The Tor Browser / annotate

media/libjpeg/simd/jiss2red-64.asm@6474c204b198 (annotated)

media/libjpeg/simd/jiss2red-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 ; jiss2red-64.asm - reduced-size 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 inverse-DCT routines that produce reduced-size
michael@0 19 ; output: either 4x4 or 2x2 pixels from an 8x8 DCT block.
michael@0 20 ; The following code is based directly on the IJG's original jidctred.c;
michael@0 21 ; see the jidctred.c for 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_4 (CONST_BITS-PASS1_BITS+1)
michael@0 34 %define DESCALE_P2_4 (CONST_BITS+PASS1_BITS+3+1)
michael@0 35 %define DESCALE_P1_2 (CONST_BITS-PASS1_BITS+2)
michael@0 36 %define DESCALE_P2_2 (CONST_BITS+PASS1_BITS+3+2)
michael@0 37
michael@0 38 %if CONST_BITS == 13
michael@0 39 F_0_211 equ 1730 ; FIX(0.211164243)
michael@0 40 F_0_509 equ 4176 ; FIX(0.509795579)
michael@0 41 F_0_601 equ 4926 ; FIX(0.601344887)
michael@0 42 F_0_720 equ 5906 ; FIX(0.720959822)
michael@0 43 F_0_765 equ 6270 ; FIX(0.765366865)
michael@0 44 F_0_850 equ 6967 ; FIX(0.850430095)
michael@0 45 F_0_899 equ 7373 ; FIX(0.899976223)
michael@0 46 F_1_061 equ 8697 ; FIX(1.061594337)
michael@0 47 F_1_272 equ 10426 ; FIX(1.272758580)
michael@0 48 F_1_451 equ 11893 ; FIX(1.451774981)
michael@0 49 F_1_847 equ 15137 ; FIX(1.847759065)
michael@0 50 F_2_172 equ 17799 ; FIX(2.172734803)
michael@0 51 F_2_562 equ 20995 ; FIX(2.562915447)
michael@0 52 F_3_624 equ 29692 ; FIX(3.624509785)
michael@0 53 %else
michael@0 54 ; NASM cannot do compile-time arithmetic on floating-point constants.
michael@0 55 %define DESCALE(x,n) (((x)+(1<<((n)-1)))>>(n))
michael@0 56 F_0_211 equ DESCALE( 226735879,30-CONST_BITS) ; FIX(0.211164243)
michael@0 57 F_0_509 equ DESCALE( 547388834,30-CONST_BITS) ; FIX(0.509795579)
michael@0 58 F_0_601 equ DESCALE( 645689155,30-CONST_BITS) ; FIX(0.601344887)
michael@0 59 F_0_720 equ DESCALE( 774124714,30-CONST_BITS) ; FIX(0.720959822)
michael@0 60 F_0_765 equ DESCALE( 821806413,30-CONST_BITS) ; FIX(0.765366865)
michael@0 61 F_0_850 equ DESCALE( 913142361,30-CONST_BITS) ; FIX(0.850430095)
michael@0 62 F_0_899 equ DESCALE( 966342111,30-CONST_BITS) ; FIX(0.899976223)
michael@0 63 F_1_061 equ DESCALE(1139878239,30-CONST_BITS) ; FIX(1.061594337)
michael@0 64 F_1_272 equ DESCALE(1366614119,30-CONST_BITS) ; FIX(1.272758580)
michael@0 65 F_1_451 equ DESCALE(1558831516,30-CONST_BITS) ; FIX(1.451774981)
michael@0 66 F_1_847 equ DESCALE(1984016188,30-CONST_BITS) ; FIX(1.847759065)
michael@0 67 F_2_172 equ DESCALE(2332956230,30-CONST_BITS) ; FIX(2.172734803)
michael@0 68 F_2_562 equ DESCALE(2751909506,30-CONST_BITS) ; FIX(2.562915447)
michael@0 69 F_3_624 equ DESCALE(3891787747,30-CONST_BITS) ; FIX(3.624509785)
michael@0 70 %endif
michael@0 71
michael@0 72 ; --------------------------------------------------------------------------
michael@0 73 SECTION SEG_CONST
michael@0 74
michael@0 75 alignz 16
michael@0 76 global EXTN(jconst_idct_red_sse2)
michael@0 77
michael@0 78 EXTN(jconst_idct_red_sse2):
michael@0 79
michael@0 80 PW_F184_MF076 times 4 dw F_1_847,-F_0_765
michael@0 81 PW_F256_F089 times 4 dw F_2_562, F_0_899
michael@0 82 PW_F106_MF217 times 4 dw F_1_061,-F_2_172
michael@0 83 PW_MF060_MF050 times 4 dw -F_0_601,-F_0_509
michael@0 84 PW_F145_MF021 times 4 dw F_1_451,-F_0_211
michael@0 85 PW_F362_MF127 times 4 dw F_3_624,-F_1_272
michael@0 86 PW_F085_MF072 times 4 dw F_0_850,-F_0_720
michael@0 87 PD_DESCALE_P1_4 times 4 dd 1 << (DESCALE_P1_4-1)
michael@0 88 PD_DESCALE_P2_4 times 4 dd 1 << (DESCALE_P2_4-1)
michael@0 89 PD_DESCALE_P1_2 times 4 dd 1 << (DESCALE_P1_2-1)
michael@0 90 PD_DESCALE_P2_2 times 4 dd 1 << (DESCALE_P2_2-1)
michael@0 91 PB_CENTERJSAMP times 16 db CENTERJSAMPLE
michael@0 92
michael@0 93 alignz 16
michael@0 94
michael@0 95 ; --------------------------------------------------------------------------
michael@0 96 SECTION SEG_TEXT
michael@0 97 BITS 64
michael@0 98 ;
michael@0 99 ; Perform dequantization and inverse DCT on one block of coefficients,
michael@0 100 ; producing a reduced-size 4x4 output block.
michael@0 101 ;
michael@0 102 ; GLOBAL(void)
michael@0 103 ; jsimd_idct_4x4_sse2 (void * dct_table, JCOEFPTR coef_block,
michael@0 104 ; JSAMPARRAY output_buf, JDIMENSION output_col)
michael@0 105 ;
michael@0 106
michael@0 107 ; r10 = void * dct_table
michael@0 108 ; r11 = JCOEFPTR coef_block
michael@0 109 ; r12 = JSAMPARRAY output_buf
michael@0 110 ; r13 = JDIMENSION output_col
michael@0 111
michael@0 112 %define original_rbp rbp+0
michael@0 113 %define wk(i) rbp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM]
michael@0 114 %define WK_NUM 2
michael@0 115
michael@0 116 align 16
michael@0 117 global EXTN(jsimd_idct_4x4_sse2)
michael@0 118
michael@0 119 EXTN(jsimd_idct_4x4_sse2):
michael@0 120 push rbp
michael@0 121 mov rax,rsp ; rax = original rbp
michael@0 122 sub rsp, byte 4
michael@0 123 and rsp, byte (-SIZEOF_XMMWORD) ; align to 128 bits
michael@0 124 mov [rsp],rax
michael@0 125 mov rbp,rsp ; rbp = aligned rbp
michael@0 126 lea rsp, [wk(0)]
michael@0 127 collect_args
michael@0 128
michael@0 129 ; ---- Pass 1: process columns from input.
michael@0 130
michael@0 131 mov rdx, r10 ; quantptr
michael@0 132 mov rsi, r11 ; inptr
michael@0 133
michael@0 134 %ifndef NO_ZERO_COLUMN_TEST_4X4_SSE2
michael@0 135 mov eax, DWORD [DWBLOCK(1,0,rsi,SIZEOF_JCOEF)]
michael@0 136 or eax, DWORD [DWBLOCK(2,0,rsi,SIZEOF_JCOEF)]
michael@0 137 jnz short .columnDCT
michael@0 138
michael@0 139 movdqa xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
michael@0 140 movdqa xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
michael@0 141 por xmm0, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
michael@0 142 por xmm1, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
michael@0 143 por xmm0, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
michael@0 144 por xmm1, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
michael@0 145 por xmm0,xmm1
michael@0 146 packsswb xmm0,xmm0
michael@0 147 packsswb xmm0,xmm0
michael@0 148 movd eax,xmm0
michael@0 149 test rax,rax
michael@0 150 jnz short .columnDCT
michael@0 151
michael@0 152 ; -- AC terms all zero
michael@0 153
michael@0 154 movdqa xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
michael@0 155 pmullw xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 156
michael@0 157 psllw xmm0,PASS1_BITS
michael@0 158
michael@0 159 movdqa xmm3,xmm0 ; xmm0=in0=(00 01 02 03 04 05 06 07)
michael@0 160 punpcklwd xmm0,xmm0 ; xmm0=(00 00 01 01 02 02 03 03)
michael@0 161 punpckhwd xmm3,xmm3 ; xmm3=(04 04 05 05 06 06 07 07)
michael@0 162
michael@0 163 pshufd xmm1,xmm0,0x50 ; xmm1=[col0 col1]=(00 00 00 00 01 01 01 01)
michael@0 164 pshufd xmm0,xmm0,0xFA ; xmm0=[col2 col3]=(02 02 02 02 03 03 03 03)
michael@0 165 pshufd xmm6,xmm3,0x50 ; xmm6=[col4 col5]=(04 04 04 04 05 05 05 05)
michael@0 166 pshufd xmm3,xmm3,0xFA ; xmm3=[col6 col7]=(06 06 06 06 07 07 07 07)
michael@0 167
michael@0 168 jmp near .column_end
michael@0 169 %endif
michael@0 170 .columnDCT:
michael@0 171
michael@0 172 ; -- Odd part
michael@0 173
michael@0 174 movdqa xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
michael@0 175 movdqa xmm1, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
michael@0 176 pmullw xmm0, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 177 pmullw xmm1, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 178 movdqa xmm2, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
michael@0 179 movdqa xmm3, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
michael@0 180 pmullw xmm2, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 181 pmullw xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 182
michael@0 183 movdqa xmm4,xmm0
michael@0 184 movdqa xmm5,xmm0
michael@0 185 punpcklwd xmm4,xmm1
michael@0 186 punpckhwd xmm5,xmm1
michael@0 187 movdqa xmm0,xmm4
michael@0 188 movdqa xmm1,xmm5
michael@0 189 pmaddwd xmm4,[rel PW_F256_F089] ; xmm4=(tmp2L)
michael@0 190 pmaddwd xmm5,[rel PW_F256_F089] ; xmm5=(tmp2H)
michael@0 191 pmaddwd xmm0,[rel PW_F106_MF217] ; xmm0=(tmp0L)
michael@0 192 pmaddwd xmm1,[rel PW_F106_MF217] ; xmm1=(tmp0H)
michael@0 193
michael@0 194 movdqa xmm6,xmm2
michael@0 195 movdqa xmm7,xmm2
michael@0 196 punpcklwd xmm6,xmm3
michael@0 197 punpckhwd xmm7,xmm3
michael@0 198 movdqa xmm2,xmm6
michael@0 199 movdqa xmm3,xmm7
michael@0 200 pmaddwd xmm6,[rel PW_MF060_MF050] ; xmm6=(tmp2L)
michael@0 201 pmaddwd xmm7,[rel PW_MF060_MF050] ; xmm7=(tmp2H)
michael@0 202 pmaddwd xmm2,[rel PW_F145_MF021] ; xmm2=(tmp0L)
michael@0 203 pmaddwd xmm3,[rel PW_F145_MF021] ; xmm3=(tmp0H)
michael@0 204
michael@0 205 paddd xmm6,xmm4 ; xmm6=tmp2L
michael@0 206 paddd xmm7,xmm5 ; xmm7=tmp2H
michael@0 207 paddd xmm2,xmm0 ; xmm2=tmp0L
michael@0 208 paddd xmm3,xmm1 ; xmm3=tmp0H
michael@0 209
michael@0 210 movdqa XMMWORD [wk(0)], xmm2 ; wk(0)=tmp0L
michael@0 211 movdqa XMMWORD [wk(1)], xmm3 ; wk(1)=tmp0H
michael@0 212
michael@0 213 ; -- Even part
michael@0 214
michael@0 215 movdqa xmm4, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
michael@0 216 movdqa xmm5, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
michael@0 217 movdqa xmm0, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
michael@0 218 pmullw xmm4, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 219 pmullw xmm5, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 220 pmullw xmm0, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 221
michael@0 222 pxor xmm1,xmm1
michael@0 223 pxor xmm2,xmm2
michael@0 224 punpcklwd xmm1,xmm4 ; xmm1=tmp0L
michael@0 225 punpckhwd xmm2,xmm4 ; xmm2=tmp0H
michael@0 226 psrad xmm1,(16-CONST_BITS-1) ; psrad xmm1,16 & pslld xmm1,CONST_BITS+1
michael@0 227 psrad xmm2,(16-CONST_BITS-1) ; psrad xmm2,16 & pslld xmm2,CONST_BITS+1
michael@0 228
michael@0 229 movdqa xmm3,xmm5 ; xmm5=in2=z2
michael@0 230 punpcklwd xmm5,xmm0 ; xmm0=in6=z3
michael@0 231 punpckhwd xmm3,xmm0
michael@0 232 pmaddwd xmm5,[rel PW_F184_MF076] ; xmm5=tmp2L
michael@0 233 pmaddwd xmm3,[rel PW_F184_MF076] ; xmm3=tmp2H
michael@0 234
michael@0 235 movdqa xmm4,xmm1
michael@0 236 movdqa xmm0,xmm2
michael@0 237 paddd xmm1,xmm5 ; xmm1=tmp10L
michael@0 238 paddd xmm2,xmm3 ; xmm2=tmp10H
michael@0 239 psubd xmm4,xmm5 ; xmm4=tmp12L
michael@0 240 psubd xmm0,xmm3 ; xmm0=tmp12H
michael@0 241
michael@0 242 ; -- Final output stage
michael@0 243
michael@0 244 movdqa xmm5,xmm1
michael@0 245 movdqa xmm3,xmm2
michael@0 246 paddd xmm1,xmm6 ; xmm1=data0L
michael@0 247 paddd xmm2,xmm7 ; xmm2=data0H
michael@0 248 psubd xmm5,xmm6 ; xmm5=data3L
michael@0 249 psubd xmm3,xmm7 ; xmm3=data3H
michael@0 250
michael@0 251 movdqa xmm6,[rel PD_DESCALE_P1_4] ; xmm6=[rel PD_DESCALE_P1_4]
michael@0 252
michael@0 253 paddd xmm1,xmm6
michael@0 254 paddd xmm2,xmm6
michael@0 255 psrad xmm1,DESCALE_P1_4
michael@0 256 psrad xmm2,DESCALE_P1_4
michael@0 257 paddd xmm5,xmm6
michael@0 258 paddd xmm3,xmm6
michael@0 259 psrad xmm5,DESCALE_P1_4
michael@0 260 psrad xmm3,DESCALE_P1_4
michael@0 261
michael@0 262 packssdw xmm1,xmm2 ; xmm1=data0=(00 01 02 03 04 05 06 07)
michael@0 263 packssdw xmm5,xmm3 ; xmm5=data3=(30 31 32 33 34 35 36 37)
michael@0 264
michael@0 265 movdqa xmm7, XMMWORD [wk(0)] ; xmm7=tmp0L
michael@0 266 movdqa xmm6, XMMWORD [wk(1)] ; xmm6=tmp0H
michael@0 267
michael@0 268 movdqa xmm2,xmm4
michael@0 269 movdqa xmm3,xmm0
michael@0 270 paddd xmm4,xmm7 ; xmm4=data1L
michael@0 271 paddd xmm0,xmm6 ; xmm0=data1H
michael@0 272 psubd xmm2,xmm7 ; xmm2=data2L
michael@0 273 psubd xmm3,xmm6 ; xmm3=data2H
michael@0 274
michael@0 275 movdqa xmm7,[rel PD_DESCALE_P1_4] ; xmm7=[rel PD_DESCALE_P1_4]
michael@0 276
michael@0 277 paddd xmm4,xmm7
michael@0 278 paddd xmm0,xmm7
michael@0 279 psrad xmm4,DESCALE_P1_4
michael@0 280 psrad xmm0,DESCALE_P1_4
michael@0 281 paddd xmm2,xmm7
michael@0 282 paddd xmm3,xmm7
michael@0 283 psrad xmm2,DESCALE_P1_4
michael@0 284 psrad xmm3,DESCALE_P1_4
michael@0 285
michael@0 286 packssdw xmm4,xmm0 ; xmm4=data1=(10 11 12 13 14 15 16 17)
michael@0 287 packssdw xmm2,xmm3 ; xmm2=data2=(20 21 22 23 24 25 26 27)
michael@0 288
michael@0 289 movdqa xmm6,xmm1 ; transpose coefficients(phase 1)
michael@0 290 punpcklwd xmm1,xmm4 ; xmm1=(00 10 01 11 02 12 03 13)
michael@0 291 punpckhwd xmm6,xmm4 ; xmm6=(04 14 05 15 06 16 07 17)
michael@0 292 movdqa xmm7,xmm2 ; transpose coefficients(phase 1)
michael@0 293 punpcklwd xmm2,xmm5 ; xmm2=(20 30 21 31 22 32 23 33)
michael@0 294 punpckhwd xmm7,xmm5 ; xmm7=(24 34 25 35 26 36 27 37)
michael@0 295
michael@0 296 movdqa xmm0,xmm1 ; transpose coefficients(phase 2)
michael@0 297 punpckldq xmm1,xmm2 ; xmm1=[col0 col1]=(00 10 20 30 01 11 21 31)
michael@0 298 punpckhdq xmm0,xmm2 ; xmm0=[col2 col3]=(02 12 22 32 03 13 23 33)
michael@0 299 movdqa xmm3,xmm6 ; transpose coefficients(phase 2)
michael@0 300 punpckldq xmm6,xmm7 ; xmm6=[col4 col5]=(04 14 24 34 05 15 25 35)
michael@0 301 punpckhdq xmm3,xmm7 ; xmm3=[col6 col7]=(06 16 26 36 07 17 27 37)
michael@0 302 .column_end:
michael@0 303
michael@0 304 ; -- Prefetch the next coefficient block
michael@0 305
michael@0 306 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
michael@0 307 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
michael@0 308 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
michael@0 309 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
michael@0 310
michael@0 311 ; ---- Pass 2: process rows, store into output array.
michael@0 312
michael@0 313 mov rax, [original_rbp]
michael@0 314 mov rdi, r12 ; (JSAMPROW *)
michael@0 315 mov rax, r13
michael@0 316
michael@0 317 ; -- Even part
michael@0 318
michael@0 319 pxor xmm4,xmm4
michael@0 320 punpcklwd xmm4,xmm1 ; xmm4=tmp0
michael@0 321 psrad xmm4,(16-CONST_BITS-1) ; psrad xmm4,16 & pslld xmm4,CONST_BITS+1
michael@0 322
michael@0 323 ; -- Odd part
michael@0 324
michael@0 325 punpckhwd xmm1,xmm0
michael@0 326 punpckhwd xmm6,xmm3
michael@0 327 movdqa xmm5,xmm1
michael@0 328 movdqa xmm2,xmm6
michael@0 329 pmaddwd xmm1,[rel PW_F256_F089] ; xmm1=(tmp2)
michael@0 330 pmaddwd xmm6,[rel PW_MF060_MF050] ; xmm6=(tmp2)
michael@0 331 pmaddwd xmm5,[rel PW_F106_MF217] ; xmm5=(tmp0)
michael@0 332 pmaddwd xmm2,[rel PW_F145_MF021] ; xmm2=(tmp0)
michael@0 333
michael@0 334 paddd xmm6,xmm1 ; xmm6=tmp2
michael@0 335 paddd xmm2,xmm5 ; xmm2=tmp0
michael@0 336
michael@0 337 ; -- Even part
michael@0 338
michael@0 339 punpcklwd xmm0,xmm3
michael@0 340 pmaddwd xmm0,[rel PW_F184_MF076] ; xmm0=tmp2
michael@0 341
michael@0 342 movdqa xmm7,xmm4
michael@0 343 paddd xmm4,xmm0 ; xmm4=tmp10
michael@0 344 psubd xmm7,xmm0 ; xmm7=tmp12
michael@0 345
michael@0 346 ; -- Final output stage
michael@0 347
michael@0 348 movdqa xmm1,[rel PD_DESCALE_P2_4] ; xmm1=[rel PD_DESCALE_P2_4]
michael@0 349
michael@0 350 movdqa xmm5,xmm4
michael@0 351 movdqa xmm3,xmm7
michael@0 352 paddd xmm4,xmm6 ; xmm4=data0=(00 10 20 30)
michael@0 353 paddd xmm7,xmm2 ; xmm7=data1=(01 11 21 31)
michael@0 354 psubd xmm5,xmm6 ; xmm5=data3=(03 13 23 33)
michael@0 355 psubd xmm3,xmm2 ; xmm3=data2=(02 12 22 32)
michael@0 356
michael@0 357 paddd xmm4,xmm1
michael@0 358 paddd xmm7,xmm1
michael@0 359 psrad xmm4,DESCALE_P2_4
michael@0 360 psrad xmm7,DESCALE_P2_4
michael@0 361 paddd xmm5,xmm1
michael@0 362 paddd xmm3,xmm1
michael@0 363 psrad xmm5,DESCALE_P2_4
michael@0 364 psrad xmm3,DESCALE_P2_4
michael@0 365
michael@0 366 packssdw xmm4,xmm3 ; xmm4=(00 10 20 30 02 12 22 32)
michael@0 367 packssdw xmm7,xmm5 ; xmm7=(01 11 21 31 03 13 23 33)
michael@0 368
michael@0 369 movdqa xmm0,xmm4 ; transpose coefficients(phase 1)
michael@0 370 punpcklwd xmm4,xmm7 ; xmm4=(00 01 10 11 20 21 30 31)
michael@0 371 punpckhwd xmm0,xmm7 ; xmm0=(02 03 12 13 22 23 32 33)
michael@0 372
michael@0 373 movdqa xmm6,xmm4 ; transpose coefficients(phase 2)
michael@0 374 punpckldq xmm4,xmm0 ; xmm4=(00 01 02 03 10 11 12 13)
michael@0 375 punpckhdq xmm6,xmm0 ; xmm6=(20 21 22 23 30 31 32 33)
michael@0 376
michael@0 377 packsswb xmm4,xmm6 ; xmm4=(00 01 02 03 10 11 12 13 20 ..)
michael@0 378 paddb xmm4,[rel PB_CENTERJSAMP]
michael@0 379
michael@0 380 pshufd xmm2,xmm4,0x39 ; xmm2=(10 11 12 13 20 21 22 23 30 ..)
michael@0 381 pshufd xmm1,xmm4,0x4E ; xmm1=(20 21 22 23 30 31 32 33 00 ..)
michael@0 382 pshufd xmm3,xmm4,0x93 ; xmm3=(30 31 32 33 00 01 02 03 10 ..)
michael@0 383
michael@0 384 mov rdx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
michael@0 385 mov rsi, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
michael@0 386 movd XMM_DWORD [rdx+rax*SIZEOF_JSAMPLE], xmm4
michael@0 387 movd XMM_DWORD [rsi+rax*SIZEOF_JSAMPLE], xmm2
michael@0 388 mov rdx, JSAMPROW [rdi+2*SIZEOF_JSAMPROW]
michael@0 389 mov rsi, JSAMPROW [rdi+3*SIZEOF_JSAMPROW]
michael@0 390 movd XMM_DWORD [rdx+rax*SIZEOF_JSAMPLE], xmm1
michael@0 391 movd XMM_DWORD [rsi+rax*SIZEOF_JSAMPLE], xmm3
michael@0 392
michael@0 393 uncollect_args
michael@0 394 mov rsp,rbp ; rsp <- aligned rbp
michael@0 395 pop rsp ; rsp <- original rbp
michael@0 396 pop rbp
michael@0 397 ret
michael@0 398
michael@0 399
michael@0 400 ; --------------------------------------------------------------------------
michael@0 401 ;
michael@0 402 ; Perform dequantization and inverse DCT on one block of coefficients,
michael@0 403 ; producing a reduced-size 2x2 output block.
michael@0 404 ;
michael@0 405 ; GLOBAL(void)
michael@0 406 ; jsimd_idct_2x2_sse2 (void * dct_table, JCOEFPTR coef_block,
michael@0 407 ; JSAMPARRAY output_buf, JDIMENSION output_col)
michael@0 408 ;
michael@0 409
michael@0 410 ; r10 = void * dct_table
michael@0 411 ; r11 = JCOEFPTR coef_block
michael@0 412 ; r12 = JSAMPARRAY output_buf
michael@0 413 ; r13 = JDIMENSION output_col
michael@0 414
michael@0 415 align 16
michael@0 416 global EXTN(jsimd_idct_2x2_sse2)
michael@0 417
michael@0 418 EXTN(jsimd_idct_2x2_sse2):
michael@0 419 push rbp
michael@0 420 mov rax,rsp
michael@0 421 mov rbp,rsp
michael@0 422 collect_args
michael@0 423 push rbx
michael@0 424
michael@0 425 ; ---- Pass 1: process columns from input.
michael@0 426
michael@0 427 mov rdx, r10 ; quantptr
michael@0 428 mov rsi, r11 ; inptr
michael@0 429
michael@0 430 ; | input: | result: |
michael@0 431 ; | 00 01 ** 03 ** 05 ** 07 | |
michael@0 432 ; | 10 11 ** 13 ** 15 ** 17 | |
michael@0 433 ; | ** ** ** ** ** ** ** ** | |
michael@0 434 ; | 30 31 ** 33 ** 35 ** 37 | A0 A1 A3 A5 A7 |
michael@0 435 ; | ** ** ** ** ** ** ** ** | B0 B1 B3 B5 B7 |
michael@0 436 ; | 50 51 ** 53 ** 55 ** 57 | |
michael@0 437 ; | ** ** ** ** ** ** ** ** | |
michael@0 438 ; | 70 71 ** 73 ** 75 ** 77 | |
michael@0 439
michael@0 440 ; -- Odd part
michael@0 441
michael@0 442 movdqa xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
michael@0 443 movdqa xmm1, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
michael@0 444 pmullw xmm0, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 445 pmullw xmm1, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 446 movdqa xmm2, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
michael@0 447 movdqa xmm3, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
michael@0 448 pmullw xmm2, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 449 pmullw xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 450
michael@0 451 ; xmm0=(10 11 ** 13 ** 15 ** 17), xmm1=(30 31 ** 33 ** 35 ** 37)
michael@0 452 ; xmm2=(50 51 ** 53 ** 55 ** 57), xmm3=(70 71 ** 73 ** 75 ** 77)
michael@0 453
michael@0 454 pcmpeqd xmm7,xmm7
michael@0 455 pslld xmm7,WORD_BIT ; xmm7={0x0000 0xFFFF 0x0000 0xFFFF ..}
michael@0 456
michael@0 457 movdqa xmm4,xmm0 ; xmm4=(10 11 ** 13 ** 15 ** 17)
michael@0 458 movdqa xmm5,xmm2 ; xmm5=(50 51 ** 53 ** 55 ** 57)
michael@0 459 punpcklwd xmm4,xmm1 ; xmm4=(10 30 11 31 ** ** 13 33)
michael@0 460 punpcklwd xmm5,xmm3 ; xmm5=(50 70 51 71 ** ** 53 73)
michael@0 461 pmaddwd xmm4,[rel PW_F362_MF127]
michael@0 462 pmaddwd xmm5,[rel PW_F085_MF072]
michael@0 463
michael@0 464 psrld xmm0,WORD_BIT ; xmm0=(11 -- 13 -- 15 -- 17 --)
michael@0 465 pand xmm1,xmm7 ; xmm1=(-- 31 -- 33 -- 35 -- 37)
michael@0 466 psrld xmm2,WORD_BIT ; xmm2=(51 -- 53 -- 55 -- 57 --)
michael@0 467 pand xmm3,xmm7 ; xmm3=(-- 71 -- 73 -- 75 -- 77)
michael@0 468 por xmm0,xmm1 ; xmm0=(11 31 13 33 15 35 17 37)
michael@0 469 por xmm2,xmm3 ; xmm2=(51 71 53 73 55 75 57 77)
michael@0 470 pmaddwd xmm0,[rel PW_F362_MF127]
michael@0 471 pmaddwd xmm2,[rel PW_F085_MF072]
michael@0 472
michael@0 473 paddd xmm4,xmm5 ; xmm4=tmp0[col0 col1 **** col3]
michael@0 474 paddd xmm0,xmm2 ; xmm0=tmp0[col1 col3 col5 col7]
michael@0 475
michael@0 476 ; -- Even part
michael@0 477
michael@0 478 movdqa xmm6, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
michael@0 479 pmullw xmm6, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 480
michael@0 481 ; xmm6=(00 01 ** 03 ** 05 ** 07)
michael@0 482
michael@0 483 movdqa xmm1,xmm6 ; xmm1=(00 01 ** 03 ** 05 ** 07)
michael@0 484 pslld xmm6,WORD_BIT ; xmm6=(-- 00 -- ** -- ** -- **)
michael@0 485 pand xmm1,xmm7 ; xmm1=(-- 01 -- 03 -- 05 -- 07)
michael@0 486 psrad xmm6,(WORD_BIT-CONST_BITS-2) ; xmm6=tmp10[col0 **** **** ****]
michael@0 487 psrad xmm1,(WORD_BIT-CONST_BITS-2) ; xmm1=tmp10[col1 col3 col5 col7]
michael@0 488
michael@0 489 ; -- Final output stage
michael@0 490
michael@0 491 movdqa xmm3,xmm6
michael@0 492 movdqa xmm5,xmm1
michael@0 493 paddd xmm6,xmm4 ; xmm6=data0[col0 **** **** ****]=(A0 ** ** **)
michael@0 494 paddd xmm1,xmm0 ; xmm1=data0[col1 col3 col5 col7]=(A1 A3 A5 A7)
michael@0 495 psubd xmm3,xmm4 ; xmm3=data1[col0 **** **** ****]=(B0 ** ** **)
michael@0 496 psubd xmm5,xmm0 ; xmm5=data1[col1 col3 col5 col7]=(B1 B3 B5 B7)
michael@0 497
michael@0 498 movdqa xmm2,[rel PD_DESCALE_P1_2] ; xmm2=[rel PD_DESCALE_P1_2]
michael@0 499
michael@0 500 punpckldq xmm6,xmm3 ; xmm6=(A0 B0 ** **)
michael@0 501
michael@0 502 movdqa xmm7,xmm1
michael@0 503 punpcklqdq xmm1,xmm5 ; xmm1=(A1 A3 B1 B3)
michael@0 504 punpckhqdq xmm7,xmm5 ; xmm7=(A5 A7 B5 B7)
michael@0 505
michael@0 506 paddd xmm6,xmm2
michael@0 507 psrad xmm6,DESCALE_P1_2
michael@0 508
michael@0 509 paddd xmm1,xmm2
michael@0 510 paddd xmm7,xmm2
michael@0 511 psrad xmm1,DESCALE_P1_2
michael@0 512 psrad xmm7,DESCALE_P1_2
michael@0 513
michael@0 514 ; -- Prefetch the next coefficient block
michael@0 515
michael@0 516 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
michael@0 517 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
michael@0 518 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
michael@0 519 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
michael@0 520
michael@0 521 ; ---- Pass 2: process rows, store into output array.
michael@0 522
michael@0 523 mov rdi, r12 ; (JSAMPROW *)
michael@0 524 mov rax, r13
michael@0 525
michael@0 526 ; | input:| result:|
michael@0 527 ; | A0 B0 | |
michael@0 528 ; | A1 B1 | C0 C1 |
michael@0 529 ; | A3 B3 | D0 D1 |
michael@0 530 ; | A5 B5 | |
michael@0 531 ; | A7 B7 | |
michael@0 532
michael@0 533 ; -- Odd part
michael@0 534
michael@0 535 packssdw xmm1,xmm1 ; xmm1=(A1 A3 B1 B3 A1 A3 B1 B3)
michael@0 536 packssdw xmm7,xmm7 ; xmm7=(A5 A7 B5 B7 A5 A7 B5 B7)
michael@0 537 pmaddwd xmm1,[rel PW_F362_MF127]
michael@0 538 pmaddwd xmm7,[rel PW_F085_MF072]
michael@0 539
michael@0 540 paddd xmm1,xmm7 ; xmm1=tmp0[row0 row1 row0 row1]
michael@0 541
michael@0 542 ; -- Even part
michael@0 543
michael@0 544 pslld xmm6,(CONST_BITS+2) ; xmm6=tmp10[row0 row1 **** ****]
michael@0 545
michael@0 546 ; -- Final output stage
michael@0 547
michael@0 548 movdqa xmm4,xmm6
michael@0 549 paddd xmm6,xmm1 ; xmm6=data0[row0 row1 **** ****]=(C0 C1 ** **)
michael@0 550 psubd xmm4,xmm1 ; xmm4=data1[row0 row1 **** ****]=(D0 D1 ** **)
michael@0 551
michael@0 552 punpckldq xmm6,xmm4 ; xmm6=(C0 D0 C1 D1)
michael@0 553
michael@0 554 paddd xmm6,[rel PD_DESCALE_P2_2]
michael@0 555 psrad xmm6,DESCALE_P2_2
michael@0 556
michael@0 557 packssdw xmm6,xmm6 ; xmm6=(C0 D0 C1 D1 C0 D0 C1 D1)
michael@0 558 packsswb xmm6,xmm6 ; xmm6=(C0 D0 C1 D1 C0 D0 C1 D1 ..)
michael@0 559 paddb xmm6,[rel PB_CENTERJSAMP]
michael@0 560
michael@0 561 pextrw ebx,xmm6,0x00 ; ebx=(C0 D0 -- --)
michael@0 562 pextrw ecx,xmm6,0x01 ; ecx=(C1 D1 -- --)
michael@0 563
michael@0 564 mov rdx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
michael@0 565 mov rsi, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
michael@0 566 mov WORD [rdx+rax*SIZEOF_JSAMPLE], bx
michael@0 567 mov WORD [rsi+rax*SIZEOF_JSAMPLE], cx
michael@0 568
michael@0 569 pop rbx
michael@0 570 uncollect_args
michael@0 571 pop rbp
michael@0 572 ret
michael@0 573
michael@0 574 ; For some reason, the OS X linker does not honor the request to align the
michael@0 575 ; segment unless we do this.
michael@0 576 align 16

Mercurial Repository: The Tor Browser

Europalab SCM Repository Server

mercurial