jfdctfst-sse2.asm (17672B)
1 ; 2 ; jfdctfst.asm - fast integer FDCT (SSE2) 3 ; 4 ; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB 5 ; Copyright (C) 2016, 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 fast, not so accurate integer implementation of 14 ; the forward DCT (Discrete Cosine Transform). The following code is 15 ; based directly on the IJG's original jfdctfst.c; see the jfdctfst.c 16 ; for more details. 17 18 %include "jsimdext.inc" 19 %include "jdct.inc" 20 21 ; -------------------------------------------------------------------------- 22 23 %define CONST_BITS 8 ; 14 is also OK. 24 25 %if CONST_BITS == 8 26 F_0_382 equ 98 ; FIX(0.382683433) 27 F_0_541 equ 139 ; FIX(0.541196100) 28 F_0_707 equ 181 ; FIX(0.707106781) 29 F_1_306 equ 334 ; FIX(1.306562965) 30 %else 31 ; NASM cannot do compile-time arithmetic on floating-point constants. 32 %define DESCALE(x, n) (((x) + (1 << ((n) - 1))) >> (n)) 33 F_0_382 equ DESCALE( 410903207, 30 - CONST_BITS) ; FIX(0.382683433) 34 F_0_541 equ DESCALE( 581104887, 30 - CONST_BITS) ; FIX(0.541196100) 35 F_0_707 equ DESCALE( 759250124, 30 - CONST_BITS) ; FIX(0.707106781) 36 F_1_306 equ DESCALE(1402911301, 30 - CONST_BITS) ; FIX(1.306562965) 37 %endif 38 39 ; -------------------------------------------------------------------------- 40 SECTION SEG_CONST 41 42 ; PRE_MULTIPLY_SCALE_BITS <= 2 (to avoid overflow) 43 ; CONST_BITS + CONST_SHIFT + PRE_MULTIPLY_SCALE_BITS == 16 (for pmulhw) 44 45 %define PRE_MULTIPLY_SCALE_BITS 2 46 %define CONST_SHIFT (16 - PRE_MULTIPLY_SCALE_BITS - CONST_BITS) 47 48 ALIGNZ 32 49 GLOBAL_DATA(jconst_fdct_ifast_sse2) 50 51 EXTN(jconst_fdct_ifast_sse2): 52 53 PW_F0707 times 8 dw F_0_707 << CONST_SHIFT 54 PW_F0382 times 8 dw F_0_382 << CONST_SHIFT 55 PW_F0541 times 8 dw F_0_541 << CONST_SHIFT 56 PW_F1306 times 8 dw F_1_306 << CONST_SHIFT 57 58 ALIGNZ 32 59 60 ; -------------------------------------------------------------------------- 61 SECTION SEG_TEXT 62 BITS 32 63 ; 64 ; Perform the forward DCT on one block of samples. 65 ; 66 ; GLOBAL(void) 67 ; jsimd_fdct_ifast_sse2(DCTELEM *data) 68 ; 69 70 %define data(b) (b) + 8 ; DCTELEM *data 71 72 %define original_ebp ebp + 0 73 %define wk(i) ebp - (WK_NUM - (i)) * SIZEOF_XMMWORD 74 ; xmmword wk[WK_NUM] 75 %define WK_NUM 2 76 77 align 32 78 GLOBAL_FUNCTION(jsimd_fdct_ifast_sse2) 79 80 EXTN(jsimd_fdct_ifast_sse2): 81 push ebp 82 mov eax, esp ; eax = original ebp 83 sub esp, byte 4 84 and esp, byte (-SIZEOF_XMMWORD) ; align to 128 bits 85 mov [esp], eax 86 mov ebp, esp ; ebp = aligned ebp 87 lea esp, [wk(0)] 88 PUSHPIC ebx 89 ; push ecx ; unused 90 ; push edx ; need not be preserved 91 ; push esi ; unused 92 ; push edi ; unused 93 94 GET_GOT ebx ; get GOT address 95 96 ; ---- Pass 1: process rows. 97 98 mov edx, POINTER [data(eax)] ; (DCTELEM *) 99 100 movdqa xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_DCTELEM)] 101 movdqa xmm1, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_DCTELEM)] 102 movdqa xmm2, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_DCTELEM)] 103 movdqa xmm3, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_DCTELEM)] 104 105 ; xmm0=(00 01 02 03 04 05 06 07), xmm2=(20 21 22 23 24 25 26 27) 106 ; xmm1=(10 11 12 13 14 15 16 17), xmm3=(30 31 32 33 34 35 36 37) 107 108 movdqa xmm4, xmm0 ; transpose coefficients(phase 1) 109 punpcklwd xmm0, xmm1 ; xmm0=(00 10 01 11 02 12 03 13) 110 punpckhwd xmm4, xmm1 ; xmm4=(04 14 05 15 06 16 07 17) 111 movdqa xmm5, xmm2 ; transpose coefficients(phase 1) 112 punpcklwd xmm2, xmm3 ; xmm2=(20 30 21 31 22 32 23 33) 113 punpckhwd xmm5, xmm3 ; xmm5=(24 34 25 35 26 36 27 37) 114 115 movdqa xmm6, XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_DCTELEM)] 116 movdqa xmm7, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_DCTELEM)] 117 movdqa xmm1, XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_DCTELEM)] 118 movdqa xmm3, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_DCTELEM)] 119 120 ; xmm6=( 4 12 20 28 36 44 52 60), xmm1=( 6 14 22 30 38 46 54 62) 121 ; xmm7=( 5 13 21 29 37 45 53 61), xmm3=( 7 15 23 31 39 47 55 63) 122 123 movdqa XMMWORD [wk(0)], xmm2 ; wk(0)=(20 30 21 31 22 32 23 33) 124 movdqa XMMWORD [wk(1)], xmm5 ; wk(1)=(24 34 25 35 26 36 27 37) 125 126 movdqa xmm2, xmm6 ; transpose coefficients(phase 1) 127 punpcklwd xmm6, xmm7 ; xmm6=(40 50 41 51 42 52 43 53) 128 punpckhwd xmm2, xmm7 ; xmm2=(44 54 45 55 46 56 47 57) 129 movdqa xmm5, xmm1 ; transpose coefficients(phase 1) 130 punpcklwd xmm1, xmm3 ; xmm1=(60 70 61 71 62 72 63 73) 131 punpckhwd xmm5, xmm3 ; xmm5=(64 74 65 75 66 76 67 77) 132 133 movdqa xmm7, xmm6 ; transpose coefficients(phase 2) 134 punpckldq xmm6, xmm1 ; xmm6=(40 50 60 70 41 51 61 71) 135 punpckhdq xmm7, xmm1 ; xmm7=(42 52 62 72 43 53 63 73) 136 movdqa xmm3, xmm2 ; transpose coefficients(phase 2) 137 punpckldq xmm2, xmm5 ; xmm2=(44 54 64 74 45 55 65 75) 138 punpckhdq xmm3, xmm5 ; xmm3=(46 56 66 76 47 57 67 77) 139 140 movdqa xmm1, XMMWORD [wk(0)] ; xmm1=(20 30 21 31 22 32 23 33) 141 movdqa xmm5, XMMWORD [wk(1)] ; xmm5=(24 34 25 35 26 36 27 37) 142 movdqa XMMWORD [wk(0)], xmm7 ; wk(0)=(42 52 62 72 43 53 63 73) 143 movdqa XMMWORD [wk(1)], xmm2 ; wk(1)=(44 54 64 74 45 55 65 75) 144 145 movdqa xmm7, xmm0 ; transpose coefficients(phase 2) 146 punpckldq xmm0, xmm1 ; xmm0=(00 10 20 30 01 11 21 31) 147 punpckhdq xmm7, xmm1 ; xmm7=(02 12 22 32 03 13 23 33) 148 movdqa xmm2, xmm4 ; transpose coefficients(phase 2) 149 punpckldq xmm4, xmm5 ; xmm4=(04 14 24 34 05 15 25 35) 150 punpckhdq xmm2, xmm5 ; xmm2=(06 16 26 36 07 17 27 37) 151 152 movdqa xmm1, xmm0 ; transpose coefficients(phase 3) 153 punpcklqdq xmm0, xmm6 ; xmm0=(00 10 20 30 40 50 60 70)=data0 154 punpckhqdq xmm1, xmm6 ; xmm1=(01 11 21 31 41 51 61 71)=data1 155 movdqa xmm5, xmm2 ; transpose coefficients(phase 3) 156 punpcklqdq xmm2, xmm3 ; xmm2=(06 16 26 36 46 56 66 76)=data6 157 punpckhqdq xmm5, xmm3 ; xmm5=(07 17 27 37 47 57 67 77)=data7 158 159 movdqa xmm6, xmm1 160 movdqa xmm3, xmm0 161 psubw xmm1, xmm2 ; xmm1=data1-data6=tmp6 162 psubw xmm0, xmm5 ; xmm0=data0-data7=tmp7 163 paddw xmm6, xmm2 ; xmm6=data1+data6=tmp1 164 paddw xmm3, xmm5 ; xmm3=data0+data7=tmp0 165 166 movdqa xmm2, XMMWORD [wk(0)] ; xmm2=(42 52 62 72 43 53 63 73) 167 movdqa xmm5, XMMWORD [wk(1)] ; xmm5=(44 54 64 74 45 55 65 75) 168 movdqa XMMWORD [wk(0)], xmm1 ; wk(0)=tmp6 169 movdqa XMMWORD [wk(1)], xmm0 ; wk(1)=tmp7 170 171 movdqa xmm1, xmm7 ; transpose coefficients(phase 3) 172 punpcklqdq xmm7, xmm2 ; xmm7=(02 12 22 32 42 52 62 72)=data2 173 punpckhqdq xmm1, xmm2 ; xmm1=(03 13 23 33 43 53 63 73)=data3 174 movdqa xmm0, xmm4 ; transpose coefficients(phase 3) 175 punpcklqdq xmm4, xmm5 ; xmm4=(04 14 24 34 44 54 64 74)=data4 176 punpckhqdq xmm0, xmm5 ; xmm0=(05 15 25 35 45 55 65 75)=data5 177 178 movdqa xmm2, xmm1 179 movdqa xmm5, xmm7 180 paddw xmm1, xmm4 ; xmm1=data3+data4=tmp3 181 paddw xmm7, xmm0 ; xmm7=data2+data5=tmp2 182 psubw xmm2, xmm4 ; xmm2=data3-data4=tmp4 183 psubw xmm5, xmm0 ; xmm5=data2-data5=tmp5 184 185 ; -- Even part 186 187 movdqa xmm4, xmm3 188 movdqa xmm0, xmm6 189 psubw xmm3, xmm1 ; xmm3=tmp13 190 psubw xmm6, xmm7 ; xmm6=tmp12 191 paddw xmm4, xmm1 ; xmm4=tmp10 192 paddw xmm0, xmm7 ; xmm0=tmp11 193 194 paddw xmm6, xmm3 195 psllw xmm6, PRE_MULTIPLY_SCALE_BITS 196 pmulhw xmm6, [GOTOFF(ebx,PW_F0707)] ; xmm6=z1 197 198 movdqa xmm1, xmm4 199 movdqa xmm7, xmm3 200 psubw xmm4, xmm0 ; xmm4=data4 201 psubw xmm3, xmm6 ; xmm3=data6 202 paddw xmm1, xmm0 ; xmm1=data0 203 paddw xmm7, xmm6 ; xmm7=data2 204 205 movdqa xmm0, XMMWORD [wk(0)] ; xmm0=tmp6 206 movdqa xmm6, XMMWORD [wk(1)] ; xmm6=tmp7 207 movdqa XMMWORD [wk(0)], xmm4 ; wk(0)=data4 208 movdqa XMMWORD [wk(1)], xmm3 ; wk(1)=data6 209 210 ; -- Odd part 211 212 paddw xmm2, xmm5 ; xmm2=tmp10 213 paddw xmm5, xmm0 ; xmm5=tmp11 214 paddw xmm0, xmm6 ; xmm0=tmp12, xmm6=tmp7 215 216 psllw xmm2, PRE_MULTIPLY_SCALE_BITS 217 psllw xmm0, PRE_MULTIPLY_SCALE_BITS 218 219 psllw xmm5, PRE_MULTIPLY_SCALE_BITS 220 pmulhw xmm5, [GOTOFF(ebx,PW_F0707)] ; xmm5=z3 221 222 movdqa xmm4, xmm2 ; xmm4=tmp10 223 psubw xmm2, xmm0 224 pmulhw xmm2, [GOTOFF(ebx,PW_F0382)] ; xmm2=z5 225 pmulhw xmm4, [GOTOFF(ebx,PW_F0541)] ; xmm4=MULTIPLY(tmp10,FIX_0_541196) 226 pmulhw xmm0, [GOTOFF(ebx,PW_F1306)] ; xmm0=MULTIPLY(tmp12,FIX_1_306562) 227 paddw xmm4, xmm2 ; xmm4=z2 228 paddw xmm0, xmm2 ; xmm0=z4 229 230 movdqa xmm3, xmm6 231 psubw xmm6, xmm5 ; xmm6=z13 232 paddw xmm3, xmm5 ; xmm3=z11 233 234 movdqa xmm2, xmm6 235 movdqa xmm5, xmm3 236 psubw xmm6, xmm4 ; xmm6=data3 237 psubw xmm3, xmm0 ; xmm3=data7 238 paddw xmm2, xmm4 ; xmm2=data5 239 paddw xmm5, xmm0 ; xmm5=data1 240 241 ; ---- Pass 2: process columns. 242 243 ; mov edx, POINTER [data(eax)] ; (DCTELEM *) 244 245 ; xmm1=(00 10 20 30 40 50 60 70), xmm7=(02 12 22 32 42 52 62 72) 246 ; xmm5=(01 11 21 31 41 51 61 71), xmm6=(03 13 23 33 43 53 63 73) 247 248 movdqa xmm4, xmm1 ; transpose coefficients(phase 1) 249 punpcklwd xmm1, xmm5 ; xmm1=(00 01 10 11 20 21 30 31) 250 punpckhwd xmm4, xmm5 ; xmm4=(40 41 50 51 60 61 70 71) 251 movdqa xmm0, xmm7 ; transpose coefficients(phase 1) 252 punpcklwd xmm7, xmm6 ; xmm7=(02 03 12 13 22 23 32 33) 253 punpckhwd xmm0, xmm6 ; xmm0=(42 43 52 53 62 63 72 73) 254 255 movdqa xmm5, XMMWORD [wk(0)] ; xmm5=col4 256 movdqa xmm6, XMMWORD [wk(1)] ; xmm6=col6 257 258 ; xmm5=(04 14 24 34 44 54 64 74), xmm6=(06 16 26 36 46 56 66 76) 259 ; xmm2=(05 15 25 35 45 55 65 75), xmm3=(07 17 27 37 47 57 67 77) 260 261 movdqa XMMWORD [wk(0)], xmm7 ; wk(0)=(02 03 12 13 22 23 32 33) 262 movdqa XMMWORD [wk(1)], xmm0 ; wk(1)=(42 43 52 53 62 63 72 73) 263 264 movdqa xmm7, xmm5 ; transpose coefficients(phase 1) 265 punpcklwd xmm5, xmm2 ; xmm5=(04 05 14 15 24 25 34 35) 266 punpckhwd xmm7, xmm2 ; xmm7=(44 45 54 55 64 65 74 75) 267 movdqa xmm0, xmm6 ; transpose coefficients(phase 1) 268 punpcklwd xmm6, xmm3 ; xmm6=(06 07 16 17 26 27 36 37) 269 punpckhwd xmm0, xmm3 ; xmm0=(46 47 56 57 66 67 76 77) 270 271 movdqa xmm2, xmm5 ; transpose coefficients(phase 2) 272 punpckldq xmm5, xmm6 ; xmm5=(04 05 06 07 14 15 16 17) 273 punpckhdq xmm2, xmm6 ; xmm2=(24 25 26 27 34 35 36 37) 274 movdqa xmm3, xmm7 ; transpose coefficients(phase 2) 275 punpckldq xmm7, xmm0 ; xmm7=(44 45 46 47 54 55 56 57) 276 punpckhdq xmm3, xmm0 ; xmm3=(64 65 66 67 74 75 76 77) 277 278 movdqa xmm6, XMMWORD [wk(0)] ; xmm6=(02 03 12 13 22 23 32 33) 279 movdqa xmm0, XMMWORD [wk(1)] ; xmm0=(42 43 52 53 62 63 72 73) 280 movdqa XMMWORD [wk(0)], xmm2 ; wk(0)=(24 25 26 27 34 35 36 37) 281 movdqa XMMWORD [wk(1)], xmm7 ; wk(1)=(44 45 46 47 54 55 56 57) 282 283 movdqa xmm2, xmm1 ; transpose coefficients(phase 2) 284 punpckldq xmm1, xmm6 ; xmm1=(00 01 02 03 10 11 12 13) 285 punpckhdq xmm2, xmm6 ; xmm2=(20 21 22 23 30 31 32 33) 286 movdqa xmm7, xmm4 ; transpose coefficients(phase 2) 287 punpckldq xmm4, xmm0 ; xmm4=(40 41 42 43 50 51 52 53) 288 punpckhdq xmm7, xmm0 ; xmm7=(60 61 62 63 70 71 72 73) 289 290 movdqa xmm6, xmm1 ; transpose coefficients(phase 3) 291 punpcklqdq xmm1, xmm5 ; xmm1=(00 01 02 03 04 05 06 07)=data0 292 punpckhqdq xmm6, xmm5 ; xmm6=(10 11 12 13 14 15 16 17)=data1 293 movdqa xmm0, xmm7 ; transpose coefficients(phase 3) 294 punpcklqdq xmm7, xmm3 ; xmm7=(60 61 62 63 64 65 66 67)=data6 295 punpckhqdq xmm0, xmm3 ; xmm0=(70 71 72 73 74 75 76 77)=data7 296 297 movdqa xmm5, xmm6 298 movdqa xmm3, xmm1 299 psubw xmm6, xmm7 ; xmm6=data1-data6=tmp6 300 psubw xmm1, xmm0 ; xmm1=data0-data7=tmp7 301 paddw xmm5, xmm7 ; xmm5=data1+data6=tmp1 302 paddw xmm3, xmm0 ; xmm3=data0+data7=tmp0 303 304 movdqa xmm7, XMMWORD [wk(0)] ; xmm7=(24 25 26 27 34 35 36 37) 305 movdqa xmm0, XMMWORD [wk(1)] ; xmm0=(44 45 46 47 54 55 56 57) 306 movdqa XMMWORD [wk(0)], xmm6 ; wk(0)=tmp6 307 movdqa XMMWORD [wk(1)], xmm1 ; wk(1)=tmp7 308 309 movdqa xmm6, xmm2 ; transpose coefficients(phase 3) 310 punpcklqdq xmm2, xmm7 ; xmm2=(20 21 22 23 24 25 26 27)=data2 311 punpckhqdq xmm6, xmm7 ; xmm6=(30 31 32 33 34 35 36 37)=data3 312 movdqa xmm1, xmm4 ; transpose coefficients(phase 3) 313 punpcklqdq xmm4, xmm0 ; xmm4=(40 41 42 43 44 45 46 47)=data4 314 punpckhqdq xmm1, xmm0 ; xmm1=(50 51 52 53 54 55 56 57)=data5 315 316 movdqa xmm7, xmm6 317 movdqa xmm0, xmm2 318 paddw xmm6, xmm4 ; xmm6=data3+data4=tmp3 319 paddw xmm2, xmm1 ; xmm2=data2+data5=tmp2 320 psubw xmm7, xmm4 ; xmm7=data3-data4=tmp4 321 psubw xmm0, xmm1 ; xmm0=data2-data5=tmp5 322 323 ; -- Even part 324 325 movdqa xmm4, xmm3 326 movdqa xmm1, xmm5 327 psubw xmm3, xmm6 ; xmm3=tmp13 328 psubw xmm5, xmm2 ; xmm5=tmp12 329 paddw xmm4, xmm6 ; xmm4=tmp10 330 paddw xmm1, xmm2 ; xmm1=tmp11 331 332 paddw xmm5, xmm3 333 psllw xmm5, PRE_MULTIPLY_SCALE_BITS 334 pmulhw xmm5, [GOTOFF(ebx,PW_F0707)] ; xmm5=z1 335 336 movdqa xmm6, xmm4 337 movdqa xmm2, xmm3 338 psubw xmm4, xmm1 ; xmm4=data4 339 psubw xmm3, xmm5 ; xmm3=data6 340 paddw xmm6, xmm1 ; xmm6=data0 341 paddw xmm2, xmm5 ; xmm2=data2 342 343 movdqa XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_DCTELEM)], xmm4 344 movdqa XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_DCTELEM)], xmm3 345 movdqa XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_DCTELEM)], xmm6 346 movdqa XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_DCTELEM)], xmm2 347 348 ; -- Odd part 349 350 movdqa xmm1, XMMWORD [wk(0)] ; xmm1=tmp6 351 movdqa xmm5, XMMWORD [wk(1)] ; xmm5=tmp7 352 353 paddw xmm7, xmm0 ; xmm7=tmp10 354 paddw xmm0, xmm1 ; xmm0=tmp11 355 paddw xmm1, xmm5 ; xmm1=tmp12, xmm5=tmp7 356 357 psllw xmm7, PRE_MULTIPLY_SCALE_BITS 358 psllw xmm1, PRE_MULTIPLY_SCALE_BITS 359 360 psllw xmm0, PRE_MULTIPLY_SCALE_BITS 361 pmulhw xmm0, [GOTOFF(ebx,PW_F0707)] ; xmm0=z3 362 363 movdqa xmm4, xmm7 ; xmm4=tmp10 364 psubw xmm7, xmm1 365 pmulhw xmm7, [GOTOFF(ebx,PW_F0382)] ; xmm7=z5 366 pmulhw xmm4, [GOTOFF(ebx,PW_F0541)] ; xmm4=MULTIPLY(tmp10,FIX_0_541196) 367 pmulhw xmm1, [GOTOFF(ebx,PW_F1306)] ; xmm1=MULTIPLY(tmp12,FIX_1_306562) 368 paddw xmm4, xmm7 ; xmm4=z2 369 paddw xmm1, xmm7 ; xmm1=z4 370 371 movdqa xmm3, xmm5 372 psubw xmm5, xmm0 ; xmm5=z13 373 paddw xmm3, xmm0 ; xmm3=z11 374 375 movdqa xmm6, xmm5 376 movdqa xmm2, xmm3 377 psubw xmm5, xmm4 ; xmm5=data3 378 psubw xmm3, xmm1 ; xmm3=data7 379 paddw xmm6, xmm4 ; xmm6=data5 380 paddw xmm2, xmm1 ; xmm2=data1 381 382 movdqa XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_DCTELEM)], xmm5 383 movdqa XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_DCTELEM)], xmm3 384 movdqa XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_DCTELEM)], xmm6 385 movdqa XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_DCTELEM)], xmm2 386 387 ; pop edi ; unused 388 ; pop esi ; unused 389 ; pop edx ; need not be preserved 390 ; pop ecx ; unused 391 POPPIC ebx 392 mov esp, ebp ; esp <- aligned ebp 393 pop esp ; esp <- original ebp 394 pop ebp 395 ret 396 397 ; For some reason, the OS X linker does not honor the request to align the 398 ; segment unless we do this. 399 align 32