1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libjpeg/simd/ji3dnflt.asm Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,452 @@
1.4 +;
1.5 +; ji3dnflt.asm - floating-point IDCT (3DNow! & MMX)
1.6 +;
1.7 +; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
1.8 +;
1.9 +; Based on
1.10 +; x86 SIMD extension for IJG JPEG library
1.11 +; Copyright (C) 1999-2006, MIYASAKA Masaru.
1.12 +; For conditions of distribution and use, see copyright notice in jsimdext.inc
1.13 +;
1.14 +; This file should be assembled with NASM (Netwide Assembler),
1.15 +; can *not* be assembled with Microsoft's MASM or any compatible
1.16 +; assembler (including Borland's Turbo Assembler).
1.17 +; NASM is available from http://nasm.sourceforge.net/ or
1.18 +; http://sourceforge.net/project/showfiles.php?group_id=6208
1.19 +;
1.20 +; This file contains a floating-point implementation of the inverse DCT
1.21 +; (Discrete Cosine Transform). The following code is based directly on
1.22 +; the IJG's original jidctflt.c; see the jidctflt.c for more details.
1.23 +;
1.24 +; [TAB8]
1.25 +
1.26 +%include "jsimdext.inc"
1.27 +%include "jdct.inc"
1.28 +
1.29 +; --------------------------------------------------------------------------
1.30 + SECTION SEG_CONST
1.31 +
1.32 + alignz 16
1.33 + global EXTN(jconst_idct_float_3dnow)
1.34 +
1.35 +EXTN(jconst_idct_float_3dnow):
1.36 +
1.37 +PD_1_414 times 2 dd 1.414213562373095048801689
1.38 +PD_1_847 times 2 dd 1.847759065022573512256366
1.39 +PD_1_082 times 2 dd 1.082392200292393968799446
1.40 +PD_2_613 times 2 dd 2.613125929752753055713286
1.41 +PD_RNDINT_MAGIC times 2 dd 100663296.0 ; (float)(0x00C00000 << 3)
1.42 +PB_CENTERJSAMP times 8 db CENTERJSAMPLE
1.43 +
1.44 + alignz 16
1.45 +
1.46 +; --------------------------------------------------------------------------
1.47 + SECTION SEG_TEXT
1.48 + BITS 32
1.49 +;
1.50 +; Perform dequantization and inverse DCT on one block of coefficients.
1.51 +;
1.52 +; GLOBAL(void)
1.53 +; jsimd_idct_float_3dnow (void * dct_table, JCOEFPTR coef_block,
1.54 +; JSAMPARRAY output_buf, JDIMENSION output_col)
1.55 +;
1.56 +
1.57 +%define dct_table(b) (b)+8 ; void * dct_table
1.58 +%define coef_block(b) (b)+12 ; JCOEFPTR coef_block
1.59 +%define output_buf(b) (b)+16 ; JSAMPARRAY output_buf
1.60 +%define output_col(b) (b)+20 ; JDIMENSION output_col
1.61 +
1.62 +%define original_ebp ebp+0
1.63 +%define wk(i) ebp-(WK_NUM-(i))*SIZEOF_MMWORD ; mmword wk[WK_NUM]
1.64 +%define WK_NUM 2
1.65 +%define workspace wk(0)-DCTSIZE2*SIZEOF_FAST_FLOAT
1.66 + ; FAST_FLOAT workspace[DCTSIZE2]
1.67 +
1.68 + align 16
1.69 + global EXTN(jsimd_idct_float_3dnow)
1.70 +
1.71 +EXTN(jsimd_idct_float_3dnow):
1.72 + push ebp
1.73 + mov eax,esp ; eax = original ebp
1.74 + sub esp, byte 4
1.75 + and esp, byte (-SIZEOF_MMWORD) ; align to 64 bits
1.76 + mov [esp],eax
1.77 + mov ebp,esp ; ebp = aligned ebp
1.78 + lea esp, [workspace]
1.79 + push ebx
1.80 +; push ecx ; need not be preserved
1.81 +; push edx ; need not be preserved
1.82 + push esi
1.83 + push edi
1.84 +
1.85 + get_GOT ebx ; get GOT address
1.86 +
1.87 + ; ---- Pass 1: process columns from input, store into work array.
1.88 +
1.89 +; mov eax, [original_ebp]
1.90 + mov edx, POINTER [dct_table(eax)] ; quantptr
1.91 + mov esi, JCOEFPTR [coef_block(eax)] ; inptr
1.92 + lea edi, [workspace] ; FAST_FLOAT * wsptr
1.93 + mov ecx, DCTSIZE/2 ; ctr
1.94 + alignx 16,7
1.95 +.columnloop:
1.96 +%ifndef NO_ZERO_COLUMN_TEST_FLOAT_3DNOW
1.97 + mov eax, DWORD [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
1.98 + or eax, DWORD [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
1.99 + jnz short .columnDCT
1.100 +
1.101 + pushpic ebx ; save GOT address
1.102 + mov ebx, DWORD [DWBLOCK(3,0,esi,SIZEOF_JCOEF)]
1.103 + mov eax, DWORD [DWBLOCK(4,0,esi,SIZEOF_JCOEF)]
1.104 + or ebx, DWORD [DWBLOCK(5,0,esi,SIZEOF_JCOEF)]
1.105 + or eax, DWORD [DWBLOCK(6,0,esi,SIZEOF_JCOEF)]
1.106 + or ebx, DWORD [DWBLOCK(7,0,esi,SIZEOF_JCOEF)]
1.107 + or eax,ebx
1.108 + poppic ebx ; restore GOT address
1.109 + jnz short .columnDCT
1.110 +
1.111 + ; -- AC terms all zero
1.112 +
1.113 + movd mm0, DWORD [DWBLOCK(0,0,esi,SIZEOF_JCOEF)]
1.114 +
1.115 + punpcklwd mm0,mm0
1.116 + psrad mm0,(DWORD_BIT-WORD_BIT)
1.117 + pi2fd mm0,mm0
1.118 +
1.119 + pfmul mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
1.120 +
1.121 + movq mm1,mm0
1.122 + punpckldq mm0,mm0
1.123 + punpckhdq mm1,mm1
1.124 +
1.125 + movq MMWORD [MMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], mm0
1.126 + movq MMWORD [MMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], mm0
1.127 + movq MMWORD [MMBLOCK(0,2,edi,SIZEOF_FAST_FLOAT)], mm0
1.128 + movq MMWORD [MMBLOCK(0,3,edi,SIZEOF_FAST_FLOAT)], mm0
1.129 + movq MMWORD [MMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], mm1
1.130 + movq MMWORD [MMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], mm1
1.131 + movq MMWORD [MMBLOCK(1,2,edi,SIZEOF_FAST_FLOAT)], mm1
1.132 + movq MMWORD [MMBLOCK(1,3,edi,SIZEOF_FAST_FLOAT)], mm1
1.133 + jmp near .nextcolumn
1.134 + alignx 16,7
1.135 +%endif
1.136 +.columnDCT:
1.137 +
1.138 + ; -- Even part
1.139 +
1.140 + movd mm0, DWORD [DWBLOCK(0,0,esi,SIZEOF_JCOEF)]
1.141 + movd mm1, DWORD [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
1.142 + movd mm2, DWORD [DWBLOCK(4,0,esi,SIZEOF_JCOEF)]
1.143 + movd mm3, DWORD [DWBLOCK(6,0,esi,SIZEOF_JCOEF)]
1.144 +
1.145 + punpcklwd mm0,mm0
1.146 + punpcklwd mm1,mm1
1.147 + psrad mm0,(DWORD_BIT-WORD_BIT)
1.148 + psrad mm1,(DWORD_BIT-WORD_BIT)
1.149 + pi2fd mm0,mm0
1.150 + pi2fd mm1,mm1
1.151 +
1.152 + pfmul mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
1.153 + pfmul mm1, MMWORD [MMBLOCK(2,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
1.154 +
1.155 + punpcklwd mm2,mm2
1.156 + punpcklwd mm3,mm3
1.157 + psrad mm2,(DWORD_BIT-WORD_BIT)
1.158 + psrad mm3,(DWORD_BIT-WORD_BIT)
1.159 + pi2fd mm2,mm2
1.160 + pi2fd mm3,mm3
1.161 +
1.162 + pfmul mm2, MMWORD [MMBLOCK(4,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
1.163 + pfmul mm3, MMWORD [MMBLOCK(6,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
1.164 +
1.165 + movq mm4,mm0
1.166 + movq mm5,mm1
1.167 + pfsub mm0,mm2 ; mm0=tmp11
1.168 + pfsub mm1,mm3
1.169 + pfadd mm4,mm2 ; mm4=tmp10
1.170 + pfadd mm5,mm3 ; mm5=tmp13
1.171 +
1.172 + pfmul mm1,[GOTOFF(ebx,PD_1_414)]
1.173 + pfsub mm1,mm5 ; mm1=tmp12
1.174 +
1.175 + movq mm6,mm4
1.176 + movq mm7,mm0
1.177 + pfsub mm4,mm5 ; mm4=tmp3
1.178 + pfsub mm0,mm1 ; mm0=tmp2
1.179 + pfadd mm6,mm5 ; mm6=tmp0
1.180 + pfadd mm7,mm1 ; mm7=tmp1
1.181 +
1.182 + movq MMWORD [wk(1)], mm4 ; tmp3
1.183 + movq MMWORD [wk(0)], mm0 ; tmp2
1.184 +
1.185 + ; -- Odd part
1.186 +
1.187 + movd mm2, DWORD [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
1.188 + movd mm3, DWORD [DWBLOCK(3,0,esi,SIZEOF_JCOEF)]
1.189 + movd mm5, DWORD [DWBLOCK(5,0,esi,SIZEOF_JCOEF)]
1.190 + movd mm1, DWORD [DWBLOCK(7,0,esi,SIZEOF_JCOEF)]
1.191 +
1.192 + punpcklwd mm2,mm2
1.193 + punpcklwd mm3,mm3
1.194 + psrad mm2,(DWORD_BIT-WORD_BIT)
1.195 + psrad mm3,(DWORD_BIT-WORD_BIT)
1.196 + pi2fd mm2,mm2
1.197 + pi2fd mm3,mm3
1.198 +
1.199 + pfmul mm2, MMWORD [MMBLOCK(1,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
1.200 + pfmul mm3, MMWORD [MMBLOCK(3,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
1.201 +
1.202 + punpcklwd mm5,mm5
1.203 + punpcklwd mm1,mm1
1.204 + psrad mm5,(DWORD_BIT-WORD_BIT)
1.205 + psrad mm1,(DWORD_BIT-WORD_BIT)
1.206 + pi2fd mm5,mm5
1.207 + pi2fd mm1,mm1
1.208 +
1.209 + pfmul mm5, MMWORD [MMBLOCK(5,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
1.210 + pfmul mm1, MMWORD [MMBLOCK(7,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
1.211 +
1.212 + movq mm4,mm2
1.213 + movq mm0,mm5
1.214 + pfadd mm2,mm1 ; mm2=z11
1.215 + pfadd mm5,mm3 ; mm5=z13
1.216 + pfsub mm4,mm1 ; mm4=z12
1.217 + pfsub mm0,mm3 ; mm0=z10
1.218 +
1.219 + movq mm1,mm2
1.220 + pfsub mm2,mm5
1.221 + pfadd mm1,mm5 ; mm1=tmp7
1.222 +
1.223 + pfmul mm2,[GOTOFF(ebx,PD_1_414)] ; mm2=tmp11
1.224 +
1.225 + movq mm3,mm0
1.226 + pfadd mm0,mm4
1.227 + pfmul mm0,[GOTOFF(ebx,PD_1_847)] ; mm0=z5
1.228 + pfmul mm3,[GOTOFF(ebx,PD_2_613)] ; mm3=(z10 * 2.613125930)
1.229 + pfmul mm4,[GOTOFF(ebx,PD_1_082)] ; mm4=(z12 * 1.082392200)
1.230 + pfsubr mm3,mm0 ; mm3=tmp12
1.231 + pfsub mm4,mm0 ; mm4=tmp10
1.232 +
1.233 + ; -- Final output stage
1.234 +
1.235 + pfsub mm3,mm1 ; mm3=tmp6
1.236 + movq mm5,mm6
1.237 + movq mm0,mm7
1.238 + pfadd mm6,mm1 ; mm6=data0=(00 01)
1.239 + pfadd mm7,mm3 ; mm7=data1=(10 11)
1.240 + pfsub mm5,mm1 ; mm5=data7=(70 71)
1.241 + pfsub mm0,mm3 ; mm0=data6=(60 61)
1.242 + pfsub mm2,mm3 ; mm2=tmp5
1.243 +
1.244 + movq mm1,mm6 ; transpose coefficients
1.245 + punpckldq mm6,mm7 ; mm6=(00 10)
1.246 + punpckhdq mm1,mm7 ; mm1=(01 11)
1.247 + movq mm3,mm0 ; transpose coefficients
1.248 + punpckldq mm0,mm5 ; mm0=(60 70)
1.249 + punpckhdq mm3,mm5 ; mm3=(61 71)
1.250 +
1.251 + movq MMWORD [MMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], mm6
1.252 + movq MMWORD [MMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], mm1
1.253 + movq MMWORD [MMBLOCK(0,3,edi,SIZEOF_FAST_FLOAT)], mm0
1.254 + movq MMWORD [MMBLOCK(1,3,edi,SIZEOF_FAST_FLOAT)], mm3
1.255 +
1.256 + movq mm7, MMWORD [wk(0)] ; mm7=tmp2
1.257 + movq mm5, MMWORD [wk(1)] ; mm5=tmp3
1.258 +
1.259 + pfadd mm4,mm2 ; mm4=tmp4
1.260 + movq mm6,mm7
1.261 + movq mm1,mm5
1.262 + pfadd mm7,mm2 ; mm7=data2=(20 21)
1.263 + pfadd mm5,mm4 ; mm5=data4=(40 41)
1.264 + pfsub mm6,mm2 ; mm6=data5=(50 51)
1.265 + pfsub mm1,mm4 ; mm1=data3=(30 31)
1.266 +
1.267 + movq mm0,mm7 ; transpose coefficients
1.268 + punpckldq mm7,mm1 ; mm7=(20 30)
1.269 + punpckhdq mm0,mm1 ; mm0=(21 31)
1.270 + movq mm3,mm5 ; transpose coefficients
1.271 + punpckldq mm5,mm6 ; mm5=(40 50)
1.272 + punpckhdq mm3,mm6 ; mm3=(41 51)
1.273 +
1.274 + movq MMWORD [MMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], mm7
1.275 + movq MMWORD [MMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], mm0
1.276 + movq MMWORD [MMBLOCK(0,2,edi,SIZEOF_FAST_FLOAT)], mm5
1.277 + movq MMWORD [MMBLOCK(1,2,edi,SIZEOF_FAST_FLOAT)], mm3
1.278 +
1.279 +.nextcolumn:
1.280 + add esi, byte 2*SIZEOF_JCOEF ; coef_block
1.281 + add edx, byte 2*SIZEOF_FLOAT_MULT_TYPE ; quantptr
1.282 + add edi, byte 2*DCTSIZE*SIZEOF_FAST_FLOAT ; wsptr
1.283 + dec ecx ; ctr
1.284 + jnz near .columnloop
1.285 +
1.286 + ; -- Prefetch the next coefficient block
1.287 +
1.288 + prefetch [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 0*32]
1.289 + prefetch [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 1*32]
1.290 + prefetch [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 2*32]
1.291 + prefetch [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 3*32]
1.292 +
1.293 + ; ---- Pass 2: process rows from work array, store into output array.
1.294 +
1.295 + mov eax, [original_ebp]
1.296 + lea esi, [workspace] ; FAST_FLOAT * wsptr
1.297 + mov edi, JSAMPARRAY [output_buf(eax)] ; (JSAMPROW *)
1.298 + mov eax, JDIMENSION [output_col(eax)]
1.299 + mov ecx, DCTSIZE/2 ; ctr
1.300 + alignx 16,7
1.301 +.rowloop:
1.302 +
1.303 + ; -- Even part
1.304 +
1.305 + movq mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_FAST_FLOAT)]
1.306 + movq mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_FAST_FLOAT)]
1.307 + movq mm2, MMWORD [MMBLOCK(4,0,esi,SIZEOF_FAST_FLOAT)]
1.308 + movq mm3, MMWORD [MMBLOCK(6,0,esi,SIZEOF_FAST_FLOAT)]
1.309 +
1.310 + movq mm4,mm0
1.311 + movq mm5,mm1
1.312 + pfsub mm0,mm2 ; mm0=tmp11
1.313 + pfsub mm1,mm3
1.314 + pfadd mm4,mm2 ; mm4=tmp10
1.315 + pfadd mm5,mm3 ; mm5=tmp13
1.316 +
1.317 + pfmul mm1,[GOTOFF(ebx,PD_1_414)]
1.318 + pfsub mm1,mm5 ; mm1=tmp12
1.319 +
1.320 + movq mm6,mm4
1.321 + movq mm7,mm0
1.322 + pfsub mm4,mm5 ; mm4=tmp3
1.323 + pfsub mm0,mm1 ; mm0=tmp2
1.324 + pfadd mm6,mm5 ; mm6=tmp0
1.325 + pfadd mm7,mm1 ; mm7=tmp1
1.326 +
1.327 + movq MMWORD [wk(1)], mm4 ; tmp3
1.328 + movq MMWORD [wk(0)], mm0 ; tmp2
1.329 +
1.330 + ; -- Odd part
1.331 +
1.332 + movq mm2, MMWORD [MMBLOCK(1,0,esi,SIZEOF_FAST_FLOAT)]
1.333 + movq mm3, MMWORD [MMBLOCK(3,0,esi,SIZEOF_FAST_FLOAT)]
1.334 + movq mm5, MMWORD [MMBLOCK(5,0,esi,SIZEOF_FAST_FLOAT)]
1.335 + movq mm1, MMWORD [MMBLOCK(7,0,esi,SIZEOF_FAST_FLOAT)]
1.336 +
1.337 + movq mm4,mm2
1.338 + movq mm0,mm5
1.339 + pfadd mm2,mm1 ; mm2=z11
1.340 + pfadd mm5,mm3 ; mm5=z13
1.341 + pfsub mm4,mm1 ; mm4=z12
1.342 + pfsub mm0,mm3 ; mm0=z10
1.343 +
1.344 + movq mm1,mm2
1.345 + pfsub mm2,mm5
1.346 + pfadd mm1,mm5 ; mm1=tmp7
1.347 +
1.348 + pfmul mm2,[GOTOFF(ebx,PD_1_414)] ; mm2=tmp11
1.349 +
1.350 + movq mm3,mm0
1.351 + pfadd mm0,mm4
1.352 + pfmul mm0,[GOTOFF(ebx,PD_1_847)] ; mm0=z5
1.353 + pfmul mm3,[GOTOFF(ebx,PD_2_613)] ; mm3=(z10 * 2.613125930)
1.354 + pfmul mm4,[GOTOFF(ebx,PD_1_082)] ; mm4=(z12 * 1.082392200)
1.355 + pfsubr mm3,mm0 ; mm3=tmp12
1.356 + pfsub mm4,mm0 ; mm4=tmp10
1.357 +
1.358 + ; -- Final output stage
1.359 +
1.360 + pfsub mm3,mm1 ; mm3=tmp6
1.361 + movq mm5,mm6
1.362 + movq mm0,mm7
1.363 + pfadd mm6,mm1 ; mm6=data0=(00 10)
1.364 + pfadd mm7,mm3 ; mm7=data1=(01 11)
1.365 + pfsub mm5,mm1 ; mm5=data7=(07 17)
1.366 + pfsub mm0,mm3 ; mm0=data6=(06 16)
1.367 + pfsub mm2,mm3 ; mm2=tmp5
1.368 +
1.369 + movq mm1,[GOTOFF(ebx,PD_RNDINT_MAGIC)] ; mm1=[PD_RNDINT_MAGIC]
1.370 + pcmpeqd mm3,mm3
1.371 + psrld mm3,WORD_BIT ; mm3={0xFFFF 0x0000 0xFFFF 0x0000}
1.372 +
1.373 + pfadd mm6,mm1 ; mm6=roundint(data0/8)=(00 ** 10 **)
1.374 + pfadd mm7,mm1 ; mm7=roundint(data1/8)=(01 ** 11 **)
1.375 + pfadd mm0,mm1 ; mm0=roundint(data6/8)=(06 ** 16 **)
1.376 + pfadd mm5,mm1 ; mm5=roundint(data7/8)=(07 ** 17 **)
1.377 +
1.378 + pand mm6,mm3 ; mm6=(00 -- 10 --)
1.379 + pslld mm7,WORD_BIT ; mm7=(-- 01 -- 11)
1.380 + pand mm0,mm3 ; mm0=(06 -- 16 --)
1.381 + pslld mm5,WORD_BIT ; mm5=(-- 07 -- 17)
1.382 + por mm6,mm7 ; mm6=(00 01 10 11)
1.383 + por mm0,mm5 ; mm0=(06 07 16 17)
1.384 +
1.385 + movq mm1, MMWORD [wk(0)] ; mm1=tmp2
1.386 + movq mm3, MMWORD [wk(1)] ; mm3=tmp3
1.387 +
1.388 + pfadd mm4,mm2 ; mm4=tmp4
1.389 + movq mm7,mm1
1.390 + movq mm5,mm3
1.391 + pfadd mm1,mm2 ; mm1=data2=(02 12)
1.392 + pfadd mm3,mm4 ; mm3=data4=(04 14)
1.393 + pfsub mm7,mm2 ; mm7=data5=(05 15)
1.394 + pfsub mm5,mm4 ; mm5=data3=(03 13)
1.395 +
1.396 + movq mm2,[GOTOFF(ebx,PD_RNDINT_MAGIC)] ; mm2=[PD_RNDINT_MAGIC]
1.397 + pcmpeqd mm4,mm4
1.398 + psrld mm4,WORD_BIT ; mm4={0xFFFF 0x0000 0xFFFF 0x0000}
1.399 +
1.400 + pfadd mm3,mm2 ; mm3=roundint(data4/8)=(04 ** 14 **)
1.401 + pfadd mm7,mm2 ; mm7=roundint(data5/8)=(05 ** 15 **)
1.402 + pfadd mm1,mm2 ; mm1=roundint(data2/8)=(02 ** 12 **)
1.403 + pfadd mm5,mm2 ; mm5=roundint(data3/8)=(03 ** 13 **)
1.404 +
1.405 + pand mm3,mm4 ; mm3=(04 -- 14 --)
1.406 + pslld mm7,WORD_BIT ; mm7=(-- 05 -- 15)
1.407 + pand mm1,mm4 ; mm1=(02 -- 12 --)
1.408 + pslld mm5,WORD_BIT ; mm5=(-- 03 -- 13)
1.409 + por mm3,mm7 ; mm3=(04 05 14 15)
1.410 + por mm1,mm5 ; mm1=(02 03 12 13)
1.411 +
1.412 + movq mm2,[GOTOFF(ebx,PB_CENTERJSAMP)] ; mm2=[PB_CENTERJSAMP]
1.413 +
1.414 + packsswb mm6,mm3 ; mm6=(00 01 10 11 04 05 14 15)
1.415 + packsswb mm1,mm0 ; mm1=(02 03 12 13 06 07 16 17)
1.416 + paddb mm6,mm2
1.417 + paddb mm1,mm2
1.418 +
1.419 + movq mm4,mm6 ; transpose coefficients(phase 2)
1.420 + punpcklwd mm6,mm1 ; mm6=(00 01 02 03 10 11 12 13)
1.421 + punpckhwd mm4,mm1 ; mm4=(04 05 06 07 14 15 16 17)
1.422 +
1.423 + movq mm7,mm6 ; transpose coefficients(phase 3)
1.424 + punpckldq mm6,mm4 ; mm6=(00 01 02 03 04 05 06 07)
1.425 + punpckhdq mm7,mm4 ; mm7=(10 11 12 13 14 15 16 17)
1.426 +
1.427 + pushpic ebx ; save GOT address
1.428 +
1.429 + mov edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
1.430 + mov ebx, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
1.431 + movq MMWORD [edx+eax*SIZEOF_JSAMPLE], mm6
1.432 + movq MMWORD [ebx+eax*SIZEOF_JSAMPLE], mm7
1.433 +
1.434 + poppic ebx ; restore GOT address
1.435 +
1.436 + add esi, byte 2*SIZEOF_FAST_FLOAT ; wsptr
1.437 + add edi, byte 2*SIZEOF_JSAMPROW
1.438 + dec ecx ; ctr
1.439 + jnz near .rowloop
1.440 +
1.441 + femms ; empty MMX/3DNow! state
1.442 +
1.443 + pop edi
1.444 + pop esi
1.445 +; pop edx ; need not be preserved
1.446 +; pop ecx ; need not be preserved
1.447 + pop ebx
1.448 + mov esp,ebp ; esp <- aligned ebp
1.449 + pop esp ; esp <- original ebp
1.450 + pop ebp
1.451 + ret
1.452 +
1.453 +; For some reason, the OS X linker does not honor the request to align the
1.454 +; segment unless we do this.
1.455 + align 16
