1; 2; jfdctint.asm - accurate integer FDCT (64-bit SSE2) 3; 4; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB 5; Copyright (C) 2009, 2016, 2020, 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; This file contains a slower but more accurate integer implementation of the 18; forward DCT (Discrete Cosine Transform). The following code is based 19; directly on the IJG's original jfdctint.c; see the jfdctint.c for 20; more details. 21 22%include "jsimdext.inc" 23%include "jdct.inc" 24 25; -------------------------------------------------------------------------- 26 27%define CONST_BITS 13 28%define PASS1_BITS 2 29 30%define DESCALE_P1 (CONST_BITS - PASS1_BITS) 31%define DESCALE_P2 (CONST_BITS + PASS1_BITS) 32 33%if CONST_BITS == 13 34F_0_298 equ 2446 ; FIX(0.298631336) 35F_0_390 equ 3196 ; FIX(0.390180644) 36F_0_541 equ 4433 ; FIX(0.541196100) 37F_0_765 equ 6270 ; FIX(0.765366865) 38F_0_899 equ 7373 ; FIX(0.899976223) 39F_1_175 equ 9633 ; FIX(1.175875602) 40F_1_501 equ 12299 ; FIX(1.501321110) 41F_1_847 equ 15137 ; FIX(1.847759065) 42F_1_961 equ 16069 ; FIX(1.961570560) 43F_2_053 equ 16819 ; FIX(2.053119869) 44F_2_562 equ 20995 ; FIX(2.562915447) 45F_3_072 equ 25172 ; FIX(3.072711026) 46%else 47; NASM cannot do compile-time arithmetic on floating-point constants. 48%define DESCALE(x, n) (((x) + (1 << ((n) - 1))) >> (n)) 49F_0_298 equ DESCALE( 320652955, 30 - CONST_BITS) ; FIX(0.298631336) 50F_0_390 equ DESCALE( 418953276, 30 - CONST_BITS) ; FIX(0.390180644) 51F_0_541 equ DESCALE( 581104887, 30 - CONST_BITS) ; FIX(0.541196100) 52F_0_765 equ DESCALE( 821806413, 30 - CONST_BITS) ; FIX(0.765366865) 53F_0_899 equ DESCALE( 966342111, 30 - CONST_BITS) ; FIX(0.899976223) 54F_1_175 equ DESCALE(1262586813, 30 - CONST_BITS) ; FIX(1.175875602) 55F_1_501 equ DESCALE(1612031267, 30 - CONST_BITS) ; FIX(1.501321110) 56F_1_847 equ DESCALE(1984016188, 30 - CONST_BITS) ; FIX(1.847759065) 57F_1_961 equ DESCALE(2106220350, 30 - CONST_BITS) ; FIX(1.961570560) 58F_2_053 equ DESCALE(2204520673, 30 - CONST_BITS) ; FIX(2.053119869) 59F_2_562 equ DESCALE(2751909506, 30 - CONST_BITS) ; FIX(2.562915447) 60F_3_072 equ DESCALE(3299298341, 30 - CONST_BITS) ; FIX(3.072711026) 61%endif 62 63; -------------------------------------------------------------------------- 64 SECTION SEG_CONST 65 66 alignz 32 67 GLOBAL_DATA(jconst_fdct_islow_sse2) 68 69EXTN(jconst_fdct_islow_sse2): 70 71PW_F130_F054 times 4 dw (F_0_541 + F_0_765), F_0_541 72PW_F054_MF130 times 4 dw F_0_541, (F_0_541 - F_1_847) 73PW_MF078_F117 times 4 dw (F_1_175 - F_1_961), F_1_175 74PW_F117_F078 times 4 dw F_1_175, (F_1_175 - F_0_390) 75PW_MF060_MF089 times 4 dw (F_0_298 - F_0_899), -F_0_899 76PW_MF089_F060 times 4 dw -F_0_899, (F_1_501 - F_0_899) 77PW_MF050_MF256 times 4 dw (F_2_053 - F_2_562), -F_2_562 78PW_MF256_F050 times 4 dw -F_2_562, (F_3_072 - F_2_562) 79PD_DESCALE_P1 times 4 dd 1 << (DESCALE_P1 - 1) 80PD_DESCALE_P2 times 4 dd 1 << (DESCALE_P2 - 1) 81PW_DESCALE_P2X times 8 dw 1 << (PASS1_BITS - 1) 82 83 alignz 32 84 85; -------------------------------------------------------------------------- 86 SECTION SEG_TEXT 87 BITS 64 88; 89; Perform the forward DCT on one block of samples. 90; 91; GLOBAL(void) 92; jsimd_fdct_islow_sse2(DCTELEM *data) 93; 94 95; r10 = DCTELEM *data 96 97%define wk(i) rbp - (WK_NUM - (i)) * SIZEOF_XMMWORD ; xmmword wk[WK_NUM] 98%define WK_NUM 6 99 100 align 32 101 GLOBAL_FUNCTION(jsimd_fdct_islow_sse2) 102 103EXTN(jsimd_fdct_islow_sse2): 104 push rbp 105 mov rax, rsp ; rax = original rbp 106 sub rsp, byte 4 107 and rsp, byte (-SIZEOF_XMMWORD) ; align to 128 bits 108 mov [rsp], rax 109 mov rbp, rsp ; rbp = aligned rbp 110 lea rsp, [wk(0)] 111 collect_args 1 112 113 ; ---- Pass 1: process rows. 114 115 mov rdx, r10 ; (DCTELEM *) 116 117 movdqa xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_DCTELEM)] 118 movdqa xmm1, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_DCTELEM)] 119 movdqa xmm2, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_DCTELEM)] 120 movdqa xmm3, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_DCTELEM)] 121 122 ; xmm0=(00 01 02 03 04 05 06 07), xmm2=(20 21 22 23 24 25 26 27) 123 ; xmm1=(10 11 12 13 14 15 16 17), xmm3=(30 31 32 33 34 35 36 37) 124 125 movdqa xmm4, xmm0 ; transpose coefficients(phase 1) 126 punpcklwd xmm0, xmm1 ; xmm0=(00 10 01 11 02 12 03 13) 127 punpckhwd xmm4, xmm1 ; xmm4=(04 14 05 15 06 16 07 17) 128 movdqa xmm5, xmm2 ; transpose coefficients(phase 1) 129 punpcklwd xmm2, xmm3 ; xmm2=(20 30 21 31 22 32 23 33) 130 punpckhwd xmm5, xmm3 ; xmm5=(24 34 25 35 26 36 27 37) 131 132 movdqa xmm6, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_DCTELEM)] 133 movdqa xmm7, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_DCTELEM)] 134 movdqa xmm1, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_DCTELEM)] 135 movdqa xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_DCTELEM)] 136 137 ; xmm6=( 4 12 20 28 36 44 52 60), xmm1=( 6 14 22 30 38 46 54 62) 138 ; xmm7=( 5 13 21 29 37 45 53 61), xmm3=( 7 15 23 31 39 47 55 63) 139 140 movdqa XMMWORD [wk(0)], xmm2 ; wk(0)=(20 30 21 31 22 32 23 33) 141 movdqa XMMWORD [wk(1)], xmm5 ; wk(1)=(24 34 25 35 26 36 27 37) 142 143 movdqa xmm2, xmm6 ; transpose coefficients(phase 1) 144 punpcklwd xmm6, xmm7 ; xmm6=(40 50 41 51 42 52 43 53) 145 punpckhwd xmm2, xmm7 ; xmm2=(44 54 45 55 46 56 47 57) 146 movdqa xmm5, xmm1 ; transpose coefficients(phase 1) 147 punpcklwd xmm1, xmm3 ; xmm1=(60 70 61 71 62 72 63 73) 148 punpckhwd xmm5, xmm3 ; xmm5=(64 74 65 75 66 76 67 77) 149 150 movdqa xmm7, xmm6 ; transpose coefficients(phase 2) 151 punpckldq xmm6, xmm1 ; xmm6=(40 50 60 70 41 51 61 71) 152 punpckhdq xmm7, xmm1 ; xmm7=(42 52 62 72 43 53 63 73) 153 movdqa xmm3, xmm2 ; transpose coefficients(phase 2) 154 punpckldq xmm2, xmm5 ; xmm2=(44 54 64 74 45 55 65 75) 155 punpckhdq xmm3, xmm5 ; xmm3=(46 56 66 76 47 57 67 77) 156 157 movdqa xmm1, XMMWORD [wk(0)] ; xmm1=(20 30 21 31 22 32 23 33) 158 movdqa xmm5, XMMWORD [wk(1)] ; xmm5=(24 34 25 35 26 36 27 37) 159 movdqa XMMWORD [wk(2)], xmm7 ; wk(2)=(42 52 62 72 43 53 63 73) 160 movdqa XMMWORD [wk(3)], xmm2 ; wk(3)=(44 54 64 74 45 55 65 75) 161 162 movdqa xmm7, xmm0 ; transpose coefficients(phase 2) 163 punpckldq xmm0, xmm1 ; xmm0=(00 10 20 30 01 11 21 31) 164 punpckhdq xmm7, xmm1 ; xmm7=(02 12 22 32 03 13 23 33) 165 movdqa xmm2, xmm4 ; transpose coefficients(phase 2) 166 punpckldq xmm4, xmm5 ; xmm4=(04 14 24 34 05 15 25 35) 167 punpckhdq xmm2, xmm5 ; xmm2=(06 16 26 36 07 17 27 37) 168 169 movdqa xmm1, xmm0 ; transpose coefficients(phase 3) 170 punpcklqdq xmm0, xmm6 ; xmm0=(00 10 20 30 40 50 60 70)=data0 171 punpckhqdq xmm1, xmm6 ; xmm1=(01 11 21 31 41 51 61 71)=data1 172 movdqa xmm5, xmm2 ; transpose coefficients(phase 3) 173 punpcklqdq xmm2, xmm3 ; xmm2=(06 16 26 36 46 56 66 76)=data6 174 punpckhqdq xmm5, xmm3 ; xmm5=(07 17 27 37 47 57 67 77)=data7 175 176 movdqa xmm6, xmm1 177 movdqa xmm3, xmm0 178 psubw xmm1, xmm2 ; xmm1=data1-data6=tmp6 179 psubw xmm0, xmm5 ; xmm0=data0-data7=tmp7 180 paddw xmm6, xmm2 ; xmm6=data1+data6=tmp1 181 paddw xmm3, xmm5 ; xmm3=data0+data7=tmp0 182 183 movdqa xmm2, XMMWORD [wk(2)] ; xmm2=(42 52 62 72 43 53 63 73) 184 movdqa xmm5, XMMWORD [wk(3)] ; xmm5=(44 54 64 74 45 55 65 75) 185 movdqa XMMWORD [wk(0)], xmm1 ; wk(0)=tmp6 186 movdqa XMMWORD [wk(1)], xmm0 ; wk(1)=tmp7 187 188 movdqa xmm1, xmm7 ; transpose coefficients(phase 3) 189 punpcklqdq xmm7, xmm2 ; xmm7=(02 12 22 32 42 52 62 72)=data2 190 punpckhqdq xmm1, xmm2 ; xmm1=(03 13 23 33 43 53 63 73)=data3 191 movdqa xmm0, xmm4 ; transpose coefficients(phase 3) 192 punpcklqdq xmm4, xmm5 ; xmm4=(04 14 24 34 44 54 64 74)=data4 193 punpckhqdq xmm0, xmm5 ; xmm0=(05 15 25 35 45 55 65 75)=data5 194 195 movdqa xmm2, xmm1 196 movdqa xmm5, xmm7 197 paddw xmm1, xmm4 ; xmm1=data3+data4=tmp3 198 paddw xmm7, xmm0 ; xmm7=data2+data5=tmp2 199 psubw xmm2, xmm4 ; xmm2=data3-data4=tmp4 200 psubw xmm5, xmm0 ; xmm5=data2-data5=tmp5 201 202 ; -- Even part 203 204 movdqa xmm4, xmm3 205 movdqa xmm0, xmm6 206 paddw xmm3, xmm1 ; xmm3=tmp10 207 paddw xmm6, xmm7 ; xmm6=tmp11 208 psubw xmm4, xmm1 ; xmm4=tmp13 209 psubw xmm0, xmm7 ; xmm0=tmp12 210 211 movdqa xmm1, xmm3 212 paddw xmm3, xmm6 ; xmm3=tmp10+tmp11 213 psubw xmm1, xmm6 ; xmm1=tmp10-tmp11 214 215 psllw xmm3, PASS1_BITS ; xmm3=data0 216 psllw xmm1, PASS1_BITS ; xmm1=data4 217 218 movdqa XMMWORD [wk(2)], xmm3 ; wk(2)=data0 219 movdqa XMMWORD [wk(3)], xmm1 ; wk(3)=data4 220 221 ; (Original) 222 ; z1 = (tmp12 + tmp13) * 0.541196100; 223 ; data2 = z1 + tmp13 * 0.765366865; 224 ; data6 = z1 + tmp12 * -1.847759065; 225 ; 226 ; (This implementation) 227 ; data2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100; 228 ; data6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065); 229 230 movdqa xmm7, xmm4 ; xmm4=tmp13 231 movdqa xmm6, xmm4 232 punpcklwd xmm7, xmm0 ; xmm0=tmp12 233 punpckhwd xmm6, xmm0 234 movdqa xmm4, xmm7 235 movdqa xmm0, xmm6 236 pmaddwd xmm7, [rel PW_F130_F054] ; xmm7=data2L 237 pmaddwd xmm6, [rel PW_F130_F054] ; xmm6=data2H 238 pmaddwd xmm4, [rel PW_F054_MF130] ; xmm4=data6L 239 pmaddwd xmm0, [rel PW_F054_MF130] ; xmm0=data6H 240 241 paddd xmm7, [rel PD_DESCALE_P1] 242 paddd xmm6, [rel PD_DESCALE_P1] 243 psrad xmm7, DESCALE_P1 244 psrad xmm6, DESCALE_P1 245 paddd xmm4, [rel PD_DESCALE_P1] 246 paddd xmm0, [rel PD_DESCALE_P1] 247 psrad xmm4, DESCALE_P1 248 psrad xmm0, DESCALE_P1 249 250 packssdw xmm7, xmm6 ; xmm7=data2 251 packssdw xmm4, xmm0 ; xmm4=data6 252 253 movdqa XMMWORD [wk(4)], xmm7 ; wk(4)=data2 254 movdqa XMMWORD [wk(5)], xmm4 ; wk(5)=data6 255 256 ; -- Odd part 257 258 movdqa xmm3, XMMWORD [wk(0)] ; xmm3=tmp6 259 movdqa xmm1, XMMWORD [wk(1)] ; xmm1=tmp7 260 261 movdqa xmm6, xmm2 ; xmm2=tmp4 262 movdqa xmm0, xmm5 ; xmm5=tmp5 263 paddw xmm6, xmm3 ; xmm6=z3 264 paddw xmm0, xmm1 ; xmm0=z4 265 266 ; (Original) 267 ; z5 = (z3 + z4) * 1.175875602; 268 ; z3 = z3 * -1.961570560; z4 = z4 * -0.390180644; 269 ; z3 += z5; z4 += z5; 270 ; 271 ; (This implementation) 272 ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602; 273 ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644); 274 275 movdqa xmm7, xmm6 276 movdqa xmm4, xmm6 277 punpcklwd xmm7, xmm0 278 punpckhwd xmm4, xmm0 279 movdqa xmm6, xmm7 280 movdqa xmm0, xmm4 281 pmaddwd xmm7, [rel PW_MF078_F117] ; xmm7=z3L 282 pmaddwd xmm4, [rel PW_MF078_F117] ; xmm4=z3H 283 pmaddwd xmm6, [rel PW_F117_F078] ; xmm6=z4L 284 pmaddwd xmm0, [rel PW_F117_F078] ; xmm0=z4H 285 286 movdqa XMMWORD [wk(0)], xmm7 ; wk(0)=z3L 287 movdqa XMMWORD [wk(1)], xmm4 ; wk(1)=z3H 288 289 ; (Original) 290 ; z1 = tmp4 + tmp7; z2 = tmp5 + tmp6; 291 ; tmp4 = tmp4 * 0.298631336; tmp5 = tmp5 * 2.053119869; 292 ; tmp6 = tmp6 * 3.072711026; tmp7 = tmp7 * 1.501321110; 293 ; z1 = z1 * -0.899976223; z2 = z2 * -2.562915447; 294 ; data7 = tmp4 + z1 + z3; data5 = tmp5 + z2 + z4; 295 ; data3 = tmp6 + z2 + z3; data1 = tmp7 + z1 + z4; 296 ; 297 ; (This implementation) 298 ; tmp4 = tmp4 * (0.298631336 - 0.899976223) + tmp7 * -0.899976223; 299 ; tmp5 = tmp5 * (2.053119869 - 2.562915447) + tmp6 * -2.562915447; 300 ; tmp6 = tmp5 * -2.562915447 + tmp6 * (3.072711026 - 2.562915447); 301 ; tmp7 = tmp4 * -0.899976223 + tmp7 * (1.501321110 - 0.899976223); 302 ; data7 = tmp4 + z3; data5 = tmp5 + z4; 303 ; data3 = tmp6 + z3; data1 = tmp7 + z4; 304 305 movdqa xmm7, xmm2 306 movdqa xmm4, xmm2 307 punpcklwd xmm7, xmm1 308 punpckhwd xmm4, xmm1 309 movdqa xmm2, xmm7 310 movdqa xmm1, xmm4 311 pmaddwd xmm7, [rel PW_MF060_MF089] ; xmm7=tmp4L 312 pmaddwd xmm4, [rel PW_MF060_MF089] ; xmm4=tmp4H 313 pmaddwd xmm2, [rel PW_MF089_F060] ; xmm2=tmp7L 314 pmaddwd xmm1, [rel PW_MF089_F060] ; xmm1=tmp7H 315 316 paddd xmm7, XMMWORD [wk(0)] ; xmm7=data7L 317 paddd xmm4, XMMWORD [wk(1)] ; xmm4=data7H 318 paddd xmm2, xmm6 ; xmm2=data1L 319 paddd xmm1, xmm0 ; xmm1=data1H 320 321 paddd xmm7, [rel PD_DESCALE_P1] 322 paddd xmm4, [rel PD_DESCALE_P1] 323 psrad xmm7, DESCALE_P1 324 psrad xmm4, DESCALE_P1 325 paddd xmm2, [rel PD_DESCALE_P1] 326 paddd xmm1, [rel PD_DESCALE_P1] 327 psrad xmm2, DESCALE_P1 328 psrad xmm1, DESCALE_P1 329 330 packssdw xmm7, xmm4 ; xmm7=data7 331 packssdw xmm2, xmm1 ; xmm2=data1 332 333 movdqa xmm4, xmm5 334 movdqa xmm1, xmm5 335 punpcklwd xmm4, xmm3 336 punpckhwd xmm1, xmm3 337 movdqa xmm5, xmm4 338 movdqa xmm3, xmm1 339 pmaddwd xmm4, [rel PW_MF050_MF256] ; xmm4=tmp5L 340 pmaddwd xmm1, [rel PW_MF050_MF256] ; xmm1=tmp5H 341 pmaddwd xmm5, [rel PW_MF256_F050] ; xmm5=tmp6L 342 pmaddwd xmm3, [rel PW_MF256_F050] ; xmm3=tmp6H 343 344 paddd xmm4, xmm6 ; xmm4=data5L 345 paddd xmm1, xmm0 ; xmm1=data5H 346 paddd xmm5, XMMWORD [wk(0)] ; xmm5=data3L 347 paddd xmm3, XMMWORD [wk(1)] ; xmm3=data3H 348 349 paddd xmm4, [rel PD_DESCALE_P1] 350 paddd xmm1, [rel PD_DESCALE_P1] 351 psrad xmm4, DESCALE_P1 352 psrad xmm1, DESCALE_P1 353 paddd xmm5, [rel PD_DESCALE_P1] 354 paddd xmm3, [rel PD_DESCALE_P1] 355 psrad xmm5, DESCALE_P1 356 psrad xmm3, DESCALE_P1 357 358 packssdw xmm4, xmm1 ; xmm4=data5 359 packssdw xmm5, xmm3 ; xmm5=data3 360 361 ; ---- Pass 2: process columns. 362 363 movdqa xmm6, XMMWORD [wk(2)] ; xmm6=col0 364 movdqa xmm0, XMMWORD [wk(4)] ; xmm0=col2 365 366 ; xmm6=(00 10 20 30 40 50 60 70), xmm0=(02 12 22 32 42 52 62 72) 367 ; xmm2=(01 11 21 31 41 51 61 71), xmm5=(03 13 23 33 43 53 63 73) 368 369 movdqa xmm1, xmm6 ; transpose coefficients(phase 1) 370 punpcklwd xmm6, xmm2 ; xmm6=(00 01 10 11 20 21 30 31) 371 punpckhwd xmm1, xmm2 ; xmm1=(40 41 50 51 60 61 70 71) 372 movdqa xmm3, xmm0 ; transpose coefficients(phase 1) 373 punpcklwd xmm0, xmm5 ; xmm0=(02 03 12 13 22 23 32 33) 374 punpckhwd xmm3, xmm5 ; xmm3=(42 43 52 53 62 63 72 73) 375 376 movdqa xmm2, XMMWORD [wk(3)] ; xmm2=col4 377 movdqa xmm5, XMMWORD [wk(5)] ; xmm5=col6 378 379 ; xmm2=(04 14 24 34 44 54 64 74), xmm5=(06 16 26 36 46 56 66 76) 380 ; xmm4=(05 15 25 35 45 55 65 75), xmm7=(07 17 27 37 47 57 67 77) 381 382 movdqa XMMWORD [wk(0)], xmm0 ; wk(0)=(02 03 12 13 22 23 32 33) 383 movdqa XMMWORD [wk(1)], xmm3 ; wk(1)=(42 43 52 53 62 63 72 73) 384 385 movdqa xmm0, xmm2 ; transpose coefficients(phase 1) 386 punpcklwd xmm2, xmm4 ; xmm2=(04 05 14 15 24 25 34 35) 387 punpckhwd xmm0, xmm4 ; xmm0=(44 45 54 55 64 65 74 75) 388 movdqa xmm3, xmm5 ; transpose coefficients(phase 1) 389 punpcklwd xmm5, xmm7 ; xmm5=(06 07 16 17 26 27 36 37) 390 punpckhwd xmm3, xmm7 ; xmm3=(46 47 56 57 66 67 76 77) 391 392 movdqa xmm4, xmm2 ; transpose coefficients(phase 2) 393 punpckldq xmm2, xmm5 ; xmm2=(04 05 06 07 14 15 16 17) 394 punpckhdq xmm4, xmm5 ; xmm4=(24 25 26 27 34 35 36 37) 395 movdqa xmm7, xmm0 ; transpose coefficients(phase 2) 396 punpckldq xmm0, xmm3 ; xmm0=(44 45 46 47 54 55 56 57) 397 punpckhdq xmm7, xmm3 ; xmm7=(64 65 66 67 74 75 76 77) 398 399 movdqa xmm5, XMMWORD [wk(0)] ; xmm5=(02 03 12 13 22 23 32 33) 400 movdqa xmm3, XMMWORD [wk(1)] ; xmm3=(42 43 52 53 62 63 72 73) 401 movdqa XMMWORD [wk(2)], xmm4 ; wk(2)=(24 25 26 27 34 35 36 37) 402 movdqa XMMWORD [wk(3)], xmm0 ; wk(3)=(44 45 46 47 54 55 56 57) 403 404 movdqa xmm4, xmm6 ; transpose coefficients(phase 2) 405 punpckldq xmm6, xmm5 ; xmm6=(00 01 02 03 10 11 12 13) 406 punpckhdq xmm4, xmm5 ; xmm4=(20 21 22 23 30 31 32 33) 407 movdqa xmm0, xmm1 ; transpose coefficients(phase 2) 408 punpckldq xmm1, xmm3 ; xmm1=(40 41 42 43 50 51 52 53) 409 punpckhdq xmm0, xmm3 ; xmm0=(60 61 62 63 70 71 72 73) 410 411 movdqa xmm5, xmm6 ; transpose coefficients(phase 3) 412 punpcklqdq xmm6, xmm2 ; xmm6=(00 01 02 03 04 05 06 07)=data0 413 punpckhqdq xmm5, xmm2 ; xmm5=(10 11 12 13 14 15 16 17)=data1 414 movdqa xmm3, xmm0 ; transpose coefficients(phase 3) 415 punpcklqdq xmm0, xmm7 ; xmm0=(60 61 62 63 64 65 66 67)=data6 416 punpckhqdq xmm3, xmm7 ; xmm3=(70 71 72 73 74 75 76 77)=data7 417 418 movdqa xmm2, xmm5 419 movdqa xmm7, xmm6 420 psubw xmm5, xmm0 ; xmm5=data1-data6=tmp6 421 psubw xmm6, xmm3 ; xmm6=data0-data7=tmp7 422 paddw xmm2, xmm0 ; xmm2=data1+data6=tmp1 423 paddw xmm7, xmm3 ; xmm7=data0+data7=tmp0 424 425 movdqa xmm0, XMMWORD [wk(2)] ; xmm0=(24 25 26 27 34 35 36 37) 426 movdqa xmm3, XMMWORD [wk(3)] ; xmm3=(44 45 46 47 54 55 56 57) 427 movdqa XMMWORD [wk(0)], xmm5 ; wk(0)=tmp6 428 movdqa XMMWORD [wk(1)], xmm6 ; wk(1)=tmp7 429 430 movdqa xmm5, xmm4 ; transpose coefficients(phase 3) 431 punpcklqdq xmm4, xmm0 ; xmm4=(20 21 22 23 24 25 26 27)=data2 432 punpckhqdq xmm5, xmm0 ; xmm5=(30 31 32 33 34 35 36 37)=data3 433 movdqa xmm6, xmm1 ; transpose coefficients(phase 3) 434 punpcklqdq xmm1, xmm3 ; xmm1=(40 41 42 43 44 45 46 47)=data4 435 punpckhqdq xmm6, xmm3 ; xmm6=(50 51 52 53 54 55 56 57)=data5 436 437 movdqa xmm0, xmm5 438 movdqa xmm3, xmm4 439 paddw xmm5, xmm1 ; xmm5=data3+data4=tmp3 440 paddw xmm4, xmm6 ; xmm4=data2+data5=tmp2 441 psubw xmm0, xmm1 ; xmm0=data3-data4=tmp4 442 psubw xmm3, xmm6 ; xmm3=data2-data5=tmp5 443 444 ; -- Even part 445 446 movdqa xmm1, xmm7 447 movdqa xmm6, xmm2 448 paddw xmm7, xmm5 ; xmm7=tmp10 449 paddw xmm2, xmm4 ; xmm2=tmp11 450 psubw xmm1, xmm5 ; xmm1=tmp13 451 psubw xmm6, xmm4 ; xmm6=tmp12 452 453 movdqa xmm5, xmm7 454 paddw xmm7, xmm2 ; xmm7=tmp10+tmp11 455 psubw xmm5, xmm2 ; xmm5=tmp10-tmp11 456 457 paddw xmm7, [rel PW_DESCALE_P2X] 458 paddw xmm5, [rel PW_DESCALE_P2X] 459 psraw xmm7, PASS1_BITS ; xmm7=data0 460 psraw xmm5, PASS1_BITS ; xmm5=data4 461 462 movdqa XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_DCTELEM)], xmm7 463 movdqa XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_DCTELEM)], xmm5 464 465 ; (Original) 466 ; z1 = (tmp12 + tmp13) * 0.541196100; 467 ; data2 = z1 + tmp13 * 0.765366865; 468 ; data6 = z1 + tmp12 * -1.847759065; 469 ; 470 ; (This implementation) 471 ; data2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100; 472 ; data6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065); 473 474 movdqa xmm4, xmm1 ; xmm1=tmp13 475 movdqa xmm2, xmm1 476 punpcklwd xmm4, xmm6 ; xmm6=tmp12 477 punpckhwd xmm2, xmm6 478 movdqa xmm1, xmm4 479 movdqa xmm6, xmm2 480 pmaddwd xmm4, [rel PW_F130_F054] ; xmm4=data2L 481 pmaddwd xmm2, [rel PW_F130_F054] ; xmm2=data2H 482 pmaddwd xmm1, [rel PW_F054_MF130] ; xmm1=data6L 483 pmaddwd xmm6, [rel PW_F054_MF130] ; xmm6=data6H 484 485 paddd xmm4, [rel PD_DESCALE_P2] 486 paddd xmm2, [rel PD_DESCALE_P2] 487 psrad xmm4, DESCALE_P2 488 psrad xmm2, DESCALE_P2 489 paddd xmm1, [rel PD_DESCALE_P2] 490 paddd xmm6, [rel PD_DESCALE_P2] 491 psrad xmm1, DESCALE_P2 492 psrad xmm6, DESCALE_P2 493 494 packssdw xmm4, xmm2 ; xmm4=data2 495 packssdw xmm1, xmm6 ; xmm1=data6 496 497 movdqa XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_DCTELEM)], xmm4 498 movdqa XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_DCTELEM)], xmm1 499 500 ; -- Odd part 501 502 movdqa xmm7, XMMWORD [wk(0)] ; xmm7=tmp6 503 movdqa xmm5, XMMWORD [wk(1)] ; xmm5=tmp7 504 505 movdqa xmm2, xmm0 ; xmm0=tmp4 506 movdqa xmm6, xmm3 ; xmm3=tmp5 507 paddw xmm2, xmm7 ; xmm2=z3 508 paddw xmm6, xmm5 ; xmm6=z4 509 510 ; (Original) 511 ; z5 = (z3 + z4) * 1.175875602; 512 ; z3 = z3 * -1.961570560; z4 = z4 * -0.390180644; 513 ; z3 += z5; z4 += z5; 514 ; 515 ; (This implementation) 516 ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602; 517 ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644); 518 519 movdqa xmm4, xmm2 520 movdqa xmm1, xmm2 521 punpcklwd xmm4, xmm6 522 punpckhwd xmm1, xmm6 523 movdqa xmm2, xmm4 524 movdqa xmm6, xmm1 525 pmaddwd xmm4, [rel PW_MF078_F117] ; xmm4=z3L 526 pmaddwd xmm1, [rel PW_MF078_F117] ; xmm1=z3H 527 pmaddwd xmm2, [rel PW_F117_F078] ; xmm2=z4L 528 pmaddwd xmm6, [rel PW_F117_F078] ; xmm6=z4H 529 530 movdqa XMMWORD [wk(0)], xmm4 ; wk(0)=z3L 531 movdqa XMMWORD [wk(1)], xmm1 ; wk(1)=z3H 532 533 ; (Original) 534 ; z1 = tmp4 + tmp7; z2 = tmp5 + tmp6; 535 ; tmp4 = tmp4 * 0.298631336; tmp5 = tmp5 * 2.053119869; 536 ; tmp6 = tmp6 * 3.072711026; tmp7 = tmp7 * 1.501321110; 537 ; z1 = z1 * -0.899976223; z2 = z2 * -2.562915447; 538 ; data7 = tmp4 + z1 + z3; data5 = tmp5 + z2 + z4; 539 ; data3 = tmp6 + z2 + z3; data1 = tmp7 + z1 + z4; 540 ; 541 ; (This implementation) 542 ; tmp4 = tmp4 * (0.298631336 - 0.899976223) + tmp7 * -0.899976223; 543 ; tmp5 = tmp5 * (2.053119869 - 2.562915447) + tmp6 * -2.562915447; 544 ; tmp6 = tmp5 * -2.562915447 + tmp6 * (3.072711026 - 2.562915447); 545 ; tmp7 = tmp4 * -0.899976223 + tmp7 * (1.501321110 - 0.899976223); 546 ; data7 = tmp4 + z3; data5 = tmp5 + z4; 547 ; data3 = tmp6 + z3; data1 = tmp7 + z4; 548 549 movdqa xmm4, xmm0 550 movdqa xmm1, xmm0 551 punpcklwd xmm4, xmm5 552 punpckhwd xmm1, xmm5 553 movdqa xmm0, xmm4 554 movdqa xmm5, xmm1 555 pmaddwd xmm4, [rel PW_MF060_MF089] ; xmm4=tmp4L 556 pmaddwd xmm1, [rel PW_MF060_MF089] ; xmm1=tmp4H 557 pmaddwd xmm0, [rel PW_MF089_F060] ; xmm0=tmp7L 558 pmaddwd xmm5, [rel PW_MF089_F060] ; xmm5=tmp7H 559 560 paddd xmm4, XMMWORD [wk(0)] ; xmm4=data7L 561 paddd xmm1, XMMWORD [wk(1)] ; xmm1=data7H 562 paddd xmm0, xmm2 ; xmm0=data1L 563 paddd xmm5, xmm6 ; xmm5=data1H 564 565 paddd xmm4, [rel PD_DESCALE_P2] 566 paddd xmm1, [rel PD_DESCALE_P2] 567 psrad xmm4, DESCALE_P2 568 psrad xmm1, DESCALE_P2 569 paddd xmm0, [rel PD_DESCALE_P2] 570 paddd xmm5, [rel PD_DESCALE_P2] 571 psrad xmm0, DESCALE_P2 572 psrad xmm5, DESCALE_P2 573 574 packssdw xmm4, xmm1 ; xmm4=data7 575 packssdw xmm0, xmm5 ; xmm0=data1 576 577 movdqa XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_DCTELEM)], xmm4 578 movdqa XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_DCTELEM)], xmm0 579 580 movdqa xmm1, xmm3 581 movdqa xmm5, xmm3 582 punpcklwd xmm1, xmm7 583 punpckhwd xmm5, xmm7 584 movdqa xmm3, xmm1 585 movdqa xmm7, xmm5 586 pmaddwd xmm1, [rel PW_MF050_MF256] ; xmm1=tmp5L 587 pmaddwd xmm5, [rel PW_MF050_MF256] ; xmm5=tmp5H 588 pmaddwd xmm3, [rel PW_MF256_F050] ; xmm3=tmp6L 589 pmaddwd xmm7, [rel PW_MF256_F050] ; xmm7=tmp6H 590 591 paddd xmm1, xmm2 ; xmm1=data5L 592 paddd xmm5, xmm6 ; xmm5=data5H 593 paddd xmm3, XMMWORD [wk(0)] ; xmm3=data3L 594 paddd xmm7, XMMWORD [wk(1)] ; xmm7=data3H 595 596 paddd xmm1, [rel PD_DESCALE_P2] 597 paddd xmm5, [rel PD_DESCALE_P2] 598 psrad xmm1, DESCALE_P2 599 psrad xmm5, DESCALE_P2 600 paddd xmm3, [rel PD_DESCALE_P2] 601 paddd xmm7, [rel PD_DESCALE_P2] 602 psrad xmm3, DESCALE_P2 603 psrad xmm7, DESCALE_P2 604 605 packssdw xmm1, xmm5 ; xmm1=data5 606 packssdw xmm3, xmm7 ; xmm3=data3 607 608 movdqa XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_DCTELEM)], xmm1 609 movdqa XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_DCTELEM)], xmm3 610 611 uncollect_args 1 612 mov rsp, rbp ; rsp <- aligned rbp 613 pop rsp ; rsp <- original rbp 614 pop rbp 615 ret 616 617; For some reason, the OS X linker does not honor the request to align the 618; segment unless we do this. 619 align 32 620