1; 2; jfdctflt.asm - floating-point FDCT (64-bit SSE) 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 floating-point implementation of the forward DCT 18; (Discrete Cosine Transform). The following code is based directly on 19; the IJG's original jfdctflt.c; see the jfdctflt.c for more details. 20 21%include "jsimdext.inc" 22%include "jdct.inc" 23 24; -------------------------------------------------------------------------- 25 26%macro unpcklps2 2 ; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(0 1 4 5) 27 shufps %1, %2, 0x44 28%endmacro 29 30%macro unpckhps2 2 ; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(2 3 6 7) 31 shufps %1, %2, 0xEE 32%endmacro 33 34; -------------------------------------------------------------------------- 35 SECTION SEG_CONST 36 37 alignz 32 38 GLOBAL_DATA(jconst_fdct_float_sse) 39 40EXTN(jconst_fdct_float_sse): 41 42PD_0_382 times 4 dd 0.382683432365089771728460 43PD_0_707 times 4 dd 0.707106781186547524400844 44PD_0_541 times 4 dd 0.541196100146196984399723 45PD_1_306 times 4 dd 1.306562964876376527856643 46 47 alignz 32 48 49; -------------------------------------------------------------------------- 50 SECTION SEG_TEXT 51 BITS 64 52; 53; Perform the forward DCT on one block of samples. 54; 55; GLOBAL(void) 56; jsimd_fdct_float_sse(FAST_FLOAT *data) 57; 58 59; r10 = FAST_FLOAT *data 60 61%define wk(i) rbp - (WK_NUM - (i)) * SIZEOF_XMMWORD ; xmmword wk[WK_NUM] 62%define WK_NUM 2 63 64 align 32 65 GLOBAL_FUNCTION(jsimd_fdct_float_sse) 66 67EXTN(jsimd_fdct_float_sse): 68 push rbp 69 mov rax, rsp ; rax = original rbp 70 sub rsp, byte 4 71 and rsp, byte (-SIZEOF_XMMWORD) ; align to 128 bits 72 mov [rsp], rax 73 mov rbp, rsp ; rbp = aligned rbp 74 lea rsp, [wk(0)] 75 collect_args 1 76 77 ; ---- Pass 1: process rows. 78 79 mov rdx, r10 ; (FAST_FLOAT *) 80 mov rcx, DCTSIZE/4 81.rowloop: 82 83 movaps xmm0, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_FAST_FLOAT)] 84 movaps xmm1, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_FAST_FLOAT)] 85 movaps xmm2, XMMWORD [XMMBLOCK(2,1,rdx,SIZEOF_FAST_FLOAT)] 86 movaps xmm3, XMMWORD [XMMBLOCK(3,1,rdx,SIZEOF_FAST_FLOAT)] 87 88 ; xmm0=(20 21 22 23), xmm2=(24 25 26 27) 89 ; xmm1=(30 31 32 33), xmm3=(34 35 36 37) 90 91 movaps xmm4, xmm0 ; transpose coefficients(phase 1) 92 unpcklps xmm0, xmm1 ; xmm0=(20 30 21 31) 93 unpckhps xmm4, xmm1 ; xmm4=(22 32 23 33) 94 movaps xmm5, xmm2 ; transpose coefficients(phase 1) 95 unpcklps xmm2, xmm3 ; xmm2=(24 34 25 35) 96 unpckhps xmm5, xmm3 ; xmm5=(26 36 27 37) 97 98 movaps xmm6, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FAST_FLOAT)] 99 movaps xmm7, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FAST_FLOAT)] 100 movaps xmm1, XMMWORD [XMMBLOCK(0,1,rdx,SIZEOF_FAST_FLOAT)] 101 movaps xmm3, XMMWORD [XMMBLOCK(1,1,rdx,SIZEOF_FAST_FLOAT)] 102 103 ; xmm6=(00 01 02 03), xmm1=(04 05 06 07) 104 ; xmm7=(10 11 12 13), xmm3=(14 15 16 17) 105 106 movaps XMMWORD [wk(0)], xmm4 ; wk(0)=(22 32 23 33) 107 movaps XMMWORD [wk(1)], xmm2 ; wk(1)=(24 34 25 35) 108 109 movaps xmm4, xmm6 ; transpose coefficients(phase 1) 110 unpcklps xmm6, xmm7 ; xmm6=(00 10 01 11) 111 unpckhps xmm4, xmm7 ; xmm4=(02 12 03 13) 112 movaps xmm2, xmm1 ; transpose coefficients(phase 1) 113 unpcklps xmm1, xmm3 ; xmm1=(04 14 05 15) 114 unpckhps xmm2, xmm3 ; xmm2=(06 16 07 17) 115 116 movaps xmm7, xmm6 ; transpose coefficients(phase 2) 117 unpcklps2 xmm6, xmm0 ; xmm6=(00 10 20 30)=data0 118 unpckhps2 xmm7, xmm0 ; xmm7=(01 11 21 31)=data1 119 movaps xmm3, xmm2 ; transpose coefficients(phase 2) 120 unpcklps2 xmm2, xmm5 ; xmm2=(06 16 26 36)=data6 121 unpckhps2 xmm3, xmm5 ; xmm3=(07 17 27 37)=data7 122 123 movaps xmm0, xmm7 124 movaps xmm5, xmm6 125 subps xmm7, xmm2 ; xmm7=data1-data6=tmp6 126 subps xmm6, xmm3 ; xmm6=data0-data7=tmp7 127 addps xmm0, xmm2 ; xmm0=data1+data6=tmp1 128 addps xmm5, xmm3 ; xmm5=data0+data7=tmp0 129 130 movaps xmm2, XMMWORD [wk(0)] ; xmm2=(22 32 23 33) 131 movaps xmm3, XMMWORD [wk(1)] ; xmm3=(24 34 25 35) 132 movaps XMMWORD [wk(0)], xmm7 ; wk(0)=tmp6 133 movaps XMMWORD [wk(1)], xmm6 ; wk(1)=tmp7 134 135 movaps xmm7, xmm4 ; transpose coefficients(phase 2) 136 unpcklps2 xmm4, xmm2 ; xmm4=(02 12 22 32)=data2 137 unpckhps2 xmm7, xmm2 ; xmm7=(03 13 23 33)=data3 138 movaps xmm6, xmm1 ; transpose coefficients(phase 2) 139 unpcklps2 xmm1, xmm3 ; xmm1=(04 14 24 34)=data4 140 unpckhps2 xmm6, xmm3 ; xmm6=(05 15 25 35)=data5 141 142 movaps xmm2, xmm7 143 movaps xmm3, xmm4 144 addps xmm7, xmm1 ; xmm7=data3+data4=tmp3 145 addps xmm4, xmm6 ; xmm4=data2+data5=tmp2 146 subps xmm2, xmm1 ; xmm2=data3-data4=tmp4 147 subps xmm3, xmm6 ; xmm3=data2-data5=tmp5 148 149 ; -- Even part 150 151 movaps xmm1, xmm5 152 movaps xmm6, xmm0 153 subps xmm5, xmm7 ; xmm5=tmp13 154 subps xmm0, xmm4 ; xmm0=tmp12 155 addps xmm1, xmm7 ; xmm1=tmp10 156 addps xmm6, xmm4 ; xmm6=tmp11 157 158 addps xmm0, xmm5 159 mulps xmm0, [rel PD_0_707] ; xmm0=z1 160 161 movaps xmm7, xmm1 162 movaps xmm4, xmm5 163 subps xmm1, xmm6 ; xmm1=data4 164 subps xmm5, xmm0 ; xmm5=data6 165 addps xmm7, xmm6 ; xmm7=data0 166 addps xmm4, xmm0 ; xmm4=data2 167 168 movaps XMMWORD [XMMBLOCK(0,1,rdx,SIZEOF_FAST_FLOAT)], xmm1 169 movaps XMMWORD [XMMBLOCK(2,1,rdx,SIZEOF_FAST_FLOAT)], xmm5 170 movaps XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FAST_FLOAT)], xmm7 171 movaps XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_FAST_FLOAT)], xmm4 172 173 ; -- Odd part 174 175 movaps xmm6, XMMWORD [wk(0)] ; xmm6=tmp6 176 movaps xmm0, XMMWORD [wk(1)] ; xmm0=tmp7 177 178 addps xmm2, xmm3 ; xmm2=tmp10 179 addps xmm3, xmm6 ; xmm3=tmp11 180 addps xmm6, xmm0 ; xmm6=tmp12, xmm0=tmp7 181 182 mulps xmm3, [rel PD_0_707] ; xmm3=z3 183 184 movaps xmm1, xmm2 ; xmm1=tmp10 185 subps xmm2, xmm6 186 mulps xmm2, [rel PD_0_382] ; xmm2=z5 187 mulps xmm1, [rel PD_0_541] ; xmm1=MULTIPLY(tmp10,FIX_0_541196) 188 mulps xmm6, [rel PD_1_306] ; xmm6=MULTIPLY(tmp12,FIX_1_306562) 189 addps xmm1, xmm2 ; xmm1=z2 190 addps xmm6, xmm2 ; xmm6=z4 191 192 movaps xmm5, xmm0 193 subps xmm0, xmm3 ; xmm0=z13 194 addps xmm5, xmm3 ; xmm5=z11 195 196 movaps xmm7, xmm0 197 movaps xmm4, xmm5 198 subps xmm0, xmm1 ; xmm0=data3 199 subps xmm5, xmm6 ; xmm5=data7 200 addps xmm7, xmm1 ; xmm7=data5 201 addps xmm4, xmm6 ; xmm4=data1 202 203 movaps XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_FAST_FLOAT)], xmm0 204 movaps XMMWORD [XMMBLOCK(3,1,rdx,SIZEOF_FAST_FLOAT)], xmm5 205 movaps XMMWORD [XMMBLOCK(1,1,rdx,SIZEOF_FAST_FLOAT)], xmm7 206 movaps XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FAST_FLOAT)], xmm4 207 208 add rdx, 4*DCTSIZE*SIZEOF_FAST_FLOAT 209 dec rcx 210 jnz near .rowloop 211 212 ; ---- Pass 2: process columns. 213 214 mov rdx, r10 ; (FAST_FLOAT *) 215 mov rcx, DCTSIZE/4 216.columnloop: 217 218 movaps xmm0, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_FAST_FLOAT)] 219 movaps xmm1, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_FAST_FLOAT)] 220 movaps xmm2, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_FAST_FLOAT)] 221 movaps xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_FAST_FLOAT)] 222 223 ; xmm0=(02 12 22 32), xmm2=(42 52 62 72) 224 ; xmm1=(03 13 23 33), xmm3=(43 53 63 73) 225 226 movaps xmm4, xmm0 ; transpose coefficients(phase 1) 227 unpcklps xmm0, xmm1 ; xmm0=(02 03 12 13) 228 unpckhps xmm4, xmm1 ; xmm4=(22 23 32 33) 229 movaps xmm5, xmm2 ; transpose coefficients(phase 1) 230 unpcklps xmm2, xmm3 ; xmm2=(42 43 52 53) 231 unpckhps xmm5, xmm3 ; xmm5=(62 63 72 73) 232 233 movaps xmm6, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FAST_FLOAT)] 234 movaps xmm7, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FAST_FLOAT)] 235 movaps xmm1, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_FAST_FLOAT)] 236 movaps xmm3, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_FAST_FLOAT)] 237 238 ; xmm6=(00 10 20 30), xmm1=(40 50 60 70) 239 ; xmm7=(01 11 21 31), xmm3=(41 51 61 71) 240 241 movaps XMMWORD [wk(0)], xmm4 ; wk(0)=(22 23 32 33) 242 movaps XMMWORD [wk(1)], xmm2 ; wk(1)=(42 43 52 53) 243 244 movaps xmm4, xmm6 ; transpose coefficients(phase 1) 245 unpcklps xmm6, xmm7 ; xmm6=(00 01 10 11) 246 unpckhps xmm4, xmm7 ; xmm4=(20 21 30 31) 247 movaps xmm2, xmm1 ; transpose coefficients(phase 1) 248 unpcklps xmm1, xmm3 ; xmm1=(40 41 50 51) 249 unpckhps xmm2, xmm3 ; xmm2=(60 61 70 71) 250 251 movaps xmm7, xmm6 ; transpose coefficients(phase 2) 252 unpcklps2 xmm6, xmm0 ; xmm6=(00 01 02 03)=data0 253 unpckhps2 xmm7, xmm0 ; xmm7=(10 11 12 13)=data1 254 movaps xmm3, xmm2 ; transpose coefficients(phase 2) 255 unpcklps2 xmm2, xmm5 ; xmm2=(60 61 62 63)=data6 256 unpckhps2 xmm3, xmm5 ; xmm3=(70 71 72 73)=data7 257 258 movaps xmm0, xmm7 259 movaps xmm5, xmm6 260 subps xmm7, xmm2 ; xmm7=data1-data6=tmp6 261 subps xmm6, xmm3 ; xmm6=data0-data7=tmp7 262 addps xmm0, xmm2 ; xmm0=data1+data6=tmp1 263 addps xmm5, xmm3 ; xmm5=data0+data7=tmp0 264 265 movaps xmm2, XMMWORD [wk(0)] ; xmm2=(22 23 32 33) 266 movaps xmm3, XMMWORD [wk(1)] ; xmm3=(42 43 52 53) 267 movaps XMMWORD [wk(0)], xmm7 ; wk(0)=tmp6 268 movaps XMMWORD [wk(1)], xmm6 ; wk(1)=tmp7 269 270 movaps xmm7, xmm4 ; transpose coefficients(phase 2) 271 unpcklps2 xmm4, xmm2 ; xmm4=(20 21 22 23)=data2 272 unpckhps2 xmm7, xmm2 ; xmm7=(30 31 32 33)=data3 273 movaps xmm6, xmm1 ; transpose coefficients(phase 2) 274 unpcklps2 xmm1, xmm3 ; xmm1=(40 41 42 43)=data4 275 unpckhps2 xmm6, xmm3 ; xmm6=(50 51 52 53)=data5 276 277 movaps xmm2, xmm7 278 movaps xmm3, xmm4 279 addps xmm7, xmm1 ; xmm7=data3+data4=tmp3 280 addps xmm4, xmm6 ; xmm4=data2+data5=tmp2 281 subps xmm2, xmm1 ; xmm2=data3-data4=tmp4 282 subps xmm3, xmm6 ; xmm3=data2-data5=tmp5 283 284 ; -- Even part 285 286 movaps xmm1, xmm5 287 movaps xmm6, xmm0 288 subps xmm5, xmm7 ; xmm5=tmp13 289 subps xmm0, xmm4 ; xmm0=tmp12 290 addps xmm1, xmm7 ; xmm1=tmp10 291 addps xmm6, xmm4 ; xmm6=tmp11 292 293 addps xmm0, xmm5 294 mulps xmm0, [rel PD_0_707] ; xmm0=z1 295 296 movaps xmm7, xmm1 297 movaps xmm4, xmm5 298 subps xmm1, xmm6 ; xmm1=data4 299 subps xmm5, xmm0 ; xmm5=data6 300 addps xmm7, xmm6 ; xmm7=data0 301 addps xmm4, xmm0 ; xmm4=data2 302 303 movaps XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_FAST_FLOAT)], xmm1 304 movaps XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_FAST_FLOAT)], xmm5 305 movaps XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FAST_FLOAT)], xmm7 306 movaps XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_FAST_FLOAT)], xmm4 307 308 ; -- Odd part 309 310 movaps xmm6, XMMWORD [wk(0)] ; xmm6=tmp6 311 movaps xmm0, XMMWORD [wk(1)] ; xmm0=tmp7 312 313 addps xmm2, xmm3 ; xmm2=tmp10 314 addps xmm3, xmm6 ; xmm3=tmp11 315 addps xmm6, xmm0 ; xmm6=tmp12, xmm0=tmp7 316 317 mulps xmm3, [rel PD_0_707] ; xmm3=z3 318 319 movaps xmm1, xmm2 ; xmm1=tmp10 320 subps xmm2, xmm6 321 mulps xmm2, [rel PD_0_382] ; xmm2=z5 322 mulps xmm1, [rel PD_0_541] ; xmm1=MULTIPLY(tmp10,FIX_0_541196) 323 mulps xmm6, [rel PD_1_306] ; xmm6=MULTIPLY(tmp12,FIX_1_306562) 324 addps xmm1, xmm2 ; xmm1=z2 325 addps xmm6, xmm2 ; xmm6=z4 326 327 movaps xmm5, xmm0 328 subps xmm0, xmm3 ; xmm0=z13 329 addps xmm5, xmm3 ; xmm5=z11 330 331 movaps xmm7, xmm0 332 movaps xmm4, xmm5 333 subps xmm0, xmm1 ; xmm0=data3 334 subps xmm5, xmm6 ; xmm5=data7 335 addps xmm7, xmm1 ; xmm7=data5 336 addps xmm4, xmm6 ; xmm4=data1 337 338 movaps XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_FAST_FLOAT)], xmm0 339 movaps XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_FAST_FLOAT)], xmm5 340 movaps XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_FAST_FLOAT)], xmm7 341 movaps XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FAST_FLOAT)], xmm4 342 343 add rdx, byte 4*SIZEOF_FAST_FLOAT 344 dec rcx 345 jnz near .columnloop 346 347 uncollect_args 1 348 mov rsp, rbp ; rsp <- aligned rbp 349 pop rsp ; rsp <- original rbp 350 pop rbp 351 ret 352 353; For some reason, the OS X linker does not honor the request to align the 354; segment unless we do this. 355 align 32 356