Chris@82
|
1 /*
|
Chris@82
|
2 * Copyright (c) 2003, 2007-14 Matteo Frigo
|
Chris@82
|
3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
|
Chris@82
|
4 *
|
Chris@82
|
5 * This program is free software; you can redistribute it and/or modify
|
Chris@82
|
6 * it under the terms of the GNU General Public License as published by
|
Chris@82
|
7 * the Free Software Foundation; either version 2 of the License, or
|
Chris@82
|
8 * (at your option) any later version.
|
Chris@82
|
9 *
|
Chris@82
|
10 * This program is distributed in the hope that it will be useful,
|
Chris@82
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
Chris@82
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
Chris@82
|
13 * GNU General Public License for more details.
|
Chris@82
|
14 *
|
Chris@82
|
15 * You should have received a copy of the GNU General Public License
|
Chris@82
|
16 * along with this program; if not, write to the Free Software
|
Chris@82
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
Chris@82
|
18 *
|
Chris@82
|
19 */
|
Chris@82
|
20
|
Chris@82
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
Chris@82
|
22 /* Generated on Thu May 24 08:05:09 EDT 2018 */
|
Chris@82
|
23
|
Chris@82
|
24 #include "dft/codelet-dft.h"
|
Chris@82
|
25
|
Chris@82
|
26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
|
Chris@82
|
27
|
Chris@82
|
28 /* Generated by: ../../../genfft/gen_notw_c.native -fma -simd -compact -variables 4 -pipeline-latency 8 -n 64 -name n2fv_64 -with-ostride 2 -include dft/simd/n2f.h -store-multiple 2 */
|
Chris@82
|
29
|
Chris@82
|
30 /*
|
Chris@82
|
31 * This function contains 456 FP additions, 258 FP multiplications,
|
Chris@82
|
32 * (or, 198 additions, 0 multiplications, 258 fused multiply/add),
|
Chris@82
|
33 * 120 stack variables, 15 constants, and 160 memory accesses
|
Chris@82
|
34 */
|
Chris@82
|
35 #include "dft/simd/n2f.h"
|
Chris@82
|
36
|
Chris@82
|
37 static void n2fv_64(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
|
Chris@82
|
38 {
|
Chris@82
|
39 DVK(KP773010453, +0.773010453362736960810906609758469800971041293);
|
Chris@82
|
40 DVK(KP820678790, +0.820678790828660330972281985331011598767386482);
|
Chris@82
|
41 DVK(KP881921264, +0.881921264348355029712756863660388349508442621);
|
Chris@82
|
42 DVK(KP534511135, +0.534511135950791641089685961295362908582039528);
|
Chris@82
|
43 DVK(KP995184726, +0.995184726672196886244836953109479921575474869);
|
Chris@82
|
44 DVK(KP098491403, +0.098491403357164253077197521291327432293052451);
|
Chris@82
|
45 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
Chris@82
|
46 DVK(KP956940335, +0.956940335732208864935797886980269969482849206);
|
Chris@82
|
47 DVK(KP303346683, +0.303346683607342391675883946941299872384187453);
|
Chris@82
|
48 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
Chris@82
|
49 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
Chris@82
|
50 DVK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
Chris@82
|
51 DVK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
Chris@82
|
52 DVK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
Chris@82
|
53 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@82
|
54 {
|
Chris@82
|
55 INT i;
|
Chris@82
|
56 const R *xi;
|
Chris@82
|
57 R *xo;
|
Chris@82
|
58 xi = ri;
|
Chris@82
|
59 xo = ro;
|
Chris@82
|
60 for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(128, is), MAKE_VOLATILE_STRIDE(128, os)) {
|
Chris@82
|
61 V T7, T26, T47, T69, T5k, T6A, T2V, T3z, Tm, T27, T5n, T6a, T2Y, T3M, T4e;
|
Chris@82
|
62 V T6B, TC, T2a, T6e, T6E, T3l, T3A, T4o, T5p, TR, T29, T6h, T6D, T3i, T3B;
|
Chris@82
|
63 V T4x, T5q, T1N, T2x, T6t, T71, T6w, T72, T1W, T2y, T39, T3H, T57, T5N, T5e;
|
Chris@82
|
64 V T5O, T3c, T3I, T1g, T2u, T6m, T6Y, T6p, T6Z, T1p, T2v, T32, T3E, T4M, T5K;
|
Chris@82
|
65 V T4T, T5L, T35, T3F;
|
Chris@82
|
66 {
|
Chris@82
|
67 V T3, T43, T25, T44, T6, T5i, T22, T45;
|
Chris@82
|
68 {
|
Chris@82
|
69 V T1, T2, T23, T24;
|
Chris@82
|
70 T1 = LD(&(xi[0]), ivs, &(xi[0]));
|
Chris@82
|
71 T2 = LD(&(xi[WS(is, 32)]), ivs, &(xi[0]));
|
Chris@82
|
72 T3 = VADD(T1, T2);
|
Chris@82
|
73 T43 = VSUB(T1, T2);
|
Chris@82
|
74 T23 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
|
Chris@82
|
75 T24 = LD(&(xi[WS(is, 40)]), ivs, &(xi[0]));
|
Chris@82
|
76 T25 = VADD(T23, T24);
|
Chris@82
|
77 T44 = VSUB(T23, T24);
|
Chris@82
|
78 }
|
Chris@82
|
79 {
|
Chris@82
|
80 V T4, T5, T20, T21;
|
Chris@82
|
81 T4 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
|
Chris@82
|
82 T5 = LD(&(xi[WS(is, 48)]), ivs, &(xi[0]));
|
Chris@82
|
83 T6 = VADD(T4, T5);
|
Chris@82
|
84 T5i = VSUB(T4, T5);
|
Chris@82
|
85 T20 = LD(&(xi[WS(is, 56)]), ivs, &(xi[0]));
|
Chris@82
|
86 T21 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
|
Chris@82
|
87 T22 = VADD(T20, T21);
|
Chris@82
|
88 T45 = VSUB(T20, T21);
|
Chris@82
|
89 }
|
Chris@82
|
90 T7 = VSUB(T3, T6);
|
Chris@82
|
91 T26 = VSUB(T22, T25);
|
Chris@82
|
92 {
|
Chris@82
|
93 V T46, T5j, T2T, T2U;
|
Chris@82
|
94 T46 = VADD(T44, T45);
|
Chris@82
|
95 T47 = VFMA(LDK(KP707106781), T46, T43);
|
Chris@82
|
96 T69 = VFNMS(LDK(KP707106781), T46, T43);
|
Chris@82
|
97 T5j = VSUB(T45, T44);
|
Chris@82
|
98 T5k = VFNMS(LDK(KP707106781), T5j, T5i);
|
Chris@82
|
99 T6A = VFMA(LDK(KP707106781), T5j, T5i);
|
Chris@82
|
100 T2T = VADD(T3, T6);
|
Chris@82
|
101 T2U = VADD(T25, T22);
|
Chris@82
|
102 T2V = VADD(T2T, T2U);
|
Chris@82
|
103 T3z = VSUB(T2T, T2U);
|
Chris@82
|
104 }
|
Chris@82
|
105 }
|
Chris@82
|
106 {
|
Chris@82
|
107 V Ta, T48, Tk, T4c, Td, T49, Th, T4b;
|
Chris@82
|
108 {
|
Chris@82
|
109 V T8, T9, Ti, Tj;
|
Chris@82
|
110 T8 = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
|
Chris@82
|
111 T9 = LD(&(xi[WS(is, 36)]), ivs, &(xi[0]));
|
Chris@82
|
112 Ta = VADD(T8, T9);
|
Chris@82
|
113 T48 = VSUB(T8, T9);
|
Chris@82
|
114 Ti = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
|
Chris@82
|
115 Tj = LD(&(xi[WS(is, 44)]), ivs, &(xi[0]));
|
Chris@82
|
116 Tk = VADD(Ti, Tj);
|
Chris@82
|
117 T4c = VSUB(Tj, Ti);
|
Chris@82
|
118 }
|
Chris@82
|
119 {
|
Chris@82
|
120 V Tb, Tc, Tf, Tg;
|
Chris@82
|
121 Tb = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
|
Chris@82
|
122 Tc = LD(&(xi[WS(is, 52)]), ivs, &(xi[0]));
|
Chris@82
|
123 Td = VADD(Tb, Tc);
|
Chris@82
|
124 T49 = VSUB(Tb, Tc);
|
Chris@82
|
125 Tf = LD(&(xi[WS(is, 60)]), ivs, &(xi[0]));
|
Chris@82
|
126 Tg = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
|
Chris@82
|
127 Th = VADD(Tf, Tg);
|
Chris@82
|
128 T4b = VSUB(Tf, Tg);
|
Chris@82
|
129 }
|
Chris@82
|
130 {
|
Chris@82
|
131 V Te, Tl, T5l, T5m;
|
Chris@82
|
132 Te = VSUB(Ta, Td);
|
Chris@82
|
133 Tl = VSUB(Th, Tk);
|
Chris@82
|
134 Tm = VADD(Te, Tl);
|
Chris@82
|
135 T27 = VSUB(Tl, Te);
|
Chris@82
|
136 T5l = VFMA(LDK(KP414213562), T48, T49);
|
Chris@82
|
137 T5m = VFMA(LDK(KP414213562), T4b, T4c);
|
Chris@82
|
138 T5n = VSUB(T5l, T5m);
|
Chris@82
|
139 T6a = VADD(T5l, T5m);
|
Chris@82
|
140 }
|
Chris@82
|
141 {
|
Chris@82
|
142 V T2W, T2X, T4a, T4d;
|
Chris@82
|
143 T2W = VADD(Ta, Td);
|
Chris@82
|
144 T2X = VADD(Th, Tk);
|
Chris@82
|
145 T2Y = VADD(T2W, T2X);
|
Chris@82
|
146 T3M = VSUB(T2X, T2W);
|
Chris@82
|
147 T4a = VFNMS(LDK(KP414213562), T49, T48);
|
Chris@82
|
148 T4d = VFNMS(LDK(KP414213562), T4c, T4b);
|
Chris@82
|
149 T4e = VADD(T4a, T4d);
|
Chris@82
|
150 T6B = VSUB(T4d, T4a);
|
Chris@82
|
151 }
|
Chris@82
|
152 }
|
Chris@82
|
153 {
|
Chris@82
|
154 V Tq, T4g, Tt, T4l, Tx, T4m, TA, T4j;
|
Chris@82
|
155 {
|
Chris@82
|
156 V To, Tp, Tr, Ts;
|
Chris@82
|
157 To = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
|
Chris@82
|
158 Tp = LD(&(xi[WS(is, 34)]), ivs, &(xi[0]));
|
Chris@82
|
159 Tq = VADD(To, Tp);
|
Chris@82
|
160 T4g = VSUB(To, Tp);
|
Chris@82
|
161 Tr = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
|
Chris@82
|
162 Ts = LD(&(xi[WS(is, 50)]), ivs, &(xi[0]));
|
Chris@82
|
163 Tt = VADD(Tr, Ts);
|
Chris@82
|
164 T4l = VSUB(Tr, Ts);
|
Chris@82
|
165 {
|
Chris@82
|
166 V Tv, Tw, T4h, Ty, Tz, T4i;
|
Chris@82
|
167 Tv = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
|
Chris@82
|
168 Tw = LD(&(xi[WS(is, 42)]), ivs, &(xi[0]));
|
Chris@82
|
169 T4h = VSUB(Tv, Tw);
|
Chris@82
|
170 Ty = LD(&(xi[WS(is, 58)]), ivs, &(xi[0]));
|
Chris@82
|
171 Tz = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
|
Chris@82
|
172 T4i = VSUB(Ty, Tz);
|
Chris@82
|
173 Tx = VADD(Tv, Tw);
|
Chris@82
|
174 T4m = VSUB(T4h, T4i);
|
Chris@82
|
175 TA = VADD(Ty, Tz);
|
Chris@82
|
176 T4j = VADD(T4h, T4i);
|
Chris@82
|
177 }
|
Chris@82
|
178 }
|
Chris@82
|
179 {
|
Chris@82
|
180 V Tu, TB, T6c, T6d;
|
Chris@82
|
181 Tu = VSUB(Tq, Tt);
|
Chris@82
|
182 TB = VSUB(Tx, TA);
|
Chris@82
|
183 TC = VFNMS(LDK(KP414213562), TB, Tu);
|
Chris@82
|
184 T2a = VFMA(LDK(KP414213562), Tu, TB);
|
Chris@82
|
185 T6c = VFNMS(LDK(KP707106781), T4m, T4l);
|
Chris@82
|
186 T6d = VFNMS(LDK(KP707106781), T4j, T4g);
|
Chris@82
|
187 T6e = VFNMS(LDK(KP668178637), T6d, T6c);
|
Chris@82
|
188 T6E = VFMA(LDK(KP668178637), T6c, T6d);
|
Chris@82
|
189 }
|
Chris@82
|
190 {
|
Chris@82
|
191 V T3j, T3k, T4k, T4n;
|
Chris@82
|
192 T3j = VADD(Tq, Tt);
|
Chris@82
|
193 T3k = VADD(Tx, TA);
|
Chris@82
|
194 T3l = VADD(T3j, T3k);
|
Chris@82
|
195 T3A = VSUB(T3j, T3k);
|
Chris@82
|
196 T4k = VFMA(LDK(KP707106781), T4j, T4g);
|
Chris@82
|
197 T4n = VFMA(LDK(KP707106781), T4m, T4l);
|
Chris@82
|
198 T4o = VFNMS(LDK(KP198912367), T4n, T4k);
|
Chris@82
|
199 T5p = VFMA(LDK(KP198912367), T4k, T4n);
|
Chris@82
|
200 }
|
Chris@82
|
201 }
|
Chris@82
|
202 {
|
Chris@82
|
203 V TF, T4p, TI, T4u, TM, T4v, TP, T4s;
|
Chris@82
|
204 {
|
Chris@82
|
205 V TD, TE, TG, TH;
|
Chris@82
|
206 TD = LD(&(xi[WS(is, 62)]), ivs, &(xi[0]));
|
Chris@82
|
207 TE = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
|
Chris@82
|
208 TF = VADD(TD, TE);
|
Chris@82
|
209 T4p = VSUB(TD, TE);
|
Chris@82
|
210 TG = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
|
Chris@82
|
211 TH = LD(&(xi[WS(is, 46)]), ivs, &(xi[0]));
|
Chris@82
|
212 TI = VADD(TG, TH);
|
Chris@82
|
213 T4u = VSUB(TH, TG);
|
Chris@82
|
214 {
|
Chris@82
|
215 V TK, TL, T4r, TN, TO, T4q;
|
Chris@82
|
216 TK = LD(&(xi[WS(is, 54)]), ivs, &(xi[0]));
|
Chris@82
|
217 TL = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
|
Chris@82
|
218 T4r = VSUB(TK, TL);
|
Chris@82
|
219 TN = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
|
Chris@82
|
220 TO = LD(&(xi[WS(is, 38)]), ivs, &(xi[0]));
|
Chris@82
|
221 T4q = VSUB(TN, TO);
|
Chris@82
|
222 TM = VADD(TK, TL);
|
Chris@82
|
223 T4v = VSUB(T4r, T4q);
|
Chris@82
|
224 TP = VADD(TN, TO);
|
Chris@82
|
225 T4s = VADD(T4q, T4r);
|
Chris@82
|
226 }
|
Chris@82
|
227 }
|
Chris@82
|
228 {
|
Chris@82
|
229 V TJ, TQ, T6f, T6g;
|
Chris@82
|
230 TJ = VSUB(TF, TI);
|
Chris@82
|
231 TQ = VSUB(TM, TP);
|
Chris@82
|
232 TR = VFNMS(LDK(KP414213562), TQ, TJ);
|
Chris@82
|
233 T29 = VFMA(LDK(KP414213562), TJ, TQ);
|
Chris@82
|
234 T6f = VFNMS(LDK(KP707106781), T4v, T4u);
|
Chris@82
|
235 T6g = VFNMS(LDK(KP707106781), T4s, T4p);
|
Chris@82
|
236 T6h = VFNMS(LDK(KP668178637), T6g, T6f);
|
Chris@82
|
237 T6D = VFMA(LDK(KP668178637), T6f, T6g);
|
Chris@82
|
238 }
|
Chris@82
|
239 {
|
Chris@82
|
240 V T3g, T3h, T4t, T4w;
|
Chris@82
|
241 T3g = VADD(TF, TI);
|
Chris@82
|
242 T3h = VADD(TP, TM);
|
Chris@82
|
243 T3i = VADD(T3g, T3h);
|
Chris@82
|
244 T3B = VSUB(T3g, T3h);
|
Chris@82
|
245 T4t = VFMA(LDK(KP707106781), T4s, T4p);
|
Chris@82
|
246 T4w = VFMA(LDK(KP707106781), T4v, T4u);
|
Chris@82
|
247 T4x = VFNMS(LDK(KP198912367), T4w, T4t);
|
Chris@82
|
248 T5q = VFMA(LDK(KP198912367), T4t, T4w);
|
Chris@82
|
249 }
|
Chris@82
|
250 }
|
Chris@82
|
251 {
|
Chris@82
|
252 V T1t, T4V, T1w, T58, T1Q, T59, T1T, T4Y, T1A, T1D, T1E, T5b, T52, T1H, T1K;
|
Chris@82
|
253 V T1L, T5c, T55;
|
Chris@82
|
254 {
|
Chris@82
|
255 V T1r, T1s, T1u, T1v;
|
Chris@82
|
256 T1r = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
257 T1s = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
258 T1t = VADD(T1r, T1s);
|
Chris@82
|
259 T4V = VSUB(T1r, T1s);
|
Chris@82
|
260 T1u = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
261 T1v = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
262 T1w = VADD(T1u, T1v);
|
Chris@82
|
263 T58 = VSUB(T1v, T1u);
|
Chris@82
|
264 }
|
Chris@82
|
265 {
|
Chris@82
|
266 V T1O, T1P, T4X, T1R, T1S, T4W;
|
Chris@82
|
267 T1O = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
268 T1P = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
269 T4X = VSUB(T1O, T1P);
|
Chris@82
|
270 T1R = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
271 T1S = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
272 T4W = VSUB(T1R, T1S);
|
Chris@82
|
273 T1Q = VADD(T1O, T1P);
|
Chris@82
|
274 T59 = VSUB(T4X, T4W);
|
Chris@82
|
275 T1T = VADD(T1R, T1S);
|
Chris@82
|
276 T4Y = VADD(T4W, T4X);
|
Chris@82
|
277 }
|
Chris@82
|
278 {
|
Chris@82
|
279 V T50, T51, T53, T54;
|
Chris@82
|
280 {
|
Chris@82
|
281 V T1y, T1z, T1B, T1C;
|
Chris@82
|
282 T1y = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
283 T1z = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
284 T1A = VADD(T1y, T1z);
|
Chris@82
|
285 T50 = VSUB(T1y, T1z);
|
Chris@82
|
286 T1B = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
287 T1C = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
288 T1D = VADD(T1B, T1C);
|
Chris@82
|
289 T51 = VSUB(T1C, T1B);
|
Chris@82
|
290 }
|
Chris@82
|
291 T1E = VSUB(T1A, T1D);
|
Chris@82
|
292 T5b = VFNMS(LDK(KP414213562), T50, T51);
|
Chris@82
|
293 T52 = VFMA(LDK(KP414213562), T51, T50);
|
Chris@82
|
294 {
|
Chris@82
|
295 V T1F, T1G, T1I, T1J;
|
Chris@82
|
296 T1F = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
297 T1G = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
298 T1H = VADD(T1F, T1G);
|
Chris@82
|
299 T53 = VSUB(T1F, T1G);
|
Chris@82
|
300 T1I = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
301 T1J = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
302 T1K = VADD(T1I, T1J);
|
Chris@82
|
303 T54 = VSUB(T1J, T1I);
|
Chris@82
|
304 }
|
Chris@82
|
305 T1L = VSUB(T1H, T1K);
|
Chris@82
|
306 T5c = VFMA(LDK(KP414213562), T53, T54);
|
Chris@82
|
307 T55 = VFNMS(LDK(KP414213562), T54, T53);
|
Chris@82
|
308 }
|
Chris@82
|
309 {
|
Chris@82
|
310 V T1x, T1M, T6r, T6s;
|
Chris@82
|
311 T1x = VSUB(T1t, T1w);
|
Chris@82
|
312 T1M = VADD(T1E, T1L);
|
Chris@82
|
313 T1N = VFMA(LDK(KP707106781), T1M, T1x);
|
Chris@82
|
314 T2x = VFNMS(LDK(KP707106781), T1M, T1x);
|
Chris@82
|
315 T6r = VFNMS(LDK(KP707106781), T4Y, T4V);
|
Chris@82
|
316 T6s = VSUB(T5c, T5b);
|
Chris@82
|
317 T6t = VFNMS(LDK(KP923879532), T6s, T6r);
|
Chris@82
|
318 T71 = VFMA(LDK(KP923879532), T6s, T6r);
|
Chris@82
|
319 }
|
Chris@82
|
320 {
|
Chris@82
|
321 V T6u, T6v, T1U, T1V;
|
Chris@82
|
322 T6u = VFNMS(LDK(KP707106781), T59, T58);
|
Chris@82
|
323 T6v = VSUB(T55, T52);
|
Chris@82
|
324 T6w = VFMA(LDK(KP923879532), T6v, T6u);
|
Chris@82
|
325 T72 = VFNMS(LDK(KP923879532), T6v, T6u);
|
Chris@82
|
326 T1U = VSUB(T1Q, T1T);
|
Chris@82
|
327 T1V = VSUB(T1L, T1E);
|
Chris@82
|
328 T1W = VFMA(LDK(KP707106781), T1V, T1U);
|
Chris@82
|
329 T2y = VFNMS(LDK(KP707106781), T1V, T1U);
|
Chris@82
|
330 }
|
Chris@82
|
331 {
|
Chris@82
|
332 V T37, T38, T4Z, T56;
|
Chris@82
|
333 T37 = VADD(T1t, T1w);
|
Chris@82
|
334 T38 = VADD(T1T, T1Q);
|
Chris@82
|
335 T39 = VADD(T37, T38);
|
Chris@82
|
336 T3H = VSUB(T37, T38);
|
Chris@82
|
337 T4Z = VFMA(LDK(KP707106781), T4Y, T4V);
|
Chris@82
|
338 T56 = VADD(T52, T55);
|
Chris@82
|
339 T57 = VFMA(LDK(KP923879532), T56, T4Z);
|
Chris@82
|
340 T5N = VFNMS(LDK(KP923879532), T56, T4Z);
|
Chris@82
|
341 }
|
Chris@82
|
342 {
|
Chris@82
|
343 V T5a, T5d, T3a, T3b;
|
Chris@82
|
344 T5a = VFMA(LDK(KP707106781), T59, T58);
|
Chris@82
|
345 T5d = VADD(T5b, T5c);
|
Chris@82
|
346 T5e = VFMA(LDK(KP923879532), T5d, T5a);
|
Chris@82
|
347 T5O = VFNMS(LDK(KP923879532), T5d, T5a);
|
Chris@82
|
348 T3a = VADD(T1A, T1D);
|
Chris@82
|
349 T3b = VADD(T1H, T1K);
|
Chris@82
|
350 T3c = VADD(T3a, T3b);
|
Chris@82
|
351 T3I = VSUB(T3b, T3a);
|
Chris@82
|
352 }
|
Chris@82
|
353 }
|
Chris@82
|
354 {
|
Chris@82
|
355 V TW, T4A, TZ, T4N, T1j, T4O, T1m, T4D, T13, T16, T17, T4Q, T4H, T1a, T1d;
|
Chris@82
|
356 V T1e, T4R, T4K;
|
Chris@82
|
357 {
|
Chris@82
|
358 V TU, TV, TX, TY;
|
Chris@82
|
359 TU = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
360 TV = LD(&(xi[WS(is, 33)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
361 TW = VADD(TU, TV);
|
Chris@82
|
362 T4A = VSUB(TU, TV);
|
Chris@82
|
363 TX = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
364 TY = LD(&(xi[WS(is, 49)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
365 TZ = VADD(TX, TY);
|
Chris@82
|
366 T4N = VSUB(TX, TY);
|
Chris@82
|
367 }
|
Chris@82
|
368 {
|
Chris@82
|
369 V T1h, T1i, T4B, T1k, T1l, T4C;
|
Chris@82
|
370 T1h = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
371 T1i = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
372 T4B = VSUB(T1h, T1i);
|
Chris@82
|
373 T1k = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
374 T1l = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
375 T4C = VSUB(T1k, T1l);
|
Chris@82
|
376 T1j = VADD(T1h, T1i);
|
Chris@82
|
377 T4O = VSUB(T4B, T4C);
|
Chris@82
|
378 T1m = VADD(T1k, T1l);
|
Chris@82
|
379 T4D = VADD(T4B, T4C);
|
Chris@82
|
380 }
|
Chris@82
|
381 {
|
Chris@82
|
382 V T4F, T4G, T4I, T4J;
|
Chris@82
|
383 {
|
Chris@82
|
384 V T11, T12, T14, T15;
|
Chris@82
|
385 T11 = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
386 T12 = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
387 T13 = VADD(T11, T12);
|
Chris@82
|
388 T4F = VSUB(T11, T12);
|
Chris@82
|
389 T14 = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
390 T15 = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
391 T16 = VADD(T14, T15);
|
Chris@82
|
392 T4G = VSUB(T14, T15);
|
Chris@82
|
393 }
|
Chris@82
|
394 T17 = VSUB(T13, T16);
|
Chris@82
|
395 T4Q = VFMA(LDK(KP414213562), T4F, T4G);
|
Chris@82
|
396 T4H = VFNMS(LDK(KP414213562), T4G, T4F);
|
Chris@82
|
397 {
|
Chris@82
|
398 V T18, T19, T1b, T1c;
|
Chris@82
|
399 T18 = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
400 T19 = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
401 T1a = VADD(T18, T19);
|
Chris@82
|
402 T4I = VSUB(T18, T19);
|
Chris@82
|
403 T1b = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
404 T1c = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
405 T1d = VADD(T1b, T1c);
|
Chris@82
|
406 T4J = VSUB(T1b, T1c);
|
Chris@82
|
407 }
|
Chris@82
|
408 T1e = VSUB(T1a, T1d);
|
Chris@82
|
409 T4R = VFNMS(LDK(KP414213562), T4I, T4J);
|
Chris@82
|
410 T4K = VFMA(LDK(KP414213562), T4J, T4I);
|
Chris@82
|
411 }
|
Chris@82
|
412 {
|
Chris@82
|
413 V T10, T1f, T6k, T6l;
|
Chris@82
|
414 T10 = VSUB(TW, TZ);
|
Chris@82
|
415 T1f = VADD(T17, T1e);
|
Chris@82
|
416 T1g = VFMA(LDK(KP707106781), T1f, T10);
|
Chris@82
|
417 T2u = VFNMS(LDK(KP707106781), T1f, T10);
|
Chris@82
|
418 T6k = VFNMS(LDK(KP707106781), T4D, T4A);
|
Chris@82
|
419 T6l = VSUB(T4Q, T4R);
|
Chris@82
|
420 T6m = VFNMS(LDK(KP923879532), T6l, T6k);
|
Chris@82
|
421 T6Y = VFMA(LDK(KP923879532), T6l, T6k);
|
Chris@82
|
422 }
|
Chris@82
|
423 {
|
Chris@82
|
424 V T6n, T6o, T1n, T1o;
|
Chris@82
|
425 T6n = VFNMS(LDK(KP707106781), T4O, T4N);
|
Chris@82
|
426 T6o = VSUB(T4H, T4K);
|
Chris@82
|
427 T6p = VFMA(LDK(KP923879532), T6o, T6n);
|
Chris@82
|
428 T6Z = VFNMS(LDK(KP923879532), T6o, T6n);
|
Chris@82
|
429 T1n = VSUB(T1j, T1m);
|
Chris@82
|
430 T1o = VSUB(T17, T1e);
|
Chris@82
|
431 T1p = VFMA(LDK(KP707106781), T1o, T1n);
|
Chris@82
|
432 T2v = VFNMS(LDK(KP707106781), T1o, T1n);
|
Chris@82
|
433 }
|
Chris@82
|
434 {
|
Chris@82
|
435 V T30, T31, T4E, T4L;
|
Chris@82
|
436 T30 = VADD(TW, TZ);
|
Chris@82
|
437 T31 = VADD(T1j, T1m);
|
Chris@82
|
438 T32 = VADD(T30, T31);
|
Chris@82
|
439 T3E = VSUB(T30, T31);
|
Chris@82
|
440 T4E = VFMA(LDK(KP707106781), T4D, T4A);
|
Chris@82
|
441 T4L = VADD(T4H, T4K);
|
Chris@82
|
442 T4M = VFMA(LDK(KP923879532), T4L, T4E);
|
Chris@82
|
443 T5K = VFNMS(LDK(KP923879532), T4L, T4E);
|
Chris@82
|
444 }
|
Chris@82
|
445 {
|
Chris@82
|
446 V T4P, T4S, T33, T34;
|
Chris@82
|
447 T4P = VFMA(LDK(KP707106781), T4O, T4N);
|
Chris@82
|
448 T4S = VADD(T4Q, T4R);
|
Chris@82
|
449 T4T = VFMA(LDK(KP923879532), T4S, T4P);
|
Chris@82
|
450 T5L = VFNMS(LDK(KP923879532), T4S, T4P);
|
Chris@82
|
451 T33 = VADD(T13, T16);
|
Chris@82
|
452 T34 = VADD(T1a, T1d);
|
Chris@82
|
453 T35 = VADD(T33, T34);
|
Chris@82
|
454 T3F = VSUB(T33, T34);
|
Chris@82
|
455 }
|
Chris@82
|
456 }
|
Chris@82
|
457 {
|
Chris@82
|
458 V T7n, T7o, T7p, T7q, T7r, T7s, T7t, T7u, T7w, T7y, T7A, T7B, T7E, T7G, T7I;
|
Chris@82
|
459 V T7J;
|
Chris@82
|
460 {
|
Chris@82
|
461 V T3t, T3x, T3w, T3y;
|
Chris@82
|
462 {
|
Chris@82
|
463 V T3r, T3s, T3u, T3v;
|
Chris@82
|
464 T3r = VADD(T2V, T2Y);
|
Chris@82
|
465 T3s = VADD(T3l, T3i);
|
Chris@82
|
466 T3t = VADD(T3r, T3s);
|
Chris@82
|
467 T3x = VSUB(T3r, T3s);
|
Chris@82
|
468 T3u = VADD(T32, T35);
|
Chris@82
|
469 T3v = VADD(T39, T3c);
|
Chris@82
|
470 T3w = VADD(T3u, T3v);
|
Chris@82
|
471 T3y = VSUB(T3v, T3u);
|
Chris@82
|
472 }
|
Chris@82
|
473 T7n = VSUB(T3t, T3w);
|
Chris@82
|
474 STM2(&(xo[64]), T7n, ovs, &(xo[0]));
|
Chris@82
|
475 T7o = VFMAI(T3y, T3x);
|
Chris@82
|
476 STM2(&(xo[32]), T7o, ovs, &(xo[0]));
|
Chris@82
|
477 T7p = VADD(T3t, T3w);
|
Chris@82
|
478 STM2(&(xo[0]), T7p, ovs, &(xo[0]));
|
Chris@82
|
479 T7q = VFNMSI(T3y, T3x);
|
Chris@82
|
480 STM2(&(xo[96]), T7q, ovs, &(xo[0]));
|
Chris@82
|
481 }
|
Chris@82
|
482 {
|
Chris@82
|
483 V T2Z, T3m, T3e, T3n, T36, T3d;
|
Chris@82
|
484 T2Z = VSUB(T2V, T2Y);
|
Chris@82
|
485 T3m = VSUB(T3i, T3l);
|
Chris@82
|
486 T36 = VSUB(T32, T35);
|
Chris@82
|
487 T3d = VSUB(T39, T3c);
|
Chris@82
|
488 T3e = VADD(T36, T3d);
|
Chris@82
|
489 T3n = VSUB(T3d, T36);
|
Chris@82
|
490 {
|
Chris@82
|
491 V T3f, T3o, T3p, T3q;
|
Chris@82
|
492 T3f = VFNMS(LDK(KP707106781), T3e, T2Z);
|
Chris@82
|
493 T3o = VFNMS(LDK(KP707106781), T3n, T3m);
|
Chris@82
|
494 T7r = VFNMSI(T3o, T3f);
|
Chris@82
|
495 STM2(&(xo[48]), T7r, ovs, &(xo[0]));
|
Chris@82
|
496 T7s = VFMAI(T3o, T3f);
|
Chris@82
|
497 STM2(&(xo[80]), T7s, ovs, &(xo[0]));
|
Chris@82
|
498 T3p = VFMA(LDK(KP707106781), T3e, T2Z);
|
Chris@82
|
499 T3q = VFMA(LDK(KP707106781), T3n, T3m);
|
Chris@82
|
500 T7t = VFNMSI(T3q, T3p);
|
Chris@82
|
501 STM2(&(xo[112]), T7t, ovs, &(xo[0]));
|
Chris@82
|
502 T7u = VFMAI(T3q, T3p);
|
Chris@82
|
503 STM2(&(xo[16]), T7u, ovs, &(xo[0]));
|
Chris@82
|
504 }
|
Chris@82
|
505 }
|
Chris@82
|
506 {
|
Chris@82
|
507 V T7v, T7x, T7z, T7C;
|
Chris@82
|
508 {
|
Chris@82
|
509 V T3D, T3V, T3O, T3Y, T3K, T3Z, T3R, T3W, T3C, T3N;
|
Chris@82
|
510 T3C = VADD(T3A, T3B);
|
Chris@82
|
511 T3D = VFMA(LDK(KP707106781), T3C, T3z);
|
Chris@82
|
512 T3V = VFNMS(LDK(KP707106781), T3C, T3z);
|
Chris@82
|
513 T3N = VSUB(T3B, T3A);
|
Chris@82
|
514 T3O = VFMA(LDK(KP707106781), T3N, T3M);
|
Chris@82
|
515 T3Y = VFNMS(LDK(KP707106781), T3N, T3M);
|
Chris@82
|
516 {
|
Chris@82
|
517 V T3G, T3J, T3P, T3Q;
|
Chris@82
|
518 T3G = VFNMS(LDK(KP414213562), T3F, T3E);
|
Chris@82
|
519 T3J = VFNMS(LDK(KP414213562), T3I, T3H);
|
Chris@82
|
520 T3K = VADD(T3G, T3J);
|
Chris@82
|
521 T3Z = VSUB(T3J, T3G);
|
Chris@82
|
522 T3P = VFMA(LDK(KP414213562), T3H, T3I);
|
Chris@82
|
523 T3Q = VFMA(LDK(KP414213562), T3E, T3F);
|
Chris@82
|
524 T3R = VSUB(T3P, T3Q);
|
Chris@82
|
525 T3W = VADD(T3Q, T3P);
|
Chris@82
|
526 }
|
Chris@82
|
527 {
|
Chris@82
|
528 V T3L, T3S, T41, T42;
|
Chris@82
|
529 T3L = VFNMS(LDK(KP923879532), T3K, T3D);
|
Chris@82
|
530 T3S = VFNMS(LDK(KP923879532), T3R, T3O);
|
Chris@82
|
531 T7v = VFNMSI(T3S, T3L);
|
Chris@82
|
532 STM2(&(xo[56]), T7v, ovs, &(xo[0]));
|
Chris@82
|
533 T7w = VFMAI(T3S, T3L);
|
Chris@82
|
534 STM2(&(xo[72]), T7w, ovs, &(xo[0]));
|
Chris@82
|
535 T41 = VFMA(LDK(KP923879532), T3W, T3V);
|
Chris@82
|
536 T42 = VFNMS(LDK(KP923879532), T3Z, T3Y);
|
Chris@82
|
537 T7x = VFNMSI(T42, T41);
|
Chris@82
|
538 STM2(&(xo[24]), T7x, ovs, &(xo[0]));
|
Chris@82
|
539 T7y = VFMAI(T42, T41);
|
Chris@82
|
540 STM2(&(xo[104]), T7y, ovs, &(xo[0]));
|
Chris@82
|
541 }
|
Chris@82
|
542 {
|
Chris@82
|
543 V T3T, T3U, T3X, T40;
|
Chris@82
|
544 T3T = VFMA(LDK(KP923879532), T3K, T3D);
|
Chris@82
|
545 T3U = VFMA(LDK(KP923879532), T3R, T3O);
|
Chris@82
|
546 T7z = VFNMSI(T3U, T3T);
|
Chris@82
|
547 STM2(&(xo[120]), T7z, ovs, &(xo[0]));
|
Chris@82
|
548 T7A = VFMAI(T3U, T3T);
|
Chris@82
|
549 STM2(&(xo[8]), T7A, ovs, &(xo[0]));
|
Chris@82
|
550 T3X = VFNMS(LDK(KP923879532), T3W, T3V);
|
Chris@82
|
551 T40 = VFMA(LDK(KP923879532), T3Z, T3Y);
|
Chris@82
|
552 T7B = VFMAI(T40, T3X);
|
Chris@82
|
553 STM2(&(xo[40]), T7B, ovs, &(xo[0]));
|
Chris@82
|
554 T7C = VFNMSI(T40, T3X);
|
Chris@82
|
555 STM2(&(xo[88]), T7C, ovs, &(xo[0]));
|
Chris@82
|
556 }
|
Chris@82
|
557 }
|
Chris@82
|
558 {
|
Chris@82
|
559 V T6X, T7f, T7b, T7g, T74, T7j, T78, T7i;
|
Chris@82
|
560 {
|
Chris@82
|
561 V T6V, T6W, T79, T7a;
|
Chris@82
|
562 T6V = VFMA(LDK(KP923879532), T6a, T69);
|
Chris@82
|
563 T6W = VADD(T6E, T6D);
|
Chris@82
|
564 T6X = VFMA(LDK(KP831469612), T6W, T6V);
|
Chris@82
|
565 T7f = VFNMS(LDK(KP831469612), T6W, T6V);
|
Chris@82
|
566 T79 = VFNMS(LDK(KP303346683), T6Y, T6Z);
|
Chris@82
|
567 T7a = VFNMS(LDK(KP303346683), T71, T72);
|
Chris@82
|
568 T7b = VSUB(T79, T7a);
|
Chris@82
|
569 T7g = VADD(T79, T7a);
|
Chris@82
|
570 }
|
Chris@82
|
571 {
|
Chris@82
|
572 V T70, T73, T76, T77;
|
Chris@82
|
573 T70 = VFMA(LDK(KP303346683), T6Z, T6Y);
|
Chris@82
|
574 T73 = VFMA(LDK(KP303346683), T72, T71);
|
Chris@82
|
575 T74 = VADD(T70, T73);
|
Chris@82
|
576 T7j = VSUB(T73, T70);
|
Chris@82
|
577 T76 = VFMA(LDK(KP923879532), T6B, T6A);
|
Chris@82
|
578 T77 = VSUB(T6e, T6h);
|
Chris@82
|
579 T78 = VFMA(LDK(KP831469612), T77, T76);
|
Chris@82
|
580 T7i = VFNMS(LDK(KP831469612), T77, T76);
|
Chris@82
|
581 }
|
Chris@82
|
582 {
|
Chris@82
|
583 V T75, T7c, T7D, T7l, T7m, T7F;
|
Chris@82
|
584 T75 = VFNMS(LDK(KP956940335), T74, T6X);
|
Chris@82
|
585 T7c = VFNMS(LDK(KP956940335), T7b, T78);
|
Chris@82
|
586 T7D = VFNMSI(T7c, T75);
|
Chris@82
|
587 STM2(&(xo[58]), T7D, ovs, &(xo[2]));
|
Chris@82
|
588 STN2(&(xo[56]), T7v, T7D, ovs);
|
Chris@82
|
589 T7E = VFMAI(T7c, T75);
|
Chris@82
|
590 STM2(&(xo[70]), T7E, ovs, &(xo[2]));
|
Chris@82
|
591 T7l = VFNMS(LDK(KP956940335), T7g, T7f);
|
Chris@82
|
592 T7m = VFNMS(LDK(KP956940335), T7j, T7i);
|
Chris@82
|
593 T7F = VFNMSI(T7m, T7l);
|
Chris@82
|
594 STM2(&(xo[26]), T7F, ovs, &(xo[2]));
|
Chris@82
|
595 STN2(&(xo[24]), T7x, T7F, ovs);
|
Chris@82
|
596 T7G = VFMAI(T7m, T7l);
|
Chris@82
|
597 STM2(&(xo[102]), T7G, ovs, &(xo[2]));
|
Chris@82
|
598 }
|
Chris@82
|
599 {
|
Chris@82
|
600 V T7d, T7e, T7H, T7h, T7k, T7K;
|
Chris@82
|
601 T7d = VFMA(LDK(KP956940335), T74, T6X);
|
Chris@82
|
602 T7e = VFMA(LDK(KP956940335), T7b, T78);
|
Chris@82
|
603 T7H = VFNMSI(T7e, T7d);
|
Chris@82
|
604 STM2(&(xo[122]), T7H, ovs, &(xo[2]));
|
Chris@82
|
605 STN2(&(xo[120]), T7z, T7H, ovs);
|
Chris@82
|
606 T7I = VFMAI(T7e, T7d);
|
Chris@82
|
607 STM2(&(xo[6]), T7I, ovs, &(xo[2]));
|
Chris@82
|
608 T7h = VFMA(LDK(KP956940335), T7g, T7f);
|
Chris@82
|
609 T7k = VFMA(LDK(KP956940335), T7j, T7i);
|
Chris@82
|
610 T7J = VFMAI(T7k, T7h);
|
Chris@82
|
611 STM2(&(xo[38]), T7J, ovs, &(xo[2]));
|
Chris@82
|
612 T7K = VFNMSI(T7k, T7h);
|
Chris@82
|
613 STM2(&(xo[90]), T7K, ovs, &(xo[2]));
|
Chris@82
|
614 STN2(&(xo[88]), T7C, T7K, ovs);
|
Chris@82
|
615 }
|
Chris@82
|
616 }
|
Chris@82
|
617 }
|
Chris@82
|
618 {
|
Chris@82
|
619 V T7L, T7N, T7P, T7S;
|
Chris@82
|
620 {
|
Chris@82
|
621 V TT, T2j, T2f, T2k, T1Y, T2n, T2c, T2m;
|
Chris@82
|
622 {
|
Chris@82
|
623 V Tn, TS, T2d, T2e;
|
Chris@82
|
624 Tn = VFMA(LDK(KP707106781), Tm, T7);
|
Chris@82
|
625 TS = VADD(TC, TR);
|
Chris@82
|
626 TT = VFMA(LDK(KP923879532), TS, Tn);
|
Chris@82
|
627 T2j = VFNMS(LDK(KP923879532), TS, Tn);
|
Chris@82
|
628 T2d = VFMA(LDK(KP198912367), T1N, T1W);
|
Chris@82
|
629 T2e = VFMA(LDK(KP198912367), T1g, T1p);
|
Chris@82
|
630 T2f = VSUB(T2d, T2e);
|
Chris@82
|
631 T2k = VADD(T2e, T2d);
|
Chris@82
|
632 }
|
Chris@82
|
633 {
|
Chris@82
|
634 V T1q, T1X, T28, T2b;
|
Chris@82
|
635 T1q = VFNMS(LDK(KP198912367), T1p, T1g);
|
Chris@82
|
636 T1X = VFNMS(LDK(KP198912367), T1W, T1N);
|
Chris@82
|
637 T1Y = VADD(T1q, T1X);
|
Chris@82
|
638 T2n = VSUB(T1X, T1q);
|
Chris@82
|
639 T28 = VFMA(LDK(KP707106781), T27, T26);
|
Chris@82
|
640 T2b = VSUB(T29, T2a);
|
Chris@82
|
641 T2c = VFMA(LDK(KP923879532), T2b, T28);
|
Chris@82
|
642 T2m = VFNMS(LDK(KP923879532), T2b, T28);
|
Chris@82
|
643 }
|
Chris@82
|
644 {
|
Chris@82
|
645 V T1Z, T2g, T7M, T2p, T2q, T7O;
|
Chris@82
|
646 T1Z = VFNMS(LDK(KP980785280), T1Y, TT);
|
Chris@82
|
647 T2g = VFNMS(LDK(KP980785280), T2f, T2c);
|
Chris@82
|
648 T7L = VFNMSI(T2g, T1Z);
|
Chris@82
|
649 STM2(&(xo[60]), T7L, ovs, &(xo[0]));
|
Chris@82
|
650 T7M = VFMAI(T2g, T1Z);
|
Chris@82
|
651 STM2(&(xo[68]), T7M, ovs, &(xo[0]));
|
Chris@82
|
652 STN2(&(xo[68]), T7M, T7E, ovs);
|
Chris@82
|
653 T2p = VFMA(LDK(KP980785280), T2k, T2j);
|
Chris@82
|
654 T2q = VFNMS(LDK(KP980785280), T2n, T2m);
|
Chris@82
|
655 T7N = VFNMSI(T2q, T2p);
|
Chris@82
|
656 STM2(&(xo[28]), T7N, ovs, &(xo[0]));
|
Chris@82
|
657 T7O = VFMAI(T2q, T2p);
|
Chris@82
|
658 STM2(&(xo[100]), T7O, ovs, &(xo[0]));
|
Chris@82
|
659 STN2(&(xo[100]), T7O, T7G, ovs);
|
Chris@82
|
660 }
|
Chris@82
|
661 {
|
Chris@82
|
662 V T2h, T2i, T7Q, T2l, T2o, T7R;
|
Chris@82
|
663 T2h = VFMA(LDK(KP980785280), T1Y, TT);
|
Chris@82
|
664 T2i = VFMA(LDK(KP980785280), T2f, T2c);
|
Chris@82
|
665 T7P = VFNMSI(T2i, T2h);
|
Chris@82
|
666 STM2(&(xo[124]), T7P, ovs, &(xo[0]));
|
Chris@82
|
667 T7Q = VFMAI(T2i, T2h);
|
Chris@82
|
668 STM2(&(xo[4]), T7Q, ovs, &(xo[0]));
|
Chris@82
|
669 STN2(&(xo[4]), T7Q, T7I, ovs);
|
Chris@82
|
670 T2l = VFNMS(LDK(KP980785280), T2k, T2j);
|
Chris@82
|
671 T2o = VFMA(LDK(KP980785280), T2n, T2m);
|
Chris@82
|
672 T7R = VFMAI(T2o, T2l);
|
Chris@82
|
673 STM2(&(xo[36]), T7R, ovs, &(xo[0]));
|
Chris@82
|
674 STN2(&(xo[36]), T7R, T7J, ovs);
|
Chris@82
|
675 T7S = VFNMSI(T2o, T2l);
|
Chris@82
|
676 STM2(&(xo[92]), T7S, ovs, &(xo[0]));
|
Chris@82
|
677 }
|
Chris@82
|
678 }
|
Chris@82
|
679 {
|
Chris@82
|
680 V T4z, T5z, T5v, T5A, T5g, T5D, T5s, T5C;
|
Chris@82
|
681 {
|
Chris@82
|
682 V T4f, T4y, T5t, T5u;
|
Chris@82
|
683 T4f = VFMA(LDK(KP923879532), T4e, T47);
|
Chris@82
|
684 T4y = VADD(T4o, T4x);
|
Chris@82
|
685 T4z = VFMA(LDK(KP980785280), T4y, T4f);
|
Chris@82
|
686 T5z = VFNMS(LDK(KP980785280), T4y, T4f);
|
Chris@82
|
687 T5t = VFMA(LDK(KP098491403), T4M, T4T);
|
Chris@82
|
688 T5u = VFMA(LDK(KP098491403), T57, T5e);
|
Chris@82
|
689 T5v = VSUB(T5t, T5u);
|
Chris@82
|
690 T5A = VADD(T5t, T5u);
|
Chris@82
|
691 }
|
Chris@82
|
692 {
|
Chris@82
|
693 V T4U, T5f, T5o, T5r;
|
Chris@82
|
694 T4U = VFNMS(LDK(KP098491403), T4T, T4M);
|
Chris@82
|
695 T5f = VFNMS(LDK(KP098491403), T5e, T57);
|
Chris@82
|
696 T5g = VADD(T4U, T5f);
|
Chris@82
|
697 T5D = VSUB(T5f, T4U);
|
Chris@82
|
698 T5o = VFMA(LDK(KP923879532), T5n, T5k);
|
Chris@82
|
699 T5r = VSUB(T5p, T5q);
|
Chris@82
|
700 T5s = VFMA(LDK(KP980785280), T5r, T5o);
|
Chris@82
|
701 T5C = VFNMS(LDK(KP980785280), T5r, T5o);
|
Chris@82
|
702 }
|
Chris@82
|
703 {
|
Chris@82
|
704 V T5h, T5w, T7T, T7U;
|
Chris@82
|
705 T5h = VFNMS(LDK(KP995184726), T5g, T4z);
|
Chris@82
|
706 T5w = VFNMS(LDK(KP995184726), T5v, T5s);
|
Chris@82
|
707 T7T = VFNMSI(T5w, T5h);
|
Chris@82
|
708 STM2(&(xo[66]), T7T, ovs, &(xo[2]));
|
Chris@82
|
709 STN2(&(xo[64]), T7n, T7T, ovs);
|
Chris@82
|
710 T7U = VFMAI(T5w, T5h);
|
Chris@82
|
711 STM2(&(xo[62]), T7U, ovs, &(xo[2]));
|
Chris@82
|
712 STN2(&(xo[60]), T7L, T7U, ovs);
|
Chris@82
|
713 }
|
Chris@82
|
714 {
|
Chris@82
|
715 V T5F, T5G, T7V, T7W;
|
Chris@82
|
716 T5F = VFMA(LDK(KP995184726), T5A, T5z);
|
Chris@82
|
717 T5G = VFMA(LDK(KP995184726), T5D, T5C);
|
Chris@82
|
718 T7V = VFMAI(T5G, T5F);
|
Chris@82
|
719 STM2(&(xo[30]), T7V, ovs, &(xo[2]));
|
Chris@82
|
720 STN2(&(xo[28]), T7N, T7V, ovs);
|
Chris@82
|
721 T7W = VFNMSI(T5G, T5F);
|
Chris@82
|
722 STM2(&(xo[98]), T7W, ovs, &(xo[2]));
|
Chris@82
|
723 STN2(&(xo[96]), T7q, T7W, ovs);
|
Chris@82
|
724 }
|
Chris@82
|
725 {
|
Chris@82
|
726 V T5x, T5y, T7X, T7Y;
|
Chris@82
|
727 T5x = VFMA(LDK(KP995184726), T5g, T4z);
|
Chris@82
|
728 T5y = VFMA(LDK(KP995184726), T5v, T5s);
|
Chris@82
|
729 T7X = VFNMSI(T5y, T5x);
|
Chris@82
|
730 STM2(&(xo[2]), T7X, ovs, &(xo[2]));
|
Chris@82
|
731 STN2(&(xo[0]), T7p, T7X, ovs);
|
Chris@82
|
732 T7Y = VFMAI(T5y, T5x);
|
Chris@82
|
733 STM2(&(xo[126]), T7Y, ovs, &(xo[2]));
|
Chris@82
|
734 STN2(&(xo[124]), T7P, T7Y, ovs);
|
Chris@82
|
735 }
|
Chris@82
|
736 {
|
Chris@82
|
737 V T5B, T5E, T7Z, T80;
|
Chris@82
|
738 T5B = VFNMS(LDK(KP995184726), T5A, T5z);
|
Chris@82
|
739 T5E = VFNMS(LDK(KP995184726), T5D, T5C);
|
Chris@82
|
740 T7Z = VFNMSI(T5E, T5B);
|
Chris@82
|
741 STM2(&(xo[34]), T7Z, ovs, &(xo[2]));
|
Chris@82
|
742 STN2(&(xo[32]), T7o, T7Z, ovs);
|
Chris@82
|
743 T80 = VFMAI(T5E, T5B);
|
Chris@82
|
744 STM2(&(xo[94]), T80, ovs, &(xo[2]));
|
Chris@82
|
745 STN2(&(xo[92]), T7S, T80, ovs);
|
Chris@82
|
746 }
|
Chris@82
|
747 }
|
Chris@82
|
748 }
|
Chris@82
|
749 {
|
Chris@82
|
750 V T82, T83, T86, T88;
|
Chris@82
|
751 {
|
Chris@82
|
752 V T6j, T6N, T6J, T6O, T6y, T6R, T6G, T6Q;
|
Chris@82
|
753 {
|
Chris@82
|
754 V T6b, T6i, T6H, T6I;
|
Chris@82
|
755 T6b = VFNMS(LDK(KP923879532), T6a, T69);
|
Chris@82
|
756 T6i = VADD(T6e, T6h);
|
Chris@82
|
757 T6j = VFNMS(LDK(KP831469612), T6i, T6b);
|
Chris@82
|
758 T6N = VFMA(LDK(KP831469612), T6i, T6b);
|
Chris@82
|
759 T6H = VFMA(LDK(KP534511135), T6m, T6p);
|
Chris@82
|
760 T6I = VFMA(LDK(KP534511135), T6t, T6w);
|
Chris@82
|
761 T6J = VSUB(T6H, T6I);
|
Chris@82
|
762 T6O = VADD(T6H, T6I);
|
Chris@82
|
763 }
|
Chris@82
|
764 {
|
Chris@82
|
765 V T6q, T6x, T6C, T6F;
|
Chris@82
|
766 T6q = VFNMS(LDK(KP534511135), T6p, T6m);
|
Chris@82
|
767 T6x = VFNMS(LDK(KP534511135), T6w, T6t);
|
Chris@82
|
768 T6y = VADD(T6q, T6x);
|
Chris@82
|
769 T6R = VSUB(T6x, T6q);
|
Chris@82
|
770 T6C = VFNMS(LDK(KP923879532), T6B, T6A);
|
Chris@82
|
771 T6F = VSUB(T6D, T6E);
|
Chris@82
|
772 T6G = VFNMS(LDK(KP831469612), T6F, T6C);
|
Chris@82
|
773 T6Q = VFMA(LDK(KP831469612), T6F, T6C);
|
Chris@82
|
774 }
|
Chris@82
|
775 {
|
Chris@82
|
776 V T6z, T6K, T81, T6T, T6U, T84;
|
Chris@82
|
777 T6z = VFNMS(LDK(KP881921264), T6y, T6j);
|
Chris@82
|
778 T6K = VFNMS(LDK(KP881921264), T6J, T6G);
|
Chris@82
|
779 T81 = VFNMSI(T6K, T6z);
|
Chris@82
|
780 STM2(&(xo[74]), T81, ovs, &(xo[2]));
|
Chris@82
|
781 STN2(&(xo[72]), T7w, T81, ovs);
|
Chris@82
|
782 T82 = VFMAI(T6K, T6z);
|
Chris@82
|
783 STM2(&(xo[54]), T82, ovs, &(xo[2]));
|
Chris@82
|
784 T6T = VFMA(LDK(KP881921264), T6O, T6N);
|
Chris@82
|
785 T6U = VFMA(LDK(KP881921264), T6R, T6Q);
|
Chris@82
|
786 T83 = VFMAI(T6U, T6T);
|
Chris@82
|
787 STM2(&(xo[22]), T83, ovs, &(xo[2]));
|
Chris@82
|
788 T84 = VFNMSI(T6U, T6T);
|
Chris@82
|
789 STM2(&(xo[106]), T84, ovs, &(xo[2]));
|
Chris@82
|
790 STN2(&(xo[104]), T7y, T84, ovs);
|
Chris@82
|
791 }
|
Chris@82
|
792 {
|
Chris@82
|
793 V T6L, T6M, T85, T6P, T6S, T87;
|
Chris@82
|
794 T6L = VFMA(LDK(KP881921264), T6y, T6j);
|
Chris@82
|
795 T6M = VFMA(LDK(KP881921264), T6J, T6G);
|
Chris@82
|
796 T85 = VFNMSI(T6M, T6L);
|
Chris@82
|
797 STM2(&(xo[10]), T85, ovs, &(xo[2]));
|
Chris@82
|
798 STN2(&(xo[8]), T7A, T85, ovs);
|
Chris@82
|
799 T86 = VFMAI(T6M, T6L);
|
Chris@82
|
800 STM2(&(xo[118]), T86, ovs, &(xo[2]));
|
Chris@82
|
801 T6P = VFNMS(LDK(KP881921264), T6O, T6N);
|
Chris@82
|
802 T6S = VFNMS(LDK(KP881921264), T6R, T6Q);
|
Chris@82
|
803 T87 = VFNMSI(T6S, T6P);
|
Chris@82
|
804 STM2(&(xo[42]), T87, ovs, &(xo[2]));
|
Chris@82
|
805 STN2(&(xo[40]), T7B, T87, ovs);
|
Chris@82
|
806 T88 = VFMAI(T6S, T6P);
|
Chris@82
|
807 STM2(&(xo[86]), T88, ovs, &(xo[2]));
|
Chris@82
|
808 }
|
Chris@82
|
809 }
|
Chris@82
|
810 {
|
Chris@82
|
811 V T89, T8c, T8d, T8f;
|
Chris@82
|
812 {
|
Chris@82
|
813 V T2t, T2L, T2H, T2M, T2A, T2P, T2E, T2O;
|
Chris@82
|
814 {
|
Chris@82
|
815 V T2r, T2s, T2F, T2G;
|
Chris@82
|
816 T2r = VFNMS(LDK(KP707106781), Tm, T7);
|
Chris@82
|
817 T2s = VADD(T2a, T29);
|
Chris@82
|
818 T2t = VFMA(LDK(KP923879532), T2s, T2r);
|
Chris@82
|
819 T2L = VFNMS(LDK(KP923879532), T2s, T2r);
|
Chris@82
|
820 T2F = VFNMS(LDK(KP668178637), T2x, T2y);
|
Chris@82
|
821 T2G = VFNMS(LDK(KP668178637), T2u, T2v);
|
Chris@82
|
822 T2H = VSUB(T2F, T2G);
|
Chris@82
|
823 T2M = VADD(T2G, T2F);
|
Chris@82
|
824 }
|
Chris@82
|
825 {
|
Chris@82
|
826 V T2w, T2z, T2C, T2D;
|
Chris@82
|
827 T2w = VFMA(LDK(KP668178637), T2v, T2u);
|
Chris@82
|
828 T2z = VFMA(LDK(KP668178637), T2y, T2x);
|
Chris@82
|
829 T2A = VADD(T2w, T2z);
|
Chris@82
|
830 T2P = VSUB(T2z, T2w);
|
Chris@82
|
831 T2C = VFNMS(LDK(KP707106781), T27, T26);
|
Chris@82
|
832 T2D = VSUB(TR, TC);
|
Chris@82
|
833 T2E = VFNMS(LDK(KP923879532), T2D, T2C);
|
Chris@82
|
834 T2O = VFMA(LDK(KP923879532), T2D, T2C);
|
Chris@82
|
835 }
|
Chris@82
|
836 {
|
Chris@82
|
837 V T2B, T2I, T8a, T2R, T2S, T8b;
|
Chris@82
|
838 T2B = VFNMS(LDK(KP831469612), T2A, T2t);
|
Chris@82
|
839 T2I = VFNMS(LDK(KP831469612), T2H, T2E);
|
Chris@82
|
840 T89 = VFNMSI(T2I, T2B);
|
Chris@82
|
841 STM2(&(xo[76]), T89, ovs, &(xo[0]));
|
Chris@82
|
842 T8a = VFMAI(T2I, T2B);
|
Chris@82
|
843 STM2(&(xo[52]), T8a, ovs, &(xo[0]));
|
Chris@82
|
844 STN2(&(xo[52]), T8a, T82, ovs);
|
Chris@82
|
845 T2R = VFNMS(LDK(KP831469612), T2M, T2L);
|
Chris@82
|
846 T2S = VFMA(LDK(KP831469612), T2P, T2O);
|
Chris@82
|
847 T8b = VFMAI(T2S, T2R);
|
Chris@82
|
848 STM2(&(xo[20]), T8b, ovs, &(xo[0]));
|
Chris@82
|
849 STN2(&(xo[20]), T8b, T83, ovs);
|
Chris@82
|
850 T8c = VFNMSI(T2S, T2R);
|
Chris@82
|
851 STM2(&(xo[108]), T8c, ovs, &(xo[0]));
|
Chris@82
|
852 }
|
Chris@82
|
853 {
|
Chris@82
|
854 V T2J, T2K, T8e, T2N, T2Q, T8g;
|
Chris@82
|
855 T2J = VFMA(LDK(KP831469612), T2A, T2t);
|
Chris@82
|
856 T2K = VFMA(LDK(KP831469612), T2H, T2E);
|
Chris@82
|
857 T8d = VFNMSI(T2K, T2J);
|
Chris@82
|
858 STM2(&(xo[12]), T8d, ovs, &(xo[0]));
|
Chris@82
|
859 T8e = VFMAI(T2K, T2J);
|
Chris@82
|
860 STM2(&(xo[116]), T8e, ovs, &(xo[0]));
|
Chris@82
|
861 STN2(&(xo[116]), T8e, T86, ovs);
|
Chris@82
|
862 T2N = VFMA(LDK(KP831469612), T2M, T2L);
|
Chris@82
|
863 T2Q = VFNMS(LDK(KP831469612), T2P, T2O);
|
Chris@82
|
864 T8f = VFNMSI(T2Q, T2N);
|
Chris@82
|
865 STM2(&(xo[44]), T8f, ovs, &(xo[0]));
|
Chris@82
|
866 T8g = VFMAI(T2Q, T2N);
|
Chris@82
|
867 STM2(&(xo[84]), T8g, ovs, &(xo[0]));
|
Chris@82
|
868 STN2(&(xo[84]), T8g, T88, ovs);
|
Chris@82
|
869 }
|
Chris@82
|
870 }
|
Chris@82
|
871 {
|
Chris@82
|
872 V T5J, T61, T5X, T62, T5Q, T65, T5U, T64;
|
Chris@82
|
873 {
|
Chris@82
|
874 V T5H, T5I, T5V, T5W;
|
Chris@82
|
875 T5H = VFNMS(LDK(KP923879532), T4e, T47);
|
Chris@82
|
876 T5I = VADD(T5p, T5q);
|
Chris@82
|
877 T5J = VFMA(LDK(KP980785280), T5I, T5H);
|
Chris@82
|
878 T61 = VFNMS(LDK(KP980785280), T5I, T5H);
|
Chris@82
|
879 T5V = VFNMS(LDK(KP820678790), T5K, T5L);
|
Chris@82
|
880 T5W = VFNMS(LDK(KP820678790), T5N, T5O);
|
Chris@82
|
881 T5X = VSUB(T5V, T5W);
|
Chris@82
|
882 T62 = VADD(T5V, T5W);
|
Chris@82
|
883 }
|
Chris@82
|
884 {
|
Chris@82
|
885 V T5M, T5P, T5S, T5T;
|
Chris@82
|
886 T5M = VFMA(LDK(KP820678790), T5L, T5K);
|
Chris@82
|
887 T5P = VFMA(LDK(KP820678790), T5O, T5N);
|
Chris@82
|
888 T5Q = VADD(T5M, T5P);
|
Chris@82
|
889 T65 = VSUB(T5P, T5M);
|
Chris@82
|
890 T5S = VFNMS(LDK(KP923879532), T5n, T5k);
|
Chris@82
|
891 T5T = VSUB(T4x, T4o);
|
Chris@82
|
892 T5U = VFMA(LDK(KP980785280), T5T, T5S);
|
Chris@82
|
893 T64 = VFNMS(LDK(KP980785280), T5T, T5S);
|
Chris@82
|
894 }
|
Chris@82
|
895 {
|
Chris@82
|
896 V T5R, T5Y, T8h, T8i;
|
Chris@82
|
897 T5R = VFNMS(LDK(KP773010453), T5Q, T5J);
|
Chris@82
|
898 T5Y = VFNMS(LDK(KP773010453), T5X, T5U);
|
Chris@82
|
899 T8h = VFNMSI(T5Y, T5R);
|
Chris@82
|
900 STM2(&(xo[50]), T8h, ovs, &(xo[2]));
|
Chris@82
|
901 STN2(&(xo[48]), T7r, T8h, ovs);
|
Chris@82
|
902 T8i = VFMAI(T5Y, T5R);
|
Chris@82
|
903 STM2(&(xo[78]), T8i, ovs, &(xo[2]));
|
Chris@82
|
904 STN2(&(xo[76]), T89, T8i, ovs);
|
Chris@82
|
905 }
|
Chris@82
|
906 {
|
Chris@82
|
907 V T67, T68, T8j, T8k;
|
Chris@82
|
908 T67 = VFNMS(LDK(KP773010453), T62, T61);
|
Chris@82
|
909 T68 = VFNMS(LDK(KP773010453), T65, T64);
|
Chris@82
|
910 T8j = VFNMSI(T68, T67);
|
Chris@82
|
911 STM2(&(xo[18]), T8j, ovs, &(xo[2]));
|
Chris@82
|
912 STN2(&(xo[16]), T7u, T8j, ovs);
|
Chris@82
|
913 T8k = VFMAI(T68, T67);
|
Chris@82
|
914 STM2(&(xo[110]), T8k, ovs, &(xo[2]));
|
Chris@82
|
915 STN2(&(xo[108]), T8c, T8k, ovs);
|
Chris@82
|
916 }
|
Chris@82
|
917 {
|
Chris@82
|
918 V T5Z, T60, T8l, T8m;
|
Chris@82
|
919 T5Z = VFMA(LDK(KP773010453), T5Q, T5J);
|
Chris@82
|
920 T60 = VFMA(LDK(KP773010453), T5X, T5U);
|
Chris@82
|
921 T8l = VFNMSI(T60, T5Z);
|
Chris@82
|
922 STM2(&(xo[114]), T8l, ovs, &(xo[2]));
|
Chris@82
|
923 STN2(&(xo[112]), T7t, T8l, ovs);
|
Chris@82
|
924 T8m = VFMAI(T60, T5Z);
|
Chris@82
|
925 STM2(&(xo[14]), T8m, ovs, &(xo[2]));
|
Chris@82
|
926 STN2(&(xo[12]), T8d, T8m, ovs);
|
Chris@82
|
927 }
|
Chris@82
|
928 {
|
Chris@82
|
929 V T63, T66, T8n, T8o;
|
Chris@82
|
930 T63 = VFMA(LDK(KP773010453), T62, T61);
|
Chris@82
|
931 T66 = VFMA(LDK(KP773010453), T65, T64);
|
Chris@82
|
932 T8n = VFMAI(T66, T63);
|
Chris@82
|
933 STM2(&(xo[46]), T8n, ovs, &(xo[2]));
|
Chris@82
|
934 STN2(&(xo[44]), T8f, T8n, ovs);
|
Chris@82
|
935 T8o = VFNMSI(T66, T63);
|
Chris@82
|
936 STM2(&(xo[82]), T8o, ovs, &(xo[2]));
|
Chris@82
|
937 STN2(&(xo[80]), T7s, T8o, ovs);
|
Chris@82
|
938 }
|
Chris@82
|
939 }
|
Chris@82
|
940 }
|
Chris@82
|
941 }
|
Chris@82
|
942 }
|
Chris@82
|
943 }
|
Chris@82
|
944 }
|
Chris@82
|
945 VLEAVE();
|
Chris@82
|
946 }
|
Chris@82
|
947
|
Chris@82
|
948 static const kdft_desc desc = { 64, XSIMD_STRING("n2fv_64"), {198, 0, 258, 0}, &GENUS, 0, 2, 0, 0 };
|
Chris@82
|
949
|
Chris@82
|
950 void XSIMD(codelet_n2fv_64) (planner *p) {
|
Chris@82
|
951 X(kdft_register) (p, n2fv_64, &desc);
|
Chris@82
|
952 }
|
Chris@82
|
953
|
Chris@82
|
954 #else
|
Chris@82
|
955
|
Chris@82
|
956 /* Generated by: ../../../genfft/gen_notw_c.native -simd -compact -variables 4 -pipeline-latency 8 -n 64 -name n2fv_64 -with-ostride 2 -include dft/simd/n2f.h -store-multiple 2 */
|
Chris@82
|
957
|
Chris@82
|
958 /*
|
Chris@82
|
959 * This function contains 456 FP additions, 124 FP multiplications,
|
Chris@82
|
960 * (or, 404 additions, 72 multiplications, 52 fused multiply/add),
|
Chris@82
|
961 * 128 stack variables, 15 constants, and 160 memory accesses
|
Chris@82
|
962 */
|
Chris@82
|
963 #include "dft/simd/n2f.h"
|
Chris@82
|
964
|
Chris@82
|
965 static void n2fv_64(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
|
Chris@82
|
966 {
|
Chris@82
|
967 DVK(KP471396736, +0.471396736825997648556387625905254377657460319);
|
Chris@82
|
968 DVK(KP881921264, +0.881921264348355029712756863660388349508442621);
|
Chris@82
|
969 DVK(KP098017140, +0.098017140329560601994195563888641845861136673);
|
Chris@82
|
970 DVK(KP995184726, +0.995184726672196886244836953109479921575474869);
|
Chris@82
|
971 DVK(KP290284677, +0.290284677254462367636192375817395274691476278);
|
Chris@82
|
972 DVK(KP956940335, +0.956940335732208864935797886980269969482849206);
|
Chris@82
|
973 DVK(KP634393284, +0.634393284163645498215171613225493370675687095);
|
Chris@82
|
974 DVK(KP773010453, +0.773010453362736960810906609758469800971041293);
|
Chris@82
|
975 DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
|
Chris@82
|
976 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
Chris@82
|
977 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
Chris@82
|
978 DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
|
Chris@82
|
979 DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
Chris@82
|
980 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
Chris@82
|
981 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
Chris@82
|
982 {
|
Chris@82
|
983 INT i;
|
Chris@82
|
984 const R *xi;
|
Chris@82
|
985 R *xo;
|
Chris@82
|
986 xi = ri;
|
Chris@82
|
987 xo = ro;
|
Chris@82
|
988 for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(128, is), MAKE_VOLATILE_STRIDE(128, os)) {
|
Chris@82
|
989 V T4p, T5q, Tb, T39, T2n, T3A, T6f, T6T, Tq, T3B, T6i, T76, T2i, T3a, T4w;
|
Chris@82
|
990 V T5r, TI, T2p, T6C, T6V, T3h, T3E, T4L, T5u, TZ, T2q, T6F, T6U, T3e, T3D;
|
Chris@82
|
991 V T4E, T5t, T23, T2N, T6t, T71, T6w, T72, T2c, T2O, T3t, T41, T5f, T5R, T5k;
|
Chris@82
|
992 V T5S, T3w, T42, T1s, T2K, T6m, T6Y, T6p, T6Z, T1B, T2L, T3m, T3Y, T4Y, T5O;
|
Chris@82
|
993 V T53, T5P, T3p, T3Z;
|
Chris@82
|
994 {
|
Chris@82
|
995 V T3, T4n, T2m, T4o, T6, T5p, T9, T5o;
|
Chris@82
|
996 {
|
Chris@82
|
997 V T1, T2, T2k, T2l;
|
Chris@82
|
998 T1 = LD(&(xi[0]), ivs, &(xi[0]));
|
Chris@82
|
999 T2 = LD(&(xi[WS(is, 32)]), ivs, &(xi[0]));
|
Chris@82
|
1000 T3 = VSUB(T1, T2);
|
Chris@82
|
1001 T4n = VADD(T1, T2);
|
Chris@82
|
1002 T2k = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
|
Chris@82
|
1003 T2l = LD(&(xi[WS(is, 48)]), ivs, &(xi[0]));
|
Chris@82
|
1004 T2m = VSUB(T2k, T2l);
|
Chris@82
|
1005 T4o = VADD(T2k, T2l);
|
Chris@82
|
1006 }
|
Chris@82
|
1007 {
|
Chris@82
|
1008 V T4, T5, T7, T8;
|
Chris@82
|
1009 T4 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
|
Chris@82
|
1010 T5 = LD(&(xi[WS(is, 40)]), ivs, &(xi[0]));
|
Chris@82
|
1011 T6 = VSUB(T4, T5);
|
Chris@82
|
1012 T5p = VADD(T4, T5);
|
Chris@82
|
1013 T7 = LD(&(xi[WS(is, 56)]), ivs, &(xi[0]));
|
Chris@82
|
1014 T8 = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
|
Chris@82
|
1015 T9 = VSUB(T7, T8);
|
Chris@82
|
1016 T5o = VADD(T7, T8);
|
Chris@82
|
1017 }
|
Chris@82
|
1018 T4p = VSUB(T4n, T4o);
|
Chris@82
|
1019 T5q = VSUB(T5o, T5p);
|
Chris@82
|
1020 {
|
Chris@82
|
1021 V Ta, T2j, T6d, T6e;
|
Chris@82
|
1022 Ta = VMUL(LDK(KP707106781), VADD(T6, T9));
|
Chris@82
|
1023 Tb = VADD(T3, Ta);
|
Chris@82
|
1024 T39 = VSUB(T3, Ta);
|
Chris@82
|
1025 T2j = VMUL(LDK(KP707106781), VSUB(T9, T6));
|
Chris@82
|
1026 T2n = VSUB(T2j, T2m);
|
Chris@82
|
1027 T3A = VADD(T2m, T2j);
|
Chris@82
|
1028 T6d = VADD(T4n, T4o);
|
Chris@82
|
1029 T6e = VADD(T5p, T5o);
|
Chris@82
|
1030 T6f = VADD(T6d, T6e);
|
Chris@82
|
1031 T6T = VSUB(T6d, T6e);
|
Chris@82
|
1032 }
|
Chris@82
|
1033 }
|
Chris@82
|
1034 {
|
Chris@82
|
1035 V Te, T4q, To, T4u, Th, T4r, Tl, T4t;
|
Chris@82
|
1036 {
|
Chris@82
|
1037 V Tc, Td, Tm, Tn;
|
Chris@82
|
1038 Tc = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
|
Chris@82
|
1039 Td = LD(&(xi[WS(is, 36)]), ivs, &(xi[0]));
|
Chris@82
|
1040 Te = VSUB(Tc, Td);
|
Chris@82
|
1041 T4q = VADD(Tc, Td);
|
Chris@82
|
1042 Tm = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
|
Chris@82
|
1043 Tn = LD(&(xi[WS(is, 44)]), ivs, &(xi[0]));
|
Chris@82
|
1044 To = VSUB(Tm, Tn);
|
Chris@82
|
1045 T4u = VADD(Tm, Tn);
|
Chris@82
|
1046 }
|
Chris@82
|
1047 {
|
Chris@82
|
1048 V Tf, Tg, Tj, Tk;
|
Chris@82
|
1049 Tf = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
|
Chris@82
|
1050 Tg = LD(&(xi[WS(is, 52)]), ivs, &(xi[0]));
|
Chris@82
|
1051 Th = VSUB(Tf, Tg);
|
Chris@82
|
1052 T4r = VADD(Tf, Tg);
|
Chris@82
|
1053 Tj = LD(&(xi[WS(is, 60)]), ivs, &(xi[0]));
|
Chris@82
|
1054 Tk = LD(&(xi[WS(is, 28)]), ivs, &(xi[0]));
|
Chris@82
|
1055 Tl = VSUB(Tj, Tk);
|
Chris@82
|
1056 T4t = VADD(Tj, Tk);
|
Chris@82
|
1057 }
|
Chris@82
|
1058 {
|
Chris@82
|
1059 V Ti, Tp, T6g, T6h;
|
Chris@82
|
1060 Ti = VFNMS(LDK(KP382683432), Th, VMUL(LDK(KP923879532), Te));
|
Chris@82
|
1061 Tp = VFMA(LDK(KP923879532), Tl, VMUL(LDK(KP382683432), To));
|
Chris@82
|
1062 Tq = VADD(Ti, Tp);
|
Chris@82
|
1063 T3B = VSUB(Tp, Ti);
|
Chris@82
|
1064 T6g = VADD(T4q, T4r);
|
Chris@82
|
1065 T6h = VADD(T4t, T4u);
|
Chris@82
|
1066 T6i = VADD(T6g, T6h);
|
Chris@82
|
1067 T76 = VSUB(T6h, T6g);
|
Chris@82
|
1068 }
|
Chris@82
|
1069 {
|
Chris@82
|
1070 V T2g, T2h, T4s, T4v;
|
Chris@82
|
1071 T2g = VFNMS(LDK(KP923879532), To, VMUL(LDK(KP382683432), Tl));
|
Chris@82
|
1072 T2h = VFMA(LDK(KP382683432), Te, VMUL(LDK(KP923879532), Th));
|
Chris@82
|
1073 T2i = VSUB(T2g, T2h);
|
Chris@82
|
1074 T3a = VADD(T2h, T2g);
|
Chris@82
|
1075 T4s = VSUB(T4q, T4r);
|
Chris@82
|
1076 T4v = VSUB(T4t, T4u);
|
Chris@82
|
1077 T4w = VMUL(LDK(KP707106781), VADD(T4s, T4v));
|
Chris@82
|
1078 T5r = VMUL(LDK(KP707106781), VSUB(T4v, T4s));
|
Chris@82
|
1079 }
|
Chris@82
|
1080 }
|
Chris@82
|
1081 {
|
Chris@82
|
1082 V Tu, T4F, TG, T4G, TB, T4J, TD, T4I;
|
Chris@82
|
1083 {
|
Chris@82
|
1084 V Ts, Tt, TE, TF;
|
Chris@82
|
1085 Ts = LD(&(xi[WS(is, 62)]), ivs, &(xi[0]));
|
Chris@82
|
1086 Tt = LD(&(xi[WS(is, 30)]), ivs, &(xi[0]));
|
Chris@82
|
1087 Tu = VSUB(Ts, Tt);
|
Chris@82
|
1088 T4F = VADD(Ts, Tt);
|
Chris@82
|
1089 TE = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
|
Chris@82
|
1090 TF = LD(&(xi[WS(is, 46)]), ivs, &(xi[0]));
|
Chris@82
|
1091 TG = VSUB(TE, TF);
|
Chris@82
|
1092 T4G = VADD(TE, TF);
|
Chris@82
|
1093 {
|
Chris@82
|
1094 V Tv, Tw, Tx, Ty, Tz, TA;
|
Chris@82
|
1095 Tv = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
|
Chris@82
|
1096 Tw = LD(&(xi[WS(is, 38)]), ivs, &(xi[0]));
|
Chris@82
|
1097 Tx = VSUB(Tv, Tw);
|
Chris@82
|
1098 Ty = LD(&(xi[WS(is, 54)]), ivs, &(xi[0]));
|
Chris@82
|
1099 Tz = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
|
Chris@82
|
1100 TA = VSUB(Ty, Tz);
|
Chris@82
|
1101 TB = VMUL(LDK(KP707106781), VADD(Tx, TA));
|
Chris@82
|
1102 T4J = VADD(Tv, Tw);
|
Chris@82
|
1103 TD = VMUL(LDK(KP707106781), VSUB(TA, Tx));
|
Chris@82
|
1104 T4I = VADD(Ty, Tz);
|
Chris@82
|
1105 }
|
Chris@82
|
1106 }
|
Chris@82
|
1107 {
|
Chris@82
|
1108 V TC, TH, T6A, T6B;
|
Chris@82
|
1109 TC = VADD(Tu, TB);
|
Chris@82
|
1110 TH = VSUB(TD, TG);
|
Chris@82
|
1111 TI = VFMA(LDK(KP195090322), TC, VMUL(LDK(KP980785280), TH));
|
Chris@82
|
1112 T2p = VFNMS(LDK(KP195090322), TH, VMUL(LDK(KP980785280), TC));
|
Chris@82
|
1113 T6A = VADD(T4F, T4G);
|
Chris@82
|
1114 T6B = VADD(T4J, T4I);
|
Chris@82
|
1115 T6C = VADD(T6A, T6B);
|
Chris@82
|
1116 T6V = VSUB(T6A, T6B);
|
Chris@82
|
1117 }
|
Chris@82
|
1118 {
|
Chris@82
|
1119 V T3f, T3g, T4H, T4K;
|
Chris@82
|
1120 T3f = VSUB(Tu, TB);
|
Chris@82
|
1121 T3g = VADD(TG, TD);
|
Chris@82
|
1122 T3h = VFNMS(LDK(KP555570233), T3g, VMUL(LDK(KP831469612), T3f));
|
Chris@82
|
1123 T3E = VFMA(LDK(KP555570233), T3f, VMUL(LDK(KP831469612), T3g));
|
Chris@82
|
1124 T4H = VSUB(T4F, T4G);
|
Chris@82
|
1125 T4K = VSUB(T4I, T4J);
|
Chris@82
|
1126 T4L = VFNMS(LDK(KP382683432), T4K, VMUL(LDK(KP923879532), T4H));
|
Chris@82
|
1127 T5u = VFMA(LDK(KP382683432), T4H, VMUL(LDK(KP923879532), T4K));
|
Chris@82
|
1128 }
|
Chris@82
|
1129 }
|
Chris@82
|
1130 {
|
Chris@82
|
1131 V TS, T4z, TW, T4y, TP, T4C, TX, T4B;
|
Chris@82
|
1132 {
|
Chris@82
|
1133 V TQ, TR, TU, TV;
|
Chris@82
|
1134 TQ = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
|
Chris@82
|
1135 TR = LD(&(xi[WS(is, 50)]), ivs, &(xi[0]));
|
Chris@82
|
1136 TS = VSUB(TQ, TR);
|
Chris@82
|
1137 T4z = VADD(TQ, TR);
|
Chris@82
|
1138 TU = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
|
Chris@82
|
1139 TV = LD(&(xi[WS(is, 34)]), ivs, &(xi[0]));
|
Chris@82
|
1140 TW = VSUB(TU, TV);
|
Chris@82
|
1141 T4y = VADD(TU, TV);
|
Chris@82
|
1142 {
|
Chris@82
|
1143 V TJ, TK, TL, TM, TN, TO;
|
Chris@82
|
1144 TJ = LD(&(xi[WS(is, 58)]), ivs, &(xi[0]));
|
Chris@82
|
1145 TK = LD(&(xi[WS(is, 26)]), ivs, &(xi[0]));
|
Chris@82
|
1146 TL = VSUB(TJ, TK);
|
Chris@82
|
1147 TM = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
|
Chris@82
|
1148 TN = LD(&(xi[WS(is, 42)]), ivs, &(xi[0]));
|
Chris@82
|
1149 TO = VSUB(TM, TN);
|
Chris@82
|
1150 TP = VMUL(LDK(KP707106781), VSUB(TL, TO));
|
Chris@82
|
1151 T4C = VADD(TM, TN);
|
Chris@82
|
1152 TX = VMUL(LDK(KP707106781), VADD(TO, TL));
|
Chris@82
|
1153 T4B = VADD(TJ, TK);
|
Chris@82
|
1154 }
|
Chris@82
|
1155 }
|
Chris@82
|
1156 {
|
Chris@82
|
1157 V TT, TY, T6D, T6E;
|
Chris@82
|
1158 TT = VSUB(TP, TS);
|
Chris@82
|
1159 TY = VADD(TW, TX);
|
Chris@82
|
1160 TZ = VFNMS(LDK(KP195090322), TY, VMUL(LDK(KP980785280), TT));
|
Chris@82
|
1161 T2q = VFMA(LDK(KP980785280), TY, VMUL(LDK(KP195090322), TT));
|
Chris@82
|
1162 T6D = VADD(T4y, T4z);
|
Chris@82
|
1163 T6E = VADD(T4C, T4B);
|
Chris@82
|
1164 T6F = VADD(T6D, T6E);
|
Chris@82
|
1165 T6U = VSUB(T6D, T6E);
|
Chris@82
|
1166 }
|
Chris@82
|
1167 {
|
Chris@82
|
1168 V T3c, T3d, T4A, T4D;
|
Chris@82
|
1169 T3c = VSUB(TW, TX);
|
Chris@82
|
1170 T3d = VADD(TS, TP);
|
Chris@82
|
1171 T3e = VFMA(LDK(KP831469612), T3c, VMUL(LDK(KP555570233), T3d));
|
Chris@82
|
1172 T3D = VFNMS(LDK(KP555570233), T3c, VMUL(LDK(KP831469612), T3d));
|
Chris@82
|
1173 T4A = VSUB(T4y, T4z);
|
Chris@82
|
1174 T4D = VSUB(T4B, T4C);
|
Chris@82
|
1175 T4E = VFMA(LDK(KP923879532), T4A, VMUL(LDK(KP382683432), T4D));
|
Chris@82
|
1176 T5t = VFNMS(LDK(KP382683432), T4A, VMUL(LDK(KP923879532), T4D));
|
Chris@82
|
1177 }
|
Chris@82
|
1178 }
|
Chris@82
|
1179 {
|
Chris@82
|
1180 V T1F, T55, T2a, T56, T1M, T5h, T27, T5g, T58, T59, T1U, T5a, T25, T5b, T5c;
|
Chris@82
|
1181 V T21, T5d, T24;
|
Chris@82
|
1182 {
|
Chris@82
|
1183 V T1D, T1E, T28, T29;
|
Chris@82
|
1184 T1D = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1185 T1E = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1186 T1F = VSUB(T1D, T1E);
|
Chris@82
|
1187 T55 = VADD(T1D, T1E);
|
Chris@82
|
1188 T28 = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1189 T29 = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1190 T2a = VSUB(T28, T29);
|
Chris@82
|
1191 T56 = VADD(T28, T29);
|
Chris@82
|
1192 }
|
Chris@82
|
1193 {
|
Chris@82
|
1194 V T1G, T1H, T1I, T1J, T1K, T1L;
|
Chris@82
|
1195 T1G = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1196 T1H = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1197 T1I = VSUB(T1G, T1H);
|
Chris@82
|
1198 T1J = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1199 T1K = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1200 T1L = VSUB(T1J, T1K);
|
Chris@82
|
1201 T1M = VMUL(LDK(KP707106781), VADD(T1I, T1L));
|
Chris@82
|
1202 T5h = VADD(T1G, T1H);
|
Chris@82
|
1203 T27 = VMUL(LDK(KP707106781), VSUB(T1L, T1I));
|
Chris@82
|
1204 T5g = VADD(T1J, T1K);
|
Chris@82
|
1205 }
|
Chris@82
|
1206 {
|
Chris@82
|
1207 V T1Q, T1T, T1X, T20;
|
Chris@82
|
1208 {
|
Chris@82
|
1209 V T1O, T1P, T1R, T1S;
|
Chris@82
|
1210 T1O = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1211 T1P = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1212 T1Q = VSUB(T1O, T1P);
|
Chris@82
|
1213 T58 = VADD(T1O, T1P);
|
Chris@82
|
1214 T1R = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1215 T1S = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1216 T1T = VSUB(T1R, T1S);
|
Chris@82
|
1217 T59 = VADD(T1R, T1S);
|
Chris@82
|
1218 }
|
Chris@82
|
1219 T1U = VFNMS(LDK(KP382683432), T1T, VMUL(LDK(KP923879532), T1Q));
|
Chris@82
|
1220 T5a = VSUB(T58, T59);
|
Chris@82
|
1221 T25 = VFMA(LDK(KP382683432), T1Q, VMUL(LDK(KP923879532), T1T));
|
Chris@82
|
1222 {
|
Chris@82
|
1223 V T1V, T1W, T1Y, T1Z;
|
Chris@82
|
1224 T1V = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1225 T1W = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1226 T1X = VSUB(T1V, T1W);
|
Chris@82
|
1227 T5b = VADD(T1V, T1W);
|
Chris@82
|
1228 T1Y = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1229 T1Z = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1230 T20 = VSUB(T1Y, T1Z);
|
Chris@82
|
1231 T5c = VADD(T1Y, T1Z);
|
Chris@82
|
1232 }
|
Chris@82
|
1233 T21 = VFMA(LDK(KP923879532), T1X, VMUL(LDK(KP382683432), T20));
|
Chris@82
|
1234 T5d = VSUB(T5b, T5c);
|
Chris@82
|
1235 T24 = VFNMS(LDK(KP923879532), T20, VMUL(LDK(KP382683432), T1X));
|
Chris@82
|
1236 }
|
Chris@82
|
1237 {
|
Chris@82
|
1238 V T1N, T22, T6r, T6s;
|
Chris@82
|
1239 T1N = VADD(T1F, T1M);
|
Chris@82
|
1240 T22 = VADD(T1U, T21);
|
Chris@82
|
1241 T23 = VSUB(T1N, T22);
|
Chris@82
|
1242 T2N = VADD(T1N, T22);
|
Chris@82
|
1243 T6r = VADD(T55, T56);
|
Chris@82
|
1244 T6s = VADD(T5h, T5g);
|
Chris@82
|
1245 T6t = VADD(T6r, T6s);
|
Chris@82
|
1246 T71 = VSUB(T6r, T6s);
|
Chris@82
|
1247 }
|
Chris@82
|
1248 {
|
Chris@82
|
1249 V T6u, T6v, T26, T2b;
|
Chris@82
|
1250 T6u = VADD(T58, T59);
|
Chris@82
|
1251 T6v = VADD(T5b, T5c);
|
Chris@82
|
1252 T6w = VADD(T6u, T6v);
|
Chris@82
|
1253 T72 = VSUB(T6v, T6u);
|
Chris@82
|
1254 T26 = VSUB(T24, T25);
|
Chris@82
|
1255 T2b = VSUB(T27, T2a);
|
Chris@82
|
1256 T2c = VSUB(T26, T2b);
|
Chris@82
|
1257 T2O = VADD(T2b, T26);
|
Chris@82
|
1258 }
|
Chris@82
|
1259 {
|
Chris@82
|
1260 V T3r, T3s, T57, T5e;
|
Chris@82
|
1261 T3r = VSUB(T1F, T1M);
|
Chris@82
|
1262 T3s = VADD(T25, T24);
|
Chris@82
|
1263 T3t = VADD(T3r, T3s);
|
Chris@82
|
1264 T41 = VSUB(T3r, T3s);
|
Chris@82
|
1265 T57 = VSUB(T55, T56);
|
Chris@82
|
1266 T5e = VMUL(LDK(KP707106781), VADD(T5a, T5d));
|
Chris@82
|
1267 T5f = VADD(T57, T5e);
|
Chris@82
|
1268 T5R = VSUB(T57, T5e);
|
Chris@82
|
1269 }
|
Chris@82
|
1270 {
|
Chris@82
|
1271 V T5i, T5j, T3u, T3v;
|
Chris@82
|
1272 T5i = VSUB(T5g, T5h);
|
Chris@82
|
1273 T5j = VMUL(LDK(KP707106781), VSUB(T5d, T5a));
|
Chris@82
|
1274 T5k = VADD(T5i, T5j);
|
Chris@82
|
1275 T5S = VSUB(T5j, T5i);
|
Chris@82
|
1276 T3u = VADD(T2a, T27);
|
Chris@82
|
1277 T3v = VSUB(T21, T1U);
|
Chris@82
|
1278 T3w = VADD(T3u, T3v);
|
Chris@82
|
1279 T42 = VSUB(T3v, T3u);
|
Chris@82
|
1280 }
|
Chris@82
|
1281 }
|
Chris@82
|
1282 {
|
Chris@82
|
1283 V T1q, T4P, T1v, T4O, T1n, T50, T1w, T4Z, T4U, T4V, T18, T4W, T1z, T4R, T4S;
|
Chris@82
|
1284 V T1f, T4T, T1y;
|
Chris@82
|
1285 {
|
Chris@82
|
1286 V T1o, T1p, T1t, T1u;
|
Chris@82
|
1287 T1o = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1288 T1p = LD(&(xi[WS(is, 49)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1289 T1q = VSUB(T1o, T1p);
|
Chris@82
|
1290 T4P = VADD(T1o, T1p);
|
Chris@82
|
1291 T1t = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1292 T1u = LD(&(xi[WS(is, 33)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1293 T1v = VSUB(T1t, T1u);
|
Chris@82
|
1294 T4O = VADD(T1t, T1u);
|
Chris@82
|
1295 }
|
Chris@82
|
1296 {
|
Chris@82
|
1297 V T1h, T1i, T1j, T1k, T1l, T1m;
|
Chris@82
|
1298 T1h = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1299 T1i = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1300 T1j = VSUB(T1h, T1i);
|
Chris@82
|
1301 T1k = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1302 T1l = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1303 T1m = VSUB(T1k, T1l);
|
Chris@82
|
1304 T1n = VMUL(LDK(KP707106781), VSUB(T1j, T1m));
|
Chris@82
|
1305 T50 = VADD(T1k, T1l);
|
Chris@82
|
1306 T1w = VMUL(LDK(KP707106781), VADD(T1m, T1j));
|
Chris@82
|
1307 T4Z = VADD(T1h, T1i);
|
Chris@82
|
1308 }
|
Chris@82
|
1309 {
|
Chris@82
|
1310 V T14, T17, T1b, T1e;
|
Chris@82
|
1311 {
|
Chris@82
|
1312 V T12, T13, T15, T16;
|
Chris@82
|
1313 T12 = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1314 T13 = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1315 T14 = VSUB(T12, T13);
|
Chris@82
|
1316 T4U = VADD(T12, T13);
|
Chris@82
|
1317 T15 = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1318 T16 = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1319 T17 = VSUB(T15, T16);
|
Chris@82
|
1320 T4V = VADD(T15, T16);
|
Chris@82
|
1321 }
|
Chris@82
|
1322 T18 = VFNMS(LDK(KP923879532), T17, VMUL(LDK(KP382683432), T14));
|
Chris@82
|
1323 T4W = VSUB(T4U, T4V);
|
Chris@82
|
1324 T1z = VFMA(LDK(KP923879532), T14, VMUL(LDK(KP382683432), T17));
|
Chris@82
|
1325 {
|
Chris@82
|
1326 V T19, T1a, T1c, T1d;
|
Chris@82
|
1327 T19 = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1328 T1a = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1329 T1b = VSUB(T19, T1a);
|
Chris@82
|
1330 T4R = VADD(T19, T1a);
|
Chris@82
|
1331 T1c = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1332 T1d = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)]));
|
Chris@82
|
1333 T1e = VSUB(T1c, T1d);
|
Chris@82
|
1334 T4S = VADD(T1c, T1d);
|
Chris@82
|
1335 }
|
Chris@82
|
1336 T1f = VFMA(LDK(KP382683432), T1b, VMUL(LDK(KP923879532), T1e));
|
Chris@82
|
1337 T4T = VSUB(T4R, T4S);
|
Chris@82
|
1338 T1y = VFNMS(LDK(KP382683432), T1e, VMUL(LDK(KP923879532), T1b));
|
Chris@82
|
1339 }
|
Chris@82
|
1340 {
|
Chris@82
|
1341 V T1g, T1r, T6k, T6l;
|
Chris@82
|
1342 T1g = VSUB(T18, T1f);
|
Chris@82
|
1343 T1r = VSUB(T1n, T1q);
|
Chris@82
|
1344 T1s = VSUB(T1g, T1r);
|
Chris@82
|
1345 T2K = VADD(T1r, T1g);
|
Chris@82
|
1346 T6k = VADD(T4O, T4P);
|
Chris@82
|
1347 T6l = VADD(T50, T4Z);
|
Chris@82
|
1348 T6m = VADD(T6k, T6l);
|
Chris@82
|
1349 T6Y = VSUB(T6k, T6l);
|
Chris@82
|
1350 }
|
Chris@82
|
1351 {
|
Chris@82
|
1352 V T6n, T6o, T1x, T1A;
|
Chris@82
|
1353 T6n = VADD(T4R, T4S);
|
Chris@82
|
1354 T6o = VADD(T4U, T4V);
|
Chris@82
|
1355 T6p = VADD(T6n, T6o);
|
Chris@82
|
1356 T6Z = VSUB(T6o, T6n);
|
Chris@82
|
1357 T1x = VADD(T1v, T1w);
|
Chris@82
|
1358 T1A = VADD(T1y, T1z);
|
Chris@82
|
1359 T1B = VSUB(T1x, T1A);
|
Chris@82
|
1360 T2L = VADD(T1x, T1A);
|
Chris@82
|
1361 }
|
Chris@82
|
1362 {
|
Chris@82
|
1363 V T3k, T3l, T4Q, T4X;
|
Chris@82
|
1364 T3k = VSUB(T1v, T1w);
|
Chris@82
|
1365 T3l = VADD(T1f, T18);
|
Chris@82
|
1366 T3m = VADD(T3k, T3l);
|
Chris@82
|
1367 T3Y = VSUB(T3k, T3l);
|
Chris@82
|
1368 T4Q = VSUB(T4O, T4P);
|
Chris@82
|
1369 T4X = VMUL(LDK(KP707106781), VADD(T4T, T4W));
|
Chris@82
|
1370 T4Y = VADD(T4Q, T4X);
|
Chris@82
|
1371 T5O = VSUB(T4Q, T4X);
|
Chris@82
|
1372 }
|
Chris@82
|
1373 {
|
Chris@82
|
1374 V T51, T52, T3n, T3o;
|
Chris@82
|
1375 T51 = VSUB(T4Z, T50);
|
Chris@82
|
1376 T52 = VMUL(LDK(KP707106781), VSUB(T4W, T4T));
|
Chris@82
|
1377 T53 = VADD(T51, T52);
|
Chris@82
|
1378 T5P = VSUB(T52, T51);
|
Chris@82
|
1379 T3n = VADD(T1q, T1n);
|
Chris@82
|
1380 T3o = VSUB(T1z, T1y);
|
Chris@82
|
1381 T3p = VADD(T3n, T3o);
|
Chris@82
|
1382 T3Z = VSUB(T3o, T3n);
|
Chris@82
|
1383 }
|
Chris@82
|
1384 }
|
Chris@82
|
1385 {
|
Chris@82
|
1386 V T7n, T7o, T7p, T7q, T7r, T7s, T7t, T7u, T7v, T7w, T7x, T7y, T7z, T7A, T7B;
|
Chris@82
|
1387 V T7C, T7D, T7E, T7F, T7G, T7H, T7I, T7J, T7K;
|
Chris@82
|
1388 {
|
Chris@82
|
1389 V T6N, T6R, T6Q, T6S;
|
Chris@82
|
1390 {
|
Chris@82
|
1391 V T6L, T6M, T6O, T6P;
|
Chris@82
|
1392 T6L = VADD(T6f, T6i);
|
Chris@82
|
1393 T6M = VADD(T6F, T6C);
|
Chris@82
|
1394 T6N = VADD(T6L, T6M);
|
Chris@82
|
1395 T6R = VSUB(T6L, T6M);
|
Chris@82
|
1396 T6O = VADD(T6m, T6p);
|
Chris@82
|
1397 T6P = VADD(T6t, T6w);
|
Chris@82
|
1398 T6Q = VADD(T6O, T6P);
|
Chris@82
|
1399 T6S = VBYI(VSUB(T6P, T6O));
|
Chris@82
|
1400 }
|
Chris@82
|
1401 T7n = VSUB(T6N, T6Q);
|
Chris@82
|
1402 STM2(&(xo[64]), T7n, ovs, &(xo[0]));
|
Chris@82
|
1403 T7o = VADD(T6R, T6S);
|
Chris@82
|
1404 STM2(&(xo[32]), T7o, ovs, &(xo[0]));
|
Chris@82
|
1405 T7p = VADD(T6N, T6Q);
|
Chris@82
|
1406 STM2(&(xo[0]), T7p, ovs, &(xo[0]));
|
Chris@82
|
1407 T7q = VSUB(T6R, T6S);
|
Chris@82
|
1408 STM2(&(xo[96]), T7q, ovs, &(xo[0]));
|
Chris@82
|
1409 }
|
Chris@82
|
1410 {
|
Chris@82
|
1411 V T6j, T6G, T6y, T6H, T6q, T6x;
|
Chris@82
|
1412 T6j = VSUB(T6f, T6i);
|
Chris@82
|
1413 T6G = VSUB(T6C, T6F);
|
Chris@82
|
1414 T6q = VSUB(T6m, T6p);
|
Chris@82
|
1415 T6x = VSUB(T6t, T6w);
|
Chris@82
|
1416 T6y = VMUL(LDK(KP707106781), VADD(T6q, T6x));
|
Chris@82
|
1417 T6H = VMUL(LDK(KP707106781), VSUB(T6x, T6q));
|
Chris@82
|
1418 {
|
Chris@82
|
1419 V T6z, T6I, T6J, T6K;
|
Chris@82
|
1420 T6z = VADD(T6j, T6y);
|
Chris@82
|
1421 T6I = VBYI(VADD(T6G, T6H));
|
Chris@82
|
1422 T7r = VSUB(T6z, T6I);
|
Chris@82
|
1423 STM2(&(xo[112]), T7r, ovs, &(xo[0]));
|
Chris@82
|
1424 T7s = VADD(T6z, T6I);
|
Chris@82
|
1425 STM2(&(xo[16]), T7s, ovs, &(xo[0]));
|
Chris@82
|
1426 T6J = VSUB(T6j, T6y);
|
Chris@82
|
1427 T6K = VBYI(VSUB(T6H, T6G));
|
Chris@82
|
1428 T7t = VSUB(T6J, T6K);
|
Chris@82
|
1429 STM2(&(xo[80]), T7t, ovs, &(xo[0]));
|
Chris@82
|
1430 T7u = VADD(T6J, T6K);
|
Chris@82
|
1431 STM2(&(xo[48]), T7u, ovs, &(xo[0]));
|
Chris@82
|
1432 }
|
Chris@82
|
1433 }
|
Chris@82
|
1434 {
|
Chris@82
|
1435 V T6X, T7i, T78, T7g, T74, T7f, T7b, T7j, T6W, T77;
|
Chris@82
|
1436 T6W = VMUL(LDK(KP707106781), VADD(T6U, T6V));
|
Chris@82
|
1437 T6X = VADD(T6T, T6W);
|
Chris@82
|
1438 T7i = VSUB(T6T, T6W);
|
Chris@82
|
1439 T77 = VMUL(LDK(KP707106781), VSUB(T6V, T6U));
|
Chris@82
|
1440 T78 = VADD(T76, T77);
|
Chris@82
|
1441 T7g = VSUB(T77, T76);
|
Chris@82
|
1442 {
|
Chris@82
|
1443 V T70, T73, T79, T7a;
|
Chris@82
|
1444 T70 = VFMA(LDK(KP923879532), T6Y, VMUL(LDK(KP382683432), T6Z));
|
Chris@82
|
1445 T73 = VFNMS(LDK(KP382683432), T72, VMUL(LDK(KP923879532), T71));
|
Chris@82
|
1446 T74 = VADD(T70, T73);
|
Chris@82
|
1447 T7f = VSUB(T73, T70);
|
Chris@82
|
1448 T79 = VFNMS(LDK(KP382683432), T6Y, VMUL(LDK(KP923879532), T6Z));
|
Chris@82
|
1449 T7a = VFMA(LDK(KP382683432), T71, VMUL(LDK(KP923879532), T72));
|
Chris@82
|
1450 T7b = VADD(T79, T7a);
|
Chris@82
|
1451 T7j = VSUB(T7a, T79);
|
Chris@82
|
1452 }
|
Chris@82
|
1453 {
|
Chris@82
|
1454 V T75, T7c, T7l, T7m;
|
Chris@82
|
1455 T75 = VADD(T6X, T74);
|
Chris@82
|
1456 T7c = VBYI(VADD(T78, T7b));
|
Chris@82
|
1457 T7v = VSUB(T75, T7c);
|
Chris@82
|
1458 STM2(&(xo[120]), T7v, ovs, &(xo[0]));
|
Chris@82
|
1459 T7w = VADD(T75, T7c);
|
Chris@82
|
1460 STM2(&(xo[8]), T7w, ovs, &(xo[0]));
|
Chris@82
|
1461 T7l = VBYI(VADD(T7g, T7f));
|
Chris@82
|
1462 T7m = VADD(T7i, T7j);
|
Chris@82
|
1463 T7x = VADD(T7l, T7m);
|
Chris@82
|
1464 STM2(&(xo[24]), T7x, ovs, &(xo[0]));
|
Chris@82
|
1465 T7y = VSUB(T7m, T7l);
|
Chris@82
|
1466 STM2(&(xo[104]), T7y, ovs, &(xo[0]));
|
Chris@82
|
1467 }
|
Chris@82
|
1468 {
|
Chris@82
|
1469 V T7d, T7e, T7h, T7k;
|
Chris@82
|
1470 T7d = VSUB(T6X, T74);
|
Chris@82
|
1471 T7e = VBYI(VSUB(T7b, T78));
|
Chris@82
|
1472 T7z = VSUB(T7d, T7e);
|
Chris@82
|
1473 STM2(&(xo[72]), T7z, ovs, &(xo[0]));
|
Chris@82
|
1474 T7A = VADD(T7d, T7e);
|
Chris@82
|
1475 STM2(&(xo[56]), T7A, ovs, &(xo[0]));
|
Chris@82
|
1476 T7h = VBYI(VSUB(T7f, T7g));
|
Chris@82
|
1477 T7k = VSUB(T7i, T7j);
|
Chris@82
|
1478 T7B = VADD(T7h, T7k);
|
Chris@82
|
1479 STM2(&(xo[40]), T7B, ovs, &(xo[0]));
|
Chris@82
|
1480 T7C = VSUB(T7k, T7h);
|
Chris@82
|
1481 STM2(&(xo[88]), T7C, ovs, &(xo[0]));
|
Chris@82
|
1482 }
|
Chris@82
|
1483 }
|
Chris@82
|
1484 {
|
Chris@82
|
1485 V T5N, T68, T61, T69, T5U, T65, T5Y, T66;
|
Chris@82
|
1486 {
|
Chris@82
|
1487 V T5L, T5M, T5Z, T60;
|
Chris@82
|
1488 T5L = VSUB(T4p, T4w);
|
Chris@82
|
1489 T5M = VSUB(T5u, T5t);
|
Chris@82
|
1490 T5N = VADD(T5L, T5M);
|
Chris@82
|
1491 T68 = VSUB(T5L, T5M);
|
Chris@82
|
1492 T5Z = VFNMS(LDK(KP555570233), T5O, VMUL(LDK(KP831469612), T5P));
|
Chris@82
|
1493 T60 = VFMA(LDK(KP555570233), T5R, VMUL(LDK(KP831469612), T5S));
|
Chris@82
|
1494 T61 = VADD(T5Z, T60);
|
Chris@82
|
1495 T69 = VSUB(T60, T5Z);
|
Chris@82
|
1496 }
|
Chris@82
|
1497 {
|
Chris@82
|
1498 V T5Q, T5T, T5W, T5X;
|
Chris@82
|
1499 T5Q = VFMA(LDK(KP831469612), T5O, VMUL(LDK(KP555570233), T5P));
|
Chris@82
|
1500 T5T = VFNMS(LDK(KP555570233), T5S, VMUL(LDK(KP831469612), T5R));
|
Chris@82
|
1501 T5U = VADD(T5Q, T5T);
|
Chris@82
|
1502 T65 = VSUB(T5T, T5Q);
|
Chris@82
|
1503 T5W = VSUB(T5r, T5q);
|
Chris@82
|
1504 T5X = VSUB(T4L, T4E);
|
Chris@82
|
1505 T5Y = VADD(T5W, T5X);
|
Chris@82
|
1506 T66 = VSUB(T5X, T5W);
|
Chris@82
|
1507 }
|
Chris@82
|
1508 {
|
Chris@82
|
1509 V T5V, T62, T6b, T6c;
|
Chris@82
|
1510 T5V = VADD(T5N, T5U);
|
Chris@82
|
1511 T62 = VBYI(VADD(T5Y, T61));
|
Chris@82
|
1512 T7D = VSUB(T5V, T62);
|
Chris@82
|
1513 STM2(&(xo[116]), T7D, ovs, &(xo[0]));
|
Chris@82
|
1514 T7E = VADD(T5V, T62);
|
Chris@82
|
1515 STM2(&(xo[12]), T7E, ovs, &(xo[0]));
|
Chris@82
|
1516 T6b = VBYI(VADD(T66, T65));
|
Chris@82
|
1517 T6c = VADD(T68, T69);
|
Chris@82
|
1518 T7F = VADD(T6b, T6c);
|
Chris@82
|
1519 STM2(&(xo[20]), T7F, ovs, &(xo[0]));
|
Chris@82
|
1520 T7G = VSUB(T6c, T6b);
|
Chris@82
|
1521 STM2(&(xo[108]), T7G, ovs, &(xo[0]));
|
Chris@82
|
1522 }
|
Chris@82
|
1523 {
|
Chris@82
|
1524 V T63, T64, T67, T6a;
|
Chris@82
|
1525 T63 = VSUB(T5N, T5U);
|
Chris@82
|
1526 T64 = VBYI(VSUB(T61, T5Y));
|
Chris@82
|
1527 T7H = VSUB(T63, T64);
|
Chris@82
|
1528 STM2(&(xo[76]), T7H, ovs, &(xo[0]));
|
Chris@82
|
1529 T7I = VADD(T63, T64);
|
Chris@82
|
1530 STM2(&(xo[52]), T7I, ovs, &(xo[0]));
|
Chris@82
|
1531 T67 = VBYI(VSUB(T65, T66));
|
Chris@82
|
1532 T6a = VSUB(T68, T69);
|
Chris@82
|
1533 T7J = VADD(T67, T6a);
|
Chris@82
|
1534 STM2(&(xo[44]), T7J, ovs, &(xo[0]));
|
Chris@82
|
1535 T7K = VSUB(T6a, T67);
|
Chris@82
|
1536 STM2(&(xo[84]), T7K, ovs, &(xo[0]));
|
Chris@82
|
1537 }
|
Chris@82
|
1538 }
|
Chris@82
|
1539 {
|
Chris@82
|
1540 V T7U, T7W, T7X, T7Z;
|
Chris@82
|
1541 {
|
Chris@82
|
1542 V T11, T2C, T2v, T2D, T2e, T2z, T2s, T2A;
|
Chris@82
|
1543 {
|
Chris@82
|
1544 V Tr, T10, T2t, T2u;
|
Chris@82
|
1545 Tr = VSUB(Tb, Tq);
|
Chris@82
|
1546 T10 = VSUB(TI, TZ);
|
Chris@82
|
1547 T11 = VADD(Tr, T10);
|
Chris@82
|
1548 T2C = VSUB(Tr, T10);
|
Chris@82
|
1549 T2t = VFNMS(LDK(KP634393284), T1B, VMUL(LDK(KP773010453), T1s));
|
Chris@82
|
1550 T2u = VFMA(LDK(KP773010453), T2c, VMUL(LDK(KP634393284), T23));
|
Chris@82
|
1551 T2v = VADD(T2t, T2u);
|
Chris@82
|
1552 T2D = VSUB(T2u, T2t);
|
Chris@82
|
1553 }
|
Chris@82
|
1554 {
|
Chris@82
|
1555 V T1C, T2d, T2o, T2r;
|
Chris@82
|
1556 T1C = VFMA(LDK(KP634393284), T1s, VMUL(LDK(KP773010453), T1B));
|
Chris@82
|
1557 T2d = VFNMS(LDK(KP634393284), T2c, VMUL(LDK(KP773010453), T23));
|
Chris@82
|
1558 T2e = VADD(T1C, T2d);
|
Chris@82
|
1559 T2z = VSUB(T2d, T1C);
|
Chris@82
|
1560 T2o = VSUB(T2i, T2n);
|
Chris@82
|
1561 T2r = VSUB(T2p, T2q);
|
Chris@82
|
1562 T2s = VADD(T2o, T2r);
|
Chris@82
|
1563 T2A = VSUB(T2r, T2o);
|
Chris@82
|
1564 }
|
Chris@82
|
1565 {
|
Chris@82
|
1566 V T2f, T2w, T7L, T7M;
|
Chris@82
|
1567 T2f = VADD(T11, T2e);
|
Chris@82
|
1568 T2w = VBYI(VADD(T2s, T2v));
|
Chris@82
|
1569 T7L = VSUB(T2f, T2w);
|
Chris@82
|
1570 STM2(&(xo[114]), T7L, ovs, &(xo[2]));
|
Chris@82
|
1571 STN2(&(xo[112]), T7r, T7L, ovs);
|
Chris@82
|
1572 T7M = VADD(T2f, T2w);
|
Chris@82
|
1573 STM2(&(xo[14]), T7M, ovs, &(xo[2]));
|
Chris@82
|
1574 STN2(&(xo[12]), T7E, T7M, ovs);
|
Chris@82
|
1575 }
|
Chris@82
|
1576 {
|
Chris@82
|
1577 V T2F, T2G, T7N, T7O;
|
Chris@82
|
1578 T2F = VBYI(VADD(T2A, T2z));
|
Chris@82
|
1579 T2G = VADD(T2C, T2D);
|
Chris@82
|
1580 T7N = VADD(T2F, T2G);
|
Chris@82
|
1581 STM2(&(xo[18]), T7N, ovs, &(xo[2]));
|
Chris@82
|
1582 STN2(&(xo[16]), T7s, T7N, ovs);
|
Chris@82
|
1583 T7O = VSUB(T2G, T2F);
|
Chris@82
|
1584 STM2(&(xo[110]), T7O, ovs, &(xo[2]));
|
Chris@82
|
1585 STN2(&(xo[108]), T7G, T7O, ovs);
|
Chris@82
|
1586 }
|
Chris@82
|
1587 {
|
Chris@82
|
1588 V T2x, T2y, T7P, T7Q;
|
Chris@82
|
1589 T2x = VSUB(T11, T2e);
|
Chris@82
|
1590 T2y = VBYI(VSUB(T2v, T2s));
|
Chris@82
|
1591 T7P = VSUB(T2x, T2y);
|
Chris@82
|
1592 STM2(&(xo[78]), T7P, ovs, &(xo[2]));
|
Chris@82
|
1593 STN2(&(xo[76]), T7H, T7P, ovs);
|
Chris@82
|
1594 T7Q = VADD(T2x, T2y);
|
Chris@82
|
1595 STM2(&(xo[50]), T7Q, ovs, &(xo[2]));
|
Chris@82
|
1596 STN2(&(xo[48]), T7u, T7Q, ovs);
|
Chris@82
|
1597 }
|
Chris@82
|
1598 {
|
Chris@82
|
1599 V T2B, T2E, T7R, T7S;
|
Chris@82
|
1600 T2B = VBYI(VSUB(T2z, T2A));
|
Chris@82
|
1601 T2E = VSUB(T2C, T2D);
|
Chris@82
|
1602 T7R = VADD(T2B, T2E);
|
Chris@82
|
1603 STM2(&(xo[46]), T7R, ovs, &(xo[2]));
|
Chris@82
|
1604 STN2(&(xo[44]), T7J, T7R, ovs);
|
Chris@82
|
1605 T7S = VSUB(T2E, T2B);
|
Chris@82
|
1606 STM2(&(xo[82]), T7S, ovs, &(xo[2]));
|
Chris@82
|
1607 STN2(&(xo[80]), T7t, T7S, ovs);
|
Chris@82
|
1608 }
|
Chris@82
|
1609 }
|
Chris@82
|
1610 {
|
Chris@82
|
1611 V T3j, T3Q, T3J, T3R, T3y, T3N, T3G, T3O;
|
Chris@82
|
1612 {
|
Chris@82
|
1613 V T3b, T3i, T3H, T3I;
|
Chris@82
|
1614 T3b = VADD(T39, T3a);
|
Chris@82
|
1615 T3i = VADD(T3e, T3h);
|
Chris@82
|
1616 T3j = VADD(T3b, T3i);
|
Chris@82
|
1617 T3Q = VSUB(T3b, T3i);
|
Chris@82
|
1618 T3H = VFNMS(LDK(KP290284677), T3m, VMUL(LDK(KP956940335), T3p));
|
Chris@82
|
1619 T3I = VFMA(LDK(KP290284677), T3t, VMUL(LDK(KP956940335), T3w));
|
Chris@82
|
1620 T3J = VADD(T3H, T3I);
|
Chris@82
|
1621 T3R = VSUB(T3I, T3H);
|
Chris@82
|
1622 }
|
Chris@82
|
1623 {
|
Chris@82
|
1624 V T3q, T3x, T3C, T3F;
|
Chris@82
|
1625 T3q = VFMA(LDK(KP956940335), T3m, VMUL(LDK(KP290284677), T3p));
|
Chris@82
|
1626 T3x = VFNMS(LDK(KP290284677), T3w, VMUL(LDK(KP956940335), T3t));
|
Chris@82
|
1627 T3y = VADD(T3q, T3x);
|
Chris@82
|
1628 T3N = VSUB(T3x, T3q);
|
Chris@82
|
1629 T3C = VADD(T3A, T3B);
|
Chris@82
|
1630 T3F = VADD(T3D, T3E);
|
Chris@82
|
1631 T3G = VADD(T3C, T3F);
|
Chris@82
|
1632 T3O = VSUB(T3F, T3C);
|
Chris@82
|
1633 }
|
Chris@82
|
1634 {
|
Chris@82
|
1635 V T3z, T3K, T7T, T3T, T3U, T7V;
|
Chris@82
|
1636 T3z = VADD(T3j, T3y);
|
Chris@82
|
1637 T3K = VBYI(VADD(T3G, T3J));
|
Chris@82
|
1638 T7T = VSUB(T3z, T3K);
|
Chris@82
|
1639 STM2(&(xo[122]), T7T, ovs, &(xo[2]));
|
Chris@82
|
1640 STN2(&(xo[120]), T7v, T7T, ovs);
|
Chris@82
|
1641 T7U = VADD(T3z, T3K);
|
Chris@82
|
1642 STM2(&(xo[6]), T7U, ovs, &(xo[2]));
|
Chris@82
|
1643 T3T = VBYI(VADD(T3O, T3N));
|
Chris@82
|
1644 T3U = VADD(T3Q, T3R);
|
Chris@82
|
1645 T7V = VADD(T3T, T3U);
|
Chris@82
|
1646 STM2(&(xo[26]), T7V, ovs, &(xo[2]));
|
Chris@82
|
1647 STN2(&(xo[24]), T7x, T7V, ovs);
|
Chris@82
|
1648 T7W = VSUB(T3U, T3T);
|
Chris@82
|
1649 STM2(&(xo[102]), T7W, ovs, &(xo[2]));
|
Chris@82
|
1650 }
|
Chris@82
|
1651 {
|
Chris@82
|
1652 V T3L, T3M, T7Y, T3P, T3S, T80;
|
Chris@82
|
1653 T3L = VSUB(T3j, T3y);
|
Chris@82
|
1654 T3M = VBYI(VSUB(T3J, T3G));
|
Chris@82
|
1655 T7X = VSUB(T3L, T3M);
|
Chris@82
|
1656 STM2(&(xo[70]), T7X, ovs, &(xo[2]));
|
Chris@82
|
1657 T7Y = VADD(T3L, T3M);
|
Chris@82
|
1658 STM2(&(xo[58]), T7Y, ovs, &(xo[2]));
|
Chris@82
|
1659 STN2(&(xo[56]), T7A, T7Y, ovs);
|
Chris@82
|
1660 T3P = VBYI(VSUB(T3N, T3O));
|
Chris@82
|
1661 T3S = VSUB(T3Q, T3R);
|
Chris@82
|
1662 T7Z = VADD(T3P, T3S);
|
Chris@82
|
1663 STM2(&(xo[38]), T7Z, ovs, &(xo[2]));
|
Chris@82
|
1664 T80 = VSUB(T3S, T3P);
|
Chris@82
|
1665 STM2(&(xo[90]), T80, ovs, &(xo[2]));
|
Chris@82
|
1666 STN2(&(xo[88]), T7C, T80, ovs);
|
Chris@82
|
1667 }
|
Chris@82
|
1668 }
|
Chris@82
|
1669 {
|
Chris@82
|
1670 V T81, T83, T86, T88;
|
Chris@82
|
1671 {
|
Chris@82
|
1672 V T4N, T5G, T5z, T5H, T5m, T5D, T5w, T5E;
|
Chris@82
|
1673 {
|
Chris@82
|
1674 V T4x, T4M, T5x, T5y;
|
Chris@82
|
1675 T4x = VADD(T4p, T4w);
|
Chris@82
|
1676 T4M = VADD(T4E, T4L);
|
Chris@82
|
1677 T4N = VADD(T4x, T4M);
|
Chris@82
|
1678 T5G = VSUB(T4x, T4M);
|
Chris@82
|
1679 T5x = VFNMS(LDK(KP195090322), T4Y, VMUL(LDK(KP980785280), T53));
|
Chris@82
|
1680 T5y = VFMA(LDK(KP195090322), T5f, VMUL(LDK(KP980785280), T5k));
|
Chris@82
|
1681 T5z = VADD(T5x, T5y);
|
Chris@82
|
1682 T5H = VSUB(T5y, T5x);
|
Chris@82
|
1683 }
|
Chris@82
|
1684 {
|
Chris@82
|
1685 V T54, T5l, T5s, T5v;
|
Chris@82
|
1686 T54 = VFMA(LDK(KP980785280), T4Y, VMUL(LDK(KP195090322), T53));
|
Chris@82
|
1687 T5l = VFNMS(LDK(KP195090322), T5k, VMUL(LDK(KP980785280), T5f));
|
Chris@82
|
1688 T5m = VADD(T54, T5l);
|
Chris@82
|
1689 T5D = VSUB(T5l, T54);
|
Chris@82
|
1690 T5s = VADD(T5q, T5r);
|
Chris@82
|
1691 T5v = VADD(T5t, T5u);
|
Chris@82
|
1692 T5w = VADD(T5s, T5v);
|
Chris@82
|
1693 T5E = VSUB(T5v, T5s);
|
Chris@82
|
1694 }
|
Chris@82
|
1695 {
|
Chris@82
|
1696 V T5n, T5A, T82, T5J, T5K, T84;
|
Chris@82
|
1697 T5n = VADD(T4N, T5m);
|
Chris@82
|
1698 T5A = VBYI(VADD(T5w, T5z));
|
Chris@82
|
1699 T81 = VSUB(T5n, T5A);
|
Chris@82
|
1700 STM2(&(xo[124]), T81, ovs, &(xo[0]));
|
Chris@82
|
1701 T82 = VADD(T5n, T5A);
|
Chris@82
|
1702 STM2(&(xo[4]), T82, ovs, &(xo[0]));
|
Chris@82
|
1703 STN2(&(xo[4]), T82, T7U, ovs);
|
Chris@82
|
1704 T5J = VBYI(VADD(T5E, T5D));
|
Chris@82
|
1705 T5K = VADD(T5G, T5H);
|
Chris@82
|
1706 T83 = VADD(T5J, T5K);
|
Chris@82
|
1707 STM2(&(xo[28]), T83, ovs, &(xo[0]));
|
Chris@82
|
1708 T84 = VSUB(T5K, T5J);
|
Chris@82
|
1709 STM2(&(xo[100]), T84, ovs, &(xo[0]));
|
Chris@82
|
1710 STN2(&(xo[100]), T84, T7W, ovs);
|
Chris@82
|
1711 }
|
Chris@82
|
1712 {
|
Chris@82
|
1713 V T5B, T5C, T85, T5F, T5I, T87;
|
Chris@82
|
1714 T5B = VSUB(T4N, T5m);
|
Chris@82
|
1715 T5C = VBYI(VSUB(T5z, T5w));
|
Chris@82
|
1716 T85 = VSUB(T5B, T5C);
|
Chris@82
|
1717 STM2(&(xo[68]), T85, ovs, &(xo[0]));
|
Chris@82
|
1718 STN2(&(xo[68]), T85, T7X, ovs);
|
Chris@82
|
1719 T86 = VADD(T5B, T5C);
|
Chris@82
|
1720 STM2(&(xo[60]), T86, ovs, &(xo[0]));
|
Chris@82
|
1721 T5F = VBYI(VSUB(T5D, T5E));
|
Chris@82
|
1722 T5I = VSUB(T5G, T5H);
|
Chris@82
|
1723 T87 = VADD(T5F, T5I);
|
Chris@82
|
1724 STM2(&(xo[36]), T87, ovs, &(xo[0]));
|
Chris@82
|
1725 STN2(&(xo[36]), T87, T7Z, ovs);
|
Chris@82
|
1726 T88 = VSUB(T5I, T5F);
|
Chris@82
|
1727 STM2(&(xo[92]), T88, ovs, &(xo[0]));
|
Chris@82
|
1728 }
|
Chris@82
|
1729 }
|
Chris@82
|
1730 {
|
Chris@82
|
1731 V T2J, T34, T2X, T35, T2Q, T31, T2U, T32;
|
Chris@82
|
1732 {
|
Chris@82
|
1733 V T2H, T2I, T2V, T2W;
|
Chris@82
|
1734 T2H = VADD(Tb, Tq);
|
Chris@82
|
1735 T2I = VADD(T2q, T2p);
|
Chris@82
|
1736 T2J = VADD(T2H, T2I);
|
Chris@82
|
1737 T34 = VSUB(T2H, T2I);
|
Chris@82
|
1738 T2V = VFNMS(LDK(KP098017140), T2L, VMUL(LDK(KP995184726), T2K));
|
Chris@82
|
1739 T2W = VFMA(LDK(KP995184726), T2O, VMUL(LDK(KP098017140), T2N));
|
Chris@82
|
1740 T2X = VADD(T2V, T2W);
|
Chris@82
|
1741 T35 = VSUB(T2W, T2V);
|
Chris@82
|
1742 }
|
Chris@82
|
1743 {
|
Chris@82
|
1744 V T2M, T2P, T2S, T2T;
|
Chris@82
|
1745 T2M = VFMA(LDK(KP098017140), T2K, VMUL(LDK(KP995184726), T2L));
|
Chris@82
|
1746 T2P = VFNMS(LDK(KP098017140), T2O, VMUL(LDK(KP995184726), T2N));
|
Chris@82
|
1747 T2Q = VADD(T2M, T2P);
|
Chris@82
|
1748 T31 = VSUB(T2P, T2M);
|
Chris@82
|
1749 T2S = VADD(T2n, T2i);
|
Chris@82
|
1750 T2T = VADD(TZ, TI);
|
Chris@82
|
1751 T2U = VADD(T2S, T2T);
|
Chris@82
|
1752 T32 = VSUB(T2T, T2S);
|
Chris@82
|
1753 }
|
Chris@82
|
1754 {
|
Chris@82
|
1755 V T2R, T2Y, T89, T8a;
|
Chris@82
|
1756 T2R = VADD(T2J, T2Q);
|
Chris@82
|
1757 T2Y = VBYI(VADD(T2U, T2X));
|
Chris@82
|
1758 T89 = VSUB(T2R, T2Y);
|
Chris@82
|
1759 STM2(&(xo[126]), T89, ovs, &(xo[2]));
|
Chris@82
|
1760 STN2(&(xo[124]), T81, T89, ovs);
|
Chris@82
|
1761 T8a = VADD(T2R, T2Y);
|
Chris@82
|
1762 STM2(&(xo[2]), T8a, ovs, &(xo[2]));
|
Chris@82
|
1763 STN2(&(xo[0]), T7p, T8a, ovs);
|
Chris@82
|
1764 }
|
Chris@82
|
1765 {
|
Chris@82
|
1766 V T37, T38, T8b, T8c;
|
Chris@82
|
1767 T37 = VBYI(VADD(T32, T31));
|
Chris@82
|
1768 T38 = VADD(T34, T35);
|
Chris@82
|
1769 T8b = VADD(T37, T38);
|
Chris@82
|
1770 STM2(&(xo[30]), T8b, ovs, &(xo[2]));
|
Chris@82
|
1771 STN2(&(xo[28]), T83, T8b, ovs);
|
Chris@82
|
1772 T8c = VSUB(T38, T37);
|
Chris@82
|
1773 STM2(&(xo[98]), T8c, ovs, &(xo[2]));
|
Chris@82
|
1774 STN2(&(xo[96]), T7q, T8c, ovs);
|
Chris@82
|
1775 }
|
Chris@82
|
1776 {
|
Chris@82
|
1777 V T2Z, T30, T8d, T8e;
|
Chris@82
|
1778 T2Z = VSUB(T2J, T2Q);
|
Chris@82
|
1779 T30 = VBYI(VSUB(T2X, T2U));
|
Chris@82
|
1780 T8d = VSUB(T2Z, T30);
|
Chris@82
|
1781 STM2(&(xo[66]), T8d, ovs, &(xo[2]));
|
Chris@82
|
1782 STN2(&(xo[64]), T7n, T8d, ovs);
|
Chris@82
|
1783 T8e = VADD(T2Z, T30);
|
Chris@82
|
1784 STM2(&(xo[62]), T8e, ovs, &(xo[2]));
|
Chris@82
|
1785 STN2(&(xo[60]), T86, T8e, ovs);
|
Chris@82
|
1786 }
|
Chris@82
|
1787 {
|
Chris@82
|
1788 V T33, T36, T8f, T8g;
|
Chris@82
|
1789 T33 = VBYI(VSUB(T31, T32));
|
Chris@82
|
1790 T36 = VSUB(T34, T35);
|
Chris@82
|
1791 T8f = VADD(T33, T36);
|
Chris@82
|
1792 STM2(&(xo[34]), T8f, ovs, &(xo[2]));
|
Chris@82
|
1793 STN2(&(xo[32]), T7o, T8f, ovs);
|
Chris@82
|
1794 T8g = VSUB(T36, T33);
|
Chris@82
|
1795 STM2(&(xo[94]), T8g, ovs, &(xo[2]));
|
Chris@82
|
1796 STN2(&(xo[92]), T88, T8g, ovs);
|
Chris@82
|
1797 }
|
Chris@82
|
1798 }
|
Chris@82
|
1799 {
|
Chris@82
|
1800 V T3X, T4i, T4b, T4j, T44, T4f, T48, T4g;
|
Chris@82
|
1801 {
|
Chris@82
|
1802 V T3V, T3W, T49, T4a;
|
Chris@82
|
1803 T3V = VSUB(T39, T3a);
|
Chris@82
|
1804 T3W = VSUB(T3E, T3D);
|
Chris@82
|
1805 T3X = VADD(T3V, T3W);
|
Chris@82
|
1806 T4i = VSUB(T3V, T3W);
|
Chris@82
|
1807 T49 = VFNMS(LDK(KP471396736), T3Y, VMUL(LDK(KP881921264), T3Z));
|
Chris@82
|
1808 T4a = VFMA(LDK(KP471396736), T41, VMUL(LDK(KP881921264), T42));
|
Chris@82
|
1809 T4b = VADD(T49, T4a);
|
Chris@82
|
1810 T4j = VSUB(T4a, T49);
|
Chris@82
|
1811 }
|
Chris@82
|
1812 {
|
Chris@82
|
1813 V T40, T43, T46, T47;
|
Chris@82
|
1814 T40 = VFMA(LDK(KP881921264), T3Y, VMUL(LDK(KP471396736), T3Z));
|
Chris@82
|
1815 T43 = VFNMS(LDK(KP471396736), T42, VMUL(LDK(KP881921264), T41));
|
Chris@82
|
1816 T44 = VADD(T40, T43);
|
Chris@82
|
1817 T4f = VSUB(T43, T40);
|
Chris@82
|
1818 T46 = VSUB(T3B, T3A);
|
Chris@82
|
1819 T47 = VSUB(T3h, T3e);
|
Chris@82
|
1820 T48 = VADD(T46, T47);
|
Chris@82
|
1821 T4g = VSUB(T47, T46);
|
Chris@82
|
1822 }
|
Chris@82
|
1823 {
|
Chris@82
|
1824 V T45, T4c, T8h, T8i;
|
Chris@82
|
1825 T45 = VADD(T3X, T44);
|
Chris@82
|
1826 T4c = VBYI(VADD(T48, T4b));
|
Chris@82
|
1827 T8h = VSUB(T45, T4c);
|
Chris@82
|
1828 STM2(&(xo[118]), T8h, ovs, &(xo[2]));
|
Chris@82
|
1829 STN2(&(xo[116]), T7D, T8h, ovs);
|
Chris@82
|
1830 T8i = VADD(T45, T4c);
|
Chris@82
|
1831 STM2(&(xo[10]), T8i, ovs, &(xo[2]));
|
Chris@82
|
1832 STN2(&(xo[8]), T7w, T8i, ovs);
|
Chris@82
|
1833 }
|
Chris@82
|
1834 {
|
Chris@82
|
1835 V T4l, T4m, T8j, T8k;
|
Chris@82
|
1836 T4l = VBYI(VADD(T4g, T4f));
|
Chris@82
|
1837 T4m = VADD(T4i, T4j);
|
Chris@82
|
1838 T8j = VADD(T4l, T4m);
|
Chris@82
|
1839 STM2(&(xo[22]), T8j, ovs, &(xo[2]));
|
Chris@82
|
1840 STN2(&(xo[20]), T7F, T8j, ovs);
|
Chris@82
|
1841 T8k = VSUB(T4m, T4l);
|
Chris@82
|
1842 STM2(&(xo[106]), T8k, ovs, &(xo[2]));
|
Chris@82
|
1843 STN2(&(xo[104]), T7y, T8k, ovs);
|
Chris@82
|
1844 }
|
Chris@82
|
1845 {
|
Chris@82
|
1846 V T4d, T4e, T8l, T8m;
|
Chris@82
|
1847 T4d = VSUB(T3X, T44);
|
Chris@82
|
1848 T4e = VBYI(VSUB(T4b, T48));
|
Chris@82
|
1849 T8l = VSUB(T4d, T4e);
|
Chris@82
|
1850 STM2(&(xo[74]), T8l, ovs, &(xo[2]));
|
Chris@82
|
1851 STN2(&(xo[72]), T7z, T8l, ovs);
|
Chris@82
|
1852 T8m = VADD(T4d, T4e);
|
Chris@82
|
1853 STM2(&(xo[54]), T8m, ovs, &(xo[2]));
|
Chris@82
|
1854 STN2(&(xo[52]), T7I, T8m, ovs);
|
Chris@82
|
1855 }
|
Chris@82
|
1856 {
|
Chris@82
|
1857 V T4h, T4k, T8n, T8o;
|
Chris@82
|
1858 T4h = VBYI(VSUB(T4f, T4g));
|
Chris@82
|
1859 T4k = VSUB(T4i, T4j);
|
Chris@82
|
1860 T8n = VADD(T4h, T4k);
|
Chris@82
|
1861 STM2(&(xo[42]), T8n, ovs, &(xo[2]));
|
Chris@82
|
1862 STN2(&(xo[40]), T7B, T8n, ovs);
|
Chris@82
|
1863 T8o = VSUB(T4k, T4h);
|
Chris@82
|
1864 STM2(&(xo[86]), T8o, ovs, &(xo[2]));
|
Chris@82
|
1865 STN2(&(xo[84]), T7K, T8o, ovs);
|
Chris@82
|
1866 }
|
Chris@82
|
1867 }
|
Chris@82
|
1868 }
|
Chris@82
|
1869 }
|
Chris@82
|
1870 }
|
Chris@82
|
1871 }
|
Chris@82
|
1872 }
|
Chris@82
|
1873 VLEAVE();
|
Chris@82
|
1874 }
|
Chris@82
|
1875
|
Chris@82
|
1876 static const kdft_desc desc = { 64, XSIMD_STRING("n2fv_64"), {404, 72, 52, 0}, &GENUS, 0, 2, 0, 0 };
|
Chris@82
|
1877
|
Chris@82
|
1878 void XSIMD(codelet_n2fv_64) (planner *p) {
|
Chris@82
|
1879 X(kdft_register) (p, n2fv_64, &desc);
|
Chris@82
|
1880 }
|
Chris@82
|
1881
|
Chris@82
|
1882 #endif
|