The Tor Browser / annotate

media/libjpeg/simd/jiss2int-64.asm@b8a032363ba2 (annotated)

media/libjpeg/simd/jiss2int-64.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 ; jiss2int-64.asm - accurate integer IDCT (64-bit SSE2)
michael@0 3 ;
michael@0 4 ; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
michael@0 5 ; Copyright 2009 D. R. Commander
michael@0 6 ;
michael@0 7 ; Based on
michael@0 8 ; x86 SIMD extension for IJG JPEG library
michael@0 9 ; Copyright (C) 1999-2006, MIYASAKA Masaru.
michael@0 10 ; For conditions of distribution and use, see copyright notice in jsimdext.inc
michael@0 11 ;
michael@0 12 ; This file should be assembled with NASM (Netwide Assembler),
michael@0 13 ; can *not* be assembled with Microsoft's MASM or any compatible
michael@0 14 ; assembler (including Borland's Turbo Assembler).
michael@0 15 ; NASM is available from http://nasm.sourceforge.net/ or
michael@0 16 ; http://sourceforge.net/project/showfiles.php?group_id=6208
michael@0 17 ;
michael@0 18 ; This file contains a slow-but-accurate integer implementation of the
michael@0 19 ; inverse DCT (Discrete Cosine Transform). The following code is based
michael@0 20 ; directly on the IJG's original jidctint.c; see the jidctint.c for
michael@0 21 ; more details.
michael@0 22 ;
michael@0 23 ; [TAB8]
michael@0 24
michael@0 25 %include "jsimdext.inc"
michael@0 26 %include "jdct.inc"
michael@0 27
michael@0 28 ; --------------------------------------------------------------------------
michael@0 29
michael@0 30 %define CONST_BITS 13
michael@0 31 %define PASS1_BITS 2
michael@0 32
michael@0 33 %define DESCALE_P1 (CONST_BITS-PASS1_BITS)
michael@0 34 %define DESCALE_P2 (CONST_BITS+PASS1_BITS+3)
michael@0 35
michael@0 36 %if CONST_BITS == 13
michael@0 37 F_0_298 equ 2446 ; FIX(0.298631336)
michael@0 38 F_0_390 equ 3196 ; FIX(0.390180644)
michael@0 39 F_0_541 equ 4433 ; FIX(0.541196100)
michael@0 40 F_0_765 equ 6270 ; FIX(0.765366865)
michael@0 41 F_0_899 equ 7373 ; FIX(0.899976223)
michael@0 42 F_1_175 equ 9633 ; FIX(1.175875602)
michael@0 43 F_1_501 equ 12299 ; FIX(1.501321110)
michael@0 44 F_1_847 equ 15137 ; FIX(1.847759065)
michael@0 45 F_1_961 equ 16069 ; FIX(1.961570560)
michael@0 46 F_2_053 equ 16819 ; FIX(2.053119869)
michael@0 47 F_2_562 equ 20995 ; FIX(2.562915447)
michael@0 48 F_3_072 equ 25172 ; FIX(3.072711026)
michael@0 49 %else
michael@0 50 ; NASM cannot do compile-time arithmetic on floating-point constants.
michael@0 51 %define DESCALE(x,n) (((x)+(1<<((n)-1)))>>(n))
michael@0 52 F_0_298 equ DESCALE( 320652955,30-CONST_BITS) ; FIX(0.298631336)
michael@0 53 F_0_390 equ DESCALE( 418953276,30-CONST_BITS) ; FIX(0.390180644)
michael@0 54 F_0_541 equ DESCALE( 581104887,30-CONST_BITS) ; FIX(0.541196100)
michael@0 55 F_0_765 equ DESCALE( 821806413,30-CONST_BITS) ; FIX(0.765366865)
michael@0 56 F_0_899 equ DESCALE( 966342111,30-CONST_BITS) ; FIX(0.899976223)
michael@0 57 F_1_175 equ DESCALE(1262586813,30-CONST_BITS) ; FIX(1.175875602)
michael@0 58 F_1_501 equ DESCALE(1612031267,30-CONST_BITS) ; FIX(1.501321110)
michael@0 59 F_1_847 equ DESCALE(1984016188,30-CONST_BITS) ; FIX(1.847759065)
michael@0 60 F_1_961 equ DESCALE(2106220350,30-CONST_BITS) ; FIX(1.961570560)
michael@0 61 F_2_053 equ DESCALE(2204520673,30-CONST_BITS) ; FIX(2.053119869)
michael@0 62 F_2_562 equ DESCALE(2751909506,30-CONST_BITS) ; FIX(2.562915447)
michael@0 63 F_3_072 equ DESCALE(3299298341,30-CONST_BITS) ; FIX(3.072711026)
michael@0 64 %endif
michael@0 65
michael@0 66 ; --------------------------------------------------------------------------
michael@0 67 SECTION SEG_CONST
michael@0 68
michael@0 69 alignz 16
michael@0 70 global EXTN(jconst_idct_islow_sse2)
michael@0 71
michael@0 72 EXTN(jconst_idct_islow_sse2):
michael@0 73
michael@0 74 PW_F130_F054 times 4 dw (F_0_541+F_0_765), F_0_541
michael@0 75 PW_F054_MF130 times 4 dw F_0_541, (F_0_541-F_1_847)
michael@0 76 PW_MF078_F117 times 4 dw (F_1_175-F_1_961), F_1_175
michael@0 77 PW_F117_F078 times 4 dw F_1_175, (F_1_175-F_0_390)
michael@0 78 PW_MF060_MF089 times 4 dw (F_0_298-F_0_899),-F_0_899
michael@0 79 PW_MF089_F060 times 4 dw -F_0_899, (F_1_501-F_0_899)
michael@0 80 PW_MF050_MF256 times 4 dw (F_2_053-F_2_562),-F_2_562
michael@0 81 PW_MF256_F050 times 4 dw -F_2_562, (F_3_072-F_2_562)
michael@0 82 PD_DESCALE_P1 times 4 dd 1 << (DESCALE_P1-1)
michael@0 83 PD_DESCALE_P2 times 4 dd 1 << (DESCALE_P2-1)
michael@0 84 PB_CENTERJSAMP times 16 db CENTERJSAMPLE
michael@0 85
michael@0 86 alignz 16
michael@0 87
michael@0 88 ; --------------------------------------------------------------------------
michael@0 89 SECTION SEG_TEXT
michael@0 90 BITS 64
michael@0 91 ;
michael@0 92 ; Perform dequantization and inverse DCT on one block of coefficients.
michael@0 93 ;
michael@0 94 ; GLOBAL(void)
michael@0 95 ; jsimd_idct_islow_sse2 (void * dct_table, JCOEFPTR coef_block,
michael@0 96 ; JSAMPARRAY output_buf, JDIMENSION output_col)
michael@0 97 ;
michael@0 98
michael@0 99 ; r10 = jpeg_component_info * compptr
michael@0 100 ; r11 = JCOEFPTR coef_block
michael@0 101 ; r12 = JSAMPARRAY output_buf
michael@0 102 ; r13 = JDIMENSION output_col
michael@0 103
michael@0 104 %define original_rbp rbp+0
michael@0 105 %define wk(i) rbp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM]
michael@0 106 %define WK_NUM 12
michael@0 107
michael@0 108 align 16
michael@0 109 global EXTN(jsimd_idct_islow_sse2)
michael@0 110
michael@0 111 EXTN(jsimd_idct_islow_sse2):
michael@0 112 push rbp
michael@0 113 mov rax,rsp ; rax = original rbp
michael@0 114 sub rsp, byte 4
michael@0 115 and rsp, byte (-SIZEOF_XMMWORD) ; align to 128 bits
michael@0 116 mov [rsp],rax
michael@0 117 mov rbp,rsp ; rbp = aligned rbp
michael@0 118 lea rsp, [wk(0)]
michael@0 119 collect_args
michael@0 120
michael@0 121 ; ---- Pass 1: process columns from input.
michael@0 122
michael@0 123 mov rdx, r10 ; quantptr
michael@0 124 mov rsi, r11 ; inptr
michael@0 125
michael@0 126 %ifndef NO_ZERO_COLUMN_TEST_ISLOW_SSE2
michael@0 127 mov eax, DWORD [DWBLOCK(1,0,rsi,SIZEOF_JCOEF)]
michael@0 128 or eax, DWORD [DWBLOCK(2,0,rsi,SIZEOF_JCOEF)]
michael@0 129 jnz near .columnDCT
michael@0 130
michael@0 131 movdqa xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
michael@0 132 movdqa xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
michael@0 133 por xmm0, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
michael@0 134 por xmm1, XMMWORD [XMMBLOCK(4,0,rsi,SIZEOF_JCOEF)]
michael@0 135 por xmm0, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
michael@0 136 por xmm1, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
michael@0 137 por xmm0, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
michael@0 138 por xmm1,xmm0
michael@0 139 packsswb xmm1,xmm1
michael@0 140 packsswb xmm1,xmm1
michael@0 141 movd eax,xmm1
michael@0 142 test rax,rax
michael@0 143 jnz short .columnDCT
michael@0 144
michael@0 145 ; -- AC terms all zero
michael@0 146
michael@0 147 movdqa xmm5, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
michael@0 148 pmullw xmm5, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 149
michael@0 150 psllw xmm5,PASS1_BITS
michael@0 151
michael@0 152 movdqa xmm4,xmm5 ; xmm5=in0=(00 01 02 03 04 05 06 07)
michael@0 153 punpcklwd xmm5,xmm5 ; xmm5=(00 00 01 01 02 02 03 03)
michael@0 154 punpckhwd xmm4,xmm4 ; xmm4=(04 04 05 05 06 06 07 07)
michael@0 155
michael@0 156 pshufd xmm7,xmm5,0x00 ; xmm7=col0=(00 00 00 00 00 00 00 00)
michael@0 157 pshufd xmm6,xmm5,0x55 ; xmm6=col1=(01 01 01 01 01 01 01 01)
michael@0 158 pshufd xmm1,xmm5,0xAA ; xmm1=col2=(02 02 02 02 02 02 02 02)
michael@0 159 pshufd xmm5,xmm5,0xFF ; xmm5=col3=(03 03 03 03 03 03 03 03)
michael@0 160 pshufd xmm0,xmm4,0x00 ; xmm0=col4=(04 04 04 04 04 04 04 04)
michael@0 161 pshufd xmm3,xmm4,0x55 ; xmm3=col5=(05 05 05 05 05 05 05 05)
michael@0 162 pshufd xmm2,xmm4,0xAA ; xmm2=col6=(06 06 06 06 06 06 06 06)
michael@0 163 pshufd xmm4,xmm4,0xFF ; xmm4=col7=(07 07 07 07 07 07 07 07)
michael@0 164
michael@0 165 movdqa XMMWORD [wk(8)], xmm6 ; wk(8)=col1
michael@0 166 movdqa XMMWORD [wk(9)], xmm5 ; wk(9)=col3
michael@0 167 movdqa XMMWORD [wk(10)], xmm3 ; wk(10)=col5
michael@0 168 movdqa XMMWORD [wk(11)], xmm4 ; wk(11)=col7
michael@0 169 jmp near .column_end
michael@0 170 %endif
michael@0 171 .columnDCT:
michael@0 172
michael@0 173 ; -- Even part
michael@0 174
michael@0 175 movdqa xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
michael@0 176 movdqa xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
michael@0 177 pmullw xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 178 pmullw xmm1, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 179 movdqa xmm2, XMMWORD [XMMBLOCK(4,0,rsi,SIZEOF_JCOEF)]
michael@0 180 movdqa xmm3, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
michael@0 181 pmullw xmm2, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 182 pmullw xmm3, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 183
michael@0 184 ; (Original)
michael@0 185 ; z1 = (z2 + z3) * 0.541196100;
michael@0 186 ; tmp2 = z1 + z3 * -1.847759065;
michael@0 187 ; tmp3 = z1 + z2 * 0.765366865;
michael@0 188 ;
michael@0 189 ; (This implementation)
michael@0 190 ; tmp2 = z2 * 0.541196100 + z3 * (0.541196100 - 1.847759065);
michael@0 191 ; tmp3 = z2 * (0.541196100 + 0.765366865) + z3 * 0.541196100;
michael@0 192
michael@0 193 movdqa xmm4,xmm1 ; xmm1=in2=z2
michael@0 194 movdqa xmm5,xmm1
michael@0 195 punpcklwd xmm4,xmm3 ; xmm3=in6=z3
michael@0 196 punpckhwd xmm5,xmm3
michael@0 197 movdqa xmm1,xmm4
michael@0 198 movdqa xmm3,xmm5
michael@0 199 pmaddwd xmm4,[rel PW_F130_F054] ; xmm4=tmp3L
michael@0 200 pmaddwd xmm5,[rel PW_F130_F054] ; xmm5=tmp3H
michael@0 201 pmaddwd xmm1,[rel PW_F054_MF130] ; xmm1=tmp2L
michael@0 202 pmaddwd xmm3,[rel PW_F054_MF130] ; xmm3=tmp2H
michael@0 203
michael@0 204 movdqa xmm6,xmm0
michael@0 205 paddw xmm0,xmm2 ; xmm0=in0+in4
michael@0 206 psubw xmm6,xmm2 ; xmm6=in0-in4
michael@0 207
michael@0 208 pxor xmm7,xmm7
michael@0 209 pxor xmm2,xmm2
michael@0 210 punpcklwd xmm7,xmm0 ; xmm7=tmp0L
michael@0 211 punpckhwd xmm2,xmm0 ; xmm2=tmp0H
michael@0 212 psrad xmm7,(16-CONST_BITS) ; psrad xmm7,16 & pslld xmm7,CONST_BITS
michael@0 213 psrad xmm2,(16-CONST_BITS) ; psrad xmm2,16 & pslld xmm2,CONST_BITS
michael@0 214
michael@0 215 movdqa xmm0,xmm7
michael@0 216 paddd xmm7,xmm4 ; xmm7=tmp10L
michael@0 217 psubd xmm0,xmm4 ; xmm0=tmp13L
michael@0 218 movdqa xmm4,xmm2
michael@0 219 paddd xmm2,xmm5 ; xmm2=tmp10H
michael@0 220 psubd xmm4,xmm5 ; xmm4=tmp13H
michael@0 221
michael@0 222 movdqa XMMWORD [wk(0)], xmm7 ; wk(0)=tmp10L
michael@0 223 movdqa XMMWORD [wk(1)], xmm2 ; wk(1)=tmp10H
michael@0 224 movdqa XMMWORD [wk(2)], xmm0 ; wk(2)=tmp13L
michael@0 225 movdqa XMMWORD [wk(3)], xmm4 ; wk(3)=tmp13H
michael@0 226
michael@0 227 pxor xmm5,xmm5
michael@0 228 pxor xmm7,xmm7
michael@0 229 punpcklwd xmm5,xmm6 ; xmm5=tmp1L
michael@0 230 punpckhwd xmm7,xmm6 ; xmm7=tmp1H
michael@0 231 psrad xmm5,(16-CONST_BITS) ; psrad xmm5,16 & pslld xmm5,CONST_BITS
michael@0 232 psrad xmm7,(16-CONST_BITS) ; psrad xmm7,16 & pslld xmm7,CONST_BITS
michael@0 233
michael@0 234 movdqa xmm2,xmm5
michael@0 235 paddd xmm5,xmm1 ; xmm5=tmp11L
michael@0 236 psubd xmm2,xmm1 ; xmm2=tmp12L
michael@0 237 movdqa xmm0,xmm7
michael@0 238 paddd xmm7,xmm3 ; xmm7=tmp11H
michael@0 239 psubd xmm0,xmm3 ; xmm0=tmp12H
michael@0 240
michael@0 241 movdqa XMMWORD [wk(4)], xmm5 ; wk(4)=tmp11L
michael@0 242 movdqa XMMWORD [wk(5)], xmm7 ; wk(5)=tmp11H
michael@0 243 movdqa XMMWORD [wk(6)], xmm2 ; wk(6)=tmp12L
michael@0 244 movdqa XMMWORD [wk(7)], xmm0 ; wk(7)=tmp12H
michael@0 245
michael@0 246 ; -- Odd part
michael@0 247
michael@0 248 movdqa xmm4, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
michael@0 249 movdqa xmm6, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
michael@0 250 pmullw xmm4, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 251 pmullw xmm6, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 252 movdqa xmm1, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
michael@0 253 movdqa xmm3, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
michael@0 254 pmullw xmm1, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 255 pmullw xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 256
michael@0 257 movdqa xmm5,xmm6
michael@0 258 movdqa xmm7,xmm4
michael@0 259 paddw xmm5,xmm3 ; xmm5=z3
michael@0 260 paddw xmm7,xmm1 ; xmm7=z4
michael@0 261
michael@0 262 ; (Original)
michael@0 263 ; z5 = (z3 + z4) * 1.175875602;
michael@0 264 ; z3 = z3 * -1.961570560; z4 = z4 * -0.390180644;
michael@0 265 ; z3 += z5; z4 += z5;
michael@0 266 ;
michael@0 267 ; (This implementation)
michael@0 268 ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
michael@0 269 ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
michael@0 270
michael@0 271 movdqa xmm2,xmm5
michael@0 272 movdqa xmm0,xmm5
michael@0 273 punpcklwd xmm2,xmm7
michael@0 274 punpckhwd xmm0,xmm7
michael@0 275 movdqa xmm5,xmm2
michael@0 276 movdqa xmm7,xmm0
michael@0 277 pmaddwd xmm2,[rel PW_MF078_F117] ; xmm2=z3L
michael@0 278 pmaddwd xmm0,[rel PW_MF078_F117] ; xmm0=z3H
michael@0 279 pmaddwd xmm5,[rel PW_F117_F078] ; xmm5=z4L
michael@0 280 pmaddwd xmm7,[rel PW_F117_F078] ; xmm7=z4H
michael@0 281
michael@0 282 movdqa XMMWORD [wk(10)], xmm2 ; wk(10)=z3L
michael@0 283 movdqa XMMWORD [wk(11)], xmm0 ; wk(11)=z3H
michael@0 284
michael@0 285 ; (Original)
michael@0 286 ; z1 = tmp0 + tmp3; z2 = tmp1 + tmp2;
michael@0 287 ; tmp0 = tmp0 * 0.298631336; tmp1 = tmp1 * 2.053119869;
michael@0 288 ; tmp2 = tmp2 * 3.072711026; tmp3 = tmp3 * 1.501321110;
michael@0 289 ; z1 = z1 * -0.899976223; z2 = z2 * -2.562915447;
michael@0 290 ; tmp0 += z1 + z3; tmp1 += z2 + z4;
michael@0 291 ; tmp2 += z2 + z3; tmp3 += z1 + z4;
michael@0 292 ;
michael@0 293 ; (This implementation)
michael@0 294 ; tmp0 = tmp0 * (0.298631336 - 0.899976223) + tmp3 * -0.899976223;
michael@0 295 ; tmp1 = tmp1 * (2.053119869 - 2.562915447) + tmp2 * -2.562915447;
michael@0 296 ; tmp2 = tmp1 * -2.562915447 + tmp2 * (3.072711026 - 2.562915447);
michael@0 297 ; tmp3 = tmp0 * -0.899976223 + tmp3 * (1.501321110 - 0.899976223);
michael@0 298 ; tmp0 += z3; tmp1 += z4;
michael@0 299 ; tmp2 += z3; tmp3 += z4;
michael@0 300
michael@0 301 movdqa xmm2,xmm3
michael@0 302 movdqa xmm0,xmm3
michael@0 303 punpcklwd xmm2,xmm4
michael@0 304 punpckhwd xmm0,xmm4
michael@0 305 movdqa xmm3,xmm2
michael@0 306 movdqa xmm4,xmm0
michael@0 307 pmaddwd xmm2,[rel PW_MF060_MF089] ; xmm2=tmp0L
michael@0 308 pmaddwd xmm0,[rel PW_MF060_MF089] ; xmm0=tmp0H
michael@0 309 pmaddwd xmm3,[rel PW_MF089_F060] ; xmm3=tmp3L
michael@0 310 pmaddwd xmm4,[rel PW_MF089_F060] ; xmm4=tmp3H
michael@0 311
michael@0 312 paddd xmm2, XMMWORD [wk(10)] ; xmm2=tmp0L
michael@0 313 paddd xmm0, XMMWORD [wk(11)] ; xmm0=tmp0H
michael@0 314 paddd xmm3,xmm5 ; xmm3=tmp3L
michael@0 315 paddd xmm4,xmm7 ; xmm4=tmp3H
michael@0 316
michael@0 317 movdqa XMMWORD [wk(8)], xmm2 ; wk(8)=tmp0L
michael@0 318 movdqa XMMWORD [wk(9)], xmm0 ; wk(9)=tmp0H
michael@0 319
michael@0 320 movdqa xmm2,xmm1
michael@0 321 movdqa xmm0,xmm1
michael@0 322 punpcklwd xmm2,xmm6
michael@0 323 punpckhwd xmm0,xmm6
michael@0 324 movdqa xmm1,xmm2
michael@0 325 movdqa xmm6,xmm0
michael@0 326 pmaddwd xmm2,[rel PW_MF050_MF256] ; xmm2=tmp1L
michael@0 327 pmaddwd xmm0,[rel PW_MF050_MF256] ; xmm0=tmp1H
michael@0 328 pmaddwd xmm1,[rel PW_MF256_F050] ; xmm1=tmp2L
michael@0 329 pmaddwd xmm6,[rel PW_MF256_F050] ; xmm6=tmp2H
michael@0 330
michael@0 331 paddd xmm2,xmm5 ; xmm2=tmp1L
michael@0 332 paddd xmm0,xmm7 ; xmm0=tmp1H
michael@0 333 paddd xmm1, XMMWORD [wk(10)] ; xmm1=tmp2L
michael@0 334 paddd xmm6, XMMWORD [wk(11)] ; xmm6=tmp2H
michael@0 335
michael@0 336 movdqa XMMWORD [wk(10)], xmm2 ; wk(10)=tmp1L
michael@0 337 movdqa XMMWORD [wk(11)], xmm0 ; wk(11)=tmp1H
michael@0 338
michael@0 339 ; -- Final output stage
michael@0 340
michael@0 341 movdqa xmm5, XMMWORD [wk(0)] ; xmm5=tmp10L
michael@0 342 movdqa xmm7, XMMWORD [wk(1)] ; xmm7=tmp10H
michael@0 343
michael@0 344 movdqa xmm2,xmm5
michael@0 345 movdqa xmm0,xmm7
michael@0 346 paddd xmm5,xmm3 ; xmm5=data0L
michael@0 347 paddd xmm7,xmm4 ; xmm7=data0H
michael@0 348 psubd xmm2,xmm3 ; xmm2=data7L
michael@0 349 psubd xmm0,xmm4 ; xmm0=data7H
michael@0 350
michael@0 351 movdqa xmm3,[rel PD_DESCALE_P1] ; xmm3=[rel PD_DESCALE_P1]
michael@0 352
michael@0 353 paddd xmm5,xmm3
michael@0 354 paddd xmm7,xmm3
michael@0 355 psrad xmm5,DESCALE_P1
michael@0 356 psrad xmm7,DESCALE_P1
michael@0 357 paddd xmm2,xmm3
michael@0 358 paddd xmm0,xmm3
michael@0 359 psrad xmm2,DESCALE_P1
michael@0 360 psrad xmm0,DESCALE_P1
michael@0 361
michael@0 362 packssdw xmm5,xmm7 ; xmm5=data0=(00 01 02 03 04 05 06 07)
michael@0 363 packssdw xmm2,xmm0 ; xmm2=data7=(70 71 72 73 74 75 76 77)
michael@0 364
michael@0 365 movdqa xmm4, XMMWORD [wk(4)] ; xmm4=tmp11L
michael@0 366 movdqa xmm3, XMMWORD [wk(5)] ; xmm3=tmp11H
michael@0 367
michael@0 368 movdqa xmm7,xmm4
michael@0 369 movdqa xmm0,xmm3
michael@0 370 paddd xmm4,xmm1 ; xmm4=data1L
michael@0 371 paddd xmm3,xmm6 ; xmm3=data1H
michael@0 372 psubd xmm7,xmm1 ; xmm7=data6L
michael@0 373 psubd xmm0,xmm6 ; xmm0=data6H
michael@0 374
michael@0 375 movdqa xmm1,[rel PD_DESCALE_P1] ; xmm1=[rel PD_DESCALE_P1]
michael@0 376
michael@0 377 paddd xmm4,xmm1
michael@0 378 paddd xmm3,xmm1
michael@0 379 psrad xmm4,DESCALE_P1
michael@0 380 psrad xmm3,DESCALE_P1
michael@0 381 paddd xmm7,xmm1
michael@0 382 paddd xmm0,xmm1
michael@0 383 psrad xmm7,DESCALE_P1
michael@0 384 psrad xmm0,DESCALE_P1
michael@0 385
michael@0 386 packssdw xmm4,xmm3 ; xmm4=data1=(10 11 12 13 14 15 16 17)
michael@0 387 packssdw xmm7,xmm0 ; xmm7=data6=(60 61 62 63 64 65 66 67)
michael@0 388
michael@0 389 movdqa xmm6,xmm5 ; transpose coefficients(phase 1)
michael@0 390 punpcklwd xmm5,xmm4 ; xmm5=(00 10 01 11 02 12 03 13)
michael@0 391 punpckhwd xmm6,xmm4 ; xmm6=(04 14 05 15 06 16 07 17)
michael@0 392 movdqa xmm1,xmm7 ; transpose coefficients(phase 1)
michael@0 393 punpcklwd xmm7,xmm2 ; xmm7=(60 70 61 71 62 72 63 73)
michael@0 394 punpckhwd xmm1,xmm2 ; xmm1=(64 74 65 75 66 76 67 77)
michael@0 395
michael@0 396 movdqa xmm3, XMMWORD [wk(6)] ; xmm3=tmp12L
michael@0 397 movdqa xmm0, XMMWORD [wk(7)] ; xmm0=tmp12H
michael@0 398 movdqa xmm4, XMMWORD [wk(10)] ; xmm4=tmp1L
michael@0 399 movdqa xmm2, XMMWORD [wk(11)] ; xmm2=tmp1H
michael@0 400
michael@0 401 movdqa XMMWORD [wk(0)], xmm5 ; wk(0)=(00 10 01 11 02 12 03 13)
michael@0 402 movdqa XMMWORD [wk(1)], xmm6 ; wk(1)=(04 14 05 15 06 16 07 17)
michael@0 403 movdqa XMMWORD [wk(4)], xmm7 ; wk(4)=(60 70 61 71 62 72 63 73)
michael@0 404 movdqa XMMWORD [wk(5)], xmm1 ; wk(5)=(64 74 65 75 66 76 67 77)
michael@0 405
michael@0 406 movdqa xmm5,xmm3
michael@0 407 movdqa xmm6,xmm0
michael@0 408 paddd xmm3,xmm4 ; xmm3=data2L
michael@0 409 paddd xmm0,xmm2 ; xmm0=data2H
michael@0 410 psubd xmm5,xmm4 ; xmm5=data5L
michael@0 411 psubd xmm6,xmm2 ; xmm6=data5H
michael@0 412
michael@0 413 movdqa xmm7,[rel PD_DESCALE_P1] ; xmm7=[rel PD_DESCALE_P1]
michael@0 414
michael@0 415 paddd xmm3,xmm7
michael@0 416 paddd xmm0,xmm7
michael@0 417 psrad xmm3,DESCALE_P1
michael@0 418 psrad xmm0,DESCALE_P1
michael@0 419 paddd xmm5,xmm7
michael@0 420 paddd xmm6,xmm7
michael@0 421 psrad xmm5,DESCALE_P1
michael@0 422 psrad xmm6,DESCALE_P1
michael@0 423
michael@0 424 packssdw xmm3,xmm0 ; xmm3=data2=(20 21 22 23 24 25 26 27)
michael@0 425 packssdw xmm5,xmm6 ; xmm5=data5=(50 51 52 53 54 55 56 57)
michael@0 426
michael@0 427 movdqa xmm1, XMMWORD [wk(2)] ; xmm1=tmp13L
michael@0 428 movdqa xmm4, XMMWORD [wk(3)] ; xmm4=tmp13H
michael@0 429 movdqa xmm2, XMMWORD [wk(8)] ; xmm2=tmp0L
michael@0 430 movdqa xmm7, XMMWORD [wk(9)] ; xmm7=tmp0H
michael@0 431
michael@0 432 movdqa xmm0,xmm1
michael@0 433 movdqa xmm6,xmm4
michael@0 434 paddd xmm1,xmm2 ; xmm1=data3L
michael@0 435 paddd xmm4,xmm7 ; xmm4=data3H
michael@0 436 psubd xmm0,xmm2 ; xmm0=data4L
michael@0 437 psubd xmm6,xmm7 ; xmm6=data4H
michael@0 438
michael@0 439 movdqa xmm2,[rel PD_DESCALE_P1] ; xmm2=[rel PD_DESCALE_P1]
michael@0 440
michael@0 441 paddd xmm1,xmm2
michael@0 442 paddd xmm4,xmm2
michael@0 443 psrad xmm1,DESCALE_P1
michael@0 444 psrad xmm4,DESCALE_P1
michael@0 445 paddd xmm0,xmm2
michael@0 446 paddd xmm6,xmm2
michael@0 447 psrad xmm0,DESCALE_P1
michael@0 448 psrad xmm6,DESCALE_P1
michael@0 449
michael@0 450 packssdw xmm1,xmm4 ; xmm1=data3=(30 31 32 33 34 35 36 37)
michael@0 451 packssdw xmm0,xmm6 ; xmm0=data4=(40 41 42 43 44 45 46 47)
michael@0 452
michael@0 453 movdqa xmm7, XMMWORD [wk(0)] ; xmm7=(00 10 01 11 02 12 03 13)
michael@0 454 movdqa xmm2, XMMWORD [wk(1)] ; xmm2=(04 14 05 15 06 16 07 17)
michael@0 455
michael@0 456 movdqa xmm4,xmm3 ; transpose coefficients(phase 1)
michael@0 457 punpcklwd xmm3,xmm1 ; xmm3=(20 30 21 31 22 32 23 33)
michael@0 458 punpckhwd xmm4,xmm1 ; xmm4=(24 34 25 35 26 36 27 37)
michael@0 459 movdqa xmm6,xmm0 ; transpose coefficients(phase 1)
michael@0 460 punpcklwd xmm0,xmm5 ; xmm0=(40 50 41 51 42 52 43 53)
michael@0 461 punpckhwd xmm6,xmm5 ; xmm6=(44 54 45 55 46 56 47 57)
michael@0 462
michael@0 463 movdqa xmm1,xmm7 ; transpose coefficients(phase 2)
michael@0 464 punpckldq xmm7,xmm3 ; xmm7=(00 10 20 30 01 11 21 31)
michael@0 465 punpckhdq xmm1,xmm3 ; xmm1=(02 12 22 32 03 13 23 33)
michael@0 466 movdqa xmm5,xmm2 ; transpose coefficients(phase 2)
michael@0 467 punpckldq xmm2,xmm4 ; xmm2=(04 14 24 34 05 15 25 35)
michael@0 468 punpckhdq xmm5,xmm4 ; xmm5=(06 16 26 36 07 17 27 37)
michael@0 469
michael@0 470 movdqa xmm3, XMMWORD [wk(4)] ; xmm3=(60 70 61 71 62 72 63 73)
michael@0 471 movdqa xmm4, XMMWORD [wk(5)] ; xmm4=(64 74 65 75 66 76 67 77)
michael@0 472
michael@0 473 movdqa XMMWORD [wk(6)], xmm2 ; wk(6)=(04 14 24 34 05 15 25 35)
michael@0 474 movdqa XMMWORD [wk(7)], xmm5 ; wk(7)=(06 16 26 36 07 17 27 37)
michael@0 475
michael@0 476 movdqa xmm2,xmm0 ; transpose coefficients(phase 2)
michael@0 477 punpckldq xmm0,xmm3 ; xmm0=(40 50 60 70 41 51 61 71)
michael@0 478 punpckhdq xmm2,xmm3 ; xmm2=(42 52 62 72 43 53 63 73)
michael@0 479 movdqa xmm5,xmm6 ; transpose coefficients(phase 2)
michael@0 480 punpckldq xmm6,xmm4 ; xmm6=(44 54 64 74 45 55 65 75)
michael@0 481 punpckhdq xmm5,xmm4 ; xmm5=(46 56 66 76 47 57 67 77)
michael@0 482
michael@0 483 movdqa xmm3,xmm7 ; transpose coefficients(phase 3)
michael@0 484 punpcklqdq xmm7,xmm0 ; xmm7=col0=(00 10 20 30 40 50 60 70)
michael@0 485 punpckhqdq xmm3,xmm0 ; xmm3=col1=(01 11 21 31 41 51 61 71)
michael@0 486 movdqa xmm4,xmm1 ; transpose coefficients(phase 3)
michael@0 487 punpcklqdq xmm1,xmm2 ; xmm1=col2=(02 12 22 32 42 52 62 72)
michael@0 488 punpckhqdq xmm4,xmm2 ; xmm4=col3=(03 13 23 33 43 53 63 73)
michael@0 489
michael@0 490 movdqa xmm0, XMMWORD [wk(6)] ; xmm0=(04 14 24 34 05 15 25 35)
michael@0 491 movdqa xmm2, XMMWORD [wk(7)] ; xmm2=(06 16 26 36 07 17 27 37)
michael@0 492
michael@0 493 movdqa XMMWORD [wk(8)], xmm3 ; wk(8)=col1
michael@0 494 movdqa XMMWORD [wk(9)], xmm4 ; wk(9)=col3
michael@0 495
michael@0 496 movdqa xmm3,xmm0 ; transpose coefficients(phase 3)
michael@0 497 punpcklqdq xmm0,xmm6 ; xmm0=col4=(04 14 24 34 44 54 64 74)
michael@0 498 punpckhqdq xmm3,xmm6 ; xmm3=col5=(05 15 25 35 45 55 65 75)
michael@0 499 movdqa xmm4,xmm2 ; transpose coefficients(phase 3)
michael@0 500 punpcklqdq xmm2,xmm5 ; xmm2=col6=(06 16 26 36 46 56 66 76)
michael@0 501 punpckhqdq xmm4,xmm5 ; xmm4=col7=(07 17 27 37 47 57 67 77)
michael@0 502
michael@0 503 movdqa XMMWORD [wk(10)], xmm3 ; wk(10)=col5
michael@0 504 movdqa XMMWORD [wk(11)], xmm4 ; wk(11)=col7
michael@0 505 .column_end:
michael@0 506
michael@0 507 ; -- Prefetch the next coefficient block
michael@0 508
michael@0 509 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
michael@0 510 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
michael@0 511 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
michael@0 512 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
michael@0 513
michael@0 514 ; ---- Pass 2: process rows from work array, store into output array.
michael@0 515
michael@0 516 mov rax, [original_rbp]
michael@0 517 mov rdi, r12 ; (JSAMPROW *)
michael@0 518 mov rax, r13
michael@0 519
michael@0 520 ; -- Even part
michael@0 521
michael@0 522 ; xmm7=col0, xmm1=col2, xmm0=col4, xmm2=col6
michael@0 523
michael@0 524 ; (Original)
michael@0 525 ; z1 = (z2 + z3) * 0.541196100;
michael@0 526 ; tmp2 = z1 + z3 * -1.847759065;
michael@0 527 ; tmp3 = z1 + z2 * 0.765366865;
michael@0 528 ;
michael@0 529 ; (This implementation)
michael@0 530 ; tmp2 = z2 * 0.541196100 + z3 * (0.541196100 - 1.847759065);
michael@0 531 ; tmp3 = z2 * (0.541196100 + 0.765366865) + z3 * 0.541196100;
michael@0 532
michael@0 533 movdqa xmm6,xmm1 ; xmm1=in2=z2
michael@0 534 movdqa xmm5,xmm1
michael@0 535 punpcklwd xmm6,xmm2 ; xmm2=in6=z3
michael@0 536 punpckhwd xmm5,xmm2
michael@0 537 movdqa xmm1,xmm6
michael@0 538 movdqa xmm2,xmm5
michael@0 539 pmaddwd xmm6,[rel PW_F130_F054] ; xmm6=tmp3L
michael@0 540 pmaddwd xmm5,[rel PW_F130_F054] ; xmm5=tmp3H
michael@0 541 pmaddwd xmm1,[rel PW_F054_MF130] ; xmm1=tmp2L
michael@0 542 pmaddwd xmm2,[rel PW_F054_MF130] ; xmm2=tmp2H
michael@0 543
michael@0 544 movdqa xmm3,xmm7
michael@0 545 paddw xmm7,xmm0 ; xmm7=in0+in4
michael@0 546 psubw xmm3,xmm0 ; xmm3=in0-in4
michael@0 547
michael@0 548 pxor xmm4,xmm4
michael@0 549 pxor xmm0,xmm0
michael@0 550 punpcklwd xmm4,xmm7 ; xmm4=tmp0L
michael@0 551 punpckhwd xmm0,xmm7 ; xmm0=tmp0H
michael@0 552 psrad xmm4,(16-CONST_BITS) ; psrad xmm4,16 & pslld xmm4,CONST_BITS
michael@0 553 psrad xmm0,(16-CONST_BITS) ; psrad xmm0,16 & pslld xmm0,CONST_BITS
michael@0 554
michael@0 555 movdqa xmm7,xmm4
michael@0 556 paddd xmm4,xmm6 ; xmm4=tmp10L
michael@0 557 psubd xmm7,xmm6 ; xmm7=tmp13L
michael@0 558 movdqa xmm6,xmm0
michael@0 559 paddd xmm0,xmm5 ; xmm0=tmp10H
michael@0 560 psubd xmm6,xmm5 ; xmm6=tmp13H
michael@0 561
michael@0 562 movdqa XMMWORD [wk(0)], xmm4 ; wk(0)=tmp10L
michael@0 563 movdqa XMMWORD [wk(1)], xmm0 ; wk(1)=tmp10H
michael@0 564 movdqa XMMWORD [wk(2)], xmm7 ; wk(2)=tmp13L
michael@0 565 movdqa XMMWORD [wk(3)], xmm6 ; wk(3)=tmp13H
michael@0 566
michael@0 567 pxor xmm5,xmm5
michael@0 568 pxor xmm4,xmm4
michael@0 569 punpcklwd xmm5,xmm3 ; xmm5=tmp1L
michael@0 570 punpckhwd xmm4,xmm3 ; xmm4=tmp1H
michael@0 571 psrad xmm5,(16-CONST_BITS) ; psrad xmm5,16 & pslld xmm5,CONST_BITS
michael@0 572 psrad xmm4,(16-CONST_BITS) ; psrad xmm4,16 & pslld xmm4,CONST_BITS
michael@0 573
michael@0 574 movdqa xmm0,xmm5
michael@0 575 paddd xmm5,xmm1 ; xmm5=tmp11L
michael@0 576 psubd xmm0,xmm1 ; xmm0=tmp12L
michael@0 577 movdqa xmm7,xmm4
michael@0 578 paddd xmm4,xmm2 ; xmm4=tmp11H
michael@0 579 psubd xmm7,xmm2 ; xmm7=tmp12H
michael@0 580
michael@0 581 movdqa XMMWORD [wk(4)], xmm5 ; wk(4)=tmp11L
michael@0 582 movdqa XMMWORD [wk(5)], xmm4 ; wk(5)=tmp11H
michael@0 583 movdqa XMMWORD [wk(6)], xmm0 ; wk(6)=tmp12L
michael@0 584 movdqa XMMWORD [wk(7)], xmm7 ; wk(7)=tmp12H
michael@0 585
michael@0 586 ; -- Odd part
michael@0 587
michael@0 588 movdqa xmm6, XMMWORD [wk(9)] ; xmm6=col3
michael@0 589 movdqa xmm3, XMMWORD [wk(8)] ; xmm3=col1
michael@0 590 movdqa xmm1, XMMWORD [wk(11)] ; xmm1=col7
michael@0 591 movdqa xmm2, XMMWORD [wk(10)] ; xmm2=col5
michael@0 592
michael@0 593 movdqa xmm5,xmm6
michael@0 594 movdqa xmm4,xmm3
michael@0 595 paddw xmm5,xmm1 ; xmm5=z3
michael@0 596 paddw xmm4,xmm2 ; xmm4=z4
michael@0 597
michael@0 598 ; (Original)
michael@0 599 ; z5 = (z3 + z4) * 1.175875602;
michael@0 600 ; z3 = z3 * -1.961570560; z4 = z4 * -0.390180644;
michael@0 601 ; z3 += z5; z4 += z5;
michael@0 602 ;
michael@0 603 ; (This implementation)
michael@0 604 ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
michael@0 605 ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
michael@0 606
michael@0 607 movdqa xmm0,xmm5
michael@0 608 movdqa xmm7,xmm5
michael@0 609 punpcklwd xmm0,xmm4
michael@0 610 punpckhwd xmm7,xmm4
michael@0 611 movdqa xmm5,xmm0
michael@0 612 movdqa xmm4,xmm7
michael@0 613 pmaddwd xmm0,[rel PW_MF078_F117] ; xmm0=z3L
michael@0 614 pmaddwd xmm7,[rel PW_MF078_F117] ; xmm7=z3H
michael@0 615 pmaddwd xmm5,[rel PW_F117_F078] ; xmm5=z4L
michael@0 616 pmaddwd xmm4,[rel PW_F117_F078] ; xmm4=z4H
michael@0 617
michael@0 618 movdqa XMMWORD [wk(10)], xmm0 ; wk(10)=z3L
michael@0 619 movdqa XMMWORD [wk(11)], xmm7 ; wk(11)=z3H
michael@0 620
michael@0 621 ; (Original)
michael@0 622 ; z1 = tmp0 + tmp3; z2 = tmp1 + tmp2;
michael@0 623 ; tmp0 = tmp0 * 0.298631336; tmp1 = tmp1 * 2.053119869;
michael@0 624 ; tmp2 = tmp2 * 3.072711026; tmp3 = tmp3 * 1.501321110;
michael@0 625 ; z1 = z1 * -0.899976223; z2 = z2 * -2.562915447;
michael@0 626 ; tmp0 += z1 + z3; tmp1 += z2 + z4;
michael@0 627 ; tmp2 += z2 + z3; tmp3 += z1 + z4;
michael@0 628 ;
michael@0 629 ; (This implementation)
michael@0 630 ; tmp0 = tmp0 * (0.298631336 - 0.899976223) + tmp3 * -0.899976223;
michael@0 631 ; tmp1 = tmp1 * (2.053119869 - 2.562915447) + tmp2 * -2.562915447;
michael@0 632 ; tmp2 = tmp1 * -2.562915447 + tmp2 * (3.072711026 - 2.562915447);
michael@0 633 ; tmp3 = tmp0 * -0.899976223 + tmp3 * (1.501321110 - 0.899976223);
michael@0 634 ; tmp0 += z3; tmp1 += z4;
michael@0 635 ; tmp2 += z3; tmp3 += z4;
michael@0 636
michael@0 637 movdqa xmm0,xmm1
michael@0 638 movdqa xmm7,xmm1
michael@0 639 punpcklwd xmm0,xmm3
michael@0 640 punpckhwd xmm7,xmm3
michael@0 641 movdqa xmm1,xmm0
michael@0 642 movdqa xmm3,xmm7
michael@0 643 pmaddwd xmm0,[rel PW_MF060_MF089] ; xmm0=tmp0L
michael@0 644 pmaddwd xmm7,[rel PW_MF060_MF089] ; xmm7=tmp0H
michael@0 645 pmaddwd xmm1,[rel PW_MF089_F060] ; xmm1=tmp3L
michael@0 646 pmaddwd xmm3,[rel PW_MF089_F060] ; xmm3=tmp3H
michael@0 647
michael@0 648 paddd xmm0, XMMWORD [wk(10)] ; xmm0=tmp0L
michael@0 649 paddd xmm7, XMMWORD [wk(11)] ; xmm7=tmp0H
michael@0 650 paddd xmm1,xmm5 ; xmm1=tmp3L
michael@0 651 paddd xmm3,xmm4 ; xmm3=tmp3H
michael@0 652
michael@0 653 movdqa XMMWORD [wk(8)], xmm0 ; wk(8)=tmp0L
michael@0 654 movdqa XMMWORD [wk(9)], xmm7 ; wk(9)=tmp0H
michael@0 655
michael@0 656 movdqa xmm0,xmm2
michael@0 657 movdqa xmm7,xmm2
michael@0 658 punpcklwd xmm0,xmm6
michael@0 659 punpckhwd xmm7,xmm6
michael@0 660 movdqa xmm2,xmm0
michael@0 661 movdqa xmm6,xmm7
michael@0 662 pmaddwd xmm0,[rel PW_MF050_MF256] ; xmm0=tmp1L
michael@0 663 pmaddwd xmm7,[rel PW_MF050_MF256] ; xmm7=tmp1H
michael@0 664 pmaddwd xmm2,[rel PW_MF256_F050] ; xmm2=tmp2L
michael@0 665 pmaddwd xmm6,[rel PW_MF256_F050] ; xmm6=tmp2H
michael@0 666
michael@0 667 paddd xmm0,xmm5 ; xmm0=tmp1L
michael@0 668 paddd xmm7,xmm4 ; xmm7=tmp1H
michael@0 669 paddd xmm2, XMMWORD [wk(10)] ; xmm2=tmp2L
michael@0 670 paddd xmm6, XMMWORD [wk(11)] ; xmm6=tmp2H
michael@0 671
michael@0 672 movdqa XMMWORD [wk(10)], xmm0 ; wk(10)=tmp1L
michael@0 673 movdqa XMMWORD [wk(11)], xmm7 ; wk(11)=tmp1H
michael@0 674
michael@0 675 ; -- Final output stage
michael@0 676
michael@0 677 movdqa xmm5, XMMWORD [wk(0)] ; xmm5=tmp10L
michael@0 678 movdqa xmm4, XMMWORD [wk(1)] ; xmm4=tmp10H
michael@0 679
michael@0 680 movdqa xmm0,xmm5
michael@0 681 movdqa xmm7,xmm4
michael@0 682 paddd xmm5,xmm1 ; xmm5=data0L
michael@0 683 paddd xmm4,xmm3 ; xmm4=data0H
michael@0 684 psubd xmm0,xmm1 ; xmm0=data7L
michael@0 685 psubd xmm7,xmm3 ; xmm7=data7H
michael@0 686
michael@0 687 movdqa xmm1,[rel PD_DESCALE_P2] ; xmm1=[rel PD_DESCALE_P2]
michael@0 688
michael@0 689 paddd xmm5,xmm1
michael@0 690 paddd xmm4,xmm1
michael@0 691 psrad xmm5,DESCALE_P2
michael@0 692 psrad xmm4,DESCALE_P2
michael@0 693 paddd xmm0,xmm1
michael@0 694 paddd xmm7,xmm1
michael@0 695 psrad xmm0,DESCALE_P2
michael@0 696 psrad xmm7,DESCALE_P2
michael@0 697
michael@0 698 packssdw xmm5,xmm4 ; xmm5=data0=(00 10 20 30 40 50 60 70)
michael@0 699 packssdw xmm0,xmm7 ; xmm0=data7=(07 17 27 37 47 57 67 77)
michael@0 700
michael@0 701 movdqa xmm3, XMMWORD [wk(4)] ; xmm3=tmp11L
michael@0 702 movdqa xmm1, XMMWORD [wk(5)] ; xmm1=tmp11H
michael@0 703
michael@0 704 movdqa xmm4,xmm3
michael@0 705 movdqa xmm7,xmm1
michael@0 706 paddd xmm3,xmm2 ; xmm3=data1L
michael@0 707 paddd xmm1,xmm6 ; xmm1=data1H
michael@0 708 psubd xmm4,xmm2 ; xmm4=data6L
michael@0 709 psubd xmm7,xmm6 ; xmm7=data6H
michael@0 710
michael@0 711 movdqa xmm2,[rel PD_DESCALE_P2] ; xmm2=[rel PD_DESCALE_P2]
michael@0 712
michael@0 713 paddd xmm3,xmm2
michael@0 714 paddd xmm1,xmm2
michael@0 715 psrad xmm3,DESCALE_P2
michael@0 716 psrad xmm1,DESCALE_P2
michael@0 717 paddd xmm4,xmm2
michael@0 718 paddd xmm7,xmm2
michael@0 719 psrad xmm4,DESCALE_P2
michael@0 720 psrad xmm7,DESCALE_P2
michael@0 721
michael@0 722 packssdw xmm3,xmm1 ; xmm3=data1=(01 11 21 31 41 51 61 71)
michael@0 723 packssdw xmm4,xmm7 ; xmm4=data6=(06 16 26 36 46 56 66 76)
michael@0 724
michael@0 725 packsswb xmm5,xmm4 ; xmm5=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76)
michael@0 726 packsswb xmm3,xmm0 ; xmm3=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77)
michael@0 727
michael@0 728 movdqa xmm6, XMMWORD [wk(6)] ; xmm6=tmp12L
michael@0 729 movdqa xmm2, XMMWORD [wk(7)] ; xmm2=tmp12H
michael@0 730 movdqa xmm1, XMMWORD [wk(10)] ; xmm1=tmp1L
michael@0 731 movdqa xmm7, XMMWORD [wk(11)] ; xmm7=tmp1H
michael@0 732
michael@0 733 movdqa XMMWORD [wk(0)], xmm5 ; wk(0)=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76)
michael@0 734 movdqa XMMWORD [wk(1)], xmm3 ; wk(1)=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77)
michael@0 735
michael@0 736 movdqa xmm4,xmm6
michael@0 737 movdqa xmm0,xmm2
michael@0 738 paddd xmm6,xmm1 ; xmm6=data2L
michael@0 739 paddd xmm2,xmm7 ; xmm2=data2H
michael@0 740 psubd xmm4,xmm1 ; xmm4=data5L
michael@0 741 psubd xmm0,xmm7 ; xmm0=data5H
michael@0 742
michael@0 743 movdqa xmm5,[rel PD_DESCALE_P2] ; xmm5=[rel PD_DESCALE_P2]
michael@0 744
michael@0 745 paddd xmm6,xmm5
michael@0 746 paddd xmm2,xmm5
michael@0 747 psrad xmm6,DESCALE_P2
michael@0 748 psrad xmm2,DESCALE_P2
michael@0 749 paddd xmm4,xmm5
michael@0 750 paddd xmm0,xmm5
michael@0 751 psrad xmm4,DESCALE_P2
michael@0 752 psrad xmm0,DESCALE_P2
michael@0 753
michael@0 754 packssdw xmm6,xmm2 ; xmm6=data2=(02 12 22 32 42 52 62 72)
michael@0 755 packssdw xmm4,xmm0 ; xmm4=data5=(05 15 25 35 45 55 65 75)
michael@0 756
michael@0 757 movdqa xmm3, XMMWORD [wk(2)] ; xmm3=tmp13L
michael@0 758 movdqa xmm1, XMMWORD [wk(3)] ; xmm1=tmp13H
michael@0 759 movdqa xmm7, XMMWORD [wk(8)] ; xmm7=tmp0L
michael@0 760 movdqa xmm5, XMMWORD [wk(9)] ; xmm5=tmp0H
michael@0 761
michael@0 762 movdqa xmm2,xmm3
michael@0 763 movdqa xmm0,xmm1
michael@0 764 paddd xmm3,xmm7 ; xmm3=data3L
michael@0 765 paddd xmm1,xmm5 ; xmm1=data3H
michael@0 766 psubd xmm2,xmm7 ; xmm2=data4L
michael@0 767 psubd xmm0,xmm5 ; xmm0=data4H
michael@0 768
michael@0 769 movdqa xmm7,[rel PD_DESCALE_P2] ; xmm7=[rel PD_DESCALE_P2]
michael@0 770
michael@0 771 paddd xmm3,xmm7
michael@0 772 paddd xmm1,xmm7
michael@0 773 psrad xmm3,DESCALE_P2
michael@0 774 psrad xmm1,DESCALE_P2
michael@0 775 paddd xmm2,xmm7
michael@0 776 paddd xmm0,xmm7
michael@0 777 psrad xmm2,DESCALE_P2
michael@0 778 psrad xmm0,DESCALE_P2
michael@0 779
michael@0 780 movdqa xmm5,[rel PB_CENTERJSAMP] ; xmm5=[rel PB_CENTERJSAMP]
michael@0 781
michael@0 782 packssdw xmm3,xmm1 ; xmm3=data3=(03 13 23 33 43 53 63 73)
michael@0 783 packssdw xmm2,xmm0 ; xmm2=data4=(04 14 24 34 44 54 64 74)
michael@0 784
michael@0 785 movdqa xmm7, XMMWORD [wk(0)] ; xmm7=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76)
michael@0 786 movdqa xmm1, XMMWORD [wk(1)] ; xmm1=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77)
michael@0 787
michael@0 788 packsswb xmm6,xmm2 ; xmm6=(02 12 22 32 42 52 62 72 04 14 24 34 44 54 64 74)
michael@0 789 packsswb xmm3,xmm4 ; xmm3=(03 13 23 33 43 53 63 73 05 15 25 35 45 55 65 75)
michael@0 790
michael@0 791 paddb xmm7,xmm5
michael@0 792 paddb xmm1,xmm5
michael@0 793 paddb xmm6,xmm5
michael@0 794 paddb xmm3,xmm5
michael@0 795
michael@0 796 movdqa xmm0,xmm7 ; transpose coefficients(phase 1)
michael@0 797 punpcklbw xmm7,xmm1 ; xmm7=(00 01 10 11 20 21 30 31 40 41 50 51 60 61 70 71)
michael@0 798 punpckhbw xmm0,xmm1 ; xmm0=(06 07 16 17 26 27 36 37 46 47 56 57 66 67 76 77)
michael@0 799 movdqa xmm2,xmm6 ; transpose coefficients(phase 1)
michael@0 800 punpcklbw xmm6,xmm3 ; xmm6=(02 03 12 13 22 23 32 33 42 43 52 53 62 63 72 73)
michael@0 801 punpckhbw xmm2,xmm3 ; xmm2=(04 05 14 15 24 25 34 35 44 45 54 55 64 65 74 75)
michael@0 802
michael@0 803 movdqa xmm4,xmm7 ; transpose coefficients(phase 2)
michael@0 804 punpcklwd xmm7,xmm6 ; xmm7=(00 01 02 03 10 11 12 13 20 21 22 23 30 31 32 33)
michael@0 805 punpckhwd xmm4,xmm6 ; xmm4=(40 41 42 43 50 51 52 53 60 61 62 63 70 71 72 73)
michael@0 806 movdqa xmm5,xmm2 ; transpose coefficients(phase 2)
michael@0 807 punpcklwd xmm2,xmm0 ; xmm2=(04 05 06 07 14 15 16 17 24 25 26 27 34 35 36 37)
michael@0 808 punpckhwd xmm5,xmm0 ; xmm5=(44 45 46 47 54 55 56 57 64 65 66 67 74 75 76 77)
michael@0 809
michael@0 810 movdqa xmm1,xmm7 ; transpose coefficients(phase 3)
michael@0 811 punpckldq xmm7,xmm2 ; xmm7=(00 01 02 03 04 05 06 07 10 11 12 13 14 15 16 17)
michael@0 812 punpckhdq xmm1,xmm2 ; xmm1=(20 21 22 23 24 25 26 27 30 31 32 33 34 35 36 37)
michael@0 813 movdqa xmm3,xmm4 ; transpose coefficients(phase 3)
michael@0 814 punpckldq xmm4,xmm5 ; xmm4=(40 41 42 43 44 45 46 47 50 51 52 53 54 55 56 57)
michael@0 815 punpckhdq xmm3,xmm5 ; xmm3=(60 61 62 63 64 65 66 67 70 71 72 73 74 75 76 77)
michael@0 816
michael@0 817 pshufd xmm6,xmm7,0x4E ; xmm6=(10 11 12 13 14 15 16 17 00 01 02 03 04 05 06 07)
michael@0 818 pshufd xmm0,xmm1,0x4E ; xmm0=(30 31 32 33 34 35 36 37 20 21 22 23 24 25 26 27)
michael@0 819 pshufd xmm2,xmm4,0x4E ; xmm2=(50 51 52 53 54 55 56 57 40 41 42 43 44 45 46 47)
michael@0 820 pshufd xmm5,xmm3,0x4E ; xmm5=(70 71 72 73 74 75 76 77 60 61 62 63 64 65 66 67)
michael@0 821
michael@0 822 mov rdx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
michael@0 823 mov rsi, JSAMPROW [rdi+2*SIZEOF_JSAMPROW]
michael@0 824 movq XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm7
michael@0 825 movq XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm1
michael@0 826 mov rdx, JSAMPROW [rdi+4*SIZEOF_JSAMPROW]
michael@0 827 mov rsi, JSAMPROW [rdi+6*SIZEOF_JSAMPROW]
michael@0 828 movq XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm4
michael@0 829 movq XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm3
michael@0 830
michael@0 831 mov rdx, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
michael@0 832 mov rsi, JSAMPROW [rdi+3*SIZEOF_JSAMPROW]
michael@0 833 movq XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm6
michael@0 834 movq XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm0
michael@0 835 mov rdx, JSAMPROW [rdi+5*SIZEOF_JSAMPROW]
michael@0 836 mov rsi, JSAMPROW [rdi+7*SIZEOF_JSAMPROW]
michael@0 837 movq XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm2
michael@0 838 movq XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm5
michael@0 839
michael@0 840 uncollect_args
michael@0 841 mov rsp,rbp ; rsp <- aligned rbp
michael@0 842 pop rsp ; rsp <- original rbp
michael@0 843 pop rbp
michael@0 844 ret
michael@0 845
michael@0 846 ; For some reason, the OS X linker does not honor the request to align the
michael@0 847 ; segment unless we do this.
michael@0 848 align 16

Mercurial Repository: The Tor Browser

Europalab SCM Repository Server

mercurial