The Tor Browser / annotate

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

media/libjpeg/simd/jimmxint.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 ; jimmxint.asm - accurate integer IDCT (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 ; inverse DCT (Discrete Cosine Transform). The following code is based
michael@0 19 ; directly on the IJG's original jidctint.c; see the jidctint.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+3)
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_idct_islow_mmx)
michael@0 70
michael@0 71 EXTN(jconst_idct_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 PB_CENTERJSAMP times 8 db CENTERJSAMPLE
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 dequantization and inverse DCT on one block of coefficients.
michael@0 92 ;
michael@0 93 ; GLOBAL(void)
michael@0 94 ; jsimd_idct_islow_mmx (void * dct_table, JCOEFPTR coef_block,
michael@0 95 ; JSAMPARRAY output_buf, JDIMENSION output_col)
michael@0 96 ;
michael@0 97
michael@0 98 %define dct_table(b) (b)+8 ; jpeg_component_info * compptr
michael@0 99 %define coef_block(b) (b)+12 ; JCOEFPTR coef_block
michael@0 100 %define output_buf(b) (b)+16 ; JSAMPARRAY output_buf
michael@0 101 %define output_col(b) (b)+20 ; JDIMENSION output_col
michael@0 102
michael@0 103 %define original_ebp ebp+0
michael@0 104 %define wk(i) ebp-(WK_NUM-(i))*SIZEOF_MMWORD ; mmword wk[WK_NUM]
michael@0 105 %define WK_NUM 12
michael@0 106 %define workspace wk(0)-DCTSIZE2*SIZEOF_JCOEF
michael@0 107 ; JCOEF workspace[DCTSIZE2]
michael@0 108
michael@0 109 align 16
michael@0 110 global EXTN(jsimd_idct_islow_mmx)
michael@0 111
michael@0 112 EXTN(jsimd_idct_islow_mmx):
michael@0 113 push ebp
michael@0 114 mov eax,esp ; eax = original ebp
michael@0 115 sub esp, byte 4
michael@0 116 and esp, byte (-SIZEOF_MMWORD) ; align to 64 bits
michael@0 117 mov [esp],eax
michael@0 118 mov ebp,esp ; ebp = aligned ebp
michael@0 119 lea esp, [workspace]
michael@0 120 push ebx
michael@0 121 ; push ecx ; need not be preserved
michael@0 122 ; push edx ; need not be preserved
michael@0 123 push esi
michael@0 124 push edi
michael@0 125
michael@0 126 get_GOT ebx ; get GOT address
michael@0 127
michael@0 128 ; ---- Pass 1: process columns from input, store into work array.
michael@0 129
michael@0 130 ; mov eax, [original_ebp]
michael@0 131 mov edx, POINTER [dct_table(eax)] ; quantptr
michael@0 132 mov esi, JCOEFPTR [coef_block(eax)] ; inptr
michael@0 133 lea edi, [workspace] ; JCOEF * wsptr
michael@0 134 mov ecx, DCTSIZE/4 ; ctr
michael@0 135 alignx 16,7
michael@0 136 .columnloop:
michael@0 137 %ifndef NO_ZERO_COLUMN_TEST_ISLOW_MMX
michael@0 138 mov eax, DWORD [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
michael@0 139 or eax, DWORD [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
michael@0 140 jnz short .columnDCT
michael@0 141
michael@0 142 movq mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
michael@0 143 movq mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
michael@0 144 por mm0, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
michael@0 145 por mm1, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
michael@0 146 por mm0, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
michael@0 147 por mm1, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
michael@0 148 por mm0, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
michael@0 149 por mm1,mm0
michael@0 150 packsswb mm1,mm1
michael@0 151 movd eax,mm1
michael@0 152 test eax,eax
michael@0 153 jnz short .columnDCT
michael@0 154
michael@0 155 ; -- AC terms all zero
michael@0 156
michael@0 157 movq mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
michael@0 158 pmullw mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 159
michael@0 160 psllw mm0,PASS1_BITS
michael@0 161
michael@0 162 movq mm2,mm0 ; mm0=in0=(00 01 02 03)
michael@0 163 punpcklwd mm0,mm0 ; mm0=(00 00 01 01)
michael@0 164 punpckhwd mm2,mm2 ; mm2=(02 02 03 03)
michael@0 165
michael@0 166 movq mm1,mm0
michael@0 167 punpckldq mm0,mm0 ; mm0=(00 00 00 00)
michael@0 168 punpckhdq mm1,mm1 ; mm1=(01 01 01 01)
michael@0 169 movq mm3,mm2
michael@0 170 punpckldq mm2,mm2 ; mm2=(02 02 02 02)
michael@0 171 punpckhdq mm3,mm3 ; mm3=(03 03 03 03)
michael@0 172
michael@0 173 movq MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm0
michael@0 174 movq MMWORD [MMBLOCK(0,1,edi,SIZEOF_JCOEF)], mm0
michael@0 175 movq MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm1
michael@0 176 movq MMWORD [MMBLOCK(1,1,edi,SIZEOF_JCOEF)], mm1
michael@0 177 movq MMWORD [MMBLOCK(2,0,edi,SIZEOF_JCOEF)], mm2
michael@0 178 movq MMWORD [MMBLOCK(2,1,edi,SIZEOF_JCOEF)], mm2
michael@0 179 movq MMWORD [MMBLOCK(3,0,edi,SIZEOF_JCOEF)], mm3
michael@0 180 movq MMWORD [MMBLOCK(3,1,edi,SIZEOF_JCOEF)], mm3
michael@0 181 jmp near .nextcolumn
michael@0 182 alignx 16,7
michael@0 183 %endif
michael@0 184 .columnDCT:
michael@0 185
michael@0 186 ; -- Even part
michael@0 187
michael@0 188 movq mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
michael@0 189 movq mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
michael@0 190 pmullw mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 191 pmullw mm1, MMWORD [MMBLOCK(2,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 192 movq mm2, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
michael@0 193 movq mm3, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
michael@0 194 pmullw mm2, MMWORD [MMBLOCK(4,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 195 pmullw mm3, MMWORD [MMBLOCK(6,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 196
michael@0 197 ; (Original)
michael@0 198 ; z1 = (z2 + z3) * 0.541196100;
michael@0 199 ; tmp2 = z1 + z3 * -1.847759065;
michael@0 200 ; tmp3 = z1 + z2 * 0.765366865;
michael@0 201 ;
michael@0 202 ; (This implementation)
michael@0 203 ; tmp2 = z2 * 0.541196100 + z3 * (0.541196100 - 1.847759065);
michael@0 204 ; tmp3 = z2 * (0.541196100 + 0.765366865) + z3 * 0.541196100;
michael@0 205
michael@0 206 movq mm4,mm1 ; mm1=in2=z2
michael@0 207 movq mm5,mm1
michael@0 208 punpcklwd mm4,mm3 ; mm3=in6=z3
michael@0 209 punpckhwd mm5,mm3
michael@0 210 movq mm1,mm4
michael@0 211 movq mm3,mm5
michael@0 212 pmaddwd mm4,[GOTOFF(ebx,PW_F130_F054)] ; mm4=tmp3L
michael@0 213 pmaddwd mm5,[GOTOFF(ebx,PW_F130_F054)] ; mm5=tmp3H
michael@0 214 pmaddwd mm1,[GOTOFF(ebx,PW_F054_MF130)] ; mm1=tmp2L
michael@0 215 pmaddwd mm3,[GOTOFF(ebx,PW_F054_MF130)] ; mm3=tmp2H
michael@0 216
michael@0 217 movq mm6,mm0
michael@0 218 paddw mm0,mm2 ; mm0=in0+in4
michael@0 219 psubw mm6,mm2 ; mm6=in0-in4
michael@0 220
michael@0 221 pxor mm7,mm7
michael@0 222 pxor mm2,mm2
michael@0 223 punpcklwd mm7,mm0 ; mm7=tmp0L
michael@0 224 punpckhwd mm2,mm0 ; mm2=tmp0H
michael@0 225 psrad mm7,(16-CONST_BITS) ; psrad mm7,16 & pslld mm7,CONST_BITS
michael@0 226 psrad mm2,(16-CONST_BITS) ; psrad mm2,16 & pslld mm2,CONST_BITS
michael@0 227
michael@0 228 movq mm0,mm7
michael@0 229 paddd mm7,mm4 ; mm7=tmp10L
michael@0 230 psubd mm0,mm4 ; mm0=tmp13L
michael@0 231 movq mm4,mm2
michael@0 232 paddd mm2,mm5 ; mm2=tmp10H
michael@0 233 psubd mm4,mm5 ; mm4=tmp13H
michael@0 234
michael@0 235 movq MMWORD [wk(0)], mm7 ; wk(0)=tmp10L
michael@0 236 movq MMWORD [wk(1)], mm2 ; wk(1)=tmp10H
michael@0 237 movq MMWORD [wk(2)], mm0 ; wk(2)=tmp13L
michael@0 238 movq MMWORD [wk(3)], mm4 ; wk(3)=tmp13H
michael@0 239
michael@0 240 pxor mm5,mm5
michael@0 241 pxor mm7,mm7
michael@0 242 punpcklwd mm5,mm6 ; mm5=tmp1L
michael@0 243 punpckhwd mm7,mm6 ; mm7=tmp1H
michael@0 244 psrad mm5,(16-CONST_BITS) ; psrad mm5,16 & pslld mm5,CONST_BITS
michael@0 245 psrad mm7,(16-CONST_BITS) ; psrad mm7,16 & pslld mm7,CONST_BITS
michael@0 246
michael@0 247 movq mm2,mm5
michael@0 248 paddd mm5,mm1 ; mm5=tmp11L
michael@0 249 psubd mm2,mm1 ; mm2=tmp12L
michael@0 250 movq mm0,mm7
michael@0 251 paddd mm7,mm3 ; mm7=tmp11H
michael@0 252 psubd mm0,mm3 ; mm0=tmp12H
michael@0 253
michael@0 254 movq MMWORD [wk(4)], mm5 ; wk(4)=tmp11L
michael@0 255 movq MMWORD [wk(5)], mm7 ; wk(5)=tmp11H
michael@0 256 movq MMWORD [wk(6)], mm2 ; wk(6)=tmp12L
michael@0 257 movq MMWORD [wk(7)], mm0 ; wk(7)=tmp12H
michael@0 258
michael@0 259 ; -- Odd part
michael@0 260
michael@0 261 movq mm4, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
michael@0 262 movq mm6, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
michael@0 263 pmullw mm4, MMWORD [MMBLOCK(1,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 264 pmullw mm6, MMWORD [MMBLOCK(3,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 265 movq mm1, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
michael@0 266 movq mm3, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
michael@0 267 pmullw mm1, MMWORD [MMBLOCK(5,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 268 pmullw mm3, MMWORD [MMBLOCK(7,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 269
michael@0 270 movq mm5,mm6
michael@0 271 movq mm7,mm4
michael@0 272 paddw mm5,mm3 ; mm5=z3
michael@0 273 paddw mm7,mm1 ; mm7=z4
michael@0 274
michael@0 275 ; (Original)
michael@0 276 ; z5 = (z3 + z4) * 1.175875602;
michael@0 277 ; z3 = z3 * -1.961570560; z4 = z4 * -0.390180644;
michael@0 278 ; z3 += z5; z4 += z5;
michael@0 279 ;
michael@0 280 ; (This implementation)
michael@0 281 ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
michael@0 282 ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
michael@0 283
michael@0 284 movq mm2,mm5
michael@0 285 movq mm0,mm5
michael@0 286 punpcklwd mm2,mm7
michael@0 287 punpckhwd mm0,mm7
michael@0 288 movq mm5,mm2
michael@0 289 movq mm7,mm0
michael@0 290 pmaddwd mm2,[GOTOFF(ebx,PW_MF078_F117)] ; mm2=z3L
michael@0 291 pmaddwd mm0,[GOTOFF(ebx,PW_MF078_F117)] ; mm0=z3H
michael@0 292 pmaddwd mm5,[GOTOFF(ebx,PW_F117_F078)] ; mm5=z4L
michael@0 293 pmaddwd mm7,[GOTOFF(ebx,PW_F117_F078)] ; mm7=z4H
michael@0 294
michael@0 295 movq MMWORD [wk(10)], mm2 ; wk(10)=z3L
michael@0 296 movq MMWORD [wk(11)], mm0 ; wk(11)=z3H
michael@0 297
michael@0 298 ; (Original)
michael@0 299 ; z1 = tmp0 + tmp3; z2 = tmp1 + tmp2;
michael@0 300 ; tmp0 = tmp0 * 0.298631336; tmp1 = tmp1 * 2.053119869;
michael@0 301 ; tmp2 = tmp2 * 3.072711026; tmp3 = tmp3 * 1.501321110;
michael@0 302 ; z1 = z1 * -0.899976223; z2 = z2 * -2.562915447;
michael@0 303 ; tmp0 += z1 + z3; tmp1 += z2 + z4;
michael@0 304 ; tmp2 += z2 + z3; tmp3 += z1 + z4;
michael@0 305 ;
michael@0 306 ; (This implementation)
michael@0 307 ; tmp0 = tmp0 * (0.298631336 - 0.899976223) + tmp3 * -0.899976223;
michael@0 308 ; tmp1 = tmp1 * (2.053119869 - 2.562915447) + tmp2 * -2.562915447;
michael@0 309 ; tmp2 = tmp1 * -2.562915447 + tmp2 * (3.072711026 - 2.562915447);
michael@0 310 ; tmp3 = tmp0 * -0.899976223 + tmp3 * (1.501321110 - 0.899976223);
michael@0 311 ; tmp0 += z3; tmp1 += z4;
michael@0 312 ; tmp2 += z3; tmp3 += z4;
michael@0 313
michael@0 314 movq mm2,mm3
michael@0 315 movq mm0,mm3
michael@0 316 punpcklwd mm2,mm4
michael@0 317 punpckhwd mm0,mm4
michael@0 318 movq mm3,mm2
michael@0 319 movq mm4,mm0
michael@0 320 pmaddwd mm2,[GOTOFF(ebx,PW_MF060_MF089)] ; mm2=tmp0L
michael@0 321 pmaddwd mm0,[GOTOFF(ebx,PW_MF060_MF089)] ; mm0=tmp0H
michael@0 322 pmaddwd mm3,[GOTOFF(ebx,PW_MF089_F060)] ; mm3=tmp3L
michael@0 323 pmaddwd mm4,[GOTOFF(ebx,PW_MF089_F060)] ; mm4=tmp3H
michael@0 324
michael@0 325 paddd mm2, MMWORD [wk(10)] ; mm2=tmp0L
michael@0 326 paddd mm0, MMWORD [wk(11)] ; mm0=tmp0H
michael@0 327 paddd mm3,mm5 ; mm3=tmp3L
michael@0 328 paddd mm4,mm7 ; mm4=tmp3H
michael@0 329
michael@0 330 movq MMWORD [wk(8)], mm2 ; wk(8)=tmp0L
michael@0 331 movq MMWORD [wk(9)], mm0 ; wk(9)=tmp0H
michael@0 332
michael@0 333 movq mm2,mm1
michael@0 334 movq mm0,mm1
michael@0 335 punpcklwd mm2,mm6
michael@0 336 punpckhwd mm0,mm6
michael@0 337 movq mm1,mm2
michael@0 338 movq mm6,mm0
michael@0 339 pmaddwd mm2,[GOTOFF(ebx,PW_MF050_MF256)] ; mm2=tmp1L
michael@0 340 pmaddwd mm0,[GOTOFF(ebx,PW_MF050_MF256)] ; mm0=tmp1H
michael@0 341 pmaddwd mm1,[GOTOFF(ebx,PW_MF256_F050)] ; mm1=tmp2L
michael@0 342 pmaddwd mm6,[GOTOFF(ebx,PW_MF256_F050)] ; mm6=tmp2H
michael@0 343
michael@0 344 paddd mm2,mm5 ; mm2=tmp1L
michael@0 345 paddd mm0,mm7 ; mm0=tmp1H
michael@0 346 paddd mm1, MMWORD [wk(10)] ; mm1=tmp2L
michael@0 347 paddd mm6, MMWORD [wk(11)] ; mm6=tmp2H
michael@0 348
michael@0 349 movq MMWORD [wk(10)], mm2 ; wk(10)=tmp1L
michael@0 350 movq MMWORD [wk(11)], mm0 ; wk(11)=tmp1H
michael@0 351
michael@0 352 ; -- Final output stage
michael@0 353
michael@0 354 movq mm5, MMWORD [wk(0)] ; mm5=tmp10L
michael@0 355 movq mm7, MMWORD [wk(1)] ; mm7=tmp10H
michael@0 356
michael@0 357 movq mm2,mm5
michael@0 358 movq mm0,mm7
michael@0 359 paddd mm5,mm3 ; mm5=data0L
michael@0 360 paddd mm7,mm4 ; mm7=data0H
michael@0 361 psubd mm2,mm3 ; mm2=data7L
michael@0 362 psubd mm0,mm4 ; mm0=data7H
michael@0 363
michael@0 364 movq mm3,[GOTOFF(ebx,PD_DESCALE_P1)] ; mm3=[PD_DESCALE_P1]
michael@0 365
michael@0 366 paddd mm5,mm3
michael@0 367 paddd mm7,mm3
michael@0 368 psrad mm5,DESCALE_P1
michael@0 369 psrad mm7,DESCALE_P1
michael@0 370 paddd mm2,mm3
michael@0 371 paddd mm0,mm3
michael@0 372 psrad mm2,DESCALE_P1
michael@0 373 psrad mm0,DESCALE_P1
michael@0 374
michael@0 375 packssdw mm5,mm7 ; mm5=data0=(00 01 02 03)
michael@0 376 packssdw mm2,mm0 ; mm2=data7=(70 71 72 73)
michael@0 377
michael@0 378 movq mm4, MMWORD [wk(4)] ; mm4=tmp11L
michael@0 379 movq mm3, MMWORD [wk(5)] ; mm3=tmp11H
michael@0 380
michael@0 381 movq mm7,mm4
michael@0 382 movq mm0,mm3
michael@0 383 paddd mm4,mm1 ; mm4=data1L
michael@0 384 paddd mm3,mm6 ; mm3=data1H
michael@0 385 psubd mm7,mm1 ; mm7=data6L
michael@0 386 psubd mm0,mm6 ; mm0=data6H
michael@0 387
michael@0 388 movq mm1,[GOTOFF(ebx,PD_DESCALE_P1)] ; mm1=[PD_DESCALE_P1]
michael@0 389
michael@0 390 paddd mm4,mm1
michael@0 391 paddd mm3,mm1
michael@0 392 psrad mm4,DESCALE_P1
michael@0 393 psrad mm3,DESCALE_P1
michael@0 394 paddd mm7,mm1
michael@0 395 paddd mm0,mm1
michael@0 396 psrad mm7,DESCALE_P1
michael@0 397 psrad mm0,DESCALE_P1
michael@0 398
michael@0 399 packssdw mm4,mm3 ; mm4=data1=(10 11 12 13)
michael@0 400 packssdw mm7,mm0 ; mm7=data6=(60 61 62 63)
michael@0 401
michael@0 402 movq mm6,mm5 ; transpose coefficients(phase 1)
michael@0 403 punpcklwd mm5,mm4 ; mm5=(00 10 01 11)
michael@0 404 punpckhwd mm6,mm4 ; mm6=(02 12 03 13)
michael@0 405 movq mm1,mm7 ; transpose coefficients(phase 1)
michael@0 406 punpcklwd mm7,mm2 ; mm7=(60 70 61 71)
michael@0 407 punpckhwd mm1,mm2 ; mm1=(62 72 63 73)
michael@0 408
michael@0 409 movq mm3, MMWORD [wk(6)] ; mm3=tmp12L
michael@0 410 movq mm0, MMWORD [wk(7)] ; mm0=tmp12H
michael@0 411 movq mm4, MMWORD [wk(10)] ; mm4=tmp1L
michael@0 412 movq mm2, MMWORD [wk(11)] ; mm2=tmp1H
michael@0 413
michael@0 414 movq MMWORD [wk(0)], mm5 ; wk(0)=(00 10 01 11)
michael@0 415 movq MMWORD [wk(1)], mm6 ; wk(1)=(02 12 03 13)
michael@0 416 movq MMWORD [wk(4)], mm7 ; wk(4)=(60 70 61 71)
michael@0 417 movq MMWORD [wk(5)], mm1 ; wk(5)=(62 72 63 73)
michael@0 418
michael@0 419 movq mm5,mm3
michael@0 420 movq mm6,mm0
michael@0 421 paddd mm3,mm4 ; mm3=data2L
michael@0 422 paddd mm0,mm2 ; mm0=data2H
michael@0 423 psubd mm5,mm4 ; mm5=data5L
michael@0 424 psubd mm6,mm2 ; mm6=data5H
michael@0 425
michael@0 426 movq mm7,[GOTOFF(ebx,PD_DESCALE_P1)] ; mm7=[PD_DESCALE_P1]
michael@0 427
michael@0 428 paddd mm3,mm7
michael@0 429 paddd mm0,mm7
michael@0 430 psrad mm3,DESCALE_P1
michael@0 431 psrad mm0,DESCALE_P1
michael@0 432 paddd mm5,mm7
michael@0 433 paddd mm6,mm7
michael@0 434 psrad mm5,DESCALE_P1
michael@0 435 psrad mm6,DESCALE_P1
michael@0 436
michael@0 437 packssdw mm3,mm0 ; mm3=data2=(20 21 22 23)
michael@0 438 packssdw mm5,mm6 ; mm5=data5=(50 51 52 53)
michael@0 439
michael@0 440 movq mm1, MMWORD [wk(2)] ; mm1=tmp13L
michael@0 441 movq mm4, MMWORD [wk(3)] ; mm4=tmp13H
michael@0 442 movq mm2, MMWORD [wk(8)] ; mm2=tmp0L
michael@0 443 movq mm7, MMWORD [wk(9)] ; mm7=tmp0H
michael@0 444
michael@0 445 movq mm0,mm1
michael@0 446 movq mm6,mm4
michael@0 447 paddd mm1,mm2 ; mm1=data3L
michael@0 448 paddd mm4,mm7 ; mm4=data3H
michael@0 449 psubd mm0,mm2 ; mm0=data4L
michael@0 450 psubd mm6,mm7 ; mm6=data4H
michael@0 451
michael@0 452 movq mm2,[GOTOFF(ebx,PD_DESCALE_P1)] ; mm2=[PD_DESCALE_P1]
michael@0 453
michael@0 454 paddd mm1,mm2
michael@0 455 paddd mm4,mm2
michael@0 456 psrad mm1,DESCALE_P1
michael@0 457 psrad mm4,DESCALE_P1
michael@0 458 paddd mm0,mm2
michael@0 459 paddd mm6,mm2
michael@0 460 psrad mm0,DESCALE_P1
michael@0 461 psrad mm6,DESCALE_P1
michael@0 462
michael@0 463 packssdw mm1,mm4 ; mm1=data3=(30 31 32 33)
michael@0 464 packssdw mm0,mm6 ; mm0=data4=(40 41 42 43)
michael@0 465
michael@0 466 movq mm7, MMWORD [wk(0)] ; mm7=(00 10 01 11)
michael@0 467 movq mm2, MMWORD [wk(1)] ; mm2=(02 12 03 13)
michael@0 468
michael@0 469 movq mm4,mm3 ; transpose coefficients(phase 1)
michael@0 470 punpcklwd mm3,mm1 ; mm3=(20 30 21 31)
michael@0 471 punpckhwd mm4,mm1 ; mm4=(22 32 23 33)
michael@0 472 movq mm6,mm0 ; transpose coefficients(phase 1)
michael@0 473 punpcklwd mm0,mm5 ; mm0=(40 50 41 51)
michael@0 474 punpckhwd mm6,mm5 ; mm6=(42 52 43 53)
michael@0 475
michael@0 476 movq mm1,mm7 ; transpose coefficients(phase 2)
michael@0 477 punpckldq mm7,mm3 ; mm7=(00 10 20 30)
michael@0 478 punpckhdq mm1,mm3 ; mm1=(01 11 21 31)
michael@0 479 movq mm5,mm2 ; transpose coefficients(phase 2)
michael@0 480 punpckldq mm2,mm4 ; mm2=(02 12 22 32)
michael@0 481 punpckhdq mm5,mm4 ; mm5=(03 13 23 33)
michael@0 482
michael@0 483 movq mm3, MMWORD [wk(4)] ; mm3=(60 70 61 71)
michael@0 484 movq mm4, MMWORD [wk(5)] ; mm4=(62 72 63 73)
michael@0 485
michael@0 486 movq MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm7
michael@0 487 movq MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm1
michael@0 488 movq MMWORD [MMBLOCK(2,0,edi,SIZEOF_JCOEF)], mm2
michael@0 489 movq MMWORD [MMBLOCK(3,0,edi,SIZEOF_JCOEF)], mm5
michael@0 490
michael@0 491 movq mm7,mm0 ; transpose coefficients(phase 2)
michael@0 492 punpckldq mm0,mm3 ; mm0=(40 50 60 70)
michael@0 493 punpckhdq mm7,mm3 ; mm7=(41 51 61 71)
michael@0 494 movq mm1,mm6 ; transpose coefficients(phase 2)
michael@0 495 punpckldq mm6,mm4 ; mm6=(42 52 62 72)
michael@0 496 punpckhdq mm1,mm4 ; mm1=(43 53 63 73)
michael@0 497
michael@0 498 movq MMWORD [MMBLOCK(0,1,edi,SIZEOF_JCOEF)], mm0
michael@0 499 movq MMWORD [MMBLOCK(1,1,edi,SIZEOF_JCOEF)], mm7
michael@0 500 movq MMWORD [MMBLOCK(2,1,edi,SIZEOF_JCOEF)], mm6
michael@0 501 movq MMWORD [MMBLOCK(3,1,edi,SIZEOF_JCOEF)], mm1
michael@0 502
michael@0 503 .nextcolumn:
michael@0 504 add esi, byte 4*SIZEOF_JCOEF ; coef_block
michael@0 505 add edx, byte 4*SIZEOF_ISLOW_MULT_TYPE ; quantptr
michael@0 506 add edi, byte 4*DCTSIZE*SIZEOF_JCOEF ; wsptr
michael@0 507 dec ecx ; ctr
michael@0 508 jnz near .columnloop
michael@0 509
michael@0 510 ; ---- Pass 2: process rows from work array, store into output array.
michael@0 511
michael@0 512 mov eax, [original_ebp]
michael@0 513 lea esi, [workspace] ; JCOEF * wsptr
michael@0 514 mov edi, JSAMPARRAY [output_buf(eax)] ; (JSAMPROW *)
michael@0 515 mov eax, JDIMENSION [output_col(eax)]
michael@0 516 mov ecx, DCTSIZE/4 ; ctr
michael@0 517 alignx 16,7
michael@0 518 .rowloop:
michael@0 519
michael@0 520 ; -- Even part
michael@0 521
michael@0 522 movq mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
michael@0 523 movq mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
michael@0 524 movq mm2, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
michael@0 525 movq mm3, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
michael@0 526
michael@0 527 ; (Original)
michael@0 528 ; z1 = (z2 + z3) * 0.541196100;
michael@0 529 ; tmp2 = z1 + z3 * -1.847759065;
michael@0 530 ; tmp3 = z1 + z2 * 0.765366865;
michael@0 531 ;
michael@0 532 ; (This implementation)
michael@0 533 ; tmp2 = z2 * 0.541196100 + z3 * (0.541196100 - 1.847759065);
michael@0 534 ; tmp3 = z2 * (0.541196100 + 0.765366865) + z3 * 0.541196100;
michael@0 535
michael@0 536 movq mm4,mm1 ; mm1=in2=z2
michael@0 537 movq mm5,mm1
michael@0 538 punpcklwd mm4,mm3 ; mm3=in6=z3
michael@0 539 punpckhwd mm5,mm3
michael@0 540 movq mm1,mm4
michael@0 541 movq mm3,mm5
michael@0 542 pmaddwd mm4,[GOTOFF(ebx,PW_F130_F054)] ; mm4=tmp3L
michael@0 543 pmaddwd mm5,[GOTOFF(ebx,PW_F130_F054)] ; mm5=tmp3H
michael@0 544 pmaddwd mm1,[GOTOFF(ebx,PW_F054_MF130)] ; mm1=tmp2L
michael@0 545 pmaddwd mm3,[GOTOFF(ebx,PW_F054_MF130)] ; mm3=tmp2H
michael@0 546
michael@0 547 movq mm6,mm0
michael@0 548 paddw mm0,mm2 ; mm0=in0+in4
michael@0 549 psubw mm6,mm2 ; mm6=in0-in4
michael@0 550
michael@0 551 pxor mm7,mm7
michael@0 552 pxor mm2,mm2
michael@0 553 punpcklwd mm7,mm0 ; mm7=tmp0L
michael@0 554 punpckhwd mm2,mm0 ; mm2=tmp0H
michael@0 555 psrad mm7,(16-CONST_BITS) ; psrad mm7,16 & pslld mm7,CONST_BITS
michael@0 556 psrad mm2,(16-CONST_BITS) ; psrad mm2,16 & pslld mm2,CONST_BITS
michael@0 557
michael@0 558 movq mm0,mm7
michael@0 559 paddd mm7,mm4 ; mm7=tmp10L
michael@0 560 psubd mm0,mm4 ; mm0=tmp13L
michael@0 561 movq mm4,mm2
michael@0 562 paddd mm2,mm5 ; mm2=tmp10H
michael@0 563 psubd mm4,mm5 ; mm4=tmp13H
michael@0 564
michael@0 565 movq MMWORD [wk(0)], mm7 ; wk(0)=tmp10L
michael@0 566 movq MMWORD [wk(1)], mm2 ; wk(1)=tmp10H
michael@0 567 movq MMWORD [wk(2)], mm0 ; wk(2)=tmp13L
michael@0 568 movq MMWORD [wk(3)], mm4 ; wk(3)=tmp13H
michael@0 569
michael@0 570 pxor mm5,mm5
michael@0 571 pxor mm7,mm7
michael@0 572 punpcklwd mm5,mm6 ; mm5=tmp1L
michael@0 573 punpckhwd mm7,mm6 ; mm7=tmp1H
michael@0 574 psrad mm5,(16-CONST_BITS) ; psrad mm5,16 & pslld mm5,CONST_BITS
michael@0 575 psrad mm7,(16-CONST_BITS) ; psrad mm7,16 & pslld mm7,CONST_BITS
michael@0 576
michael@0 577 movq mm2,mm5
michael@0 578 paddd mm5,mm1 ; mm5=tmp11L
michael@0 579 psubd mm2,mm1 ; mm2=tmp12L
michael@0 580 movq mm0,mm7
michael@0 581 paddd mm7,mm3 ; mm7=tmp11H
michael@0 582 psubd mm0,mm3 ; mm0=tmp12H
michael@0 583
michael@0 584 movq MMWORD [wk(4)], mm5 ; wk(4)=tmp11L
michael@0 585 movq MMWORD [wk(5)], mm7 ; wk(5)=tmp11H
michael@0 586 movq MMWORD [wk(6)], mm2 ; wk(6)=tmp12L
michael@0 587 movq MMWORD [wk(7)], mm0 ; wk(7)=tmp12H
michael@0 588
michael@0 589 ; -- Odd part
michael@0 590
michael@0 591 movq mm4, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
michael@0 592 movq mm6, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
michael@0 593 movq mm1, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
michael@0 594 movq mm3, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
michael@0 595
michael@0 596 movq mm5,mm6
michael@0 597 movq mm7,mm4
michael@0 598 paddw mm5,mm3 ; mm5=z3
michael@0 599 paddw mm7,mm1 ; mm7=z4
michael@0 600
michael@0 601 ; (Original)
michael@0 602 ; z5 = (z3 + z4) * 1.175875602;
michael@0 603 ; z3 = z3 * -1.961570560; z4 = z4 * -0.390180644;
michael@0 604 ; z3 += z5; z4 += z5;
michael@0 605 ;
michael@0 606 ; (This implementation)
michael@0 607 ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
michael@0 608 ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
michael@0 609
michael@0 610 movq mm2,mm5
michael@0 611 movq mm0,mm5
michael@0 612 punpcklwd mm2,mm7
michael@0 613 punpckhwd mm0,mm7
michael@0 614 movq mm5,mm2
michael@0 615 movq mm7,mm0
michael@0 616 pmaddwd mm2,[GOTOFF(ebx,PW_MF078_F117)] ; mm2=z3L
michael@0 617 pmaddwd mm0,[GOTOFF(ebx,PW_MF078_F117)] ; mm0=z3H
michael@0 618 pmaddwd mm5,[GOTOFF(ebx,PW_F117_F078)] ; mm5=z4L
michael@0 619 pmaddwd mm7,[GOTOFF(ebx,PW_F117_F078)] ; mm7=z4H
michael@0 620
michael@0 621 movq MMWORD [wk(10)], mm2 ; wk(10)=z3L
michael@0 622 movq MMWORD [wk(11)], mm0 ; wk(11)=z3H
michael@0 623
michael@0 624 ; (Original)
michael@0 625 ; z1 = tmp0 + tmp3; z2 = tmp1 + tmp2;
michael@0 626 ; tmp0 = tmp0 * 0.298631336; tmp1 = tmp1 * 2.053119869;
michael@0 627 ; tmp2 = tmp2 * 3.072711026; tmp3 = tmp3 * 1.501321110;
michael@0 628 ; z1 = z1 * -0.899976223; z2 = z2 * -2.562915447;
michael@0 629 ; tmp0 += z1 + z3; tmp1 += z2 + z4;
michael@0 630 ; tmp2 += z2 + z3; tmp3 += z1 + z4;
michael@0 631 ;
michael@0 632 ; (This implementation)
michael@0 633 ; tmp0 = tmp0 * (0.298631336 - 0.899976223) + tmp3 * -0.899976223;
michael@0 634 ; tmp1 = tmp1 * (2.053119869 - 2.562915447) + tmp2 * -2.562915447;
michael@0 635 ; tmp2 = tmp1 * -2.562915447 + tmp2 * (3.072711026 - 2.562915447);
michael@0 636 ; tmp3 = tmp0 * -0.899976223 + tmp3 * (1.501321110 - 0.899976223);
michael@0 637 ; tmp0 += z3; tmp1 += z4;
michael@0 638 ; tmp2 += z3; tmp3 += z4;
michael@0 639
michael@0 640 movq mm2,mm3
michael@0 641 movq mm0,mm3
michael@0 642 punpcklwd mm2,mm4
michael@0 643 punpckhwd mm0,mm4
michael@0 644 movq mm3,mm2
michael@0 645 movq mm4,mm0
michael@0 646 pmaddwd mm2,[GOTOFF(ebx,PW_MF060_MF089)] ; mm2=tmp0L
michael@0 647 pmaddwd mm0,[GOTOFF(ebx,PW_MF060_MF089)] ; mm0=tmp0H
michael@0 648 pmaddwd mm3,[GOTOFF(ebx,PW_MF089_F060)] ; mm3=tmp3L
michael@0 649 pmaddwd mm4,[GOTOFF(ebx,PW_MF089_F060)] ; mm4=tmp3H
michael@0 650
michael@0 651 paddd mm2, MMWORD [wk(10)] ; mm2=tmp0L
michael@0 652 paddd mm0, MMWORD [wk(11)] ; mm0=tmp0H
michael@0 653 paddd mm3,mm5 ; mm3=tmp3L
michael@0 654 paddd mm4,mm7 ; mm4=tmp3H
michael@0 655
michael@0 656 movq MMWORD [wk(8)], mm2 ; wk(8)=tmp0L
michael@0 657 movq MMWORD [wk(9)], mm0 ; wk(9)=tmp0H
michael@0 658
michael@0 659 movq mm2,mm1
michael@0 660 movq mm0,mm1
michael@0 661 punpcklwd mm2,mm6
michael@0 662 punpckhwd mm0,mm6
michael@0 663 movq mm1,mm2
michael@0 664 movq mm6,mm0
michael@0 665 pmaddwd mm2,[GOTOFF(ebx,PW_MF050_MF256)] ; mm2=tmp1L
michael@0 666 pmaddwd mm0,[GOTOFF(ebx,PW_MF050_MF256)] ; mm0=tmp1H
michael@0 667 pmaddwd mm1,[GOTOFF(ebx,PW_MF256_F050)] ; mm1=tmp2L
michael@0 668 pmaddwd mm6,[GOTOFF(ebx,PW_MF256_F050)] ; mm6=tmp2H
michael@0 669
michael@0 670 paddd mm2,mm5 ; mm2=tmp1L
michael@0 671 paddd mm0,mm7 ; mm0=tmp1H
michael@0 672 paddd mm1, MMWORD [wk(10)] ; mm1=tmp2L
michael@0 673 paddd mm6, MMWORD [wk(11)] ; mm6=tmp2H
michael@0 674
michael@0 675 movq MMWORD [wk(10)], mm2 ; wk(10)=tmp1L
michael@0 676 movq MMWORD [wk(11)], mm0 ; wk(11)=tmp1H
michael@0 677
michael@0 678 ; -- Final output stage
michael@0 679
michael@0 680 movq mm5, MMWORD [wk(0)] ; mm5=tmp10L
michael@0 681 movq mm7, MMWORD [wk(1)] ; mm7=tmp10H
michael@0 682
michael@0 683 movq mm2,mm5
michael@0 684 movq mm0,mm7
michael@0 685 paddd mm5,mm3 ; mm5=data0L
michael@0 686 paddd mm7,mm4 ; mm7=data0H
michael@0 687 psubd mm2,mm3 ; mm2=data7L
michael@0 688 psubd mm0,mm4 ; mm0=data7H
michael@0 689
michael@0 690 movq mm3,[GOTOFF(ebx,PD_DESCALE_P2)] ; mm3=[PD_DESCALE_P2]
michael@0 691
michael@0 692 paddd mm5,mm3
michael@0 693 paddd mm7,mm3
michael@0 694 psrad mm5,DESCALE_P2
michael@0 695 psrad mm7,DESCALE_P2
michael@0 696 paddd mm2,mm3
michael@0 697 paddd mm0,mm3
michael@0 698 psrad mm2,DESCALE_P2
michael@0 699 psrad mm0,DESCALE_P2
michael@0 700
michael@0 701 packssdw mm5,mm7 ; mm5=data0=(00 10 20 30)
michael@0 702 packssdw mm2,mm0 ; mm2=data7=(07 17 27 37)
michael@0 703
michael@0 704 movq mm4, MMWORD [wk(4)] ; mm4=tmp11L
michael@0 705 movq mm3, MMWORD [wk(5)] ; mm3=tmp11H
michael@0 706
michael@0 707 movq mm7,mm4
michael@0 708 movq mm0,mm3
michael@0 709 paddd mm4,mm1 ; mm4=data1L
michael@0 710 paddd mm3,mm6 ; mm3=data1H
michael@0 711 psubd mm7,mm1 ; mm7=data6L
michael@0 712 psubd mm0,mm6 ; mm0=data6H
michael@0 713
michael@0 714 movq mm1,[GOTOFF(ebx,PD_DESCALE_P2)] ; mm1=[PD_DESCALE_P2]
michael@0 715
michael@0 716 paddd mm4,mm1
michael@0 717 paddd mm3,mm1
michael@0 718 psrad mm4,DESCALE_P2
michael@0 719 psrad mm3,DESCALE_P2
michael@0 720 paddd mm7,mm1
michael@0 721 paddd mm0,mm1
michael@0 722 psrad mm7,DESCALE_P2
michael@0 723 psrad mm0,DESCALE_P2
michael@0 724
michael@0 725 packssdw mm4,mm3 ; mm4=data1=(01 11 21 31)
michael@0 726 packssdw mm7,mm0 ; mm7=data6=(06 16 26 36)
michael@0 727
michael@0 728 packsswb mm5,mm7 ; mm5=(00 10 20 30 06 16 26 36)
michael@0 729 packsswb mm4,mm2 ; mm4=(01 11 21 31 07 17 27 37)
michael@0 730
michael@0 731 movq mm6, MMWORD [wk(6)] ; mm6=tmp12L
michael@0 732 movq mm1, MMWORD [wk(7)] ; mm1=tmp12H
michael@0 733 movq mm3, MMWORD [wk(10)] ; mm3=tmp1L
michael@0 734 movq mm0, MMWORD [wk(11)] ; mm0=tmp1H
michael@0 735
michael@0 736 movq MMWORD [wk(0)], mm5 ; wk(0)=(00 10 20 30 06 16 26 36)
michael@0 737 movq MMWORD [wk(1)], mm4 ; wk(1)=(01 11 21 31 07 17 27 37)
michael@0 738
michael@0 739 movq mm7,mm6
michael@0 740 movq mm2,mm1
michael@0 741 paddd mm6,mm3 ; mm6=data2L
michael@0 742 paddd mm1,mm0 ; mm1=data2H
michael@0 743 psubd mm7,mm3 ; mm7=data5L
michael@0 744 psubd mm2,mm0 ; mm2=data5H
michael@0 745
michael@0 746 movq mm5,[GOTOFF(ebx,PD_DESCALE_P2)] ; mm5=[PD_DESCALE_P2]
michael@0 747
michael@0 748 paddd mm6,mm5
michael@0 749 paddd mm1,mm5
michael@0 750 psrad mm6,DESCALE_P2
michael@0 751 psrad mm1,DESCALE_P2
michael@0 752 paddd mm7,mm5
michael@0 753 paddd mm2,mm5
michael@0 754 psrad mm7,DESCALE_P2
michael@0 755 psrad mm2,DESCALE_P2
michael@0 756
michael@0 757 packssdw mm6,mm1 ; mm6=data2=(02 12 22 32)
michael@0 758 packssdw mm7,mm2 ; mm7=data5=(05 15 25 35)
michael@0 759
michael@0 760 movq mm4, MMWORD [wk(2)] ; mm4=tmp13L
michael@0 761 movq mm3, MMWORD [wk(3)] ; mm3=tmp13H
michael@0 762 movq mm0, MMWORD [wk(8)] ; mm0=tmp0L
michael@0 763 movq mm5, MMWORD [wk(9)] ; mm5=tmp0H
michael@0 764
michael@0 765 movq mm1,mm4
michael@0 766 movq mm2,mm3
michael@0 767 paddd mm4,mm0 ; mm4=data3L
michael@0 768 paddd mm3,mm5 ; mm3=data3H
michael@0 769 psubd mm1,mm0 ; mm1=data4L
michael@0 770 psubd mm2,mm5 ; mm2=data4H
michael@0 771
michael@0 772 movq mm0,[GOTOFF(ebx,PD_DESCALE_P2)] ; mm0=[PD_DESCALE_P2]
michael@0 773
michael@0 774 paddd mm4,mm0
michael@0 775 paddd mm3,mm0
michael@0 776 psrad mm4,DESCALE_P2
michael@0 777 psrad mm3,DESCALE_P2
michael@0 778 paddd mm1,mm0
michael@0 779 paddd mm2,mm0
michael@0 780 psrad mm1,DESCALE_P2
michael@0 781 psrad mm2,DESCALE_P2
michael@0 782
michael@0 783 movq mm5,[GOTOFF(ebx,PB_CENTERJSAMP)] ; mm5=[PB_CENTERJSAMP]
michael@0 784
michael@0 785 packssdw mm4,mm3 ; mm4=data3=(03 13 23 33)
michael@0 786 packssdw mm1,mm2 ; mm1=data4=(04 14 24 34)
michael@0 787
michael@0 788 movq mm0, MMWORD [wk(0)] ; mm0=(00 10 20 30 06 16 26 36)
michael@0 789 movq mm3, MMWORD [wk(1)] ; mm3=(01 11 21 31 07 17 27 37)
michael@0 790
michael@0 791 packsswb mm6,mm1 ; mm6=(02 12 22 32 04 14 24 34)
michael@0 792 packsswb mm4,mm7 ; mm4=(03 13 23 33 05 15 25 35)
michael@0 793
michael@0 794 paddb mm0,mm5
michael@0 795 paddb mm3,mm5
michael@0 796 paddb mm6,mm5
michael@0 797 paddb mm4,mm5
michael@0 798
michael@0 799 movq mm2,mm0 ; transpose coefficients(phase 1)
michael@0 800 punpcklbw mm0,mm3 ; mm0=(00 01 10 11 20 21 30 31)
michael@0 801 punpckhbw mm2,mm3 ; mm2=(06 07 16 17 26 27 36 37)
michael@0 802 movq mm1,mm6 ; transpose coefficients(phase 1)
michael@0 803 punpcklbw mm6,mm4 ; mm6=(02 03 12 13 22 23 32 33)
michael@0 804 punpckhbw mm1,mm4 ; mm1=(04 05 14 15 24 25 34 35)
michael@0 805
michael@0 806 movq mm7,mm0 ; transpose coefficients(phase 2)
michael@0 807 punpcklwd mm0,mm6 ; mm0=(00 01 02 03 10 11 12 13)
michael@0 808 punpckhwd mm7,mm6 ; mm7=(20 21 22 23 30 31 32 33)
michael@0 809 movq mm5,mm1 ; transpose coefficients(phase 2)
michael@0 810 punpcklwd mm1,mm2 ; mm1=(04 05 06 07 14 15 16 17)
michael@0 811 punpckhwd mm5,mm2 ; mm5=(24 25 26 27 34 35 36 37)
michael@0 812
michael@0 813 movq mm3,mm0 ; transpose coefficients(phase 3)
michael@0 814 punpckldq mm0,mm1 ; mm0=(00 01 02 03 04 05 06 07)
michael@0 815 punpckhdq mm3,mm1 ; mm3=(10 11 12 13 14 15 16 17)
michael@0 816 movq mm4,mm7 ; transpose coefficients(phase 3)
michael@0 817 punpckldq mm7,mm5 ; mm7=(20 21 22 23 24 25 26 27)
michael@0 818 punpckhdq mm4,mm5 ; mm4=(30 31 32 33 34 35 36 37)
michael@0 819
michael@0 820 pushpic ebx ; save GOT address
michael@0 821
michael@0 822 mov edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
michael@0 823 mov ebx, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
michael@0 824 movq MMWORD [edx+eax*SIZEOF_JSAMPLE], mm0
michael@0 825 movq MMWORD [ebx+eax*SIZEOF_JSAMPLE], mm3
michael@0 826 mov edx, JSAMPROW [edi+2*SIZEOF_JSAMPROW]
michael@0 827 mov ebx, JSAMPROW [edi+3*SIZEOF_JSAMPROW]
michael@0 828 movq MMWORD [edx+eax*SIZEOF_JSAMPLE], mm7
michael@0 829 movq MMWORD [ebx+eax*SIZEOF_JSAMPLE], mm4
michael@0 830
michael@0 831 poppic ebx ; restore GOT address
michael@0 832
michael@0 833 add esi, byte 4*SIZEOF_JCOEF ; wsptr
michael@0 834 add edi, byte 4*SIZEOF_JSAMPROW
michael@0 835 dec ecx ; ctr
michael@0 836 jnz near .rowloop
michael@0 837
michael@0 838 emms ; empty MMX state
michael@0 839
michael@0 840 pop edi
michael@0 841 pop esi
michael@0 842 ; pop edx ; need not be preserved
michael@0 843 ; pop ecx ; need not be preserved
michael@0 844 pop ebx
michael@0 845 mov esp,ebp ; esp <- aligned ebp
michael@0 846 pop esp ; esp <- original ebp
michael@0 847 pop ebp
michael@0 848 ret
michael@0 849
michael@0 850 ; For some reason, the OS X linker does not honor the request to align the
michael@0 851 ; segment unless we do this.
michael@0 852 align 16

Mercurial Repository: The Tor Browser

Europalab SCM Repository Server

mercurial