The Tor Browser / annotate

media/libjpeg/simd/jfmmxint.asm@b8a032363ba2 (annotated)

media/libjpeg/simd/jfmmxint.asm

Thu, 22 Jan 2015 13:21:57 +0100

author
Michael Schloh von Bennewitz <michael@schloh.com>
date
Thu, 22 Jan 2015 13:21:57 +0100
branch
TOR_BUG_9701
changeset 15
b8a032363ba2
permissions
-rw-r--r--

Incorporate requested changes from Mozilla in review:
https://bugzilla.mozilla.org/show_bug.cgi?id=1123480#c6

michael@0 1 ;
michael@0 2 ; jfmmxint.asm - accurate integer FDCT (MMX)
michael@0 3 ;
michael@0 4 ; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
michael@0 5 ;
michael@0 6 ; Based on
michael@0 7 ; x86 SIMD extension for IJG JPEG library
michael@0 8 ; Copyright (C) 1999-2006, MIYASAKA Masaru.
michael@0 9 ; For conditions of distribution and use, see copyright notice in jsimdext.inc
michael@0 10 ;
michael@0 11 ; This file should be assembled with NASM (Netwide Assembler),
michael@0 12 ; can *not* be assembled with Microsoft's MASM or any compatible
michael@0 13 ; assembler (including Borland's Turbo Assembler).
michael@0 14 ; NASM is available from http://nasm.sourceforge.net/ or
michael@0 15 ; http://sourceforge.net/project/showfiles.php?group_id=6208
michael@0 16 ;
michael@0 17 ; This file contains a slow-but-accurate integer implementation of the
michael@0 18 ; forward DCT (Discrete Cosine Transform). The following code is based
michael@0 19 ; directly on the IJG's original jfdctint.c; see the jfdctint.c for
michael@0 20 ; more details.
michael@0 21 ;
michael@0 22 ; [TAB8]
michael@0 23
michael@0 24 %include "jsimdext.inc"
michael@0 25 %include "jdct.inc"
michael@0 26
michael@0 27 ; --------------------------------------------------------------------------
michael@0 28
michael@0 29 %define CONST_BITS 13
michael@0 30 %define PASS1_BITS 2
michael@0 31
michael@0 32 %define DESCALE_P1 (CONST_BITS-PASS1_BITS)
michael@0 33 %define DESCALE_P2 (CONST_BITS+PASS1_BITS)
michael@0 34
michael@0 35 %if CONST_BITS == 13
michael@0 36 F_0_298 equ 2446 ; FIX(0.298631336)
michael@0 37 F_0_390 equ 3196 ; FIX(0.390180644)
michael@0 38 F_0_541 equ 4433 ; FIX(0.541196100)
michael@0 39 F_0_765 equ 6270 ; FIX(0.765366865)
michael@0 40 F_0_899 equ 7373 ; FIX(0.899976223)
michael@0 41 F_1_175 equ 9633 ; FIX(1.175875602)
michael@0 42 F_1_501 equ 12299 ; FIX(1.501321110)
michael@0 43 F_1_847 equ 15137 ; FIX(1.847759065)
michael@0 44 F_1_961 equ 16069 ; FIX(1.961570560)
michael@0 45 F_2_053 equ 16819 ; FIX(2.053119869)
michael@0 46 F_2_562 equ 20995 ; FIX(2.562915447)
michael@0 47 F_3_072 equ 25172 ; FIX(3.072711026)
michael@0 48 %else
michael@0 49 ; NASM cannot do compile-time arithmetic on floating-point constants.
michael@0 50 %define DESCALE(x,n) (((x)+(1<<((n)-1)))>>(n))
michael@0 51 F_0_298 equ DESCALE( 320652955,30-CONST_BITS) ; FIX(0.298631336)
michael@0 52 F_0_390 equ DESCALE( 418953276,30-CONST_BITS) ; FIX(0.390180644)
michael@0 53 F_0_541 equ DESCALE( 581104887,30-CONST_BITS) ; FIX(0.541196100)
michael@0 54 F_0_765 equ DESCALE( 821806413,30-CONST_BITS) ; FIX(0.765366865)
michael@0 55 F_0_899 equ DESCALE( 966342111,30-CONST_BITS) ; FIX(0.899976223)
michael@0 56 F_1_175 equ DESCALE(1262586813,30-CONST_BITS) ; FIX(1.175875602)
michael@0 57 F_1_501 equ DESCALE(1612031267,30-CONST_BITS) ; FIX(1.501321110)
michael@0 58 F_1_847 equ DESCALE(1984016188,30-CONST_BITS) ; FIX(1.847759065)
michael@0 59 F_1_961 equ DESCALE(2106220350,30-CONST_BITS) ; FIX(1.961570560)
michael@0 60 F_2_053 equ DESCALE(2204520673,30-CONST_BITS) ; FIX(2.053119869)
michael@0 61 F_2_562 equ DESCALE(2751909506,30-CONST_BITS) ; FIX(2.562915447)
michael@0 62 F_3_072 equ DESCALE(3299298341,30-CONST_BITS) ; FIX(3.072711026)
michael@0 63 %endif
michael@0 64
michael@0 65 ; --------------------------------------------------------------------------
michael@0 66 SECTION SEG_CONST
michael@0 67
michael@0 68 alignz 16
michael@0 69 global EXTN(jconst_fdct_islow_mmx)
michael@0 70
michael@0 71 EXTN(jconst_fdct_islow_mmx):
michael@0 72
michael@0 73 PW_F130_F054 times 2 dw (F_0_541+F_0_765), F_0_541
michael@0 74 PW_F054_MF130 times 2 dw F_0_541, (F_0_541-F_1_847)
michael@0 75 PW_MF078_F117 times 2 dw (F_1_175-F_1_961), F_1_175
michael@0 76 PW_F117_F078 times 2 dw F_1_175, (F_1_175-F_0_390)
michael@0 77 PW_MF060_MF089 times 2 dw (F_0_298-F_0_899),-F_0_899
michael@0 78 PW_MF089_F060 times 2 dw -F_0_899, (F_1_501-F_0_899)
michael@0 79 PW_MF050_MF256 times 2 dw (F_2_053-F_2_562),-F_2_562
michael@0 80 PW_MF256_F050 times 2 dw -F_2_562, (F_3_072-F_2_562)
michael@0 81 PD_DESCALE_P1 times 2 dd 1 << (DESCALE_P1-1)
michael@0 82 PD_DESCALE_P2 times 2 dd 1 << (DESCALE_P2-1)
michael@0 83 PW_DESCALE_P2X times 4 dw 1 << (PASS1_BITS-1)
michael@0 84
michael@0 85 alignz 16
michael@0 86
michael@0 87 ; --------------------------------------------------------------------------
michael@0 88 SECTION SEG_TEXT
michael@0 89 BITS 32
michael@0 90 ;
michael@0 91 ; Perform the forward DCT on one block of samples.
michael@0 92 ;
michael@0 93 ; GLOBAL(void)
michael@0 94 ; jsimd_fdct_islow_mmx (DCTELEM * data)
michael@0 95 ;
michael@0 96
michael@0 97 %define data(b) (b)+8 ; DCTELEM * data
michael@0 98
michael@0 99 %define original_ebp ebp+0
michael@0 100 %define wk(i) ebp-(WK_NUM-(i))*SIZEOF_MMWORD ; mmword wk[WK_NUM]
michael@0 101 %define WK_NUM 2
michael@0 102
michael@0 103 align 16
michael@0 104 global EXTN(jsimd_fdct_islow_mmx)
michael@0 105
michael@0 106 EXTN(jsimd_fdct_islow_mmx):
michael@0 107 push ebp
michael@0 108 mov eax,esp ; eax = original ebp
michael@0 109 sub esp, byte 4
michael@0 110 and esp, byte (-SIZEOF_MMWORD) ; align to 64 bits
michael@0 111 mov [esp],eax
michael@0 112 mov ebp,esp ; ebp = aligned ebp
michael@0 113 lea esp, [wk(0)]
michael@0 114 pushpic ebx
michael@0 115 ; push ecx ; need not be preserved
michael@0 116 ; push edx ; need not be preserved
michael@0 117 ; push esi ; unused
michael@0 118 ; push edi ; unused
michael@0 119
michael@0 120 get_GOT ebx ; get GOT address
michael@0 121
michael@0 122 ; ---- Pass 1: process rows.
michael@0 123
michael@0 124 mov edx, POINTER [data(eax)] ; (DCTELEM *)
michael@0 125 mov ecx, DCTSIZE/4
michael@0 126 alignx 16,7
michael@0 127 .rowloop:
michael@0 128
michael@0 129 movq mm0, MMWORD [MMBLOCK(2,0,edx,SIZEOF_DCTELEM)]
michael@0 130 movq mm1, MMWORD [MMBLOCK(3,0,edx,SIZEOF_DCTELEM)]
michael@0 131 movq mm2, MMWORD [MMBLOCK(2,1,edx,SIZEOF_DCTELEM)]
michael@0 132 movq mm3, MMWORD [MMBLOCK(3,1,edx,SIZEOF_DCTELEM)]
michael@0 133
michael@0 134 ; mm0=(20 21 22 23), mm2=(24 25 26 27)
michael@0 135 ; mm1=(30 31 32 33), mm3=(34 35 36 37)
michael@0 136
michael@0 137 movq mm4,mm0 ; transpose coefficients(phase 1)
michael@0 138 punpcklwd mm0,mm1 ; mm0=(20 30 21 31)
michael@0 139 punpckhwd mm4,mm1 ; mm4=(22 32 23 33)
michael@0 140 movq mm5,mm2 ; transpose coefficients(phase 1)
michael@0 141 punpcklwd mm2,mm3 ; mm2=(24 34 25 35)
michael@0 142 punpckhwd mm5,mm3 ; mm5=(26 36 27 37)
michael@0 143
michael@0 144 movq mm6, MMWORD [MMBLOCK(0,0,edx,SIZEOF_DCTELEM)]
michael@0 145 movq mm7, MMWORD [MMBLOCK(1,0,edx,SIZEOF_DCTELEM)]
michael@0 146 movq mm1, MMWORD [MMBLOCK(0,1,edx,SIZEOF_DCTELEM)]
michael@0 147 movq mm3, MMWORD [MMBLOCK(1,1,edx,SIZEOF_DCTELEM)]
michael@0 148
michael@0 149 ; mm6=(00 01 02 03), mm1=(04 05 06 07)
michael@0 150 ; mm7=(10 11 12 13), mm3=(14 15 16 17)
michael@0 151
michael@0 152 movq MMWORD [wk(0)], mm4 ; wk(0)=(22 32 23 33)
michael@0 153 movq MMWORD [wk(1)], mm2 ; wk(1)=(24 34 25 35)
michael@0 154
michael@0 155 movq mm4,mm6 ; transpose coefficients(phase 1)
michael@0 156 punpcklwd mm6,mm7 ; mm6=(00 10 01 11)
michael@0 157 punpckhwd mm4,mm7 ; mm4=(02 12 03 13)
michael@0 158 movq mm2,mm1 ; transpose coefficients(phase 1)
michael@0 159 punpcklwd mm1,mm3 ; mm1=(04 14 05 15)
michael@0 160 punpckhwd mm2,mm3 ; mm2=(06 16 07 17)
michael@0 161
michael@0 162 movq mm7,mm6 ; transpose coefficients(phase 2)
michael@0 163 punpckldq mm6,mm0 ; mm6=(00 10 20 30)=data0
michael@0 164 punpckhdq mm7,mm0 ; mm7=(01 11 21 31)=data1
michael@0 165 movq mm3,mm2 ; transpose coefficients(phase 2)
michael@0 166 punpckldq mm2,mm5 ; mm2=(06 16 26 36)=data6
michael@0 167 punpckhdq mm3,mm5 ; mm3=(07 17 27 37)=data7
michael@0 168
michael@0 169 movq mm0,mm7
michael@0 170 movq mm5,mm6
michael@0 171 psubw mm7,mm2 ; mm7=data1-data6=tmp6
michael@0 172 psubw mm6,mm3 ; mm6=data0-data7=tmp7
michael@0 173 paddw mm0,mm2 ; mm0=data1+data6=tmp1
michael@0 174 paddw mm5,mm3 ; mm5=data0+data7=tmp0
michael@0 175
michael@0 176 movq mm2, MMWORD [wk(0)] ; mm2=(22 32 23 33)
michael@0 177 movq mm3, MMWORD [wk(1)] ; mm3=(24 34 25 35)
michael@0 178 movq MMWORD [wk(0)], mm7 ; wk(0)=tmp6
michael@0 179 movq MMWORD [wk(1)], mm6 ; wk(1)=tmp7
michael@0 180
michael@0 181 movq mm7,mm4 ; transpose coefficients(phase 2)
michael@0 182 punpckldq mm4,mm2 ; mm4=(02 12 22 32)=data2
michael@0 183 punpckhdq mm7,mm2 ; mm7=(03 13 23 33)=data3
michael@0 184 movq mm6,mm1 ; transpose coefficients(phase 2)
michael@0 185 punpckldq mm1,mm3 ; mm1=(04 14 24 34)=data4
michael@0 186 punpckhdq mm6,mm3 ; mm6=(05 15 25 35)=data5
michael@0 187
michael@0 188 movq mm2,mm7
michael@0 189 movq mm3,mm4
michael@0 190 paddw mm7,mm1 ; mm7=data3+data4=tmp3
michael@0 191 paddw mm4,mm6 ; mm4=data2+data5=tmp2
michael@0 192 psubw mm2,mm1 ; mm2=data3-data4=tmp4
michael@0 193 psubw mm3,mm6 ; mm3=data2-data5=tmp5
michael@0 194
michael@0 195 ; -- Even part
michael@0 196
michael@0 197 movq mm1,mm5
michael@0 198 movq mm6,mm0
michael@0 199 paddw mm5,mm7 ; mm5=tmp10
michael@0 200 paddw mm0,mm4 ; mm0=tmp11
michael@0 201 psubw mm1,mm7 ; mm1=tmp13
michael@0 202 psubw mm6,mm4 ; mm6=tmp12
michael@0 203
michael@0 204 movq mm7,mm5
michael@0 205 paddw mm5,mm0 ; mm5=tmp10+tmp11
michael@0 206 psubw mm7,mm0 ; mm7=tmp10-tmp11
michael@0 207
michael@0 208 psllw mm5,PASS1_BITS ; mm5=data0
michael@0 209 psllw mm7,PASS1_BITS ; mm7=data4
michael@0 210
michael@0 211 movq MMWORD [MMBLOCK(0,0,edx,SIZEOF_DCTELEM)], mm5
michael@0 212 movq MMWORD [MMBLOCK(0,1,edx,SIZEOF_DCTELEM)], mm7
michael@0 213
michael@0 214 ; (Original)
michael@0 215 ; z1 = (tmp12 + tmp13) * 0.541196100;
michael@0 216 ; data2 = z1 + tmp13 * 0.765366865;
michael@0 217 ; data6 = z1 + tmp12 * -1.847759065;
michael@0 218 ;
michael@0 219 ; (This implementation)
michael@0 220 ; data2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100;
michael@0 221 ; data6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065);
michael@0 222
michael@0 223 movq mm4,mm1 ; mm1=tmp13
michael@0 224 movq mm0,mm1
michael@0 225 punpcklwd mm4,mm6 ; mm6=tmp12
michael@0 226 punpckhwd mm0,mm6
michael@0 227 movq mm1,mm4
michael@0 228 movq mm6,mm0
michael@0 229 pmaddwd mm4,[GOTOFF(ebx,PW_F130_F054)] ; mm4=data2L
michael@0 230 pmaddwd mm0,[GOTOFF(ebx,PW_F130_F054)] ; mm0=data2H
michael@0 231 pmaddwd mm1,[GOTOFF(ebx,PW_F054_MF130)] ; mm1=data6L
michael@0 232 pmaddwd mm6,[GOTOFF(ebx,PW_F054_MF130)] ; mm6=data6H
michael@0 233
michael@0 234 paddd mm4,[GOTOFF(ebx,PD_DESCALE_P1)]
michael@0 235 paddd mm0,[GOTOFF(ebx,PD_DESCALE_P1)]
michael@0 236 psrad mm4,DESCALE_P1
michael@0 237 psrad mm0,DESCALE_P1
michael@0 238 paddd mm1,[GOTOFF(ebx,PD_DESCALE_P1)]
michael@0 239 paddd mm6,[GOTOFF(ebx,PD_DESCALE_P1)]
michael@0 240 psrad mm1,DESCALE_P1
michael@0 241 psrad mm6,DESCALE_P1
michael@0 242
michael@0 243 packssdw mm4,mm0 ; mm4=data2
michael@0 244 packssdw mm1,mm6 ; mm1=data6
michael@0 245
michael@0 246 movq MMWORD [MMBLOCK(2,0,edx,SIZEOF_DCTELEM)], mm4
michael@0 247 movq MMWORD [MMBLOCK(2,1,edx,SIZEOF_DCTELEM)], mm1
michael@0 248
michael@0 249 ; -- Odd part
michael@0 250
michael@0 251 movq mm5, MMWORD [wk(0)] ; mm5=tmp6
michael@0 252 movq mm7, MMWORD [wk(1)] ; mm7=tmp7
michael@0 253
michael@0 254 movq mm0,mm2 ; mm2=tmp4
michael@0 255 movq mm6,mm3 ; mm3=tmp5
michael@0 256 paddw mm0,mm5 ; mm0=z3
michael@0 257 paddw mm6,mm7 ; mm6=z4
michael@0 258
michael@0 259 ; (Original)
michael@0 260 ; z5 = (z3 + z4) * 1.175875602;
michael@0 261 ; z3 = z3 * -1.961570560; z4 = z4 * -0.390180644;
michael@0 262 ; z3 += z5; z4 += z5;
michael@0 263 ;
michael@0 264 ; (This implementation)
michael@0 265 ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
michael@0 266 ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
michael@0 267
michael@0 268 movq mm4,mm0
michael@0 269 movq mm1,mm0
michael@0 270 punpcklwd mm4,mm6
michael@0 271 punpckhwd mm1,mm6
michael@0 272 movq mm0,mm4
michael@0 273 movq mm6,mm1
michael@0 274 pmaddwd mm4,[GOTOFF(ebx,PW_MF078_F117)] ; mm4=z3L
michael@0 275 pmaddwd mm1,[GOTOFF(ebx,PW_MF078_F117)] ; mm1=z3H
michael@0 276 pmaddwd mm0,[GOTOFF(ebx,PW_F117_F078)] ; mm0=z4L
michael@0 277 pmaddwd mm6,[GOTOFF(ebx,PW_F117_F078)] ; mm6=z4H
michael@0 278
michael@0 279 movq MMWORD [wk(0)], mm4 ; wk(0)=z3L
michael@0 280 movq MMWORD [wk(1)], mm1 ; wk(1)=z3H
michael@0 281
michael@0 282 ; (Original)
michael@0 283 ; z1 = tmp4 + tmp7; z2 = tmp5 + tmp6;
michael@0 284 ; tmp4 = tmp4 * 0.298631336; tmp5 = tmp5 * 2.053119869;
michael@0 285 ; tmp6 = tmp6 * 3.072711026; tmp7 = tmp7 * 1.501321110;
michael@0 286 ; z1 = z1 * -0.899976223; z2 = z2 * -2.562915447;
michael@0 287 ; data7 = tmp4 + z1 + z3; data5 = tmp5 + z2 + z4;
michael@0 288 ; data3 = tmp6 + z2 + z3; data1 = tmp7 + z1 + z4;
michael@0 289 ;
michael@0 290 ; (This implementation)
michael@0 291 ; tmp4 = tmp4 * (0.298631336 - 0.899976223) + tmp7 * -0.899976223;
michael@0 292 ; tmp5 = tmp5 * (2.053119869 - 2.562915447) + tmp6 * -2.562915447;
michael@0 293 ; tmp6 = tmp5 * -2.562915447 + tmp6 * (3.072711026 - 2.562915447);
michael@0 294 ; tmp7 = tmp4 * -0.899976223 + tmp7 * (1.501321110 - 0.899976223);
michael@0 295 ; data7 = tmp4 + z3; data5 = tmp5 + z4;
michael@0 296 ; data3 = tmp6 + z3; data1 = tmp7 + z4;
michael@0 297
michael@0 298 movq mm4,mm2
michael@0 299 movq mm1,mm2
michael@0 300 punpcklwd mm4,mm7
michael@0 301 punpckhwd mm1,mm7
michael@0 302 movq mm2,mm4
michael@0 303 movq mm7,mm1
michael@0 304 pmaddwd mm4,[GOTOFF(ebx,PW_MF060_MF089)] ; mm4=tmp4L
michael@0 305 pmaddwd mm1,[GOTOFF(ebx,PW_MF060_MF089)] ; mm1=tmp4H
michael@0 306 pmaddwd mm2,[GOTOFF(ebx,PW_MF089_F060)] ; mm2=tmp7L
michael@0 307 pmaddwd mm7,[GOTOFF(ebx,PW_MF089_F060)] ; mm7=tmp7H
michael@0 308
michael@0 309 paddd mm4, MMWORD [wk(0)] ; mm4=data7L
michael@0 310 paddd mm1, MMWORD [wk(1)] ; mm1=data7H
michael@0 311 paddd mm2,mm0 ; mm2=data1L
michael@0 312 paddd mm7,mm6 ; mm7=data1H
michael@0 313
michael@0 314 paddd mm4,[GOTOFF(ebx,PD_DESCALE_P1)]
michael@0 315 paddd mm1,[GOTOFF(ebx,PD_DESCALE_P1)]
michael@0 316 psrad mm4,DESCALE_P1
michael@0 317 psrad mm1,DESCALE_P1
michael@0 318 paddd mm2,[GOTOFF(ebx,PD_DESCALE_P1)]
michael@0 319 paddd mm7,[GOTOFF(ebx,PD_DESCALE_P1)]
michael@0 320 psrad mm2,DESCALE_P1
michael@0 321 psrad mm7,DESCALE_P1
michael@0 322
michael@0 323 packssdw mm4,mm1 ; mm4=data7
michael@0 324 packssdw mm2,mm7 ; mm2=data1
michael@0 325
michael@0 326 movq MMWORD [MMBLOCK(3,1,edx,SIZEOF_DCTELEM)], mm4
michael@0 327 movq MMWORD [MMBLOCK(1,0,edx,SIZEOF_DCTELEM)], mm2
michael@0 328
michael@0 329 movq mm1,mm3
michael@0 330 movq mm7,mm3
michael@0 331 punpcklwd mm1,mm5
michael@0 332 punpckhwd mm7,mm5
michael@0 333 movq mm3,mm1
michael@0 334 movq mm5,mm7
michael@0 335 pmaddwd mm1,[GOTOFF(ebx,PW_MF050_MF256)] ; mm1=tmp5L
michael@0 336 pmaddwd mm7,[GOTOFF(ebx,PW_MF050_MF256)] ; mm7=tmp5H
michael@0 337 pmaddwd mm3,[GOTOFF(ebx,PW_MF256_F050)] ; mm3=tmp6L
michael@0 338 pmaddwd mm5,[GOTOFF(ebx,PW_MF256_F050)] ; mm5=tmp6H
michael@0 339
michael@0 340 paddd mm1,mm0 ; mm1=data5L
michael@0 341 paddd mm7,mm6 ; mm7=data5H
michael@0 342 paddd mm3, MMWORD [wk(0)] ; mm3=data3L
michael@0 343 paddd mm5, MMWORD [wk(1)] ; mm5=data3H
michael@0 344
michael@0 345 paddd mm1,[GOTOFF(ebx,PD_DESCALE_P1)]
michael@0 346 paddd mm7,[GOTOFF(ebx,PD_DESCALE_P1)]
michael@0 347 psrad mm1,DESCALE_P1
michael@0 348 psrad mm7,DESCALE_P1
michael@0 349 paddd mm3,[GOTOFF(ebx,PD_DESCALE_P1)]
michael@0 350 paddd mm5,[GOTOFF(ebx,PD_DESCALE_P1)]
michael@0 351 psrad mm3,DESCALE_P1
michael@0 352 psrad mm5,DESCALE_P1
michael@0 353
michael@0 354 packssdw mm1,mm7 ; mm1=data5
michael@0 355 packssdw mm3,mm5 ; mm3=data3
michael@0 356
michael@0 357 movq MMWORD [MMBLOCK(1,1,edx,SIZEOF_DCTELEM)], mm1
michael@0 358 movq MMWORD [MMBLOCK(3,0,edx,SIZEOF_DCTELEM)], mm3
michael@0 359
michael@0 360 add edx, byte 4*DCTSIZE*SIZEOF_DCTELEM
michael@0 361 dec ecx
michael@0 362 jnz near .rowloop
michael@0 363
michael@0 364 ; ---- Pass 2: process columns.
michael@0 365
michael@0 366 mov edx, POINTER [data(eax)] ; (DCTELEM *)
michael@0 367 mov ecx, DCTSIZE/4
michael@0 368 alignx 16,7
michael@0 369 .columnloop:
michael@0 370
michael@0 371 movq mm0, MMWORD [MMBLOCK(2,0,edx,SIZEOF_DCTELEM)]
michael@0 372 movq mm1, MMWORD [MMBLOCK(3,0,edx,SIZEOF_DCTELEM)]
michael@0 373 movq mm2, MMWORD [MMBLOCK(6,0,edx,SIZEOF_DCTELEM)]
michael@0 374 movq mm3, MMWORD [MMBLOCK(7,0,edx,SIZEOF_DCTELEM)]
michael@0 375
michael@0 376 ; mm0=(02 12 22 32), mm2=(42 52 62 72)
michael@0 377 ; mm1=(03 13 23 33), mm3=(43 53 63 73)
michael@0 378
michael@0 379 movq mm4,mm0 ; transpose coefficients(phase 1)
michael@0 380 punpcklwd mm0,mm1 ; mm0=(02 03 12 13)
michael@0 381 punpckhwd mm4,mm1 ; mm4=(22 23 32 33)
michael@0 382 movq mm5,mm2 ; transpose coefficients(phase 1)
michael@0 383 punpcklwd mm2,mm3 ; mm2=(42 43 52 53)
michael@0 384 punpckhwd mm5,mm3 ; mm5=(62 63 72 73)
michael@0 385
michael@0 386 movq mm6, MMWORD [MMBLOCK(0,0,edx,SIZEOF_DCTELEM)]
michael@0 387 movq mm7, MMWORD [MMBLOCK(1,0,edx,SIZEOF_DCTELEM)]
michael@0 388 movq mm1, MMWORD [MMBLOCK(4,0,edx,SIZEOF_DCTELEM)]
michael@0 389 movq mm3, MMWORD [MMBLOCK(5,0,edx,SIZEOF_DCTELEM)]
michael@0 390
michael@0 391 ; mm6=(00 10 20 30), mm1=(40 50 60 70)
michael@0 392 ; mm7=(01 11 21 31), mm3=(41 51 61 71)
michael@0 393
michael@0 394 movq MMWORD [wk(0)], mm4 ; wk(0)=(22 23 32 33)
michael@0 395 movq MMWORD [wk(1)], mm2 ; wk(1)=(42 43 52 53)
michael@0 396
michael@0 397 movq mm4,mm6 ; transpose coefficients(phase 1)
michael@0 398 punpcklwd mm6,mm7 ; mm6=(00 01 10 11)
michael@0 399 punpckhwd mm4,mm7 ; mm4=(20 21 30 31)
michael@0 400 movq mm2,mm1 ; transpose coefficients(phase 1)
michael@0 401 punpcklwd mm1,mm3 ; mm1=(40 41 50 51)
michael@0 402 punpckhwd mm2,mm3 ; mm2=(60 61 70 71)
michael@0 403
michael@0 404 movq mm7,mm6 ; transpose coefficients(phase 2)
michael@0 405 punpckldq mm6,mm0 ; mm6=(00 01 02 03)=data0
michael@0 406 punpckhdq mm7,mm0 ; mm7=(10 11 12 13)=data1
michael@0 407 movq mm3,mm2 ; transpose coefficients(phase 2)
michael@0 408 punpckldq mm2,mm5 ; mm2=(60 61 62 63)=data6
michael@0 409 punpckhdq mm3,mm5 ; mm3=(70 71 72 73)=data7
michael@0 410
michael@0 411 movq mm0,mm7
michael@0 412 movq mm5,mm6
michael@0 413 psubw mm7,mm2 ; mm7=data1-data6=tmp6
michael@0 414 psubw mm6,mm3 ; mm6=data0-data7=tmp7
michael@0 415 paddw mm0,mm2 ; mm0=data1+data6=tmp1
michael@0 416 paddw mm5,mm3 ; mm5=data0+data7=tmp0
michael@0 417
michael@0 418 movq mm2, MMWORD [wk(0)] ; mm2=(22 23 32 33)
michael@0 419 movq mm3, MMWORD [wk(1)] ; mm3=(42 43 52 53)
michael@0 420 movq MMWORD [wk(0)], mm7 ; wk(0)=tmp6
michael@0 421 movq MMWORD [wk(1)], mm6 ; wk(1)=tmp7
michael@0 422
michael@0 423 movq mm7,mm4 ; transpose coefficients(phase 2)
michael@0 424 punpckldq mm4,mm2 ; mm4=(20 21 22 23)=data2
michael@0 425 punpckhdq mm7,mm2 ; mm7=(30 31 32 33)=data3
michael@0 426 movq mm6,mm1 ; transpose coefficients(phase 2)
michael@0 427 punpckldq mm1,mm3 ; mm1=(40 41 42 43)=data4
michael@0 428 punpckhdq mm6,mm3 ; mm6=(50 51 52 53)=data5
michael@0 429
michael@0 430 movq mm2,mm7
michael@0 431 movq mm3,mm4
michael@0 432 paddw mm7,mm1 ; mm7=data3+data4=tmp3
michael@0 433 paddw mm4,mm6 ; mm4=data2+data5=tmp2
michael@0 434 psubw mm2,mm1 ; mm2=data3-data4=tmp4
michael@0 435 psubw mm3,mm6 ; mm3=data2-data5=tmp5
michael@0 436
michael@0 437 ; -- Even part
michael@0 438
michael@0 439 movq mm1,mm5
michael@0 440 movq mm6,mm0
michael@0 441 paddw mm5,mm7 ; mm5=tmp10
michael@0 442 paddw mm0,mm4 ; mm0=tmp11
michael@0 443 psubw mm1,mm7 ; mm1=tmp13
michael@0 444 psubw mm6,mm4 ; mm6=tmp12
michael@0 445
michael@0 446 movq mm7,mm5
michael@0 447 paddw mm5,mm0 ; mm5=tmp10+tmp11
michael@0 448 psubw mm7,mm0 ; mm7=tmp10-tmp11
michael@0 449
michael@0 450 paddw mm5,[GOTOFF(ebx,PW_DESCALE_P2X)]
michael@0 451 paddw mm7,[GOTOFF(ebx,PW_DESCALE_P2X)]
michael@0 452 psraw mm5,PASS1_BITS ; mm5=data0
michael@0 453 psraw mm7,PASS1_BITS ; mm7=data4
michael@0 454
michael@0 455 movq MMWORD [MMBLOCK(0,0,edx,SIZEOF_DCTELEM)], mm5
michael@0 456 movq MMWORD [MMBLOCK(4,0,edx,SIZEOF_DCTELEM)], mm7
michael@0 457
michael@0 458 ; (Original)
michael@0 459 ; z1 = (tmp12 + tmp13) * 0.541196100;
michael@0 460 ; data2 = z1 + tmp13 * 0.765366865;
michael@0 461 ; data6 = z1 + tmp12 * -1.847759065;
michael@0 462 ;
michael@0 463 ; (This implementation)
michael@0 464 ; data2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100;
michael@0 465 ; data6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065);
michael@0 466
michael@0 467 movq mm4,mm1 ; mm1=tmp13
michael@0 468 movq mm0,mm1
michael@0 469 punpcklwd mm4,mm6 ; mm6=tmp12
michael@0 470 punpckhwd mm0,mm6
michael@0 471 movq mm1,mm4
michael@0 472 movq mm6,mm0
michael@0 473 pmaddwd mm4,[GOTOFF(ebx,PW_F130_F054)] ; mm4=data2L
michael@0 474 pmaddwd mm0,[GOTOFF(ebx,PW_F130_F054)] ; mm0=data2H
michael@0 475 pmaddwd mm1,[GOTOFF(ebx,PW_F054_MF130)] ; mm1=data6L
michael@0 476 pmaddwd mm6,[GOTOFF(ebx,PW_F054_MF130)] ; mm6=data6H
michael@0 477
michael@0 478 paddd mm4,[GOTOFF(ebx,PD_DESCALE_P2)]
michael@0 479 paddd mm0,[GOTOFF(ebx,PD_DESCALE_P2)]
michael@0 480 psrad mm4,DESCALE_P2
michael@0 481 psrad mm0,DESCALE_P2
michael@0 482 paddd mm1,[GOTOFF(ebx,PD_DESCALE_P2)]
michael@0 483 paddd mm6,[GOTOFF(ebx,PD_DESCALE_P2)]
michael@0 484 psrad mm1,DESCALE_P2
michael@0 485 psrad mm6,DESCALE_P2
michael@0 486
michael@0 487 packssdw mm4,mm0 ; mm4=data2
michael@0 488 packssdw mm1,mm6 ; mm1=data6
michael@0 489
michael@0 490 movq MMWORD [MMBLOCK(2,0,edx,SIZEOF_DCTELEM)], mm4
michael@0 491 movq MMWORD [MMBLOCK(6,0,edx,SIZEOF_DCTELEM)], mm1
michael@0 492
michael@0 493 ; -- Odd part
michael@0 494
michael@0 495 movq mm5, MMWORD [wk(0)] ; mm5=tmp6
michael@0 496 movq mm7, MMWORD [wk(1)] ; mm7=tmp7
michael@0 497
michael@0 498 movq mm0,mm2 ; mm2=tmp4
michael@0 499 movq mm6,mm3 ; mm3=tmp5
michael@0 500 paddw mm0,mm5 ; mm0=z3
michael@0 501 paddw mm6,mm7 ; mm6=z4
michael@0 502
michael@0 503 ; (Original)
michael@0 504 ; z5 = (z3 + z4) * 1.175875602;
michael@0 505 ; z3 = z3 * -1.961570560; z4 = z4 * -0.390180644;
michael@0 506 ; z3 += z5; z4 += z5;
michael@0 507 ;
michael@0 508 ; (This implementation)
michael@0 509 ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
michael@0 510 ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
michael@0 511
michael@0 512 movq mm4,mm0
michael@0 513 movq mm1,mm0
michael@0 514 punpcklwd mm4,mm6
michael@0 515 punpckhwd mm1,mm6
michael@0 516 movq mm0,mm4
michael@0 517 movq mm6,mm1
michael@0 518 pmaddwd mm4,[GOTOFF(ebx,PW_MF078_F117)] ; mm4=z3L
michael@0 519 pmaddwd mm1,[GOTOFF(ebx,PW_MF078_F117)] ; mm1=z3H
michael@0 520 pmaddwd mm0,[GOTOFF(ebx,PW_F117_F078)] ; mm0=z4L
michael@0 521 pmaddwd mm6,[GOTOFF(ebx,PW_F117_F078)] ; mm6=z4H
michael@0 522
michael@0 523 movq MMWORD [wk(0)], mm4 ; wk(0)=z3L
michael@0 524 movq MMWORD [wk(1)], mm1 ; wk(1)=z3H
michael@0 525
michael@0 526 ; (Original)
michael@0 527 ; z1 = tmp4 + tmp7; z2 = tmp5 + tmp6;
michael@0 528 ; tmp4 = tmp4 * 0.298631336; tmp5 = tmp5 * 2.053119869;
michael@0 529 ; tmp6 = tmp6 * 3.072711026; tmp7 = tmp7 * 1.501321110;
michael@0 530 ; z1 = z1 * -0.899976223; z2 = z2 * -2.562915447;
michael@0 531 ; data7 = tmp4 + z1 + z3; data5 = tmp5 + z2 + z4;
michael@0 532 ; data3 = tmp6 + z2 + z3; data1 = tmp7 + z1 + z4;
michael@0 533 ;
michael@0 534 ; (This implementation)
michael@0 535 ; tmp4 = tmp4 * (0.298631336 - 0.899976223) + tmp7 * -0.899976223;
michael@0 536 ; tmp5 = tmp5 * (2.053119869 - 2.562915447) + tmp6 * -2.562915447;
michael@0 537 ; tmp6 = tmp5 * -2.562915447 + tmp6 * (3.072711026 - 2.562915447);
michael@0 538 ; tmp7 = tmp4 * -0.899976223 + tmp7 * (1.501321110 - 0.899976223);
michael@0 539 ; data7 = tmp4 + z3; data5 = tmp5 + z4;
michael@0 540 ; data3 = tmp6 + z3; data1 = tmp7 + z4;
michael@0 541
michael@0 542 movq mm4,mm2
michael@0 543 movq mm1,mm2
michael@0 544 punpcklwd mm4,mm7
michael@0 545 punpckhwd mm1,mm7
michael@0 546 movq mm2,mm4
michael@0 547 movq mm7,mm1
michael@0 548 pmaddwd mm4,[GOTOFF(ebx,PW_MF060_MF089)] ; mm4=tmp4L
michael@0 549 pmaddwd mm1,[GOTOFF(ebx,PW_MF060_MF089)] ; mm1=tmp4H
michael@0 550 pmaddwd mm2,[GOTOFF(ebx,PW_MF089_F060)] ; mm2=tmp7L
michael@0 551 pmaddwd mm7,[GOTOFF(ebx,PW_MF089_F060)] ; mm7=tmp7H
michael@0 552
michael@0 553 paddd mm4, MMWORD [wk(0)] ; mm4=data7L
michael@0 554 paddd mm1, MMWORD [wk(1)] ; mm1=data7H
michael@0 555 paddd mm2,mm0 ; mm2=data1L
michael@0 556 paddd mm7,mm6 ; mm7=data1H
michael@0 557
michael@0 558 paddd mm4,[GOTOFF(ebx,PD_DESCALE_P2)]
michael@0 559 paddd mm1,[GOTOFF(ebx,PD_DESCALE_P2)]
michael@0 560 psrad mm4,DESCALE_P2
michael@0 561 psrad mm1,DESCALE_P2
michael@0 562 paddd mm2,[GOTOFF(ebx,PD_DESCALE_P2)]
michael@0 563 paddd mm7,[GOTOFF(ebx,PD_DESCALE_P2)]
michael@0 564 psrad mm2,DESCALE_P2
michael@0 565 psrad mm7,DESCALE_P2
michael@0 566
michael@0 567 packssdw mm4,mm1 ; mm4=data7
michael@0 568 packssdw mm2,mm7 ; mm2=data1
michael@0 569
michael@0 570 movq MMWORD [MMBLOCK(7,0,edx,SIZEOF_DCTELEM)], mm4
michael@0 571 movq MMWORD [MMBLOCK(1,0,edx,SIZEOF_DCTELEM)], mm2
michael@0 572
michael@0 573 movq mm1,mm3
michael@0 574 movq mm7,mm3
michael@0 575 punpcklwd mm1,mm5
michael@0 576 punpckhwd mm7,mm5
michael@0 577 movq mm3,mm1
michael@0 578 movq mm5,mm7
michael@0 579 pmaddwd mm1,[GOTOFF(ebx,PW_MF050_MF256)] ; mm1=tmp5L
michael@0 580 pmaddwd mm7,[GOTOFF(ebx,PW_MF050_MF256)] ; mm7=tmp5H
michael@0 581 pmaddwd mm3,[GOTOFF(ebx,PW_MF256_F050)] ; mm3=tmp6L
michael@0 582 pmaddwd mm5,[GOTOFF(ebx,PW_MF256_F050)] ; mm5=tmp6H
michael@0 583
michael@0 584 paddd mm1,mm0 ; mm1=data5L
michael@0 585 paddd mm7,mm6 ; mm7=data5H
michael@0 586 paddd mm3, MMWORD [wk(0)] ; mm3=data3L
michael@0 587 paddd mm5, MMWORD [wk(1)] ; mm5=data3H
michael@0 588
michael@0 589 paddd mm1,[GOTOFF(ebx,PD_DESCALE_P2)]
michael@0 590 paddd mm7,[GOTOFF(ebx,PD_DESCALE_P2)]
michael@0 591 psrad mm1,DESCALE_P2
michael@0 592 psrad mm7,DESCALE_P2
michael@0 593 paddd mm3,[GOTOFF(ebx,PD_DESCALE_P2)]
michael@0 594 paddd mm5,[GOTOFF(ebx,PD_DESCALE_P2)]
michael@0 595 psrad mm3,DESCALE_P2
michael@0 596 psrad mm5,DESCALE_P2
michael@0 597
michael@0 598 packssdw mm1,mm7 ; mm1=data5
michael@0 599 packssdw mm3,mm5 ; mm3=data3
michael@0 600
michael@0 601 movq MMWORD [MMBLOCK(5,0,edx,SIZEOF_DCTELEM)], mm1
michael@0 602 movq MMWORD [MMBLOCK(3,0,edx,SIZEOF_DCTELEM)], mm3
michael@0 603
michael@0 604 add edx, byte 4*SIZEOF_DCTELEM
michael@0 605 dec ecx
michael@0 606 jnz near .columnloop
michael@0 607
michael@0 608 emms ; empty MMX state
michael@0 609
michael@0 610 ; pop edi ; unused
michael@0 611 ; pop esi ; unused
michael@0 612 ; pop edx ; need not be preserved
michael@0 613 ; pop ecx ; need not be preserved
michael@0 614 poppic ebx
michael@0 615 mov esp,ebp ; esp <- aligned ebp
michael@0 616 pop esp ; esp <- original ebp
michael@0 617 pop ebp
michael@0 618 ret
michael@0 619
michael@0 620 ; For some reason, the OS X linker does not honor the request to align the
michael@0 621 ; segment unless we do this.
michael@0 622 align 16

Mercurial Repository: The Tor Browser

Europalab SCM Repository Server

mercurial