1; 2; jfdctfst.asm - fast integer FDCT (SSE2) 3; 4; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB 5; Copyright (C) 2016, 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 fast, not so accurate integer implementation of 18; the forward DCT (Discrete Cosine Transform). The following code is 19; based directly on the IJG's original jfdctfst.c; see the jfdctfst.c 20; for more details. 21; 22; [TAB8] 23 24%include "jsimdext.inc" 25%include "jdct.inc" 26 27; -------------------------------------------------------------------------- 28 29%define CONST_BITS 8 ; 14 is also OK. 30 31%if CONST_BITS == 8 32F_0_382 equ 98 ; FIX(0.382683433) 33F_0_541 equ 139 ; FIX(0.541196100) 34F_0_707 equ 181 ; FIX(0.707106781) 35F_1_306 equ 334 ; FIX(1.306562965) 36%else 37; NASM cannot do compile-time arithmetic on floating-point constants. 38%define DESCALE(x, n) (((x) + (1 << ((n) - 1))) >> (n)) 39F_0_382 equ DESCALE( 410903207, 30 - CONST_BITS) ; FIX(0.382683433) 40F_0_541 equ DESCALE( 581104887, 30 - CONST_BITS) ; FIX(0.541196100) 41F_0_707 equ DESCALE( 759250124, 30 - CONST_BITS) ; FIX(0.707106781) 42F_1_306 equ DESCALE(1402911301, 30 - CONST_BITS) ; FIX(1.306562965) 43%endif 44 45; -------------------------------------------------------------------------- 46 SECTION SEG_CONST 47 48; PRE_MULTIPLY_SCALE_BITS <= 2 (to avoid overflow) 49; CONST_BITS + CONST_SHIFT + PRE_MULTIPLY_SCALE_BITS == 16 (for pmulhw) 50 51%define PRE_MULTIPLY_SCALE_BITS 2 52%define CONST_SHIFT (16 - PRE_MULTIPLY_SCALE_BITS - CONST_BITS) 53 54 alignz 32 55 GLOBAL_DATA(jconst_fdct_ifast_sse2) 56 57EXTN(jconst_fdct_ifast_sse2): 58 59PW_F0707 times 8 dw F_0_707 << CONST_SHIFT 60PW_F0382 times 8 dw F_0_382 << CONST_SHIFT 61PW_F0541 times 8 dw F_0_541 << CONST_SHIFT 62PW_F1306 times 8 dw F_1_306 << CONST_SHIFT 63 64 alignz 32 65 66; -------------------------------------------------------------------------- 67 SECTION SEG_TEXT 68 BITS 32 69; 70; Perform the forward DCT on one block of samples. 71; 72; GLOBAL(void) 73; jsimd_fdct_ifast_sse2(DCTELEM *data) 74; 75 76%define data(b) (b) + 8 ; DCTELEM *data 77 78%define original_ebp ebp + 0 79%define wk(i) ebp - (WK_NUM - (i)) * SIZEOF_XMMWORD 80 ; xmmword wk[WK_NUM] 81%define WK_NUM 2 82 83 align 32 84 GLOBAL_FUNCTION(jsimd_fdct_ifast_sse2) 85 86EXTN(jsimd_fdct_ifast_sse2): 87 push ebp 88 mov eax, esp ; eax = original ebp 89 sub esp, byte 4 90 and esp, byte (-SIZEOF_XMMWORD) ; align to 128 bits 91 mov [esp], eax 92 mov ebp, esp ; ebp = aligned ebp 93 lea esp, [wk(0)] 94 pushpic ebx 95; push ecx ; unused 96; push edx ; need not be preserved 97; push esi ; unused 98; push edi ; unused 99 100 get_GOT ebx ; get GOT address 101 102 ; ---- Pass 1: process rows. 103 104 mov edx, POINTER [data(eax)] ; (DCTELEM *) 105 106 movdqa xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_DCTELEM)] 107 movdqa xmm1, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_DCTELEM)] 108 movdqa xmm2, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_DCTELEM)] 109 movdqa xmm3, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_DCTELEM)] 110 111 ; xmm0=(00 01 02 03 04 05 06 07), xmm2=(20 21 22 23 24 25 26 27) 112 ; xmm1=(10 11 12 13 14 15 16 17), xmm3=(30 31 32 33 34 35 36 37) 113 114 movdqa xmm4, xmm0 ; transpose coefficients(phase 1) 115 punpcklwd xmm0, xmm1 ; xmm0=(00 10 01 11 02 12 03 13) 116 punpckhwd xmm4, xmm1 ; xmm4=(04 14 05 15 06 16 07 17) 117 movdqa xmm5, xmm2 ; transpose coefficients(phase 1) 118 punpcklwd xmm2, xmm3 ; xmm2=(20 30 21 31 22 32 23 33) 119 punpckhwd xmm5, xmm3 ; xmm5=(24 34 25 35 26 36 27 37) 120 121 movdqa xmm6, XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_DCTELEM)] 122 movdqa xmm7, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_DCTELEM)] 123 movdqa xmm1, XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_DCTELEM)] 124 movdqa xmm3, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_DCTELEM)] 125 126 ; xmm6=( 4 12 20 28 36 44 52 60), xmm1=( 6 14 22 30 38 46 54 62) 127 ; xmm7=( 5 13 21 29 37 45 53 61), xmm3=( 7 15 23 31 39 47 55 63) 128 129 movdqa XMMWORD [wk(0)], xmm2 ; wk(0)=(20 30 21 31 22 32 23 33) 130 movdqa XMMWORD [wk(1)], xmm5 ; wk(1)=(24 34 25 35 26 36 27 37) 131 132 movdqa xmm2, xmm6 ; transpose coefficients(phase 1) 133 punpcklwd xmm6, xmm7 ; xmm6=(40 50 41 51 42 52 43 53) 134 punpckhwd xmm2, xmm7 ; xmm2=(44 54 45 55 46 56 47 57) 135 movdqa xmm5, xmm1 ; transpose coefficients(phase 1) 136 punpcklwd xmm1, xmm3 ; xmm1=(60 70 61 71 62 72 63 73) 137 punpckhwd xmm5, xmm3 ; xmm5=(64 74 65 75 66 76 67 77) 138 139 movdqa xmm7, xmm6 ; transpose coefficients(phase 2) 140 punpckldq xmm6, xmm1 ; xmm6=(40 50 60 70 41 51 61 71) 141 punpckhdq xmm7, xmm1 ; xmm7=(42 52 62 72 43 53 63 73) 142 movdqa xmm3, xmm2 ; transpose coefficients(phase 2) 143 punpckldq xmm2, xmm5 ; xmm2=(44 54 64 74 45 55 65 75) 144 punpckhdq xmm3, xmm5 ; xmm3=(46 56 66 76 47 57 67 77) 145 146 movdqa xmm1, XMMWORD [wk(0)] ; xmm1=(20 30 21 31 22 32 23 33) 147 movdqa xmm5, XMMWORD [wk(1)] ; xmm5=(24 34 25 35 26 36 27 37) 148 movdqa XMMWORD [wk(0)], xmm7 ; wk(0)=(42 52 62 72 43 53 63 73) 149 movdqa XMMWORD [wk(1)], xmm2 ; wk(1)=(44 54 64 74 45 55 65 75) 150 151 movdqa xmm7, xmm0 ; transpose coefficients(phase 2) 152 punpckldq xmm0, xmm1 ; xmm0=(00 10 20 30 01 11 21 31) 153 punpckhdq xmm7, xmm1 ; xmm7=(02 12 22 32 03 13 23 33) 154 movdqa xmm2, xmm4 ; transpose coefficients(phase 2) 155 punpckldq xmm4, xmm5 ; xmm4=(04 14 24 34 05 15 25 35) 156 punpckhdq xmm2, xmm5 ; xmm2=(06 16 26 36 07 17 27 37) 157 158 movdqa xmm1, xmm0 ; transpose coefficients(phase 3) 159 punpcklqdq xmm0, xmm6 ; xmm0=(00 10 20 30 40 50 60 70)=data0 160 punpckhqdq xmm1, xmm6 ; xmm1=(01 11 21 31 41 51 61 71)=data1 161 movdqa xmm5, xmm2 ; transpose coefficients(phase 3) 162 punpcklqdq xmm2, xmm3 ; xmm2=(06 16 26 36 46 56 66 76)=data6 163 punpckhqdq xmm5, xmm3 ; xmm5=(07 17 27 37 47 57 67 77)=data7 164 165 movdqa xmm6, xmm1 166 movdqa xmm3, xmm0 167 psubw xmm1, xmm2 ; xmm1=data1-data6=tmp6 168 psubw xmm0, xmm5 ; xmm0=data0-data7=tmp7 169 paddw xmm6, xmm2 ; xmm6=data1+data6=tmp1 170 paddw xmm3, xmm5 ; xmm3=data0+data7=tmp0 171 172 movdqa xmm2, XMMWORD [wk(0)] ; xmm2=(42 52 62 72 43 53 63 73) 173 movdqa xmm5, XMMWORD [wk(1)] ; xmm5=(44 54 64 74 45 55 65 75) 174 movdqa XMMWORD [wk(0)], xmm1 ; wk(0)=tmp6 175 movdqa XMMWORD [wk(1)], xmm0 ; wk(1)=tmp7 176 177 movdqa xmm1, xmm7 ; transpose coefficients(phase 3) 178 punpcklqdq xmm7, xmm2 ; xmm7=(02 12 22 32 42 52 62 72)=data2 179 punpckhqdq xmm1, xmm2 ; xmm1=(03 13 23 33 43 53 63 73)=data3 180 movdqa xmm0, xmm4 ; transpose coefficients(phase 3) 181 punpcklqdq xmm4, xmm5 ; xmm4=(04 14 24 34 44 54 64 74)=data4 182 punpckhqdq xmm0, xmm5 ; xmm0=(05 15 25 35 45 55 65 75)=data5 183 184 movdqa xmm2, xmm1 185 movdqa xmm5, xmm7 186 paddw xmm1, xmm4 ; xmm1=data3+data4=tmp3 187 paddw xmm7, xmm0 ; xmm7=data2+data5=tmp2 188 psubw xmm2, xmm4 ; xmm2=data3-data4=tmp4 189 psubw xmm5, xmm0 ; xmm5=data2-data5=tmp5 190 191 ; -- Even part 192 193 movdqa xmm4, xmm3 194 movdqa xmm0, xmm6 195 psubw xmm3, xmm1 ; xmm3=tmp13 196 psubw xmm6, xmm7 ; xmm6=tmp12 197 paddw xmm4, xmm1 ; xmm4=tmp10 198 paddw xmm0, xmm7 ; xmm0=tmp11 199 200 paddw xmm6, xmm3 201 psllw xmm6, PRE_MULTIPLY_SCALE_BITS 202 pmulhw xmm6, [GOTOFF(ebx,PW_F0707)] ; xmm6=z1 203 204 movdqa xmm1, xmm4 205 movdqa xmm7, xmm3 206 psubw xmm4, xmm0 ; xmm4=data4 207 psubw xmm3, xmm6 ; xmm3=data6 208 paddw xmm1, xmm0 ; xmm1=data0 209 paddw xmm7, xmm6 ; xmm7=data2 210 211 movdqa xmm0, XMMWORD [wk(0)] ; xmm0=tmp6 212 movdqa xmm6, XMMWORD [wk(1)] ; xmm6=tmp7 213 movdqa XMMWORD [wk(0)], xmm4 ; wk(0)=data4 214 movdqa XMMWORD [wk(1)], xmm3 ; wk(1)=data6 215 216 ; -- Odd part 217 218 paddw xmm2, xmm5 ; xmm2=tmp10 219 paddw xmm5, xmm0 ; xmm5=tmp11 220 paddw xmm0, xmm6 ; xmm0=tmp12, xmm6=tmp7 221 222 psllw xmm2, PRE_MULTIPLY_SCALE_BITS 223 psllw xmm0, PRE_MULTIPLY_SCALE_BITS 224 225 psllw xmm5, PRE_MULTIPLY_SCALE_BITS 226 pmulhw xmm5, [GOTOFF(ebx,PW_F0707)] ; xmm5=z3 227 228 movdqa xmm4, xmm2 ; xmm4=tmp10 229 psubw xmm2, xmm0 230 pmulhw xmm2, [GOTOFF(ebx,PW_F0382)] ; xmm2=z5 231 pmulhw xmm4, [GOTOFF(ebx,PW_F0541)] ; xmm4=MULTIPLY(tmp10,FIX_0_541196) 232 pmulhw xmm0, [GOTOFF(ebx,PW_F1306)] ; xmm0=MULTIPLY(tmp12,FIX_1_306562) 233 paddw xmm4, xmm2 ; xmm4=z2 234 paddw xmm0, xmm2 ; xmm0=z4 235 236 movdqa xmm3, xmm6 237 psubw xmm6, xmm5 ; xmm6=z13 238 paddw xmm3, xmm5 ; xmm3=z11 239 240 movdqa xmm2, xmm6 241 movdqa xmm5, xmm3 242 psubw xmm6, xmm4 ; xmm6=data3 243 psubw xmm3, xmm0 ; xmm3=data7 244 paddw xmm2, xmm4 ; xmm2=data5 245 paddw xmm5, xmm0 ; xmm5=data1 246 247 ; ---- Pass 2: process columns. 248 249; mov edx, POINTER [data(eax)] ; (DCTELEM *) 250 251 ; xmm1=(00 10 20 30 40 50 60 70), xmm7=(02 12 22 32 42 52 62 72) 252 ; xmm5=(01 11 21 31 41 51 61 71), xmm6=(03 13 23 33 43 53 63 73) 253 254 movdqa xmm4, xmm1 ; transpose coefficients(phase 1) 255 punpcklwd xmm1, xmm5 ; xmm1=(00 01 10 11 20 21 30 31) 256 punpckhwd xmm4, xmm5 ; xmm4=(40 41 50 51 60 61 70 71) 257 movdqa xmm0, xmm7 ; transpose coefficients(phase 1) 258 punpcklwd xmm7, xmm6 ; xmm7=(02 03 12 13 22 23 32 33) 259 punpckhwd xmm0, xmm6 ; xmm0=(42 43 52 53 62 63 72 73) 260 261 movdqa xmm5, XMMWORD [wk(0)] ; xmm5=col4 262 movdqa xmm6, XMMWORD [wk(1)] ; xmm6=col6 263 264 ; xmm5=(04 14 24 34 44 54 64 74), xmm6=(06 16 26 36 46 56 66 76) 265 ; xmm2=(05 15 25 35 45 55 65 75), xmm3=(07 17 27 37 47 57 67 77) 266 267 movdqa XMMWORD [wk(0)], xmm7 ; wk(0)=(02 03 12 13 22 23 32 33) 268 movdqa XMMWORD [wk(1)], xmm0 ; wk(1)=(42 43 52 53 62 63 72 73) 269 270 movdqa xmm7, xmm5 ; transpose coefficients(phase 1) 271 punpcklwd xmm5, xmm2 ; xmm5=(04 05 14 15 24 25 34 35) 272 punpckhwd xmm7, xmm2 ; xmm7=(44 45 54 55 64 65 74 75) 273 movdqa xmm0, xmm6 ; transpose coefficients(phase 1) 274 punpcklwd xmm6, xmm3 ; xmm6=(06 07 16 17 26 27 36 37) 275 punpckhwd xmm0, xmm3 ; xmm0=(46 47 56 57 66 67 76 77) 276 277 movdqa xmm2, xmm5 ; transpose coefficients(phase 2) 278 punpckldq xmm5, xmm6 ; xmm5=(04 05 06 07 14 15 16 17) 279 punpckhdq xmm2, xmm6 ; xmm2=(24 25 26 27 34 35 36 37) 280 movdqa xmm3, xmm7 ; transpose coefficients(phase 2) 281 punpckldq xmm7, xmm0 ; xmm7=(44 45 46 47 54 55 56 57) 282 punpckhdq xmm3, xmm0 ; xmm3=(64 65 66 67 74 75 76 77) 283 284 movdqa xmm6, XMMWORD [wk(0)] ; xmm6=(02 03 12 13 22 23 32 33) 285 movdqa xmm0, XMMWORD [wk(1)] ; xmm0=(42 43 52 53 62 63 72 73) 286 movdqa XMMWORD [wk(0)], xmm2 ; wk(0)=(24 25 26 27 34 35 36 37) 287 movdqa XMMWORD [wk(1)], xmm7 ; wk(1)=(44 45 46 47 54 55 56 57) 288 289 movdqa xmm2, xmm1 ; transpose coefficients(phase 2) 290 punpckldq xmm1, xmm6 ; xmm1=(00 01 02 03 10 11 12 13) 291 punpckhdq xmm2, xmm6 ; xmm2=(20 21 22 23 30 31 32 33) 292 movdqa xmm7, xmm4 ; transpose coefficients(phase 2) 293 punpckldq xmm4, xmm0 ; xmm4=(40 41 42 43 50 51 52 53) 294 punpckhdq xmm7, xmm0 ; xmm7=(60 61 62 63 70 71 72 73) 295 296 movdqa xmm6, xmm1 ; transpose coefficients(phase 3) 297 punpcklqdq xmm1, xmm5 ; xmm1=(00 01 02 03 04 05 06 07)=data0 298 punpckhqdq xmm6, xmm5 ; xmm6=(10 11 12 13 14 15 16 17)=data1 299 movdqa xmm0, xmm7 ; transpose coefficients(phase 3) 300 punpcklqdq xmm7, xmm3 ; xmm7=(60 61 62 63 64 65 66 67)=data6 301 punpckhqdq xmm0, xmm3 ; xmm0=(70 71 72 73 74 75 76 77)=data7 302 303 movdqa xmm5, xmm6 304 movdqa xmm3, xmm1 305 psubw xmm6, xmm7 ; xmm6=data1-data6=tmp6 306 psubw xmm1, xmm0 ; xmm1=data0-data7=tmp7 307 paddw xmm5, xmm7 ; xmm5=data1+data6=tmp1 308 paddw xmm3, xmm0 ; xmm3=data0+data7=tmp0 309 310 movdqa xmm7, XMMWORD [wk(0)] ; xmm7=(24 25 26 27 34 35 36 37) 311 movdqa xmm0, XMMWORD [wk(1)] ; xmm0=(44 45 46 47 54 55 56 57) 312 movdqa XMMWORD [wk(0)], xmm6 ; wk(0)=tmp6 313 movdqa XMMWORD [wk(1)], xmm1 ; wk(1)=tmp7 314 315 movdqa xmm6, xmm2 ; transpose coefficients(phase 3) 316 punpcklqdq xmm2, xmm7 ; xmm2=(20 21 22 23 24 25 26 27)=data2 317 punpckhqdq xmm6, xmm7 ; xmm6=(30 31 32 33 34 35 36 37)=data3 318 movdqa xmm1, xmm4 ; transpose coefficients(phase 3) 319 punpcklqdq xmm4, xmm0 ; xmm4=(40 41 42 43 44 45 46 47)=data4 320 punpckhqdq xmm1, xmm0 ; xmm1=(50 51 52 53 54 55 56 57)=data5 321 322 movdqa xmm7, xmm6 323 movdqa xmm0, xmm2 324 paddw xmm6, xmm4 ; xmm6=data3+data4=tmp3 325 paddw xmm2, xmm1 ; xmm2=data2+data5=tmp2 326 psubw xmm7, xmm4 ; xmm7=data3-data4=tmp4 327 psubw xmm0, xmm1 ; xmm0=data2-data5=tmp5 328 329 ; -- Even part 330 331 movdqa xmm4, xmm3 332 movdqa xmm1, xmm5 333 psubw xmm3, xmm6 ; xmm3=tmp13 334 psubw xmm5, xmm2 ; xmm5=tmp12 335 paddw xmm4, xmm6 ; xmm4=tmp10 336 paddw xmm1, xmm2 ; xmm1=tmp11 337 338 paddw xmm5, xmm3 339 psllw xmm5, PRE_MULTIPLY_SCALE_BITS 340 pmulhw xmm5, [GOTOFF(ebx,PW_F0707)] ; xmm5=z1 341 342 movdqa xmm6, xmm4 343 movdqa xmm2, xmm3 344 psubw xmm4, xmm1 ; xmm4=data4 345 psubw xmm3, xmm5 ; xmm3=data6 346 paddw xmm6, xmm1 ; xmm6=data0 347 paddw xmm2, xmm5 ; xmm2=data2 348 349 movdqa XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_DCTELEM)], xmm4 350 movdqa XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_DCTELEM)], xmm3 351 movdqa XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_DCTELEM)], xmm6 352 movdqa XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_DCTELEM)], xmm2 353 354 ; -- Odd part 355 356 movdqa xmm1, XMMWORD [wk(0)] ; xmm1=tmp6 357 movdqa xmm5, XMMWORD [wk(1)] ; xmm5=tmp7 358 359 paddw xmm7, xmm0 ; xmm7=tmp10 360 paddw xmm0, xmm1 ; xmm0=tmp11 361 paddw xmm1, xmm5 ; xmm1=tmp12, xmm5=tmp7 362 363 psllw xmm7, PRE_MULTIPLY_SCALE_BITS 364 psllw xmm1, PRE_MULTIPLY_SCALE_BITS 365 366 psllw xmm0, PRE_MULTIPLY_SCALE_BITS 367 pmulhw xmm0, [GOTOFF(ebx,PW_F0707)] ; xmm0=z3 368 369 movdqa xmm4, xmm7 ; xmm4=tmp10 370 psubw xmm7, xmm1 371 pmulhw xmm7, [GOTOFF(ebx,PW_F0382)] ; xmm7=z5 372 pmulhw xmm4, [GOTOFF(ebx,PW_F0541)] ; xmm4=MULTIPLY(tmp10,FIX_0_541196) 373 pmulhw xmm1, [GOTOFF(ebx,PW_F1306)] ; xmm1=MULTIPLY(tmp12,FIX_1_306562) 374 paddw xmm4, xmm7 ; xmm4=z2 375 paddw xmm1, xmm7 ; xmm1=z4 376 377 movdqa xmm3, xmm5 378 psubw xmm5, xmm0 ; xmm5=z13 379 paddw xmm3, xmm0 ; xmm3=z11 380 381 movdqa xmm6, xmm5 382 movdqa xmm2, xmm3 383 psubw xmm5, xmm4 ; xmm5=data3 384 psubw xmm3, xmm1 ; xmm3=data7 385 paddw xmm6, xmm4 ; xmm6=data5 386 paddw xmm2, xmm1 ; xmm2=data1 387 388 movdqa XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_DCTELEM)], xmm5 389 movdqa XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_DCTELEM)], xmm3 390 movdqa XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_DCTELEM)], xmm6 391 movdqa XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_DCTELEM)], xmm2 392 393; pop edi ; unused 394; pop esi ; unused 395; pop edx ; need not be preserved 396; pop ecx ; unused 397 poppic ebx 398 mov esp, ebp ; esp <- aligned ebp 399 pop esp ; esp <- original ebp 400 pop ebp 401 ret 402 403; For some reason, the OS X linker does not honor the request to align the 404; segment unless we do this. 405 align 32 406