1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libjpeg/simd/jfss2int-64.asm Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,622 @@
1.4 +;
1.5 +; jfss2int-64.asm - accurate integer FDCT (64-bit SSE2)
1.6 +;
1.7 +; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
1.8 +; Copyright 2009 D. R. Commander
1.9 +;
1.10 +; Based on
1.11 +; x86 SIMD extension for IJG JPEG library
1.12 +; Copyright (C) 1999-2006, MIYASAKA Masaru.
1.13 +; For conditions of distribution and use, see copyright notice in jsimdext.inc
1.14 +;
1.15 +; This file should be assembled with NASM (Netwide Assembler),
1.16 +; can *not* be assembled with Microsoft's MASM or any compatible
1.17 +; assembler (including Borland's Turbo Assembler).
1.18 +; NASM is available from http://nasm.sourceforge.net/ or
1.19 +; http://sourceforge.net/project/showfiles.php?group_id=6208
1.20 +;
1.21 +; This file contains a slow-but-accurate integer implementation of the
1.22 +; forward DCT (Discrete Cosine Transform). The following code is based
1.23 +; directly on the IJG's original jfdctint.c; see the jfdctint.c for
1.24 +; more details.
1.25 +;
1.26 +; [TAB8]
1.27 +
1.28 +%include "jsimdext.inc"
1.29 +%include "jdct.inc"
1.30 +
1.31 +; --------------------------------------------------------------------------
1.32 +
1.33 +%define CONST_BITS 13
1.34 +%define PASS1_BITS 2
1.35 +
1.36 +%define DESCALE_P1 (CONST_BITS-PASS1_BITS)
1.37 +%define DESCALE_P2 (CONST_BITS+PASS1_BITS)
1.38 +
1.39 +%if CONST_BITS == 13
1.40 +F_0_298 equ 2446 ; FIX(0.298631336)
1.41 +F_0_390 equ 3196 ; FIX(0.390180644)
1.42 +F_0_541 equ 4433 ; FIX(0.541196100)
1.43 +F_0_765 equ 6270 ; FIX(0.765366865)
1.44 +F_0_899 equ 7373 ; FIX(0.899976223)
1.45 +F_1_175 equ 9633 ; FIX(1.175875602)
1.46 +F_1_501 equ 12299 ; FIX(1.501321110)
1.47 +F_1_847 equ 15137 ; FIX(1.847759065)
1.48 +F_1_961 equ 16069 ; FIX(1.961570560)
1.49 +F_2_053 equ 16819 ; FIX(2.053119869)
1.50 +F_2_562 equ 20995 ; FIX(2.562915447)
1.51 +F_3_072 equ 25172 ; FIX(3.072711026)
1.52 +%else
1.53 +; NASM cannot do compile-time arithmetic on floating-point constants.
1.54 +%define DESCALE(x,n) (((x)+(1<<((n)-1)))>>(n))
1.55 +F_0_298 equ DESCALE( 320652955,30-CONST_BITS) ; FIX(0.298631336)
1.56 +F_0_390 equ DESCALE( 418953276,30-CONST_BITS) ; FIX(0.390180644)
1.57 +F_0_541 equ DESCALE( 581104887,30-CONST_BITS) ; FIX(0.541196100)
1.58 +F_0_765 equ DESCALE( 821806413,30-CONST_BITS) ; FIX(0.765366865)
1.59 +F_0_899 equ DESCALE( 966342111,30-CONST_BITS) ; FIX(0.899976223)
1.60 +F_1_175 equ DESCALE(1262586813,30-CONST_BITS) ; FIX(1.175875602)
1.61 +F_1_501 equ DESCALE(1612031267,30-CONST_BITS) ; FIX(1.501321110)
1.62 +F_1_847 equ DESCALE(1984016188,30-CONST_BITS) ; FIX(1.847759065)
1.63 +F_1_961 equ DESCALE(2106220350,30-CONST_BITS) ; FIX(1.961570560)
1.64 +F_2_053 equ DESCALE(2204520673,30-CONST_BITS) ; FIX(2.053119869)
1.65 +F_2_562 equ DESCALE(2751909506,30-CONST_BITS) ; FIX(2.562915447)
1.66 +F_3_072 equ DESCALE(3299298341,30-CONST_BITS) ; FIX(3.072711026)
1.67 +%endif
1.68 +
1.69 +; --------------------------------------------------------------------------
1.70 + SECTION SEG_CONST
1.71 +
1.72 + alignz 16
1.73 + global EXTN(jconst_fdct_islow_sse2)
1.74 +
1.75 +EXTN(jconst_fdct_islow_sse2):
1.76 +
1.77 +PW_F130_F054 times 4 dw (F_0_541+F_0_765), F_0_541
1.78 +PW_F054_MF130 times 4 dw F_0_541, (F_0_541-F_1_847)
1.79 +PW_MF078_F117 times 4 dw (F_1_175-F_1_961), F_1_175
1.80 +PW_F117_F078 times 4 dw F_1_175, (F_1_175-F_0_390)
1.81 +PW_MF060_MF089 times 4 dw (F_0_298-F_0_899),-F_0_899
1.82 +PW_MF089_F060 times 4 dw -F_0_899, (F_1_501-F_0_899)
1.83 +PW_MF050_MF256 times 4 dw (F_2_053-F_2_562),-F_2_562
1.84 +PW_MF256_F050 times 4 dw -F_2_562, (F_3_072-F_2_562)
1.85 +PD_DESCALE_P1 times 4 dd 1 << (DESCALE_P1-1)
1.86 +PD_DESCALE_P2 times 4 dd 1 << (DESCALE_P2-1)
1.87 +PW_DESCALE_P2X times 8 dw 1 << (PASS1_BITS-1)
1.88 +
1.89 + alignz 16
1.90 +
1.91 +; --------------------------------------------------------------------------
1.92 + SECTION SEG_TEXT
1.93 + BITS 64
1.94 +;
1.95 +; Perform the forward DCT on one block of samples.
1.96 +;
1.97 +; GLOBAL(void)
1.98 +; jsimd_fdct_islow_sse2 (DCTELEM * data)
1.99 +;
1.100 +
1.101 +; r10 = DCTELEM * data
1.102 +
1.103 +%define wk(i) rbp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM]
1.104 +%define WK_NUM 6
1.105 +
1.106 + align 16
1.107 + global EXTN(jsimd_fdct_islow_sse2)
1.108 +
1.109 +EXTN(jsimd_fdct_islow_sse2):
1.110 + push rbp
1.111 + mov rax,rsp ; rax = original rbp
1.112 + sub rsp, byte 4
1.113 + and rsp, byte (-SIZEOF_XMMWORD) ; align to 128 bits
1.114 + mov [rsp],rax
1.115 + mov rbp,rsp ; rbp = aligned rbp
1.116 + lea rsp, [wk(0)]
1.117 + collect_args
1.118 +
1.119 + ; ---- Pass 1: process rows.
1.120 +
1.121 + mov rdx, r10 ; (DCTELEM *)
1.122 +
1.123 + movdqa xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_DCTELEM)]
1.124 + movdqa xmm1, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_DCTELEM)]
1.125 + movdqa xmm2, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_DCTELEM)]
1.126 + movdqa xmm3, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_DCTELEM)]
1.127 +
1.128 + ; xmm0=(00 01 02 03 04 05 06 07), xmm2=(20 21 22 23 24 25 26 27)
1.129 + ; xmm1=(10 11 12 13 14 15 16 17), xmm3=(30 31 32 33 34 35 36 37)
1.130 +
1.131 + movdqa xmm4,xmm0 ; transpose coefficients(phase 1)
1.132 + punpcklwd xmm0,xmm1 ; xmm0=(00 10 01 11 02 12 03 13)
1.133 + punpckhwd xmm4,xmm1 ; xmm4=(04 14 05 15 06 16 07 17)
1.134 + movdqa xmm5,xmm2 ; transpose coefficients(phase 1)
1.135 + punpcklwd xmm2,xmm3 ; xmm2=(20 30 21 31 22 32 23 33)
1.136 + punpckhwd xmm5,xmm3 ; xmm5=(24 34 25 35 26 36 27 37)
1.137 +
1.138 + movdqa xmm6, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_DCTELEM)]
1.139 + movdqa xmm7, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_DCTELEM)]
1.140 + movdqa xmm1, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_DCTELEM)]
1.141 + movdqa xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_DCTELEM)]
1.142 +
1.143 + ; xmm6=( 4 12 20 28 36 44 52 60), xmm1=( 6 14 22 30 38 46 54 62)
1.144 + ; xmm7=( 5 13 21 29 37 45 53 61), xmm3=( 7 15 23 31 39 47 55 63)
1.145 +
1.146 + movdqa XMMWORD [wk(0)], xmm2 ; wk(0)=(20 30 21 31 22 32 23 33)
1.147 + movdqa XMMWORD [wk(1)], xmm5 ; wk(1)=(24 34 25 35 26 36 27 37)
1.148 +
1.149 + movdqa xmm2,xmm6 ; transpose coefficients(phase 1)
1.150 + punpcklwd xmm6,xmm7 ; xmm6=(40 50 41 51 42 52 43 53)
1.151 + punpckhwd xmm2,xmm7 ; xmm2=(44 54 45 55 46 56 47 57)
1.152 + movdqa xmm5,xmm1 ; transpose coefficients(phase 1)
1.153 + punpcklwd xmm1,xmm3 ; xmm1=(60 70 61 71 62 72 63 73)
1.154 + punpckhwd xmm5,xmm3 ; xmm5=(64 74 65 75 66 76 67 77)
1.155 +
1.156 + movdqa xmm7,xmm6 ; transpose coefficients(phase 2)
1.157 + punpckldq xmm6,xmm1 ; xmm6=(40 50 60 70 41 51 61 71)
1.158 + punpckhdq xmm7,xmm1 ; xmm7=(42 52 62 72 43 53 63 73)
1.159 + movdqa xmm3,xmm2 ; transpose coefficients(phase 2)
1.160 + punpckldq xmm2,xmm5 ; xmm2=(44 54 64 74 45 55 65 75)
1.161 + punpckhdq xmm3,xmm5 ; xmm3=(46 56 66 76 47 57 67 77)
1.162 +
1.163 + movdqa xmm1, XMMWORD [wk(0)] ; xmm1=(20 30 21 31 22 32 23 33)
1.164 + movdqa xmm5, XMMWORD [wk(1)] ; xmm5=(24 34 25 35 26 36 27 37)
1.165 + movdqa XMMWORD [wk(2)], xmm7 ; wk(2)=(42 52 62 72 43 53 63 73)
1.166 + movdqa XMMWORD [wk(3)], xmm2 ; wk(3)=(44 54 64 74 45 55 65 75)
1.167 +
1.168 + movdqa xmm7,xmm0 ; transpose coefficients(phase 2)
1.169 + punpckldq xmm0,xmm1 ; xmm0=(00 10 20 30 01 11 21 31)
1.170 + punpckhdq xmm7,xmm1 ; xmm7=(02 12 22 32 03 13 23 33)
1.171 + movdqa xmm2,xmm4 ; transpose coefficients(phase 2)
1.172 + punpckldq xmm4,xmm5 ; xmm4=(04 14 24 34 05 15 25 35)
1.173 + punpckhdq xmm2,xmm5 ; xmm2=(06 16 26 36 07 17 27 37)
1.174 +
1.175 + movdqa xmm1,xmm0 ; transpose coefficients(phase 3)
1.176 + punpcklqdq xmm0,xmm6 ; xmm0=(00 10 20 30 40 50 60 70)=data0
1.177 + punpckhqdq xmm1,xmm6 ; xmm1=(01 11 21 31 41 51 61 71)=data1
1.178 + movdqa xmm5,xmm2 ; transpose coefficients(phase 3)
1.179 + punpcklqdq xmm2,xmm3 ; xmm2=(06 16 26 36 46 56 66 76)=data6
1.180 + punpckhqdq xmm5,xmm3 ; xmm5=(07 17 27 37 47 57 67 77)=data7
1.181 +
1.182 + movdqa xmm6,xmm1
1.183 + movdqa xmm3,xmm0
1.184 + psubw xmm1,xmm2 ; xmm1=data1-data6=tmp6
1.185 + psubw xmm0,xmm5 ; xmm0=data0-data7=tmp7
1.186 + paddw xmm6,xmm2 ; xmm6=data1+data6=tmp1
1.187 + paddw xmm3,xmm5 ; xmm3=data0+data7=tmp0
1.188 +
1.189 + movdqa xmm2, XMMWORD [wk(2)] ; xmm2=(42 52 62 72 43 53 63 73)
1.190 + movdqa xmm5, XMMWORD [wk(3)] ; xmm5=(44 54 64 74 45 55 65 75)
1.191 + movdqa XMMWORD [wk(0)], xmm1 ; wk(0)=tmp6
1.192 + movdqa XMMWORD [wk(1)], xmm0 ; wk(1)=tmp7
1.193 +
1.194 + movdqa xmm1,xmm7 ; transpose coefficients(phase 3)
1.195 + punpcklqdq xmm7,xmm2 ; xmm7=(02 12 22 32 42 52 62 72)=data2
1.196 + punpckhqdq xmm1,xmm2 ; xmm1=(03 13 23 33 43 53 63 73)=data3
1.197 + movdqa xmm0,xmm4 ; transpose coefficients(phase 3)
1.198 + punpcklqdq xmm4,xmm5 ; xmm4=(04 14 24 34 44 54 64 74)=data4
1.199 + punpckhqdq xmm0,xmm5 ; xmm0=(05 15 25 35 45 55 65 75)=data5
1.200 +
1.201 + movdqa xmm2,xmm1
1.202 + movdqa xmm5,xmm7
1.203 + paddw xmm1,xmm4 ; xmm1=data3+data4=tmp3
1.204 + paddw xmm7,xmm0 ; xmm7=data2+data5=tmp2
1.205 + psubw xmm2,xmm4 ; xmm2=data3-data4=tmp4
1.206 + psubw xmm5,xmm0 ; xmm5=data2-data5=tmp5
1.207 +
1.208 + ; -- Even part
1.209 +
1.210 + movdqa xmm4,xmm3
1.211 + movdqa xmm0,xmm6
1.212 + paddw xmm3,xmm1 ; xmm3=tmp10
1.213 + paddw xmm6,xmm7 ; xmm6=tmp11
1.214 + psubw xmm4,xmm1 ; xmm4=tmp13
1.215 + psubw xmm0,xmm7 ; xmm0=tmp12
1.216 +
1.217 + movdqa xmm1,xmm3
1.218 + paddw xmm3,xmm6 ; xmm3=tmp10+tmp11
1.219 + psubw xmm1,xmm6 ; xmm1=tmp10-tmp11
1.220 +
1.221 + psllw xmm3,PASS1_BITS ; xmm3=data0
1.222 + psllw xmm1,PASS1_BITS ; xmm1=data4
1.223 +
1.224 + movdqa XMMWORD [wk(2)], xmm3 ; wk(2)=data0
1.225 + movdqa XMMWORD [wk(3)], xmm1 ; wk(3)=data4
1.226 +
1.227 + ; (Original)
1.228 + ; z1 = (tmp12 + tmp13) * 0.541196100;
1.229 + ; data2 = z1 + tmp13 * 0.765366865;
1.230 + ; data6 = z1 + tmp12 * -1.847759065;
1.231 + ;
1.232 + ; (This implementation)
1.233 + ; data2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100;
1.234 + ; data6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065);
1.235 +
1.236 + movdqa xmm7,xmm4 ; xmm4=tmp13
1.237 + movdqa xmm6,xmm4
1.238 + punpcklwd xmm7,xmm0 ; xmm0=tmp12
1.239 + punpckhwd xmm6,xmm0
1.240 + movdqa xmm4,xmm7
1.241 + movdqa xmm0,xmm6
1.242 + pmaddwd xmm7,[rel PW_F130_F054] ; xmm7=data2L
1.243 + pmaddwd xmm6,[rel PW_F130_F054] ; xmm6=data2H
1.244 + pmaddwd xmm4,[rel PW_F054_MF130] ; xmm4=data6L
1.245 + pmaddwd xmm0,[rel PW_F054_MF130] ; xmm0=data6H
1.246 +
1.247 + paddd xmm7,[rel PD_DESCALE_P1]
1.248 + paddd xmm6,[rel PD_DESCALE_P1]
1.249 + psrad xmm7,DESCALE_P1
1.250 + psrad xmm6,DESCALE_P1
1.251 + paddd xmm4,[rel PD_DESCALE_P1]
1.252 + paddd xmm0,[rel PD_DESCALE_P1]
1.253 + psrad xmm4,DESCALE_P1
1.254 + psrad xmm0,DESCALE_P1
1.255 +
1.256 + packssdw xmm7,xmm6 ; xmm7=data2
1.257 + packssdw xmm4,xmm0 ; xmm4=data6
1.258 +
1.259 + movdqa XMMWORD [wk(4)], xmm7 ; wk(4)=data2
1.260 + movdqa XMMWORD [wk(5)], xmm4 ; wk(5)=data6
1.261 +
1.262 + ; -- Odd part
1.263 +
1.264 + movdqa xmm3, XMMWORD [wk(0)] ; xmm3=tmp6
1.265 + movdqa xmm1, XMMWORD [wk(1)] ; xmm1=tmp7
1.266 +
1.267 + movdqa xmm6,xmm2 ; xmm2=tmp4
1.268 + movdqa xmm0,xmm5 ; xmm5=tmp5
1.269 + paddw xmm6,xmm3 ; xmm6=z3
1.270 + paddw xmm0,xmm1 ; xmm0=z4
1.271 +
1.272 + ; (Original)
1.273 + ; z5 = (z3 + z4) * 1.175875602;
1.274 + ; z3 = z3 * -1.961570560; z4 = z4 * -0.390180644;
1.275 + ; z3 += z5; z4 += z5;
1.276 + ;
1.277 + ; (This implementation)
1.278 + ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
1.279 + ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
1.280 +
1.281 + movdqa xmm7,xmm6
1.282 + movdqa xmm4,xmm6
1.283 + punpcklwd xmm7,xmm0
1.284 + punpckhwd xmm4,xmm0
1.285 + movdqa xmm6,xmm7
1.286 + movdqa xmm0,xmm4
1.287 + pmaddwd xmm7,[rel PW_MF078_F117] ; xmm7=z3L
1.288 + pmaddwd xmm4,[rel PW_MF078_F117] ; xmm4=z3H
1.289 + pmaddwd xmm6,[rel PW_F117_F078] ; xmm6=z4L
1.290 + pmaddwd xmm0,[rel PW_F117_F078] ; xmm0=z4H
1.291 +
1.292 + movdqa XMMWORD [wk(0)], xmm7 ; wk(0)=z3L
1.293 + movdqa XMMWORD [wk(1)], xmm4 ; wk(1)=z3H
1.294 +
1.295 + ; (Original)
1.296 + ; z1 = tmp4 + tmp7; z2 = tmp5 + tmp6;
1.297 + ; tmp4 = tmp4 * 0.298631336; tmp5 = tmp5 * 2.053119869;
1.298 + ; tmp6 = tmp6 * 3.072711026; tmp7 = tmp7 * 1.501321110;
1.299 + ; z1 = z1 * -0.899976223; z2 = z2 * -2.562915447;
1.300 + ; data7 = tmp4 + z1 + z3; data5 = tmp5 + z2 + z4;
1.301 + ; data3 = tmp6 + z2 + z3; data1 = tmp7 + z1 + z4;
1.302 + ;
1.303 + ; (This implementation)
1.304 + ; tmp4 = tmp4 * (0.298631336 - 0.899976223) + tmp7 * -0.899976223;
1.305 + ; tmp5 = tmp5 * (2.053119869 - 2.562915447) + tmp6 * -2.562915447;
1.306 + ; tmp6 = tmp5 * -2.562915447 + tmp6 * (3.072711026 - 2.562915447);
1.307 + ; tmp7 = tmp4 * -0.899976223 + tmp7 * (1.501321110 - 0.899976223);
1.308 + ; data7 = tmp4 + z3; data5 = tmp5 + z4;
1.309 + ; data3 = tmp6 + z3; data1 = tmp7 + z4;
1.310 +
1.311 + movdqa xmm7,xmm2
1.312 + movdqa xmm4,xmm2
1.313 + punpcklwd xmm7,xmm1
1.314 + punpckhwd xmm4,xmm1
1.315 + movdqa xmm2,xmm7
1.316 + movdqa xmm1,xmm4
1.317 + pmaddwd xmm7,[rel PW_MF060_MF089] ; xmm7=tmp4L
1.318 + pmaddwd xmm4,[rel PW_MF060_MF089] ; xmm4=tmp4H
1.319 + pmaddwd xmm2,[rel PW_MF089_F060] ; xmm2=tmp7L
1.320 + pmaddwd xmm1,[rel PW_MF089_F060] ; xmm1=tmp7H
1.321 +
1.322 + paddd xmm7, XMMWORD [wk(0)] ; xmm7=data7L
1.323 + paddd xmm4, XMMWORD [wk(1)] ; xmm4=data7H
1.324 + paddd xmm2,xmm6 ; xmm2=data1L
1.325 + paddd xmm1,xmm0 ; xmm1=data1H
1.326 +
1.327 + paddd xmm7,[rel PD_DESCALE_P1]
1.328 + paddd xmm4,[rel PD_DESCALE_P1]
1.329 + psrad xmm7,DESCALE_P1
1.330 + psrad xmm4,DESCALE_P1
1.331 + paddd xmm2,[rel PD_DESCALE_P1]
1.332 + paddd xmm1,[rel PD_DESCALE_P1]
1.333 + psrad xmm2,DESCALE_P1
1.334 + psrad xmm1,DESCALE_P1
1.335 +
1.336 + packssdw xmm7,xmm4 ; xmm7=data7
1.337 + packssdw xmm2,xmm1 ; xmm2=data1
1.338 +
1.339 + movdqa xmm4,xmm5
1.340 + movdqa xmm1,xmm5
1.341 + punpcklwd xmm4,xmm3
1.342 + punpckhwd xmm1,xmm3
1.343 + movdqa xmm5,xmm4
1.344 + movdqa xmm3,xmm1
1.345 + pmaddwd xmm4,[rel PW_MF050_MF256] ; xmm4=tmp5L
1.346 + pmaddwd xmm1,[rel PW_MF050_MF256] ; xmm1=tmp5H
1.347 + pmaddwd xmm5,[rel PW_MF256_F050] ; xmm5=tmp6L
1.348 + pmaddwd xmm3,[rel PW_MF256_F050] ; xmm3=tmp6H
1.349 +
1.350 + paddd xmm4,xmm6 ; xmm4=data5L
1.351 + paddd xmm1,xmm0 ; xmm1=data5H
1.352 + paddd xmm5, XMMWORD [wk(0)] ; xmm5=data3L
1.353 + paddd xmm3, XMMWORD [wk(1)] ; xmm3=data3H
1.354 +
1.355 + paddd xmm4,[rel PD_DESCALE_P1]
1.356 + paddd xmm1,[rel PD_DESCALE_P1]
1.357 + psrad xmm4,DESCALE_P1
1.358 + psrad xmm1,DESCALE_P1
1.359 + paddd xmm5,[rel PD_DESCALE_P1]
1.360 + paddd xmm3,[rel PD_DESCALE_P1]
1.361 + psrad xmm5,DESCALE_P1
1.362 + psrad xmm3,DESCALE_P1
1.363 +
1.364 + packssdw xmm4,xmm1 ; xmm4=data5
1.365 + packssdw xmm5,xmm3 ; xmm5=data3
1.366 +
1.367 + ; ---- Pass 2: process columns.
1.368 +
1.369 + movdqa xmm6, XMMWORD [wk(2)] ; xmm6=col0
1.370 + movdqa xmm0, XMMWORD [wk(4)] ; xmm0=col2
1.371 +
1.372 + ; xmm6=(00 10 20 30 40 50 60 70), xmm0=(02 12 22 32 42 52 62 72)
1.373 + ; xmm2=(01 11 21 31 41 51 61 71), xmm5=(03 13 23 33 43 53 63 73)
1.374 +
1.375 + movdqa xmm1,xmm6 ; transpose coefficients(phase 1)
1.376 + punpcklwd xmm6,xmm2 ; xmm6=(00 01 10 11 20 21 30 31)
1.377 + punpckhwd xmm1,xmm2 ; xmm1=(40 41 50 51 60 61 70 71)
1.378 + movdqa xmm3,xmm0 ; transpose coefficients(phase 1)
1.379 + punpcklwd xmm0,xmm5 ; xmm0=(02 03 12 13 22 23 32 33)
1.380 + punpckhwd xmm3,xmm5 ; xmm3=(42 43 52 53 62 63 72 73)
1.381 +
1.382 + movdqa xmm2, XMMWORD [wk(3)] ; xmm2=col4
1.383 + movdqa xmm5, XMMWORD [wk(5)] ; xmm5=col6
1.384 +
1.385 + ; xmm2=(04 14 24 34 44 54 64 74), xmm5=(06 16 26 36 46 56 66 76)
1.386 + ; xmm4=(05 15 25 35 45 55 65 75), xmm7=(07 17 27 37 47 57 67 77)
1.387 +
1.388 + movdqa XMMWORD [wk(0)], xmm0 ; wk(0)=(02 03 12 13 22 23 32 33)
1.389 + movdqa XMMWORD [wk(1)], xmm3 ; wk(1)=(42 43 52 53 62 63 72 73)
1.390 +
1.391 + movdqa xmm0,xmm2 ; transpose coefficients(phase 1)
1.392 + punpcklwd xmm2,xmm4 ; xmm2=(04 05 14 15 24 25 34 35)
1.393 + punpckhwd xmm0,xmm4 ; xmm0=(44 45 54 55 64 65 74 75)
1.394 + movdqa xmm3,xmm5 ; transpose coefficients(phase 1)
1.395 + punpcklwd xmm5,xmm7 ; xmm5=(06 07 16 17 26 27 36 37)
1.396 + punpckhwd xmm3,xmm7 ; xmm3=(46 47 56 57 66 67 76 77)
1.397 +
1.398 + movdqa xmm4,xmm2 ; transpose coefficients(phase 2)
1.399 + punpckldq xmm2,xmm5 ; xmm2=(04 05 06 07 14 15 16 17)
1.400 + punpckhdq xmm4,xmm5 ; xmm4=(24 25 26 27 34 35 36 37)
1.401 + movdqa xmm7,xmm0 ; transpose coefficients(phase 2)
1.402 + punpckldq xmm0,xmm3 ; xmm0=(44 45 46 47 54 55 56 57)
1.403 + punpckhdq xmm7,xmm3 ; xmm7=(64 65 66 67 74 75 76 77)
1.404 +
1.405 + movdqa xmm5, XMMWORD [wk(0)] ; xmm5=(02 03 12 13 22 23 32 33)
1.406 + movdqa xmm3, XMMWORD [wk(1)] ; xmm3=(42 43 52 53 62 63 72 73)
1.407 + movdqa XMMWORD [wk(2)], xmm4 ; wk(2)=(24 25 26 27 34 35 36 37)
1.408 + movdqa XMMWORD [wk(3)], xmm0 ; wk(3)=(44 45 46 47 54 55 56 57)
1.409 +
1.410 + movdqa xmm4,xmm6 ; transpose coefficients(phase 2)
1.411 + punpckldq xmm6,xmm5 ; xmm6=(00 01 02 03 10 11 12 13)
1.412 + punpckhdq xmm4,xmm5 ; xmm4=(20 21 22 23 30 31 32 33)
1.413 + movdqa xmm0,xmm1 ; transpose coefficients(phase 2)
1.414 + punpckldq xmm1,xmm3 ; xmm1=(40 41 42 43 50 51 52 53)
1.415 + punpckhdq xmm0,xmm3 ; xmm0=(60 61 62 63 70 71 72 73)
1.416 +
1.417 + movdqa xmm5,xmm6 ; transpose coefficients(phase 3)
1.418 + punpcklqdq xmm6,xmm2 ; xmm6=(00 01 02 03 04 05 06 07)=data0
1.419 + punpckhqdq xmm5,xmm2 ; xmm5=(10 11 12 13 14 15 16 17)=data1
1.420 + movdqa xmm3,xmm0 ; transpose coefficients(phase 3)
1.421 + punpcklqdq xmm0,xmm7 ; xmm0=(60 61 62 63 64 65 66 67)=data6
1.422 + punpckhqdq xmm3,xmm7 ; xmm3=(70 71 72 73 74 75 76 77)=data7
1.423 +
1.424 + movdqa xmm2,xmm5
1.425 + movdqa xmm7,xmm6
1.426 + psubw xmm5,xmm0 ; xmm5=data1-data6=tmp6
1.427 + psubw xmm6,xmm3 ; xmm6=data0-data7=tmp7
1.428 + paddw xmm2,xmm0 ; xmm2=data1+data6=tmp1
1.429 + paddw xmm7,xmm3 ; xmm7=data0+data7=tmp0
1.430 +
1.431 + movdqa xmm0, XMMWORD [wk(2)] ; xmm0=(24 25 26 27 34 35 36 37)
1.432 + movdqa xmm3, XMMWORD [wk(3)] ; xmm3=(44 45 46 47 54 55 56 57)
1.433 + movdqa XMMWORD [wk(0)], xmm5 ; wk(0)=tmp6
1.434 + movdqa XMMWORD [wk(1)], xmm6 ; wk(1)=tmp7
1.435 +
1.436 + movdqa xmm5,xmm4 ; transpose coefficients(phase 3)
1.437 + punpcklqdq xmm4,xmm0 ; xmm4=(20 21 22 23 24 25 26 27)=data2
1.438 + punpckhqdq xmm5,xmm0 ; xmm5=(30 31 32 33 34 35 36 37)=data3
1.439 + movdqa xmm6,xmm1 ; transpose coefficients(phase 3)
1.440 + punpcklqdq xmm1,xmm3 ; xmm1=(40 41 42 43 44 45 46 47)=data4
1.441 + punpckhqdq xmm6,xmm3 ; xmm6=(50 51 52 53 54 55 56 57)=data5
1.442 +
1.443 + movdqa xmm0,xmm5
1.444 + movdqa xmm3,xmm4
1.445 + paddw xmm5,xmm1 ; xmm5=data3+data4=tmp3
1.446 + paddw xmm4,xmm6 ; xmm4=data2+data5=tmp2
1.447 + psubw xmm0,xmm1 ; xmm0=data3-data4=tmp4
1.448 + psubw xmm3,xmm6 ; xmm3=data2-data5=tmp5
1.449 +
1.450 + ; -- Even part
1.451 +
1.452 + movdqa xmm1,xmm7
1.453 + movdqa xmm6,xmm2
1.454 + paddw xmm7,xmm5 ; xmm7=tmp10
1.455 + paddw xmm2,xmm4 ; xmm2=tmp11
1.456 + psubw xmm1,xmm5 ; xmm1=tmp13
1.457 + psubw xmm6,xmm4 ; xmm6=tmp12
1.458 +
1.459 + movdqa xmm5,xmm7
1.460 + paddw xmm7,xmm2 ; xmm7=tmp10+tmp11
1.461 + psubw xmm5,xmm2 ; xmm5=tmp10-tmp11
1.462 +
1.463 + paddw xmm7,[rel PW_DESCALE_P2X]
1.464 + paddw xmm5,[rel PW_DESCALE_P2X]
1.465 + psraw xmm7,PASS1_BITS ; xmm7=data0
1.466 + psraw xmm5,PASS1_BITS ; xmm5=data4
1.467 +
1.468 + movdqa XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_DCTELEM)], xmm7
1.469 + movdqa XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_DCTELEM)], xmm5
1.470 +
1.471 + ; (Original)
1.472 + ; z1 = (tmp12 + tmp13) * 0.541196100;
1.473 + ; data2 = z1 + tmp13 * 0.765366865;
1.474 + ; data6 = z1 + tmp12 * -1.847759065;
1.475 + ;
1.476 + ; (This implementation)
1.477 + ; data2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100;
1.478 + ; data6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065);
1.479 +
1.480 + movdqa xmm4,xmm1 ; xmm1=tmp13
1.481 + movdqa xmm2,xmm1
1.482 + punpcklwd xmm4,xmm6 ; xmm6=tmp12
1.483 + punpckhwd xmm2,xmm6
1.484 + movdqa xmm1,xmm4
1.485 + movdqa xmm6,xmm2
1.486 + pmaddwd xmm4,[rel PW_F130_F054] ; xmm4=data2L
1.487 + pmaddwd xmm2,[rel PW_F130_F054] ; xmm2=data2H
1.488 + pmaddwd xmm1,[rel PW_F054_MF130] ; xmm1=data6L
1.489 + pmaddwd xmm6,[rel PW_F054_MF130] ; xmm6=data6H
1.490 +
1.491 + paddd xmm4,[rel PD_DESCALE_P2]
1.492 + paddd xmm2,[rel PD_DESCALE_P2]
1.493 + psrad xmm4,DESCALE_P2
1.494 + psrad xmm2,DESCALE_P2
1.495 + paddd xmm1,[rel PD_DESCALE_P2]
1.496 + paddd xmm6,[rel PD_DESCALE_P2]
1.497 + psrad xmm1,DESCALE_P2
1.498 + psrad xmm6,DESCALE_P2
1.499 +
1.500 + packssdw xmm4,xmm2 ; xmm4=data2
1.501 + packssdw xmm1,xmm6 ; xmm1=data6
1.502 +
1.503 + movdqa XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_DCTELEM)], xmm4
1.504 + movdqa XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_DCTELEM)], xmm1
1.505 +
1.506 + ; -- Odd part
1.507 +
1.508 + movdqa xmm7, XMMWORD [wk(0)] ; xmm7=tmp6
1.509 + movdqa xmm5, XMMWORD [wk(1)] ; xmm5=tmp7
1.510 +
1.511 + movdqa xmm2,xmm0 ; xmm0=tmp4
1.512 + movdqa xmm6,xmm3 ; xmm3=tmp5
1.513 + paddw xmm2,xmm7 ; xmm2=z3
1.514 + paddw xmm6,xmm5 ; xmm6=z4
1.515 +
1.516 + ; (Original)
1.517 + ; z5 = (z3 + z4) * 1.175875602;
1.518 + ; z3 = z3 * -1.961570560; z4 = z4 * -0.390180644;
1.519 + ; z3 += z5; z4 += z5;
1.520 + ;
1.521 + ; (This implementation)
1.522 + ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
1.523 + ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
1.524 +
1.525 + movdqa xmm4,xmm2
1.526 + movdqa xmm1,xmm2
1.527 + punpcklwd xmm4,xmm6
1.528 + punpckhwd xmm1,xmm6
1.529 + movdqa xmm2,xmm4
1.530 + movdqa xmm6,xmm1
1.531 + pmaddwd xmm4,[rel PW_MF078_F117] ; xmm4=z3L
1.532 + pmaddwd xmm1,[rel PW_MF078_F117] ; xmm1=z3H
1.533 + pmaddwd xmm2,[rel PW_F117_F078] ; xmm2=z4L
1.534 + pmaddwd xmm6,[rel PW_F117_F078] ; xmm6=z4H
1.535 +
1.536 + movdqa XMMWORD [wk(0)], xmm4 ; wk(0)=z3L
1.537 + movdqa XMMWORD [wk(1)], xmm1 ; wk(1)=z3H
1.538 +
1.539 + ; (Original)
1.540 + ; z1 = tmp4 + tmp7; z2 = tmp5 + tmp6;
1.541 + ; tmp4 = tmp4 * 0.298631336; tmp5 = tmp5 * 2.053119869;
1.542 + ; tmp6 = tmp6 * 3.072711026; tmp7 = tmp7 * 1.501321110;
1.543 + ; z1 = z1 * -0.899976223; z2 = z2 * -2.562915447;
1.544 + ; data7 = tmp4 + z1 + z3; data5 = tmp5 + z2 + z4;
1.545 + ; data3 = tmp6 + z2 + z3; data1 = tmp7 + z1 + z4;
1.546 + ;
1.547 + ; (This implementation)
1.548 + ; tmp4 = tmp4 * (0.298631336 - 0.899976223) + tmp7 * -0.899976223;
1.549 + ; tmp5 = tmp5 * (2.053119869 - 2.562915447) + tmp6 * -2.562915447;
1.550 + ; tmp6 = tmp5 * -2.562915447 + tmp6 * (3.072711026 - 2.562915447);
1.551 + ; tmp7 = tmp4 * -0.899976223 + tmp7 * (1.501321110 - 0.899976223);
1.552 + ; data7 = tmp4 + z3; data5 = tmp5 + z4;
1.553 + ; data3 = tmp6 + z3; data1 = tmp7 + z4;
1.554 +
1.555 + movdqa xmm4,xmm0
1.556 + movdqa xmm1,xmm0
1.557 + punpcklwd xmm4,xmm5
1.558 + punpckhwd xmm1,xmm5
1.559 + movdqa xmm0,xmm4
1.560 + movdqa xmm5,xmm1
1.561 + pmaddwd xmm4,[rel PW_MF060_MF089] ; xmm4=tmp4L
1.562 + pmaddwd xmm1,[rel PW_MF060_MF089] ; xmm1=tmp4H
1.563 + pmaddwd xmm0,[rel PW_MF089_F060] ; xmm0=tmp7L
1.564 + pmaddwd xmm5,[rel PW_MF089_F060] ; xmm5=tmp7H
1.565 +
1.566 + paddd xmm4, XMMWORD [wk(0)] ; xmm4=data7L
1.567 + paddd xmm1, XMMWORD [wk(1)] ; xmm1=data7H
1.568 + paddd xmm0,xmm2 ; xmm0=data1L
1.569 + paddd xmm5,xmm6 ; xmm5=data1H
1.570 +
1.571 + paddd xmm4,[rel PD_DESCALE_P2]
1.572 + paddd xmm1,[rel PD_DESCALE_P2]
1.573 + psrad xmm4,DESCALE_P2
1.574 + psrad xmm1,DESCALE_P2
1.575 + paddd xmm0,[rel PD_DESCALE_P2]
1.576 + paddd xmm5,[rel PD_DESCALE_P2]
1.577 + psrad xmm0,DESCALE_P2
1.578 + psrad xmm5,DESCALE_P2
1.579 +
1.580 + packssdw xmm4,xmm1 ; xmm4=data7
1.581 + packssdw xmm0,xmm5 ; xmm0=data1
1.582 +
1.583 + movdqa XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_DCTELEM)], xmm4
1.584 + movdqa XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_DCTELEM)], xmm0
1.585 +
1.586 + movdqa xmm1,xmm3
1.587 + movdqa xmm5,xmm3
1.588 + punpcklwd xmm1,xmm7
1.589 + punpckhwd xmm5,xmm7
1.590 + movdqa xmm3,xmm1
1.591 + movdqa xmm7,xmm5
1.592 + pmaddwd xmm1,[rel PW_MF050_MF256] ; xmm1=tmp5L
1.593 + pmaddwd xmm5,[rel PW_MF050_MF256] ; xmm5=tmp5H
1.594 + pmaddwd xmm3,[rel PW_MF256_F050] ; xmm3=tmp6L
1.595 + pmaddwd xmm7,[rel PW_MF256_F050] ; xmm7=tmp6H
1.596 +
1.597 + paddd xmm1,xmm2 ; xmm1=data5L
1.598 + paddd xmm5,xmm6 ; xmm5=data5H
1.599 + paddd xmm3, XMMWORD [wk(0)] ; xmm3=data3L
1.600 + paddd xmm7, XMMWORD [wk(1)] ; xmm7=data3H
1.601 +
1.602 + paddd xmm1,[rel PD_DESCALE_P2]
1.603 + paddd xmm5,[rel PD_DESCALE_P2]
1.604 + psrad xmm1,DESCALE_P2
1.605 + psrad xmm5,DESCALE_P2
1.606 + paddd xmm3,[rel PD_DESCALE_P2]
1.607 + paddd xmm7,[rel PD_DESCALE_P2]
1.608 + psrad xmm3,DESCALE_P2
1.609 + psrad xmm7,DESCALE_P2
1.610 +
1.611 + packssdw xmm1,xmm5 ; xmm1=data5
1.612 + packssdw xmm3,xmm7 ; xmm3=data3
1.613 +
1.614 + movdqa XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_DCTELEM)], xmm1
1.615 + movdqa XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_DCTELEM)], xmm3
1.616 +
1.617 + uncollect_args
1.618 + mov rsp,rbp ; rsp <- aligned rbp
1.619 + pop rsp ; rsp <- original rbp
1.620 + pop rbp
1.621 + ret
1.622 +
1.623 +; For some reason, the OS X linker does not honor the request to align the
1.624 +; segment unless we do this.
1.625 + align 16
