The Tor Browser / annotate

media/libjpeg/simd/jiss2red.asm@b8a032363ba2 (annotated)

media/libjpeg/simd/jiss2red.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 ; jiss2red.asm - reduced-size IDCT (SSE2)
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 inverse-DCT routines that produce reduced-size
michael@0 18 ; output: either 4x4 or 2x2 pixels from an 8x8 DCT block.
michael@0 19 ; The following code is based directly on the IJG's original jidctred.c;
michael@0 20 ; see the jidctred.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 %define CONST_BITS 13
michael@0 30 %define PASS1_BITS 2
michael@0 31
michael@0 32 %define DESCALE_P1_4 (CONST_BITS-PASS1_BITS+1)
michael@0 33 %define DESCALE_P2_4 (CONST_BITS+PASS1_BITS+3+1)
michael@0 34 %define DESCALE_P1_2 (CONST_BITS-PASS1_BITS+2)
michael@0 35 %define DESCALE_P2_2 (CONST_BITS+PASS1_BITS+3+2)
michael@0 36
michael@0 37 %if CONST_BITS == 13
michael@0 38 F_0_211 equ 1730 ; FIX(0.211164243)
michael@0 39 F_0_509 equ 4176 ; FIX(0.509795579)
michael@0 40 F_0_601 equ 4926 ; FIX(0.601344887)
michael@0 41 F_0_720 equ 5906 ; FIX(0.720959822)
michael@0 42 F_0_765 equ 6270 ; FIX(0.765366865)
michael@0 43 F_0_850 equ 6967 ; FIX(0.850430095)
michael@0 44 F_0_899 equ 7373 ; FIX(0.899976223)
michael@0 45 F_1_061 equ 8697 ; FIX(1.061594337)
michael@0 46 F_1_272 equ 10426 ; FIX(1.272758580)
michael@0 47 F_1_451 equ 11893 ; FIX(1.451774981)
michael@0 48 F_1_847 equ 15137 ; FIX(1.847759065)
michael@0 49 F_2_172 equ 17799 ; FIX(2.172734803)
michael@0 50 F_2_562 equ 20995 ; FIX(2.562915447)
michael@0 51 F_3_624 equ 29692 ; FIX(3.624509785)
michael@0 52 %else
michael@0 53 ; NASM cannot do compile-time arithmetic on floating-point constants.
michael@0 54 %define DESCALE(x,n) (((x)+(1<<((n)-1)))>>(n))
michael@0 55 F_0_211 equ DESCALE( 226735879,30-CONST_BITS) ; FIX(0.211164243)
michael@0 56 F_0_509 equ DESCALE( 547388834,30-CONST_BITS) ; FIX(0.509795579)
michael@0 57 F_0_601 equ DESCALE( 645689155,30-CONST_BITS) ; FIX(0.601344887)
michael@0 58 F_0_720 equ DESCALE( 774124714,30-CONST_BITS) ; FIX(0.720959822)
michael@0 59 F_0_765 equ DESCALE( 821806413,30-CONST_BITS) ; FIX(0.765366865)
michael@0 60 F_0_850 equ DESCALE( 913142361,30-CONST_BITS) ; FIX(0.850430095)
michael@0 61 F_0_899 equ DESCALE( 966342111,30-CONST_BITS) ; FIX(0.899976223)
michael@0 62 F_1_061 equ DESCALE(1139878239,30-CONST_BITS) ; FIX(1.061594337)
michael@0 63 F_1_272 equ DESCALE(1366614119,30-CONST_BITS) ; FIX(1.272758580)
michael@0 64 F_1_451 equ DESCALE(1558831516,30-CONST_BITS) ; FIX(1.451774981)
michael@0 65 F_1_847 equ DESCALE(1984016188,30-CONST_BITS) ; FIX(1.847759065)
michael@0 66 F_2_172 equ DESCALE(2332956230,30-CONST_BITS) ; FIX(2.172734803)
michael@0 67 F_2_562 equ DESCALE(2751909506,30-CONST_BITS) ; FIX(2.562915447)
michael@0 68 F_3_624 equ DESCALE(3891787747,30-CONST_BITS) ; FIX(3.624509785)
michael@0 69 %endif
michael@0 70
michael@0 71 ; --------------------------------------------------------------------------
michael@0 72 SECTION SEG_CONST
michael@0 73
michael@0 74 alignz 16
michael@0 75 global EXTN(jconst_idct_red_sse2)
michael@0 76
michael@0 77 EXTN(jconst_idct_red_sse2):
michael@0 78
michael@0 79 PW_F184_MF076 times 4 dw F_1_847,-F_0_765
michael@0 80 PW_F256_F089 times 4 dw F_2_562, F_0_899
michael@0 81 PW_F106_MF217 times 4 dw F_1_061,-F_2_172
michael@0 82 PW_MF060_MF050 times 4 dw -F_0_601,-F_0_509
michael@0 83 PW_F145_MF021 times 4 dw F_1_451,-F_0_211
michael@0 84 PW_F362_MF127 times 4 dw F_3_624,-F_1_272
michael@0 85 PW_F085_MF072 times 4 dw F_0_850,-F_0_720
michael@0 86 PD_DESCALE_P1_4 times 4 dd 1 << (DESCALE_P1_4-1)
michael@0 87 PD_DESCALE_P2_4 times 4 dd 1 << (DESCALE_P2_4-1)
michael@0 88 PD_DESCALE_P1_2 times 4 dd 1 << (DESCALE_P1_2-1)
michael@0 89 PD_DESCALE_P2_2 times 4 dd 1 << (DESCALE_P2_2-1)
michael@0 90 PB_CENTERJSAMP times 16 db CENTERJSAMPLE
michael@0 91
michael@0 92 alignz 16
michael@0 93
michael@0 94 ; --------------------------------------------------------------------------
michael@0 95 SECTION SEG_TEXT
michael@0 96 BITS 32
michael@0 97 ;
michael@0 98 ; Perform dequantization and inverse DCT on one block of coefficients,
michael@0 99 ; producing a reduced-size 4x4 output block.
michael@0 100 ;
michael@0 101 ; GLOBAL(void)
michael@0 102 ; jsimd_idct_4x4_sse2 (void * dct_table, JCOEFPTR coef_block,
michael@0 103 ; JSAMPARRAY output_buf, JDIMENSION output_col)
michael@0 104 ;
michael@0 105
michael@0 106 %define dct_table(b) (b)+8 ; void * dct_table
michael@0 107 %define coef_block(b) (b)+12 ; JCOEFPTR coef_block
michael@0 108 %define output_buf(b) (b)+16 ; JSAMPARRAY output_buf
michael@0 109 %define output_col(b) (b)+20 ; JDIMENSION output_col
michael@0 110
michael@0 111 %define original_ebp ebp+0
michael@0 112 %define wk(i) ebp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM]
michael@0 113 %define WK_NUM 2
michael@0 114
michael@0 115 align 16
michael@0 116 global EXTN(jsimd_idct_4x4_sse2)
michael@0 117
michael@0 118 EXTN(jsimd_idct_4x4_sse2):
michael@0 119 push ebp
michael@0 120 mov eax,esp ; eax = original ebp
michael@0 121 sub esp, byte 4
michael@0 122 and esp, byte (-SIZEOF_XMMWORD) ; align to 128 bits
michael@0 123 mov [esp],eax
michael@0 124 mov ebp,esp ; ebp = aligned ebp
michael@0 125 lea esp, [wk(0)]
michael@0 126 pushpic ebx
michael@0 127 ; push ecx ; unused
michael@0 128 ; push edx ; need not be preserved
michael@0 129 push esi
michael@0 130 push edi
michael@0 131
michael@0 132 get_GOT ebx ; get GOT address
michael@0 133
michael@0 134 ; ---- Pass 1: process columns from input.
michael@0 135
michael@0 136 ; mov eax, [original_ebp]
michael@0 137 mov edx, POINTER [dct_table(eax)] ; quantptr
michael@0 138 mov esi, JCOEFPTR [coef_block(eax)] ; inptr
michael@0 139
michael@0 140 %ifndef NO_ZERO_COLUMN_TEST_4X4_SSE2
michael@0 141 mov eax, DWORD [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
michael@0 142 or eax, DWORD [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
michael@0 143 jnz short .columnDCT
michael@0 144
michael@0 145 movdqa xmm0, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_JCOEF)]
michael@0 146 movdqa xmm1, XMMWORD [XMMBLOCK(2,0,esi,SIZEOF_JCOEF)]
michael@0 147 por xmm0, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_JCOEF)]
michael@0 148 por xmm1, XMMWORD [XMMBLOCK(5,0,esi,SIZEOF_JCOEF)]
michael@0 149 por xmm0, XMMWORD [XMMBLOCK(6,0,esi,SIZEOF_JCOEF)]
michael@0 150 por xmm1, XMMWORD [XMMBLOCK(7,0,esi,SIZEOF_JCOEF)]
michael@0 151 por xmm0,xmm1
michael@0 152 packsswb xmm0,xmm0
michael@0 153 packsswb xmm0,xmm0
michael@0 154 movd eax,xmm0
michael@0 155 test eax,eax
michael@0 156 jnz short .columnDCT
michael@0 157
michael@0 158 ; -- AC terms all zero
michael@0 159
michael@0 160 movdqa xmm0, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_JCOEF)]
michael@0 161 pmullw xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 162
michael@0 163 psllw xmm0,PASS1_BITS
michael@0 164
michael@0 165 movdqa xmm3,xmm0 ; xmm0=in0=(00 01 02 03 04 05 06 07)
michael@0 166 punpcklwd xmm0,xmm0 ; xmm0=(00 00 01 01 02 02 03 03)
michael@0 167 punpckhwd xmm3,xmm3 ; xmm3=(04 04 05 05 06 06 07 07)
michael@0 168
michael@0 169 pshufd xmm1,xmm0,0x50 ; xmm1=[col0 col1]=(00 00 00 00 01 01 01 01)
michael@0 170 pshufd xmm0,xmm0,0xFA ; xmm0=[col2 col3]=(02 02 02 02 03 03 03 03)
michael@0 171 pshufd xmm6,xmm3,0x50 ; xmm6=[col4 col5]=(04 04 04 04 05 05 05 05)
michael@0 172 pshufd xmm3,xmm3,0xFA ; xmm3=[col6 col7]=(06 06 06 06 07 07 07 07)
michael@0 173
michael@0 174 jmp near .column_end
michael@0 175 alignx 16,7
michael@0 176 %endif
michael@0 177 .columnDCT:
michael@0 178
michael@0 179 ; -- Odd part
michael@0 180
michael@0 181 movdqa xmm0, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_JCOEF)]
michael@0 182 movdqa xmm1, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_JCOEF)]
michael@0 183 pmullw xmm0, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 184 pmullw xmm1, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 185 movdqa xmm2, XMMWORD [XMMBLOCK(5,0,esi,SIZEOF_JCOEF)]
michael@0 186 movdqa xmm3, XMMWORD [XMMBLOCK(7,0,esi,SIZEOF_JCOEF)]
michael@0 187 pmullw xmm2, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 188 pmullw xmm3, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 189
michael@0 190 movdqa xmm4,xmm0
michael@0 191 movdqa xmm5,xmm0
michael@0 192 punpcklwd xmm4,xmm1
michael@0 193 punpckhwd xmm5,xmm1
michael@0 194 movdqa xmm0,xmm4
michael@0 195 movdqa xmm1,xmm5
michael@0 196 pmaddwd xmm4,[GOTOFF(ebx,PW_F256_F089)] ; xmm4=(tmp2L)
michael@0 197 pmaddwd xmm5,[GOTOFF(ebx,PW_F256_F089)] ; xmm5=(tmp2H)
michael@0 198 pmaddwd xmm0,[GOTOFF(ebx,PW_F106_MF217)] ; xmm0=(tmp0L)
michael@0 199 pmaddwd xmm1,[GOTOFF(ebx,PW_F106_MF217)] ; xmm1=(tmp0H)
michael@0 200
michael@0 201 movdqa xmm6,xmm2
michael@0 202 movdqa xmm7,xmm2
michael@0 203 punpcklwd xmm6,xmm3
michael@0 204 punpckhwd xmm7,xmm3
michael@0 205 movdqa xmm2,xmm6
michael@0 206 movdqa xmm3,xmm7
michael@0 207 pmaddwd xmm6,[GOTOFF(ebx,PW_MF060_MF050)] ; xmm6=(tmp2L)
michael@0 208 pmaddwd xmm7,[GOTOFF(ebx,PW_MF060_MF050)] ; xmm7=(tmp2H)
michael@0 209 pmaddwd xmm2,[GOTOFF(ebx,PW_F145_MF021)] ; xmm2=(tmp0L)
michael@0 210 pmaddwd xmm3,[GOTOFF(ebx,PW_F145_MF021)] ; xmm3=(tmp0H)
michael@0 211
michael@0 212 paddd xmm6,xmm4 ; xmm6=tmp2L
michael@0 213 paddd xmm7,xmm5 ; xmm7=tmp2H
michael@0 214 paddd xmm2,xmm0 ; xmm2=tmp0L
michael@0 215 paddd xmm3,xmm1 ; xmm3=tmp0H
michael@0 216
michael@0 217 movdqa XMMWORD [wk(0)], xmm2 ; wk(0)=tmp0L
michael@0 218 movdqa XMMWORD [wk(1)], xmm3 ; wk(1)=tmp0H
michael@0 219
michael@0 220 ; -- Even part
michael@0 221
michael@0 222 movdqa xmm4, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_JCOEF)]
michael@0 223 movdqa xmm5, XMMWORD [XMMBLOCK(2,0,esi,SIZEOF_JCOEF)]
michael@0 224 movdqa xmm0, XMMWORD [XMMBLOCK(6,0,esi,SIZEOF_JCOEF)]
michael@0 225 pmullw xmm4, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 226 pmullw xmm5, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 227 pmullw xmm0, XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 228
michael@0 229 pxor xmm1,xmm1
michael@0 230 pxor xmm2,xmm2
michael@0 231 punpcklwd xmm1,xmm4 ; xmm1=tmp0L
michael@0 232 punpckhwd xmm2,xmm4 ; xmm2=tmp0H
michael@0 233 psrad xmm1,(16-CONST_BITS-1) ; psrad xmm1,16 & pslld xmm1,CONST_BITS+1
michael@0 234 psrad xmm2,(16-CONST_BITS-1) ; psrad xmm2,16 & pslld xmm2,CONST_BITS+1
michael@0 235
michael@0 236 movdqa xmm3,xmm5 ; xmm5=in2=z2
michael@0 237 punpcklwd xmm5,xmm0 ; xmm0=in6=z3
michael@0 238 punpckhwd xmm3,xmm0
michael@0 239 pmaddwd xmm5,[GOTOFF(ebx,PW_F184_MF076)] ; xmm5=tmp2L
michael@0 240 pmaddwd xmm3,[GOTOFF(ebx,PW_F184_MF076)] ; xmm3=tmp2H
michael@0 241
michael@0 242 movdqa xmm4,xmm1
michael@0 243 movdqa xmm0,xmm2
michael@0 244 paddd xmm1,xmm5 ; xmm1=tmp10L
michael@0 245 paddd xmm2,xmm3 ; xmm2=tmp10H
michael@0 246 psubd xmm4,xmm5 ; xmm4=tmp12L
michael@0 247 psubd xmm0,xmm3 ; xmm0=tmp12H
michael@0 248
michael@0 249 ; -- Final output stage
michael@0 250
michael@0 251 movdqa xmm5,xmm1
michael@0 252 movdqa xmm3,xmm2
michael@0 253 paddd xmm1,xmm6 ; xmm1=data0L
michael@0 254 paddd xmm2,xmm7 ; xmm2=data0H
michael@0 255 psubd xmm5,xmm6 ; xmm5=data3L
michael@0 256 psubd xmm3,xmm7 ; xmm3=data3H
michael@0 257
michael@0 258 movdqa xmm6,[GOTOFF(ebx,PD_DESCALE_P1_4)] ; xmm6=[PD_DESCALE_P1_4]
michael@0 259
michael@0 260 paddd xmm1,xmm6
michael@0 261 paddd xmm2,xmm6
michael@0 262 psrad xmm1,DESCALE_P1_4
michael@0 263 psrad xmm2,DESCALE_P1_4
michael@0 264 paddd xmm5,xmm6
michael@0 265 paddd xmm3,xmm6
michael@0 266 psrad xmm5,DESCALE_P1_4
michael@0 267 psrad xmm3,DESCALE_P1_4
michael@0 268
michael@0 269 packssdw xmm1,xmm2 ; xmm1=data0=(00 01 02 03 04 05 06 07)
michael@0 270 packssdw xmm5,xmm3 ; xmm5=data3=(30 31 32 33 34 35 36 37)
michael@0 271
michael@0 272 movdqa xmm7, XMMWORD [wk(0)] ; xmm7=tmp0L
michael@0 273 movdqa xmm6, XMMWORD [wk(1)] ; xmm6=tmp0H
michael@0 274
michael@0 275 movdqa xmm2,xmm4
michael@0 276 movdqa xmm3,xmm0
michael@0 277 paddd xmm4,xmm7 ; xmm4=data1L
michael@0 278 paddd xmm0,xmm6 ; xmm0=data1H
michael@0 279 psubd xmm2,xmm7 ; xmm2=data2L
michael@0 280 psubd xmm3,xmm6 ; xmm3=data2H
michael@0 281
michael@0 282 movdqa xmm7,[GOTOFF(ebx,PD_DESCALE_P1_4)] ; xmm7=[PD_DESCALE_P1_4]
michael@0 283
michael@0 284 paddd xmm4,xmm7
michael@0 285 paddd xmm0,xmm7
michael@0 286 psrad xmm4,DESCALE_P1_4
michael@0 287 psrad xmm0,DESCALE_P1_4
michael@0 288 paddd xmm2,xmm7
michael@0 289 paddd xmm3,xmm7
michael@0 290 psrad xmm2,DESCALE_P1_4
michael@0 291 psrad xmm3,DESCALE_P1_4
michael@0 292
michael@0 293 packssdw xmm4,xmm0 ; xmm4=data1=(10 11 12 13 14 15 16 17)
michael@0 294 packssdw xmm2,xmm3 ; xmm2=data2=(20 21 22 23 24 25 26 27)
michael@0 295
michael@0 296 movdqa xmm6,xmm1 ; transpose coefficients(phase 1)
michael@0 297 punpcklwd xmm1,xmm4 ; xmm1=(00 10 01 11 02 12 03 13)
michael@0 298 punpckhwd xmm6,xmm4 ; xmm6=(04 14 05 15 06 16 07 17)
michael@0 299 movdqa xmm7,xmm2 ; transpose coefficients(phase 1)
michael@0 300 punpcklwd xmm2,xmm5 ; xmm2=(20 30 21 31 22 32 23 33)
michael@0 301 punpckhwd xmm7,xmm5 ; xmm7=(24 34 25 35 26 36 27 37)
michael@0 302
michael@0 303 movdqa xmm0,xmm1 ; transpose coefficients(phase 2)
michael@0 304 punpckldq xmm1,xmm2 ; xmm1=[col0 col1]=(00 10 20 30 01 11 21 31)
michael@0 305 punpckhdq xmm0,xmm2 ; xmm0=[col2 col3]=(02 12 22 32 03 13 23 33)
michael@0 306 movdqa xmm3,xmm6 ; transpose coefficients(phase 2)
michael@0 307 punpckldq xmm6,xmm7 ; xmm6=[col4 col5]=(04 14 24 34 05 15 25 35)
michael@0 308 punpckhdq xmm3,xmm7 ; xmm3=[col6 col7]=(06 16 26 36 07 17 27 37)
michael@0 309 .column_end:
michael@0 310
michael@0 311 ; -- Prefetch the next coefficient block
michael@0 312
michael@0 313 prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
michael@0 314 prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
michael@0 315 prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
michael@0 316 prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
michael@0 317
michael@0 318 ; ---- Pass 2: process rows, store into output array.
michael@0 319
michael@0 320 mov eax, [original_ebp]
michael@0 321 mov edi, JSAMPARRAY [output_buf(eax)] ; (JSAMPROW *)
michael@0 322 mov eax, JDIMENSION [output_col(eax)]
michael@0 323
michael@0 324 ; -- Even part
michael@0 325
michael@0 326 pxor xmm4,xmm4
michael@0 327 punpcklwd xmm4,xmm1 ; xmm4=tmp0
michael@0 328 psrad xmm4,(16-CONST_BITS-1) ; psrad xmm4,16 & pslld xmm4,CONST_BITS+1
michael@0 329
michael@0 330 ; -- Odd part
michael@0 331
michael@0 332 punpckhwd xmm1,xmm0
michael@0 333 punpckhwd xmm6,xmm3
michael@0 334 movdqa xmm5,xmm1
michael@0 335 movdqa xmm2,xmm6
michael@0 336 pmaddwd xmm1,[GOTOFF(ebx,PW_F256_F089)] ; xmm1=(tmp2)
michael@0 337 pmaddwd xmm6,[GOTOFF(ebx,PW_MF060_MF050)] ; xmm6=(tmp2)
michael@0 338 pmaddwd xmm5,[GOTOFF(ebx,PW_F106_MF217)] ; xmm5=(tmp0)
michael@0 339 pmaddwd xmm2,[GOTOFF(ebx,PW_F145_MF021)] ; xmm2=(tmp0)
michael@0 340
michael@0 341 paddd xmm6,xmm1 ; xmm6=tmp2
michael@0 342 paddd xmm2,xmm5 ; xmm2=tmp0
michael@0 343
michael@0 344 ; -- Even part
michael@0 345
michael@0 346 punpcklwd xmm0,xmm3
michael@0 347 pmaddwd xmm0,[GOTOFF(ebx,PW_F184_MF076)] ; xmm0=tmp2
michael@0 348
michael@0 349 movdqa xmm7,xmm4
michael@0 350 paddd xmm4,xmm0 ; xmm4=tmp10
michael@0 351 psubd xmm7,xmm0 ; xmm7=tmp12
michael@0 352
michael@0 353 ; -- Final output stage
michael@0 354
michael@0 355 movdqa xmm1,[GOTOFF(ebx,PD_DESCALE_P2_4)] ; xmm1=[PD_DESCALE_P2_4]
michael@0 356
michael@0 357 movdqa xmm5,xmm4
michael@0 358 movdqa xmm3,xmm7
michael@0 359 paddd xmm4,xmm6 ; xmm4=data0=(00 10 20 30)
michael@0 360 paddd xmm7,xmm2 ; xmm7=data1=(01 11 21 31)
michael@0 361 psubd xmm5,xmm6 ; xmm5=data3=(03 13 23 33)
michael@0 362 psubd xmm3,xmm2 ; xmm3=data2=(02 12 22 32)
michael@0 363
michael@0 364 paddd xmm4,xmm1
michael@0 365 paddd xmm7,xmm1
michael@0 366 psrad xmm4,DESCALE_P2_4
michael@0 367 psrad xmm7,DESCALE_P2_4
michael@0 368 paddd xmm5,xmm1
michael@0 369 paddd xmm3,xmm1
michael@0 370 psrad xmm5,DESCALE_P2_4
michael@0 371 psrad xmm3,DESCALE_P2_4
michael@0 372
michael@0 373 packssdw xmm4,xmm3 ; xmm4=(00 10 20 30 02 12 22 32)
michael@0 374 packssdw xmm7,xmm5 ; xmm7=(01 11 21 31 03 13 23 33)
michael@0 375
michael@0 376 movdqa xmm0,xmm4 ; transpose coefficients(phase 1)
michael@0 377 punpcklwd xmm4,xmm7 ; xmm4=(00 01 10 11 20 21 30 31)
michael@0 378 punpckhwd xmm0,xmm7 ; xmm0=(02 03 12 13 22 23 32 33)
michael@0 379
michael@0 380 movdqa xmm6,xmm4 ; transpose coefficients(phase 2)
michael@0 381 punpckldq xmm4,xmm0 ; xmm4=(00 01 02 03 10 11 12 13)
michael@0 382 punpckhdq xmm6,xmm0 ; xmm6=(20 21 22 23 30 31 32 33)
michael@0 383
michael@0 384 packsswb xmm4,xmm6 ; xmm4=(00 01 02 03 10 11 12 13 20 ..)
michael@0 385 paddb xmm4,[GOTOFF(ebx,PB_CENTERJSAMP)]
michael@0 386
michael@0 387 pshufd xmm2,xmm4,0x39 ; xmm2=(10 11 12 13 20 21 22 23 30 ..)
michael@0 388 pshufd xmm1,xmm4,0x4E ; xmm1=(20 21 22 23 30 31 32 33 00 ..)
michael@0 389 pshufd xmm3,xmm4,0x93 ; xmm3=(30 31 32 33 00 01 02 03 10 ..)
michael@0 390
michael@0 391 mov edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
michael@0 392 mov esi, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
michael@0 393 movd XMM_DWORD [edx+eax*SIZEOF_JSAMPLE], xmm4
michael@0 394 movd XMM_DWORD [esi+eax*SIZEOF_JSAMPLE], xmm2
michael@0 395 mov edx, JSAMPROW [edi+2*SIZEOF_JSAMPROW]
michael@0 396 mov esi, JSAMPROW [edi+3*SIZEOF_JSAMPROW]
michael@0 397 movd XMM_DWORD [edx+eax*SIZEOF_JSAMPLE], xmm1
michael@0 398 movd XMM_DWORD [esi+eax*SIZEOF_JSAMPLE], xmm3
michael@0 399
michael@0 400 pop edi
michael@0 401 pop esi
michael@0 402 ; pop edx ; need not be preserved
michael@0 403 ; pop ecx ; unused
michael@0 404 poppic ebx
michael@0 405 mov esp,ebp ; esp <- aligned ebp
michael@0 406 pop esp ; esp <- original ebp
michael@0 407 pop ebp
michael@0 408 ret
michael@0 409
michael@0 410
michael@0 411 ; --------------------------------------------------------------------------
michael@0 412 ;
michael@0 413 ; Perform dequantization and inverse DCT on one block of coefficients,
michael@0 414 ; producing a reduced-size 2x2 output block.
michael@0 415 ;
michael@0 416 ; GLOBAL(void)
michael@0 417 ; jsimd_idct_2x2_sse2 (void * dct_table, JCOEFPTR coef_block,
michael@0 418 ; JSAMPARRAY output_buf, JDIMENSION output_col)
michael@0 419 ;
michael@0 420
michael@0 421 %define dct_table(b) (b)+8 ; void * dct_table
michael@0 422 %define coef_block(b) (b)+12 ; JCOEFPTR coef_block
michael@0 423 %define output_buf(b) (b)+16 ; JSAMPARRAY output_buf
michael@0 424 %define output_col(b) (b)+20 ; JDIMENSION output_col
michael@0 425
michael@0 426 align 16
michael@0 427 global EXTN(jsimd_idct_2x2_sse2)
michael@0 428
michael@0 429 EXTN(jsimd_idct_2x2_sse2):
michael@0 430 push ebp
michael@0 431 mov ebp,esp
michael@0 432 push ebx
michael@0 433 ; push ecx ; need not be preserved
michael@0 434 ; push edx ; need not be preserved
michael@0 435 push esi
michael@0 436 push edi
michael@0 437
michael@0 438 get_GOT ebx ; get GOT address
michael@0 439
michael@0 440 ; ---- Pass 1: process columns from input.
michael@0 441
michael@0 442 mov edx, POINTER [dct_table(ebp)] ; quantptr
michael@0 443 mov esi, JCOEFPTR [coef_block(ebp)] ; inptr
michael@0 444
michael@0 445 ; | input: | result: |
michael@0 446 ; | 00 01 ** 03 ** 05 ** 07 | |
michael@0 447 ; | 10 11 ** 13 ** 15 ** 17 | |
michael@0 448 ; | ** ** ** ** ** ** ** ** | |
michael@0 449 ; | 30 31 ** 33 ** 35 ** 37 | A0 A1 A3 A5 A7 |
michael@0 450 ; | ** ** ** ** ** ** ** ** | B0 B1 B3 B5 B7 |
michael@0 451 ; | 50 51 ** 53 ** 55 ** 57 | |
michael@0 452 ; | ** ** ** ** ** ** ** ** | |
michael@0 453 ; | 70 71 ** 73 ** 75 ** 77 | |
michael@0 454
michael@0 455 ; -- Odd part
michael@0 456
michael@0 457 movdqa xmm0, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_JCOEF)]
michael@0 458 movdqa xmm1, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_JCOEF)]
michael@0 459 pmullw xmm0, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 460 pmullw xmm1, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 461 movdqa xmm2, XMMWORD [XMMBLOCK(5,0,esi,SIZEOF_JCOEF)]
michael@0 462 movdqa xmm3, XMMWORD [XMMBLOCK(7,0,esi,SIZEOF_JCOEF)]
michael@0 463 pmullw xmm2, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 464 pmullw xmm3, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 465
michael@0 466 ; xmm0=(10 11 ** 13 ** 15 ** 17), xmm1=(30 31 ** 33 ** 35 ** 37)
michael@0 467 ; xmm2=(50 51 ** 53 ** 55 ** 57), xmm3=(70 71 ** 73 ** 75 ** 77)
michael@0 468
michael@0 469 pcmpeqd xmm7,xmm7
michael@0 470 pslld xmm7,WORD_BIT ; xmm7={0x0000 0xFFFF 0x0000 0xFFFF ..}
michael@0 471
michael@0 472 movdqa xmm4,xmm0 ; xmm4=(10 11 ** 13 ** 15 ** 17)
michael@0 473 movdqa xmm5,xmm2 ; xmm5=(50 51 ** 53 ** 55 ** 57)
michael@0 474 punpcklwd xmm4,xmm1 ; xmm4=(10 30 11 31 ** ** 13 33)
michael@0 475 punpcklwd xmm5,xmm3 ; xmm5=(50 70 51 71 ** ** 53 73)
michael@0 476 pmaddwd xmm4,[GOTOFF(ebx,PW_F362_MF127)]
michael@0 477 pmaddwd xmm5,[GOTOFF(ebx,PW_F085_MF072)]
michael@0 478
michael@0 479 psrld xmm0,WORD_BIT ; xmm0=(11 -- 13 -- 15 -- 17 --)
michael@0 480 pand xmm1,xmm7 ; xmm1=(-- 31 -- 33 -- 35 -- 37)
michael@0 481 psrld xmm2,WORD_BIT ; xmm2=(51 -- 53 -- 55 -- 57 --)
michael@0 482 pand xmm3,xmm7 ; xmm3=(-- 71 -- 73 -- 75 -- 77)
michael@0 483 por xmm0,xmm1 ; xmm0=(11 31 13 33 15 35 17 37)
michael@0 484 por xmm2,xmm3 ; xmm2=(51 71 53 73 55 75 57 77)
michael@0 485 pmaddwd xmm0,[GOTOFF(ebx,PW_F362_MF127)]
michael@0 486 pmaddwd xmm2,[GOTOFF(ebx,PW_F085_MF072)]
michael@0 487
michael@0 488 paddd xmm4,xmm5 ; xmm4=tmp0[col0 col1 **** col3]
michael@0 489 paddd xmm0,xmm2 ; xmm0=tmp0[col1 col3 col5 col7]
michael@0 490
michael@0 491 ; -- Even part
michael@0 492
michael@0 493 movdqa xmm6, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_JCOEF)]
michael@0 494 pmullw xmm6, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
michael@0 495
michael@0 496 ; xmm6=(00 01 ** 03 ** 05 ** 07)
michael@0 497
michael@0 498 movdqa xmm1,xmm6 ; xmm1=(00 01 ** 03 ** 05 ** 07)
michael@0 499 pslld xmm6,WORD_BIT ; xmm6=(-- 00 -- ** -- ** -- **)
michael@0 500 pand xmm1,xmm7 ; xmm1=(-- 01 -- 03 -- 05 -- 07)
michael@0 501 psrad xmm6,(WORD_BIT-CONST_BITS-2) ; xmm6=tmp10[col0 **** **** ****]
michael@0 502 psrad xmm1,(WORD_BIT-CONST_BITS-2) ; xmm1=tmp10[col1 col3 col5 col7]
michael@0 503
michael@0 504 ; -- Final output stage
michael@0 505
michael@0 506 movdqa xmm3,xmm6
michael@0 507 movdqa xmm5,xmm1
michael@0 508 paddd xmm6,xmm4 ; xmm6=data0[col0 **** **** ****]=(A0 ** ** **)
michael@0 509 paddd xmm1,xmm0 ; xmm1=data0[col1 col3 col5 col7]=(A1 A3 A5 A7)
michael@0 510 psubd xmm3,xmm4 ; xmm3=data1[col0 **** **** ****]=(B0 ** ** **)
michael@0 511 psubd xmm5,xmm0 ; xmm5=data1[col1 col3 col5 col7]=(B1 B3 B5 B7)
michael@0 512
michael@0 513 movdqa xmm2,[GOTOFF(ebx,PD_DESCALE_P1_2)] ; xmm2=[PD_DESCALE_P1_2]
michael@0 514
michael@0 515 punpckldq xmm6,xmm3 ; xmm6=(A0 B0 ** **)
michael@0 516
michael@0 517 movdqa xmm7,xmm1
michael@0 518 punpcklqdq xmm1,xmm5 ; xmm1=(A1 A3 B1 B3)
michael@0 519 punpckhqdq xmm7,xmm5 ; xmm7=(A5 A7 B5 B7)
michael@0 520
michael@0 521 paddd xmm6,xmm2
michael@0 522 psrad xmm6,DESCALE_P1_2
michael@0 523
michael@0 524 paddd xmm1,xmm2
michael@0 525 paddd xmm7,xmm2
michael@0 526 psrad xmm1,DESCALE_P1_2
michael@0 527 psrad xmm7,DESCALE_P1_2
michael@0 528
michael@0 529 ; -- Prefetch the next coefficient block
michael@0 530
michael@0 531 prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
michael@0 532 prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
michael@0 533 prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
michael@0 534 prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
michael@0 535
michael@0 536 ; ---- Pass 2: process rows, store into output array.
michael@0 537
michael@0 538 mov edi, JSAMPARRAY [output_buf(ebp)] ; (JSAMPROW *)
michael@0 539 mov eax, JDIMENSION [output_col(ebp)]
michael@0 540
michael@0 541 ; | input:| result:|
michael@0 542 ; | A0 B0 | |
michael@0 543 ; | A1 B1 | C0 C1 |
michael@0 544 ; | A3 B3 | D0 D1 |
michael@0 545 ; | A5 B5 | |
michael@0 546 ; | A7 B7 | |
michael@0 547
michael@0 548 ; -- Odd part
michael@0 549
michael@0 550 packssdw xmm1,xmm1 ; xmm1=(A1 A3 B1 B3 A1 A3 B1 B3)
michael@0 551 packssdw xmm7,xmm7 ; xmm7=(A5 A7 B5 B7 A5 A7 B5 B7)
michael@0 552 pmaddwd xmm1,[GOTOFF(ebx,PW_F362_MF127)]
michael@0 553 pmaddwd xmm7,[GOTOFF(ebx,PW_F085_MF072)]
michael@0 554
michael@0 555 paddd xmm1,xmm7 ; xmm1=tmp0[row0 row1 row0 row1]
michael@0 556
michael@0 557 ; -- Even part
michael@0 558
michael@0 559 pslld xmm6,(CONST_BITS+2) ; xmm6=tmp10[row0 row1 **** ****]
michael@0 560
michael@0 561 ; -- Final output stage
michael@0 562
michael@0 563 movdqa xmm4,xmm6
michael@0 564 paddd xmm6,xmm1 ; xmm6=data0[row0 row1 **** ****]=(C0 C1 ** **)
michael@0 565 psubd xmm4,xmm1 ; xmm4=data1[row0 row1 **** ****]=(D0 D1 ** **)
michael@0 566
michael@0 567 punpckldq xmm6,xmm4 ; xmm6=(C0 D0 C1 D1)
michael@0 568
michael@0 569 paddd xmm6,[GOTOFF(ebx,PD_DESCALE_P2_2)]
michael@0 570 psrad xmm6,DESCALE_P2_2
michael@0 571
michael@0 572 packssdw xmm6,xmm6 ; xmm6=(C0 D0 C1 D1 C0 D0 C1 D1)
michael@0 573 packsswb xmm6,xmm6 ; xmm6=(C0 D0 C1 D1 C0 D0 C1 D1 ..)
michael@0 574 paddb xmm6,[GOTOFF(ebx,PB_CENTERJSAMP)]
michael@0 575
michael@0 576 pextrw ebx,xmm6,0x00 ; ebx=(C0 D0 -- --)
michael@0 577 pextrw ecx,xmm6,0x01 ; ecx=(C1 D1 -- --)
michael@0 578
michael@0 579 mov edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
michael@0 580 mov esi, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
michael@0 581 mov WORD [edx+eax*SIZEOF_JSAMPLE], bx
michael@0 582 mov WORD [esi+eax*SIZEOF_JSAMPLE], cx
michael@0 583
michael@0 584 pop edi
michael@0 585 pop esi
michael@0 586 ; pop edx ; need not be preserved
michael@0 587 ; pop ecx ; need not be preserved
michael@0 588 pop ebx
michael@0 589 pop ebp
michael@0 590 ret
michael@0 591
michael@0 592 ; For some reason, the OS X linker does not honor the request to align the
michael@0 593 ; segment unless we do this.
michael@0 594 align 16

Mercurial Repository: The Tor Browser

Europalab SCM Repository Server

mercurial