The Tor Browser / annotate

media/libjpeg/simd/jiss2flt-64.asm@6474c204b198 (annotated)

media/libjpeg/simd/jiss2flt-64.asm

Wed, 31 Dec 2014 06:09:35 +0100

author
Michael Schloh von Bennewitz <michael@schloh.com>
date
Wed, 31 Dec 2014 06:09:35 +0100
changeset 0
6474c204b198
permissions
-rw-r--r--

Cloned upstream origin tor-browser at tor-browser-31.3.0esr-4.5-1-build1
revision ID fc1c9ff7c1b2defdbc039f12214767608f46423f for hacking purpose.

michael@0 1 ;
michael@0 2 ; jiss2flt-64.asm - floating-point IDCT (64-bit SSE & 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 floating-point implementation of the inverse DCT
michael@0 19 ; (Discrete Cosine Transform). The following code is based directly on
michael@0 20 ; the IJG's original jidctflt.c; see the jidctflt.c for 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 %macro unpcklps2 2 ; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(0 1 4 5)
michael@0 30 shufps %1,%2,0x44
michael@0 31 %endmacro
michael@0 32
michael@0 33 %macro unpckhps2 2 ; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(2 3 6 7)
michael@0 34 shufps %1,%2,0xEE
michael@0 35 %endmacro
michael@0 36
michael@0 37 ; --------------------------------------------------------------------------
michael@0 38 SECTION SEG_CONST
michael@0 39
michael@0 40 alignz 16
michael@0 41 global EXTN(jconst_idct_float_sse2)
michael@0 42
michael@0 43 EXTN(jconst_idct_float_sse2):
michael@0 44
michael@0 45 PD_1_414 times 4 dd 1.414213562373095048801689
michael@0 46 PD_1_847 times 4 dd 1.847759065022573512256366
michael@0 47 PD_1_082 times 4 dd 1.082392200292393968799446
michael@0 48 PD_M2_613 times 4 dd -2.613125929752753055713286
michael@0 49 PD_RNDINT_MAGIC times 4 dd 100663296.0 ; (float)(0x00C00000 << 3)
michael@0 50 PB_CENTERJSAMP times 16 db CENTERJSAMPLE
michael@0 51
michael@0 52 alignz 16
michael@0 53
michael@0 54 ; --------------------------------------------------------------------------
michael@0 55 SECTION SEG_TEXT
michael@0 56 BITS 64
michael@0 57 ;
michael@0 58 ; Perform dequantization and inverse DCT on one block of coefficients.
michael@0 59 ;
michael@0 60 ; GLOBAL(void)
michael@0 61 ; jsimd_idct_float_sse2 (void * dct_table, JCOEFPTR coef_block,
michael@0 62 ; JSAMPARRAY output_buf, JDIMENSION output_col)
michael@0 63 ;
michael@0 64
michael@0 65 ; r10 = void * dct_table
michael@0 66 ; r11 = JCOEFPTR coef_block
michael@0 67 ; r12 = JSAMPARRAY output_buf
michael@0 68 ; r13 = JDIMENSION output_col
michael@0 69
michael@0 70 %define original_rbp rbp+0
michael@0 71 %define wk(i) rbp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM]
michael@0 72 %define WK_NUM 2
michael@0 73 %define workspace wk(0)-DCTSIZE2*SIZEOF_FAST_FLOAT
michael@0 74 ; FAST_FLOAT workspace[DCTSIZE2]
michael@0 75
michael@0 76 align 16
michael@0 77 global EXTN(jsimd_idct_float_sse2)
michael@0 78
michael@0 79 EXTN(jsimd_idct_float_sse2):
michael@0 80 push rbp
michael@0 81 mov rax,rsp ; rax = original rbp
michael@0 82 sub rsp, byte 4
michael@0 83 and rsp, byte (-SIZEOF_XMMWORD) ; align to 128 bits
michael@0 84 mov [rsp],rax
michael@0 85 mov rbp,rsp ; rbp = aligned rbp
michael@0 86 lea rsp, [workspace]
michael@0 87 collect_args
michael@0 88 push rbx
michael@0 89
michael@0 90 ; ---- Pass 1: process columns from input, store into work array.
michael@0 91
michael@0 92 mov rdx, r10 ; quantptr
michael@0 93 mov rsi, r11 ; inptr
michael@0 94 lea rdi, [workspace] ; FAST_FLOAT * wsptr
michael@0 95 mov rcx, DCTSIZE/4 ; ctr
michael@0 96 .columnloop:
michael@0 97 %ifndef NO_ZERO_COLUMN_TEST_FLOAT_SSE
michael@0 98 mov eax, DWORD [DWBLOCK(1,0,rsi,SIZEOF_JCOEF)]
michael@0 99 or eax, DWORD [DWBLOCK(2,0,rsi,SIZEOF_JCOEF)]
michael@0 100 jnz near .columnDCT
michael@0 101
michael@0 102 movq xmm1, XMM_MMWORD [MMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
michael@0 103 movq xmm2, XMM_MMWORD [MMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
michael@0 104 movq xmm3, XMM_MMWORD [MMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
michael@0 105 movq xmm4, XMM_MMWORD [MMBLOCK(4,0,rsi,SIZEOF_JCOEF)]
michael@0 106 movq xmm5, XMM_MMWORD [MMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
michael@0 107 movq xmm6, XMM_MMWORD [MMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
michael@0 108 movq xmm7, XMM_MMWORD [MMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
michael@0 109 por xmm1,xmm2
michael@0 110 por xmm3,xmm4
michael@0 111 por xmm5,xmm6
michael@0 112 por xmm1,xmm3
michael@0 113 por xmm5,xmm7
michael@0 114 por xmm1,xmm5
michael@0 115 packsswb xmm1,xmm1
michael@0 116 movd eax,xmm1
michael@0 117 test rax,rax
michael@0 118 jnz short .columnDCT
michael@0 119
michael@0 120 ; -- AC terms all zero
michael@0 121
michael@0 122 movq xmm0, XMM_MMWORD [MMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
michael@0 123
michael@0 124 punpcklwd xmm0,xmm0 ; xmm0=(00 00 01 01 02 02 03 03)
michael@0 125 psrad xmm0,(DWORD_BIT-WORD_BIT) ; xmm0=in0=(00 01 02 03)
michael@0 126 cvtdq2ps xmm0,xmm0 ; xmm0=in0=(00 01 02 03)
michael@0 127
michael@0 128 mulps xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FLOAT_MULT_TYPE)]
michael@0 129
michael@0 130 movaps xmm1,xmm0
michael@0 131 movaps xmm2,xmm0
michael@0 132 movaps xmm3,xmm0
michael@0 133
michael@0 134 shufps xmm0,xmm0,0x00 ; xmm0=(00 00 00 00)
michael@0 135 shufps xmm1,xmm1,0x55 ; xmm1=(01 01 01 01)
michael@0 136 shufps xmm2,xmm2,0xAA ; xmm2=(02 02 02 02)
michael@0 137 shufps xmm3,xmm3,0xFF ; xmm3=(03 03 03 03)
michael@0 138
michael@0 139 movaps XMMWORD [XMMBLOCK(0,0,rdi,SIZEOF_FAST_FLOAT)], xmm0
michael@0 140 movaps XMMWORD [XMMBLOCK(0,1,rdi,SIZEOF_FAST_FLOAT)], xmm0
michael@0 141 movaps XMMWORD [XMMBLOCK(1,0,rdi,SIZEOF_FAST_FLOAT)], xmm1
michael@0 142 movaps XMMWORD [XMMBLOCK(1,1,rdi,SIZEOF_FAST_FLOAT)], xmm1
michael@0 143 movaps XMMWORD [XMMBLOCK(2,0,rdi,SIZEOF_FAST_FLOAT)], xmm2
michael@0 144 movaps XMMWORD [XMMBLOCK(2,1,rdi,SIZEOF_FAST_FLOAT)], xmm2
michael@0 145 movaps XMMWORD [XMMBLOCK(3,0,rdi,SIZEOF_FAST_FLOAT)], xmm3
michael@0 146 movaps XMMWORD [XMMBLOCK(3,1,rdi,SIZEOF_FAST_FLOAT)], xmm3
michael@0 147 jmp near .nextcolumn
michael@0 148 %endif
michael@0 149 .columnDCT:
michael@0 150
michael@0 151 ; -- Even part
michael@0 152
michael@0 153 movq xmm0, XMM_MMWORD [MMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
michael@0 154 movq xmm1, XMM_MMWORD [MMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
michael@0 155 movq xmm2, XMM_MMWORD [MMBLOCK(4,0,rsi,SIZEOF_JCOEF)]
michael@0 156 movq xmm3, XMM_MMWORD [MMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
michael@0 157
michael@0 158 punpcklwd xmm0,xmm0 ; xmm0=(00 00 01 01 02 02 03 03)
michael@0 159 punpcklwd xmm1,xmm1 ; xmm1=(20 20 21 21 22 22 23 23)
michael@0 160 psrad xmm0,(DWORD_BIT-WORD_BIT) ; xmm0=in0=(00 01 02 03)
michael@0 161 psrad xmm1,(DWORD_BIT-WORD_BIT) ; xmm1=in2=(20 21 22 23)
michael@0 162 cvtdq2ps xmm0,xmm0 ; xmm0=in0=(00 01 02 03)
michael@0 163 cvtdq2ps xmm1,xmm1 ; xmm1=in2=(20 21 22 23)
michael@0 164
michael@0 165 punpcklwd xmm2,xmm2 ; xmm2=(40 40 41 41 42 42 43 43)
michael@0 166 punpcklwd xmm3,xmm3 ; xmm3=(60 60 61 61 62 62 63 63)
michael@0 167 psrad xmm2,(DWORD_BIT-WORD_BIT) ; xmm2=in4=(40 41 42 43)
michael@0 168 psrad xmm3,(DWORD_BIT-WORD_BIT) ; xmm3=in6=(60 61 62 63)
michael@0 169 cvtdq2ps xmm2,xmm2 ; xmm2=in4=(40 41 42 43)
michael@0 170 cvtdq2ps xmm3,xmm3 ; xmm3=in6=(60 61 62 63)
michael@0 171
michael@0 172 mulps xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FLOAT_MULT_TYPE)]
michael@0 173 mulps xmm1, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_FLOAT_MULT_TYPE)]
michael@0 174 mulps xmm2, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_FLOAT_MULT_TYPE)]
michael@0 175 mulps xmm3, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_FLOAT_MULT_TYPE)]
michael@0 176
michael@0 177 movaps xmm4,xmm0
michael@0 178 movaps xmm5,xmm1
michael@0 179 subps xmm0,xmm2 ; xmm0=tmp11
michael@0 180 subps xmm1,xmm3
michael@0 181 addps xmm4,xmm2 ; xmm4=tmp10
michael@0 182 addps xmm5,xmm3 ; xmm5=tmp13
michael@0 183
michael@0 184 mulps xmm1,[rel PD_1_414]
michael@0 185 subps xmm1,xmm5 ; xmm1=tmp12
michael@0 186
michael@0 187 movaps xmm6,xmm4
michael@0 188 movaps xmm7,xmm0
michael@0 189 subps xmm4,xmm5 ; xmm4=tmp3
michael@0 190 subps xmm0,xmm1 ; xmm0=tmp2
michael@0 191 addps xmm6,xmm5 ; xmm6=tmp0
michael@0 192 addps xmm7,xmm1 ; xmm7=tmp1
michael@0 193
michael@0 194 movaps XMMWORD [wk(1)], xmm4 ; tmp3
michael@0 195 movaps XMMWORD [wk(0)], xmm0 ; tmp2
michael@0 196
michael@0 197 ; -- Odd part
michael@0 198
michael@0 199 movq xmm2, XMM_MMWORD [MMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
michael@0 200 movq xmm3, XMM_MMWORD [MMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
michael@0 201 movq xmm5, XMM_MMWORD [MMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
michael@0 202 movq xmm1, XMM_MMWORD [MMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
michael@0 203
michael@0 204 punpcklwd xmm2,xmm2 ; xmm2=(10 10 11 11 12 12 13 13)
michael@0 205 punpcklwd xmm3,xmm3 ; xmm3=(30 30 31 31 32 32 33 33)
michael@0 206 psrad xmm2,(DWORD_BIT-WORD_BIT) ; xmm2=in1=(10 11 12 13)
michael@0 207 psrad xmm3,(DWORD_BIT-WORD_BIT) ; xmm3=in3=(30 31 32 33)
michael@0 208 cvtdq2ps xmm2,xmm2 ; xmm2=in1=(10 11 12 13)
michael@0 209 cvtdq2ps xmm3,xmm3 ; xmm3=in3=(30 31 32 33)
michael@0 210
michael@0 211 punpcklwd xmm5,xmm5 ; xmm5=(50 50 51 51 52 52 53 53)
michael@0 212 punpcklwd xmm1,xmm1 ; xmm1=(70 70 71 71 72 72 73 73)
michael@0 213 psrad xmm5,(DWORD_BIT-WORD_BIT) ; xmm5=in5=(50 51 52 53)
michael@0 214 psrad xmm1,(DWORD_BIT-WORD_BIT) ; xmm1=in7=(70 71 72 73)
michael@0 215 cvtdq2ps xmm5,xmm5 ; xmm5=in5=(50 51 52 53)
michael@0 216 cvtdq2ps xmm1,xmm1 ; xmm1=in7=(70 71 72 73)
michael@0 217
michael@0 218 mulps xmm2, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FLOAT_MULT_TYPE)]
michael@0 219 mulps xmm3, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_FLOAT_MULT_TYPE)]
michael@0 220 mulps xmm5, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_FLOAT_MULT_TYPE)]
michael@0 221 mulps xmm1, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_FLOAT_MULT_TYPE)]
michael@0 222
michael@0 223 movaps xmm4,xmm2
michael@0 224 movaps xmm0,xmm5
michael@0 225 addps xmm2,xmm1 ; xmm2=z11
michael@0 226 addps xmm5,xmm3 ; xmm5=z13
michael@0 227 subps xmm4,xmm1 ; xmm4=z12
michael@0 228 subps xmm0,xmm3 ; xmm0=z10
michael@0 229
michael@0 230 movaps xmm1,xmm2
michael@0 231 subps xmm2,xmm5
michael@0 232 addps xmm1,xmm5 ; xmm1=tmp7
michael@0 233
michael@0 234 mulps xmm2,[rel PD_1_414] ; xmm2=tmp11
michael@0 235
michael@0 236 movaps xmm3,xmm0
michael@0 237 addps xmm0,xmm4
michael@0 238 mulps xmm0,[rel PD_1_847] ; xmm0=z5
michael@0 239 mulps xmm3,[rel PD_M2_613] ; xmm3=(z10 * -2.613125930)
michael@0 240 mulps xmm4,[rel PD_1_082] ; xmm4=(z12 * 1.082392200)
michael@0 241 addps xmm3,xmm0 ; xmm3=tmp12
michael@0 242 subps xmm4,xmm0 ; xmm4=tmp10
michael@0 243
michael@0 244 ; -- Final output stage
michael@0 245
michael@0 246 subps xmm3,xmm1 ; xmm3=tmp6
michael@0 247 movaps xmm5,xmm6
michael@0 248 movaps xmm0,xmm7
michael@0 249 addps xmm6,xmm1 ; xmm6=data0=(00 01 02 03)
michael@0 250 addps xmm7,xmm3 ; xmm7=data1=(10 11 12 13)
michael@0 251 subps xmm5,xmm1 ; xmm5=data7=(70 71 72 73)
michael@0 252 subps xmm0,xmm3 ; xmm0=data6=(60 61 62 63)
michael@0 253 subps xmm2,xmm3 ; xmm2=tmp5
michael@0 254
michael@0 255 movaps xmm1,xmm6 ; transpose coefficients(phase 1)
michael@0 256 unpcklps xmm6,xmm7 ; xmm6=(00 10 01 11)
michael@0 257 unpckhps xmm1,xmm7 ; xmm1=(02 12 03 13)
michael@0 258 movaps xmm3,xmm0 ; transpose coefficients(phase 1)
michael@0 259 unpcklps xmm0,xmm5 ; xmm0=(60 70 61 71)
michael@0 260 unpckhps xmm3,xmm5 ; xmm3=(62 72 63 73)
michael@0 261
michael@0 262 movaps xmm7, XMMWORD [wk(0)] ; xmm7=tmp2
michael@0 263 movaps xmm5, XMMWORD [wk(1)] ; xmm5=tmp3
michael@0 264
michael@0 265 movaps XMMWORD [wk(0)], xmm0 ; wk(0)=(60 70 61 71)
michael@0 266 movaps XMMWORD [wk(1)], xmm3 ; wk(1)=(62 72 63 73)
michael@0 267
michael@0 268 addps xmm4,xmm2 ; xmm4=tmp4
michael@0 269 movaps xmm0,xmm7
michael@0 270 movaps xmm3,xmm5
michael@0 271 addps xmm7,xmm2 ; xmm7=data2=(20 21 22 23)
michael@0 272 addps xmm5,xmm4 ; xmm5=data4=(40 41 42 43)
michael@0 273 subps xmm0,xmm2 ; xmm0=data5=(50 51 52 53)
michael@0 274 subps xmm3,xmm4 ; xmm3=data3=(30 31 32 33)
michael@0 275
michael@0 276 movaps xmm2,xmm7 ; transpose coefficients(phase 1)
michael@0 277 unpcklps xmm7,xmm3 ; xmm7=(20 30 21 31)
michael@0 278 unpckhps xmm2,xmm3 ; xmm2=(22 32 23 33)
michael@0 279 movaps xmm4,xmm5 ; transpose coefficients(phase 1)
michael@0 280 unpcklps xmm5,xmm0 ; xmm5=(40 50 41 51)
michael@0 281 unpckhps xmm4,xmm0 ; xmm4=(42 52 43 53)
michael@0 282
michael@0 283 movaps xmm3,xmm6 ; transpose coefficients(phase 2)
michael@0 284 unpcklps2 xmm6,xmm7 ; xmm6=(00 10 20 30)
michael@0 285 unpckhps2 xmm3,xmm7 ; xmm3=(01 11 21 31)
michael@0 286 movaps xmm0,xmm1 ; transpose coefficients(phase 2)
michael@0 287 unpcklps2 xmm1,xmm2 ; xmm1=(02 12 22 32)
michael@0 288 unpckhps2 xmm0,xmm2 ; xmm0=(03 13 23 33)
michael@0 289
michael@0 290 movaps xmm7, XMMWORD [wk(0)] ; xmm7=(60 70 61 71)
michael@0 291 movaps xmm2, XMMWORD [wk(1)] ; xmm2=(62 72 63 73)
michael@0 292
michael@0 293 movaps XMMWORD [XMMBLOCK(0,0,rdi,SIZEOF_FAST_FLOAT)], xmm6
michael@0 294 movaps XMMWORD [XMMBLOCK(1,0,rdi,SIZEOF_FAST_FLOAT)], xmm3
michael@0 295 movaps XMMWORD [XMMBLOCK(2,0,rdi,SIZEOF_FAST_FLOAT)], xmm1
michael@0 296 movaps XMMWORD [XMMBLOCK(3,0,rdi,SIZEOF_FAST_FLOAT)], xmm0
michael@0 297
michael@0 298 movaps xmm6,xmm5 ; transpose coefficients(phase 2)
michael@0 299 unpcklps2 xmm5,xmm7 ; xmm5=(40 50 60 70)
michael@0 300 unpckhps2 xmm6,xmm7 ; xmm6=(41 51 61 71)
michael@0 301 movaps xmm3,xmm4 ; transpose coefficients(phase 2)
michael@0 302 unpcklps2 xmm4,xmm2 ; xmm4=(42 52 62 72)
michael@0 303 unpckhps2 xmm3,xmm2 ; xmm3=(43 53 63 73)
michael@0 304
michael@0 305 movaps XMMWORD [XMMBLOCK(0,1,rdi,SIZEOF_FAST_FLOAT)], xmm5
michael@0 306 movaps XMMWORD [XMMBLOCK(1,1,rdi,SIZEOF_FAST_FLOAT)], xmm6
michael@0 307 movaps XMMWORD [XMMBLOCK(2,1,rdi,SIZEOF_FAST_FLOAT)], xmm4
michael@0 308 movaps XMMWORD [XMMBLOCK(3,1,rdi,SIZEOF_FAST_FLOAT)], xmm3
michael@0 309
michael@0 310 .nextcolumn:
michael@0 311 add rsi, byte 4*SIZEOF_JCOEF ; coef_block
michael@0 312 add rdx, byte 4*SIZEOF_FLOAT_MULT_TYPE ; quantptr
michael@0 313 add rdi, 4*DCTSIZE*SIZEOF_FAST_FLOAT ; wsptr
michael@0 314 dec rcx ; ctr
michael@0 315 jnz near .columnloop
michael@0 316
michael@0 317 ; -- Prefetch the next coefficient block
michael@0 318
michael@0 319 prefetchnta [rsi + (DCTSIZE2-8)*SIZEOF_JCOEF + 0*32]
michael@0 320 prefetchnta [rsi + (DCTSIZE2-8)*SIZEOF_JCOEF + 1*32]
michael@0 321 prefetchnta [rsi + (DCTSIZE2-8)*SIZEOF_JCOEF + 2*32]
michael@0 322 prefetchnta [rsi + (DCTSIZE2-8)*SIZEOF_JCOEF + 3*32]
michael@0 323
michael@0 324 ; ---- Pass 2: process rows from work array, store into output array.
michael@0 325
michael@0 326 mov rax, [original_rbp]
michael@0 327 lea rsi, [workspace] ; FAST_FLOAT * wsptr
michael@0 328 mov rdi, r12 ; (JSAMPROW *)
michael@0 329 mov rax, r13
michael@0 330 mov rcx, DCTSIZE/4 ; ctr
michael@0 331 .rowloop:
michael@0 332
michael@0 333 ; -- Even part
michael@0 334
michael@0 335 movaps xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_FAST_FLOAT)]
michael@0 336 movaps xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_FAST_FLOAT)]
michael@0 337 movaps xmm2, XMMWORD [XMMBLOCK(4,0,rsi,SIZEOF_FAST_FLOAT)]
michael@0 338 movaps xmm3, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_FAST_FLOAT)]
michael@0 339
michael@0 340 movaps xmm4,xmm0
michael@0 341 movaps xmm5,xmm1
michael@0 342 subps xmm0,xmm2 ; xmm0=tmp11
michael@0 343 subps xmm1,xmm3
michael@0 344 addps xmm4,xmm2 ; xmm4=tmp10
michael@0 345 addps xmm5,xmm3 ; xmm5=tmp13
michael@0 346
michael@0 347 mulps xmm1,[rel PD_1_414]
michael@0 348 subps xmm1,xmm5 ; xmm1=tmp12
michael@0 349
michael@0 350 movaps xmm6,xmm4
michael@0 351 movaps xmm7,xmm0
michael@0 352 subps xmm4,xmm5 ; xmm4=tmp3
michael@0 353 subps xmm0,xmm1 ; xmm0=tmp2
michael@0 354 addps xmm6,xmm5 ; xmm6=tmp0
michael@0 355 addps xmm7,xmm1 ; xmm7=tmp1
michael@0 356
michael@0 357 movaps XMMWORD [wk(1)], xmm4 ; tmp3
michael@0 358 movaps XMMWORD [wk(0)], xmm0 ; tmp2
michael@0 359
michael@0 360 ; -- Odd part
michael@0 361
michael@0 362 movaps xmm2, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_FAST_FLOAT)]
michael@0 363 movaps xmm3, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_FAST_FLOAT)]
michael@0 364 movaps xmm5, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_FAST_FLOAT)]
michael@0 365 movaps xmm1, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_FAST_FLOAT)]
michael@0 366
michael@0 367 movaps xmm4,xmm2
michael@0 368 movaps xmm0,xmm5
michael@0 369 addps xmm2,xmm1 ; xmm2=z11
michael@0 370 addps xmm5,xmm3 ; xmm5=z13
michael@0 371 subps xmm4,xmm1 ; xmm4=z12
michael@0 372 subps xmm0,xmm3 ; xmm0=z10
michael@0 373
michael@0 374 movaps xmm1,xmm2
michael@0 375 subps xmm2,xmm5
michael@0 376 addps xmm1,xmm5 ; xmm1=tmp7
michael@0 377
michael@0 378 mulps xmm2,[rel PD_1_414] ; xmm2=tmp11
michael@0 379
michael@0 380 movaps xmm3,xmm0
michael@0 381 addps xmm0,xmm4
michael@0 382 mulps xmm0,[rel PD_1_847] ; xmm0=z5
michael@0 383 mulps xmm3,[rel PD_M2_613] ; xmm3=(z10 * -2.613125930)
michael@0 384 mulps xmm4,[rel PD_1_082] ; xmm4=(z12 * 1.082392200)
michael@0 385 addps xmm3,xmm0 ; xmm3=tmp12
michael@0 386 subps xmm4,xmm0 ; xmm4=tmp10
michael@0 387
michael@0 388 ; -- Final output stage
michael@0 389
michael@0 390 subps xmm3,xmm1 ; xmm3=tmp6
michael@0 391 movaps xmm5,xmm6
michael@0 392 movaps xmm0,xmm7
michael@0 393 addps xmm6,xmm1 ; xmm6=data0=(00 10 20 30)
michael@0 394 addps xmm7,xmm3 ; xmm7=data1=(01 11 21 31)
michael@0 395 subps xmm5,xmm1 ; xmm5=data7=(07 17 27 37)
michael@0 396 subps xmm0,xmm3 ; xmm0=data6=(06 16 26 36)
michael@0 397 subps xmm2,xmm3 ; xmm2=tmp5
michael@0 398
michael@0 399 movaps xmm1,[rel PD_RNDINT_MAGIC] ; xmm1=[rel PD_RNDINT_MAGIC]
michael@0 400 pcmpeqd xmm3,xmm3
michael@0 401 psrld xmm3,WORD_BIT ; xmm3={0xFFFF 0x0000 0xFFFF 0x0000 ..}
michael@0 402
michael@0 403 addps xmm6,xmm1 ; xmm6=roundint(data0/8)=(00 ** 10 ** 20 ** 30 **)
michael@0 404 addps xmm7,xmm1 ; xmm7=roundint(data1/8)=(01 ** 11 ** 21 ** 31 **)
michael@0 405 addps xmm0,xmm1 ; xmm0=roundint(data6/8)=(06 ** 16 ** 26 ** 36 **)
michael@0 406 addps xmm5,xmm1 ; xmm5=roundint(data7/8)=(07 ** 17 ** 27 ** 37 **)
michael@0 407
michael@0 408 pand xmm6,xmm3 ; xmm6=(00 -- 10 -- 20 -- 30 --)
michael@0 409 pslld xmm7,WORD_BIT ; xmm7=(-- 01 -- 11 -- 21 -- 31)
michael@0 410 pand xmm0,xmm3 ; xmm0=(06 -- 16 -- 26 -- 36 --)
michael@0 411 pslld xmm5,WORD_BIT ; xmm5=(-- 07 -- 17 -- 27 -- 37)
michael@0 412 por xmm6,xmm7 ; xmm6=(00 01 10 11 20 21 30 31)
michael@0 413 por xmm0,xmm5 ; xmm0=(06 07 16 17 26 27 36 37)
michael@0 414
michael@0 415 movaps xmm1, XMMWORD [wk(0)] ; xmm1=tmp2
michael@0 416 movaps xmm3, XMMWORD [wk(1)] ; xmm3=tmp3
michael@0 417
michael@0 418 addps xmm4,xmm2 ; xmm4=tmp4
michael@0 419 movaps xmm7,xmm1
michael@0 420 movaps xmm5,xmm3
michael@0 421 addps xmm1,xmm2 ; xmm1=data2=(02 12 22 32)
michael@0 422 addps xmm3,xmm4 ; xmm3=data4=(04 14 24 34)
michael@0 423 subps xmm7,xmm2 ; xmm7=data5=(05 15 25 35)
michael@0 424 subps xmm5,xmm4 ; xmm5=data3=(03 13 23 33)
michael@0 425
michael@0 426 movaps xmm2,[rel PD_RNDINT_MAGIC] ; xmm2=[rel PD_RNDINT_MAGIC]
michael@0 427 pcmpeqd xmm4,xmm4
michael@0 428 psrld xmm4,WORD_BIT ; xmm4={0xFFFF 0x0000 0xFFFF 0x0000 ..}
michael@0 429
michael@0 430 addps xmm3,xmm2 ; xmm3=roundint(data4/8)=(04 ** 14 ** 24 ** 34 **)
michael@0 431 addps xmm7,xmm2 ; xmm7=roundint(data5/8)=(05 ** 15 ** 25 ** 35 **)
michael@0 432 addps xmm1,xmm2 ; xmm1=roundint(data2/8)=(02 ** 12 ** 22 ** 32 **)
michael@0 433 addps xmm5,xmm2 ; xmm5=roundint(data3/8)=(03 ** 13 ** 23 ** 33 **)
michael@0 434
michael@0 435 pand xmm3,xmm4 ; xmm3=(04 -- 14 -- 24 -- 34 --)
michael@0 436 pslld xmm7,WORD_BIT ; xmm7=(-- 05 -- 15 -- 25 -- 35)
michael@0 437 pand xmm1,xmm4 ; xmm1=(02 -- 12 -- 22 -- 32 --)
michael@0 438 pslld xmm5,WORD_BIT ; xmm5=(-- 03 -- 13 -- 23 -- 33)
michael@0 439 por xmm3,xmm7 ; xmm3=(04 05 14 15 24 25 34 35)
michael@0 440 por xmm1,xmm5 ; xmm1=(02 03 12 13 22 23 32 33)
michael@0 441
michael@0 442 movdqa xmm2,[rel PB_CENTERJSAMP] ; xmm2=[rel PB_CENTERJSAMP]
michael@0 443
michael@0 444 packsswb xmm6,xmm3 ; xmm6=(00 01 10 11 20 21 30 31 04 05 14 15 24 25 34 35)
michael@0 445 packsswb xmm1,xmm0 ; xmm1=(02 03 12 13 22 23 32 33 06 07 16 17 26 27 36 37)
michael@0 446 paddb xmm6,xmm2
michael@0 447 paddb xmm1,xmm2
michael@0 448
michael@0 449 movdqa xmm4,xmm6 ; transpose coefficients(phase 2)
michael@0 450 punpcklwd xmm6,xmm1 ; xmm6=(00 01 02 03 10 11 12 13 20 21 22 23 30 31 32 33)
michael@0 451 punpckhwd xmm4,xmm1 ; xmm4=(04 05 06 07 14 15 16 17 24 25 26 27 34 35 36 37)
michael@0 452
michael@0 453 movdqa xmm7,xmm6 ; transpose coefficients(phase 3)
michael@0 454 punpckldq xmm6,xmm4 ; xmm6=(00 01 02 03 04 05 06 07 10 11 12 13 14 15 16 17)
michael@0 455 punpckhdq xmm7,xmm4 ; xmm7=(20 21 22 23 24 25 26 27 30 31 32 33 34 35 36 37)
michael@0 456
michael@0 457 pshufd xmm5,xmm6,0x4E ; xmm5=(10 11 12 13 14 15 16 17 00 01 02 03 04 05 06 07)
michael@0 458 pshufd xmm3,xmm7,0x4E ; xmm3=(30 31 32 33 34 35 36 37 20 21 22 23 24 25 26 27)
michael@0 459
michael@0 460 mov rdx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
michael@0 461 mov rbx, JSAMPROW [rdi+2*SIZEOF_JSAMPROW]
michael@0 462 movq XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm6
michael@0 463 movq XMM_MMWORD [rbx+rax*SIZEOF_JSAMPLE], xmm7
michael@0 464 mov rdx, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
michael@0 465 mov rbx, JSAMPROW [rdi+3*SIZEOF_JSAMPROW]
michael@0 466 movq XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm5
michael@0 467 movq XMM_MMWORD [rbx+rax*SIZEOF_JSAMPLE], xmm3
michael@0 468
michael@0 469 add rsi, byte 4*SIZEOF_FAST_FLOAT ; wsptr
michael@0 470 add rdi, byte 4*SIZEOF_JSAMPROW
michael@0 471 dec rcx ; ctr
michael@0 472 jnz near .rowloop
michael@0 473
michael@0 474 pop rbx
michael@0 475 uncollect_args
michael@0 476 mov rsp,rbp ; rsp <- aligned rbp
michael@0 477 pop rsp ; rsp <- original rbp
michael@0 478 pop rbp
michael@0 479 ret
michael@0 480
michael@0 481 ; For some reason, the OS X linker does not honor the request to align the
michael@0 482 ; segment unless we do this.
michael@0 483 align 16

Mercurial Repository: The Tor Browser

Europalab SCM Repository Server

mercurial