIZEOF_JSAMPROW] 472 mov esi, JSAMPROW [edi+6*SIZEOF_JSAMPROW] 473 movq XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE], xmm4 474 movq XMM_MMWORD [esi+eax*SIZEOF_JSAMPLE], xmm7 475 476 mov edx, JSAMPROW [edi+1*SIZEOF_JSAMPROW] 477 mov esi, JSAMPROW [edi+3*SIZEOF_JSAMPROW] 478 movq XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE], xmm5 479 movq XMM_MMWORD [esi+eax*SIZEOF_JSAMPLE], xmm0 480 mov edx, JSAMPROW [edi+5*SIZEOF_JSAMPROW] 481 mov esi, JSAMPROW [edi+7*SIZEOF_JSAMPROW] 482 movq XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE], xmm6 483 movq XMM_MMWORD [esi+eax*SIZEOF_JSAMPLE], xmm2 484 485 pop edi 486 pop esi 487 ; pop edx ; need not be preserved 488 ; pop ecx ; unused 489 POPPIC ebx 490 mov esp, ebp ; esp <- aligned ebp 491 pop esp ; esp <- original ebp 492 pop ebp 493 ret 494 495 ; For some reason, the OS X linker does not honor the request to align the 496 ; segment unless we do this. 497 align 32