1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libjpeg/simd/jf3dnflt.asm Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,320 @@
1.4 +;
1.5 +; jf3dnflt.asm - floating-point FDCT (3DNow!)
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 forward DCT
1.21 +; (Discrete Cosine Transform). The following code is based directly on
1.22 +; the IJG's original jfdctflt.c; see the jfdctflt.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_fdct_float_3dnow)
1.34 +
1.35 +EXTN(jconst_fdct_float_3dnow):
1.36 +
1.37 +PD_0_382 times 2 dd 0.382683432365089771728460
1.38 +PD_0_707 times 2 dd 0.707106781186547524400844
1.39 +PD_0_541 times 2 dd 0.541196100146196984399723
1.40 +PD_1_306 times 2 dd 1.306562964876376527856643
1.41 +
1.42 + alignz 16
1.43 +
1.44 +; --------------------------------------------------------------------------
1.45 + SECTION SEG_TEXT
1.46 + BITS 32
1.47 +;
1.48 +; Perform the forward DCT on one block of samples.
1.49 +;
1.50 +; GLOBAL(void)
1.51 +; jsimd_fdct_float_3dnow (FAST_FLOAT * data)
1.52 +;
1.53 +
1.54 +%define data(b) (b)+8 ; FAST_FLOAT * data
1.55 +
1.56 +%define original_ebp ebp+0
1.57 +%define wk(i) ebp-(WK_NUM-(i))*SIZEOF_MMWORD ; mmword wk[WK_NUM]
1.58 +%define WK_NUM 2
1.59 +
1.60 + align 16
1.61 + global EXTN(jsimd_fdct_float_3dnow)
1.62 +
1.63 +EXTN(jsimd_fdct_float_3dnow):
1.64 + push ebp
1.65 + mov eax,esp ; eax = original ebp
1.66 + sub esp, byte 4
1.67 + and esp, byte (-SIZEOF_MMWORD) ; align to 64 bits
1.68 + mov [esp],eax
1.69 + mov ebp,esp ; ebp = aligned ebp
1.70 + lea esp, [wk(0)]
1.71 + pushpic ebx
1.72 +; push ecx ; need not be preserved
1.73 +; push edx ; need not be preserved
1.74 +; push esi ; unused
1.75 +; push edi ; unused
1.76 +
1.77 + get_GOT ebx ; get GOT address
1.78 +
1.79 + ; ---- Pass 1: process rows.
1.80 +
1.81 + mov edx, POINTER [data(eax)] ; (FAST_FLOAT *)
1.82 + mov ecx, DCTSIZE/2
1.83 + alignx 16,7
1.84 +.rowloop:
1.85 +
1.86 + movq mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)]
1.87 + movq mm1, MMWORD [MMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)]
1.88 + movq mm2, MMWORD [MMBLOCK(0,3,edx,SIZEOF_FAST_FLOAT)]
1.89 + movq mm3, MMWORD [MMBLOCK(1,3,edx,SIZEOF_FAST_FLOAT)]
1.90 +
1.91 + ; mm0=(00 01), mm1=(10 11), mm2=(06 07), mm3=(16 17)
1.92 +
1.93 + movq mm4,mm0 ; transpose coefficients
1.94 + punpckldq mm0,mm1 ; mm0=(00 10)=data0
1.95 + punpckhdq mm4,mm1 ; mm4=(01 11)=data1
1.96 + movq mm5,mm2 ; transpose coefficients
1.97 + punpckldq mm2,mm3 ; mm2=(06 16)=data6
1.98 + punpckhdq mm5,mm3 ; mm5=(07 17)=data7
1.99 +
1.100 + movq mm6,mm4
1.101 + movq mm7,mm0
1.102 + pfsub mm4,mm2 ; mm4=data1-data6=tmp6
1.103 + pfsub mm0,mm5 ; mm0=data0-data7=tmp7
1.104 + pfadd mm6,mm2 ; mm6=data1+data6=tmp1
1.105 + pfadd mm7,mm5 ; mm7=data0+data7=tmp0
1.106 +
1.107 + movq mm1, MMWORD [MMBLOCK(0,1,edx,SIZEOF_FAST_FLOAT)]
1.108 + movq mm3, MMWORD [MMBLOCK(1,1,edx,SIZEOF_FAST_FLOAT)]
1.109 + movq mm2, MMWORD [MMBLOCK(0,2,edx,SIZEOF_FAST_FLOAT)]
1.110 + movq mm5, MMWORD [MMBLOCK(1,2,edx,SIZEOF_FAST_FLOAT)]
1.111 +
1.112 + ; mm1=(02 03), mm3=(12 13), mm2=(04 05), mm5=(14 15)
1.113 +
1.114 + movq MMWORD [wk(0)], mm4 ; wk(0)=tmp6
1.115 + movq MMWORD [wk(1)], mm0 ; wk(1)=tmp7
1.116 +
1.117 + movq mm4,mm1 ; transpose coefficients
1.118 + punpckldq mm1,mm3 ; mm1=(02 12)=data2
1.119 + punpckhdq mm4,mm3 ; mm4=(03 13)=data3
1.120 + movq mm0,mm2 ; transpose coefficients
1.121 + punpckldq mm2,mm5 ; mm2=(04 14)=data4
1.122 + punpckhdq mm0,mm5 ; mm0=(05 15)=data5
1.123 +
1.124 + movq mm3,mm4
1.125 + movq mm5,mm1
1.126 + pfadd mm4,mm2 ; mm4=data3+data4=tmp3
1.127 + pfadd mm1,mm0 ; mm1=data2+data5=tmp2
1.128 + pfsub mm3,mm2 ; mm3=data3-data4=tmp4
1.129 + pfsub mm5,mm0 ; mm5=data2-data5=tmp5
1.130 +
1.131 + ; -- Even part
1.132 +
1.133 + movq mm2,mm7
1.134 + movq mm0,mm6
1.135 + pfsub mm7,mm4 ; mm7=tmp13
1.136 + pfsub mm6,mm1 ; mm6=tmp12
1.137 + pfadd mm2,mm4 ; mm2=tmp10
1.138 + pfadd mm0,mm1 ; mm0=tmp11
1.139 +
1.140 + pfadd mm6,mm7
1.141 + pfmul mm6,[GOTOFF(ebx,PD_0_707)] ; mm6=z1
1.142 +
1.143 + movq mm4,mm2
1.144 + movq mm1,mm7
1.145 + pfsub mm2,mm0 ; mm2=data4
1.146 + pfsub mm7,mm6 ; mm7=data6
1.147 + pfadd mm4,mm0 ; mm4=data0
1.148 + pfadd mm1,mm6 ; mm1=data2
1.149 +
1.150 + movq MMWORD [MMBLOCK(0,2,edx,SIZEOF_FAST_FLOAT)], mm2
1.151 + movq MMWORD [MMBLOCK(0,3,edx,SIZEOF_FAST_FLOAT)], mm7
1.152 + movq MMWORD [MMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)], mm4
1.153 + movq MMWORD [MMBLOCK(0,1,edx,SIZEOF_FAST_FLOAT)], mm1
1.154 +
1.155 + ; -- Odd part
1.156 +
1.157 + movq mm0, MMWORD [wk(0)] ; mm0=tmp6
1.158 + movq mm6, MMWORD [wk(1)] ; mm6=tmp7
1.159 +
1.160 + pfadd mm3,mm5 ; mm3=tmp10
1.161 + pfadd mm5,mm0 ; mm5=tmp11
1.162 + pfadd mm0,mm6 ; mm0=tmp12, mm6=tmp7
1.163 +
1.164 + pfmul mm5,[GOTOFF(ebx,PD_0_707)] ; mm5=z3
1.165 +
1.166 + movq mm2,mm3 ; mm2=tmp10
1.167 + pfsub mm3,mm0
1.168 + pfmul mm3,[GOTOFF(ebx,PD_0_382)] ; mm3=z5
1.169 + pfmul mm2,[GOTOFF(ebx,PD_0_541)] ; mm2=MULTIPLY(tmp10,FIX_0_54119610)
1.170 + pfmul mm0,[GOTOFF(ebx,PD_1_306)] ; mm0=MULTIPLY(tmp12,FIX_1_30656296)
1.171 + pfadd mm2,mm3 ; mm2=z2
1.172 + pfadd mm0,mm3 ; mm0=z4
1.173 +
1.174 + movq mm7,mm6
1.175 + pfsub mm6,mm5 ; mm6=z13
1.176 + pfadd mm7,mm5 ; mm7=z11
1.177 +
1.178 + movq mm4,mm6
1.179 + movq mm1,mm7
1.180 + pfsub mm6,mm2 ; mm6=data3
1.181 + pfsub mm7,mm0 ; mm7=data7
1.182 + pfadd mm4,mm2 ; mm4=data5
1.183 + pfadd mm1,mm0 ; mm1=data1
1.184 +
1.185 + movq MMWORD [MMBLOCK(1,1,edx,SIZEOF_FAST_FLOAT)], mm6
1.186 + movq MMWORD [MMBLOCK(1,3,edx,SIZEOF_FAST_FLOAT)], mm7
1.187 + movq MMWORD [MMBLOCK(1,2,edx,SIZEOF_FAST_FLOAT)], mm4
1.188 + movq MMWORD [MMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)], mm1
1.189 +
1.190 + add edx, byte 2*DCTSIZE*SIZEOF_FAST_FLOAT
1.191 + dec ecx
1.192 + jnz near .rowloop
1.193 +
1.194 + ; ---- Pass 2: process columns.
1.195 +
1.196 + mov edx, POINTER [data(eax)] ; (FAST_FLOAT *)
1.197 + mov ecx, DCTSIZE/2
1.198 + alignx 16,7
1.199 +.columnloop:
1.200 +
1.201 + movq mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)]
1.202 + movq mm1, MMWORD [MMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)]
1.203 + movq mm2, MMWORD [MMBLOCK(6,0,edx,SIZEOF_FAST_FLOAT)]
1.204 + movq mm3, MMWORD [MMBLOCK(7,0,edx,SIZEOF_FAST_FLOAT)]
1.205 +
1.206 + ; mm0=(00 10), mm1=(01 11), mm2=(60 70), mm3=(61 71)
1.207 +
1.208 + movq mm4,mm0 ; transpose coefficients
1.209 + punpckldq mm0,mm1 ; mm0=(00 01)=data0
1.210 + punpckhdq mm4,mm1 ; mm4=(10 11)=data1
1.211 + movq mm5,mm2 ; transpose coefficients
1.212 + punpckldq mm2,mm3 ; mm2=(60 61)=data6
1.213 + punpckhdq mm5,mm3 ; mm5=(70 71)=data7
1.214 +
1.215 + movq mm6,mm4
1.216 + movq mm7,mm0
1.217 + pfsub mm4,mm2 ; mm4=data1-data6=tmp6
1.218 + pfsub mm0,mm5 ; mm0=data0-data7=tmp7
1.219 + pfadd mm6,mm2 ; mm6=data1+data6=tmp1
1.220 + pfadd mm7,mm5 ; mm7=data0+data7=tmp0
1.221 +
1.222 + movq mm1, MMWORD [MMBLOCK(2,0,edx,SIZEOF_FAST_FLOAT)]
1.223 + movq mm3, MMWORD [MMBLOCK(3,0,edx,SIZEOF_FAST_FLOAT)]
1.224 + movq mm2, MMWORD [MMBLOCK(4,0,edx,SIZEOF_FAST_FLOAT)]
1.225 + movq mm5, MMWORD [MMBLOCK(5,0,edx,SIZEOF_FAST_FLOAT)]
1.226 +
1.227 + ; mm1=(20 30), mm3=(21 31), mm2=(40 50), mm5=(41 51)
1.228 +
1.229 + movq MMWORD [wk(0)], mm4 ; wk(0)=tmp6
1.230 + movq MMWORD [wk(1)], mm0 ; wk(1)=tmp7
1.231 +
1.232 + movq mm4,mm1 ; transpose coefficients
1.233 + punpckldq mm1,mm3 ; mm1=(20 21)=data2
1.234 + punpckhdq mm4,mm3 ; mm4=(30 31)=data3
1.235 + movq mm0,mm2 ; transpose coefficients
1.236 + punpckldq mm2,mm5 ; mm2=(40 41)=data4
1.237 + punpckhdq mm0,mm5 ; mm0=(50 51)=data5
1.238 +
1.239 + movq mm3,mm4
1.240 + movq mm5,mm1
1.241 + pfadd mm4,mm2 ; mm4=data3+data4=tmp3
1.242 + pfadd mm1,mm0 ; mm1=data2+data5=tmp2
1.243 + pfsub mm3,mm2 ; mm3=data3-data4=tmp4
1.244 + pfsub mm5,mm0 ; mm5=data2-data5=tmp5
1.245 +
1.246 + ; -- Even part
1.247 +
1.248 + movq mm2,mm7
1.249 + movq mm0,mm6
1.250 + pfsub mm7,mm4 ; mm7=tmp13
1.251 + pfsub mm6,mm1 ; mm6=tmp12
1.252 + pfadd mm2,mm4 ; mm2=tmp10
1.253 + pfadd mm0,mm1 ; mm0=tmp11
1.254 +
1.255 + pfadd mm6,mm7
1.256 + pfmul mm6,[GOTOFF(ebx,PD_0_707)] ; mm6=z1
1.257 +
1.258 + movq mm4,mm2
1.259 + movq mm1,mm7
1.260 + pfsub mm2,mm0 ; mm2=data4
1.261 + pfsub mm7,mm6 ; mm7=data6
1.262 + pfadd mm4,mm0 ; mm4=data0
1.263 + pfadd mm1,mm6 ; mm1=data2
1.264 +
1.265 + movq MMWORD [MMBLOCK(4,0,edx,SIZEOF_FAST_FLOAT)], mm2
1.266 + movq MMWORD [MMBLOCK(6,0,edx,SIZEOF_FAST_FLOAT)], mm7
1.267 + movq MMWORD [MMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)], mm4
1.268 + movq MMWORD [MMBLOCK(2,0,edx,SIZEOF_FAST_FLOAT)], mm1
1.269 +
1.270 + ; -- Odd part
1.271 +
1.272 + movq mm0, MMWORD [wk(0)] ; mm0=tmp6
1.273 + movq mm6, MMWORD [wk(1)] ; mm6=tmp7
1.274 +
1.275 + pfadd mm3,mm5 ; mm3=tmp10
1.276 + pfadd mm5,mm0 ; mm5=tmp11
1.277 + pfadd mm0,mm6 ; mm0=tmp12, mm6=tmp7
1.278 +
1.279 + pfmul mm5,[GOTOFF(ebx,PD_0_707)] ; mm5=z3
1.280 +
1.281 + movq mm2,mm3 ; mm2=tmp10
1.282 + pfsub mm3,mm0
1.283 + pfmul mm3,[GOTOFF(ebx,PD_0_382)] ; mm3=z5
1.284 + pfmul mm2,[GOTOFF(ebx,PD_0_541)] ; mm2=MULTIPLY(tmp10,FIX_0_54119610)
1.285 + pfmul mm0,[GOTOFF(ebx,PD_1_306)] ; mm0=MULTIPLY(tmp12,FIX_1_30656296)
1.286 + pfadd mm2,mm3 ; mm2=z2
1.287 + pfadd mm0,mm3 ; mm0=z4
1.288 +
1.289 + movq mm7,mm6
1.290 + pfsub mm6,mm5 ; mm6=z13
1.291 + pfadd mm7,mm5 ; mm7=z11
1.292 +
1.293 + movq mm4,mm6
1.294 + movq mm1,mm7
1.295 + pfsub mm6,mm2 ; mm6=data3
1.296 + pfsub mm7,mm0 ; mm7=data7
1.297 + pfadd mm4,mm2 ; mm4=data5
1.298 + pfadd mm1,mm0 ; mm1=data1
1.299 +
1.300 + movq MMWORD [MMBLOCK(3,0,edx,SIZEOF_FAST_FLOAT)], mm6
1.301 + movq MMWORD [MMBLOCK(7,0,edx,SIZEOF_FAST_FLOAT)], mm7
1.302 + movq MMWORD [MMBLOCK(5,0,edx,SIZEOF_FAST_FLOAT)], mm4
1.303 + movq MMWORD [MMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)], mm1
1.304 +
1.305 + add edx, byte 2*SIZEOF_FAST_FLOAT
1.306 + dec ecx
1.307 + jnz near .columnloop
1.308 +
1.309 + femms ; empty MMX/3DNow! state
1.310 +
1.311 +; pop edi ; unused
1.312 +; pop esi ; unused
1.313 +; pop edx ; need not be preserved
1.314 +; pop ecx ; need not be preserved
1.315 + poppic ebx
1.316 + mov esp,ebp ; esp <- aligned ebp
1.317 + pop esp ; esp <- original ebp
1.318 + pop ebp
1.319 + ret
1.320 +
1.321 +; For some reason, the OS X linker does not honor the request to align the
1.322 +; segment unless we do this.
1.323 + align 16
