1; 2; jfdctfst.asm - fast integer FDCT (64-bit SSE2) 3; 4; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB 5; Copyright (C) 2009, 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%include "jsimdext.inc" 23%include "jdct.inc" 24 25; -------------------------------------------------------------------------- 26 27%define CONST_BITS 8 ; 14 is also OK. 28 29%if CONST_BITS == 8 30F_0_382 equ 98 ; FIX(0.382683433) 31F_0_541 equ 139 ; FIX(0.541196100) 32F_0_707 equ 181 ; FIX(0.707106781) 33F_1_306 equ 334 ; FIX(1.306562965) 34%else 35; NASM cannot do compile-time arithmetic on floating-point constants. 36%define DESCALE(x, n) (((x) + (1 << ((n) - 1))) >> (n)) 37F_0_382 equ DESCALE( 410903207, 30 - CONST_BITS) ; FIX(0.382683433) 38F_0_541 equ DESCALE( 581104887, 30 - CONST_BITS) ; FIX(0.541196100) 39F_0_707 equ DESCALE( 759250124, 30 - CONST_BITS) ; FIX(0.707106781) 40F_1_306 equ DESCALE(1402911301, 30 - CONST_BITS) ; FIX(1.306562965) 41%endif 42 43; -------------------------------------------------------------------------- 44 SECTION SEG_CONST 45 46; PRE_MULTIPLY_SCALE_BITS <= 2 (to avoid overflow) 47; CONST_BITS + CONST_SHIFT + PRE_MULTIPLY_SCALE_BITS == 16 (for pmulhw) 48 49%define PRE_MULTIPLY_SCALE_BITS 2 50%define CONST_SHIFT (16 - PRE_MULTIPLY_SCALE_BITS - CONST_BITS) 51 52 alignz 32 53 GLOBAL_DATA(jconst_fdct_ifast_sse2) 54 55EXTN(jconst_fdct_ifast_sse2): 56 57PW_F0707 times 8 dw F_0_707 << CONST_SHIFT 58PW_F0382 times 8 dw F_0_382 << CONST_SHIFT 59PW_F0541 times 8 dw F_0_541 << CONST_SHIFT 60PW_F1306 times 8 dw F_1_306 << CONST_SHIFT 61 62 alignz 32 63 64; -------------------------------------------------------------------------- 65 SECTION SEG_TEXT 66 BITS 64 67; 68; Perform the forward DCT on one block of samples. 69; 70; GLOBAL(void) 71; jsimd_fdct_ifast_sse2(DCTELEM *data) 72; 73 74; r10 = DCTELEM *data 75 76%define wk(i) rbp - (WK_NUM - (i)) * SIZEOF_XMMWORD ; xmmword wk[WK_NUM] 77%define WK_NUM 2 78 79 align 32 80 GLOBAL_FUNCTION(jsimd_fdct_ifast_sse2) 81 82EXTN(jsimd_fdct_ifast_sse2): 83 push rbp 84 mov rax, rsp ; rax = original rbp 85 sub rsp, byte 4 86 and rsp, byte (-SIZEOF_XMMWORD) ; align to 128 bits 87 mov [rsp], rax 88 mov rbp, rsp ; rbp = aligned rbp 89 lea rsp, [wk(0)] 90 collect_args 1 91 92 ; ---- Pass 1: process rows. 93 94 mov rdx, r10 ; (DCTELEM *) 95 96 movdqa xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_DCTELEM)] 97 movdqa xmm1, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_DCTELEM)] 98 movdqa xmm2, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_DCTELEM)] 99 movdqa xmm3, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_DCTELEM)] 100 101 ; xmm0=(00 01 02 03 04 05 06 07), xmm2=(20 21 22 23 24 25 26 27) 102 ; xmm1=(10 11 12 13 14 15 16 17), xmm3=(30 31 32 33 34 35 36 37) 103 104 movdqa xmm4, xmm0 ; transpose coefficients(phase 1) 105 punpcklwd xmm0, xmm1 ; xmm0=(00 10 01 11 02 12 03 13) 106 punpckhwd xmm4, xmm1 ; xmm4=(04 14 05 15 06 16 07 17) 107 movdqa xmm5, xmm2 ; transpose coefficients(phase 1) 108 punpcklwd xmm2, xmm3 ; xmm2=(20 30 21 31 22 32 23 33) 109 punpckhwd xmm5, xmm3 ; xmm5=(24 34 25 35 26 36 27 37) 110 111 movdqa xmm6, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_DCTELEM)] 112 movdqa xmm7, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_DCTELEM)] 113 movdqa xmm1, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_DCTELEM)] 114 movdqa xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_DCTELEM)] 115 116 ; xmm6=( 4 12 20 28 36 44 52 60), xmm1=( 6 14 22 30 38 46 54 62) 117 ; xmm7=( 5 13 21 29 37 45 53 61), xmm3=( 7 15 23 31 39 47 55 63) 118 119 movdqa XMMWORD [wk(0)], xmm2 ; wk(0)=(20 30 21 31 22 32 23 33) 120 movdqa XMMWORD [wk(1)], xmm5 ; wk(1)=(24 34 25 35 26 36 27 37) 121 122 movdqa xmm2, xmm6 ; transpose coefficients(phase 1) 123 punpcklwd xmm6, xmm7 ; xmm6=(40 50 41 51 42 52 43 53) 124 punpckhwd xmm2, xmm7 ; xmm2=(44 54 45 55 46 56 47 57) 125 movdqa xmm5, xmm1 ; transpose coefficients(phase 1) 126 punpcklwd xmm1, xmm3 ; xmm1=(60 70 61 71 62 72 63 73) 127 punpckhwd xmm5, xmm3 ; xmm5=(64 74 65 75 66 76 67 77) 128 129 movdqa xmm7, xmm6 ; transpose coefficients(phase 2) 130 punpckldq xmm6, xmm1 ; xmm6=(40 50 60 70 41 51 61 71) 131 punpckhdq xmm7, xmm1 ; xmm7=(42 52 62 72 43 53 63 73) 132 movdqa xmm3, xmm2 ; transpose coefficients(phase 2) 133 punpckldq xmm2, xmm5 ; xmm2=(44 54 64 74 45 55 65 75) 134 punpckhdq xmm3, xmm5 ; xmm3=(46 56 66 76 47 57 67 77) 135 136 movdqa xmm1, XMMWORD [wk(0)] ; xmm1=(20 30 21 31 22 32 23 33) 137 movdqa xmm5, XMMWORD [wk(1)] ; xmm5=(24 34 25 35 26 36 27 37) 138 movdqa XMMWORD [wk(0)], xmm7 ; wk(0)=(42 52 62 72 43 53 63 73) 139 movdqa XMMWORD [wk(1)], xmm2 ; wk(1)=(44 54 64 74 45 55 65 75) 140 141 movdqa xmm7, xmm0 ; transpose coefficients(phase 2) 142 punpckldq xmm0, xmm1 ; xmm0=(00 10 20 30 01 11 21 31) 143 punpckhdq xmm7, xmm1 ; xmm7=(02 12 22 32 03 13 23 33) 144 movdqa xmm2, xmm4 ; transpose coefficients(phase 2) 145 punpckldq xmm4, xmm5 ; xmm4=(04 14 24 34 05 15 25 35) 146 punpckhdq xmm2, xmm5 ; xmm2=(06 16 26 36 07 17 27 37) 147 148 movdqa xmm1, xmm0 ; transpose coefficients(phase 3) 149 punpcklqdq xmm0, xmm6 ; xmm0=(00 10 20 30 40 50 60 70)=data0 150 punpckhqdq xmm1, xmm6 ; xmm1=(01 11 21 31 41 51 61 71)=data1 151 movdqa xmm5, xmm2 ; transpose coefficients(phase 3) 152 punpcklqdq xmm2, xmm3 ; xmm2=(06 16 26 36 46 56 66 76)=data6 153 punpckhqdq xmm5, xmm3 ; xmm5=(07 17 27 37 47 57 67 77)=data7 154 155 movdqa xmm6, xmm1 156 movdqa xmm3, xmm0 157 psubw xmm1, xmm2 ; xmm1=data1-data6=tmp6 158 psubw xmm0, xmm5 ; xmm0=data0-data7=tmp7 159 paddw xmm6, xmm2 ; xmm6=data1+data6=tmp1 160 paddw xmm3, xmm5 ; xmm3=data0+data7=tmp0 161 162 movdqa xmm2, XMMWORD [wk(0)] ; xmm2=(42 52 62 72 43 53 63 73) 163 movdqa xmm5, XMMWORD [wk(1)] ; xmm5=(44 54 64 74 45 55 65 75) 164 movdqa XMMWORD [wk(0)], xmm1 ; wk(0)=tmp6 165 movdqa XMMWORD [wk(1)], xmm0 ; wk(1)=tmp7 166 167 movdqa xmm1, xmm7 ; transpose coefficients(phase 3) 168 punpcklqdq xmm7, xmm2 ; xmm7=(02 12 22 32 42 52 62 72)=data2 169 punpckhqdq xmm1, xmm2 ; xmm1=(03 13 23 33 43 53 63 73)=data3 170 movdqa xmm0, xmm4 ; transpose coefficients(phase 3) 171 punpcklqdq xmm4, xmm5 ; xmm4=(04 14 24 34 44 54 64 74)=data4 172 punpckhqdq xmm0, xmm5 ; xmm0=(05 15 25 35 45 55 65 75)=data5 173 174 movdqa xmm2, xmm1 175 movdqa xmm5, xmm7 176 paddw xmm1, xmm4 ; xmm1=data3+data4=tmp3 177 paddw xmm7, xmm0 ; xmm7=data2+data5=tmp2 178 psubw xmm2, xmm4 ; xmm2=data3-data4=tmp4 179 psubw xmm5, xmm0 ; xmm5=data2-data5=tmp5 180 181 ; -- Even part 182 183 movdqa xmm4, xmm3 184 movdqa xmm0, xmm6 185 psubw xmm3, xmm1 ; xmm3=tmp13 186 psubw xmm6, xmm7 ; xmm6=tmp12 187 paddw xmm4, xmm1 ; xmm4=tmp10 188 paddw xmm0, xmm7 ; xmm0=tmp11 189 190 paddw xmm6, xmm3 191 psllw xmm6, PRE_MULTIPLY_SCALE_BITS 192 pmulhw xmm6, [rel PW_F0707] ; xmm6=z1 193 194 movdqa xmm1, xmm4 195 movdqa xmm7, xmm3 196 psubw xmm4, xmm0 ; xmm4=data4 197 psubw xmm3, xmm6 ; xmm3=data6 198 paddw xmm1, xmm0 ; xmm1=data0 199 paddw xmm7, xmm6 ; xmm7=data2 200 201 movdqa xmm0, XMMWORD [wk(0)] ; xmm0=tmp6 202 movdqa xmm6, XMMWORD [wk(1)] ; xmm6=tmp7 203 movdqa XMMWORD [wk(0)], xmm4 ; wk(0)=data4 204 movdqa XMMWORD [wk(1)], xmm3 ; wk(1)=data6 205 206 ; -- Odd part 207 208 paddw xmm2, xmm5 ; xmm2=tmp10 209 paddw xmm5, xmm0 ; xmm5=tmp11 210 paddw xmm0, xmm6 ; xmm0=tmp12, xmm6=tmp7 211 212 psllw xmm2, PRE_MULTIPLY_SCALE_BITS 213 psllw xmm0, PRE_MULTIPLY_SCALE_BITS 214 215 psllw xmm5, PRE_MULTIPLY_SCALE_BITS 216 pmulhw xmm5, [rel PW_F0707] ; xmm5=z3 217 218 movdqa xmm4, xmm2 ; xmm4=tmp10 219 psubw xmm2, xmm0 220 pmulhw xmm2, [rel PW_F0382] ; xmm2=z5 221 pmulhw xmm4, [rel PW_F0541] ; xmm4=MULTIPLY(tmp10,FIX_0_541196) 222 pmulhw xmm0, [rel PW_F1306] ; xmm0=MULTIPLY(tmp12,FIX_1_306562) 223 paddw xmm4, xmm2 ; xmm4=z2 224 paddw xmm0, xmm2 ; xmm0=z4 225 226 movdqa xmm3, xmm6 227 psubw xmm6, xmm5 ; xmm6=z13 228 paddw xmm3, xmm5 ; xmm3=z11 229 230 movdqa xmm2, xmm6 231 movdqa xmm5, xmm3 232 psubw xmm6, xmm4 ; xmm6=data3 233 psubw xmm3, xmm0 ; xmm3=data7 234 paddw xmm2, xmm4 ; xmm2=data5 235 paddw xmm5, xmm0 ; xmm5=data1 236 237 ; ---- Pass 2: process columns. 238 239 ; xmm1=(00 10 20 30 40 50 60 70), xmm7=(02 12 22 32 42 52 62 72) 240 ; xmm5=(01 11 21 31 41 51 61 71), xmm6=(03 13 23 33 43 53 63 73) 241 242 movdqa xmm4, xmm1 ; transpose coefficients(phase 1) 243 punpcklwd xmm1, xmm5 ; xmm1=(00 01 10 11 20 21 30 31) 244 punpckhwd xmm4, xmm5 ; xmm4=(40 41 50 51 60 61 70 71) 245 movdqa xmm0, xmm7 ; transpose coefficients(phase 1) 246 punpcklwd xmm7, xmm6 ; xmm7=(02 03 12 13 22 23 32 33) 247 punpckhwd xmm0, xmm6 ; xmm0=(42 43 52 53 62 63 72 73) 248 249 movdqa xmm5, XMMWORD [wk(0)] ; xmm5=col4 250 movdqa xmm6, XMMWORD [wk(1)] ; xmm6=col6 251 252 ; xmm5=(04 14 24 34 44 54 64 74), xmm6=(06 16 26 36 46 56 66 76) 253 ; xmm2=(05 15 25 35 45 55 65 75), xmm3=(07 17 27 37 47 57 67 77) 254 255 movdqa XMMWORD [wk(0)], xmm7 ; wk(0)=(02 03 12 13 22 23 32 33) 256 movdqa XMMWORD [wk(1)], xmm0 ; wk(1)=(42 43 52 53 62 63 72 73) 257 258 movdqa xmm7, xmm5 ; transpose coefficients(phase 1) 259 punpcklwd xmm5, xmm2 ; xmm5=(04 05 14 15 24 25 34 35) 260 punpckhwd xmm7, xmm2 ; xmm7=(44 45 54 55 64 65 74 75) 261 movdqa xmm0, xmm6 ; transpose coefficients(phase 1) 262 punpcklwd xmm6, xmm3 ; xmm6=(06 07 16 17 26 27 36 37) 263 punpckhwd xmm0, xmm3 ; xmm0=(46 47 56 57 66 67 76 77) 264 265 movdqa xmm2, xmm5 ; transpose coefficients(phase 2) 266 punpckldq xmm5, xmm6 ; xmm5=(04 05 06 07 14 15 16 17) 267 punpckhdq xmm2, xmm6 ; xmm2=(24 25 26 27 34 35 36 37) 268 movdqa xmm3, xmm7 ; transpose coefficients(phase 2) 269 punpckldq xmm7, xmm0 ; xmm7=(44 45 46 47 54 55 56 57) 270 punpckhdq xmm3, xmm0 ; xmm3=(64 65 66 67 74 75 76 77) 271 272 movdqa xmm6, XMMWORD [wk(0)] ; xmm6=(02 03 12 13 22 23 32 33) 273 movdqa xmm0, XMMWORD [wk(1)] ; xmm0=(42 43 52 53 62 63 72 73) 274 movdqa XMMWORD [wk(0)], xmm2 ; wk(0)=(24 25 26 27 34 35 36 37) 275 movdqa XMMWORD [wk(1)], xmm7 ; wk(1)=(44 45 46 47 54 55 56 57) 276 277 movdqa xmm2, xmm1 ; transpose coefficients(phase 2) 278 punpckldq xmm1, xmm6 ; xmm1=(00 01 02 03 10 11 12 13) 279 punpckhdq xmm2, xmm6 ; xmm2=(20 21 22 23 30 31 32 33) 280 movdqa xmm7, xmm4 ; transpose coefficients(phase 2) 281 punpckldq xmm4, xmm0 ; xmm4=(40 41 42 43 50 51 52 53) 282 punpckhdq xmm7, xmm0 ; xmm7=(60 61 62 63 70 71 72 73) 283 284 movdqa xmm6, xmm1 ; transpose coefficients(phase 3) 285 punpcklqdq xmm1, xmm5 ; xmm1=(00 01 02 03 04 05 06 07)=data0 286 punpckhqdq xmm6, xmm5 ; xmm6=(10 11 12 13 14 15 16 17)=data1 287 movdqa xmm0, xmm7 ; transpose coefficients(phase 3) 288 punpcklqdq xmm7, xmm3 ; xmm7=(60 61 62 63 64 65 66 67)=data6 289 punpckhqdq xmm0, xmm3 ; xmm0=(70 71 72 73 74 75 76 77)=data7 290 291 movdqa xmm5, xmm6 292 movdqa xmm3, xmm1 293 psubw xmm6, xmm7 ; xmm6=data1-data6=tmp6 294 psubw xmm1, xmm0 ; xmm1=data0-data7=tmp7 295 paddw xmm5, xmm7 ; xmm5=data1+data6=tmp1 296 paddw xmm3, xmm0 ; xmm3=data0+data7=tmp0 297 298 movdqa xmm7, XMMWORD [wk(0)] ; xmm7=(24 25 26 27 34 35 36 37) 299 movdqa xmm0, XMMWORD [wk(1)] ; xmm0=(44 45 46 47 54 55 56 57) 300 movdqa XMMWORD [wk(0)], xmm6 ; wk(0)=tmp6 301 movdqa XMMWORD [wk(1)], xmm1 ; wk(1)=tmp7 302 303 movdqa xmm6, xmm2 ; transpose coefficients(phase 3) 304 punpcklqdq xmm2, xmm7 ; xmm2=(20 21 22 23 24 25 26 27)=data2 305 punpckhqdq xmm6, xmm7 ; xmm6=(30 31 32 33 34 35 36 37)=data3 306 movdqa xmm1, xmm4 ; transpose coefficients(phase 3) 307 punpcklqdq xmm4, xmm0 ; xmm4=(40 41 42 43 44 45 46 47)=data4 308 punpckhqdq xmm1, xmm0 ; xmm1=(50 51 52 53 54 55 56 57)=data5 309 310 movdqa xmm7, xmm6 311 movdqa xmm0, xmm2 312 paddw xmm6, xmm4 ; xmm6=data3+data4=tmp3 313 paddw xmm2, xmm1 ; xmm2=data2+data5=tmp2 314 psubw xmm7, xmm4 ; xmm7=data3-data4=tmp4 315 psubw xmm0, xmm1 ; xmm0=data2-data5=tmp5 316 317 ; -- Even part 318 319 movdqa xmm4, xmm3 320 movdqa xmm1, xmm5 321 psubw xmm3, xmm6 ; xmm3=tmp13 322 psubw xmm5, xmm2 ; xmm5=tmp12 323 paddw xmm4, xmm6 ; xmm4=tmp10 324 paddw xmm1, xmm2 ; xmm1=tmp11 325 326 paddw xmm5, xmm3 327 psllw xmm5, PRE_MULTIPLY_SCALE_BITS 328 pmulhw xmm5, [rel PW_F0707] ; xmm5=z1 329 330 movdqa xmm6, xmm4 331 movdqa xmm2, xmm3 332 psubw xmm4, xmm1 ; xmm4=data4 333 psubw xmm3, xmm5 ; xmm3=data6 334 paddw xmm6, xmm1 ; xmm6=data0 335 paddw xmm2, xmm5 ; xmm2=data2 336 337 movdqa XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_DCTELEM)], xmm4 338 movdqa XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_DCTELEM)], xmm3 339 movdqa XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_DCTELEM)], xmm6 340 movdqa XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_DCTELEM)], xmm2 341 342 ; -- Odd part 343 344 movdqa xmm1, XMMWORD [wk(0)] ; xmm1=tmp6 345 movdqa xmm5, XMMWORD [wk(1)] ; xmm5=tmp7 346 347 paddw xmm7, xmm0 ; xmm7=tmp10 348 paddw xmm0, xmm1 ; xmm0=tmp11 349 paddw xmm1, xmm5 ; xmm1=tmp12, xmm5=tmp7 350 351 psllw xmm7, PRE_MULTIPLY_SCALE_BITS 352 psllw xmm1, PRE_MULTIPLY_SCALE_BITS 353 354 psllw xmm0, PRE_MULTIPLY_SCALE_BITS 355 pmulhw xmm0, [rel PW_F0707] ; xmm0=z3 356 357 movdqa xmm4, xmm7 ; xmm4=tmp10 358 psubw xmm7, xmm1 359 pmulhw xmm7, [rel PW_F0382] ; xmm7=z5 360 pmulhw xmm4, [rel PW_F0541] ; xmm4=MULTIPLY(tmp10,FIX_0_541196) 361 pmulhw xmm1, [rel PW_F1306] ; xmm1=MULTIPLY(tmp12,FIX_1_306562) 362 paddw xmm4, xmm7 ; xmm4=z2 363 paddw xmm1, xmm7 ; xmm1=z4 364 365 movdqa xmm3, xmm5 366 psubw xmm5, xmm0 ; xmm5=z13 367 paddw xmm3, xmm0 ; xmm3=z11 368 369 movdqa xmm6, xmm5 370 movdqa xmm2, xmm3 371 psubw xmm5, xmm4 ; xmm5=data3 372 psubw xmm3, xmm1 ; xmm3=data7 373 paddw xmm6, xmm4 ; xmm6=data5 374 paddw xmm2, xmm1 ; xmm2=data1 375 376 movdqa XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_DCTELEM)], xmm5 377 movdqa XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_DCTELEM)], xmm3 378 movdqa XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_DCTELEM)], xmm6 379 movdqa XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_DCTELEM)], xmm2 380 381 uncollect_args 1 382 mov rsp, rbp ; rsp <- aligned rbp 383 pop rsp ; rsp <- original rbp 384 pop rbp 385 ret 386 387; For some reason, the OS X linker does not honor the request to align the 388; segment unless we do this. 389 align 32 390