1; 2; jfdctflt.asm - floating-point FDCT (SSE) 3; 4; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB 5; 6; Based on the x86 SIMD extension for IJG JPEG library 7; Copyright (C) 1999-2006, MIYASAKA Masaru. 8; For conditions of distribution and use, see copyright notice in jsimdext.inc 9; 10; This file should be assembled with NASM (Netwide Assembler), 11; can *not* be assembled with Microsoft's MASM or any compatible 12; assembler (including Borland's Turbo Assembler). 13; NASM is available from http://nasm.sourceforge.net/ or 14; http://sourceforge.net/project/showfiles.php?group_id=6208 15; 16; This file contains a floating-point implementation of the forward DCT 17; (Discrete Cosine Transform). The following code is based directly on 18; the IJG's original jfdctflt.c; see the jfdctflt.c for more details. 19; 20; [TAB8] 21 22%include "jsimdext.inc" 23%include "jdct.inc" 24 25; -------------------------------------------------------------------------- 26 27%macro unpcklps2 2 ; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(0 1 4 5) 28 shufps %1,%2,0x44 29%endmacro 30 31%macro unpckhps2 2 ; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(2 3 6 7) 32 shufps %1,%2,0xEE 33%endmacro 34 35; -------------------------------------------------------------------------- 36 SECTION SEG_CONST 37 38 alignz 16 39 global EXTN(jconst_fdct_float_sse) 40 41EXTN(jconst_fdct_float_sse): 42 43PD_0_382 times 4 dd 0.382683432365089771728460 44PD_0_707 times 4 dd 0.707106781186547524400844 45PD_0_541 times 4 dd 0.541196100146196984399723 46PD_1_306 times 4 dd 1.306562964876376527856643 47 48 alignz 16 49 50; -------------------------------------------------------------------------- 51 SECTION SEG_TEXT 52 BITS 32 53; 54; Perform the forward DCT on one block of samples. 55; 56; GLOBAL(void) 57; jsimd_fdct_float_sse (FAST_FLOAT *data) 58; 59 60%define data(b) (b)+8 ; FAST_FLOAT *data 61 62%define original_ebp ebp+0 63%define wk(i) ebp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM] 64%define WK_NUM 2 65 66 align 16 67 global EXTN(jsimd_fdct_float_sse) 68 69EXTN(jsimd_fdct_float_sse): 70 push ebp 71 mov eax,esp ; eax = original ebp 72 sub esp, byte 4 73 and esp, byte (-SIZEOF_XMMWORD) ; align to 128 bits 74 mov [esp],eax 75 mov ebp,esp ; ebp = aligned ebp 76 lea esp, [wk(0)] 77 pushpic ebx 78; push ecx ; need not be preserved 79; push edx ; need not be preserved 80; push esi ; unused 81; push edi ; unused 82 83 get_GOT ebx ; get GOT address 84 85 ; ---- Pass 1: process rows. 86 87 mov edx, POINTER [data(eax)] ; (FAST_FLOAT *) 88 mov ecx, DCTSIZE/4 89 alignx 16,7 90.rowloop: 91 92 movaps xmm0, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_FAST_FLOAT)] 93 movaps xmm1, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_FAST_FLOAT)] 94 movaps xmm2, XMMWORD [XMMBLOCK(2,1,edx,SIZEOF_FAST_FLOAT)] 95 movaps xmm3, XMMWORD [XMMBLOCK(3,1,edx,SIZEOF_FAST_FLOAT)] 96 97 ; xmm0=(20 21 22 23), xmm2=(24 25 26 27) 98 ; xmm1=(30 31 32 33), xmm3=(34 35 36 37) 99 100 movaps xmm4,xmm0 ; transpose coefficients(phase 1) 101 unpcklps xmm0,xmm1 ; xmm0=(20 30 21 31) 102 unpckhps xmm4,xmm1 ; xmm4=(22 32 23 33) 103 movaps xmm5,xmm2 ; transpose coefficients(phase 1) 104 unpcklps xmm2,xmm3 ; xmm2=(24 34 25 35) 105 unpckhps xmm5,xmm3 ; xmm5=(26 36 27 37) 106 107 movaps xmm6, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)] 108 movaps xmm7, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)] 109 movaps xmm1, XMMWORD [XMMBLOCK(0,1,edx,SIZEOF_FAST_FLOAT)] 110 movaps xmm3, XMMWORD [XMMBLOCK(1,1,edx,SIZEOF_FAST_FLOAT)] 111 112 ; xmm6=(00 01 02 03), xmm1=(04 05 06 07) 113 ; xmm7=(10 11 12 13), xmm3=(14 15 16 17) 114 115 movaps XMMWORD [wk(0)], xmm4 ; wk(0)=(22 32 23 33) 116 movaps XMMWORD [wk(1)], xmm2 ; wk(1)=(24 34 25 35) 117 118 movaps xmm4,xmm6 ; transpose coefficients(phase 1) 119 unpcklps xmm6,xmm7 ; xmm6=(00 10 01 11) 120 unpckhps xmm4,xmm7 ; xmm4=(02 12 03 13) 121 movaps xmm2,xmm1 ; transpose coefficients(phase 1) 122 unpcklps xmm1,xmm3 ; xmm1=(04 14 05 15) 123 unpckhps xmm2,xmm3 ; xmm2=(06 16 07 17) 124 125 movaps xmm7,xmm6 ; transpose coefficients(phase 2) 126 unpcklps2 xmm6,xmm0 ; xmm6=(00 10 20 30)=data0 127 unpckhps2 xmm7,xmm0 ; xmm7=(01 11 21 31)=data1 128 movaps xmm3,xmm2 ; transpose coefficients(phase 2) 129 unpcklps2 xmm2,xmm5 ; xmm2=(06 16 26 36)=data6 130 unpckhps2 xmm3,xmm5 ; xmm3=(07 17 27 37)=data7 131 132 movaps xmm0,xmm7 133 movaps xmm5,xmm6 134 subps xmm7,xmm2 ; xmm7=data1-data6=tmp6 135 subps xmm6,xmm3 ; xmm6=data0-data7=tmp7 136 addps xmm0,xmm2 ; xmm0=data1+data6=tmp1 137 addps xmm5,xmm3 ; xmm5=data0+data7=tmp0 138 139 movaps xmm2, XMMWORD [wk(0)] ; xmm2=(22 32 23 33) 140 movaps xmm3, XMMWORD [wk(1)] ; xmm3=(24 34 25 35) 141 movaps XMMWORD [wk(0)], xmm7 ; wk(0)=tmp6 142 movaps XMMWORD [wk(1)], xmm6 ; wk(1)=tmp7 143 144 movaps xmm7,xmm4 ; transpose coefficients(phase 2) 145 unpcklps2 xmm4,xmm2 ; xmm4=(02 12 22 32)=data2 146 unpckhps2 xmm7,xmm2 ; xmm7=(03 13 23 33)=data3 147 movaps xmm6,xmm1 ; transpose coefficients(phase 2) 148 unpcklps2 xmm1,xmm3 ; xmm1=(04 14 24 34)=data4 149 unpckhps2 xmm6,xmm3 ; xmm6=(05 15 25 35)=data5 150 151 movaps xmm2,xmm7 152 movaps xmm3,xmm4 153 addps xmm7,xmm1 ; xmm7=data3+data4=tmp3 154 addps xmm4,xmm6 ; xmm4=data2+data5=tmp2 155 subps xmm2,xmm1 ; xmm2=data3-data4=tmp4 156 subps xmm3,xmm6 ; xmm3=data2-data5=tmp5 157 158 ; -- Even part 159 160 movaps xmm1,xmm5 161 movaps xmm6,xmm0 162 subps xmm5,xmm7 ; xmm5=tmp13 163 subps xmm0,xmm4 ; xmm0=tmp12 164 addps xmm1,xmm7 ; xmm1=tmp10 165 addps xmm6,xmm4 ; xmm6=tmp11 166 167 addps xmm0,xmm5 168 mulps xmm0,[GOTOFF(ebx,PD_0_707)] ; xmm0=z1 169 170 movaps xmm7,xmm1 171 movaps xmm4,xmm5 172 subps xmm1,xmm6 ; xmm1=data4 173 subps xmm5,xmm0 ; xmm5=data6 174 addps xmm7,xmm6 ; xmm7=data0 175 addps xmm4,xmm0 ; xmm4=data2 176 177 movaps XMMWORD [XMMBLOCK(0,1,edx,SIZEOF_FAST_FLOAT)], xmm1 178 movaps XMMWORD [XMMBLOCK(2,1,edx,SIZEOF_FAST_FLOAT)], xmm5 179 movaps XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)], xmm7 180 movaps XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_FAST_FLOAT)], xmm4 181 182 ; -- Odd part 183 184 movaps xmm6, XMMWORD [wk(0)] ; xmm6=tmp6 185 movaps xmm0, XMMWORD [wk(1)] ; xmm0=tmp7 186 187 addps xmm2,xmm3 ; xmm2=tmp10 188 addps xmm3,xmm6 ; xmm3=tmp11 189 addps xmm6,xmm0 ; xmm6=tmp12, xmm0=tmp7 190 191 mulps xmm3,[GOTOFF(ebx,PD_0_707)] ; xmm3=z3 192 193 movaps xmm1,xmm2 ; xmm1=tmp10 194 subps xmm2,xmm6 195 mulps xmm2,[GOTOFF(ebx,PD_0_382)] ; xmm2=z5 196 mulps xmm1,[GOTOFF(ebx,PD_0_541)] ; xmm1=MULTIPLY(tmp10,FIX_0_541196) 197 mulps xmm6,[GOTOFF(ebx,PD_1_306)] ; xmm6=MULTIPLY(tmp12,FIX_1_306562) 198 addps xmm1,xmm2 ; xmm1=z2 199 addps xmm6,xmm2 ; xmm6=z4 200 201 movaps xmm5,xmm0 202 subps xmm0,xmm3 ; xmm0=z13 203 addps xmm5,xmm3 ; xmm5=z11 204 205 movaps xmm7,xmm0 206 movaps xmm4,xmm5 207 subps xmm0,xmm1 ; xmm0=data3 208 subps xmm5,xmm6 ; xmm5=data7 209 addps xmm7,xmm1 ; xmm7=data5 210 addps xmm4,xmm6 ; xmm4=data1 211 212 movaps XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_FAST_FLOAT)], xmm0 213 movaps XMMWORD [XMMBLOCK(3,1,edx,SIZEOF_FAST_FLOAT)], xmm5 214 movaps XMMWORD [XMMBLOCK(1,1,edx,SIZEOF_FAST_FLOAT)], xmm7 215 movaps XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)], xmm4 216 217 add edx, 4*DCTSIZE*SIZEOF_FAST_FLOAT 218 dec ecx 219 jnz near .rowloop 220 221 ; ---- Pass 2: process columns. 222 223 mov edx, POINTER [data(eax)] ; (FAST_FLOAT *) 224 mov ecx, DCTSIZE/4 225 alignx 16,7 226.columnloop: 227 228 movaps xmm0, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_FAST_FLOAT)] 229 movaps xmm1, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_FAST_FLOAT)] 230 movaps xmm2, XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_FAST_FLOAT)] 231 movaps xmm3, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_FAST_FLOAT)] 232 233 ; xmm0=(02 12 22 32), xmm2=(42 52 62 72) 234 ; xmm1=(03 13 23 33), xmm3=(43 53 63 73) 235 236 movaps xmm4,xmm0 ; transpose coefficients(phase 1) 237 unpcklps xmm0,xmm1 ; xmm0=(02 03 12 13) 238 unpckhps xmm4,xmm1 ; xmm4=(22 23 32 33) 239 movaps xmm5,xmm2 ; transpose coefficients(phase 1) 240 unpcklps xmm2,xmm3 ; xmm2=(42 43 52 53) 241 unpckhps xmm5,xmm3 ; xmm5=(62 63 72 73) 242 243 movaps xmm6, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)] 244 movaps xmm7, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)] 245 movaps xmm1, XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_FAST_FLOAT)] 246 movaps xmm3, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_FAST_FLOAT)] 247 248 ; xmm6=(00 10 20 30), xmm1=(40 50 60 70) 249 ; xmm7=(01 11 21 31), xmm3=(41 51 61 71) 250 251 movaps XMMWORD [wk(0)], xmm4 ; wk(0)=(22 23 32 33) 252 movaps XMMWORD [wk(1)], xmm2 ; wk(1)=(42 43 52 53) 253 254 movaps xmm4,xmm6 ; transpose coefficients(phase 1) 255 unpcklps xmm6,xmm7 ; xmm6=(00 01 10 11) 256 unpckhps xmm4,xmm7 ; xmm4=(20 21 30 31) 257 movaps xmm2,xmm1 ; transpose coefficients(phase 1) 258 unpcklps xmm1,xmm3 ; xmm1=(40 41 50 51) 259 unpckhps xmm2,xmm3 ; xmm2=(60 61 70 71) 260 261 movaps xmm7,xmm6 ; transpose coefficients(phase 2) 262 unpcklps2 xmm6,xmm0 ; xmm6=(00 01 02 03)=data0 263 unpckhps2 xmm7,xmm0 ; xmm7=(10 11 12 13)=data1 264 movaps xmm3,xmm2 ; transpose coefficients(phase 2) 265 unpcklps2 xmm2,xmm5 ; xmm2=(60 61 62 63)=data6 266 unpckhps2 xmm3,xmm5 ; xmm3=(70 71 72 73)=data7 267 268 movaps xmm0,xmm7 269 movaps xmm5,xmm6 270 subps xmm7,xmm2 ; xmm7=data1-data6=tmp6 271 subps xmm6,xmm3 ; xmm6=data0-data7=tmp7 272 addps xmm0,xmm2 ; xmm0=data1+data6=tmp1 273 addps xmm5,xmm3 ; xmm5=data0+data7=tmp0 274 275 movaps xmm2, XMMWORD [wk(0)] ; xmm2=(22 23 32 33) 276 movaps xmm3, XMMWORD [wk(1)] ; xmm3=(42 43 52 53) 277 movaps XMMWORD [wk(0)], xmm7 ; wk(0)=tmp6 278 movaps XMMWORD [wk(1)], xmm6 ; wk(1)=tmp7 279 280 movaps xmm7,xmm4 ; transpose coefficients(phase 2) 281 unpcklps2 xmm4,xmm2 ; xmm4=(20 21 22 23)=data2 282 unpckhps2 xmm7,xmm2 ; xmm7=(30 31 32 33)=data3 283 movaps xmm6,xmm1 ; transpose coefficients(phase 2) 284 unpcklps2 xmm1,xmm3 ; xmm1=(40 41 42 43)=data4 285 unpckhps2 xmm6,xmm3 ; xmm6=(50 51 52 53)=data5 286 287 movaps xmm2,xmm7 288 movaps xmm3,xmm4 289 addps xmm7,xmm1 ; xmm7=data3+data4=tmp3 290 addps xmm4,xmm6 ; xmm4=data2+data5=tmp2 291 subps xmm2,xmm1 ; xmm2=data3-data4=tmp4 292 subps xmm3,xmm6 ; xmm3=data2-data5=tmp5 293 294 ; -- Even part 295 296 movaps xmm1,xmm5 297 movaps xmm6,xmm0 298 subps xmm5,xmm7 ; xmm5=tmp13 299 subps xmm0,xmm4 ; xmm0=tmp12 300 addps xmm1,xmm7 ; xmm1=tmp10 301 addps xmm6,xmm4 ; xmm6=tmp11 302 303 addps xmm0,xmm5 304 mulps xmm0,[GOTOFF(ebx,PD_0_707)] ; xmm0=z1 305 306 movaps xmm7,xmm1 307 movaps xmm4,xmm5 308 subps xmm1,xmm6 ; xmm1=data4 309 subps xmm5,xmm0 ; xmm5=data6 310 addps xmm7,xmm6 ; xmm7=data0 311 addps xmm4,xmm0 ; xmm4=data2 312 313 movaps XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_FAST_FLOAT)], xmm1 314 movaps XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_FAST_FLOAT)], xmm5 315 movaps XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)], xmm7 316 movaps XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_FAST_FLOAT)], xmm4 317 318 ; -- Odd part 319 320 movaps xmm6, XMMWORD [wk(0)] ; xmm6=tmp6 321 movaps xmm0, XMMWORD [wk(1)] ; xmm0=tmp7 322 323 addps xmm2,xmm3 ; xmm2=tmp10 324 addps xmm3,xmm6 ; xmm3=tmp11 325 addps xmm6,xmm0 ; xmm6=tmp12, xmm0=tmp7 326 327 mulps xmm3,[GOTOFF(ebx,PD_0_707)] ; xmm3=z3 328 329 movaps xmm1,xmm2 ; xmm1=tmp10 330 subps xmm2,xmm6 331 mulps xmm2,[GOTOFF(ebx,PD_0_382)] ; xmm2=z5 332 mulps xmm1,[GOTOFF(ebx,PD_0_541)] ; xmm1=MULTIPLY(tmp10,FIX_0_541196) 333 mulps xmm6,[GOTOFF(ebx,PD_1_306)] ; xmm6=MULTIPLY(tmp12,FIX_1_306562) 334 addps xmm1,xmm2 ; xmm1=z2 335 addps xmm6,xmm2 ; xmm6=z4 336 337 movaps xmm5,xmm0 338 subps xmm0,xmm3 ; xmm0=z13 339 addps xmm5,xmm3 ; xmm5=z11 340 341 movaps xmm7,xmm0 342 movaps xmm4,xmm5 343 subps xmm0,xmm1 ; xmm0=data3 344 subps xmm5,xmm6 ; xmm5=data7 345 addps xmm7,xmm1 ; xmm7=data5 346 addps xmm4,xmm6 ; xmm4=data1 347 348 movaps XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_FAST_FLOAT)], xmm0 349 movaps XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_FAST_FLOAT)], xmm5 350 movaps XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_FAST_FLOAT)], xmm7 351 movaps XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)], xmm4 352 353 add edx, byte 4*SIZEOF_FAST_FLOAT 354 dec ecx 355 jnz near .columnloop 356 357; pop edi ; unused 358; pop esi ; unused 359; pop edx ; need not be preserved 360; pop ecx ; need not be preserved 361 poppic ebx 362 mov esp,ebp ; esp <- aligned ebp 363 pop esp ; esp <- original ebp 364 pop ebp 365 ret 366 367; For some reason, the OS X linker does not honor the request to align the 368; segment unless we do this. 369 align 16 370