jfdctflt-3dn.asm (12061B)
1 ; 2 ; jfdctflt.asm - floating-point FDCT (3DNow!) 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 floating-point implementation of the forward DCT 14 ; (Discrete Cosine Transform). The following code is based directly on 15 ; the IJG's original jfdctflt.c; see the jfdctflt.c for more details. 16 17 %include "jsimdext.inc" 18 %include "jdct.inc" 19 20 ; -------------------------------------------------------------------------- 21 SECTION SEG_CONST 22 23 ALIGNZ 32 24 GLOBAL_DATA(jconst_fdct_float_3dnow) 25 26 EXTN(jconst_fdct_float_3dnow): 27 28 PD_0_382 times 2 dd 0.382683432365089771728460 29 PD_0_707 times 2 dd 0.707106781186547524400844 30 PD_0_541 times 2 dd 0.541196100146196984399723 31 PD_1_306 times 2 dd 1.306562964876376527856643 32 33 ALIGNZ 32 34 35 ; -------------------------------------------------------------------------- 36 SECTION SEG_TEXT 37 BITS 32 38 ; 39 ; Perform the forward DCT on one block of samples. 40 ; 41 ; GLOBAL(void) 42 ; jsimd_fdct_float_3dnow(FAST_FLOAT *data) 43 ; 44 45 %define data(b) (b) + 8 ; FAST_FLOAT *data 46 47 %define original_ebp ebp + 0 48 %define wk(i) ebp - (WK_NUM - (i)) * SIZEOF_MMWORD ; mmword wk[WK_NUM] 49 %define WK_NUM 2 50 51 align 32 52 GLOBAL_FUNCTION(jsimd_fdct_float_3dnow) 53 54 EXTN(jsimd_fdct_float_3dnow): 55 push ebp 56 mov eax, esp ; eax = original ebp 57 sub esp, byte 4 58 and esp, byte (-SIZEOF_MMWORD) ; align to 64 bits 59 mov [esp], eax 60 mov ebp, esp ; ebp = aligned ebp 61 lea esp, [wk(0)] 62 PUSHPIC ebx 63 ; push ecx ; need not be preserved 64 ; push edx ; need not be preserved 65 ; push esi ; unused 66 ; push edi ; unused 67 68 GET_GOT ebx ; get GOT address 69 70 ; ---- Pass 1: process rows. 71 72 mov edx, POINTER [data(eax)] ; (FAST_FLOAT *) 73 mov ecx, DCTSIZE/2 74 ALIGNX 16, 7 75 .rowloop: 76 77 movq mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)] 78 movq mm1, MMWORD [MMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)] 79 movq mm2, MMWORD [MMBLOCK(0,3,edx,SIZEOF_FAST_FLOAT)] 80 movq mm3, MMWORD [MMBLOCK(1,3,edx,SIZEOF_FAST_FLOAT)] 81 82 ; mm0=(00 01), mm1=(10 11), mm2=(06 07), mm3=(16 17) 83 84 movq mm4, mm0 ; transpose coefficients 85 punpckldq mm0, mm1 ; mm0=(00 10)=data0 86 punpckhdq mm4, mm1 ; mm4=(01 11)=data1 87 movq mm5, mm2 ; transpose coefficients 88 punpckldq mm2, mm3 ; mm2=(06 16)=data6 89 punpckhdq mm5, mm3 ; mm5=(07 17)=data7 90 91 movq mm6, mm4 92 movq mm7, mm0 93 pfsub mm4, mm2 ; mm4=data1-data6=tmp6 94 pfsub mm0, mm5 ; mm0=data0-data7=tmp7 95 pfadd mm6, mm2 ; mm6=data1+data6=tmp1 96 pfadd mm7, mm5 ; mm7=data0+data7=tmp0 97 98 movq mm1, MMWORD [MMBLOCK(0,1,edx,SIZEOF_FAST_FLOAT)] 99 movq mm3, MMWORD [MMBLOCK(1,1,edx,SIZEOF_FAST_FLOAT)] 100 movq mm2, MMWORD [MMBLOCK(0,2,edx,SIZEOF_FAST_FLOAT)] 101 movq mm5, MMWORD [MMBLOCK(1,2,edx,SIZEOF_FAST_FLOAT)] 102 103 ; mm1=(02 03), mm3=(12 13), mm2=(04 05), mm5=(14 15) 104 105 movq MMWORD [wk(0)], mm4 ; wk(0)=tmp6 106 movq MMWORD [wk(1)], mm0 ; wk(1)=tmp7 107 108 movq mm4, mm1 ; transpose coefficients 109 punpckldq mm1, mm3 ; mm1=(02 12)=data2 110 punpckhdq mm4, mm3 ; mm4=(03 13)=data3 111 movq mm0, mm2 ; transpose coefficients 112 punpckldq mm2, mm5 ; mm2=(04 14)=data4 113 punpckhdq mm0, mm5 ; mm0=(05 15)=data5 114 115 movq mm3, mm4 116 movq mm5, mm1 117 pfadd mm4, mm2 ; mm4=data3+data4=tmp3 118 pfadd mm1, mm0 ; mm1=data2+data5=tmp2 119 pfsub mm3, mm2 ; mm3=data3-data4=tmp4 120 pfsub mm5, mm0 ; mm5=data2-data5=tmp5 121 122 ; -- Even part 123 124 movq mm2, mm7 125 movq mm0, mm6 126 pfsub mm7, mm4 ; mm7=tmp13 127 pfsub mm6, mm1 ; mm6=tmp12 128 pfadd mm2, mm4 ; mm2=tmp10 129 pfadd mm0, mm1 ; mm0=tmp11 130 131 pfadd mm6, mm7 132 pfmul mm6, [GOTOFF(ebx,PD_0_707)] ; mm6=z1 133 134 movq mm4, mm2 135 movq mm1, mm7 136 pfsub mm2, mm0 ; mm2=data4 137 pfsub mm7, mm6 ; mm7=data6 138 pfadd mm4, mm0 ; mm4=data0 139 pfadd mm1, mm6 ; mm1=data2 140 141 movq MMWORD [MMBLOCK(0,2,edx,SIZEOF_FAST_FLOAT)], mm2 142 movq MMWORD [MMBLOCK(0,3,edx,SIZEOF_FAST_FLOAT)], mm7 143 movq MMWORD [MMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)], mm4 144 movq MMWORD [MMBLOCK(0,1,edx,SIZEOF_FAST_FLOAT)], mm1 145 146 ; -- Odd part 147 148 movq mm0, MMWORD [wk(0)] ; mm0=tmp6 149 movq mm6, MMWORD [wk(1)] ; mm6=tmp7 150 151 pfadd mm3, mm5 ; mm3=tmp10 152 pfadd mm5, mm0 ; mm5=tmp11 153 pfadd mm0, mm6 ; mm0=tmp12, mm6=tmp7 154 155 pfmul mm5, [GOTOFF(ebx,PD_0_707)] ; mm5=z3 156 157 movq mm2, mm3 ; mm2=tmp10 158 pfsub mm3, mm0 159 pfmul mm3, [GOTOFF(ebx,PD_0_382)] ; mm3=z5 160 pfmul mm2, [GOTOFF(ebx,PD_0_541)] ; mm2=MULTIPLY(tmp10,FIX_0_54119610) 161 pfmul mm0, [GOTOFF(ebx,PD_1_306)] ; mm0=MULTIPLY(tmp12,FIX_1_30656296) 162 pfadd mm2, mm3 ; mm2=z2 163 pfadd mm0, mm3 ; mm0=z4 164 165 movq mm7, mm6 166 pfsub mm6, mm5 ; mm6=z13 167 pfadd mm7, mm5 ; mm7=z11 168 169 movq mm4, mm6 170 movq mm1, mm7 171 pfsub mm6, mm2 ; mm6=data3 172 pfsub mm7, mm0 ; mm7=data7 173 pfadd mm4, mm2 ; mm4=data5 174 pfadd mm1, mm0 ; mm1=data1 175 176 movq MMWORD [MMBLOCK(1,1,edx,SIZEOF_FAST_FLOAT)], mm6 177 movq MMWORD [MMBLOCK(1,3,edx,SIZEOF_FAST_FLOAT)], mm7 178 movq MMWORD [MMBLOCK(1,2,edx,SIZEOF_FAST_FLOAT)], mm4 179 movq MMWORD [MMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)], mm1 180 181 add edx, byte 2*DCTSIZE*SIZEOF_FAST_FLOAT 182 dec ecx 183 jnz near .rowloop 184 185 ; ---- Pass 2: process columns. 186 187 mov edx, POINTER [data(eax)] ; (FAST_FLOAT *) 188 mov ecx, DCTSIZE/2 189 ALIGNX 16, 7 190 .columnloop: 191 192 movq mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)] 193 movq mm1, MMWORD [MMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)] 194 movq mm2, MMWORD [MMBLOCK(6,0,edx,SIZEOF_FAST_FLOAT)] 195 movq mm3, MMWORD [MMBLOCK(7,0,edx,SIZEOF_FAST_FLOAT)] 196 197 ; mm0=(00 10), mm1=(01 11), mm2=(60 70), mm3=(61 71) 198 199 movq mm4, mm0 ; transpose coefficients 200 punpckldq mm0, mm1 ; mm0=(00 01)=data0 201 punpckhdq mm4, mm1 ; mm4=(10 11)=data1 202 movq mm5, mm2 ; transpose coefficients 203 punpckldq mm2, mm3 ; mm2=(60 61)=data6 204 punpckhdq mm5, mm3 ; mm5=(70 71)=data7 205 206 movq mm6, mm4 207 movq mm7, mm0 208 pfsub mm4, mm2 ; mm4=data1-data6=tmp6 209 pfsub mm0, mm5 ; mm0=data0-data7=tmp7 210 pfadd mm6, mm2 ; mm6=data1+data6=tmp1 211 pfadd mm7, mm5 ; mm7=data0+data7=tmp0 212 213 movq mm1, MMWORD [MMBLOCK(2,0,edx,SIZEOF_FAST_FLOAT)] 214 movq mm3, MMWORD [MMBLOCK(3,0,edx,SIZEOF_FAST_FLOAT)] 215 movq mm2, MMWORD [MMBLOCK(4,0,edx,SIZEOF_FAST_FLOAT)] 216 movq mm5, MMWORD [MMBLOCK(5,0,edx,SIZEOF_FAST_FLOAT)] 217 218 ; mm1=(20 30), mm3=(21 31), mm2=(40 50), mm5=(41 51) 219 220 movq MMWORD [wk(0)], mm4 ; wk(0)=tmp6 221 movq MMWORD [wk(1)], mm0 ; wk(1)=tmp7 222 223 movq mm4, mm1 ; transpose coefficients 224 punpckldq mm1, mm3 ; mm1=(20 21)=data2 225 punpckhdq mm4, mm3 ; mm4=(30 31)=data3 226 movq mm0, mm2 ; transpose coefficients 227 punpckldq mm2, mm5 ; mm2=(40 41)=data4 228 punpckhdq mm0, mm5 ; mm0=(50 51)=data5 229 230 movq mm3, mm4 231 movq mm5, mm1 232 pfadd mm4, mm2 ; mm4=data3+data4=tmp3 233 pfadd mm1, mm0 ; mm1=data2+data5=tmp2 234 pfsub mm3, mm2 ; mm3=data3-data4=tmp4 235 pfsub mm5, mm0 ; mm5=data2-data5=tmp5 236 237 ; -- Even part 238 239 movq mm2, mm7 240 movq mm0, mm6 241 pfsub mm7, mm4 ; mm7=tmp13 242 pfsub mm6, mm1 ; mm6=tmp12 243 pfadd mm2, mm4 ; mm2=tmp10 244 pfadd mm0, mm1 ; mm0=tmp11 245 246 pfadd mm6, mm7 247 pfmul mm6, [GOTOFF(ebx,PD_0_707)] ; mm6=z1 248 249 movq mm4, mm2 250 movq mm1, mm7 251 pfsub mm2, mm0 ; mm2=data4 252 pfsub mm7, mm6 ; mm7=data6 253 pfadd mm4, mm0 ; mm4=data0 254 pfadd mm1, mm6 ; mm1=data2 255 256 movq MMWORD [MMBLOCK(4,0,edx,SIZEOF_FAST_FLOAT)], mm2 257 movq MMWORD [MMBLOCK(6,0,edx,SIZEOF_FAST_FLOAT)], mm7 258 movq MMWORD [MMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)], mm4 259 movq MMWORD [MMBLOCK(2,0,edx,SIZEOF_FAST_FLOAT)], mm1 260 261 ; -- Odd part 262 263 movq mm0, MMWORD [wk(0)] ; mm0=tmp6 264 movq mm6, MMWORD [wk(1)] ; mm6=tmp7 265 266 pfadd mm3, mm5 ; mm3=tmp10 267 pfadd mm5, mm0 ; mm5=tmp11 268 pfadd mm0, mm6 ; mm0=tmp12, mm6=tmp7 269 270 pfmul mm5, [GOTOFF(ebx,PD_0_707)] ; mm5=z3 271 272 movq mm2, mm3 ; mm2=tmp10 273 pfsub mm3, mm0 274 pfmul mm3, [GOTOFF(ebx,PD_0_382)] ; mm3=z5 275 pfmul mm2, [GOTOFF(ebx,PD_0_541)] ; mm2=MULTIPLY(tmp10,FIX_0_54119610) 276 pfmul mm0, [GOTOFF(ebx,PD_1_306)] ; mm0=MULTIPLY(tmp12,FIX_1_30656296) 277 pfadd mm2, mm3 ; mm2=z2 278 pfadd mm0, mm3 ; mm0=z4 279 280 movq mm7, mm6 281 pfsub mm6, mm5 ; mm6=z13 282 pfadd mm7, mm5 ; mm7=z11 283 284 movq mm4, mm6 285 movq mm1, mm7 286 pfsub mm6, mm2 ; mm6=data3 287 pfsub mm7, mm0 ; mm7=data7 288 pfadd mm4, mm2 ; mm4=data5 289 pfadd mm1, mm0 ; mm1=data1 290 291 movq MMWORD [MMBLOCK(3,0,edx,SIZEOF_FAST_FLOAT)], mm6 292 movq MMWORD [MMBLOCK(7,0,edx,SIZEOF_FAST_FLOAT)], mm7 293 movq MMWORD [MMBLOCK(5,0,edx,SIZEOF_FAST_FLOAT)], mm4 294 movq MMWORD [MMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)], mm1 295 296 add edx, byte 2*SIZEOF_FAST_FLOAT 297 dec ecx 298 jnz near .columnloop 299 300 femms ; empty MMX/3DNow! state 301 302 ; pop edi ; unused 303 ; pop esi ; unused 304 ; pop edx ; need not be preserved 305 ; pop ecx ; need not be preserved 306 POPPIC ebx 307 mov esp, ebp ; esp <- aligned ebp 308 pop esp ; esp <- original ebp 309 pop ebp 310 ret 311 312 ; For some reason, the OS X linker does not honor the request to align the 313 ; segment unless we do this. 314 align 32