The Tor Browser / annotate

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

media/libjpeg/simd/jisseflt.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 ; jisseflt.asm - floating-point IDCT (SSE & MMX)
michael@0 3 ;
michael@0 4 ; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
michael@0 5 ;
michael@0 6 ; Based on
michael@0 7 ; x86 SIMD extension for IJG JPEG library
michael@0 8 ; Copyright (C) 1999-2006, MIYASAKA Masaru.
michael@0 9 ; For conditions of distribution and use, see copyright notice in jsimdext.inc
michael@0 10 ;
michael@0 11 ; This file should be assembled with NASM (Netwide Assembler),
michael@0 12 ; can *not* be assembled with Microsoft's MASM or any compatible
michael@0 13 ; assembler (including Borland's Turbo Assembler).
michael@0 14 ; NASM is available from http://nasm.sourceforge.net/ or
michael@0 15 ; http://sourceforge.net/project/showfiles.php?group_id=6208
michael@0 16 ;
michael@0 17 ; This file contains a floating-point implementation of the inverse DCT
michael@0 18 ; (Discrete Cosine Transform). The following code is based directly on
michael@0 19 ; the IJG's original jidctflt.c; see the jidctflt.c for more details.
michael@0 20 ;
michael@0 21 ; [TAB8]
michael@0 22
michael@0 23 %include "jsimdext.inc"
michael@0 24 %include "jdct.inc"
michael@0 25
michael@0 26 ; --------------------------------------------------------------------------
michael@0 27
michael@0 28 %macro unpcklps2 2 ; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(0 1 4 5)
michael@0 29 shufps %1,%2,0x44
michael@0 30 %endmacro
michael@0 31
michael@0 32 %macro unpckhps2 2 ; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(2 3 6 7)
michael@0 33 shufps %1,%2,0xEE
michael@0 34 %endmacro
michael@0 35
michael@0 36 ; --------------------------------------------------------------------------
michael@0 37 SECTION SEG_CONST
michael@0 38
michael@0 39 alignz 16
michael@0 40 global EXTN(jconst_idct_float_sse)
michael@0 41
michael@0 42 EXTN(jconst_idct_float_sse):
michael@0 43
michael@0 44 PD_1_414 times 4 dd 1.414213562373095048801689
michael@0 45 PD_1_847 times 4 dd 1.847759065022573512256366
michael@0 46 PD_1_082 times 4 dd 1.082392200292393968799446
michael@0 47 PD_M2_613 times 4 dd -2.613125929752753055713286
michael@0 48 PD_0_125 times 4 dd 0.125 ; 1/8
michael@0 49 PB_CENTERJSAMP times 8 db CENTERJSAMPLE
michael@0 50
michael@0 51 alignz 16
michael@0 52
michael@0 53 ; --------------------------------------------------------------------------
michael@0 54 SECTION SEG_TEXT
michael@0 55 BITS 32
michael@0 56 ;
michael@0 57 ; Perform dequantization and inverse DCT on one block of coefficients.
michael@0 58 ;
michael@0 59 ; GLOBAL(void)
michael@0 60 ; jsimd_idct_float_sse (void * dct_table, JCOEFPTR coef_block,
michael@0 61 ; JSAMPARRAY output_buf, JDIMENSION output_col)
michael@0 62 ;
michael@0 63
michael@0 64 %define dct_table(b) (b)+8 ; void * dct_table
michael@0 65 %define coef_block(b) (b)+12 ; JCOEFPTR coef_block
michael@0 66 %define output_buf(b) (b)+16 ; JSAMPARRAY output_buf
michael@0 67 %define output_col(b) (b)+20 ; JDIMENSION output_col
michael@0 68
michael@0 69 %define original_ebp ebp+0
michael@0 70 %define wk(i) ebp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM]
michael@0 71 %define WK_NUM 2
michael@0 72 %define workspace wk(0)-DCTSIZE2*SIZEOF_FAST_FLOAT
michael@0 73 ; FAST_FLOAT workspace[DCTSIZE2]
michael@0 74
michael@0 75 align 16
michael@0 76 global EXTN(jsimd_idct_float_sse)
michael@0 77
michael@0 78 EXTN(jsimd_idct_float_sse):
michael@0 79 push ebp
michael@0 80 mov eax,esp ; eax = original ebp
michael@0 81 sub esp, byte 4
michael@0 82 and esp, byte (-SIZEOF_XMMWORD) ; align to 128 bits
michael@0 83 mov [esp],eax
michael@0 84 mov ebp,esp ; ebp = aligned ebp
michael@0 85 lea esp, [workspace]
michael@0 86 push ebx
michael@0 87 ; push ecx ; need not be preserved
michael@0 88 ; push edx ; need not be preserved
michael@0 89 push esi
michael@0 90 push edi
michael@0 91
michael@0 92 get_GOT ebx ; get GOT address
michael@0 93
michael@0 94 ; ---- Pass 1: process columns from input, store into work array.
michael@0 95
michael@0 96 ; mov eax, [original_ebp]
michael@0 97 mov edx, POINTER [dct_table(eax)] ; quantptr
michael@0 98 mov esi, JCOEFPTR [coef_block(eax)] ; inptr
michael@0 99 lea edi, [workspace] ; FAST_FLOAT * wsptr
michael@0 100 mov ecx, DCTSIZE/4 ; ctr
michael@0 101 alignx 16,7
michael@0 102 .columnloop:
michael@0 103 %ifndef NO_ZERO_COLUMN_TEST_FLOAT_SSE
michael@0 104 mov eax, DWORD [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
michael@0 105 or eax, DWORD [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
michael@0 106 jnz near .columnDCT
michael@0 107
michael@0 108 movq mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
michael@0 109 movq mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
michael@0 110 por mm0, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
michael@0 111 por mm1, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
michael@0 112 por mm0, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
michael@0 113 por mm1, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
michael@0 114 por mm0, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
michael@0 115 por mm1,mm0
michael@0 116 packsswb mm1,mm1
michael@0 117 movd eax,mm1
michael@0 118 test eax,eax
michael@0 119 jnz short .columnDCT
michael@0 120
michael@0 121 ; -- AC terms all zero
michael@0 122
michael@0 123 movq mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
michael@0 124
michael@0 125 punpckhwd mm1,mm0 ; mm1=(** 02 ** 03)
michael@0 126 punpcklwd mm0,mm0 ; mm0=(00 00 01 01)
michael@0 127 psrad mm1,(DWORD_BIT-WORD_BIT) ; mm1=in0H=(02 03)
michael@0 128 psrad mm0,(DWORD_BIT-WORD_BIT) ; mm0=in0L=(00 01)
michael@0 129 cvtpi2ps xmm3,mm1 ; xmm3=(02 03 ** **)
michael@0 130 cvtpi2ps xmm0,mm0 ; xmm0=(00 01 ** **)
michael@0 131 movlhps xmm0,xmm3 ; xmm0=in0=(00 01 02 03)
michael@0 132
michael@0 133 mulps xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
michael@0 134
michael@0 135 movaps xmm1,xmm0
michael@0 136 movaps xmm2,xmm0
michael@0 137 movaps xmm3,xmm0
michael@0 138
michael@0 139 shufps xmm0,xmm0,0x00 ; xmm0=(00 00 00 00)
michael@0 140 shufps xmm1,xmm1,0x55 ; xmm1=(01 01 01 01)
michael@0 141 shufps xmm2,xmm2,0xAA ; xmm2=(02 02 02 02)
michael@0 142 shufps xmm3,xmm3,0xFF ; xmm3=(03 03 03 03)
michael@0 143
michael@0 144 movaps XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], xmm0
michael@0 145 movaps XMMWORD [XMMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], xmm0
michael@0 146 movaps XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], xmm1
michael@0 147 movaps XMMWORD [XMMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], xmm1
michael@0 148 movaps XMMWORD [XMMBLOCK(2,0,edi,SIZEOF_FAST_FLOAT)], xmm2
michael@0 149 movaps XMMWORD [XMMBLOCK(2,1,edi,SIZEOF_FAST_FLOAT)], xmm2
michael@0 150 movaps XMMWORD [XMMBLOCK(3,0,edi,SIZEOF_FAST_FLOAT)], xmm3
michael@0 151 movaps XMMWORD [XMMBLOCK(3,1,edi,SIZEOF_FAST_FLOAT)], xmm3
michael@0 152 jmp near .nextcolumn
michael@0 153 alignx 16,7
michael@0 154 %endif
michael@0 155 .columnDCT:
michael@0 156
michael@0 157 ; -- Even part
michael@0 158
michael@0 159 movq mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
michael@0 160 movq mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
michael@0 161 movq mm2, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
michael@0 162 movq mm3, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
michael@0 163
michael@0 164 punpckhwd mm4,mm0 ; mm4=(** 02 ** 03)
michael@0 165 punpcklwd mm0,mm0 ; mm0=(00 00 01 01)
michael@0 166 punpckhwd mm5,mm1 ; mm5=(** 22 ** 23)
michael@0 167 punpcklwd mm1,mm1 ; mm1=(20 20 21 21)
michael@0 168
michael@0 169 psrad mm4,(DWORD_BIT-WORD_BIT) ; mm4=in0H=(02 03)
michael@0 170 psrad mm0,(DWORD_BIT-WORD_BIT) ; mm0=in0L=(00 01)
michael@0 171 cvtpi2ps xmm4,mm4 ; xmm4=(02 03 ** **)
michael@0 172 cvtpi2ps xmm0,mm0 ; xmm0=(00 01 ** **)
michael@0 173 psrad mm5,(DWORD_BIT-WORD_BIT) ; mm5=in2H=(22 23)
michael@0 174 psrad mm1,(DWORD_BIT-WORD_BIT) ; mm1=in2L=(20 21)
michael@0 175 cvtpi2ps xmm5,mm5 ; xmm5=(22 23 ** **)
michael@0 176 cvtpi2ps xmm1,mm1 ; xmm1=(20 21 ** **)
michael@0 177
michael@0 178 punpckhwd mm6,mm2 ; mm6=(** 42 ** 43)
michael@0 179 punpcklwd mm2,mm2 ; mm2=(40 40 41 41)
michael@0 180 punpckhwd mm7,mm3 ; mm7=(** 62 ** 63)
michael@0 181 punpcklwd mm3,mm3 ; mm3=(60 60 61 61)
michael@0 182
michael@0 183 psrad mm6,(DWORD_BIT-WORD_BIT) ; mm6=in4H=(42 43)
michael@0 184 psrad mm2,(DWORD_BIT-WORD_BIT) ; mm2=in4L=(40 41)
michael@0 185 cvtpi2ps xmm6,mm6 ; xmm6=(42 43 ** **)
michael@0 186 cvtpi2ps xmm2,mm2 ; xmm2=(40 41 ** **)
michael@0 187 psrad mm7,(DWORD_BIT-WORD_BIT) ; mm7=in6H=(62 63)
michael@0 188 psrad mm3,(DWORD_BIT-WORD_BIT) ; mm3=in6L=(60 61)
michael@0 189 cvtpi2ps xmm7,mm7 ; xmm7=(62 63 ** **)
michael@0 190 cvtpi2ps xmm3,mm3 ; xmm3=(60 61 ** **)
michael@0 191
michael@0 192 movlhps xmm0,xmm4 ; xmm0=in0=(00 01 02 03)
michael@0 193 movlhps xmm1,xmm5 ; xmm1=in2=(20 21 22 23)
michael@0 194 mulps xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
michael@0 195 mulps xmm1, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
michael@0 196
michael@0 197 movlhps xmm2,xmm6 ; xmm2=in4=(40 41 42 43)
michael@0 198 movlhps xmm3,xmm7 ; xmm3=in6=(60 61 62 63)
michael@0 199 mulps xmm2, XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
michael@0 200 mulps xmm3, XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
michael@0 201
michael@0 202 movaps xmm4,xmm0
michael@0 203 movaps xmm5,xmm1
michael@0 204 subps xmm0,xmm2 ; xmm0=tmp11
michael@0 205 subps xmm1,xmm3
michael@0 206 addps xmm4,xmm2 ; xmm4=tmp10
michael@0 207 addps xmm5,xmm3 ; xmm5=tmp13
michael@0 208
michael@0 209 mulps xmm1,[GOTOFF(ebx,PD_1_414)]
michael@0 210 subps xmm1,xmm5 ; xmm1=tmp12
michael@0 211
michael@0 212 movaps xmm6,xmm4
michael@0 213 movaps xmm7,xmm0
michael@0 214 subps xmm4,xmm5 ; xmm4=tmp3
michael@0 215 subps xmm0,xmm1 ; xmm0=tmp2
michael@0 216 addps xmm6,xmm5 ; xmm6=tmp0
michael@0 217 addps xmm7,xmm1 ; xmm7=tmp1
michael@0 218
michael@0 219 movaps XMMWORD [wk(1)], xmm4 ; tmp3
michael@0 220 movaps XMMWORD [wk(0)], xmm0 ; tmp2
michael@0 221
michael@0 222 ; -- Odd part
michael@0 223
michael@0 224 movq mm4, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
michael@0 225 movq mm0, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
michael@0 226 movq mm5, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
michael@0 227 movq mm1, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
michael@0 228
michael@0 229 punpckhwd mm6,mm4 ; mm6=(** 12 ** 13)
michael@0 230 punpcklwd mm4,mm4 ; mm4=(10 10 11 11)
michael@0 231 punpckhwd mm2,mm0 ; mm2=(** 32 ** 33)
michael@0 232 punpcklwd mm0,mm0 ; mm0=(30 30 31 31)
michael@0 233
michael@0 234 psrad mm6,(DWORD_BIT-WORD_BIT) ; mm6=in1H=(12 13)
michael@0 235 psrad mm4,(DWORD_BIT-WORD_BIT) ; mm4=in1L=(10 11)
michael@0 236 cvtpi2ps xmm4,mm6 ; xmm4=(12 13 ** **)
michael@0 237 cvtpi2ps xmm2,mm4 ; xmm2=(10 11 ** **)
michael@0 238 psrad mm2,(DWORD_BIT-WORD_BIT) ; mm2=in3H=(32 33)
michael@0 239 psrad mm0,(DWORD_BIT-WORD_BIT) ; mm0=in3L=(30 31)
michael@0 240 cvtpi2ps xmm0,mm2 ; xmm0=(32 33 ** **)
michael@0 241 cvtpi2ps xmm3,mm0 ; xmm3=(30 31 ** **)
michael@0 242
michael@0 243 punpckhwd mm7,mm5 ; mm7=(** 52 ** 53)
michael@0 244 punpcklwd mm5,mm5 ; mm5=(50 50 51 51)
michael@0 245 punpckhwd mm3,mm1 ; mm3=(** 72 ** 73)
michael@0 246 punpcklwd mm1,mm1 ; mm1=(70 70 71 71)
michael@0 247
michael@0 248 movlhps xmm2,xmm4 ; xmm2=in1=(10 11 12 13)
michael@0 249 movlhps xmm3,xmm0 ; xmm3=in3=(30 31 32 33)
michael@0 250
michael@0 251 psrad mm7,(DWORD_BIT-WORD_BIT) ; mm7=in5H=(52 53)
michael@0 252 psrad mm5,(DWORD_BIT-WORD_BIT) ; mm5=in5L=(50 51)
michael@0 253 cvtpi2ps xmm4,mm7 ; xmm4=(52 53 ** **)
michael@0 254 cvtpi2ps xmm5,mm5 ; xmm5=(50 51 ** **)
michael@0 255 psrad mm3,(DWORD_BIT-WORD_BIT) ; mm3=in7H=(72 73)
michael@0 256 psrad mm1,(DWORD_BIT-WORD_BIT) ; mm1=in7L=(70 71)
michael@0 257 cvtpi2ps xmm0,mm3 ; xmm0=(72 73 ** **)
michael@0 258 cvtpi2ps xmm1,mm1 ; xmm1=(70 71 ** **)
michael@0 259
michael@0 260 mulps xmm2, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
michael@0 261 mulps xmm3, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
michael@0 262
michael@0 263 movlhps xmm5,xmm4 ; xmm5=in5=(50 51 52 53)
michael@0 264 movlhps xmm1,xmm0 ; xmm1=in7=(70 71 72 73)
michael@0 265 mulps xmm5, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
michael@0 266 mulps xmm1, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
michael@0 267
michael@0 268 movaps xmm4,xmm2
michael@0 269 movaps xmm0,xmm5
michael@0 270 addps xmm2,xmm1 ; xmm2=z11
michael@0 271 addps xmm5,xmm3 ; xmm5=z13
michael@0 272 subps xmm4,xmm1 ; xmm4=z12
michael@0 273 subps xmm0,xmm3 ; xmm0=z10
michael@0 274
michael@0 275 movaps xmm1,xmm2
michael@0 276 subps xmm2,xmm5
michael@0 277 addps xmm1,xmm5 ; xmm1=tmp7
michael@0 278
michael@0 279 mulps xmm2,[GOTOFF(ebx,PD_1_414)] ; xmm2=tmp11
michael@0 280
michael@0 281 movaps xmm3,xmm0
michael@0 282 addps xmm0,xmm4
michael@0 283 mulps xmm0,[GOTOFF(ebx,PD_1_847)] ; xmm0=z5
michael@0 284 mulps xmm3,[GOTOFF(ebx,PD_M2_613)] ; xmm3=(z10 * -2.613125930)
michael@0 285 mulps xmm4,[GOTOFF(ebx,PD_1_082)] ; xmm4=(z12 * 1.082392200)
michael@0 286 addps xmm3,xmm0 ; xmm3=tmp12
michael@0 287 subps xmm4,xmm0 ; xmm4=tmp10
michael@0 288
michael@0 289 ; -- Final output stage
michael@0 290
michael@0 291 subps xmm3,xmm1 ; xmm3=tmp6
michael@0 292 movaps xmm5,xmm6
michael@0 293 movaps xmm0,xmm7
michael@0 294 addps xmm6,xmm1 ; xmm6=data0=(00 01 02 03)
michael@0 295 addps xmm7,xmm3 ; xmm7=data1=(10 11 12 13)
michael@0 296 subps xmm5,xmm1 ; xmm5=data7=(70 71 72 73)
michael@0 297 subps xmm0,xmm3 ; xmm0=data6=(60 61 62 63)
michael@0 298 subps xmm2,xmm3 ; xmm2=tmp5
michael@0 299
michael@0 300 movaps xmm1,xmm6 ; transpose coefficients(phase 1)
michael@0 301 unpcklps xmm6,xmm7 ; xmm6=(00 10 01 11)
michael@0 302 unpckhps xmm1,xmm7 ; xmm1=(02 12 03 13)
michael@0 303 movaps xmm3,xmm0 ; transpose coefficients(phase 1)
michael@0 304 unpcklps xmm0,xmm5 ; xmm0=(60 70 61 71)
michael@0 305 unpckhps xmm3,xmm5 ; xmm3=(62 72 63 73)
michael@0 306
michael@0 307 movaps xmm7, XMMWORD [wk(0)] ; xmm7=tmp2
michael@0 308 movaps xmm5, XMMWORD [wk(1)] ; xmm5=tmp3
michael@0 309
michael@0 310 movaps XMMWORD [wk(0)], xmm0 ; wk(0)=(60 70 61 71)
michael@0 311 movaps XMMWORD [wk(1)], xmm3 ; wk(1)=(62 72 63 73)
michael@0 312
michael@0 313 addps xmm4,xmm2 ; xmm4=tmp4
michael@0 314 movaps xmm0,xmm7
michael@0 315 movaps xmm3,xmm5
michael@0 316 addps xmm7,xmm2 ; xmm7=data2=(20 21 22 23)
michael@0 317 addps xmm5,xmm4 ; xmm5=data4=(40 41 42 43)
michael@0 318 subps xmm0,xmm2 ; xmm0=data5=(50 51 52 53)
michael@0 319 subps xmm3,xmm4 ; xmm3=data3=(30 31 32 33)
michael@0 320
michael@0 321 movaps xmm2,xmm7 ; transpose coefficients(phase 1)
michael@0 322 unpcklps xmm7,xmm3 ; xmm7=(20 30 21 31)
michael@0 323 unpckhps xmm2,xmm3 ; xmm2=(22 32 23 33)
michael@0 324 movaps xmm4,xmm5 ; transpose coefficients(phase 1)
michael@0 325 unpcklps xmm5,xmm0 ; xmm5=(40 50 41 51)
michael@0 326 unpckhps xmm4,xmm0 ; xmm4=(42 52 43 53)
michael@0 327
michael@0 328 movaps xmm3,xmm6 ; transpose coefficients(phase 2)
michael@0 329 unpcklps2 xmm6,xmm7 ; xmm6=(00 10 20 30)
michael@0 330 unpckhps2 xmm3,xmm7 ; xmm3=(01 11 21 31)
michael@0 331 movaps xmm0,xmm1 ; transpose coefficients(phase 2)
michael@0 332 unpcklps2 xmm1,xmm2 ; xmm1=(02 12 22 32)
michael@0 333 unpckhps2 xmm0,xmm2 ; xmm0=(03 13 23 33)
michael@0 334
michael@0 335 movaps xmm7, XMMWORD [wk(0)] ; xmm7=(60 70 61 71)
michael@0 336 movaps xmm2, XMMWORD [wk(1)] ; xmm2=(62 72 63 73)
michael@0 337
michael@0 338 movaps XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], xmm6
michael@0 339 movaps XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], xmm3
michael@0 340 movaps XMMWORD [XMMBLOCK(2,0,edi,SIZEOF_FAST_FLOAT)], xmm1
michael@0 341 movaps XMMWORD [XMMBLOCK(3,0,edi,SIZEOF_FAST_FLOAT)], xmm0
michael@0 342
michael@0 343 movaps xmm6,xmm5 ; transpose coefficients(phase 2)
michael@0 344 unpcklps2 xmm5,xmm7 ; xmm5=(40 50 60 70)
michael@0 345 unpckhps2 xmm6,xmm7 ; xmm6=(41 51 61 71)
michael@0 346 movaps xmm3,xmm4 ; transpose coefficients(phase 2)
michael@0 347 unpcklps2 xmm4,xmm2 ; xmm4=(42 52 62 72)
michael@0 348 unpckhps2 xmm3,xmm2 ; xmm3=(43 53 63 73)
michael@0 349
michael@0 350 movaps XMMWORD [XMMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], xmm5
michael@0 351 movaps XMMWORD [XMMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], xmm6
michael@0 352 movaps XMMWORD [XMMBLOCK(2,1,edi,SIZEOF_FAST_FLOAT)], xmm4
michael@0 353 movaps XMMWORD [XMMBLOCK(3,1,edi,SIZEOF_FAST_FLOAT)], xmm3
michael@0 354
michael@0 355 .nextcolumn:
michael@0 356 add esi, byte 4*SIZEOF_JCOEF ; coef_block
michael@0 357 add edx, byte 4*SIZEOF_FLOAT_MULT_TYPE ; quantptr
michael@0 358 add edi, 4*DCTSIZE*SIZEOF_FAST_FLOAT ; wsptr
michael@0 359 dec ecx ; ctr
michael@0 360 jnz near .columnloop
michael@0 361
michael@0 362 ; -- Prefetch the next coefficient block
michael@0 363
michael@0 364 prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 0*32]
michael@0 365 prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 1*32]
michael@0 366 prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 2*32]
michael@0 367 prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 3*32]
michael@0 368
michael@0 369 ; ---- Pass 2: process rows from work array, store into output array.
michael@0 370
michael@0 371 mov eax, [original_ebp]
michael@0 372 lea esi, [workspace] ; FAST_FLOAT * wsptr
michael@0 373 mov edi, JSAMPARRAY [output_buf(eax)] ; (JSAMPROW *)
michael@0 374 mov eax, JDIMENSION [output_col(eax)]
michael@0 375 mov ecx, DCTSIZE/4 ; ctr
michael@0 376 alignx 16,7
michael@0 377 .rowloop:
michael@0 378
michael@0 379 ; -- Even part
michael@0 380
michael@0 381 movaps xmm0, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_FAST_FLOAT)]
michael@0 382 movaps xmm1, XMMWORD [XMMBLOCK(2,0,esi,SIZEOF_FAST_FLOAT)]
michael@0 383 movaps xmm2, XMMWORD [XMMBLOCK(4,0,esi,SIZEOF_FAST_FLOAT)]
michael@0 384 movaps xmm3, XMMWORD [XMMBLOCK(6,0,esi,SIZEOF_FAST_FLOAT)]
michael@0 385
michael@0 386 movaps xmm4,xmm0
michael@0 387 movaps xmm5,xmm1
michael@0 388 subps xmm0,xmm2 ; xmm0=tmp11
michael@0 389 subps xmm1,xmm3
michael@0 390 addps xmm4,xmm2 ; xmm4=tmp10
michael@0 391 addps xmm5,xmm3 ; xmm5=tmp13
michael@0 392
michael@0 393 mulps xmm1,[GOTOFF(ebx,PD_1_414)]
michael@0 394 subps xmm1,xmm5 ; xmm1=tmp12
michael@0 395
michael@0 396 movaps xmm6,xmm4
michael@0 397 movaps xmm7,xmm0
michael@0 398 subps xmm4,xmm5 ; xmm4=tmp3
michael@0 399 subps xmm0,xmm1 ; xmm0=tmp2
michael@0 400 addps xmm6,xmm5 ; xmm6=tmp0
michael@0 401 addps xmm7,xmm1 ; xmm7=tmp1
michael@0 402
michael@0 403 movaps XMMWORD [wk(1)], xmm4 ; tmp3
michael@0 404 movaps XMMWORD [wk(0)], xmm0 ; tmp2
michael@0 405
michael@0 406 ; -- Odd part
michael@0 407
michael@0 408 movaps xmm2, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_FAST_FLOAT)]
michael@0 409 movaps xmm3, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_FAST_FLOAT)]
michael@0 410 movaps xmm5, XMMWORD [XMMBLOCK(5,0,esi,SIZEOF_FAST_FLOAT)]
michael@0 411 movaps xmm1, XMMWORD [XMMBLOCK(7,0,esi,SIZEOF_FAST_FLOAT)]
michael@0 412
michael@0 413 movaps xmm4,xmm2
michael@0 414 movaps xmm0,xmm5
michael@0 415 addps xmm2,xmm1 ; xmm2=z11
michael@0 416 addps xmm5,xmm3 ; xmm5=z13
michael@0 417 subps xmm4,xmm1 ; xmm4=z12
michael@0 418 subps xmm0,xmm3 ; xmm0=z10
michael@0 419
michael@0 420 movaps xmm1,xmm2
michael@0 421 subps xmm2,xmm5
michael@0 422 addps xmm1,xmm5 ; xmm1=tmp7
michael@0 423
michael@0 424 mulps xmm2,[GOTOFF(ebx,PD_1_414)] ; xmm2=tmp11
michael@0 425
michael@0 426 movaps xmm3,xmm0
michael@0 427 addps xmm0,xmm4
michael@0 428 mulps xmm0,[GOTOFF(ebx,PD_1_847)] ; xmm0=z5
michael@0 429 mulps xmm3,[GOTOFF(ebx,PD_M2_613)] ; xmm3=(z10 * -2.613125930)
michael@0 430 mulps xmm4,[GOTOFF(ebx,PD_1_082)] ; xmm4=(z12 * 1.082392200)
michael@0 431 addps xmm3,xmm0 ; xmm3=tmp12
michael@0 432 subps xmm4,xmm0 ; xmm4=tmp10
michael@0 433
michael@0 434 ; -- Final output stage
michael@0 435
michael@0 436 subps xmm3,xmm1 ; xmm3=tmp6
michael@0 437 movaps xmm5,xmm6
michael@0 438 movaps xmm0,xmm7
michael@0 439 addps xmm6,xmm1 ; xmm6=data0=(00 10 20 30)
michael@0 440 addps xmm7,xmm3 ; xmm7=data1=(01 11 21 31)
michael@0 441 subps xmm5,xmm1 ; xmm5=data7=(07 17 27 37)
michael@0 442 subps xmm0,xmm3 ; xmm0=data6=(06 16 26 36)
michael@0 443 subps xmm2,xmm3 ; xmm2=tmp5
michael@0 444
michael@0 445 movaps xmm1,[GOTOFF(ebx,PD_0_125)] ; xmm1=[PD_0_125]
michael@0 446
michael@0 447 mulps xmm6,xmm1 ; descale(1/8)
michael@0 448 mulps xmm7,xmm1 ; descale(1/8)
michael@0 449 mulps xmm5,xmm1 ; descale(1/8)
michael@0 450 mulps xmm0,xmm1 ; descale(1/8)
michael@0 451
michael@0 452 movhlps xmm3,xmm6
michael@0 453 movhlps xmm1,xmm7
michael@0 454 cvtps2pi mm0,xmm6 ; round to int32, mm0=data0L=(00 10)
michael@0 455 cvtps2pi mm1,xmm7 ; round to int32, mm1=data1L=(01 11)
michael@0 456 cvtps2pi mm2,xmm3 ; round to int32, mm2=data0H=(20 30)
michael@0 457 cvtps2pi mm3,xmm1 ; round to int32, mm3=data1H=(21 31)
michael@0 458 packssdw mm0,mm2 ; mm0=data0=(00 10 20 30)
michael@0 459 packssdw mm1,mm3 ; mm1=data1=(01 11 21 31)
michael@0 460
michael@0 461 movhlps xmm6,xmm5
michael@0 462 movhlps xmm7,xmm0
michael@0 463 cvtps2pi mm4,xmm5 ; round to int32, mm4=data7L=(07 17)
michael@0 464 cvtps2pi mm5,xmm0 ; round to int32, mm5=data6L=(06 16)
michael@0 465 cvtps2pi mm6,xmm6 ; round to int32, mm6=data7H=(27 37)
michael@0 466 cvtps2pi mm7,xmm7 ; round to int32, mm7=data6H=(26 36)
michael@0 467 packssdw mm4,mm6 ; mm4=data7=(07 17 27 37)
michael@0 468 packssdw mm5,mm7 ; mm5=data6=(06 16 26 36)
michael@0 469
michael@0 470 packsswb mm0,mm5 ; mm0=(00 10 20 30 06 16 26 36)
michael@0 471 packsswb mm1,mm4 ; mm1=(01 11 21 31 07 17 27 37)
michael@0 472
michael@0 473 movaps xmm3, XMMWORD [wk(0)] ; xmm3=tmp2
michael@0 474 movaps xmm1, XMMWORD [wk(1)] ; xmm1=tmp3
michael@0 475
michael@0 476 movaps xmm6,[GOTOFF(ebx,PD_0_125)] ; xmm6=[PD_0_125]
michael@0 477
michael@0 478 addps xmm4,xmm2 ; xmm4=tmp4
michael@0 479 movaps xmm5,xmm3
michael@0 480 movaps xmm0,xmm1
michael@0 481 addps xmm3,xmm2 ; xmm3=data2=(02 12 22 32)
michael@0 482 addps xmm1,xmm4 ; xmm1=data4=(04 14 24 34)
michael@0 483 subps xmm5,xmm2 ; xmm5=data5=(05 15 25 35)
michael@0 484 subps xmm0,xmm4 ; xmm0=data3=(03 13 23 33)
michael@0 485
michael@0 486 mulps xmm3,xmm6 ; descale(1/8)
michael@0 487 mulps xmm1,xmm6 ; descale(1/8)
michael@0 488 mulps xmm5,xmm6 ; descale(1/8)
michael@0 489 mulps xmm0,xmm6 ; descale(1/8)
michael@0 490
michael@0 491 movhlps xmm7,xmm3
michael@0 492 movhlps xmm2,xmm1
michael@0 493 cvtps2pi mm2,xmm3 ; round to int32, mm2=data2L=(02 12)
michael@0 494 cvtps2pi mm3,xmm1 ; round to int32, mm3=data4L=(04 14)
michael@0 495 cvtps2pi mm6,xmm7 ; round to int32, mm6=data2H=(22 32)
michael@0 496 cvtps2pi mm7,xmm2 ; round to int32, mm7=data4H=(24 34)
michael@0 497 packssdw mm2,mm6 ; mm2=data2=(02 12 22 32)
michael@0 498 packssdw mm3,mm7 ; mm3=data4=(04 14 24 34)
michael@0 499
michael@0 500 movhlps xmm4,xmm5
michael@0 501 movhlps xmm6,xmm0
michael@0 502 cvtps2pi mm5,xmm5 ; round to int32, mm5=data5L=(05 15)
michael@0 503 cvtps2pi mm4,xmm0 ; round to int32, mm4=data3L=(03 13)
michael@0 504 cvtps2pi mm6,xmm4 ; round to int32, mm6=data5H=(25 35)
michael@0 505 cvtps2pi mm7,xmm6 ; round to int32, mm7=data3H=(23 33)
michael@0 506 packssdw mm5,mm6 ; mm5=data5=(05 15 25 35)
michael@0 507 packssdw mm4,mm7 ; mm4=data3=(03 13 23 33)
michael@0 508
michael@0 509 movq mm6,[GOTOFF(ebx,PB_CENTERJSAMP)] ; mm6=[PB_CENTERJSAMP]
michael@0 510
michael@0 511 packsswb mm2,mm3 ; mm2=(02 12 22 32 04 14 24 34)
michael@0 512 packsswb mm4,mm5 ; mm4=(03 13 23 33 05 15 25 35)
michael@0 513
michael@0 514 paddb mm0,mm6
michael@0 515 paddb mm1,mm6
michael@0 516 paddb mm2,mm6
michael@0 517 paddb mm4,mm6
michael@0 518
michael@0 519 movq mm7,mm0 ; transpose coefficients(phase 1)
michael@0 520 punpcklbw mm0,mm1 ; mm0=(00 01 10 11 20 21 30 31)
michael@0 521 punpckhbw mm7,mm1 ; mm7=(06 07 16 17 26 27 36 37)
michael@0 522 movq mm3,mm2 ; transpose coefficients(phase 1)
michael@0 523 punpcklbw mm2,mm4 ; mm2=(02 03 12 13 22 23 32 33)
michael@0 524 punpckhbw mm3,mm4 ; mm3=(04 05 14 15 24 25 34 35)
michael@0 525
michael@0 526 movq mm5,mm0 ; transpose coefficients(phase 2)
michael@0 527 punpcklwd mm0,mm2 ; mm0=(00 01 02 03 10 11 12 13)
michael@0 528 punpckhwd mm5,mm2 ; mm5=(20 21 22 23 30 31 32 33)
michael@0 529 movq mm6,mm3 ; transpose coefficients(phase 2)
michael@0 530 punpcklwd mm3,mm7 ; mm3=(04 05 06 07 14 15 16 17)
michael@0 531 punpckhwd mm6,mm7 ; mm6=(24 25 26 27 34 35 36 37)
michael@0 532
michael@0 533 movq mm1,mm0 ; transpose coefficients(phase 3)
michael@0 534 punpckldq mm0,mm3 ; mm0=(00 01 02 03 04 05 06 07)
michael@0 535 punpckhdq mm1,mm3 ; mm1=(10 11 12 13 14 15 16 17)
michael@0 536 movq mm4,mm5 ; transpose coefficients(phase 3)
michael@0 537 punpckldq mm5,mm6 ; mm5=(20 21 22 23 24 25 26 27)
michael@0 538 punpckhdq mm4,mm6 ; mm4=(30 31 32 33 34 35 36 37)
michael@0 539
michael@0 540 pushpic ebx ; save GOT address
michael@0 541
michael@0 542 mov edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
michael@0 543 mov ebx, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
michael@0 544 movq MMWORD [edx+eax*SIZEOF_JSAMPLE], mm0
michael@0 545 movq MMWORD [ebx+eax*SIZEOF_JSAMPLE], mm1
michael@0 546 mov edx, JSAMPROW [edi+2*SIZEOF_JSAMPROW]
michael@0 547 mov ebx, JSAMPROW [edi+3*SIZEOF_JSAMPROW]
michael@0 548 movq MMWORD [edx+eax*SIZEOF_JSAMPLE], mm5
michael@0 549 movq MMWORD [ebx+eax*SIZEOF_JSAMPLE], mm4
michael@0 550
michael@0 551 poppic ebx ; restore GOT address
michael@0 552
michael@0 553 add esi, byte 4*SIZEOF_FAST_FLOAT ; wsptr
michael@0 554 add edi, byte 4*SIZEOF_JSAMPROW
michael@0 555 dec ecx ; ctr
michael@0 556 jnz near .rowloop
michael@0 557
michael@0 558 emms ; empty MMX state
michael@0 559
michael@0 560 pop edi
michael@0 561 pop esi
michael@0 562 ; pop edx ; need not be preserved
michael@0 563 ; pop ecx ; need not be preserved
michael@0 564 pop ebx
michael@0 565 mov esp,ebp ; esp <- aligned ebp
michael@0 566 pop esp ; esp <- original ebp
michael@0 567 pop ebp
michael@0 568 ret
michael@0 569
michael@0 570 ; For some reason, the OS X linker does not honor the request to align the
michael@0 571 ; segment unless we do this.
michael@0 572 align 16

Mercurial Repository: The Tor Browser

Europalab SCM Repository Server

mercurial