1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libjpeg/simd/jfss2fst.asm Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,404 @@
1.4 +;
1.5 +; jfss2fst.asm - fast integer FDCT (SSE2)
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 fast, not so accurate integer implementation of
1.21 +; the forward DCT (Discrete Cosine Transform). The following code is
1.22 +; based directly on the IJG's original jfdctfst.c; see the jfdctfst.c
1.23 +; for more details.
1.24 +;
1.25 +; [TAB8]
1.26 +
1.27 +%include "jsimdext.inc"
1.28 +%include "jdct.inc"
1.29 +
1.30 +; --------------------------------------------------------------------------
1.31 +
1.32 +%define CONST_BITS 8 ; 14 is also OK.
1.33 +
1.34 +%if CONST_BITS == 8
1.35 +F_0_382 equ 98 ; FIX(0.382683433)
1.36 +F_0_541 equ 139 ; FIX(0.541196100)
1.37 +F_0_707 equ 181 ; FIX(0.707106781)
1.38 +F_1_306 equ 334 ; FIX(1.306562965)
1.39 +%else
1.40 +; NASM cannot do compile-time arithmetic on floating-point constants.
1.41 +%define DESCALE(x,n) (((x)+(1<<((n)-1)))>>(n))
1.42 +F_0_382 equ DESCALE( 410903207,30-CONST_BITS) ; FIX(0.382683433)
1.43 +F_0_541 equ DESCALE( 581104887,30-CONST_BITS) ; FIX(0.541196100)
1.44 +F_0_707 equ DESCALE( 759250124,30-CONST_BITS) ; FIX(0.707106781)
1.45 +F_1_306 equ DESCALE(1402911301,30-CONST_BITS) ; FIX(1.306562965)
1.46 +%endif
1.47 +
1.48 +; --------------------------------------------------------------------------
1.49 + SECTION SEG_CONST
1.50 +
1.51 +; PRE_MULTIPLY_SCALE_BITS <= 2 (to avoid overflow)
1.52 +; CONST_BITS + CONST_SHIFT + PRE_MULTIPLY_SCALE_BITS == 16 (for pmulhw)
1.53 +
1.54 +%define PRE_MULTIPLY_SCALE_BITS 2
1.55 +%define CONST_SHIFT (16 - PRE_MULTIPLY_SCALE_BITS - CONST_BITS)
1.56 +
1.57 + alignz 16
1.58 + global EXTN(jconst_fdct_ifast_sse2)
1.59 +
1.60 +EXTN(jconst_fdct_ifast_sse2):
1.61 +
1.62 +PW_F0707 times 8 dw F_0_707 << CONST_SHIFT
1.63 +PW_F0382 times 8 dw F_0_382 << CONST_SHIFT
1.64 +PW_F0541 times 8 dw F_0_541 << CONST_SHIFT
1.65 +PW_F1306 times 8 dw F_1_306 << CONST_SHIFT
1.66 +
1.67 + alignz 16
1.68 +
1.69 +; --------------------------------------------------------------------------
1.70 + SECTION SEG_TEXT
1.71 + BITS 32
1.72 +;
1.73 +; Perform the forward DCT on one block of samples.
1.74 +;
1.75 +; GLOBAL(void)
1.76 +; jsimd_fdct_ifast_sse2 (DCTELEM * data)
1.77 +;
1.78 +
1.79 +%define data(b) (b)+8 ; DCTELEM * data
1.80 +
1.81 +%define original_ebp ebp+0
1.82 +%define wk(i) ebp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM]
1.83 +%define WK_NUM 2
1.84 +
1.85 + align 16
1.86 + global EXTN(jsimd_fdct_ifast_sse2)
1.87 +
1.88 +EXTN(jsimd_fdct_ifast_sse2):
1.89 + push ebp
1.90 + mov eax,esp ; eax = original ebp
1.91 + sub esp, byte 4
1.92 + and esp, byte (-SIZEOF_XMMWORD) ; align to 128 bits
1.93 + mov [esp],eax
1.94 + mov ebp,esp ; ebp = aligned ebp
1.95 + lea esp, [wk(0)]
1.96 + pushpic ebx
1.97 +; push ecx ; unused
1.98 +; push edx ; need not be preserved
1.99 +; push esi ; unused
1.100 +; push edi ; unused
1.101 +
1.102 + get_GOT ebx ; get GOT address
1.103 +
1.104 + ; ---- Pass 1: process rows.
1.105 +
1.106 + mov edx, POINTER [data(eax)] ; (DCTELEM *)
1.107 +
1.108 + movdqa xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_DCTELEM)]
1.109 + movdqa xmm1, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_DCTELEM)]
1.110 + movdqa xmm2, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_DCTELEM)]
1.111 + movdqa xmm3, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_DCTELEM)]
1.112 +
1.113 + ; xmm0=(00 01 02 03 04 05 06 07), xmm2=(20 21 22 23 24 25 26 27)
1.114 + ; xmm1=(10 11 12 13 14 15 16 17), xmm3=(30 31 32 33 34 35 36 37)
1.115 +
1.116 + movdqa xmm4,xmm0 ; transpose coefficients(phase 1)
1.117 + punpcklwd xmm0,xmm1 ; xmm0=(00 10 01 11 02 12 03 13)
1.118 + punpckhwd xmm4,xmm1 ; xmm4=(04 14 05 15 06 16 07 17)
1.119 + movdqa xmm5,xmm2 ; transpose coefficients(phase 1)
1.120 + punpcklwd xmm2,xmm3 ; xmm2=(20 30 21 31 22 32 23 33)
1.121 + punpckhwd xmm5,xmm3 ; xmm5=(24 34 25 35 26 36 27 37)
1.122 +
1.123 + movdqa xmm6, XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_DCTELEM)]
1.124 + movdqa xmm7, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_DCTELEM)]
1.125 + movdqa xmm1, XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_DCTELEM)]
1.126 + movdqa xmm3, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_DCTELEM)]
1.127 +
1.128 + ; xmm6=( 4 12 20 28 36 44 52 60), xmm1=( 6 14 22 30 38 46 54 62)
1.129 + ; xmm7=( 5 13 21 29 37 45 53 61), xmm3=( 7 15 23 31 39 47 55 63)
1.130 +
1.131 + movdqa XMMWORD [wk(0)], xmm2 ; wk(0)=(20 30 21 31 22 32 23 33)
1.132 + movdqa XMMWORD [wk(1)], xmm5 ; wk(1)=(24 34 25 35 26 36 27 37)
1.133 +
1.134 + movdqa xmm2,xmm6 ; transpose coefficients(phase 1)
1.135 + punpcklwd xmm6,xmm7 ; xmm6=(40 50 41 51 42 52 43 53)
1.136 + punpckhwd xmm2,xmm7 ; xmm2=(44 54 45 55 46 56 47 57)
1.137 + movdqa xmm5,xmm1 ; transpose coefficients(phase 1)
1.138 + punpcklwd xmm1,xmm3 ; xmm1=(60 70 61 71 62 72 63 73)
1.139 + punpckhwd xmm5,xmm3 ; xmm5=(64 74 65 75 66 76 67 77)
1.140 +
1.141 + movdqa xmm7,xmm6 ; transpose coefficients(phase 2)
1.142 + punpckldq xmm6,xmm1 ; xmm6=(40 50 60 70 41 51 61 71)
1.143 + punpckhdq xmm7,xmm1 ; xmm7=(42 52 62 72 43 53 63 73)
1.144 + movdqa xmm3,xmm2 ; transpose coefficients(phase 2)
1.145 + punpckldq xmm2,xmm5 ; xmm2=(44 54 64 74 45 55 65 75)
1.146 + punpckhdq xmm3,xmm5 ; xmm3=(46 56 66 76 47 57 67 77)
1.147 +
1.148 + movdqa xmm1, XMMWORD [wk(0)] ; xmm1=(20 30 21 31 22 32 23 33)
1.149 + movdqa xmm5, XMMWORD [wk(1)] ; xmm5=(24 34 25 35 26 36 27 37)
1.150 + movdqa XMMWORD [wk(0)], xmm7 ; wk(0)=(42 52 62 72 43 53 63 73)
1.151 + movdqa XMMWORD [wk(1)], xmm2 ; wk(1)=(44 54 64 74 45 55 65 75)
1.152 +
1.153 + movdqa xmm7,xmm0 ; transpose coefficients(phase 2)
1.154 + punpckldq xmm0,xmm1 ; xmm0=(00 10 20 30 01 11 21 31)
1.155 + punpckhdq xmm7,xmm1 ; xmm7=(02 12 22 32 03 13 23 33)
1.156 + movdqa xmm2,xmm4 ; transpose coefficients(phase 2)
1.157 + punpckldq xmm4,xmm5 ; xmm4=(04 14 24 34 05 15 25 35)
1.158 + punpckhdq xmm2,xmm5 ; xmm2=(06 16 26 36 07 17 27 37)
1.159 +
1.160 + movdqa xmm1,xmm0 ; transpose coefficients(phase 3)
1.161 + punpcklqdq xmm0,xmm6 ; xmm0=(00 10 20 30 40 50 60 70)=data0
1.162 + punpckhqdq xmm1,xmm6 ; xmm1=(01 11 21 31 41 51 61 71)=data1
1.163 + movdqa xmm5,xmm2 ; transpose coefficients(phase 3)
1.164 + punpcklqdq xmm2,xmm3 ; xmm2=(06 16 26 36 46 56 66 76)=data6
1.165 + punpckhqdq xmm5,xmm3 ; xmm5=(07 17 27 37 47 57 67 77)=data7
1.166 +
1.167 + movdqa xmm6,xmm1
1.168 + movdqa xmm3,xmm0
1.169 + psubw xmm1,xmm2 ; xmm1=data1-data6=tmp6
1.170 + psubw xmm0,xmm5 ; xmm0=data0-data7=tmp7
1.171 + paddw xmm6,xmm2 ; xmm6=data1+data6=tmp1
1.172 + paddw xmm3,xmm5 ; xmm3=data0+data7=tmp0
1.173 +
1.174 + movdqa xmm2, XMMWORD [wk(0)] ; xmm2=(42 52 62 72 43 53 63 73)
1.175 + movdqa xmm5, XMMWORD [wk(1)] ; xmm5=(44 54 64 74 45 55 65 75)
1.176 + movdqa XMMWORD [wk(0)], xmm1 ; wk(0)=tmp6
1.177 + movdqa XMMWORD [wk(1)], xmm0 ; wk(1)=tmp7
1.178 +
1.179 + movdqa xmm1,xmm7 ; transpose coefficients(phase 3)
1.180 + punpcklqdq xmm7,xmm2 ; xmm7=(02 12 22 32 42 52 62 72)=data2
1.181 + punpckhqdq xmm1,xmm2 ; xmm1=(03 13 23 33 43 53 63 73)=data3
1.182 + movdqa xmm0,xmm4 ; transpose coefficients(phase 3)
1.183 + punpcklqdq xmm4,xmm5 ; xmm4=(04 14 24 34 44 54 64 74)=data4
1.184 + punpckhqdq xmm0,xmm5 ; xmm0=(05 15 25 35 45 55 65 75)=data5
1.185 +
1.186 + movdqa xmm2,xmm1
1.187 + movdqa xmm5,xmm7
1.188 + paddw xmm1,xmm4 ; xmm1=data3+data4=tmp3
1.189 + paddw xmm7,xmm0 ; xmm7=data2+data5=tmp2
1.190 + psubw xmm2,xmm4 ; xmm2=data3-data4=tmp4
1.191 + psubw xmm5,xmm0 ; xmm5=data2-data5=tmp5
1.192 +
1.193 + ; -- Even part
1.194 +
1.195 + movdqa xmm4,xmm3
1.196 + movdqa xmm0,xmm6
1.197 + psubw xmm3,xmm1 ; xmm3=tmp13
1.198 + psubw xmm6,xmm7 ; xmm6=tmp12
1.199 + paddw xmm4,xmm1 ; xmm4=tmp10
1.200 + paddw xmm0,xmm7 ; xmm0=tmp11
1.201 +
1.202 + paddw xmm6,xmm3
1.203 + psllw xmm6,PRE_MULTIPLY_SCALE_BITS
1.204 + pmulhw xmm6,[GOTOFF(ebx,PW_F0707)] ; xmm6=z1
1.205 +
1.206 + movdqa xmm1,xmm4
1.207 + movdqa xmm7,xmm3
1.208 + psubw xmm4,xmm0 ; xmm4=data4
1.209 + psubw xmm3,xmm6 ; xmm3=data6
1.210 + paddw xmm1,xmm0 ; xmm1=data0
1.211 + paddw xmm7,xmm6 ; xmm7=data2
1.212 +
1.213 + movdqa xmm0, XMMWORD [wk(0)] ; xmm0=tmp6
1.214 + movdqa xmm6, XMMWORD [wk(1)] ; xmm6=tmp7
1.215 + movdqa XMMWORD [wk(0)], xmm4 ; wk(0)=data4
1.216 + movdqa XMMWORD [wk(1)], xmm3 ; wk(1)=data6
1.217 +
1.218 + ; -- Odd part
1.219 +
1.220 + paddw xmm2,xmm5 ; xmm2=tmp10
1.221 + paddw xmm5,xmm0 ; xmm5=tmp11
1.222 + paddw xmm0,xmm6 ; xmm0=tmp12, xmm6=tmp7
1.223 +
1.224 + psllw xmm2,PRE_MULTIPLY_SCALE_BITS
1.225 + psllw xmm0,PRE_MULTIPLY_SCALE_BITS
1.226 +
1.227 + psllw xmm5,PRE_MULTIPLY_SCALE_BITS
1.228 + pmulhw xmm5,[GOTOFF(ebx,PW_F0707)] ; xmm5=z3
1.229 +
1.230 + movdqa xmm4,xmm2 ; xmm4=tmp10
1.231 + psubw xmm2,xmm0
1.232 + pmulhw xmm2,[GOTOFF(ebx,PW_F0382)] ; xmm2=z5
1.233 + pmulhw xmm4,[GOTOFF(ebx,PW_F0541)] ; xmm4=MULTIPLY(tmp10,FIX_0_541196)
1.234 + pmulhw xmm0,[GOTOFF(ebx,PW_F1306)] ; xmm0=MULTIPLY(tmp12,FIX_1_306562)
1.235 + paddw xmm4,xmm2 ; xmm4=z2
1.236 + paddw xmm0,xmm2 ; xmm0=z4
1.237 +
1.238 + movdqa xmm3,xmm6
1.239 + psubw xmm6,xmm5 ; xmm6=z13
1.240 + paddw xmm3,xmm5 ; xmm3=z11
1.241 +
1.242 + movdqa xmm2,xmm6
1.243 + movdqa xmm5,xmm3
1.244 + psubw xmm6,xmm4 ; xmm6=data3
1.245 + psubw xmm3,xmm0 ; xmm3=data7
1.246 + paddw xmm2,xmm4 ; xmm2=data5
1.247 + paddw xmm5,xmm0 ; xmm5=data1
1.248 +
1.249 + ; ---- Pass 2: process columns.
1.250 +
1.251 +; mov edx, POINTER [data(eax)] ; (DCTELEM *)
1.252 +
1.253 + ; xmm1=(00 10 20 30 40 50 60 70), xmm7=(02 12 22 32 42 52 62 72)
1.254 + ; xmm5=(01 11 21 31 41 51 61 71), xmm6=(03 13 23 33 43 53 63 73)
1.255 +
1.256 + movdqa xmm4,xmm1 ; transpose coefficients(phase 1)
1.257 + punpcklwd xmm1,xmm5 ; xmm1=(00 01 10 11 20 21 30 31)
1.258 + punpckhwd xmm4,xmm5 ; xmm4=(40 41 50 51 60 61 70 71)
1.259 + movdqa xmm0,xmm7 ; transpose coefficients(phase 1)
1.260 + punpcklwd xmm7,xmm6 ; xmm7=(02 03 12 13 22 23 32 33)
1.261 + punpckhwd xmm0,xmm6 ; xmm0=(42 43 52 53 62 63 72 73)
1.262 +
1.263 + movdqa xmm5, XMMWORD [wk(0)] ; xmm5=col4
1.264 + movdqa xmm6, XMMWORD [wk(1)] ; xmm6=col6
1.265 +
1.266 + ; xmm5=(04 14 24 34 44 54 64 74), xmm6=(06 16 26 36 46 56 66 76)
1.267 + ; xmm2=(05 15 25 35 45 55 65 75), xmm3=(07 17 27 37 47 57 67 77)
1.268 +
1.269 + movdqa XMMWORD [wk(0)], xmm7 ; wk(0)=(02 03 12 13 22 23 32 33)
1.270 + movdqa XMMWORD [wk(1)], xmm0 ; wk(1)=(42 43 52 53 62 63 72 73)
1.271 +
1.272 + movdqa xmm7,xmm5 ; transpose coefficients(phase 1)
1.273 + punpcklwd xmm5,xmm2 ; xmm5=(04 05 14 15 24 25 34 35)
1.274 + punpckhwd xmm7,xmm2 ; xmm7=(44 45 54 55 64 65 74 75)
1.275 + movdqa xmm0,xmm6 ; transpose coefficients(phase 1)
1.276 + punpcklwd xmm6,xmm3 ; xmm6=(06 07 16 17 26 27 36 37)
1.277 + punpckhwd xmm0,xmm3 ; xmm0=(46 47 56 57 66 67 76 77)
1.278 +
1.279 + movdqa xmm2,xmm5 ; transpose coefficients(phase 2)
1.280 + punpckldq xmm5,xmm6 ; xmm5=(04 05 06 07 14 15 16 17)
1.281 + punpckhdq xmm2,xmm6 ; xmm2=(24 25 26 27 34 35 36 37)
1.282 + movdqa xmm3,xmm7 ; transpose coefficients(phase 2)
1.283 + punpckldq xmm7,xmm0 ; xmm7=(44 45 46 47 54 55 56 57)
1.284 + punpckhdq xmm3,xmm0 ; xmm3=(64 65 66 67 74 75 76 77)
1.285 +
1.286 + movdqa xmm6, XMMWORD [wk(0)] ; xmm6=(02 03 12 13 22 23 32 33)
1.287 + movdqa xmm0, XMMWORD [wk(1)] ; xmm0=(42 43 52 53 62 63 72 73)
1.288 + movdqa XMMWORD [wk(0)], xmm2 ; wk(0)=(24 25 26 27 34 35 36 37)
1.289 + movdqa XMMWORD [wk(1)], xmm7 ; wk(1)=(44 45 46 47 54 55 56 57)
1.290 +
1.291 + movdqa xmm2,xmm1 ; transpose coefficients(phase 2)
1.292 + punpckldq xmm1,xmm6 ; xmm1=(00 01 02 03 10 11 12 13)
1.293 + punpckhdq xmm2,xmm6 ; xmm2=(20 21 22 23 30 31 32 33)
1.294 + movdqa xmm7,xmm4 ; transpose coefficients(phase 2)
1.295 + punpckldq xmm4,xmm0 ; xmm4=(40 41 42 43 50 51 52 53)
1.296 + punpckhdq xmm7,xmm0 ; xmm7=(60 61 62 63 70 71 72 73)
1.297 +
1.298 + movdqa xmm6,xmm1 ; transpose coefficients(phase 3)
1.299 + punpcklqdq xmm1,xmm5 ; xmm1=(00 01 02 03 04 05 06 07)=data0
1.300 + punpckhqdq xmm6,xmm5 ; xmm6=(10 11 12 13 14 15 16 17)=data1
1.301 + movdqa xmm0,xmm7 ; transpose coefficients(phase 3)
1.302 + punpcklqdq xmm7,xmm3 ; xmm7=(60 61 62 63 64 65 66 67)=data6
1.303 + punpckhqdq xmm0,xmm3 ; xmm0=(70 71 72 73 74 75 76 77)=data7
1.304 +
1.305 + movdqa xmm5,xmm6
1.306 + movdqa xmm3,xmm1
1.307 + psubw xmm6,xmm7 ; xmm6=data1-data6=tmp6
1.308 + psubw xmm1,xmm0 ; xmm1=data0-data7=tmp7
1.309 + paddw xmm5,xmm7 ; xmm5=data1+data6=tmp1
1.310 + paddw xmm3,xmm0 ; xmm3=data0+data7=tmp0
1.311 +
1.312 + movdqa xmm7, XMMWORD [wk(0)] ; xmm7=(24 25 26 27 34 35 36 37)
1.313 + movdqa xmm0, XMMWORD [wk(1)] ; xmm0=(44 45 46 47 54 55 56 57)
1.314 + movdqa XMMWORD [wk(0)], xmm6 ; wk(0)=tmp6
1.315 + movdqa XMMWORD [wk(1)], xmm1 ; wk(1)=tmp7
1.316 +
1.317 + movdqa xmm6,xmm2 ; transpose coefficients(phase 3)
1.318 + punpcklqdq xmm2,xmm7 ; xmm2=(20 21 22 23 24 25 26 27)=data2
1.319 + punpckhqdq xmm6,xmm7 ; xmm6=(30 31 32 33 34 35 36 37)=data3
1.320 + movdqa xmm1,xmm4 ; transpose coefficients(phase 3)
1.321 + punpcklqdq xmm4,xmm0 ; xmm4=(40 41 42 43 44 45 46 47)=data4
1.322 + punpckhqdq xmm1,xmm0 ; xmm1=(50 51 52 53 54 55 56 57)=data5
1.323 +
1.324 + movdqa xmm7,xmm6
1.325 + movdqa xmm0,xmm2
1.326 + paddw xmm6,xmm4 ; xmm6=data3+data4=tmp3
1.327 + paddw xmm2,xmm1 ; xmm2=data2+data5=tmp2
1.328 + psubw xmm7,xmm4 ; xmm7=data3-data4=tmp4
1.329 + psubw xmm0,xmm1 ; xmm0=data2-data5=tmp5
1.330 +
1.331 + ; -- Even part
1.332 +
1.333 + movdqa xmm4,xmm3
1.334 + movdqa xmm1,xmm5
1.335 + psubw xmm3,xmm6 ; xmm3=tmp13
1.336 + psubw xmm5,xmm2 ; xmm5=tmp12
1.337 + paddw xmm4,xmm6 ; xmm4=tmp10
1.338 + paddw xmm1,xmm2 ; xmm1=tmp11
1.339 +
1.340 + paddw xmm5,xmm3
1.341 + psllw xmm5,PRE_MULTIPLY_SCALE_BITS
1.342 + pmulhw xmm5,[GOTOFF(ebx,PW_F0707)] ; xmm5=z1
1.343 +
1.344 + movdqa xmm6,xmm4
1.345 + movdqa xmm2,xmm3
1.346 + psubw xmm4,xmm1 ; xmm4=data4
1.347 + psubw xmm3,xmm5 ; xmm3=data6
1.348 + paddw xmm6,xmm1 ; xmm6=data0
1.349 + paddw xmm2,xmm5 ; xmm2=data2
1.350 +
1.351 + movdqa XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_DCTELEM)], xmm4
1.352 + movdqa XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_DCTELEM)], xmm3
1.353 + movdqa XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_DCTELEM)], xmm6
1.354 + movdqa XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_DCTELEM)], xmm2
1.355 +
1.356 + ; -- Odd part
1.357 +
1.358 + movdqa xmm1, XMMWORD [wk(0)] ; xmm1=tmp6
1.359 + movdqa xmm5, XMMWORD [wk(1)] ; xmm5=tmp7
1.360 +
1.361 + paddw xmm7,xmm0 ; xmm7=tmp10
1.362 + paddw xmm0,xmm1 ; xmm0=tmp11
1.363 + paddw xmm1,xmm5 ; xmm1=tmp12, xmm5=tmp7
1.364 +
1.365 + psllw xmm7,PRE_MULTIPLY_SCALE_BITS
1.366 + psllw xmm1,PRE_MULTIPLY_SCALE_BITS
1.367 +
1.368 + psllw xmm0,PRE_MULTIPLY_SCALE_BITS
1.369 + pmulhw xmm0,[GOTOFF(ebx,PW_F0707)] ; xmm0=z3
1.370 +
1.371 + movdqa xmm4,xmm7 ; xmm4=tmp10
1.372 + psubw xmm7,xmm1
1.373 + pmulhw xmm7,[GOTOFF(ebx,PW_F0382)] ; xmm7=z5
1.374 + pmulhw xmm4,[GOTOFF(ebx,PW_F0541)] ; xmm4=MULTIPLY(tmp10,FIX_0_541196)
1.375 + pmulhw xmm1,[GOTOFF(ebx,PW_F1306)] ; xmm1=MULTIPLY(tmp12,FIX_1_306562)
1.376 + paddw xmm4,xmm7 ; xmm4=z2
1.377 + paddw xmm1,xmm7 ; xmm1=z4
1.378 +
1.379 + movdqa xmm3,xmm5
1.380 + psubw xmm5,xmm0 ; xmm5=z13
1.381 + paddw xmm3,xmm0 ; xmm3=z11
1.382 +
1.383 + movdqa xmm6,xmm5
1.384 + movdqa xmm2,xmm3
1.385 + psubw xmm5,xmm4 ; xmm5=data3
1.386 + psubw xmm3,xmm1 ; xmm3=data7
1.387 + paddw xmm6,xmm4 ; xmm6=data5
1.388 + paddw xmm2,xmm1 ; xmm2=data1
1.389 +
1.390 + movdqa XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_DCTELEM)], xmm5
1.391 + movdqa XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_DCTELEM)], xmm3
1.392 + movdqa XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_DCTELEM)], xmm6
1.393 + movdqa XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_DCTELEM)], xmm2
1.394 +
1.395 +; pop edi ; unused
1.396 +; pop esi ; unused
1.397 +; pop edx ; need not be preserved
1.398 +; pop ecx ; unused
1.399 + poppic ebx
1.400 + mov esp,ebp ; esp <- aligned ebp
1.401 + pop esp ; esp <- original ebp
1.402 + pop ebp
1.403 + ret
1.404 +
1.405 +; For some reason, the OS X linker does not honor the request to align the
1.406 +; segment unless we do this.
1.407 + align 16
