d@0
|
1 /*
|
d@0
|
2 * Copyright (c) 2003, 2007-8 Matteo Frigo
|
d@0
|
3 * Copyright (c) 2003, 2007-8 Massachusetts Institute of Technology
|
d@0
|
4 *
|
d@0
|
5 * This program is free software; you can redistribute it and/or modify
|
d@0
|
6 * it under the terms of the GNU General Public License as published by
|
d@0
|
7 * the Free Software Foundation; either version 2 of the License, or
|
d@0
|
8 * (at your option) any later version.
|
d@0
|
9 *
|
d@0
|
10 * This program is distributed in the hope that it will be useful,
|
d@0
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
d@0
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
d@0
|
13 * GNU General Public License for more details.
|
d@0
|
14 *
|
d@0
|
15 * You should have received a copy of the GNU General Public License
|
d@0
|
16 * along with this program; if not, write to the Free Software
|
d@0
|
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
|
d@0
|
18 *
|
d@0
|
19 */
|
d@0
|
20 /* Generated by: ../../genfft/gen_notw_c -standalone -fma -reorder-insns -simd -compact -variables 100000 -with-ostride 2 -include fftw-spu.h -store-multiple 2 -n 32 -name X(spu_n2fv_32) */
|
d@0
|
21
|
d@0
|
22 /*
|
d@0
|
23 * This function contains 186 FP additions, 98 FP multiplications,
|
d@0
|
24 * (or, 88 additions, 0 multiplications, 98 fused multiply/add),
|
d@0
|
25 * 228 stack variables, 7 constants, and 80 memory accesses
|
d@0
|
26 */
|
d@0
|
27 #include "fftw-spu.h"
|
d@0
|
28
|
d@0
|
29 void X(spu_n2fv_32) (const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs) {
|
d@0
|
30 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
d@0
|
31 DVK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
d@0
|
32 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
d@0
|
33 DVK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
d@0
|
34 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
d@0
|
35 DVK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
d@0
|
36 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
d@0
|
37 INT i;
|
d@0
|
38 const R *xi;
|
d@0
|
39 R *xo;
|
d@0
|
40 xi = ri;
|
d@0
|
41 xo = ro;
|
d@0
|
42 for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(os)) {
|
d@0
|
43 V T2x, T1T, T1W, T2K, T16, T1A, T1p, Tb, TT, T1v, T1w, TY, T2t, T2b, T2H;
|
d@0
|
44 V T2N, T1s, TC, TH, T1t, T24, T2s, T2O, T2E, T2g, T2j, T1B, Tq, T1q, T19;
|
d@0
|
45 V T2L, T2A, T1R, T3, T1S, T14, T1U, T6, T1V, T9, T1, T2, T12, T13, T4;
|
d@0
|
46 V T5, T7, T8, Ta, T15, T25, TL, T26, TW, T28, TO, T29, TR, TJ, TK;
|
d@0
|
47 V TU, TV, TM, TN, TP, TQ, TX, TS, T2F, T2G, T27, T2a, Tu, T1Y, TF;
|
d@0
|
48 V T1Z, Tx, T21, TA, T22, Ts, Tt, TD, TE, Tv, Tw, Ty, Tz, TG, TB;
|
d@0
|
49 V T2C, T2D, T20, T23, T2h, Te, T2f, To, T2i, Th, T2e, Tl, Tc, Td, Tm;
|
d@0
|
50 V Tn, Tf, Tg, Tj, Tk, T2z, T2y, T17, T18, Ti, Tp, T31, T32, T33, T34;
|
d@0
|
51 V T35, T36, T37, T38, T2B, T2P, T2M, T2I, T39, T3a, T3b, T3c, T2p, T2q, T2n;
|
d@0
|
52 V T2d, T2o, T2m, T2l, T1X, T2c, T2k, T2v, T2r, T2w, T2u, T2Z, T2V, T30, T2Y;
|
d@0
|
53 V T2T, T2U, T2W, T2X, T3d, T3e, T3f, T3g, T2R, T2J, T2S, T2Q, T1J, T1r, T1C;
|
d@0
|
54 V T1M, T1K, T1F, T1N, T1y, T1u, T1x, T1D, T1E, T1H, T1I, T1L, T1O, T1z, T1G;
|
d@0
|
55 V T3h, T3i, T1P, T1Q, T3j, T3k, T3l, T3m, T3n, T3o, T1h, Tr, T1a, T1k, T1i;
|
d@0
|
56 V T1d, T1l, T10, TI, TZ, T1b, T1c, T1f, T1g, T1j, T1m, T11, T1e, T3p, T3q;
|
d@0
|
57 V T1n, T1o, T3r, T3s, T3t, T3u, T3v, T3w;
|
d@0
|
58 T1 = LD(&(xi[0]), ivs, &(xi[0]));
|
d@0
|
59 T2 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
|
d@0
|
60 T1R = VADD(T1, T2);
|
d@0
|
61 T3 = VSUB(T1, T2);
|
d@0
|
62 T12 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
|
d@0
|
63 T13 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
|
d@0
|
64 T1S = VADD(T12, T13);
|
d@0
|
65 T14 = VSUB(T12, T13);
|
d@0
|
66 T4 = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
|
d@0
|
67 T5 = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
|
d@0
|
68 T1U = VADD(T4, T5);
|
d@0
|
69 T6 = VSUB(T4, T5);
|
d@0
|
70 T7 = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
|
d@0
|
71 T8 = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
|
d@0
|
72 T1V = VADD(T7, T8);
|
d@0
|
73 T9 = VSUB(T7, T8);
|
d@0
|
74 T2x = VSUB(T1R, T1S);
|
d@0
|
75 T1T = VADD(T1R, T1S);
|
d@0
|
76 T1W = VADD(T1U, T1V);
|
d@0
|
77 T2K = VSUB(T1V, T1U);
|
d@0
|
78 Ta = VADD(T6, T9);
|
d@0
|
79 T15 = VSUB(T9, T6);
|
d@0
|
80 T16 = VFNMS(LDK(KP707106781), T15, T14);
|
d@0
|
81 T1A = VFMA(LDK(KP707106781), T15, T14);
|
d@0
|
82 T1p = VFNMS(LDK(KP707106781), Ta, T3);
|
d@0
|
83 Tb = VFMA(LDK(KP707106781), Ta, T3);
|
d@0
|
84 TJ = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
|
d@0
|
85 TK = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
|
d@0
|
86 T25 = VADD(TJ, TK);
|
d@0
|
87 TL = VSUB(TJ, TK);
|
d@0
|
88 TU = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
|
d@0
|
89 TV = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
|
d@0
|
90 T26 = VADD(TV, TU);
|
d@0
|
91 TW = VSUB(TU, TV);
|
d@0
|
92 TM = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
|
d@0
|
93 TN = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
|
d@0
|
94 T28 = VADD(TM, TN);
|
d@0
|
95 TO = VSUB(TM, TN);
|
d@0
|
96 TP = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
|
d@0
|
97 TQ = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
|
d@0
|
98 T29 = VADD(TP, TQ);
|
d@0
|
99 TR = VSUB(TP, TQ);
|
d@0
|
100 TX = VSUB(TR, TO);
|
d@0
|
101 TS = VADD(TO, TR);
|
d@0
|
102 TT = VFMA(LDK(KP707106781), TS, TL);
|
d@0
|
103 T1v = VFNMS(LDK(KP707106781), TS, TL);
|
d@0
|
104 T1w = VFNMS(LDK(KP707106781), TX, TW);
|
d@0
|
105 TY = VFMA(LDK(KP707106781), TX, TW);
|
d@0
|
106 T2F = VSUB(T25, T26);
|
d@0
|
107 T27 = VADD(T25, T26);
|
d@0
|
108 T2a = VADD(T28, T29);
|
d@0
|
109 T2G = VSUB(T29, T28);
|
d@0
|
110 T2t = VADD(T27, T2a);
|
d@0
|
111 T2b = VSUB(T27, T2a);
|
d@0
|
112 T2H = VFNMS(LDK(KP414213562), T2G, T2F);
|
d@0
|
113 T2N = VFMA(LDK(KP414213562), T2F, T2G);
|
d@0
|
114 Ts = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
|
d@0
|
115 Tt = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
|
d@0
|
116 Tu = VSUB(Ts, Tt);
|
d@0
|
117 T1Y = VADD(Ts, Tt);
|
d@0
|
118 TD = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
|
d@0
|
119 TE = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
|
d@0
|
120 TF = VSUB(TD, TE);
|
d@0
|
121 T1Z = VADD(TD, TE);
|
d@0
|
122 Tv = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
|
d@0
|
123 Tw = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
|
d@0
|
124 Tx = VSUB(Tv, Tw);
|
d@0
|
125 T21 = VADD(Tv, Tw);
|
d@0
|
126 Ty = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
|
d@0
|
127 Tz = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
|
d@0
|
128 TA = VSUB(Ty, Tz);
|
d@0
|
129 T22 = VADD(Ty, Tz);
|
d@0
|
130 TB = VADD(Tx, TA);
|
d@0
|
131 TG = VSUB(Tx, TA);
|
d@0
|
132 T1s = VFNMS(LDK(KP707106781), TB, Tu);
|
d@0
|
133 TC = VFMA(LDK(KP707106781), TB, Tu);
|
d@0
|
134 TH = VFMA(LDK(KP707106781), TG, TF);
|
d@0
|
135 T1t = VFNMS(LDK(KP707106781), TG, TF);
|
d@0
|
136 T20 = VADD(T1Y, T1Z);
|
d@0
|
137 T2C = VSUB(T1Y, T1Z);
|
d@0
|
138 T2D = VSUB(T21, T22);
|
d@0
|
139 T23 = VADD(T21, T22);
|
d@0
|
140 T24 = VSUB(T20, T23);
|
d@0
|
141 T2s = VADD(T20, T23);
|
d@0
|
142 T2O = VFMA(LDK(KP414213562), T2C, T2D);
|
d@0
|
143 T2E = VFNMS(LDK(KP414213562), T2D, T2C);
|
d@0
|
144 Tc = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
|
d@0
|
145 Td = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
|
d@0
|
146 T2h = VADD(Tc, Td);
|
d@0
|
147 Te = VSUB(Tc, Td);
|
d@0
|
148 Tm = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
|
d@0
|
149 Tn = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
|
d@0
|
150 T2f = VADD(Tn, Tm);
|
d@0
|
151 To = VSUB(Tm, Tn);
|
d@0
|
152 Tf = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
|
d@0
|
153 Tg = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
|
d@0
|
154 T2i = VADD(Tf, Tg);
|
d@0
|
155 Th = VSUB(Tf, Tg);
|
d@0
|
156 Tj = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
|
d@0
|
157 Tk = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
|
d@0
|
158 T2e = VADD(Tj, Tk);
|
d@0
|
159 Tl = VSUB(Tj, Tk);
|
d@0
|
160 T2z = VSUB(T2e, T2f);
|
d@0
|
161 T2g = VADD(T2e, T2f);
|
d@0
|
162 T2j = VADD(T2h, T2i);
|
d@0
|
163 T2y = VSUB(T2h, T2i);
|
d@0
|
164 T17 = VFMA(LDK(KP414213562), Te, Th);
|
d@0
|
165 Ti = VFNMS(LDK(KP414213562), Th, Te);
|
d@0
|
166 Tp = VFNMS(LDK(KP414213562), To, Tl);
|
d@0
|
167 T18 = VFMA(LDK(KP414213562), Tl, To);
|
d@0
|
168 T1B = VSUB(Tp, Ti);
|
d@0
|
169 Tq = VADD(Ti, Tp);
|
d@0
|
170 T1q = VADD(T17, T18);
|
d@0
|
171 T19 = VSUB(T17, T18);
|
d@0
|
172 T2L = VSUB(T2z, T2y);
|
d@0
|
173 T2A = VADD(T2y, T2z);
|
d@0
|
174 T2p = VADD(T1T, T1W);
|
d@0
|
175 T1X = VSUB(T1T, T1W);
|
d@0
|
176 T2c = VADD(T24, T2b);
|
d@0
|
177 T2l = VSUB(T2b, T24);
|
d@0
|
178 T2n = VFMA(LDK(KP707106781), T2c, T1X);
|
d@0
|
179 T2d = VFNMS(LDK(KP707106781), T2c, T1X);
|
d@0
|
180 T2q = VADD(T2j, T2g);
|
d@0
|
181 T2k = VSUB(T2g, T2j);
|
d@0
|
182 T2o = VFMA(LDK(KP707106781), T2l, T2k);
|
d@0
|
183 T2m = VFNMS(LDK(KP707106781), T2l, T2k);
|
d@0
|
184 T31 = VFNMSI(T2m, T2d);
|
d@0
|
185 STM2(&(xo[24]), T31, ovs, &(xo[0]));
|
d@0
|
186 T32 = VFMAI(T2o, T2n);
|
d@0
|
187 STM2(&(xo[8]), T32, ovs, &(xo[0]));
|
d@0
|
188 T33 = VFMAI(T2m, T2d);
|
d@0
|
189 STM2(&(xo[40]), T33, ovs, &(xo[0]));
|
d@0
|
190 T34 = VFNMSI(T2o, T2n);
|
d@0
|
191 STM2(&(xo[56]), T34, ovs, &(xo[0]));
|
d@0
|
192 T2v = VSUB(T2p, T2q);
|
d@0
|
193 T2r = VADD(T2p, T2q);
|
d@0
|
194 T2w = VSUB(T2t, T2s);
|
d@0
|
195 T2u = VADD(T2s, T2t);
|
d@0
|
196 T35 = VSUB(T2r, T2u);
|
d@0
|
197 STM2(&(xo[32]), T35, ovs, &(xo[0]));
|
d@0
|
198 T36 = VFMAI(T2w, T2v);
|
d@0
|
199 STM2(&(xo[16]), T36, ovs, &(xo[0]));
|
d@0
|
200 T37 = VADD(T2r, T2u);
|
d@0
|
201 STM2(&(xo[0]), T37, ovs, &(xo[0]));
|
d@0
|
202 T38 = VFNMSI(T2w, T2v);
|
d@0
|
203 STM2(&(xo[48]), T38, ovs, &(xo[0]));
|
d@0
|
204 T2B = VFMA(LDK(KP707106781), T2A, T2x);
|
d@0
|
205 T2T = VFNMS(LDK(KP707106781), T2A, T2x);
|
d@0
|
206 T2U = VADD(T2O, T2N);
|
d@0
|
207 T2P = VSUB(T2N, T2O);
|
d@0
|
208 T2Z = VFMA(LDK(KP923879532), T2U, T2T);
|
d@0
|
209 T2V = VFNMS(LDK(KP923879532), T2U, T2T);
|
d@0
|
210 T2M = VFMA(LDK(KP707106781), T2L, T2K);
|
d@0
|
211 T2W = VFNMS(LDK(KP707106781), T2L, T2K);
|
d@0
|
212 T2X = VSUB(T2H, T2E);
|
d@0
|
213 T2I = VADD(T2E, T2H);
|
d@0
|
214 T30 = VFNMS(LDK(KP923879532), T2X, T2W);
|
d@0
|
215 T2Y = VFMA(LDK(KP923879532), T2X, T2W);
|
d@0
|
216 T39 = VFMAI(T2Y, T2V);
|
d@0
|
217 STM2(&(xo[20]), T39, ovs, &(xo[0]));
|
d@0
|
218 T3a = VFMAI(T30, T2Z);
|
d@0
|
219 STM2(&(xo[52]), T3a, ovs, &(xo[0]));
|
d@0
|
220 T3b = VFNMSI(T2Y, T2V);
|
d@0
|
221 STM2(&(xo[44]), T3b, ovs, &(xo[0]));
|
d@0
|
222 T3c = VFNMSI(T30, T2Z);
|
d@0
|
223 STM2(&(xo[12]), T3c, ovs, &(xo[0]));
|
d@0
|
224 T2R = VFMA(LDK(KP923879532), T2I, T2B);
|
d@0
|
225 T2J = VFNMS(LDK(KP923879532), T2I, T2B);
|
d@0
|
226 T2S = VFMA(LDK(KP923879532), T2P, T2M);
|
d@0
|
227 T2Q = VFNMS(LDK(KP923879532), T2P, T2M);
|
d@0
|
228 T3d = VFNMSI(T2Q, T2J);
|
d@0
|
229 STM2(&(xo[28]), T3d, ovs, &(xo[0]));
|
d@0
|
230 T3e = VFMAI(T2S, T2R);
|
d@0
|
231 STM2(&(xo[4]), T3e, ovs, &(xo[0]));
|
d@0
|
232 T3f = VFMAI(T2Q, T2J);
|
d@0
|
233 STM2(&(xo[36]), T3f, ovs, &(xo[0]));
|
d@0
|
234 T3g = VFNMSI(T2S, T2R);
|
d@0
|
235 STM2(&(xo[60]), T3g, ovs, &(xo[0]));
|
d@0
|
236 T1J = VFNMS(LDK(KP923879532), T1q, T1p);
|
d@0
|
237 T1r = VFMA(LDK(KP923879532), T1q, T1p);
|
d@0
|
238 T1C = VFMA(LDK(KP923879532), T1B, T1A);
|
d@0
|
239 T1M = VFNMS(LDK(KP923879532), T1B, T1A);
|
d@0
|
240 T1u = VFMA(LDK(KP668178637), T1t, T1s);
|
d@0
|
241 T1D = VFNMS(LDK(KP668178637), T1s, T1t);
|
d@0
|
242 T1E = VFNMS(LDK(KP668178637), T1v, T1w);
|
d@0
|
243 T1x = VFMA(LDK(KP668178637), T1w, T1v);
|
d@0
|
244 T1K = VADD(T1D, T1E);
|
d@0
|
245 T1F = VSUB(T1D, T1E);
|
d@0
|
246 T1N = VSUB(T1x, T1u);
|
d@0
|
247 T1y = VADD(T1u, T1x);
|
d@0
|
248 T1H = VFMA(LDK(KP831469612), T1y, T1r);
|
d@0
|
249 T1z = VFNMS(LDK(KP831469612), T1y, T1r);
|
d@0
|
250 T1G = VFNMS(LDK(KP831469612), T1F, T1C);
|
d@0
|
251 T1I = VFMA(LDK(KP831469612), T1F, T1C);
|
d@0
|
252 T3h = VFNMSI(T1G, T1z);
|
d@0
|
253 STM2(&(xo[26]), T3h, ovs, &(xo[2]));
|
d@0
|
254 STN2(&(xo[24]), T31, T3h, ovs);
|
d@0
|
255 T3i = VFMAI(T1G, T1z);
|
d@0
|
256 STM2(&(xo[38]), T3i, ovs, &(xo[2]));
|
d@0
|
257 STN2(&(xo[36]), T3f, T3i, ovs);
|
d@0
|
258 T1L = VFMA(LDK(KP831469612), T1K, T1J);
|
d@0
|
259 T1P = VFNMS(LDK(KP831469612), T1K, T1J);
|
d@0
|
260 T1Q = VFNMS(LDK(KP831469612), T1N, T1M);
|
d@0
|
261 T1O = VFMA(LDK(KP831469612), T1N, T1M);
|
d@0
|
262 T3j = VFNMSI(T1Q, T1P);
|
d@0
|
263 STM2(&(xo[10]), T3j, ovs, &(xo[2]));
|
d@0
|
264 STN2(&(xo[8]), T32, T3j, ovs);
|
d@0
|
265 T3k = VFMAI(T1Q, T1P);
|
d@0
|
266 STM2(&(xo[54]), T3k, ovs, &(xo[2]));
|
d@0
|
267 STN2(&(xo[52]), T3a, T3k, ovs);
|
d@0
|
268 T3l = VFNMSI(T1I, T1H);
|
d@0
|
269 STM2(&(xo[58]), T3l, ovs, &(xo[2]));
|
d@0
|
270 STN2(&(xo[56]), T34, T3l, ovs);
|
d@0
|
271 T3m = VFMAI(T1I, T1H);
|
d@0
|
272 STM2(&(xo[6]), T3m, ovs, &(xo[2]));
|
d@0
|
273 STN2(&(xo[4]), T3e, T3m, ovs);
|
d@0
|
274 T3n = VFMAI(T1O, T1L);
|
d@0
|
275 STM2(&(xo[22]), T3n, ovs, &(xo[2]));
|
d@0
|
276 STN2(&(xo[20]), T39, T3n, ovs);
|
d@0
|
277 T3o = VFNMSI(T1O, T1L);
|
d@0
|
278 STM2(&(xo[42]), T3o, ovs, &(xo[2]));
|
d@0
|
279 STN2(&(xo[40]), T33, T3o, ovs);
|
d@0
|
280 T1h = VFNMS(LDK(KP923879532), Tq, Tb);
|
d@0
|
281 Tr = VFMA(LDK(KP923879532), Tq, Tb);
|
d@0
|
282 T1a = VFMA(LDK(KP923879532), T19, T16);
|
d@0
|
283 T1k = VFNMS(LDK(KP923879532), T19, T16);
|
d@0
|
284 TI = VFNMS(LDK(KP198912367), TH, TC);
|
d@0
|
285 T1b = VFMA(LDK(KP198912367), TC, TH);
|
d@0
|
286 T1c = VFMA(LDK(KP198912367), TT, TY);
|
d@0
|
287 TZ = VFNMS(LDK(KP198912367), TY, TT);
|
d@0
|
288 T1i = VADD(T1b, T1c);
|
d@0
|
289 T1d = VSUB(T1b, T1c);
|
d@0
|
290 T1l = VSUB(TZ, TI);
|
d@0
|
291 T10 = VADD(TI, TZ);
|
d@0
|
292 T1f = VFMA(LDK(KP980785280), T10, Tr);
|
d@0
|
293 T11 = VFNMS(LDK(KP980785280), T10, Tr);
|
d@0
|
294 T1e = VFNMS(LDK(KP980785280), T1d, T1a);
|
d@0
|
295 T1g = VFMA(LDK(KP980785280), T1d, T1a);
|
d@0
|
296 T3p = VFNMSI(T1e, T11);
|
d@0
|
297 STM2(&(xo[34]), T3p, ovs, &(xo[2]));
|
d@0
|
298 STN2(&(xo[32]), T35, T3p, ovs);
|
d@0
|
299 T3q = VFMAI(T1e, T11);
|
d@0
|
300 STM2(&(xo[30]), T3q, ovs, &(xo[2]));
|
d@0
|
301 STN2(&(xo[28]), T3d, T3q, ovs);
|
d@0
|
302 T1j = VFNMS(LDK(KP980785280), T1i, T1h);
|
d@0
|
303 T1n = VFMA(LDK(KP980785280), T1i, T1h);
|
d@0
|
304 T1o = VFMA(LDK(KP980785280), T1l, T1k);
|
d@0
|
305 T1m = VFNMS(LDK(KP980785280), T1l, T1k);
|
d@0
|
306 T3r = VFMAI(T1o, T1n);
|
d@0
|
307 STM2(&(xo[14]), T3r, ovs, &(xo[2]));
|
d@0
|
308 STN2(&(xo[12]), T3c, T3r, ovs);
|
d@0
|
309 T3s = VFNMSI(T1o, T1n);
|
d@0
|
310 STM2(&(xo[50]), T3s, ovs, &(xo[2]));
|
d@0
|
311 STN2(&(xo[48]), T38, T3s, ovs);
|
d@0
|
312 T3t = VFNMSI(T1g, T1f);
|
d@0
|
313 STM2(&(xo[2]), T3t, ovs, &(xo[2]));
|
d@0
|
314 STN2(&(xo[0]), T37, T3t, ovs);
|
d@0
|
315 T3u = VFMAI(T1g, T1f);
|
d@0
|
316 STM2(&(xo[62]), T3u, ovs, &(xo[2]));
|
d@0
|
317 STN2(&(xo[60]), T3g, T3u, ovs);
|
d@0
|
318 T3v = VFNMSI(T1m, T1j);
|
d@0
|
319 STM2(&(xo[18]), T3v, ovs, &(xo[2]));
|
d@0
|
320 STN2(&(xo[16]), T36, T3v, ovs);
|
d@0
|
321 T3w = VFMAI(T1m, T1j);
|
d@0
|
322 STM2(&(xo[46]), T3w, ovs, &(xo[2]));
|
d@0
|
323 STN2(&(xo[44]), T3b, T3w, ovs);
|
d@0
|
324 }
|
d@0
|
325 }
|