1; 2; jfdctfst.asm - fast integer FDCT (MMX) 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%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_mmx) 54 55EXTN(jconst_fdct_ifast_mmx): 56 57PW_F0707 times 4 dw F_0_707 << CONST_SHIFT 58PW_F0382 times 4 dw F_0_382 << CONST_SHIFT 59PW_F0541 times 4 dw F_0_541 << CONST_SHIFT 60PW_F1306 times 4 dw F_1_306 << CONST_SHIFT 61 62 alignz 32 63 64; -------------------------------------------------------------------------- 65 SECTION SEG_TEXT 66 BITS 32 67; 68; Perform the forward DCT on one block of samples. 69; 70; GLOBAL(void) 71; jsimd_fdct_ifast_mmx(DCTELEM *data) 72; 73 74%define data(b) (b) + 8 ; DCTELEM *data 75 76%define original_ebp ebp + 0 77%define wk(i) ebp - (WK_NUM - (i)) * SIZEOF_MMWORD ; mmword wk[WK_NUM] 78%define WK_NUM 2 79 80 align 32 81 GLOBAL_FUNCTION(jsimd_fdct_ifast_mmx) 82 83EXTN(jsimd_fdct_ifast_mmx): 84 push ebp 85 mov eax, esp ; eax = original ebp 86 sub esp, byte 4 87 and esp, byte (-SIZEOF_MMWORD) ; align to 64 bits 88 mov [esp], eax 89 mov ebp, esp ; ebp = aligned ebp 90 lea esp, [wk(0)] 91 pushpic ebx 92; push ecx ; need not be preserved 93; push edx ; need not be preserved 94; push esi ; unused 95; push edi ; unused 96 97 get_GOT ebx ; get GOT address 98 99 ; ---- Pass 1: process rows. 100 101 mov edx, POINTER [data(eax)] ; (DCTELEM *) 102 mov ecx, DCTSIZE/4 103 alignx 16, 7 104.rowloop: 105 106 movq mm0, MMWORD [MMBLOCK(2,0,edx,SIZEOF_DCTELEM)] 107 movq mm1, MMWORD [MMBLOCK(3,0,edx,SIZEOF_DCTELEM)] 108 movq mm2, MMWORD [MMBLOCK(2,1,edx,SIZEOF_DCTELEM)] 109 movq mm3, MMWORD [MMBLOCK(3,1,edx,SIZEOF_DCTELEM)] 110 111 ; mm0=(20 21 22 23), mm2=(24 25 26 27) 112 ; mm1=(30 31 32 33), mm3=(34 35 36 37) 113 114 movq mm4, mm0 ; transpose coefficients(phase 1) 115 punpcklwd mm0, mm1 ; mm0=(20 30 21 31) 116 punpckhwd mm4, mm1 ; mm4=(22 32 23 33) 117 movq mm5, mm2 ; transpose coefficients(phase 1) 118 punpcklwd mm2, mm3 ; mm2=(24 34 25 35) 119 punpckhwd mm5, mm3 ; mm5=(26 36 27 37) 120 121 movq mm6, MMWORD [MMBLOCK(0,0,edx,SIZEOF_DCTELEM)] 122 movq mm7, MMWORD [MMBLOCK(1,0,edx,SIZEOF_DCTELEM)] 123 movq mm1, MMWORD [MMBLOCK(0,1,edx,SIZEOF_DCTELEM)] 124 movq mm3, MMWORD [MMBLOCK(1,1,edx,SIZEOF_DCTELEM)] 125 126 ; mm6=(00 01 02 03), mm1=(04 05 06 07) 127 ; mm7=(10 11 12 13), mm3=(14 15 16 17) 128 129 movq MMWORD [wk(0)], mm4 ; wk(0)=(22 32 23 33) 130 movq MMWORD [wk(1)], mm2 ; wk(1)=(24 34 25 35) 131 132 movq mm4, mm6 ; transpose coefficients(phase 1) 133 punpcklwd mm6, mm7 ; mm6=(00 10 01 11) 134 punpckhwd mm4, mm7 ; mm4=(02 12 03 13) 135 movq mm2, mm1 ; transpose coefficients(phase 1) 136 punpcklwd mm1, mm3 ; mm1=(04 14 05 15) 137 punpckhwd mm2, mm3 ; mm2=(06 16 07 17) 138 139 movq mm7, mm6 ; transpose coefficients(phase 2) 140 punpckldq mm6, mm0 ; mm6=(00 10 20 30)=data0 141 punpckhdq mm7, mm0 ; mm7=(01 11 21 31)=data1 142 movq mm3, mm2 ; transpose coefficients(phase 2) 143 punpckldq mm2, mm5 ; mm2=(06 16 26 36)=data6 144 punpckhdq mm3, mm5 ; mm3=(07 17 27 37)=data7 145 146 movq mm0, mm7 147 movq mm5, mm6 148 psubw mm7, mm2 ; mm7=data1-data6=tmp6 149 psubw mm6, mm3 ; mm6=data0-data7=tmp7 150 paddw mm0, mm2 ; mm0=data1+data6=tmp1 151 paddw mm5, mm3 ; mm5=data0+data7=tmp0 152 153 movq mm2, MMWORD [wk(0)] ; mm2=(22 32 23 33) 154 movq mm3, MMWORD [wk(1)] ; mm3=(24 34 25 35) 155 movq MMWORD [wk(0)], mm7 ; wk(0)=tmp6 156 movq MMWORD [wk(1)], mm6 ; wk(1)=tmp7 157 158 movq mm7, mm4 ; transpose coefficients(phase 2) 159 punpckldq mm4, mm2 ; mm4=(02 12 22 32)=data2 160 punpckhdq mm7, mm2 ; mm7=(03 13 23 33)=data3 161 movq mm6, mm1 ; transpose coefficients(phase 2) 162 punpckldq mm1, mm3 ; mm1=(04 14 24 34)=data4 163 punpckhdq mm6, mm3 ; mm6=(05 15 25 35)=data5 164 165 movq mm2, mm7 166 movq mm3, mm4 167 paddw mm7, mm1 ; mm7=data3+data4=tmp3 168 paddw mm4, mm6 ; mm4=data2+data5=tmp2 169 psubw mm2, mm1 ; mm2=data3-data4=tmp4 170 psubw mm3, mm6 ; mm3=data2-data5=tmp5 171 172 ; -- Even part 173 174 movq mm1, mm5 175 movq mm6, mm0 176 psubw mm5, mm7 ; mm5=tmp13 177 psubw mm0, mm4 ; mm0=tmp12 178 paddw mm1, mm7 ; mm1=tmp10 179 paddw mm6, mm4 ; mm6=tmp11 180 181 paddw mm0, mm5 182 psllw mm0, PRE_MULTIPLY_SCALE_BITS 183 pmulhw mm0, [GOTOFF(ebx,PW_F0707)] ; mm0=z1 184 185 movq mm7, mm1 186 movq mm4, mm5 187 psubw mm1, mm6 ; mm1=data4 188 psubw mm5, mm0 ; mm5=data6 189 paddw mm7, mm6 ; mm7=data0 190 paddw mm4, mm0 ; mm4=data2 191 192 movq MMWORD [MMBLOCK(0,1,edx,SIZEOF_DCTELEM)], mm1 193 movq MMWORD [MMBLOCK(2,1,edx,SIZEOF_DCTELEM)], mm5 194 movq MMWORD [MMBLOCK(0,0,edx,SIZEOF_DCTELEM)], mm7 195 movq MMWORD [MMBLOCK(2,0,edx,SIZEOF_DCTELEM)], mm4 196 197 ; -- Odd part 198 199 movq mm6, MMWORD [wk(0)] ; mm6=tmp6 200 movq mm0, MMWORD [wk(1)] ; mm0=tmp7 201 202 paddw mm2, mm3 ; mm2=tmp10 203 paddw mm3, mm6 ; mm3=tmp11 204 paddw mm6, mm0 ; mm6=tmp12, mm0=tmp7 205 206 psllw mm2, PRE_MULTIPLY_SCALE_BITS 207 psllw mm6, PRE_MULTIPLY_SCALE_BITS 208 209 psllw mm3, PRE_MULTIPLY_SCALE_BITS 210 pmulhw mm3, [GOTOFF(ebx,PW_F0707)] ; mm3=z3 211 212 movq mm1, mm2 ; mm1=tmp10 213 psubw mm2, mm6 214 pmulhw mm2, [GOTOFF(ebx,PW_F0382)] ; mm2=z5 215 pmulhw mm1, [GOTOFF(ebx,PW_F0541)] ; mm1=MULTIPLY(tmp10,FIX_0_54119610) 216 pmulhw mm6, [GOTOFF(ebx,PW_F1306)] ; mm6=MULTIPLY(tmp12,FIX_1_30656296) 217 paddw mm1, mm2 ; mm1=z2 218 paddw mm6, mm2 ; mm6=z4 219 220 movq mm5, mm0 221 psubw mm0, mm3 ; mm0=z13 222 paddw mm5, mm3 ; mm5=z11 223 224 movq mm7, mm0 225 movq mm4, mm5 226 psubw mm0, mm1 ; mm0=data3 227 psubw mm5, mm6 ; mm5=data7 228 paddw mm7, mm1 ; mm7=data5 229 paddw mm4, mm6 ; mm4=data1 230 231 movq MMWORD [MMBLOCK(3,0,edx,SIZEOF_DCTELEM)], mm0 232 movq MMWORD [MMBLOCK(3,1,edx,SIZEOF_DCTELEM)], mm5 233 movq MMWORD [MMBLOCK(1,1,edx,SIZEOF_DCTELEM)], mm7 234 movq MMWORD [MMBLOCK(1,0,edx,SIZEOF_DCTELEM)], mm4 235 236 add edx, byte 4*DCTSIZE*SIZEOF_DCTELEM 237 dec ecx 238 jnz near .rowloop 239 240 ; ---- Pass 2: process columns. 241 242 mov edx, POINTER [data(eax)] ; (DCTELEM *) 243 mov ecx, DCTSIZE/4 244 alignx 16, 7 245.columnloop: 246 247 movq mm0, MMWORD [MMBLOCK(2,0,edx,SIZEOF_DCTELEM)] 248 movq mm1, MMWORD [MMBLOCK(3,0,edx,SIZEOF_DCTELEM)] 249 movq mm2, MMWORD [MMBLOCK(6,0,edx,SIZEOF_DCTELEM)] 250 movq mm3, MMWORD [MMBLOCK(7,0,edx,SIZEOF_DCTELEM)] 251 252 ; mm0=(02 12 22 32), mm2=(42 52 62 72) 253 ; mm1=(03 13 23 33), mm3=(43 53 63 73) 254 255 movq mm4, mm0 ; transpose coefficients(phase 1) 256 punpcklwd mm0, mm1 ; mm0=(02 03 12 13) 257 punpckhwd mm4, mm1 ; mm4=(22 23 32 33) 258 movq mm5, mm2 ; transpose coefficients(phase 1) 259 punpcklwd mm2, mm3 ; mm2=(42 43 52 53) 260 punpckhwd mm5, mm3 ; mm5=(62 63 72 73) 261 262 movq mm6, MMWORD [MMBLOCK(0,0,edx,SIZEOF_DCTELEM)] 263 movq mm7, MMWORD [MMBLOCK(1,0,edx,SIZEOF_DCTELEM)] 264 movq mm1, MMWORD [MMBLOCK(4,0,edx,SIZEOF_DCTELEM)] 265 movq mm3, MMWORD [MMBLOCK(5,0,edx,SIZEOF_DCTELEM)] 266 267 ; mm6=(00 10 20 30), mm1=(40 50 60 70) 268 ; mm7=(01 11 21 31), mm3=(41 51 61 71) 269 270 movq MMWORD [wk(0)], mm4 ; wk(0)=(22 23 32 33) 271 movq MMWORD [wk(1)], mm2 ; wk(1)=(42 43 52 53) 272 273 movq mm4, mm6 ; transpose coefficients(phase 1) 274 punpcklwd mm6, mm7 ; mm6=(00 01 10 11) 275 punpckhwd mm4, mm7 ; mm4=(20 21 30 31) 276 movq mm2, mm1 ; transpose coefficients(phase 1) 277 punpcklwd mm1, mm3 ; mm1=(40 41 50 51) 278 punpckhwd mm2, mm3 ; mm2=(60 61 70 71) 279 280 movq mm7, mm6 ; transpose coefficients(phase 2) 281 punpckldq mm6, mm0 ; mm6=(00 01 02 03)=data0 282 punpckhdq mm7, mm0 ; mm7=(10 11 12 13)=data1 283 movq mm3, mm2 ; transpose coefficients(phase 2) 284 punpckldq mm2, mm5 ; mm2=(60 61 62 63)=data6 285 punpckhdq mm3, mm5 ; mm3=(70 71 72 73)=data7 286 287 movq mm0, mm7 288 movq mm5, mm6 289 psubw mm7, mm2 ; mm7=data1-data6=tmp6 290 psubw mm6, mm3 ; mm6=data0-data7=tmp7 291 paddw mm0, mm2 ; mm0=data1+data6=tmp1 292 paddw mm5, mm3 ; mm5=data0+data7=tmp0 293 294 movq mm2, MMWORD [wk(0)] ; mm2=(22 23 32 33) 295 movq mm3, MMWORD [wk(1)] ; mm3=(42 43 52 53) 296 movq MMWORD [wk(0)], mm7 ; wk(0)=tmp6 297 movq MMWORD [wk(1)], mm6 ; wk(1)=tmp7 298 299 movq mm7, mm4 ; transpose coefficients(phase 2) 300 punpckldq mm4, mm2 ; mm4=(20 21 22 23)=data2 301 punpckhdq mm7, mm2 ; mm7=(30 31 32 33)=data3 302 movq mm6, mm1 ; transpose coefficients(phase 2) 303 punpckldq mm1, mm3 ; mm1=(40 41 42 43)=data4 304 punpckhdq mm6, mm3 ; mm6=(50 51 52 53)=data5 305 306 movq mm2, mm7 307 movq mm3, mm4 308 paddw mm7, mm1 ; mm7=data3+data4=tmp3 309 paddw mm4, mm6 ; mm4=data2+data5=tmp2 310 psubw mm2, mm1 ; mm2=data3-data4=tmp4 311 psubw mm3, mm6 ; mm3=data2-data5=tmp5 312 313 ; -- Even part 314 315 movq mm1, mm5 316 movq mm6, mm0 317 psubw mm5, mm7 ; mm5=tmp13 318 psubw mm0, mm4 ; mm0=tmp12 319 paddw mm1, mm7 ; mm1=tmp10 320 paddw mm6, mm4 ; mm6=tmp11 321 322 paddw mm0, mm5 323 psllw mm0, PRE_MULTIPLY_SCALE_BITS 324 pmulhw mm0, [GOTOFF(ebx,PW_F0707)] ; mm0=z1 325 326 movq mm7, mm1 327 movq mm4, mm5 328 psubw mm1, mm6 ; mm1=data4 329 psubw mm5, mm0 ; mm5=data6 330 paddw mm7, mm6 ; mm7=data0 331 paddw mm4, mm0 ; mm4=data2 332 333 movq MMWORD [MMBLOCK(4,0,edx,SIZEOF_DCTELEM)], mm1 334 movq MMWORD [MMBLOCK(6,0,edx,SIZEOF_DCTELEM)], mm5 335 movq MMWORD [MMBLOCK(0,0,edx,SIZEOF_DCTELEM)], mm7 336 movq MMWORD [MMBLOCK(2,0,edx,SIZEOF_DCTELEM)], mm4 337 338 ; -- Odd part 339 340 movq mm6, MMWORD [wk(0)] ; mm6=tmp6 341 movq mm0, MMWORD [wk(1)] ; mm0=tmp7 342 343 paddw mm2, mm3 ; mm2=tmp10 344 paddw mm3, mm6 ; mm3=tmp11 345 paddw mm6, mm0 ; mm6=tmp12, mm0=tmp7 346 347 psllw mm2, PRE_MULTIPLY_SCALE_BITS 348 psllw mm6, PRE_MULTIPLY_SCALE_BITS 349 350 psllw mm3, PRE_MULTIPLY_SCALE_BITS 351 pmulhw mm3, [GOTOFF(ebx,PW_F0707)] ; mm3=z3 352 353 movq mm1, mm2 ; mm1=tmp10 354 psubw mm2, mm6 355 pmulhw mm2, [GOTOFF(ebx,PW_F0382)] ; mm2=z5 356 pmulhw mm1, [GOTOFF(ebx,PW_F0541)] ; mm1=MULTIPLY(tmp10,FIX_0_54119610) 357 pmulhw mm6, [GOTOFF(ebx,PW_F1306)] ; mm6=MULTIPLY(tmp12,FIX_1_30656296) 358 paddw mm1, mm2 ; mm1=z2 359 paddw mm6, mm2 ; mm6=z4 360 361 movq mm5, mm0 362 psubw mm0, mm3 ; mm0=z13 363 paddw mm5, mm3 ; mm5=z11 364 365 movq mm7, mm0 366 movq mm4, mm5 367 psubw mm0, mm1 ; mm0=data3 368 psubw mm5, mm6 ; mm5=data7 369 paddw mm7, mm1 ; mm7=data5 370 paddw mm4, mm6 ; mm4=data1 371 372 movq MMWORD [MMBLOCK(3,0,edx,SIZEOF_DCTELEM)], mm0 373 movq MMWORD [MMBLOCK(7,0,edx,SIZEOF_DCTELEM)], mm5 374 movq MMWORD [MMBLOCK(5,0,edx,SIZEOF_DCTELEM)], mm7 375 movq MMWORD [MMBLOCK(1,0,edx,SIZEOF_DCTELEM)], mm4 376 377 add edx, byte 4*SIZEOF_DCTELEM 378 dec ecx 379 jnz near .columnloop 380 381 emms ; empty MMX state 382 383; pop edi ; unused 384; pop esi ; unused 385; pop edx ; need not be preserved 386; pop ecx ; need not be preserved 387 poppic ebx 388 mov esp, ebp ; esp <- aligned ebp 389 pop esp ; esp <- original ebp 390 pop ebp 391 ret 392 393; For some reason, the OS X linker does not honor the request to align the 394; segment unless we do this. 395 align 32 396