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, 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 slow-but-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; [TAB8] 23 24%include "jsimdext.inc" 25%include "jdct.inc" 26 27; -------------------------------------------------------------------------- 28 29%define CONST_BITS 13 30%define PASS1_BITS 2 31 32%define DESCALE_P1 (CONST_BITS-PASS1_BITS) 33%define DESCALE_P2 (CONST_BITS+PASS1_BITS) 34 35%if CONST_BITS == 13 36F_0_298 equ 2446 ; FIX(0.298631336) 37F_0_390 equ 3196 ; FIX(0.390180644) 38F_0_541 equ 4433 ; FIX(0.541196100) 39F_0_765 equ 6270 ; FIX(0.765366865) 40F_0_899 equ 7373 ; FIX(0.899976223) 41F_1_175 equ 9633 ; FIX(1.175875602) 42F_1_501 equ 12299 ; FIX(1.501321110) 43F_1_847 equ 15137 ; FIX(1.847759065) 44F_1_961 equ 16069 ; FIX(1.961570560) 45F_2_053 equ 16819 ; FIX(2.053119869) 46F_2_562 equ 20995 ; FIX(2.562915447) 47F_3_072 equ 25172 ; FIX(3.072711026) 48%else 49; NASM cannot do compile-time arithmetic on floating-point constants. 50%define DESCALE(x,n) (((x)+(1<<((n)-1)))>>(n)) 51F_0_298 equ DESCALE( 320652955,30-CONST_BITS) ; FIX(0.298631336) 52F_0_390 equ DESCALE( 418953276,30-CONST_BITS) ; FIX(0.390180644) 53F_0_541 equ DESCALE( 581104887,30-CONST_BITS) ; FIX(0.541196100) 54F_0_765 equ DESCALE( 821806413,30-CONST_BITS) ; FIX(0.765366865) 55F_0_899 equ DESCALE( 966342111,30-CONST_BITS) ; FIX(0.899976223) 56F_1_175 equ DESCALE(1262586813,30-CONST_BITS) ; FIX(1.175875602) 57F_1_501 equ DESCALE(1612031267,30-CONST_BITS) ; FIX(1.501321110) 58F_1_847 equ DESCALE(1984016188,30-CONST_BITS) ; FIX(1.847759065) 59F_1_961 equ DESCALE(2106220350,30-CONST_BITS) ; FIX(1.961570560) 60F_2_053 equ DESCALE(2204520673,30-CONST_BITS) ; FIX(2.053119869) 61F_2_562 equ DESCALE(2751909506,30-CONST_BITS) ; FIX(2.562915447) 62F_3_072 equ DESCALE(3299298341,30-CONST_BITS) ; FIX(3.072711026) 63%endif 64 65; -------------------------------------------------------------------------- 66 SECTION SEG_CONST 67 68 alignz 16 69 global EXTN(jconst_fdct_islow_sse2) 70 71EXTN(jconst_fdct_islow_sse2): 72 73PW_F130_F054 times 4 dw (F_0_541+F_0_765), F_0_541 74PW_F054_MF130 times 4 dw F_0_541, (F_0_541-F_1_847) 75PW_MF078_F117 times 4 dw (F_1_175-F_1_961), F_1_175 76PW_F117_F078 times 4 dw F_1_175, (F_1_175-F_0_390) 77PW_MF060_MF089 times 4 dw (F_0_298-F_0_899),-F_0_899 78PW_MF089_F060 times 4 dw -F_0_899, (F_1_501-F_0_899) 79PW_MF050_MF256 times 4 dw (F_2_053-F_2_562),-F_2_562 80PW_MF256_F050 times 4 dw -F_2_562, (F_3_072-F_2_562) 81PD_DESCALE_P1 times 4 dd 1 << (DESCALE_P1-1) 82PD_DESCALE_P2 times 4 dd 1 << (DESCALE_P2-1) 83PW_DESCALE_P2X times 8 dw 1 << (PASS1_BITS-1) 84 85 alignz 16 86 87; -------------------------------------------------------------------------- 88 SECTION SEG_TEXT 89 BITS 64 90; 91; Perform the forward DCT on one block of samples. 92; 93; GLOBAL(void) 94; jsimd_fdct_islow_sse2 (DCTELEM *data) 95; 96 97; r10 = DCTELEM *data 98 99%define wk(i) rbp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM] 100%define WK_NUM 6 101 102 align 16 103 global EXTN(jsimd_fdct_islow_sse2) 104 105EXTN(jsimd_fdct_islow_sse2): 106 push rbp 107 mov rax,rsp ; rax = original rbp 108 sub rsp, byte 4 109 and rsp, byte (-SIZEOF_XMMWORD) ; align to 128 bits 110 mov [rsp],rax 111 mov rbp,rsp ; rbp = aligned rbp 112 lea rsp, [wk(0)] 113 collect_args 114 115 ; ---- Pass 1: process rows. 116 117 mov rdx, r10 ; (DCTELEM *) 118 119 movdqa xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_DCTELEM)] 120 movdqa xmm1, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_DCTELEM)] 121 movdqa xmm2, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_DCTELEM)] 122 movdqa xmm3, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_DCTELEM)] 123 124 ; xmm0=(00 01 02 03 04 05 06 07), xmm2=(20 21 22 23 24 25 26 27) 125 ; xmm1=(10 11 12 13 14 15 16 17), xmm3=(30 31 32 33 34 35 36 37) 126 127 movdqa xmm4,xmm0 ; transpose coefficients(phase 1) 128 punpcklwd xmm0,xmm1 ; xmm0=(00 10 01 11 02 12 03 13) 129 punpckhwd xmm4,xmm1 ; xmm4=(04 14 05 15 06 16 07 17) 130 movdqa xmm5,xmm2 ; transpose coefficients(phase 1) 131 punpcklwd xmm2,xmm3 ; xmm2=(20 30 21 31 22 32 23 33) 132 punpckhwd xmm5,xmm3 ; xmm5=(24 34 25 35 26 36 27 37) 133 134 movdqa xmm6, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_DCTELEM)] 135 movdqa xmm7, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_DCTELEM)] 136 movdqa xmm1, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_DCTELEM)] 137 movdqa xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_DCTELEM)] 138 139 ; xmm6=( 4 12 20 28 36 44 52 60), xmm1=( 6 14 22 30 38 46 54 62) 140 ; xmm7=( 5 13 21 29 37 45 53 61), xmm3=( 7 15 23 31 39 47 55 63) 141 142 movdqa XMMWORD [wk(0)], xmm2 ; wk(0)=(20 30 21 31 22 32 23 33) 143 movdqa XMMWORD [wk(1)], xmm5 ; wk(1)=(24 34 25 35 26 36 27 37) 144 145 movdqa xmm2,xmm6 ; transpose coefficients(phase 1) 146 punpcklwd xmm6,xmm7 ; xmm6=(40 50 41 51 42 52 43 53) 147 punpckhwd xmm2,xmm7 ; xmm2=(44 54 45 55 46 56 47 57) 148 movdqa xmm5,xmm1 ; transpose coefficients(phase 1) 149 punpcklwd xmm1,xmm3 ; xmm1=(60 70 61 71 62 72 63 73) 150 punpckhwd xmm5,xmm3 ; xmm5=(64 74 65 75 66 76 67 77) 151 152 movdqa xmm7,xmm6 ; transpose coefficients(phase 2) 153 punpckldq xmm6,xmm1 ; xmm6=(40 50 60 70 41 51 61 71) 154 punpckhdq xmm7,xmm1 ; xmm7=(42 52 62 72 43 53 63 73) 155 movdqa xmm3,xmm2 ; transpose coefficients(phase 2) 156 punpckldq xmm2,xmm5 ; xmm2=(44 54 64 74 45 55 65 75) 157 punpckhdq xmm3,xmm5 ; xmm3=(46 56 66 76 47 57 67 77) 158 159 movdqa xmm1, XMMWORD [wk(0)] ; xmm1=(20 30 21 31 22 32 23 33) 160 movdqa xmm5, XMMWORD [wk(1)] ; xmm5=(24 34 25 35 26 36 27 37) 161 movdqa XMMWORD [wk(2)], xmm7 ; wk(2)=(42 52 62 72 43 53 63 73) 162 movdqa XMMWORD [wk(3)], xmm2 ; wk(3)=(44 54 64 74 45 55 65 75) 163 164 movdqa xmm7,xmm0 ; transpose coefficients(phase 2) 165 punpckldq xmm0,xmm1 ; xmm0=(00 10 20 30 01 11 21 31) 166 punpckhdq xmm7,xmm1 ; xmm7=(02 12 22 32 03 13 23 33) 167 movdqa xmm2,xmm4 ; transpose coefficients(phase 2) 168 punpckldq xmm4,xmm5 ; xmm4=(04 14 24 34 05 15 25 35) 169 punpckhdq xmm2,xmm5 ; xmm2=(06 16 26 36 07 17 27 37) 170 171 movdqa xmm1,xmm0 ; transpose coefficients(phase 3) 172 punpcklqdq xmm0,xmm6 ; xmm0=(00 10 20 30 40 50 60 70)=data0 173 punpckhqdq xmm1,xmm6 ; xmm1=(01 11 21 31 41 51 61 71)=data1 174 movdqa xmm5,xmm2 ; transpose coefficients(phase 3) 175 punpcklqdq xmm2,xmm3 ; xmm2=(06 16 26 36 46 56 66 76)=data6 176 punpckhqdq xmm5,xmm3 ; xmm5=(07 17 27 37 47 57 67 77)=data7 177 178 movdqa xmm6,xmm1 179 movdqa xmm3,xmm0 180 psubw xmm1,xmm2 ; xmm1=data1-data6=tmp6 181 psubw xmm0,xmm5 ; xmm0=data0-data7=tmp7 182 paddw xmm6,xmm2 ; xmm6=data1+data6=tmp1 183 paddw xmm3,xmm5 ; xmm3=data0+data7=tmp0 184 185 movdqa xmm2, XMMWORD [wk(2)] ; xmm2=(42 52 62 72 43 53 63 73) 186 movdqa xmm5, XMMWORD [wk(3)] ; xmm5=(44 54 64 74 45 55 65 75) 187 movdqa XMMWORD [wk(0)], xmm1 ; wk(0)=tmp6 188 movdqa XMMWORD [wk(1)], xmm0 ; wk(1)=tmp7 189 190 movdqa xmm1,xmm7 ; transpose coefficients(phase 3) 191 punpcklqdq xmm7,xmm2 ; xmm7=(02 12 22 32 42 52 62 72)=data2 192 punpckhqdq xmm1,xmm2 ; xmm1=(03 13 23 33 43 53 63 73)=data3 193 movdqa xmm0,xmm4 ; transpose coefficients(phase 3) 194 punpcklqdq xmm4,xmm5 ; xmm4=(04 14 24 34 44 54 64 74)=data4 195 punpckhqdq xmm0,xmm5 ; xmm0=(05 15 25 35 45 55 65 75)=data5 196 197 movdqa xmm2,xmm1 198 movdqa xmm5,xmm7 199 paddw xmm1,xmm4 ; xmm1=data3+data4=tmp3 200 paddw xmm7,xmm0 ; xmm7=data2+data5=tmp2 201 psubw xmm2,xmm4 ; xmm2=data3-data4=tmp4 202 psubw xmm5,xmm0 ; xmm5=data2-data5=tmp5 203 204 ; -- Even part 205 206 movdqa xmm4,xmm3 207 movdqa xmm0,xmm6 208 paddw xmm3,xmm1 ; xmm3=tmp10 209 paddw xmm6,xmm7 ; xmm6=tmp11 210 psubw xmm4,xmm1 ; xmm4=tmp13 211 psubw xmm0,xmm7 ; xmm0=tmp12 212 213 movdqa xmm1,xmm3 214 paddw xmm3,xmm6 ; xmm3=tmp10+tmp11 215 psubw xmm1,xmm6 ; xmm1=tmp10-tmp11 216 217 psllw xmm3,PASS1_BITS ; xmm3=data0 218 psllw xmm1,PASS1_BITS ; xmm1=data4 219 220 movdqa XMMWORD [wk(2)], xmm3 ; wk(2)=data0 221 movdqa XMMWORD [wk(3)], xmm1 ; wk(3)=data4 222 223 ; (Original) 224 ; z1 = (tmp12 + tmp13) * 0.541196100; 225 ; data2 = z1 + tmp13 * 0.765366865; 226 ; data6 = z1 + tmp12 * -1.847759065; 227 ; 228 ; (This implementation) 229 ; data2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100; 230 ; data6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065); 231 232 movdqa xmm7,xmm4 ; xmm4=tmp13 233 movdqa xmm6,xmm4 234 punpcklwd xmm7,xmm0 ; xmm0=tmp12 235 punpckhwd xmm6,xmm0 236 movdqa xmm4,xmm7 237 movdqa xmm0,xmm6 238 pmaddwd xmm7,[rel PW_F130_F054] ; xmm7=data2L 239 pmaddwd xmm6,[rel PW_F130_F054] ; xmm6=data2H 240 pmaddwd xmm4,[rel PW_F054_MF130] ; xmm4=data6L 241 pmaddwd xmm0,[rel PW_F054_MF130] ; xmm0=data6H 242 243 paddd xmm7,[rel PD_DESCALE_P1] 244 paddd xmm6,[rel PD_DESCALE_P1] 245 psrad xmm7,DESCALE_P1 246 psrad xmm6,DESCALE_P1 247 paddd xmm4,[rel PD_DESCALE_P1] 248 paddd xmm0,[rel PD_DESCALE_P1] 249 psrad xmm4,DESCALE_P1 250 psrad xmm0,DESCALE_P1 251 252 packssdw xmm7,xmm6 ; xmm7=data2 253 packssdw xmm4,xmm0 ; xmm4=data6 254 255 movdqa XMMWORD [wk(4)], xmm7 ; wk(4)=data2 256 movdqa XMMWORD [wk(5)], xmm4 ; wk(5)=data6 257 258 ; -- Odd part 259 260 movdqa xmm3, XMMWORD [wk(0)] ; xmm3=tmp6 261 movdqa xmm1, XMMWORD [wk(1)] ; xmm1=tmp7 262 263 movdqa xmm6,xmm2 ; xmm2=tmp4 264 movdqa xmm0,xmm5 ; xmm5=tmp5 265 paddw xmm6,xmm3 ; xmm6=z3 266 paddw xmm0,xmm1 ; xmm0=z4 267 268 ; (Original) 269 ; z5 = (z3 + z4) * 1.175875602; 270 ; z3 = z3 * -1.961570560; z4 = z4 * -0.390180644; 271 ; z3 += z5; z4 += z5; 272 ; 273 ; (This implementation) 274 ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602; 275 ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644); 276 277 movdqa xmm7,xmm6 278 movdqa xmm4,xmm6 279 punpcklwd xmm7,xmm0 280 punpckhwd xmm4,xmm0 281 movdqa xmm6,xmm7 282 movdqa xmm0,xmm4 283 pmaddwd xmm7,[rel PW_MF078_F117] ; xmm7=z3L 284 pmaddwd xmm4,[rel PW_MF078_F117] ; xmm4=z3H 285 pmaddwd xmm6,[rel PW_F117_F078] ; xmm6=z4L 286 pmaddwd xmm0,[rel PW_F117_F078] ; xmm0=z4H 287 288 movdqa XMMWORD [wk(0)], xmm7 ; wk(0)=z3L 289 movdqa XMMWORD [wk(1)], xmm4 ; wk(1)=z3H 290 291 ; (Original) 292 ; z1 = tmp4 + tmp7; z2 = tmp5 + tmp6; 293 ; tmp4 = tmp4 * 0.298631336; tmp5 = tmp5 * 2.053119869; 294 ; tmp6 = tmp6 * 3.072711026; tmp7 = tmp7 * 1.501321110; 295 ; z1 = z1 * -0.899976223; z2 = z2 * -2.562915447; 296 ; data7 = tmp4 + z1 + z3; data5 = tmp5 + z2 + z4; 297 ; data3 = tmp6 + z2 + z3; data1 = tmp7 + z1 + z4; 298 ; 299 ; (This implementation) 300 ; tmp4 = tmp4 * (0.298631336 - 0.899976223) + tmp7 * -0.899976223; 301 ; tmp5 = tmp5 * (2.053119869 - 2.562915447) + tmp6 * -2.562915447; 302 ; tmp6 = tmp5 * -2.562915447 + tmp6 * (3.072711026 - 2.562915447); 303 ; tmp7 = tmp4 * -0.899976223 + tmp7 * (1.501321110 - 0.899976223); 304 ; data7 = tmp4 + z3; data5 = tmp5 + z4; 305 ; data3 = tmp6 + z3; data1 = tmp7 + z4; 306 307 movdqa xmm7,xmm2 308 movdqa xmm4,xmm2 309 punpcklwd xmm7,xmm1 310 punpckhwd xmm4,xmm1 311 movdqa xmm2,xmm7 312 movdqa xmm1,xmm4 313 pmaddwd xmm7,[rel PW_MF060_MF089] ; xmm7=tmp4L 314 pmaddwd xmm4,[rel PW_MF060_MF089] ; xmm4=tmp4H 315 pmaddwd xmm2,[rel PW_MF089_F060] ; xmm2=tmp7L 316 pmaddwd xmm1,[rel PW_MF089_F060] ; xmm1=tmp7H 317 318 paddd xmm7, XMMWORD [wk(0)] ; xmm7=data7L 319 paddd xmm4, XMMWORD [wk(1)] ; xmm4=data7H 320 paddd xmm2,xmm6 ; xmm2=data1L 321 paddd xmm1,xmm0 ; xmm1=data1H 322 323 paddd xmm7,[rel PD_DESCALE_P1] 324 paddd xmm4,[rel PD_DESCALE_P1] 325 psrad xmm7,DESCALE_P1 326 psrad xmm4,DESCALE_P1 327 paddd xmm2,[rel PD_DESCALE_P1] 328 paddd xmm1,[rel PD_DESCALE_P1] 329 psrad xmm2,DESCALE_P1 330 psrad xmm1,DESCALE_P1 331 332 packssdw xmm7,xmm4 ; xmm7=data7 333 packssdw xmm2,xmm1 ; xmm2=data1 334 335 movdqa xmm4,xmm5 336 movdqa xmm1,xmm5 337 punpcklwd xmm4,xmm3 338 punpckhwd xmm1,xmm3 339 movdqa xmm5,xmm4 340 movdqa xmm3,xmm1 341 pmaddwd xmm4,[rel PW_MF050_MF256] ; xmm4=tmp5L 342 pmaddwd xmm1,[rel PW_MF050_MF256] ; xmm1=tmp5H 343 pmaddwd xmm5,[rel PW_MF256_F050] ; xmm5=tmp6L 344 pmaddwd xmm3,[rel PW_MF256_F050] ; xmm3=tmp6H 345 346 paddd xmm4,xmm6 ; xmm4=data5L 347 paddd xmm1,xmm0 ; xmm1=data5H 348 paddd xmm5, XMMWORD [wk(0)] ; xmm5=data3L 349 paddd xmm3, XMMWORD [wk(1)] ; xmm3=data3H 350 351 paddd xmm4,[rel PD_DESCALE_P1] 352 paddd xmm1,[rel PD_DESCALE_P1] 353 psrad xmm4,DESCALE_P1 354 psrad xmm1,DESCALE_P1 355 paddd xmm5,[rel PD_DESCALE_P1] 356 paddd xmm3,[rel PD_DESCALE_P1] 357 psrad xmm5,DESCALE_P1 358 psrad xmm3,DESCALE_P1 359 360 packssdw xmm4,xmm1 ; xmm4=data5 361 packssdw xmm5,xmm3 ; xmm5=data3 362 363 ; ---- Pass 2: process columns. 364 365 movdqa xmm6, XMMWORD [wk(2)] ; xmm6=col0 366 movdqa xmm0, XMMWORD [wk(4)] ; xmm0=col2 367 368 ; xmm6=(00 10 20 30 40 50 60 70), xmm0=(02 12 22 32 42 52 62 72) 369 ; xmm2=(01 11 21 31 41 51 61 71), xmm5=(03 13 23 33 43 53 63 73) 370 371 movdqa xmm1,xmm6 ; transpose coefficients(phase 1) 372 punpcklwd xmm6,xmm2 ; xmm6=(00 01 10 11 20 21 30 31) 373 punpckhwd xmm1,xmm2 ; xmm1=(40 41 50 51 60 61 70 71) 374 movdqa xmm3,xmm0 ; transpose coefficients(phase 1) 375 punpcklwd xmm0,xmm5 ; xmm0=(02 03 12 13 22 23 32 33) 376 punpckhwd xmm3,xmm5 ; xmm3=(42 43 52 53 62 63 72 73) 377 378 movdqa xmm2, XMMWORD [wk(3)] ; xmm2=col4 379 movdqa xmm5, XMMWORD [wk(5)] ; xmm5=col6 380 381 ; xmm2=(04 14 24 34 44 54 64 74), xmm5=(06 16 26 36 46 56 66 76) 382 ; xmm4=(05 15 25 35 45 55 65 75), xmm7=(07 17 27 37 47 57 67 77) 383 384 movdqa XMMWORD [wk(0)], xmm0 ; wk(0)=(02 03 12 13 22 23 32 33) 385 movdqa XMMWORD [wk(1)], xmm3 ; wk(1)=(42 43 52 53 62 63 72 73) 386 387 movdqa xmm0,xmm2 ; transpose coefficients(phase 1) 388 punpcklwd xmm2,xmm4 ; xmm2=(04 05 14 15 24 25 34 35) 389 punpckhwd xmm0,xmm4 ; xmm0=(44 45 54 55 64 65 74 75) 390 movdqa xmm3,xmm5 ; transpose coefficients(phase 1) 391 punpcklwd xmm5,xmm7 ; xmm5=(06 07 16 17 26 27 36 37) 392 punpckhwd xmm3,xmm7 ; xmm3=(46 47 56 57 66 67 76 77) 393 394 movdqa xmm4,xmm2 ; transpose coefficients(phase 2) 395 punpckldq xmm2,xmm5 ; xmm2=(04 05 06 07 14 15 16 17) 396 punpckhdq xmm4,xmm5 ; xmm4=(24 25 26 27 34 35 36 37) 397 movdqa xmm7,xmm0 ; transpose coefficients(phase 2) 398 punpckldq xmm0,xmm3 ; xmm0=(44 45 46 47 54 55 56 57) 399 punpckhdq xmm7,xmm3 ; xmm7=(64 65 66 67 74 75 76 77) 400 401 movdqa xmm5, XMMWORD [wk(0)] ; xmm5=(02 03 12 13 22 23 32 33) 402 movdqa xmm3, XMMWORD [wk(1)] ; xmm3=(42 43 52 53 62 63 72 73) 403 movdqa XMMWORD [wk(2)], xmm4 ; wk(2)=(24 25 26 27 34 35 36 37) 404 movdqa XMMWORD [wk(3)], xmm0 ; wk(3)=(44 45 46 47 54 55 56 57) 405 406 movdqa xmm4,xmm6 ; transpose coefficients(phase 2) 407 punpckldq xmm6,xmm5 ; xmm6=(00 01 02 03 10 11 12 13) 408 punpckhdq xmm4,xmm5 ; xmm4=(20 21 22 23 30 31 32 33) 409 movdqa xmm0,xmm1 ; transpose coefficients(phase 2) 410 punpckldq xmm1,xmm3 ; xmm1=(40 41 42 43 50 51 52 53) 411 punpckhdq xmm0,xmm3 ; xmm0=(60 61 62 63 70 71 72 73) 412 413 movdqa xmm5,xmm6 ; transpose coefficients(phase 3) 414 punpcklqdq xmm6,xmm2 ; xmm6=(00 01 02 03 04 05 06 07)=data0 415 punpckhqdq xmm5,xmm2 ; xmm5=(10 11 12 13 14 15 16 17)=data1 416 movdqa xmm3,xmm0 ; transpose coefficients(phase 3) 417 punpcklqdq xmm0,xmm7 ; xmm0=(60 61 62 63 64 65 66 67)=data6 418 punpckhqdq xmm3,xmm7 ; xmm3=(70 71 72 73 74 75 76 77)=data7 419 420 movdqa xmm2,xmm5 421 movdqa xmm7,xmm6 422 psubw xmm5,xmm0 ; xmm5=data1-data6=tmp6 423 psubw xmm6,xmm3 ; xmm6=data0-data7=tmp7 424 paddw xmm2,xmm0 ; xmm2=data1+data6=tmp1 425 paddw xmm7,xmm3 ; xmm7=data0+data7=tmp0 426 427 movdqa xmm0, XMMWORD [wk(2)] ; xmm0=(24 25 26 27 34 35 36 37) 428 movdqa xmm3, XMMWORD [wk(3)] ; xmm3=(44 45 46 47 54 55 56 57) 429 movdqa XMMWORD [wk(0)], xmm5 ; wk(0)=tmp6 430 movdqa XMMWORD [wk(1)], xmm6 ; wk(1)=tmp7 431 432 movdqa xmm5,xmm4 ; transpose coefficients(phase 3) 433 punpcklqdq xmm4,xmm0 ; xmm4=(20 21 22 23 24 25 26 27)=data2 434 punpckhqdq xmm5,xmm0 ; xmm5=(30 31 32 33 34 35 36 37)=data3 435 movdqa xmm6,xmm1 ; transpose coefficients(phase 3) 436 punpcklqdq xmm1,xmm3 ; xmm1=(40 41 42 43 44 45 46 47)=data4 437 punpckhqdq xmm6,xmm3 ; xmm6=(50 51 52 53 54 55 56 57)=data5 438 439 movdqa xmm0,xmm5 440 movdqa xmm3,xmm4 441 paddw xmm5,xmm1 ; xmm5=data3+data4=tmp3 442 paddw xmm4,xmm6 ; xmm4=data2+data5=tmp2 443 psubw xmm0,xmm1 ; xmm0=data3-data4=tmp4 444 psubw xmm3,xmm6 ; xmm3=data2-data5=tmp5 445 446 ; -- Even part 447 448 movdqa xmm1,xmm7 449 movdqa xmm6,xmm2 450 paddw xmm7,xmm5 ; xmm7=tmp10 451 paddw xmm2,xmm4 ; xmm2=tmp11 452 psubw xmm1,xmm5 ; xmm1=tmp13 453 psubw xmm6,xmm4 ; xmm6=tmp12 454 455 movdqa xmm5,xmm7 456 paddw xmm7,xmm2 ; xmm7=tmp10+tmp11 457 psubw xmm5,xmm2 ; xmm5=tmp10-tmp11 458 459 paddw xmm7,[rel PW_DESCALE_P2X] 460 paddw xmm5,[rel PW_DESCALE_P2X] 461 psraw xmm7,PASS1_BITS ; xmm7=data0 462 psraw xmm5,PASS1_BITS ; xmm5=data4 463 464 movdqa XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_DCTELEM)], xmm7 465 movdqa XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_DCTELEM)], xmm5 466 467 ; (Original) 468 ; z1 = (tmp12 + tmp13) * 0.541196100; 469 ; data2 = z1 + tmp13 * 0.765366865; 470 ; data6 = z1 + tmp12 * -1.847759065; 471 ; 472 ; (This implementation) 473 ; data2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100; 474 ; data6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065); 475 476 movdqa xmm4,xmm1 ; xmm1=tmp13 477 movdqa xmm2,xmm1 478 punpcklwd xmm4,xmm6 ; xmm6=tmp12 479 punpckhwd xmm2,xmm6 480 movdqa xmm1,xmm4 481 movdqa xmm6,xmm2 482 pmaddwd xmm4,[rel PW_F130_F054] ; xmm4=data2L 483 pmaddwd xmm2,[rel PW_F130_F054] ; xmm2=data2H 484 pmaddwd xmm1,[rel PW_F054_MF130] ; xmm1=data6L 485 pmaddwd xmm6,[rel PW_F054_MF130] ; xmm6=data6H 486 487 paddd xmm4,[rel PD_DESCALE_P2] 488 paddd xmm2,[rel PD_DESCALE_P2] 489 psrad xmm4,DESCALE_P2 490 psrad xmm2,DESCALE_P2 491 paddd xmm1,[rel PD_DESCALE_P2] 492 paddd xmm6,[rel PD_DESCALE_P2] 493 psrad xmm1,DESCALE_P2 494 psrad xmm6,DESCALE_P2 495 496 packssdw xmm4,xmm2 ; xmm4=data2 497 packssdw xmm1,xmm6 ; xmm1=data6 498 499 movdqa XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_DCTELEM)], xmm4 500 movdqa XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_DCTELEM)], xmm1 501 502 ; -- Odd part 503 504 movdqa xmm7, XMMWORD [wk(0)] ; xmm7=tmp6 505 movdqa xmm5, XMMWORD [wk(1)] ; xmm5=tmp7 506 507 movdqa xmm2,xmm0 ; xmm0=tmp4 508 movdqa xmm6,xmm3 ; xmm3=tmp5 509 paddw xmm2,xmm7 ; xmm2=z3 510 paddw xmm6,xmm5 ; xmm6=z4 511 512 ; (Original) 513 ; z5 = (z3 + z4) * 1.175875602; 514 ; z3 = z3 * -1.961570560; z4 = z4 * -0.390180644; 515 ; z3 += z5; z4 += z5; 516 ; 517 ; (This implementation) 518 ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602; 519 ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644); 520 521 movdqa xmm4,xmm2 522 movdqa xmm1,xmm2 523 punpcklwd xmm4,xmm6 524 punpckhwd xmm1,xmm6 525 movdqa xmm2,xmm4 526 movdqa xmm6,xmm1 527 pmaddwd xmm4,[rel PW_MF078_F117] ; xmm4=z3L 528 pmaddwd xmm1,[rel PW_MF078_F117] ; xmm1=z3H 529 pmaddwd xmm2,[rel PW_F117_F078] ; xmm2=z4L 530 pmaddwd xmm6,[rel PW_F117_F078] ; xmm6=z4H 531 532 movdqa XMMWORD [wk(0)], xmm4 ; wk(0)=z3L 533 movdqa XMMWORD [wk(1)], xmm1 ; wk(1)=z3H 534 535 ; (Original) 536 ; z1 = tmp4 + tmp7; z2 = tmp5 + tmp6; 537 ; tmp4 = tmp4 * 0.298631336; tmp5 = tmp5 * 2.053119869; 538 ; tmp6 = tmp6 * 3.072711026; tmp7 = tmp7 * 1.501321110; 539 ; z1 = z1 * -0.899976223; z2 = z2 * -2.562915447; 540 ; data7 = tmp4 + z1 + z3; data5 = tmp5 + z2 + z4; 541 ; data3 = tmp6 + z2 + z3; data1 = tmp7 + z1 + z4; 542 ; 543 ; (This implementation) 544 ; tmp4 = tmp4 * (0.298631336 - 0.899976223) + tmp7 * -0.899976223; 545 ; tmp5 = tmp5 * (2.053119869 - 2.562915447) + tmp6 * -2.562915447; 546 ; tmp6 = tmp5 * -2.562915447 + tmp6 * (3.072711026 - 2.562915447); 547 ; tmp7 = tmp4 * -0.899976223 + tmp7 * (1.501321110 - 0.899976223); 548 ; data7 = tmp4 + z3; data5 = tmp5 + z4; 549 ; data3 = tmp6 + z3; data1 = tmp7 + z4; 550 551 movdqa xmm4,xmm0 552 movdqa xmm1,xmm0 553 punpcklwd xmm4,xmm5 554 punpckhwd xmm1,xmm5 555 movdqa xmm0,xmm4 556 movdqa xmm5,xmm1 557 pmaddwd xmm4,[rel PW_MF060_MF089] ; xmm4=tmp4L 558 pmaddwd xmm1,[rel PW_MF060_MF089] ; xmm1=tmp4H 559 pmaddwd xmm0,[rel PW_MF089_F060] ; xmm0=tmp7L 560 pmaddwd xmm5,[rel PW_MF089_F060] ; xmm5=tmp7H 561 562 paddd xmm4, XMMWORD [wk(0)] ; xmm4=data7L 563 paddd xmm1, XMMWORD [wk(1)] ; xmm1=data7H 564 paddd xmm0,xmm2 ; xmm0=data1L 565 paddd xmm5,xmm6 ; xmm5=data1H 566 567 paddd xmm4,[rel PD_DESCALE_P2] 568 paddd xmm1,[rel PD_DESCALE_P2] 569 psrad xmm4,DESCALE_P2 570 psrad xmm1,DESCALE_P2 571 paddd xmm0,[rel PD_DESCALE_P2] 572 paddd xmm5,[rel PD_DESCALE_P2] 573 psrad xmm0,DESCALE_P2 574 psrad xmm5,DESCALE_P2 575 576 packssdw xmm4,xmm1 ; xmm4=data7 577 packssdw xmm0,xmm5 ; xmm0=data1 578 579 movdqa XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_DCTELEM)], xmm4 580 movdqa XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_DCTELEM)], xmm0 581 582 movdqa xmm1,xmm3 583 movdqa xmm5,xmm3 584 punpcklwd xmm1,xmm7 585 punpckhwd xmm5,xmm7 586 movdqa xmm3,xmm1 587 movdqa xmm7,xmm5 588 pmaddwd xmm1,[rel PW_MF050_MF256] ; xmm1=tmp5L 589 pmaddwd xmm5,[rel PW_MF050_MF256] ; xmm5=tmp5H 590 pmaddwd xmm3,[rel PW_MF256_F050] ; xmm3=tmp6L 591 pmaddwd xmm7,[rel PW_MF256_F050] ; xmm7=tmp6H 592 593 paddd xmm1,xmm2 ; xmm1=data5L 594 paddd xmm5,xmm6 ; xmm5=data5H 595 paddd xmm3, XMMWORD [wk(0)] ; xmm3=data3L 596 paddd xmm7, XMMWORD [wk(1)] ; xmm7=data3H 597 598 paddd xmm1,[rel PD_DESCALE_P2] 599 paddd xmm5,[rel PD_DESCALE_P2] 600 psrad xmm1,DESCALE_P2 601 psrad xmm5,DESCALE_P2 602 paddd xmm3,[rel PD_DESCALE_P2] 603 paddd xmm7,[rel PD_DESCALE_P2] 604 psrad xmm3,DESCALE_P2 605 psrad xmm7,DESCALE_P2 606 607 packssdw xmm1,xmm5 ; xmm1=data5 608 packssdw xmm3,xmm7 ; xmm3=data3 609 610 movdqa XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_DCTELEM)], xmm1 611 movdqa XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_DCTELEM)], xmm3 612 613 uncollect_args 614 mov rsp,rbp ; rsp <- aligned rbp 615 pop rsp ; rsp <- original rbp 616 pop rbp 617 ret 618 619; For some reason, the OS X linker does not honor the request to align the 620; segment unless we do this. 621 align 16 622