jidctint-avx2.asm (18793B)
1 ; 2 ; jidctint.asm - accurate integer IDCT (AVX2) 3 ; 4 ; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB 5 ; Copyright (C) 2009, 2016, 2018, 2020, 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 slower but more accurate integer implementation of the 14 ; inverse DCT (Discrete Cosine Transform). The following code is based 15 ; directly on the IJG's original jidctint.c; see the jidctint.c for 16 ; 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 (CONST_BITS - PASS1_BITS) 27 %define DESCALE_P2 (CONST_BITS + PASS1_BITS + 3) 28 29 %if CONST_BITS == 13 30 F_0_298 equ 2446 ; FIX(0.298631336) 31 F_0_390 equ 3196 ; FIX(0.390180644) 32 F_0_541 equ 4433 ; FIX(0.541196100) 33 F_0_765 equ 6270 ; FIX(0.765366865) 34 F_0_899 equ 7373 ; FIX(0.899976223) 35 F_1_175 equ 9633 ; FIX(1.175875602) 36 F_1_501 equ 12299 ; FIX(1.501321110) 37 F_1_847 equ 15137 ; FIX(1.847759065) 38 F_1_961 equ 16069 ; FIX(1.961570560) 39 F_2_053 equ 16819 ; FIX(2.053119869) 40 F_2_562 equ 20995 ; FIX(2.562915447) 41 F_3_072 equ 25172 ; FIX(3.072711026) 42 %else 43 ; NASM cannot do compile-time arithmetic on floating-point constants. 44 %define DESCALE(x, n) (((x) + (1 << ((n) - 1))) >> (n)) 45 F_0_298 equ DESCALE( 320652955, 30 - CONST_BITS) ; FIX(0.298631336) 46 F_0_390 equ DESCALE( 418953276, 30 - CONST_BITS) ; FIX(0.390180644) 47 F_0_541 equ DESCALE( 581104887, 30 - CONST_BITS) ; FIX(0.541196100) 48 F_0_765 equ DESCALE( 821806413, 30 - CONST_BITS) ; FIX(0.765366865) 49 F_0_899 equ DESCALE( 966342111, 30 - CONST_BITS) ; FIX(0.899976223) 50 F_1_175 equ DESCALE(1262586813, 30 - CONST_BITS) ; FIX(1.175875602) 51 F_1_501 equ DESCALE(1612031267, 30 - CONST_BITS) ; FIX(1.501321110) 52 F_1_847 equ DESCALE(1984016188, 30 - CONST_BITS) ; FIX(1.847759065) 53 F_1_961 equ DESCALE(2106220350, 30 - CONST_BITS) ; FIX(1.961570560) 54 F_2_053 equ DESCALE(2204520673, 30 - CONST_BITS) ; FIX(2.053119869) 55 F_2_562 equ DESCALE(2751909506, 30 - CONST_BITS) ; FIX(2.562915447) 56 F_3_072 equ DESCALE(3299298341, 30 - CONST_BITS) ; FIX(3.072711026) 57 %endif 58 59 ; -------------------------------------------------------------------------- 60 ; In-place 8x8x16-bit inverse matrix transpose using AVX2 instructions 61 ; %1-%4: Input/output registers 62 ; %5-%8: Temp registers 63 64 %macro DOTRANSPOSE 8 65 ; %5=(00 10 20 30 40 50 60 70 01 11 21 31 41 51 61 71) 66 ; %6=(03 13 23 33 43 53 63 73 02 12 22 32 42 52 62 72) 67 ; %7=(04 14 24 34 44 54 64 74 05 15 25 35 45 55 65 75) 68 ; %8=(07 17 27 37 47 57 67 77 06 16 26 36 46 56 66 76) 69 70 vpermq %5, %1, 0xD8 71 vpermq %6, %2, 0x72 72 vpermq %7, %3, 0xD8 73 vpermq %8, %4, 0x72 74 ; transpose coefficients(phase 1) 75 ; %5=(00 10 20 30 01 11 21 31 40 50 60 70 41 51 61 71) 76 ; %6=(02 12 22 32 03 13 23 33 42 52 62 72 43 53 63 73) 77 ; %7=(04 14 24 34 05 15 25 35 44 54 64 74 45 55 65 75) 78 ; %8=(06 16 26 36 07 17 27 37 46 56 66 76 47 57 67 77) 79 80 vpunpcklwd %1, %5, %6 81 vpunpckhwd %2, %5, %6 82 vpunpcklwd %3, %7, %8 83 vpunpckhwd %4, %7, %8 84 ; transpose coefficients(phase 2) 85 ; %1=(00 02 10 12 20 22 30 32 40 42 50 52 60 62 70 72) 86 ; %2=(01 03 11 13 21 23 31 33 41 43 51 53 61 63 71 73) 87 ; %3=(04 06 14 16 24 26 34 36 44 46 54 56 64 66 74 76) 88 ; %4=(05 07 15 17 25 27 35 37 45 47 55 57 65 67 75 77) 89 90 vpunpcklwd %5, %1, %2 91 vpunpcklwd %6, %3, %4 92 vpunpckhwd %7, %1, %2 93 vpunpckhwd %8, %3, %4 94 ; transpose coefficients(phase 3) 95 ; %5=(00 01 02 03 10 11 12 13 40 41 42 43 50 51 52 53) 96 ; %6=(04 05 06 07 14 15 16 17 44 45 46 47 54 55 56 57) 97 ; %7=(20 21 22 23 30 31 32 33 60 61 62 63 70 71 72 73) 98 ; %8=(24 25 26 27 34 35 36 37 64 65 66 67 74 75 76 77) 99 100 vpunpcklqdq %1, %5, %6 101 vpunpckhqdq %2, %5, %6 102 vpunpcklqdq %3, %7, %8 103 vpunpckhqdq %4, %7, %8 104 ; transpose coefficients(phase 4) 105 ; %1=(00 01 02 03 04 05 06 07 40 41 42 43 44 45 46 47) 106 ; %2=(10 11 12 13 14 15 16 17 50 51 52 53 54 55 56 57) 107 ; %3=(20 21 22 23 24 25 26 27 60 61 62 63 64 65 66 67) 108 ; %4=(30 31 32 33 34 35 36 37 70 71 72 73 74 75 76 77) 109 %endmacro 110 111 ; -------------------------------------------------------------------------- 112 ; In-place 8x8x16-bit accurate integer inverse DCT using AVX2 instructions 113 ; %1-%4: Input/output registers 114 ; %5-%12: Temp registers 115 ; %9: Pass (1 or 2) 116 117 %macro DODCT 13 118 ; -- Even part 119 120 ; (Original) 121 ; z1 = (z2 + z3) * 0.541196100; 122 ; tmp2 = z1 + z3 * -1.847759065; 123 ; tmp3 = z1 + z2 * 0.765366865; 124 ; 125 ; (This implementation) 126 ; tmp2 = z2 * 0.541196100 + z3 * (0.541196100 - 1.847759065); 127 ; tmp3 = z2 * (0.541196100 + 0.765366865) + z3 * 0.541196100; 128 129 vperm2i128 %6, %3, %3, 0x01 ; %6=in6_2 130 vpunpcklwd %5, %3, %6 ; %5=in26_62L 131 vpunpckhwd %6, %3, %6 ; %6=in26_62H 132 vpmaddwd %5, %5, [GOTOFF(ebx,PW_F130_F054_MF130_F054)] ; %5=tmp3_2L 133 vpmaddwd %6, %6, [GOTOFF(ebx,PW_F130_F054_MF130_F054)] ; %6=tmp3_2H 134 135 vperm2i128 %7, %1, %1, 0x01 ; %7=in4_0 136 vpsignw %1, %1, [GOTOFF(ebx,PW_1_NEG1)] 137 vpaddw %7, %7, %1 ; %7=(in0+in4)_(in0-in4) 138 139 vpxor %1, %1, %1 140 vpunpcklwd %8, %1, %7 ; %8=tmp0_1L 141 vpunpckhwd %1, %1, %7 ; %1=tmp0_1H 142 vpsrad %8, %8, (16-CONST_BITS) ; vpsrad %8,16 & vpslld %8,CONST_BITS 143 vpsrad %1, %1, (16-CONST_BITS) ; vpsrad %1,16 & vpslld %1,CONST_BITS 144 145 vpsubd %3, %8, %5 146 vmovdqu %11, %3 ; %11=tmp0_1L-tmp3_2L=tmp13_12L 147 vpaddd %3, %8, %5 148 vmovdqu %9, %3 ; %9=tmp0_1L+tmp3_2L=tmp10_11L 149 vpsubd %3, %1, %6 150 vmovdqu %12, %3 ; %12=tmp0_1H-tmp3_2H=tmp13_12H 151 vpaddd %3, %1, %6 152 vmovdqu %10, %3 ; %10=tmp0_1H+tmp3_2H=tmp10_11H 153 154 ; -- Odd part 155 156 vpaddw %1, %4, %2 ; %1=in7_5+in3_1=z3_4 157 158 ; (Original) 159 ; z5 = (z3 + z4) * 1.175875602; 160 ; z3 = z3 * -1.961570560; z4 = z4 * -0.390180644; 161 ; z3 += z5; z4 += z5; 162 ; 163 ; (This implementation) 164 ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602; 165 ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644); 166 167 vperm2i128 %8, %1, %1, 0x01 ; %8=z4_3 168 vpunpcklwd %7, %1, %8 ; %7=z34_43L 169 vpunpckhwd %8, %1, %8 ; %8=z34_43H 170 vpmaddwd %7, %7, [GOTOFF(ebx,PW_MF078_F117_F078_F117)] ; %7=z3_4L 171 vpmaddwd %8, %8, [GOTOFF(ebx,PW_MF078_F117_F078_F117)] ; %8=z3_4H 172 173 ; (Original) 174 ; z1 = tmp0 + tmp3; z2 = tmp1 + tmp2; 175 ; tmp0 = tmp0 * 0.298631336; tmp1 = tmp1 * 2.053119869; 176 ; tmp2 = tmp2 * 3.072711026; tmp3 = tmp3 * 1.501321110; 177 ; z1 = z1 * -0.899976223; z2 = z2 * -2.562915447; 178 ; tmp0 += z1 + z3; tmp1 += z2 + z4; 179 ; tmp2 += z2 + z3; tmp3 += z1 + z4; 180 ; 181 ; (This implementation) 182 ; tmp0 = tmp0 * (0.298631336 - 0.899976223) + tmp3 * -0.899976223; 183 ; tmp1 = tmp1 * (2.053119869 - 2.562915447) + tmp2 * -2.562915447; 184 ; tmp2 = tmp1 * -2.562915447 + tmp2 * (3.072711026 - 2.562915447); 185 ; tmp3 = tmp0 * -0.899976223 + tmp3 * (1.501321110 - 0.899976223); 186 ; tmp0 += z3; tmp1 += z4; 187 ; tmp2 += z3; tmp3 += z4; 188 189 vperm2i128 %2, %2, %2, 0x01 ; %2=in1_3 190 vpunpcklwd %3, %4, %2 ; %3=in71_53L 191 vpunpckhwd %4, %4, %2 ; %4=in71_53H 192 193 vpmaddwd %5, %3, [GOTOFF(ebx,PW_MF060_MF089_MF050_MF256)] ; %5=tmp0_1L 194 vpmaddwd %6, %4, [GOTOFF(ebx,PW_MF060_MF089_MF050_MF256)] ; %6=tmp0_1H 195 vpaddd %5, %5, %7 ; %5=tmp0_1L+z3_4L=tmp0_1L 196 vpaddd %6, %6, %8 ; %6=tmp0_1H+z3_4H=tmp0_1H 197 198 vpmaddwd %3, %3, [GOTOFF(ebx,PW_MF089_F060_MF256_F050)] ; %3=tmp3_2L 199 vpmaddwd %4, %4, [GOTOFF(ebx,PW_MF089_F060_MF256_F050)] ; %4=tmp3_2H 200 vperm2i128 %7, %7, %7, 0x01 ; %7=z4_3L 201 vperm2i128 %8, %8, %8, 0x01 ; %8=z4_3H 202 vpaddd %7, %3, %7 ; %7=tmp3_2L+z4_3L=tmp3_2L 203 vpaddd %8, %4, %8 ; %8=tmp3_2H+z4_3H=tmp3_2H 204 205 ; -- Final output stage 206 207 vmovdqu %3, %9 208 vmovdqu %4, %10 209 210 vpaddd %1, %3, %7 ; %1=tmp10_11L+tmp3_2L=data0_1L 211 vpaddd %2, %4, %8 ; %2=tmp10_11H+tmp3_2H=data0_1H 212 vpaddd %1, %1, [GOTOFF(ebx,PD_DESCALE_P %+ %13)] 213 vpaddd %2, %2, [GOTOFF(ebx,PD_DESCALE_P %+ %13)] 214 vpsrad %1, %1, DESCALE_P %+ %13 215 vpsrad %2, %2, DESCALE_P %+ %13 216 vpackssdw %1, %1, %2 ; %1=data0_1 217 218 vpsubd %3, %3, %7 ; %3=tmp10_11L-tmp3_2L=data7_6L 219 vpsubd %4, %4, %8 ; %4=tmp10_11H-tmp3_2H=data7_6H 220 vpaddd %3, %3, [GOTOFF(ebx,PD_DESCALE_P %+ %13)] 221 vpaddd %4, %4, [GOTOFF(ebx,PD_DESCALE_P %+ %13)] 222 vpsrad %3, %3, DESCALE_P %+ %13 223 vpsrad %4, %4, DESCALE_P %+ %13 224 vpackssdw %4, %3, %4 ; %4=data7_6 225 226 vmovdqu %7, %11 227 vmovdqu %8, %12 228 229 vpaddd %2, %7, %5 ; %7=tmp13_12L+tmp0_1L=data3_2L 230 vpaddd %3, %8, %6 ; %8=tmp13_12H+tmp0_1H=data3_2H 231 vpaddd %2, %2, [GOTOFF(ebx,PD_DESCALE_P %+ %13)] 232 vpaddd %3, %3, [GOTOFF(ebx,PD_DESCALE_P %+ %13)] 233 vpsrad %2, %2, DESCALE_P %+ %13 234 vpsrad %3, %3, DESCALE_P %+ %13 235 vpackssdw %2, %2, %3 ; %2=data3_2 236 237 vpsubd %3, %7, %5 ; %7=tmp13_12L-tmp0_1L=data4_5L 238 vpsubd %6, %8, %6 ; %8=tmp13_12H-tmp0_1H=data4_5H 239 vpaddd %3, %3, [GOTOFF(ebx,PD_DESCALE_P %+ %13)] 240 vpaddd %6, %6, [GOTOFF(ebx,PD_DESCALE_P %+ %13)] 241 vpsrad %3, %3, DESCALE_P %+ %13 242 vpsrad %6, %6, DESCALE_P %+ %13 243 vpackssdw %3, %3, %6 ; %3=data4_5 244 %endmacro 245 246 ; -------------------------------------------------------------------------- 247 SECTION SEG_CONST 248 249 ALIGNZ 32 250 GLOBAL_DATA(jconst_idct_islow_avx2) 251 252 EXTN(jconst_idct_islow_avx2): 253 254 PW_F130_F054_MF130_F054 times 4 dw (F_0_541 + F_0_765), F_0_541 255 times 4 dw (F_0_541 - F_1_847), F_0_541 256 PW_MF078_F117_F078_F117 times 4 dw (F_1_175 - F_1_961), F_1_175 257 times 4 dw (F_1_175 - F_0_390), F_1_175 258 PW_MF060_MF089_MF050_MF256 times 4 dw (F_0_298 - F_0_899), -F_0_899 259 times 4 dw (F_2_053 - F_2_562), -F_2_562 260 PW_MF089_F060_MF256_F050 times 4 dw -F_0_899, (F_1_501 - F_0_899) 261 times 4 dw -F_2_562, (F_3_072 - F_2_562) 262 PD_DESCALE_P1 times 8 dd 1 << (DESCALE_P1 - 1) 263 PD_DESCALE_P2 times 8 dd 1 << (DESCALE_P2 - 1) 264 PB_CENTERJSAMP times 32 db CENTERJSAMPLE 265 PW_1_NEG1 times 8 dw 1 266 times 8 dw -1 267 268 ALIGNZ 32 269 270 ; -------------------------------------------------------------------------- 271 SECTION SEG_TEXT 272 BITS 32 273 ; 274 ; Perform dequantization and inverse DCT on one block of coefficients. 275 ; 276 ; GLOBAL(void) 277 ; jsimd_idct_islow_avx2(void *dct_table, JCOEFPTR coef_block, 278 ; JSAMPARRAY output_buf, JDIMENSION output_col) 279 ; 280 281 %define dct_table(b) (b) + 8 ; jpeg_component_info *compptr 282 %define coef_block(b) (b) + 12 ; JCOEFPTR coef_block 283 %define output_buf(b) (b) + 16 ; JSAMPARRAY output_buf 284 %define output_col(b) (b) + 20 ; JDIMENSION output_col 285 286 %define original_ebp ebp + 0 287 %define wk(i) ebp - (WK_NUM - (i)) * SIZEOF_YMMWORD 288 ; ymmword wk[WK_NUM] 289 %define WK_NUM 4 290 291 align 32 292 GLOBAL_FUNCTION(jsimd_idct_islow_avx2) 293 294 EXTN(jsimd_idct_islow_avx2): 295 push ebp 296 mov eax, esp ; eax = original ebp 297 sub esp, byte 4 298 and esp, byte (-SIZEOF_XMMWORD) ; align to 128 bits 299 mov [esp], eax 300 mov ebp, esp ; ebp = aligned ebp 301 lea esp, [wk(0)] 302 PUSHPIC ebx 303 ; push ecx ; unused 304 ; push edx ; need not be preserved 305 push esi 306 push edi 307 308 GET_GOT ebx ; get GOT address 309 310 ; ---- Pass 1: process columns. 311 312 ; mov eax, [original_ebp] 313 mov edx, POINTER [dct_table(eax)] ; quantptr 314 mov esi, JCOEFPTR [coef_block(eax)] ; inptr 315 316 %ifndef NO_ZERO_COLUMN_TEST_ISLOW_AVX2 317 mov eax, dword [DWBLOCK(1,0,esi,SIZEOF_JCOEF)] 318 or eax, dword [DWBLOCK(2,0,esi,SIZEOF_JCOEF)] 319 jnz near .columnDCT 320 321 movdqa xmm0, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_JCOEF)] 322 movdqa xmm1, XMMWORD [XMMBLOCK(2,0,esi,SIZEOF_JCOEF)] 323 vpor xmm0, xmm0, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_JCOEF)] 324 vpor xmm1, xmm1, XMMWORD [XMMBLOCK(4,0,esi,SIZEOF_JCOEF)] 325 vpor xmm0, xmm0, XMMWORD [XMMBLOCK(5,0,esi,SIZEOF_JCOEF)] 326 vpor xmm1, xmm1, XMMWORD [XMMBLOCK(6,0,esi,SIZEOF_JCOEF)] 327 vpor xmm0, xmm0, XMMWORD [XMMBLOCK(7,0,esi,SIZEOF_JCOEF)] 328 vpor xmm1, xmm1, xmm0 329 vpacksswb xmm1, xmm1, xmm1 330 vpacksswb xmm1, xmm1, xmm1 331 movd eax, xmm1 332 test eax, eax 333 jnz short .columnDCT 334 335 ; -- AC terms all zero 336 337 movdqa xmm5, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_JCOEF)] 338 vpmullw xmm5, xmm5, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)] 339 340 vpsllw xmm5, xmm5, PASS1_BITS 341 342 vpunpcklwd xmm4, xmm5, xmm5 ; xmm4=(00 00 01 01 02 02 03 03) 343 vpunpckhwd xmm5, xmm5, xmm5 ; xmm5=(04 04 05 05 06 06 07 07) 344 vinserti128 ymm4, ymm4, xmm5, 1 345 346 vpshufd ymm0, ymm4, 0x00 ; ymm0=col0_4=(00 00 00 00 00 00 00 00 04 04 04 04 04 04 04 04) 347 vpshufd ymm1, ymm4, 0x55 ; ymm1=col1_5=(01 01 01 01 01 01 01 01 05 05 05 05 05 05 05 05) 348 vpshufd ymm2, ymm4, 0xAA ; ymm2=col2_6=(02 02 02 02 02 02 02 02 06 06 06 06 06 06 06 06) 349 vpshufd ymm3, ymm4, 0xFF ; ymm3=col3_7=(03 03 03 03 03 03 03 03 07 07 07 07 07 07 07 07) 350 351 jmp near .column_end 352 ALIGNX 16, 7 353 %endif 354 .columnDCT: 355 356 vmovdqu ymm4, YMMWORD [YMMBLOCK(0,0,esi,SIZEOF_JCOEF)] ; ymm4=in0_1 357 vmovdqu ymm5, YMMWORD [YMMBLOCK(2,0,esi,SIZEOF_JCOEF)] ; ymm5=in2_3 358 vmovdqu ymm6, YMMWORD [YMMBLOCK(4,0,esi,SIZEOF_JCOEF)] ; ymm6=in4_5 359 vmovdqu ymm7, YMMWORD [YMMBLOCK(6,0,esi,SIZEOF_JCOEF)] ; ymm7=in6_7 360 vpmullw ymm4, ymm4, YMMWORD [YMMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)] 361 vpmullw ymm5, ymm5, YMMWORD [YMMBLOCK(2,0,edx,SIZEOF_ISLOW_MULT_TYPE)] 362 vpmullw ymm6, ymm6, YMMWORD [YMMBLOCK(4,0,edx,SIZEOF_ISLOW_MULT_TYPE)] 363 vpmullw ymm7, ymm7, YMMWORD [YMMBLOCK(6,0,edx,SIZEOF_ISLOW_MULT_TYPE)] 364 365 vperm2i128 ymm0, ymm4, ymm6, 0x20 ; ymm0=in0_4 366 vperm2i128 ymm1, ymm5, ymm4, 0x31 ; ymm1=in3_1 367 vperm2i128 ymm2, ymm5, ymm7, 0x20 ; ymm2=in2_6 368 vperm2i128 ymm3, ymm7, ymm6, 0x31 ; ymm3=in7_5 369 370 DODCT ymm0, ymm1, ymm2, ymm3, ymm4, ymm5, ymm6, ymm7, XMMWORD [wk(0)], XMMWORD [wk(1)], XMMWORD [wk(2)], XMMWORD [wk(3)], 1 371 ; ymm0=data0_1, ymm1=data3_2, ymm2=data4_5, ymm3=data7_6 372 373 DOTRANSPOSE ymm0, ymm1, ymm2, ymm3, ymm4, ymm5, ymm6, ymm7 374 ; ymm0=data0_4, ymm1=data1_5, ymm2=data2_6, ymm3=data3_7 375 376 .column_end: 377 378 ; -- Prefetch the next coefficient block 379 380 prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 0*32] 381 prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 1*32] 382 prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 2*32] 383 prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 3*32] 384 385 ; ---- Pass 2: process rows. 386 387 mov eax, [original_ebp] 388 mov edi, JSAMPARRAY [output_buf(eax)] ; (JSAMPROW *) 389 mov eax, JDIMENSION [output_col(eax)] 390 391 vperm2i128 ymm4, ymm3, ymm1, 0x31 ; ymm3=in7_5 392 vperm2i128 ymm1, ymm3, ymm1, 0x20 ; ymm1=in3_1 393 394 DODCT ymm0, ymm1, ymm2, ymm4, ymm3, ymm5, ymm6, ymm7, XMMWORD [wk(0)], XMMWORD [wk(1)], XMMWORD [wk(2)], XMMWORD [wk(3)], 2 395 ; ymm0=data0_1, ymm1=data3_2, ymm2=data4_5, ymm4=data7_6 396 397 DOTRANSPOSE ymm0, ymm1, ymm2, ymm4, ymm3, ymm5, ymm6, ymm7 398 ; ymm0=data0_4, ymm1=data1_5, ymm2=data2_6, ymm4=data3_7 399 400 vpacksswb ymm0, ymm0, ymm1 ; ymm0=data01_45 401 vpacksswb ymm1, ymm2, ymm4 ; ymm1=data23_67 402 vpaddb ymm0, ymm0, [GOTOFF(ebx,PB_CENTERJSAMP)] 403 vpaddb ymm1, ymm1, [GOTOFF(ebx,PB_CENTERJSAMP)] 404 405 vextracti128 xmm6, ymm1, 1 ; xmm3=data67 406 vextracti128 xmm4, ymm0, 1 ; xmm2=data45 407 vextracti128 xmm2, ymm1, 0 ; xmm1=data23 408 vextracti128 xmm0, ymm0, 0 ; xmm0=data01 409 410 vpshufd xmm1, xmm0, 0x4E ; xmm1=(10 11 12 13 14 15 16 17 00 01 02 03 04 05 06 07) 411 vpshufd xmm3, xmm2, 0x4E ; xmm3=(30 31 32 33 34 35 36 37 20 21 22 23 24 25 26 27) 412 vpshufd xmm5, xmm4, 0x4E ; xmm5=(50 51 52 53 54 55 56 57 40 41 42 43 44 45 46 47) 413 vpshufd xmm7, xmm6, 0x4E ; xmm7=(70 71 72 73 74 75 76 77 60 61 62 63 64 65 66 67) 414 415 vzeroupper 416 417 mov edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW] ; (JSAMPLE *) 418 mov esi, JSAMPROW [edi+1*SIZEOF_JSAMPROW] ; (JSAMPLE *) 419 movq XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE], xmm0 420 movq XMM_MMWORD [esi+eax*SIZEOF_JSAMPLE], xmm1 421 422 mov edx, JSAMPROW [edi+2*SIZEOF_JSAMPROW] ; (JSAMPLE *) 423 mov esi, JSAMPROW [edi+3*SIZEOF_JSAMPROW] ; (JSAMPLE *) 424 movq XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE], xmm2 425 movq XMM_MMWORD [esi+eax*SIZEOF_JSAMPLE], xmm3 426 427 mov edx, JSAMPROW [edi+4*SIZEOF_JSAMPROW] ; (JSAMPLE *) 428 mov esi, JSAMPROW [edi+5*SIZEOF_JSAMPROW] ; (JSAMPLE *) 429 movq XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE], xmm4 430 movq XMM_MMWORD [esi+eax*SIZEOF_JSAMPLE], xmm5 431 432 mov edx, JSAMPROW [edi+6*SIZEOF_JSAMPROW] ; (JSAMPLE *) 433 mov esi, JSAMPROW [edi+7*SIZEOF_JSAMPROW] ; (JSAMPLE *) 434 movq XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE], xmm6 435 movq XMM_MMWORD [esi+eax*SIZEOF_JSAMPLE], xmm7 436 437 pop edi 438 pop esi 439 ; pop edx ; need not be preserved 440 ; pop ecx ; unused 441 POPPIC ebx 442 mov esp, ebp ; esp <- aligned ebp 443 pop esp ; esp <- original ebp 444 pop ebp 445 ret 446 447 ; For some reason, the OS X linker does not honor the request to align the 448 ; segment unless we do this. 449 align 32