cannam@167
|
1 /*
|
cannam@167
|
2 * Copyright (c) 2003, 2007-14 Matteo Frigo
|
cannam@167
|
3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
|
cannam@167
|
4 *
|
cannam@167
|
5 * This program is free software; you can redistribute it and/or modify
|
cannam@167
|
6 * it under the terms of the GNU General Public License as published by
|
cannam@167
|
7 * the Free Software Foundation; either version 2 of the License, or
|
cannam@167
|
8 * (at your option) any later version.
|
cannam@167
|
9 *
|
cannam@167
|
10 * This program is distributed in the hope that it will be useful,
|
cannam@167
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
cannam@167
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
cannam@167
|
13 * GNU General Public License for more details.
|
cannam@167
|
14 *
|
cannam@167
|
15 * You should have received a copy of the GNU General Public License
|
cannam@167
|
16 * along with this program; if not, write to the Free Software
|
cannam@167
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
cannam@167
|
18 *
|
cannam@167
|
19 */
|
cannam@167
|
20
|
cannam@167
|
21 /* This file was automatically generated --- DO NOT EDIT */
|
cannam@167
|
22 /* Generated on Thu May 24 08:06:10 EDT 2018 */
|
cannam@167
|
23
|
cannam@167
|
24 #include "dft/codelet-dft.h"
|
cannam@167
|
25
|
cannam@167
|
26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
|
cannam@167
|
27
|
cannam@167
|
28 /* Generated by: ../../../genfft/gen_twiddle.native -fma -simd -compact -variables 4 -pipeline-latency 8 -n 32 -name t1sv_32 -include dft/simd/ts.h */
|
cannam@167
|
29
|
cannam@167
|
30 /*
|
cannam@167
|
31 * This function contains 434 FP additions, 260 FP multiplications,
|
cannam@167
|
32 * (or, 236 additions, 62 multiplications, 198 fused multiply/add),
|
cannam@167
|
33 * 102 stack variables, 7 constants, and 128 memory accesses
|
cannam@167
|
34 */
|
cannam@167
|
35 #include "dft/simd/ts.h"
|
cannam@167
|
36
|
cannam@167
|
37 static void t1sv_32(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
38 {
|
cannam@167
|
39 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
cannam@167
|
40 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
cannam@167
|
41 DVK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
cannam@167
|
42 DVK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
cannam@167
|
43 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
44 DVK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
cannam@167
|
45 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
46 {
|
cannam@167
|
47 INT m;
|
cannam@167
|
48 for (m = mb, W = W + (mb * 62); m < me; m = m + (2 * VL), ri = ri + ((2 * VL) * ms), ii = ii + ((2 * VL) * ms), W = W + ((2 * VL) * 62), MAKE_VOLATILE_STRIDE(64, rs)) {
|
cannam@167
|
49 V T8, T8x, T3w, T87, Tl, T8y, T3B, T83, Tz, T6F, T3J, T5T, TM, T6G, T3Q;
|
cannam@167
|
50 V T5U, T11, T1e, T6M, T6J, T6K, T6L, T3Z, T5X, T46, T5Y, T1s, T1F, T6O, T6P;
|
cannam@167
|
51 V T6Q, T6R, T4e, T60, T4l, T61, T32, T7b, T78, T7N, T54, T6f, T5r, T6c, T29;
|
cannam@167
|
52 V T70, T6X, T7I, T4v, T68, T4S, T65, T3t, T79, T7e, T7O, T5b, T5s, T5i, T5t;
|
cannam@167
|
53 V T2A, T6Y, T73, T7J, T4C, T4T, T4J, T4U;
|
cannam@167
|
54 {
|
cannam@167
|
55 V T1, T86, T3, T6, T4, T84, T2, T7, T85, T5;
|
cannam@167
|
56 T1 = LD(&(ri[0]), ms, &(ri[0]));
|
cannam@167
|
57 T86 = LD(&(ii[0]), ms, &(ii[0]));
|
cannam@167
|
58 T3 = LD(&(ri[WS(rs, 16)]), ms, &(ri[0]));
|
cannam@167
|
59 T6 = LD(&(ii[WS(rs, 16)]), ms, &(ii[0]));
|
cannam@167
|
60 T2 = LDW(&(W[TWVL * 30]));
|
cannam@167
|
61 T4 = VMUL(T2, T3);
|
cannam@167
|
62 T84 = VMUL(T2, T6);
|
cannam@167
|
63 T5 = LDW(&(W[TWVL * 31]));
|
cannam@167
|
64 T7 = VFMA(T5, T6, T4);
|
cannam@167
|
65 T85 = VFNMS(T5, T3, T84);
|
cannam@167
|
66 T8 = VADD(T1, T7);
|
cannam@167
|
67 T8x = VSUB(T86, T85);
|
cannam@167
|
68 T3w = VSUB(T1, T7);
|
cannam@167
|
69 T87 = VADD(T85, T86);
|
cannam@167
|
70 }
|
cannam@167
|
71 {
|
cannam@167
|
72 V Ta, Td, Tb, T3x, Tg, Tj, Th, T3z, T9, Tf;
|
cannam@167
|
73 Ta = LD(&(ri[WS(rs, 8)]), ms, &(ri[0]));
|
cannam@167
|
74 Td = LD(&(ii[WS(rs, 8)]), ms, &(ii[0]));
|
cannam@167
|
75 T9 = LDW(&(W[TWVL * 14]));
|
cannam@167
|
76 Tb = VMUL(T9, Ta);
|
cannam@167
|
77 T3x = VMUL(T9, Td);
|
cannam@167
|
78 Tg = LD(&(ri[WS(rs, 24)]), ms, &(ri[0]));
|
cannam@167
|
79 Tj = LD(&(ii[WS(rs, 24)]), ms, &(ii[0]));
|
cannam@167
|
80 Tf = LDW(&(W[TWVL * 46]));
|
cannam@167
|
81 Th = VMUL(Tf, Tg);
|
cannam@167
|
82 T3z = VMUL(Tf, Tj);
|
cannam@167
|
83 {
|
cannam@167
|
84 V Te, T3y, Tk, T3A, Tc, Ti;
|
cannam@167
|
85 Tc = LDW(&(W[TWVL * 15]));
|
cannam@167
|
86 Te = VFMA(Tc, Td, Tb);
|
cannam@167
|
87 T3y = VFNMS(Tc, Ta, T3x);
|
cannam@167
|
88 Ti = LDW(&(W[TWVL * 47]));
|
cannam@167
|
89 Tk = VFMA(Ti, Tj, Th);
|
cannam@167
|
90 T3A = VFNMS(Ti, Tg, T3z);
|
cannam@167
|
91 Tl = VADD(Te, Tk);
|
cannam@167
|
92 T8y = VSUB(Te, Tk);
|
cannam@167
|
93 T3B = VSUB(T3y, T3A);
|
cannam@167
|
94 T83 = VADD(T3y, T3A);
|
cannam@167
|
95 }
|
cannam@167
|
96 }
|
cannam@167
|
97 {
|
cannam@167
|
98 V Ts, T3F, Ty, T3H, T3D, T3I;
|
cannam@167
|
99 {
|
cannam@167
|
100 V To, Tr, Tp, T3E, Tn, Tq;
|
cannam@167
|
101 To = LD(&(ri[WS(rs, 4)]), ms, &(ri[0]));
|
cannam@167
|
102 Tr = LD(&(ii[WS(rs, 4)]), ms, &(ii[0]));
|
cannam@167
|
103 Tn = LDW(&(W[TWVL * 6]));
|
cannam@167
|
104 Tp = VMUL(Tn, To);
|
cannam@167
|
105 T3E = VMUL(Tn, Tr);
|
cannam@167
|
106 Tq = LDW(&(W[TWVL * 7]));
|
cannam@167
|
107 Ts = VFMA(Tq, Tr, Tp);
|
cannam@167
|
108 T3F = VFNMS(Tq, To, T3E);
|
cannam@167
|
109 }
|
cannam@167
|
110 {
|
cannam@167
|
111 V Tu, Tx, Tv, T3G, Tt, Tw;
|
cannam@167
|
112 Tu = LD(&(ri[WS(rs, 20)]), ms, &(ri[0]));
|
cannam@167
|
113 Tx = LD(&(ii[WS(rs, 20)]), ms, &(ii[0]));
|
cannam@167
|
114 Tt = LDW(&(W[TWVL * 38]));
|
cannam@167
|
115 Tv = VMUL(Tt, Tu);
|
cannam@167
|
116 T3G = VMUL(Tt, Tx);
|
cannam@167
|
117 Tw = LDW(&(W[TWVL * 39]));
|
cannam@167
|
118 Ty = VFMA(Tw, Tx, Tv);
|
cannam@167
|
119 T3H = VFNMS(Tw, Tu, T3G);
|
cannam@167
|
120 }
|
cannam@167
|
121 Tz = VADD(Ts, Ty);
|
cannam@167
|
122 T6F = VADD(T3F, T3H);
|
cannam@167
|
123 T3D = VSUB(Ts, Ty);
|
cannam@167
|
124 T3I = VSUB(T3F, T3H);
|
cannam@167
|
125 T3J = VADD(T3D, T3I);
|
cannam@167
|
126 T5T = VSUB(T3I, T3D);
|
cannam@167
|
127 }
|
cannam@167
|
128 {
|
cannam@167
|
129 V TF, T3M, TL, T3O, T3K, T3P;
|
cannam@167
|
130 {
|
cannam@167
|
131 V TB, TE, TC, T3L, TA, TD;
|
cannam@167
|
132 TB = LD(&(ri[WS(rs, 28)]), ms, &(ri[0]));
|
cannam@167
|
133 TE = LD(&(ii[WS(rs, 28)]), ms, &(ii[0]));
|
cannam@167
|
134 TA = LDW(&(W[TWVL * 54]));
|
cannam@167
|
135 TC = VMUL(TA, TB);
|
cannam@167
|
136 T3L = VMUL(TA, TE);
|
cannam@167
|
137 TD = LDW(&(W[TWVL * 55]));
|
cannam@167
|
138 TF = VFMA(TD, TE, TC);
|
cannam@167
|
139 T3M = VFNMS(TD, TB, T3L);
|
cannam@167
|
140 }
|
cannam@167
|
141 {
|
cannam@167
|
142 V TH, TK, TI, T3N, TG, TJ;
|
cannam@167
|
143 TH = LD(&(ri[WS(rs, 12)]), ms, &(ri[0]));
|
cannam@167
|
144 TK = LD(&(ii[WS(rs, 12)]), ms, &(ii[0]));
|
cannam@167
|
145 TG = LDW(&(W[TWVL * 22]));
|
cannam@167
|
146 TI = VMUL(TG, TH);
|
cannam@167
|
147 T3N = VMUL(TG, TK);
|
cannam@167
|
148 TJ = LDW(&(W[TWVL * 23]));
|
cannam@167
|
149 TL = VFMA(TJ, TK, TI);
|
cannam@167
|
150 T3O = VFNMS(TJ, TH, T3N);
|
cannam@167
|
151 }
|
cannam@167
|
152 TM = VADD(TF, TL);
|
cannam@167
|
153 T6G = VADD(T3M, T3O);
|
cannam@167
|
154 T3K = VSUB(TF, TL);
|
cannam@167
|
155 T3P = VSUB(T3M, T3O);
|
cannam@167
|
156 T3Q = VSUB(T3K, T3P);
|
cannam@167
|
157 T5U = VADD(T3K, T3P);
|
cannam@167
|
158 }
|
cannam@167
|
159 {
|
cannam@167
|
160 V TU, T3U, T1d, T44, T10, T3W, T17, T42;
|
cannam@167
|
161 {
|
cannam@167
|
162 V TQ, TT, TR, T3T, TP, TS;
|
cannam@167
|
163 TQ = LD(&(ri[WS(rs, 2)]), ms, &(ri[0]));
|
cannam@167
|
164 TT = LD(&(ii[WS(rs, 2)]), ms, &(ii[0]));
|
cannam@167
|
165 TP = LDW(&(W[TWVL * 2]));
|
cannam@167
|
166 TR = VMUL(TP, TQ);
|
cannam@167
|
167 T3T = VMUL(TP, TT);
|
cannam@167
|
168 TS = LDW(&(W[TWVL * 3]));
|
cannam@167
|
169 TU = VFMA(TS, TT, TR);
|
cannam@167
|
170 T3U = VFNMS(TS, TQ, T3T);
|
cannam@167
|
171 }
|
cannam@167
|
172 {
|
cannam@167
|
173 V T19, T1c, T1a, T43, T18, T1b;
|
cannam@167
|
174 T19 = LD(&(ri[WS(rs, 26)]), ms, &(ri[0]));
|
cannam@167
|
175 T1c = LD(&(ii[WS(rs, 26)]), ms, &(ii[0]));
|
cannam@167
|
176 T18 = LDW(&(W[TWVL * 50]));
|
cannam@167
|
177 T1a = VMUL(T18, T19);
|
cannam@167
|
178 T43 = VMUL(T18, T1c);
|
cannam@167
|
179 T1b = LDW(&(W[TWVL * 51]));
|
cannam@167
|
180 T1d = VFMA(T1b, T1c, T1a);
|
cannam@167
|
181 T44 = VFNMS(T1b, T19, T43);
|
cannam@167
|
182 }
|
cannam@167
|
183 {
|
cannam@167
|
184 V TW, TZ, TX, T3V, TV, TY;
|
cannam@167
|
185 TW = LD(&(ri[WS(rs, 18)]), ms, &(ri[0]));
|
cannam@167
|
186 TZ = LD(&(ii[WS(rs, 18)]), ms, &(ii[0]));
|
cannam@167
|
187 TV = LDW(&(W[TWVL * 34]));
|
cannam@167
|
188 TX = VMUL(TV, TW);
|
cannam@167
|
189 T3V = VMUL(TV, TZ);
|
cannam@167
|
190 TY = LDW(&(W[TWVL * 35]));
|
cannam@167
|
191 T10 = VFMA(TY, TZ, TX);
|
cannam@167
|
192 T3W = VFNMS(TY, TW, T3V);
|
cannam@167
|
193 }
|
cannam@167
|
194 {
|
cannam@167
|
195 V T13, T16, T14, T41, T12, T15;
|
cannam@167
|
196 T13 = LD(&(ri[WS(rs, 10)]), ms, &(ri[0]));
|
cannam@167
|
197 T16 = LD(&(ii[WS(rs, 10)]), ms, &(ii[0]));
|
cannam@167
|
198 T12 = LDW(&(W[TWVL * 18]));
|
cannam@167
|
199 T14 = VMUL(T12, T13);
|
cannam@167
|
200 T41 = VMUL(T12, T16);
|
cannam@167
|
201 T15 = LDW(&(W[TWVL * 19]));
|
cannam@167
|
202 T17 = VFMA(T15, T16, T14);
|
cannam@167
|
203 T42 = VFNMS(T15, T13, T41);
|
cannam@167
|
204 }
|
cannam@167
|
205 T11 = VADD(TU, T10);
|
cannam@167
|
206 T1e = VADD(T17, T1d);
|
cannam@167
|
207 T6M = VSUB(T11, T1e);
|
cannam@167
|
208 T6J = VADD(T3U, T3W);
|
cannam@167
|
209 T6K = VADD(T42, T44);
|
cannam@167
|
210 T6L = VSUB(T6J, T6K);
|
cannam@167
|
211 {
|
cannam@167
|
212 V T3X, T3Y, T40, T45;
|
cannam@167
|
213 T3X = VSUB(T3U, T3W);
|
cannam@167
|
214 T3Y = VSUB(T17, T1d);
|
cannam@167
|
215 T3Z = VSUB(T3X, T3Y);
|
cannam@167
|
216 T5X = VADD(T3X, T3Y);
|
cannam@167
|
217 T40 = VSUB(TU, T10);
|
cannam@167
|
218 T45 = VSUB(T42, T44);
|
cannam@167
|
219 T46 = VADD(T40, T45);
|
cannam@167
|
220 T5Y = VSUB(T40, T45);
|
cannam@167
|
221 }
|
cannam@167
|
222 }
|
cannam@167
|
223 {
|
cannam@167
|
224 V T1l, T49, T1E, T4j, T1r, T4b, T1y, T4h;
|
cannam@167
|
225 {
|
cannam@167
|
226 V T1h, T1k, T1i, T48, T1g, T1j;
|
cannam@167
|
227 T1h = LD(&(ri[WS(rs, 30)]), ms, &(ri[0]));
|
cannam@167
|
228 T1k = LD(&(ii[WS(rs, 30)]), ms, &(ii[0]));
|
cannam@167
|
229 T1g = LDW(&(W[TWVL * 58]));
|
cannam@167
|
230 T1i = VMUL(T1g, T1h);
|
cannam@167
|
231 T48 = VMUL(T1g, T1k);
|
cannam@167
|
232 T1j = LDW(&(W[TWVL * 59]));
|
cannam@167
|
233 T1l = VFMA(T1j, T1k, T1i);
|
cannam@167
|
234 T49 = VFNMS(T1j, T1h, T48);
|
cannam@167
|
235 }
|
cannam@167
|
236 {
|
cannam@167
|
237 V T1A, T1D, T1B, T4i, T1z, T1C;
|
cannam@167
|
238 T1A = LD(&(ri[WS(rs, 22)]), ms, &(ri[0]));
|
cannam@167
|
239 T1D = LD(&(ii[WS(rs, 22)]), ms, &(ii[0]));
|
cannam@167
|
240 T1z = LDW(&(W[TWVL * 42]));
|
cannam@167
|
241 T1B = VMUL(T1z, T1A);
|
cannam@167
|
242 T4i = VMUL(T1z, T1D);
|
cannam@167
|
243 T1C = LDW(&(W[TWVL * 43]));
|
cannam@167
|
244 T1E = VFMA(T1C, T1D, T1B);
|
cannam@167
|
245 T4j = VFNMS(T1C, T1A, T4i);
|
cannam@167
|
246 }
|
cannam@167
|
247 {
|
cannam@167
|
248 V T1n, T1q, T1o, T4a, T1m, T1p;
|
cannam@167
|
249 T1n = LD(&(ri[WS(rs, 14)]), ms, &(ri[0]));
|
cannam@167
|
250 T1q = LD(&(ii[WS(rs, 14)]), ms, &(ii[0]));
|
cannam@167
|
251 T1m = LDW(&(W[TWVL * 26]));
|
cannam@167
|
252 T1o = VMUL(T1m, T1n);
|
cannam@167
|
253 T4a = VMUL(T1m, T1q);
|
cannam@167
|
254 T1p = LDW(&(W[TWVL * 27]));
|
cannam@167
|
255 T1r = VFMA(T1p, T1q, T1o);
|
cannam@167
|
256 T4b = VFNMS(T1p, T1n, T4a);
|
cannam@167
|
257 }
|
cannam@167
|
258 {
|
cannam@167
|
259 V T1u, T1x, T1v, T4g, T1t, T1w;
|
cannam@167
|
260 T1u = LD(&(ri[WS(rs, 6)]), ms, &(ri[0]));
|
cannam@167
|
261 T1x = LD(&(ii[WS(rs, 6)]), ms, &(ii[0]));
|
cannam@167
|
262 T1t = LDW(&(W[TWVL * 10]));
|
cannam@167
|
263 T1v = VMUL(T1t, T1u);
|
cannam@167
|
264 T4g = VMUL(T1t, T1x);
|
cannam@167
|
265 T1w = LDW(&(W[TWVL * 11]));
|
cannam@167
|
266 T1y = VFMA(T1w, T1x, T1v);
|
cannam@167
|
267 T4h = VFNMS(T1w, T1u, T4g);
|
cannam@167
|
268 }
|
cannam@167
|
269 T1s = VADD(T1l, T1r);
|
cannam@167
|
270 T1F = VADD(T1y, T1E);
|
cannam@167
|
271 T6O = VSUB(T1s, T1F);
|
cannam@167
|
272 T6P = VADD(T49, T4b);
|
cannam@167
|
273 T6Q = VADD(T4h, T4j);
|
cannam@167
|
274 T6R = VSUB(T6P, T6Q);
|
cannam@167
|
275 {
|
cannam@167
|
276 V T4c, T4d, T4f, T4k;
|
cannam@167
|
277 T4c = VSUB(T49, T4b);
|
cannam@167
|
278 T4d = VSUB(T1y, T1E);
|
cannam@167
|
279 T4e = VSUB(T4c, T4d);
|
cannam@167
|
280 T60 = VADD(T4c, T4d);
|
cannam@167
|
281 T4f = VSUB(T1l, T1r);
|
cannam@167
|
282 T4k = VSUB(T4h, T4j);
|
cannam@167
|
283 T4l = VADD(T4f, T4k);
|
cannam@167
|
284 T61 = VSUB(T4f, T4k);
|
cannam@167
|
285 }
|
cannam@167
|
286 }
|
cannam@167
|
287 {
|
cannam@167
|
288 V T2H, T4Z, T30, T5p, T2N, T51, T2U, T5n;
|
cannam@167
|
289 {
|
cannam@167
|
290 V T2D, T2G, T2E, T4Y, T2C, T2F;
|
cannam@167
|
291 T2D = LD(&(ri[WS(rs, 31)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
292 T2G = LD(&(ii[WS(rs, 31)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
293 T2C = LDW(&(W[TWVL * 60]));
|
cannam@167
|
294 T2E = VMUL(T2C, T2D);
|
cannam@167
|
295 T4Y = VMUL(T2C, T2G);
|
cannam@167
|
296 T2F = LDW(&(W[TWVL * 61]));
|
cannam@167
|
297 T2H = VFMA(T2F, T2G, T2E);
|
cannam@167
|
298 T4Z = VFNMS(T2F, T2D, T4Y);
|
cannam@167
|
299 }
|
cannam@167
|
300 {
|
cannam@167
|
301 V T2W, T2Z, T2X, T5o, T2V, T2Y;
|
cannam@167
|
302 T2W = LD(&(ri[WS(rs, 23)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
303 T2Z = LD(&(ii[WS(rs, 23)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
304 T2V = LDW(&(W[TWVL * 44]));
|
cannam@167
|
305 T2X = VMUL(T2V, T2W);
|
cannam@167
|
306 T5o = VMUL(T2V, T2Z);
|
cannam@167
|
307 T2Y = LDW(&(W[TWVL * 45]));
|
cannam@167
|
308 T30 = VFMA(T2Y, T2Z, T2X);
|
cannam@167
|
309 T5p = VFNMS(T2Y, T2W, T5o);
|
cannam@167
|
310 }
|
cannam@167
|
311 {
|
cannam@167
|
312 V T2J, T2M, T2K, T50, T2I, T2L;
|
cannam@167
|
313 T2J = LD(&(ri[WS(rs, 15)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
314 T2M = LD(&(ii[WS(rs, 15)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
315 T2I = LDW(&(W[TWVL * 28]));
|
cannam@167
|
316 T2K = VMUL(T2I, T2J);
|
cannam@167
|
317 T50 = VMUL(T2I, T2M);
|
cannam@167
|
318 T2L = LDW(&(W[TWVL * 29]));
|
cannam@167
|
319 T2N = VFMA(T2L, T2M, T2K);
|
cannam@167
|
320 T51 = VFNMS(T2L, T2J, T50);
|
cannam@167
|
321 }
|
cannam@167
|
322 {
|
cannam@167
|
323 V T2Q, T2T, T2R, T5m, T2P, T2S;
|
cannam@167
|
324 T2Q = LD(&(ri[WS(rs, 7)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
325 T2T = LD(&(ii[WS(rs, 7)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
326 T2P = LDW(&(W[TWVL * 12]));
|
cannam@167
|
327 T2R = VMUL(T2P, T2Q);
|
cannam@167
|
328 T5m = VMUL(T2P, T2T);
|
cannam@167
|
329 T2S = LDW(&(W[TWVL * 13]));
|
cannam@167
|
330 T2U = VFMA(T2S, T2T, T2R);
|
cannam@167
|
331 T5n = VFNMS(T2S, T2Q, T5m);
|
cannam@167
|
332 }
|
cannam@167
|
333 {
|
cannam@167
|
334 V T2O, T31, T76, T77;
|
cannam@167
|
335 T2O = VADD(T2H, T2N);
|
cannam@167
|
336 T31 = VADD(T2U, T30);
|
cannam@167
|
337 T32 = VADD(T2O, T31);
|
cannam@167
|
338 T7b = VSUB(T2O, T31);
|
cannam@167
|
339 T76 = VADD(T4Z, T51);
|
cannam@167
|
340 T77 = VADD(T5n, T5p);
|
cannam@167
|
341 T78 = VSUB(T76, T77);
|
cannam@167
|
342 T7N = VADD(T76, T77);
|
cannam@167
|
343 }
|
cannam@167
|
344 {
|
cannam@167
|
345 V T52, T53, T5l, T5q;
|
cannam@167
|
346 T52 = VSUB(T4Z, T51);
|
cannam@167
|
347 T53 = VSUB(T2U, T30);
|
cannam@167
|
348 T54 = VSUB(T52, T53);
|
cannam@167
|
349 T6f = VADD(T52, T53);
|
cannam@167
|
350 T5l = VSUB(T2H, T2N);
|
cannam@167
|
351 T5q = VSUB(T5n, T5p);
|
cannam@167
|
352 T5r = VADD(T5l, T5q);
|
cannam@167
|
353 T6c = VSUB(T5l, T5q);
|
cannam@167
|
354 }
|
cannam@167
|
355 }
|
cannam@167
|
356 {
|
cannam@167
|
357 V T1O, T4q, T27, T4Q, T1U, T4s, T21, T4O;
|
cannam@167
|
358 {
|
cannam@167
|
359 V T1K, T1N, T1L, T4p, T1J, T1M;
|
cannam@167
|
360 T1K = LD(&(ri[WS(rs, 1)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
361 T1N = LD(&(ii[WS(rs, 1)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
362 T1J = LDW(&(W[0]));
|
cannam@167
|
363 T1L = VMUL(T1J, T1K);
|
cannam@167
|
364 T4p = VMUL(T1J, T1N);
|
cannam@167
|
365 T1M = LDW(&(W[TWVL * 1]));
|
cannam@167
|
366 T1O = VFMA(T1M, T1N, T1L);
|
cannam@167
|
367 T4q = VFNMS(T1M, T1K, T4p);
|
cannam@167
|
368 }
|
cannam@167
|
369 {
|
cannam@167
|
370 V T23, T26, T24, T4P, T22, T25;
|
cannam@167
|
371 T23 = LD(&(ri[WS(rs, 25)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
372 T26 = LD(&(ii[WS(rs, 25)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
373 T22 = LDW(&(W[TWVL * 48]));
|
cannam@167
|
374 T24 = VMUL(T22, T23);
|
cannam@167
|
375 T4P = VMUL(T22, T26);
|
cannam@167
|
376 T25 = LDW(&(W[TWVL * 49]));
|
cannam@167
|
377 T27 = VFMA(T25, T26, T24);
|
cannam@167
|
378 T4Q = VFNMS(T25, T23, T4P);
|
cannam@167
|
379 }
|
cannam@167
|
380 {
|
cannam@167
|
381 V T1Q, T1T, T1R, T4r, T1P, T1S;
|
cannam@167
|
382 T1Q = LD(&(ri[WS(rs, 17)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
383 T1T = LD(&(ii[WS(rs, 17)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
384 T1P = LDW(&(W[TWVL * 32]));
|
cannam@167
|
385 T1R = VMUL(T1P, T1Q);
|
cannam@167
|
386 T4r = VMUL(T1P, T1T);
|
cannam@167
|
387 T1S = LDW(&(W[TWVL * 33]));
|
cannam@167
|
388 T1U = VFMA(T1S, T1T, T1R);
|
cannam@167
|
389 T4s = VFNMS(T1S, T1Q, T4r);
|
cannam@167
|
390 }
|
cannam@167
|
391 {
|
cannam@167
|
392 V T1X, T20, T1Y, T4N, T1W, T1Z;
|
cannam@167
|
393 T1X = LD(&(ri[WS(rs, 9)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
394 T20 = LD(&(ii[WS(rs, 9)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
395 T1W = LDW(&(W[TWVL * 16]));
|
cannam@167
|
396 T1Y = VMUL(T1W, T1X);
|
cannam@167
|
397 T4N = VMUL(T1W, T20);
|
cannam@167
|
398 T1Z = LDW(&(W[TWVL * 17]));
|
cannam@167
|
399 T21 = VFMA(T1Z, T20, T1Y);
|
cannam@167
|
400 T4O = VFNMS(T1Z, T1X, T4N);
|
cannam@167
|
401 }
|
cannam@167
|
402 {
|
cannam@167
|
403 V T1V, T28, T6V, T6W;
|
cannam@167
|
404 T1V = VADD(T1O, T1U);
|
cannam@167
|
405 T28 = VADD(T21, T27);
|
cannam@167
|
406 T29 = VADD(T1V, T28);
|
cannam@167
|
407 T70 = VSUB(T1V, T28);
|
cannam@167
|
408 T6V = VADD(T4q, T4s);
|
cannam@167
|
409 T6W = VADD(T4O, T4Q);
|
cannam@167
|
410 T6X = VSUB(T6V, T6W);
|
cannam@167
|
411 T7I = VADD(T6V, T6W);
|
cannam@167
|
412 }
|
cannam@167
|
413 {
|
cannam@167
|
414 V T4t, T4u, T4M, T4R;
|
cannam@167
|
415 T4t = VSUB(T4q, T4s);
|
cannam@167
|
416 T4u = VSUB(T21, T27);
|
cannam@167
|
417 T4v = VSUB(T4t, T4u);
|
cannam@167
|
418 T68 = VADD(T4t, T4u);
|
cannam@167
|
419 T4M = VSUB(T1O, T1U);
|
cannam@167
|
420 T4R = VSUB(T4O, T4Q);
|
cannam@167
|
421 T4S = VADD(T4M, T4R);
|
cannam@167
|
422 T65 = VSUB(T4M, T4R);
|
cannam@167
|
423 }
|
cannam@167
|
424 }
|
cannam@167
|
425 {
|
cannam@167
|
426 V T38, T56, T3r, T5g, T3e, T58, T3l, T5e;
|
cannam@167
|
427 {
|
cannam@167
|
428 V T34, T37, T35, T55, T33, T36;
|
cannam@167
|
429 T34 = LD(&(ri[WS(rs, 3)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
430 T37 = LD(&(ii[WS(rs, 3)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
431 T33 = LDW(&(W[TWVL * 4]));
|
cannam@167
|
432 T35 = VMUL(T33, T34);
|
cannam@167
|
433 T55 = VMUL(T33, T37);
|
cannam@167
|
434 T36 = LDW(&(W[TWVL * 5]));
|
cannam@167
|
435 T38 = VFMA(T36, T37, T35);
|
cannam@167
|
436 T56 = VFNMS(T36, T34, T55);
|
cannam@167
|
437 }
|
cannam@167
|
438 {
|
cannam@167
|
439 V T3n, T3q, T3o, T5f, T3m, T3p;
|
cannam@167
|
440 T3n = LD(&(ri[WS(rs, 11)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
441 T3q = LD(&(ii[WS(rs, 11)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
442 T3m = LDW(&(W[TWVL * 20]));
|
cannam@167
|
443 T3o = VMUL(T3m, T3n);
|
cannam@167
|
444 T5f = VMUL(T3m, T3q);
|
cannam@167
|
445 T3p = LDW(&(W[TWVL * 21]));
|
cannam@167
|
446 T3r = VFMA(T3p, T3q, T3o);
|
cannam@167
|
447 T5g = VFNMS(T3p, T3n, T5f);
|
cannam@167
|
448 }
|
cannam@167
|
449 {
|
cannam@167
|
450 V T3a, T3d, T3b, T57, T39, T3c;
|
cannam@167
|
451 T3a = LD(&(ri[WS(rs, 19)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
452 T3d = LD(&(ii[WS(rs, 19)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
453 T39 = LDW(&(W[TWVL * 36]));
|
cannam@167
|
454 T3b = VMUL(T39, T3a);
|
cannam@167
|
455 T57 = VMUL(T39, T3d);
|
cannam@167
|
456 T3c = LDW(&(W[TWVL * 37]));
|
cannam@167
|
457 T3e = VFMA(T3c, T3d, T3b);
|
cannam@167
|
458 T58 = VFNMS(T3c, T3a, T57);
|
cannam@167
|
459 }
|
cannam@167
|
460 {
|
cannam@167
|
461 V T3h, T3k, T3i, T5d, T3g, T3j;
|
cannam@167
|
462 T3h = LD(&(ri[WS(rs, 27)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
463 T3k = LD(&(ii[WS(rs, 27)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
464 T3g = LDW(&(W[TWVL * 52]));
|
cannam@167
|
465 T3i = VMUL(T3g, T3h);
|
cannam@167
|
466 T5d = VMUL(T3g, T3k);
|
cannam@167
|
467 T3j = LDW(&(W[TWVL * 53]));
|
cannam@167
|
468 T3l = VFMA(T3j, T3k, T3i);
|
cannam@167
|
469 T5e = VFNMS(T3j, T3h, T5d);
|
cannam@167
|
470 }
|
cannam@167
|
471 {
|
cannam@167
|
472 V T3f, T3s, T7c, T7d;
|
cannam@167
|
473 T3f = VADD(T38, T3e);
|
cannam@167
|
474 T3s = VADD(T3l, T3r);
|
cannam@167
|
475 T3t = VADD(T3f, T3s);
|
cannam@167
|
476 T79 = VSUB(T3s, T3f);
|
cannam@167
|
477 T7c = VADD(T56, T58);
|
cannam@167
|
478 T7d = VADD(T5e, T5g);
|
cannam@167
|
479 T7e = VSUB(T7c, T7d);
|
cannam@167
|
480 T7O = VADD(T7c, T7d);
|
cannam@167
|
481 }
|
cannam@167
|
482 {
|
cannam@167
|
483 V T59, T5a, T5c, T5h;
|
cannam@167
|
484 T59 = VSUB(T56, T58);
|
cannam@167
|
485 T5a = VSUB(T38, T3e);
|
cannam@167
|
486 T5b = VSUB(T59, T5a);
|
cannam@167
|
487 T5s = VADD(T5a, T59);
|
cannam@167
|
488 T5c = VSUB(T3l, T3r);
|
cannam@167
|
489 T5h = VSUB(T5e, T5g);
|
cannam@167
|
490 T5i = VADD(T5c, T5h);
|
cannam@167
|
491 T5t = VSUB(T5c, T5h);
|
cannam@167
|
492 }
|
cannam@167
|
493 }
|
cannam@167
|
494 {
|
cannam@167
|
495 V T2f, T4x, T2y, T4H, T2l, T4z, T2s, T4F;
|
cannam@167
|
496 {
|
cannam@167
|
497 V T2b, T2e, T2c, T4w, T2a, T2d;
|
cannam@167
|
498 T2b = LD(&(ri[WS(rs, 5)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
499 T2e = LD(&(ii[WS(rs, 5)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
500 T2a = LDW(&(W[TWVL * 8]));
|
cannam@167
|
501 T2c = VMUL(T2a, T2b);
|
cannam@167
|
502 T4w = VMUL(T2a, T2e);
|
cannam@167
|
503 T2d = LDW(&(W[TWVL * 9]));
|
cannam@167
|
504 T2f = VFMA(T2d, T2e, T2c);
|
cannam@167
|
505 T4x = VFNMS(T2d, T2b, T4w);
|
cannam@167
|
506 }
|
cannam@167
|
507 {
|
cannam@167
|
508 V T2u, T2x, T2v, T4G, T2t, T2w;
|
cannam@167
|
509 T2u = LD(&(ri[WS(rs, 13)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
510 T2x = LD(&(ii[WS(rs, 13)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
511 T2t = LDW(&(W[TWVL * 24]));
|
cannam@167
|
512 T2v = VMUL(T2t, T2u);
|
cannam@167
|
513 T4G = VMUL(T2t, T2x);
|
cannam@167
|
514 T2w = LDW(&(W[TWVL * 25]));
|
cannam@167
|
515 T2y = VFMA(T2w, T2x, T2v);
|
cannam@167
|
516 T4H = VFNMS(T2w, T2u, T4G);
|
cannam@167
|
517 }
|
cannam@167
|
518 {
|
cannam@167
|
519 V T2h, T2k, T2i, T4y, T2g, T2j;
|
cannam@167
|
520 T2h = LD(&(ri[WS(rs, 21)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
521 T2k = LD(&(ii[WS(rs, 21)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
522 T2g = LDW(&(W[TWVL * 40]));
|
cannam@167
|
523 T2i = VMUL(T2g, T2h);
|
cannam@167
|
524 T4y = VMUL(T2g, T2k);
|
cannam@167
|
525 T2j = LDW(&(W[TWVL * 41]));
|
cannam@167
|
526 T2l = VFMA(T2j, T2k, T2i);
|
cannam@167
|
527 T4z = VFNMS(T2j, T2h, T4y);
|
cannam@167
|
528 }
|
cannam@167
|
529 {
|
cannam@167
|
530 V T2o, T2r, T2p, T4E, T2n, T2q;
|
cannam@167
|
531 T2o = LD(&(ri[WS(rs, 29)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
532 T2r = LD(&(ii[WS(rs, 29)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
533 T2n = LDW(&(W[TWVL * 56]));
|
cannam@167
|
534 T2p = VMUL(T2n, T2o);
|
cannam@167
|
535 T4E = VMUL(T2n, T2r);
|
cannam@167
|
536 T2q = LDW(&(W[TWVL * 57]));
|
cannam@167
|
537 T2s = VFMA(T2q, T2r, T2p);
|
cannam@167
|
538 T4F = VFNMS(T2q, T2o, T4E);
|
cannam@167
|
539 }
|
cannam@167
|
540 {
|
cannam@167
|
541 V T2m, T2z, T71, T72;
|
cannam@167
|
542 T2m = VADD(T2f, T2l);
|
cannam@167
|
543 T2z = VADD(T2s, T2y);
|
cannam@167
|
544 T2A = VADD(T2m, T2z);
|
cannam@167
|
545 T6Y = VSUB(T2z, T2m);
|
cannam@167
|
546 T71 = VADD(T4x, T4z);
|
cannam@167
|
547 T72 = VADD(T4F, T4H);
|
cannam@167
|
548 T73 = VSUB(T71, T72);
|
cannam@167
|
549 T7J = VADD(T71, T72);
|
cannam@167
|
550 }
|
cannam@167
|
551 {
|
cannam@167
|
552 V T4A, T4B, T4D, T4I;
|
cannam@167
|
553 T4A = VSUB(T4x, T4z);
|
cannam@167
|
554 T4B = VSUB(T2f, T2l);
|
cannam@167
|
555 T4C = VSUB(T4A, T4B);
|
cannam@167
|
556 T4T = VADD(T4B, T4A);
|
cannam@167
|
557 T4D = VSUB(T2s, T2y);
|
cannam@167
|
558 T4I = VSUB(T4F, T4H);
|
cannam@167
|
559 T4J = VADD(T4D, T4I);
|
cannam@167
|
560 T4U = VSUB(T4D, T4I);
|
cannam@167
|
561 }
|
cannam@167
|
562 }
|
cannam@167
|
563 {
|
cannam@167
|
564 V TO, T7C, T7Z, T80, T89, T8e, T1H, T8d, T3v, T8b, T7L, T7T, T7Q, T7U, T7F;
|
cannam@167
|
565 V T81;
|
cannam@167
|
566 {
|
cannam@167
|
567 V Tm, TN, T7X, T7Y;
|
cannam@167
|
568 Tm = VADD(T8, Tl);
|
cannam@167
|
569 TN = VADD(Tz, TM);
|
cannam@167
|
570 TO = VADD(Tm, TN);
|
cannam@167
|
571 T7C = VSUB(Tm, TN);
|
cannam@167
|
572 T7X = VADD(T7I, T7J);
|
cannam@167
|
573 T7Y = VADD(T7N, T7O);
|
cannam@167
|
574 T7Z = VSUB(T7X, T7Y);
|
cannam@167
|
575 T80 = VADD(T7X, T7Y);
|
cannam@167
|
576 }
|
cannam@167
|
577 {
|
cannam@167
|
578 V T82, T88, T1f, T1G;
|
cannam@167
|
579 T82 = VADD(T6F, T6G);
|
cannam@167
|
580 T88 = VADD(T83, T87);
|
cannam@167
|
581 T89 = VADD(T82, T88);
|
cannam@167
|
582 T8e = VSUB(T88, T82);
|
cannam@167
|
583 T1f = VADD(T11, T1e);
|
cannam@167
|
584 T1G = VADD(T1s, T1F);
|
cannam@167
|
585 T1H = VADD(T1f, T1G);
|
cannam@167
|
586 T8d = VSUB(T1G, T1f);
|
cannam@167
|
587 }
|
cannam@167
|
588 {
|
cannam@167
|
589 V T2B, T3u, T7H, T7K;
|
cannam@167
|
590 T2B = VADD(T29, T2A);
|
cannam@167
|
591 T3u = VADD(T32, T3t);
|
cannam@167
|
592 T3v = VADD(T2B, T3u);
|
cannam@167
|
593 T8b = VSUB(T3u, T2B);
|
cannam@167
|
594 T7H = VSUB(T29, T2A);
|
cannam@167
|
595 T7K = VSUB(T7I, T7J);
|
cannam@167
|
596 T7L = VADD(T7H, T7K);
|
cannam@167
|
597 T7T = VSUB(T7K, T7H);
|
cannam@167
|
598 }
|
cannam@167
|
599 {
|
cannam@167
|
600 V T7M, T7P, T7D, T7E;
|
cannam@167
|
601 T7M = VSUB(T32, T3t);
|
cannam@167
|
602 T7P = VSUB(T7N, T7O);
|
cannam@167
|
603 T7Q = VSUB(T7M, T7P);
|
cannam@167
|
604 T7U = VADD(T7M, T7P);
|
cannam@167
|
605 T7D = VADD(T6J, T6K);
|
cannam@167
|
606 T7E = VADD(T6P, T6Q);
|
cannam@167
|
607 T7F = VSUB(T7D, T7E);
|
cannam@167
|
608 T81 = VADD(T7D, T7E);
|
cannam@167
|
609 }
|
cannam@167
|
610 {
|
cannam@167
|
611 V T1I, T8a, T7W, T8c;
|
cannam@167
|
612 T1I = VADD(TO, T1H);
|
cannam@167
|
613 ST(&(ri[WS(rs, 16)]), VSUB(T1I, T3v), ms, &(ri[0]));
|
cannam@167
|
614 ST(&(ri[0]), VADD(T1I, T3v), ms, &(ri[0]));
|
cannam@167
|
615 T8a = VADD(T81, T89);
|
cannam@167
|
616 ST(&(ii[0]), VADD(T80, T8a), ms, &(ii[0]));
|
cannam@167
|
617 ST(&(ii[WS(rs, 16)]), VSUB(T8a, T80), ms, &(ii[0]));
|
cannam@167
|
618 T7W = VSUB(TO, T1H);
|
cannam@167
|
619 ST(&(ri[WS(rs, 24)]), VSUB(T7W, T7Z), ms, &(ri[0]));
|
cannam@167
|
620 ST(&(ri[WS(rs, 8)]), VADD(T7W, T7Z), ms, &(ri[0]));
|
cannam@167
|
621 T8c = VSUB(T89, T81);
|
cannam@167
|
622 ST(&(ii[WS(rs, 8)]), VADD(T8b, T8c), ms, &(ii[0]));
|
cannam@167
|
623 ST(&(ii[WS(rs, 24)]), VSUB(T8c, T8b), ms, &(ii[0]));
|
cannam@167
|
624 }
|
cannam@167
|
625 {
|
cannam@167
|
626 V T7G, T7R, T8f, T8g;
|
cannam@167
|
627 T7G = VADD(T7C, T7F);
|
cannam@167
|
628 T7R = VADD(T7L, T7Q);
|
cannam@167
|
629 ST(&(ri[WS(rs, 20)]), VFNMS(LDK(KP707106781), T7R, T7G), ms, &(ri[0]));
|
cannam@167
|
630 ST(&(ri[WS(rs, 4)]), VFMA(LDK(KP707106781), T7R, T7G), ms, &(ri[0]));
|
cannam@167
|
631 T8f = VADD(T8d, T8e);
|
cannam@167
|
632 T8g = VADD(T7T, T7U);
|
cannam@167
|
633 ST(&(ii[WS(rs, 4)]), VFMA(LDK(KP707106781), T8g, T8f), ms, &(ii[0]));
|
cannam@167
|
634 ST(&(ii[WS(rs, 20)]), VFNMS(LDK(KP707106781), T8g, T8f), ms, &(ii[0]));
|
cannam@167
|
635 }
|
cannam@167
|
636 {
|
cannam@167
|
637 V T7S, T7V, T8h, T8i;
|
cannam@167
|
638 T7S = VSUB(T7C, T7F);
|
cannam@167
|
639 T7V = VSUB(T7T, T7U);
|
cannam@167
|
640 ST(&(ri[WS(rs, 28)]), VFNMS(LDK(KP707106781), T7V, T7S), ms, &(ri[0]));
|
cannam@167
|
641 ST(&(ri[WS(rs, 12)]), VFMA(LDK(KP707106781), T7V, T7S), ms, &(ri[0]));
|
cannam@167
|
642 T8h = VSUB(T8e, T8d);
|
cannam@167
|
643 T8i = VSUB(T7Q, T7L);
|
cannam@167
|
644 ST(&(ii[WS(rs, 12)]), VFMA(LDK(KP707106781), T8i, T8h), ms, &(ii[0]));
|
cannam@167
|
645 ST(&(ii[WS(rs, 28)]), VFNMS(LDK(KP707106781), T8i, T8h), ms, &(ii[0]));
|
cannam@167
|
646 }
|
cannam@167
|
647 }
|
cannam@167
|
648 {
|
cannam@167
|
649 V T6I, T7m, T7w, T7A, T8l, T8r, T6T, T8m, T75, T7j, T7p, T8s, T7t, T7z, T7g;
|
cannam@167
|
650 V T7k;
|
cannam@167
|
651 {
|
cannam@167
|
652 V T6E, T6H, T7u, T7v;
|
cannam@167
|
653 T6E = VSUB(T8, Tl);
|
cannam@167
|
654 T6H = VSUB(T6F, T6G);
|
cannam@167
|
655 T6I = VSUB(T6E, T6H);
|
cannam@167
|
656 T7m = VADD(T6E, T6H);
|
cannam@167
|
657 T7u = VADD(T7b, T7e);
|
cannam@167
|
658 T7v = VADD(T78, T79);
|
cannam@167
|
659 T7w = VFNMS(LDK(KP414213562), T7v, T7u);
|
cannam@167
|
660 T7A = VFMA(LDK(KP414213562), T7u, T7v);
|
cannam@167
|
661 }
|
cannam@167
|
662 {
|
cannam@167
|
663 V T8j, T8k, T6N, T6S;
|
cannam@167
|
664 T8j = VSUB(TM, Tz);
|
cannam@167
|
665 T8k = VSUB(T87, T83);
|
cannam@167
|
666 T8l = VADD(T8j, T8k);
|
cannam@167
|
667 T8r = VSUB(T8k, T8j);
|
cannam@167
|
668 T6N = VSUB(T6L, T6M);
|
cannam@167
|
669 T6S = VADD(T6O, T6R);
|
cannam@167
|
670 T6T = VSUB(T6N, T6S);
|
cannam@167
|
671 T8m = VADD(T6N, T6S);
|
cannam@167
|
672 }
|
cannam@167
|
673 {
|
cannam@167
|
674 V T6Z, T74, T7n, T7o;
|
cannam@167
|
675 T6Z = VSUB(T6X, T6Y);
|
cannam@167
|
676 T74 = VSUB(T70, T73);
|
cannam@167
|
677 T75 = VFMA(LDK(KP414213562), T74, T6Z);
|
cannam@167
|
678 T7j = VFNMS(LDK(KP414213562), T6Z, T74);
|
cannam@167
|
679 T7n = VADD(T6M, T6L);
|
cannam@167
|
680 T7o = VSUB(T6O, T6R);
|
cannam@167
|
681 T7p = VADD(T7n, T7o);
|
cannam@167
|
682 T8s = VSUB(T7o, T7n);
|
cannam@167
|
683 }
|
cannam@167
|
684 {
|
cannam@167
|
685 V T7r, T7s, T7a, T7f;
|
cannam@167
|
686 T7r = VADD(T70, T73);
|
cannam@167
|
687 T7s = VADD(T6X, T6Y);
|
cannam@167
|
688 T7t = VFMA(LDK(KP414213562), T7s, T7r);
|
cannam@167
|
689 T7z = VFNMS(LDK(KP414213562), T7r, T7s);
|
cannam@167
|
690 T7a = VSUB(T78, T79);
|
cannam@167
|
691 T7f = VSUB(T7b, T7e);
|
cannam@167
|
692 T7g = VFNMS(LDK(KP414213562), T7f, T7a);
|
cannam@167
|
693 T7k = VFMA(LDK(KP414213562), T7a, T7f);
|
cannam@167
|
694 }
|
cannam@167
|
695 {
|
cannam@167
|
696 V T6U, T7h, T8t, T8u;
|
cannam@167
|
697 T6U = VFMA(LDK(KP707106781), T6T, T6I);
|
cannam@167
|
698 T7h = VSUB(T75, T7g);
|
cannam@167
|
699 ST(&(ri[WS(rs, 22)]), VFNMS(LDK(KP923879532), T7h, T6U), ms, &(ri[0]));
|
cannam@167
|
700 ST(&(ri[WS(rs, 6)]), VFMA(LDK(KP923879532), T7h, T6U), ms, &(ri[0]));
|
cannam@167
|
701 T8t = VFMA(LDK(KP707106781), T8s, T8r);
|
cannam@167
|
702 T8u = VSUB(T7k, T7j);
|
cannam@167
|
703 ST(&(ii[WS(rs, 6)]), VFMA(LDK(KP923879532), T8u, T8t), ms, &(ii[0]));
|
cannam@167
|
704 ST(&(ii[WS(rs, 22)]), VFNMS(LDK(KP923879532), T8u, T8t), ms, &(ii[0]));
|
cannam@167
|
705 }
|
cannam@167
|
706 {
|
cannam@167
|
707 V T7i, T7l, T8v, T8w;
|
cannam@167
|
708 T7i = VFNMS(LDK(KP707106781), T6T, T6I);
|
cannam@167
|
709 T7l = VADD(T7j, T7k);
|
cannam@167
|
710 ST(&(ri[WS(rs, 14)]), VFNMS(LDK(KP923879532), T7l, T7i), ms, &(ri[0]));
|
cannam@167
|
711 ST(&(ri[WS(rs, 30)]), VFMA(LDK(KP923879532), T7l, T7i), ms, &(ri[0]));
|
cannam@167
|
712 T8v = VFNMS(LDK(KP707106781), T8s, T8r);
|
cannam@167
|
713 T8w = VADD(T75, T7g);
|
cannam@167
|
714 ST(&(ii[WS(rs, 14)]), VFNMS(LDK(KP923879532), T8w, T8v), ms, &(ii[0]));
|
cannam@167
|
715 ST(&(ii[WS(rs, 30)]), VFMA(LDK(KP923879532), T8w, T8v), ms, &(ii[0]));
|
cannam@167
|
716 }
|
cannam@167
|
717 {
|
cannam@167
|
718 V T7q, T7x, T8n, T8o;
|
cannam@167
|
719 T7q = VFMA(LDK(KP707106781), T7p, T7m);
|
cannam@167
|
720 T7x = VADD(T7t, T7w);
|
cannam@167
|
721 ST(&(ri[WS(rs, 18)]), VFNMS(LDK(KP923879532), T7x, T7q), ms, &(ri[0]));
|
cannam@167
|
722 ST(&(ri[WS(rs, 2)]), VFMA(LDK(KP923879532), T7x, T7q), ms, &(ri[0]));
|
cannam@167
|
723 T8n = VFMA(LDK(KP707106781), T8m, T8l);
|
cannam@167
|
724 T8o = VADD(T7z, T7A);
|
cannam@167
|
725 ST(&(ii[WS(rs, 2)]), VFMA(LDK(KP923879532), T8o, T8n), ms, &(ii[0]));
|
cannam@167
|
726 ST(&(ii[WS(rs, 18)]), VFNMS(LDK(KP923879532), T8o, T8n), ms, &(ii[0]));
|
cannam@167
|
727 }
|
cannam@167
|
728 {
|
cannam@167
|
729 V T7y, T7B, T8p, T8q;
|
cannam@167
|
730 T7y = VFNMS(LDK(KP707106781), T7p, T7m);
|
cannam@167
|
731 T7B = VSUB(T7z, T7A);
|
cannam@167
|
732 ST(&(ri[WS(rs, 26)]), VFNMS(LDK(KP923879532), T7B, T7y), ms, &(ri[0]));
|
cannam@167
|
733 ST(&(ri[WS(rs, 10)]), VFMA(LDK(KP923879532), T7B, T7y), ms, &(ri[0]));
|
cannam@167
|
734 T8p = VFNMS(LDK(KP707106781), T8m, T8l);
|
cannam@167
|
735 T8q = VSUB(T7w, T7t);
|
cannam@167
|
736 ST(&(ii[WS(rs, 10)]), VFMA(LDK(KP923879532), T8q, T8p), ms, &(ii[0]));
|
cannam@167
|
737 ST(&(ii[WS(rs, 26)]), VFNMS(LDK(KP923879532), T8q, T8p), ms, &(ii[0]));
|
cannam@167
|
738 }
|
cannam@167
|
739 }
|
cannam@167
|
740 {
|
cannam@167
|
741 V T3S, T5C, T4n, T8C, T8B, T8H, T5F, T8I, T5w, T5Q, T5A, T5M, T4X, T5P, T5z;
|
cannam@167
|
742 V T5J;
|
cannam@167
|
743 {
|
cannam@167
|
744 V T3C, T3R, T5D, T5E;
|
cannam@167
|
745 T3C = VADD(T3w, T3B);
|
cannam@167
|
746 T3R = VADD(T3J, T3Q);
|
cannam@167
|
747 T3S = VFNMS(LDK(KP707106781), T3R, T3C);
|
cannam@167
|
748 T5C = VFMA(LDK(KP707106781), T3R, T3C);
|
cannam@167
|
749 {
|
cannam@167
|
750 V T47, T4m, T8z, T8A;
|
cannam@167
|
751 T47 = VFNMS(LDK(KP414213562), T46, T3Z);
|
cannam@167
|
752 T4m = VFMA(LDK(KP414213562), T4l, T4e);
|
cannam@167
|
753 T4n = VSUB(T47, T4m);
|
cannam@167
|
754 T8C = VADD(T47, T4m);
|
cannam@167
|
755 T8z = VSUB(T8x, T8y);
|
cannam@167
|
756 T8A = VADD(T5T, T5U);
|
cannam@167
|
757 T8B = VFMA(LDK(KP707106781), T8A, T8z);
|
cannam@167
|
758 T8H = VFNMS(LDK(KP707106781), T8A, T8z);
|
cannam@167
|
759 }
|
cannam@167
|
760 T5D = VFMA(LDK(KP414213562), T3Z, T46);
|
cannam@167
|
761 T5E = VFNMS(LDK(KP414213562), T4e, T4l);
|
cannam@167
|
762 T5F = VADD(T5D, T5E);
|
cannam@167
|
763 T8I = VSUB(T5E, T5D);
|
cannam@167
|
764 {
|
cannam@167
|
765 V T5k, T5L, T5v, T5K, T5j, T5u;
|
cannam@167
|
766 T5j = VADD(T5b, T5i);
|
cannam@167
|
767 T5k = VFNMS(LDK(KP707106781), T5j, T54);
|
cannam@167
|
768 T5L = VFMA(LDK(KP707106781), T5j, T54);
|
cannam@167
|
769 T5u = VADD(T5s, T5t);
|
cannam@167
|
770 T5v = VFNMS(LDK(KP707106781), T5u, T5r);
|
cannam@167
|
771 T5K = VFMA(LDK(KP707106781), T5u, T5r);
|
cannam@167
|
772 T5w = VFNMS(LDK(KP668178637), T5v, T5k);
|
cannam@167
|
773 T5Q = VFMA(LDK(KP198912367), T5K, T5L);
|
cannam@167
|
774 T5A = VFMA(LDK(KP668178637), T5k, T5v);
|
cannam@167
|
775 T5M = VFNMS(LDK(KP198912367), T5L, T5K);
|
cannam@167
|
776 }
|
cannam@167
|
777 {
|
cannam@167
|
778 V T4L, T5I, T4W, T5H, T4K, T4V;
|
cannam@167
|
779 T4K = VADD(T4C, T4J);
|
cannam@167
|
780 T4L = VFNMS(LDK(KP707106781), T4K, T4v);
|
cannam@167
|
781 T5I = VFMA(LDK(KP707106781), T4K, T4v);
|
cannam@167
|
782 T4V = VADD(T4T, T4U);
|
cannam@167
|
783 T4W = VFNMS(LDK(KP707106781), T4V, T4S);
|
cannam@167
|
784 T5H = VFMA(LDK(KP707106781), T4V, T4S);
|
cannam@167
|
785 T4X = VFMA(LDK(KP668178637), T4W, T4L);
|
cannam@167
|
786 T5P = VFNMS(LDK(KP198912367), T5H, T5I);
|
cannam@167
|
787 T5z = VFNMS(LDK(KP668178637), T4L, T4W);
|
cannam@167
|
788 T5J = VFMA(LDK(KP198912367), T5I, T5H);
|
cannam@167
|
789 }
|
cannam@167
|
790 }
|
cannam@167
|
791 {
|
cannam@167
|
792 V T4o, T5x, T8J, T8K;
|
cannam@167
|
793 T4o = VFMA(LDK(KP923879532), T4n, T3S);
|
cannam@167
|
794 T5x = VSUB(T4X, T5w);
|
cannam@167
|
795 ST(&(ri[WS(rs, 21)]), VFNMS(LDK(KP831469612), T5x, T4o), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
796 ST(&(ri[WS(rs, 5)]), VFMA(LDK(KP831469612), T5x, T4o), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
797 T8J = VFMA(LDK(KP923879532), T8I, T8H);
|
cannam@167
|
798 T8K = VSUB(T5A, T5z);
|
cannam@167
|
799 ST(&(ii[WS(rs, 5)]), VFMA(LDK(KP831469612), T8K, T8J), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
800 ST(&(ii[WS(rs, 21)]), VFNMS(LDK(KP831469612), T8K, T8J), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
801 }
|
cannam@167
|
802 {
|
cannam@167
|
803 V T5y, T5B, T8L, T8M;
|
cannam@167
|
804 T5y = VFNMS(LDK(KP923879532), T4n, T3S);
|
cannam@167
|
805 T5B = VADD(T5z, T5A);
|
cannam@167
|
806 ST(&(ri[WS(rs, 13)]), VFNMS(LDK(KP831469612), T5B, T5y), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
807 ST(&(ri[WS(rs, 29)]), VFMA(LDK(KP831469612), T5B, T5y), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
808 T8L = VFNMS(LDK(KP923879532), T8I, T8H);
|
cannam@167
|
809 T8M = VADD(T4X, T5w);
|
cannam@167
|
810 ST(&(ii[WS(rs, 13)]), VFNMS(LDK(KP831469612), T8M, T8L), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
811 ST(&(ii[WS(rs, 29)]), VFMA(LDK(KP831469612), T8M, T8L), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
812 }
|
cannam@167
|
813 {
|
cannam@167
|
814 V T5G, T5N, T8D, T8E;
|
cannam@167
|
815 T5G = VFMA(LDK(KP923879532), T5F, T5C);
|
cannam@167
|
816 T5N = VADD(T5J, T5M);
|
cannam@167
|
817 ST(&(ri[WS(rs, 17)]), VFNMS(LDK(KP980785280), T5N, T5G), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
818 ST(&(ri[WS(rs, 1)]), VFMA(LDK(KP980785280), T5N, T5G), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
819 T8D = VFMA(LDK(KP923879532), T8C, T8B);
|
cannam@167
|
820 T8E = VADD(T5P, T5Q);
|
cannam@167
|
821 ST(&(ii[WS(rs, 1)]), VFMA(LDK(KP980785280), T8E, T8D), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
822 ST(&(ii[WS(rs, 17)]), VFNMS(LDK(KP980785280), T8E, T8D), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
823 }
|
cannam@167
|
824 {
|
cannam@167
|
825 V T5O, T5R, T8F, T8G;
|
cannam@167
|
826 T5O = VFNMS(LDK(KP923879532), T5F, T5C);
|
cannam@167
|
827 T5R = VSUB(T5P, T5Q);
|
cannam@167
|
828 ST(&(ri[WS(rs, 25)]), VFNMS(LDK(KP980785280), T5R, T5O), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
829 ST(&(ri[WS(rs, 9)]), VFMA(LDK(KP980785280), T5R, T5O), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
830 T8F = VFNMS(LDK(KP923879532), T8C, T8B);
|
cannam@167
|
831 T8G = VSUB(T5M, T5J);
|
cannam@167
|
832 ST(&(ii[WS(rs, 9)]), VFMA(LDK(KP980785280), T8G, T8F), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
833 ST(&(ii[WS(rs, 25)]), VFNMS(LDK(KP980785280), T8G, T8F), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
834 }
|
cannam@167
|
835 }
|
cannam@167
|
836 {
|
cannam@167
|
837 V T5W, T6o, T63, T8W, T8P, T8V, T6r, T8Q, T6i, T6C, T6m, T6y, T6b, T6B, T6l;
|
cannam@167
|
838 V T6v;
|
cannam@167
|
839 {
|
cannam@167
|
840 V T5S, T5V, T6p, T6q;
|
cannam@167
|
841 T5S = VSUB(T3w, T3B);
|
cannam@167
|
842 T5V = VSUB(T5T, T5U);
|
cannam@167
|
843 T5W = VFMA(LDK(KP707106781), T5V, T5S);
|
cannam@167
|
844 T6o = VFNMS(LDK(KP707106781), T5V, T5S);
|
cannam@167
|
845 {
|
cannam@167
|
846 V T5Z, T62, T8N, T8O;
|
cannam@167
|
847 T5Z = VFMA(LDK(KP414213562), T5Y, T5X);
|
cannam@167
|
848 T62 = VFNMS(LDK(KP414213562), T61, T60);
|
cannam@167
|
849 T63 = VSUB(T5Z, T62);
|
cannam@167
|
850 T8W = VADD(T5Z, T62);
|
cannam@167
|
851 T8N = VADD(T8y, T8x);
|
cannam@167
|
852 T8O = VSUB(T3Q, T3J);
|
cannam@167
|
853 T8P = VFMA(LDK(KP707106781), T8O, T8N);
|
cannam@167
|
854 T8V = VFNMS(LDK(KP707106781), T8O, T8N);
|
cannam@167
|
855 }
|
cannam@167
|
856 T6p = VFNMS(LDK(KP414213562), T5X, T5Y);
|
cannam@167
|
857 T6q = VFMA(LDK(KP414213562), T60, T61);
|
cannam@167
|
858 T6r = VADD(T6p, T6q);
|
cannam@167
|
859 T8Q = VSUB(T6q, T6p);
|
cannam@167
|
860 {
|
cannam@167
|
861 V T6e, T6x, T6h, T6w, T6d, T6g;
|
cannam@167
|
862 T6d = VSUB(T5i, T5b);
|
cannam@167
|
863 T6e = VFNMS(LDK(KP707106781), T6d, T6c);
|
cannam@167
|
864 T6x = VFMA(LDK(KP707106781), T6d, T6c);
|
cannam@167
|
865 T6g = VSUB(T5s, T5t);
|
cannam@167
|
866 T6h = VFNMS(LDK(KP707106781), T6g, T6f);
|
cannam@167
|
867 T6w = VFMA(LDK(KP707106781), T6g, T6f);
|
cannam@167
|
868 T6i = VFNMS(LDK(KP668178637), T6h, T6e);
|
cannam@167
|
869 T6C = VFMA(LDK(KP198912367), T6w, T6x);
|
cannam@167
|
870 T6m = VFMA(LDK(KP668178637), T6e, T6h);
|
cannam@167
|
871 T6y = VFNMS(LDK(KP198912367), T6x, T6w);
|
cannam@167
|
872 }
|
cannam@167
|
873 {
|
cannam@167
|
874 V T67, T6u, T6a, T6t, T66, T69;
|
cannam@167
|
875 T66 = VSUB(T4J, T4C);
|
cannam@167
|
876 T67 = VFNMS(LDK(KP707106781), T66, T65);
|
cannam@167
|
877 T6u = VFMA(LDK(KP707106781), T66, T65);
|
cannam@167
|
878 T69 = VSUB(T4T, T4U);
|
cannam@167
|
879 T6a = VFNMS(LDK(KP707106781), T69, T68);
|
cannam@167
|
880 T6t = VFMA(LDK(KP707106781), T69, T68);
|
cannam@167
|
881 T6b = VFMA(LDK(KP668178637), T6a, T67);
|
cannam@167
|
882 T6B = VFNMS(LDK(KP198912367), T6t, T6u);
|
cannam@167
|
883 T6l = VFNMS(LDK(KP668178637), T67, T6a);
|
cannam@167
|
884 T6v = VFMA(LDK(KP198912367), T6u, T6t);
|
cannam@167
|
885 }
|
cannam@167
|
886 }
|
cannam@167
|
887 {
|
cannam@167
|
888 V T64, T6j, T8R, T8S;
|
cannam@167
|
889 T64 = VFMA(LDK(KP923879532), T63, T5W);
|
cannam@167
|
890 T6j = VADD(T6b, T6i);
|
cannam@167
|
891 ST(&(ri[WS(rs, 19)]), VFNMS(LDK(KP831469612), T6j, T64), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
892 ST(&(ri[WS(rs, 3)]), VFMA(LDK(KP831469612), T6j, T64), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
893 T8R = VFMA(LDK(KP923879532), T8Q, T8P);
|
cannam@167
|
894 T8S = VADD(T6l, T6m);
|
cannam@167
|
895 ST(&(ii[WS(rs, 3)]), VFMA(LDK(KP831469612), T8S, T8R), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
896 ST(&(ii[WS(rs, 19)]), VFNMS(LDK(KP831469612), T8S, T8R), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
897 }
|
cannam@167
|
898 {
|
cannam@167
|
899 V T6k, T6n, T8T, T8U;
|
cannam@167
|
900 T6k = VFNMS(LDK(KP923879532), T63, T5W);
|
cannam@167
|
901 T6n = VSUB(T6l, T6m);
|
cannam@167
|
902 ST(&(ri[WS(rs, 27)]), VFNMS(LDK(KP831469612), T6n, T6k), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
903 ST(&(ri[WS(rs, 11)]), VFMA(LDK(KP831469612), T6n, T6k), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
904 T8T = VFNMS(LDK(KP923879532), T8Q, T8P);
|
cannam@167
|
905 T8U = VSUB(T6i, T6b);
|
cannam@167
|
906 ST(&(ii[WS(rs, 11)]), VFMA(LDK(KP831469612), T8U, T8T), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
907 ST(&(ii[WS(rs, 27)]), VFNMS(LDK(KP831469612), T8U, T8T), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
908 }
|
cannam@167
|
909 {
|
cannam@167
|
910 V T6s, T6z, T8X, T8Y;
|
cannam@167
|
911 T6s = VFNMS(LDK(KP923879532), T6r, T6o);
|
cannam@167
|
912 T6z = VSUB(T6v, T6y);
|
cannam@167
|
913 ST(&(ri[WS(rs, 23)]), VFNMS(LDK(KP980785280), T6z, T6s), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
914 ST(&(ri[WS(rs, 7)]), VFMA(LDK(KP980785280), T6z, T6s), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
915 T8X = VFNMS(LDK(KP923879532), T8W, T8V);
|
cannam@167
|
916 T8Y = VSUB(T6C, T6B);
|
cannam@167
|
917 ST(&(ii[WS(rs, 7)]), VFMA(LDK(KP980785280), T8Y, T8X), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
918 ST(&(ii[WS(rs, 23)]), VFNMS(LDK(KP980785280), T8Y, T8X), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
919 }
|
cannam@167
|
920 {
|
cannam@167
|
921 V T6A, T6D, T8Z, T90;
|
cannam@167
|
922 T6A = VFMA(LDK(KP923879532), T6r, T6o);
|
cannam@167
|
923 T6D = VADD(T6B, T6C);
|
cannam@167
|
924 ST(&(ri[WS(rs, 15)]), VFNMS(LDK(KP980785280), T6D, T6A), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
925 ST(&(ri[WS(rs, 31)]), VFMA(LDK(KP980785280), T6D, T6A), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
926 T8Z = VFMA(LDK(KP923879532), T8W, T8V);
|
cannam@167
|
927 T90 = VADD(T6v, T6y);
|
cannam@167
|
928 ST(&(ii[WS(rs, 15)]), VFNMS(LDK(KP980785280), T90, T8Z), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
929 ST(&(ii[WS(rs, 31)]), VFMA(LDK(KP980785280), T90, T8Z), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
930 }
|
cannam@167
|
931 }
|
cannam@167
|
932 }
|
cannam@167
|
933 }
|
cannam@167
|
934 VLEAVE();
|
cannam@167
|
935 }
|
cannam@167
|
936
|
cannam@167
|
937 static const tw_instr twinstr[] = {
|
cannam@167
|
938 VTW(0, 1),
|
cannam@167
|
939 VTW(0, 2),
|
cannam@167
|
940 VTW(0, 3),
|
cannam@167
|
941 VTW(0, 4),
|
cannam@167
|
942 VTW(0, 5),
|
cannam@167
|
943 VTW(0, 6),
|
cannam@167
|
944 VTW(0, 7),
|
cannam@167
|
945 VTW(0, 8),
|
cannam@167
|
946 VTW(0, 9),
|
cannam@167
|
947 VTW(0, 10),
|
cannam@167
|
948 VTW(0, 11),
|
cannam@167
|
949 VTW(0, 12),
|
cannam@167
|
950 VTW(0, 13),
|
cannam@167
|
951 VTW(0, 14),
|
cannam@167
|
952 VTW(0, 15),
|
cannam@167
|
953 VTW(0, 16),
|
cannam@167
|
954 VTW(0, 17),
|
cannam@167
|
955 VTW(0, 18),
|
cannam@167
|
956 VTW(0, 19),
|
cannam@167
|
957 VTW(0, 20),
|
cannam@167
|
958 VTW(0, 21),
|
cannam@167
|
959 VTW(0, 22),
|
cannam@167
|
960 VTW(0, 23),
|
cannam@167
|
961 VTW(0, 24),
|
cannam@167
|
962 VTW(0, 25),
|
cannam@167
|
963 VTW(0, 26),
|
cannam@167
|
964 VTW(0, 27),
|
cannam@167
|
965 VTW(0, 28),
|
cannam@167
|
966 VTW(0, 29),
|
cannam@167
|
967 VTW(0, 30),
|
cannam@167
|
968 VTW(0, 31),
|
cannam@167
|
969 {TW_NEXT, (2 * VL), 0}
|
cannam@167
|
970 };
|
cannam@167
|
971
|
cannam@167
|
972 static const ct_desc desc = { 32, XSIMD_STRING("t1sv_32"), twinstr, &GENUS, {236, 62, 198, 0}, 0, 0, 0 };
|
cannam@167
|
973
|
cannam@167
|
974 void XSIMD(codelet_t1sv_32) (planner *p) {
|
cannam@167
|
975 X(kdft_dit_register) (p, t1sv_32, &desc);
|
cannam@167
|
976 }
|
cannam@167
|
977 #else
|
cannam@167
|
978
|
cannam@167
|
979 /* Generated by: ../../../genfft/gen_twiddle.native -simd -compact -variables 4 -pipeline-latency 8 -n 32 -name t1sv_32 -include dft/simd/ts.h */
|
cannam@167
|
980
|
cannam@167
|
981 /*
|
cannam@167
|
982 * This function contains 434 FP additions, 208 FP multiplications,
|
cannam@167
|
983 * (or, 340 additions, 114 multiplications, 94 fused multiply/add),
|
cannam@167
|
984 * 96 stack variables, 7 constants, and 128 memory accesses
|
cannam@167
|
985 */
|
cannam@167
|
986 #include "dft/simd/ts.h"
|
cannam@167
|
987
|
cannam@167
|
988 static void t1sv_32(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
989 {
|
cannam@167
|
990 DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
|
cannam@167
|
991 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
cannam@167
|
992 DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
|
cannam@167
|
993 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
cannam@167
|
994 DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
cannam@167
|
995 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
996 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
997 {
|
cannam@167
|
998 INT m;
|
cannam@167
|
999 for (m = mb, W = W + (mb * 62); m < me; m = m + (2 * VL), ri = ri + ((2 * VL) * ms), ii = ii + ((2 * VL) * ms), W = W + ((2 * VL) * 62), MAKE_VOLATILE_STRIDE(64, rs)) {
|
cannam@167
|
1000 V Tj, T5F, T7C, T7Q, T35, T4T, T78, T7m, T1Q, T61, T5Y, T6J, T3K, T59, T41;
|
cannam@167
|
1001 V T56, T2B, T67, T6e, T6O, T4b, T5d, T4s, T5g, TG, T7l, T5I, T73, T3a, T4U;
|
cannam@167
|
1002 V T3f, T4V, T14, T5N, T5M, T6E, T3m, T4Y, T3r, T4Z, T1r, T5P, T5S, T6F, T3x;
|
cannam@167
|
1003 V T51, T3C, T52, T2d, T5Z, T64, T6K, T3V, T57, T44, T5a, T2Y, T6f, T6a, T6P;
|
cannam@167
|
1004 V T4m, T5h, T4v, T5e;
|
cannam@167
|
1005 {
|
cannam@167
|
1006 V T1, T76, T6, T75, Tc, T32, Th, T33;
|
cannam@167
|
1007 T1 = LD(&(ri[0]), ms, &(ri[0]));
|
cannam@167
|
1008 T76 = LD(&(ii[0]), ms, &(ii[0]));
|
cannam@167
|
1009 {
|
cannam@167
|
1010 V T3, T5, T2, T4;
|
cannam@167
|
1011 T3 = LD(&(ri[WS(rs, 16)]), ms, &(ri[0]));
|
cannam@167
|
1012 T5 = LD(&(ii[WS(rs, 16)]), ms, &(ii[0]));
|
cannam@167
|
1013 T2 = LDW(&(W[TWVL * 30]));
|
cannam@167
|
1014 T4 = LDW(&(W[TWVL * 31]));
|
cannam@167
|
1015 T6 = VFMA(T2, T3, VMUL(T4, T5));
|
cannam@167
|
1016 T75 = VFNMS(T4, T3, VMUL(T2, T5));
|
cannam@167
|
1017 }
|
cannam@167
|
1018 {
|
cannam@167
|
1019 V T9, Tb, T8, Ta;
|
cannam@167
|
1020 T9 = LD(&(ri[WS(rs, 8)]), ms, &(ri[0]));
|
cannam@167
|
1021 Tb = LD(&(ii[WS(rs, 8)]), ms, &(ii[0]));
|
cannam@167
|
1022 T8 = LDW(&(W[TWVL * 14]));
|
cannam@167
|
1023 Ta = LDW(&(W[TWVL * 15]));
|
cannam@167
|
1024 Tc = VFMA(T8, T9, VMUL(Ta, Tb));
|
cannam@167
|
1025 T32 = VFNMS(Ta, T9, VMUL(T8, Tb));
|
cannam@167
|
1026 }
|
cannam@167
|
1027 {
|
cannam@167
|
1028 V Te, Tg, Td, Tf;
|
cannam@167
|
1029 Te = LD(&(ri[WS(rs, 24)]), ms, &(ri[0]));
|
cannam@167
|
1030 Tg = LD(&(ii[WS(rs, 24)]), ms, &(ii[0]));
|
cannam@167
|
1031 Td = LDW(&(W[TWVL * 46]));
|
cannam@167
|
1032 Tf = LDW(&(W[TWVL * 47]));
|
cannam@167
|
1033 Th = VFMA(Td, Te, VMUL(Tf, Tg));
|
cannam@167
|
1034 T33 = VFNMS(Tf, Te, VMUL(Td, Tg));
|
cannam@167
|
1035 }
|
cannam@167
|
1036 {
|
cannam@167
|
1037 V T7, Ti, T7A, T7B;
|
cannam@167
|
1038 T7 = VADD(T1, T6);
|
cannam@167
|
1039 Ti = VADD(Tc, Th);
|
cannam@167
|
1040 Tj = VADD(T7, Ti);
|
cannam@167
|
1041 T5F = VSUB(T7, Ti);
|
cannam@167
|
1042 T7A = VSUB(T76, T75);
|
cannam@167
|
1043 T7B = VSUB(Tc, Th);
|
cannam@167
|
1044 T7C = VSUB(T7A, T7B);
|
cannam@167
|
1045 T7Q = VADD(T7B, T7A);
|
cannam@167
|
1046 }
|
cannam@167
|
1047 {
|
cannam@167
|
1048 V T31, T34, T74, T77;
|
cannam@167
|
1049 T31 = VSUB(T1, T6);
|
cannam@167
|
1050 T34 = VSUB(T32, T33);
|
cannam@167
|
1051 T35 = VSUB(T31, T34);
|
cannam@167
|
1052 T4T = VADD(T31, T34);
|
cannam@167
|
1053 T74 = VADD(T32, T33);
|
cannam@167
|
1054 T77 = VADD(T75, T76);
|
cannam@167
|
1055 T78 = VADD(T74, T77);
|
cannam@167
|
1056 T7m = VSUB(T77, T74);
|
cannam@167
|
1057 }
|
cannam@167
|
1058 }
|
cannam@167
|
1059 {
|
cannam@167
|
1060 V T1y, T3G, T1O, T3Z, T1D, T3H, T1J, T3Y;
|
cannam@167
|
1061 {
|
cannam@167
|
1062 V T1v, T1x, T1u, T1w;
|
cannam@167
|
1063 T1v = LD(&(ri[WS(rs, 1)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1064 T1x = LD(&(ii[WS(rs, 1)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1065 T1u = LDW(&(W[0]));
|
cannam@167
|
1066 T1w = LDW(&(W[TWVL * 1]));
|
cannam@167
|
1067 T1y = VFMA(T1u, T1v, VMUL(T1w, T1x));
|
cannam@167
|
1068 T3G = VFNMS(T1w, T1v, VMUL(T1u, T1x));
|
cannam@167
|
1069 }
|
cannam@167
|
1070 {
|
cannam@167
|
1071 V T1L, T1N, T1K, T1M;
|
cannam@167
|
1072 T1L = LD(&(ri[WS(rs, 25)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1073 T1N = LD(&(ii[WS(rs, 25)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1074 T1K = LDW(&(W[TWVL * 48]));
|
cannam@167
|
1075 T1M = LDW(&(W[TWVL * 49]));
|
cannam@167
|
1076 T1O = VFMA(T1K, T1L, VMUL(T1M, T1N));
|
cannam@167
|
1077 T3Z = VFNMS(T1M, T1L, VMUL(T1K, T1N));
|
cannam@167
|
1078 }
|
cannam@167
|
1079 {
|
cannam@167
|
1080 V T1A, T1C, T1z, T1B;
|
cannam@167
|
1081 T1A = LD(&(ri[WS(rs, 17)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1082 T1C = LD(&(ii[WS(rs, 17)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1083 T1z = LDW(&(W[TWVL * 32]));
|
cannam@167
|
1084 T1B = LDW(&(W[TWVL * 33]));
|
cannam@167
|
1085 T1D = VFMA(T1z, T1A, VMUL(T1B, T1C));
|
cannam@167
|
1086 T3H = VFNMS(T1B, T1A, VMUL(T1z, T1C));
|
cannam@167
|
1087 }
|
cannam@167
|
1088 {
|
cannam@167
|
1089 V T1G, T1I, T1F, T1H;
|
cannam@167
|
1090 T1G = LD(&(ri[WS(rs, 9)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1091 T1I = LD(&(ii[WS(rs, 9)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1092 T1F = LDW(&(W[TWVL * 16]));
|
cannam@167
|
1093 T1H = LDW(&(W[TWVL * 17]));
|
cannam@167
|
1094 T1J = VFMA(T1F, T1G, VMUL(T1H, T1I));
|
cannam@167
|
1095 T3Y = VFNMS(T1H, T1G, VMUL(T1F, T1I));
|
cannam@167
|
1096 }
|
cannam@167
|
1097 {
|
cannam@167
|
1098 V T1E, T1P, T5W, T5X;
|
cannam@167
|
1099 T1E = VADD(T1y, T1D);
|
cannam@167
|
1100 T1P = VADD(T1J, T1O);
|
cannam@167
|
1101 T1Q = VADD(T1E, T1P);
|
cannam@167
|
1102 T61 = VSUB(T1E, T1P);
|
cannam@167
|
1103 T5W = VADD(T3G, T3H);
|
cannam@167
|
1104 T5X = VADD(T3Y, T3Z);
|
cannam@167
|
1105 T5Y = VSUB(T5W, T5X);
|
cannam@167
|
1106 T6J = VADD(T5W, T5X);
|
cannam@167
|
1107 }
|
cannam@167
|
1108 {
|
cannam@167
|
1109 V T3I, T3J, T3X, T40;
|
cannam@167
|
1110 T3I = VSUB(T3G, T3H);
|
cannam@167
|
1111 T3J = VSUB(T1J, T1O);
|
cannam@167
|
1112 T3K = VADD(T3I, T3J);
|
cannam@167
|
1113 T59 = VSUB(T3I, T3J);
|
cannam@167
|
1114 T3X = VSUB(T1y, T1D);
|
cannam@167
|
1115 T40 = VSUB(T3Y, T3Z);
|
cannam@167
|
1116 T41 = VSUB(T3X, T40);
|
cannam@167
|
1117 T56 = VADD(T3X, T40);
|
cannam@167
|
1118 }
|
cannam@167
|
1119 }
|
cannam@167
|
1120 {
|
cannam@167
|
1121 V T2j, T4o, T2z, T49, T2o, T4p, T2u, T48;
|
cannam@167
|
1122 {
|
cannam@167
|
1123 V T2g, T2i, T2f, T2h;
|
cannam@167
|
1124 T2g = LD(&(ri[WS(rs, 31)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1125 T2i = LD(&(ii[WS(rs, 31)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1126 T2f = LDW(&(W[TWVL * 60]));
|
cannam@167
|
1127 T2h = LDW(&(W[TWVL * 61]));
|
cannam@167
|
1128 T2j = VFMA(T2f, T2g, VMUL(T2h, T2i));
|
cannam@167
|
1129 T4o = VFNMS(T2h, T2g, VMUL(T2f, T2i));
|
cannam@167
|
1130 }
|
cannam@167
|
1131 {
|
cannam@167
|
1132 V T2w, T2y, T2v, T2x;
|
cannam@167
|
1133 T2w = LD(&(ri[WS(rs, 23)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1134 T2y = LD(&(ii[WS(rs, 23)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1135 T2v = LDW(&(W[TWVL * 44]));
|
cannam@167
|
1136 T2x = LDW(&(W[TWVL * 45]));
|
cannam@167
|
1137 T2z = VFMA(T2v, T2w, VMUL(T2x, T2y));
|
cannam@167
|
1138 T49 = VFNMS(T2x, T2w, VMUL(T2v, T2y));
|
cannam@167
|
1139 }
|
cannam@167
|
1140 {
|
cannam@167
|
1141 V T2l, T2n, T2k, T2m;
|
cannam@167
|
1142 T2l = LD(&(ri[WS(rs, 15)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1143 T2n = LD(&(ii[WS(rs, 15)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1144 T2k = LDW(&(W[TWVL * 28]));
|
cannam@167
|
1145 T2m = LDW(&(W[TWVL * 29]));
|
cannam@167
|
1146 T2o = VFMA(T2k, T2l, VMUL(T2m, T2n));
|
cannam@167
|
1147 T4p = VFNMS(T2m, T2l, VMUL(T2k, T2n));
|
cannam@167
|
1148 }
|
cannam@167
|
1149 {
|
cannam@167
|
1150 V T2r, T2t, T2q, T2s;
|
cannam@167
|
1151 T2r = LD(&(ri[WS(rs, 7)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1152 T2t = LD(&(ii[WS(rs, 7)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1153 T2q = LDW(&(W[TWVL * 12]));
|
cannam@167
|
1154 T2s = LDW(&(W[TWVL * 13]));
|
cannam@167
|
1155 T2u = VFMA(T2q, T2r, VMUL(T2s, T2t));
|
cannam@167
|
1156 T48 = VFNMS(T2s, T2r, VMUL(T2q, T2t));
|
cannam@167
|
1157 }
|
cannam@167
|
1158 {
|
cannam@167
|
1159 V T2p, T2A, T6c, T6d;
|
cannam@167
|
1160 T2p = VADD(T2j, T2o);
|
cannam@167
|
1161 T2A = VADD(T2u, T2z);
|
cannam@167
|
1162 T2B = VADD(T2p, T2A);
|
cannam@167
|
1163 T67 = VSUB(T2p, T2A);
|
cannam@167
|
1164 T6c = VADD(T4o, T4p);
|
cannam@167
|
1165 T6d = VADD(T48, T49);
|
cannam@167
|
1166 T6e = VSUB(T6c, T6d);
|
cannam@167
|
1167 T6O = VADD(T6c, T6d);
|
cannam@167
|
1168 }
|
cannam@167
|
1169 {
|
cannam@167
|
1170 V T47, T4a, T4q, T4r;
|
cannam@167
|
1171 T47 = VSUB(T2j, T2o);
|
cannam@167
|
1172 T4a = VSUB(T48, T49);
|
cannam@167
|
1173 T4b = VSUB(T47, T4a);
|
cannam@167
|
1174 T5d = VADD(T47, T4a);
|
cannam@167
|
1175 T4q = VSUB(T4o, T4p);
|
cannam@167
|
1176 T4r = VSUB(T2u, T2z);
|
cannam@167
|
1177 T4s = VADD(T4q, T4r);
|
cannam@167
|
1178 T5g = VSUB(T4q, T4r);
|
cannam@167
|
1179 }
|
cannam@167
|
1180 }
|
cannam@167
|
1181 {
|
cannam@167
|
1182 V To, T36, TE, T3d, Tt, T37, Tz, T3c;
|
cannam@167
|
1183 {
|
cannam@167
|
1184 V Tl, Tn, Tk, Tm;
|
cannam@167
|
1185 Tl = LD(&(ri[WS(rs, 4)]), ms, &(ri[0]));
|
cannam@167
|
1186 Tn = LD(&(ii[WS(rs, 4)]), ms, &(ii[0]));
|
cannam@167
|
1187 Tk = LDW(&(W[TWVL * 6]));
|
cannam@167
|
1188 Tm = LDW(&(W[TWVL * 7]));
|
cannam@167
|
1189 To = VFMA(Tk, Tl, VMUL(Tm, Tn));
|
cannam@167
|
1190 T36 = VFNMS(Tm, Tl, VMUL(Tk, Tn));
|
cannam@167
|
1191 }
|
cannam@167
|
1192 {
|
cannam@167
|
1193 V TB, TD, TA, TC;
|
cannam@167
|
1194 TB = LD(&(ri[WS(rs, 12)]), ms, &(ri[0]));
|
cannam@167
|
1195 TD = LD(&(ii[WS(rs, 12)]), ms, &(ii[0]));
|
cannam@167
|
1196 TA = LDW(&(W[TWVL * 22]));
|
cannam@167
|
1197 TC = LDW(&(W[TWVL * 23]));
|
cannam@167
|
1198 TE = VFMA(TA, TB, VMUL(TC, TD));
|
cannam@167
|
1199 T3d = VFNMS(TC, TB, VMUL(TA, TD));
|
cannam@167
|
1200 }
|
cannam@167
|
1201 {
|
cannam@167
|
1202 V Tq, Ts, Tp, Tr;
|
cannam@167
|
1203 Tq = LD(&(ri[WS(rs, 20)]), ms, &(ri[0]));
|
cannam@167
|
1204 Ts = LD(&(ii[WS(rs, 20)]), ms, &(ii[0]));
|
cannam@167
|
1205 Tp = LDW(&(W[TWVL * 38]));
|
cannam@167
|
1206 Tr = LDW(&(W[TWVL * 39]));
|
cannam@167
|
1207 Tt = VFMA(Tp, Tq, VMUL(Tr, Ts));
|
cannam@167
|
1208 T37 = VFNMS(Tr, Tq, VMUL(Tp, Ts));
|
cannam@167
|
1209 }
|
cannam@167
|
1210 {
|
cannam@167
|
1211 V Tw, Ty, Tv, Tx;
|
cannam@167
|
1212 Tw = LD(&(ri[WS(rs, 28)]), ms, &(ri[0]));
|
cannam@167
|
1213 Ty = LD(&(ii[WS(rs, 28)]), ms, &(ii[0]));
|
cannam@167
|
1214 Tv = LDW(&(W[TWVL * 54]));
|
cannam@167
|
1215 Tx = LDW(&(W[TWVL * 55]));
|
cannam@167
|
1216 Tz = VFMA(Tv, Tw, VMUL(Tx, Ty));
|
cannam@167
|
1217 T3c = VFNMS(Tx, Tw, VMUL(Tv, Ty));
|
cannam@167
|
1218 }
|
cannam@167
|
1219 {
|
cannam@167
|
1220 V Tu, TF, T5G, T5H;
|
cannam@167
|
1221 Tu = VADD(To, Tt);
|
cannam@167
|
1222 TF = VADD(Tz, TE);
|
cannam@167
|
1223 TG = VADD(Tu, TF);
|
cannam@167
|
1224 T7l = VSUB(TF, Tu);
|
cannam@167
|
1225 T5G = VADD(T36, T37);
|
cannam@167
|
1226 T5H = VADD(T3c, T3d);
|
cannam@167
|
1227 T5I = VSUB(T5G, T5H);
|
cannam@167
|
1228 T73 = VADD(T5G, T5H);
|
cannam@167
|
1229 }
|
cannam@167
|
1230 {
|
cannam@167
|
1231 V T38, T39, T3b, T3e;
|
cannam@167
|
1232 T38 = VSUB(T36, T37);
|
cannam@167
|
1233 T39 = VSUB(To, Tt);
|
cannam@167
|
1234 T3a = VSUB(T38, T39);
|
cannam@167
|
1235 T4U = VADD(T39, T38);
|
cannam@167
|
1236 T3b = VSUB(Tz, TE);
|
cannam@167
|
1237 T3e = VSUB(T3c, T3d);
|
cannam@167
|
1238 T3f = VADD(T3b, T3e);
|
cannam@167
|
1239 T4V = VSUB(T3b, T3e);
|
cannam@167
|
1240 }
|
cannam@167
|
1241 }
|
cannam@167
|
1242 {
|
cannam@167
|
1243 V TM, T3i, T12, T3p, TR, T3j, TX, T3o;
|
cannam@167
|
1244 {
|
cannam@167
|
1245 V TJ, TL, TI, TK;
|
cannam@167
|
1246 TJ = LD(&(ri[WS(rs, 2)]), ms, &(ri[0]));
|
cannam@167
|
1247 TL = LD(&(ii[WS(rs, 2)]), ms, &(ii[0]));
|
cannam@167
|
1248 TI = LDW(&(W[TWVL * 2]));
|
cannam@167
|
1249 TK = LDW(&(W[TWVL * 3]));
|
cannam@167
|
1250 TM = VFMA(TI, TJ, VMUL(TK, TL));
|
cannam@167
|
1251 T3i = VFNMS(TK, TJ, VMUL(TI, TL));
|
cannam@167
|
1252 }
|
cannam@167
|
1253 {
|
cannam@167
|
1254 V TZ, T11, TY, T10;
|
cannam@167
|
1255 TZ = LD(&(ri[WS(rs, 26)]), ms, &(ri[0]));
|
cannam@167
|
1256 T11 = LD(&(ii[WS(rs, 26)]), ms, &(ii[0]));
|
cannam@167
|
1257 TY = LDW(&(W[TWVL * 50]));
|
cannam@167
|
1258 T10 = LDW(&(W[TWVL * 51]));
|
cannam@167
|
1259 T12 = VFMA(TY, TZ, VMUL(T10, T11));
|
cannam@167
|
1260 T3p = VFNMS(T10, TZ, VMUL(TY, T11));
|
cannam@167
|
1261 }
|
cannam@167
|
1262 {
|
cannam@167
|
1263 V TO, TQ, TN, TP;
|
cannam@167
|
1264 TO = LD(&(ri[WS(rs, 18)]), ms, &(ri[0]));
|
cannam@167
|
1265 TQ = LD(&(ii[WS(rs, 18)]), ms, &(ii[0]));
|
cannam@167
|
1266 TN = LDW(&(W[TWVL * 34]));
|
cannam@167
|
1267 TP = LDW(&(W[TWVL * 35]));
|
cannam@167
|
1268 TR = VFMA(TN, TO, VMUL(TP, TQ));
|
cannam@167
|
1269 T3j = VFNMS(TP, TO, VMUL(TN, TQ));
|
cannam@167
|
1270 }
|
cannam@167
|
1271 {
|
cannam@167
|
1272 V TU, TW, TT, TV;
|
cannam@167
|
1273 TU = LD(&(ri[WS(rs, 10)]), ms, &(ri[0]));
|
cannam@167
|
1274 TW = LD(&(ii[WS(rs, 10)]), ms, &(ii[0]));
|
cannam@167
|
1275 TT = LDW(&(W[TWVL * 18]));
|
cannam@167
|
1276 TV = LDW(&(W[TWVL * 19]));
|
cannam@167
|
1277 TX = VFMA(TT, TU, VMUL(TV, TW));
|
cannam@167
|
1278 T3o = VFNMS(TV, TU, VMUL(TT, TW));
|
cannam@167
|
1279 }
|
cannam@167
|
1280 {
|
cannam@167
|
1281 V TS, T13, T5K, T5L;
|
cannam@167
|
1282 TS = VADD(TM, TR);
|
cannam@167
|
1283 T13 = VADD(TX, T12);
|
cannam@167
|
1284 T14 = VADD(TS, T13);
|
cannam@167
|
1285 T5N = VSUB(TS, T13);
|
cannam@167
|
1286 T5K = VADD(T3i, T3j);
|
cannam@167
|
1287 T5L = VADD(T3o, T3p);
|
cannam@167
|
1288 T5M = VSUB(T5K, T5L);
|
cannam@167
|
1289 T6E = VADD(T5K, T5L);
|
cannam@167
|
1290 }
|
cannam@167
|
1291 {
|
cannam@167
|
1292 V T3k, T3l, T3n, T3q;
|
cannam@167
|
1293 T3k = VSUB(T3i, T3j);
|
cannam@167
|
1294 T3l = VSUB(TX, T12);
|
cannam@167
|
1295 T3m = VADD(T3k, T3l);
|
cannam@167
|
1296 T4Y = VSUB(T3k, T3l);
|
cannam@167
|
1297 T3n = VSUB(TM, TR);
|
cannam@167
|
1298 T3q = VSUB(T3o, T3p);
|
cannam@167
|
1299 T3r = VSUB(T3n, T3q);
|
cannam@167
|
1300 T4Z = VADD(T3n, T3q);
|
cannam@167
|
1301 }
|
cannam@167
|
1302 }
|
cannam@167
|
1303 {
|
cannam@167
|
1304 V T19, T3t, T1p, T3A, T1e, T3u, T1k, T3z;
|
cannam@167
|
1305 {
|
cannam@167
|
1306 V T16, T18, T15, T17;
|
cannam@167
|
1307 T16 = LD(&(ri[WS(rs, 30)]), ms, &(ri[0]));
|
cannam@167
|
1308 T18 = LD(&(ii[WS(rs, 30)]), ms, &(ii[0]));
|
cannam@167
|
1309 T15 = LDW(&(W[TWVL * 58]));
|
cannam@167
|
1310 T17 = LDW(&(W[TWVL * 59]));
|
cannam@167
|
1311 T19 = VFMA(T15, T16, VMUL(T17, T18));
|
cannam@167
|
1312 T3t = VFNMS(T17, T16, VMUL(T15, T18));
|
cannam@167
|
1313 }
|
cannam@167
|
1314 {
|
cannam@167
|
1315 V T1m, T1o, T1l, T1n;
|
cannam@167
|
1316 T1m = LD(&(ri[WS(rs, 22)]), ms, &(ri[0]));
|
cannam@167
|
1317 T1o = LD(&(ii[WS(rs, 22)]), ms, &(ii[0]));
|
cannam@167
|
1318 T1l = LDW(&(W[TWVL * 42]));
|
cannam@167
|
1319 T1n = LDW(&(W[TWVL * 43]));
|
cannam@167
|
1320 T1p = VFMA(T1l, T1m, VMUL(T1n, T1o));
|
cannam@167
|
1321 T3A = VFNMS(T1n, T1m, VMUL(T1l, T1o));
|
cannam@167
|
1322 }
|
cannam@167
|
1323 {
|
cannam@167
|
1324 V T1b, T1d, T1a, T1c;
|
cannam@167
|
1325 T1b = LD(&(ri[WS(rs, 14)]), ms, &(ri[0]));
|
cannam@167
|
1326 T1d = LD(&(ii[WS(rs, 14)]), ms, &(ii[0]));
|
cannam@167
|
1327 T1a = LDW(&(W[TWVL * 26]));
|
cannam@167
|
1328 T1c = LDW(&(W[TWVL * 27]));
|
cannam@167
|
1329 T1e = VFMA(T1a, T1b, VMUL(T1c, T1d));
|
cannam@167
|
1330 T3u = VFNMS(T1c, T1b, VMUL(T1a, T1d));
|
cannam@167
|
1331 }
|
cannam@167
|
1332 {
|
cannam@167
|
1333 V T1h, T1j, T1g, T1i;
|
cannam@167
|
1334 T1h = LD(&(ri[WS(rs, 6)]), ms, &(ri[0]));
|
cannam@167
|
1335 T1j = LD(&(ii[WS(rs, 6)]), ms, &(ii[0]));
|
cannam@167
|
1336 T1g = LDW(&(W[TWVL * 10]));
|
cannam@167
|
1337 T1i = LDW(&(W[TWVL * 11]));
|
cannam@167
|
1338 T1k = VFMA(T1g, T1h, VMUL(T1i, T1j));
|
cannam@167
|
1339 T3z = VFNMS(T1i, T1h, VMUL(T1g, T1j));
|
cannam@167
|
1340 }
|
cannam@167
|
1341 {
|
cannam@167
|
1342 V T1f, T1q, T5Q, T5R;
|
cannam@167
|
1343 T1f = VADD(T19, T1e);
|
cannam@167
|
1344 T1q = VADD(T1k, T1p);
|
cannam@167
|
1345 T1r = VADD(T1f, T1q);
|
cannam@167
|
1346 T5P = VSUB(T1f, T1q);
|
cannam@167
|
1347 T5Q = VADD(T3t, T3u);
|
cannam@167
|
1348 T5R = VADD(T3z, T3A);
|
cannam@167
|
1349 T5S = VSUB(T5Q, T5R);
|
cannam@167
|
1350 T6F = VADD(T5Q, T5R);
|
cannam@167
|
1351 }
|
cannam@167
|
1352 {
|
cannam@167
|
1353 V T3v, T3w, T3y, T3B;
|
cannam@167
|
1354 T3v = VSUB(T3t, T3u);
|
cannam@167
|
1355 T3w = VSUB(T1k, T1p);
|
cannam@167
|
1356 T3x = VADD(T3v, T3w);
|
cannam@167
|
1357 T51 = VSUB(T3v, T3w);
|
cannam@167
|
1358 T3y = VSUB(T19, T1e);
|
cannam@167
|
1359 T3B = VSUB(T3z, T3A);
|
cannam@167
|
1360 T3C = VSUB(T3y, T3B);
|
cannam@167
|
1361 T52 = VADD(T3y, T3B);
|
cannam@167
|
1362 }
|
cannam@167
|
1363 }
|
cannam@167
|
1364 {
|
cannam@167
|
1365 V T1V, T3R, T20, T3S, T3Q, T3T, T26, T3M, T2b, T3N, T3L, T3O;
|
cannam@167
|
1366 {
|
cannam@167
|
1367 V T1S, T1U, T1R, T1T;
|
cannam@167
|
1368 T1S = LD(&(ri[WS(rs, 5)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1369 T1U = LD(&(ii[WS(rs, 5)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1370 T1R = LDW(&(W[TWVL * 8]));
|
cannam@167
|
1371 T1T = LDW(&(W[TWVL * 9]));
|
cannam@167
|
1372 T1V = VFMA(T1R, T1S, VMUL(T1T, T1U));
|
cannam@167
|
1373 T3R = VFNMS(T1T, T1S, VMUL(T1R, T1U));
|
cannam@167
|
1374 }
|
cannam@167
|
1375 {
|
cannam@167
|
1376 V T1X, T1Z, T1W, T1Y;
|
cannam@167
|
1377 T1X = LD(&(ri[WS(rs, 21)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1378 T1Z = LD(&(ii[WS(rs, 21)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1379 T1W = LDW(&(W[TWVL * 40]));
|
cannam@167
|
1380 T1Y = LDW(&(W[TWVL * 41]));
|
cannam@167
|
1381 T20 = VFMA(T1W, T1X, VMUL(T1Y, T1Z));
|
cannam@167
|
1382 T3S = VFNMS(T1Y, T1X, VMUL(T1W, T1Z));
|
cannam@167
|
1383 }
|
cannam@167
|
1384 T3Q = VSUB(T1V, T20);
|
cannam@167
|
1385 T3T = VSUB(T3R, T3S);
|
cannam@167
|
1386 {
|
cannam@167
|
1387 V T23, T25, T22, T24;
|
cannam@167
|
1388 T23 = LD(&(ri[WS(rs, 29)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1389 T25 = LD(&(ii[WS(rs, 29)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1390 T22 = LDW(&(W[TWVL * 56]));
|
cannam@167
|
1391 T24 = LDW(&(W[TWVL * 57]));
|
cannam@167
|
1392 T26 = VFMA(T22, T23, VMUL(T24, T25));
|
cannam@167
|
1393 T3M = VFNMS(T24, T23, VMUL(T22, T25));
|
cannam@167
|
1394 }
|
cannam@167
|
1395 {
|
cannam@167
|
1396 V T28, T2a, T27, T29;
|
cannam@167
|
1397 T28 = LD(&(ri[WS(rs, 13)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1398 T2a = LD(&(ii[WS(rs, 13)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1399 T27 = LDW(&(W[TWVL * 24]));
|
cannam@167
|
1400 T29 = LDW(&(W[TWVL * 25]));
|
cannam@167
|
1401 T2b = VFMA(T27, T28, VMUL(T29, T2a));
|
cannam@167
|
1402 T3N = VFNMS(T29, T28, VMUL(T27, T2a));
|
cannam@167
|
1403 }
|
cannam@167
|
1404 T3L = VSUB(T26, T2b);
|
cannam@167
|
1405 T3O = VSUB(T3M, T3N);
|
cannam@167
|
1406 {
|
cannam@167
|
1407 V T21, T2c, T62, T63;
|
cannam@167
|
1408 T21 = VADD(T1V, T20);
|
cannam@167
|
1409 T2c = VADD(T26, T2b);
|
cannam@167
|
1410 T2d = VADD(T21, T2c);
|
cannam@167
|
1411 T5Z = VSUB(T2c, T21);
|
cannam@167
|
1412 T62 = VADD(T3R, T3S);
|
cannam@167
|
1413 T63 = VADD(T3M, T3N);
|
cannam@167
|
1414 T64 = VSUB(T62, T63);
|
cannam@167
|
1415 T6K = VADD(T62, T63);
|
cannam@167
|
1416 }
|
cannam@167
|
1417 {
|
cannam@167
|
1418 V T3P, T3U, T42, T43;
|
cannam@167
|
1419 T3P = VSUB(T3L, T3O);
|
cannam@167
|
1420 T3U = VADD(T3Q, T3T);
|
cannam@167
|
1421 T3V = VMUL(LDK(KP707106781), VSUB(T3P, T3U));
|
cannam@167
|
1422 T57 = VMUL(LDK(KP707106781), VADD(T3U, T3P));
|
cannam@167
|
1423 T42 = VSUB(T3T, T3Q);
|
cannam@167
|
1424 T43 = VADD(T3L, T3O);
|
cannam@167
|
1425 T44 = VMUL(LDK(KP707106781), VSUB(T42, T43));
|
cannam@167
|
1426 T5a = VMUL(LDK(KP707106781), VADD(T42, T43));
|
cannam@167
|
1427 }
|
cannam@167
|
1428 }
|
cannam@167
|
1429 {
|
cannam@167
|
1430 V T2G, T4c, T2L, T4d, T4e, T4f, T2R, T4i, T2W, T4j, T4h, T4k;
|
cannam@167
|
1431 {
|
cannam@167
|
1432 V T2D, T2F, T2C, T2E;
|
cannam@167
|
1433 T2D = LD(&(ri[WS(rs, 3)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1434 T2F = LD(&(ii[WS(rs, 3)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1435 T2C = LDW(&(W[TWVL * 4]));
|
cannam@167
|
1436 T2E = LDW(&(W[TWVL * 5]));
|
cannam@167
|
1437 T2G = VFMA(T2C, T2D, VMUL(T2E, T2F));
|
cannam@167
|
1438 T4c = VFNMS(T2E, T2D, VMUL(T2C, T2F));
|
cannam@167
|
1439 }
|
cannam@167
|
1440 {
|
cannam@167
|
1441 V T2I, T2K, T2H, T2J;
|
cannam@167
|
1442 T2I = LD(&(ri[WS(rs, 19)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1443 T2K = LD(&(ii[WS(rs, 19)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1444 T2H = LDW(&(W[TWVL * 36]));
|
cannam@167
|
1445 T2J = LDW(&(W[TWVL * 37]));
|
cannam@167
|
1446 T2L = VFMA(T2H, T2I, VMUL(T2J, T2K));
|
cannam@167
|
1447 T4d = VFNMS(T2J, T2I, VMUL(T2H, T2K));
|
cannam@167
|
1448 }
|
cannam@167
|
1449 T4e = VSUB(T4c, T4d);
|
cannam@167
|
1450 T4f = VSUB(T2G, T2L);
|
cannam@167
|
1451 {
|
cannam@167
|
1452 V T2O, T2Q, T2N, T2P;
|
cannam@167
|
1453 T2O = LD(&(ri[WS(rs, 27)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1454 T2Q = LD(&(ii[WS(rs, 27)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1455 T2N = LDW(&(W[TWVL * 52]));
|
cannam@167
|
1456 T2P = LDW(&(W[TWVL * 53]));
|
cannam@167
|
1457 T2R = VFMA(T2N, T2O, VMUL(T2P, T2Q));
|
cannam@167
|
1458 T4i = VFNMS(T2P, T2O, VMUL(T2N, T2Q));
|
cannam@167
|
1459 }
|
cannam@167
|
1460 {
|
cannam@167
|
1461 V T2T, T2V, T2S, T2U;
|
cannam@167
|
1462 T2T = LD(&(ri[WS(rs, 11)]), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1463 T2V = LD(&(ii[WS(rs, 11)]), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1464 T2S = LDW(&(W[TWVL * 20]));
|
cannam@167
|
1465 T2U = LDW(&(W[TWVL * 21]));
|
cannam@167
|
1466 T2W = VFMA(T2S, T2T, VMUL(T2U, T2V));
|
cannam@167
|
1467 T4j = VFNMS(T2U, T2T, VMUL(T2S, T2V));
|
cannam@167
|
1468 }
|
cannam@167
|
1469 T4h = VSUB(T2R, T2W);
|
cannam@167
|
1470 T4k = VSUB(T4i, T4j);
|
cannam@167
|
1471 {
|
cannam@167
|
1472 V T2M, T2X, T68, T69;
|
cannam@167
|
1473 T2M = VADD(T2G, T2L);
|
cannam@167
|
1474 T2X = VADD(T2R, T2W);
|
cannam@167
|
1475 T2Y = VADD(T2M, T2X);
|
cannam@167
|
1476 T6f = VSUB(T2X, T2M);
|
cannam@167
|
1477 T68 = VADD(T4c, T4d);
|
cannam@167
|
1478 T69 = VADD(T4i, T4j);
|
cannam@167
|
1479 T6a = VSUB(T68, T69);
|
cannam@167
|
1480 T6P = VADD(T68, T69);
|
cannam@167
|
1481 }
|
cannam@167
|
1482 {
|
cannam@167
|
1483 V T4g, T4l, T4t, T4u;
|
cannam@167
|
1484 T4g = VSUB(T4e, T4f);
|
cannam@167
|
1485 T4l = VADD(T4h, T4k);
|
cannam@167
|
1486 T4m = VMUL(LDK(KP707106781), VSUB(T4g, T4l));
|
cannam@167
|
1487 T5h = VMUL(LDK(KP707106781), VADD(T4g, T4l));
|
cannam@167
|
1488 T4t = VSUB(T4h, T4k);
|
cannam@167
|
1489 T4u = VADD(T4f, T4e);
|
cannam@167
|
1490 T4v = VMUL(LDK(KP707106781), VSUB(T4t, T4u));
|
cannam@167
|
1491 T5e = VMUL(LDK(KP707106781), VADD(T4u, T4t));
|
cannam@167
|
1492 }
|
cannam@167
|
1493 }
|
cannam@167
|
1494 {
|
cannam@167
|
1495 V T1t, T6X, T7a, T7c, T30, T7b, T70, T71;
|
cannam@167
|
1496 {
|
cannam@167
|
1497 V TH, T1s, T72, T79;
|
cannam@167
|
1498 TH = VADD(Tj, TG);
|
cannam@167
|
1499 T1s = VADD(T14, T1r);
|
cannam@167
|
1500 T1t = VADD(TH, T1s);
|
cannam@167
|
1501 T6X = VSUB(TH, T1s);
|
cannam@167
|
1502 T72 = VADD(T6E, T6F);
|
cannam@167
|
1503 T79 = VADD(T73, T78);
|
cannam@167
|
1504 T7a = VADD(T72, T79);
|
cannam@167
|
1505 T7c = VSUB(T79, T72);
|
cannam@167
|
1506 }
|
cannam@167
|
1507 {
|
cannam@167
|
1508 V T2e, T2Z, T6Y, T6Z;
|
cannam@167
|
1509 T2e = VADD(T1Q, T2d);
|
cannam@167
|
1510 T2Z = VADD(T2B, T2Y);
|
cannam@167
|
1511 T30 = VADD(T2e, T2Z);
|
cannam@167
|
1512 T7b = VSUB(T2Z, T2e);
|
cannam@167
|
1513 T6Y = VADD(T6J, T6K);
|
cannam@167
|
1514 T6Z = VADD(T6O, T6P);
|
cannam@167
|
1515 T70 = VSUB(T6Y, T6Z);
|
cannam@167
|
1516 T71 = VADD(T6Y, T6Z);
|
cannam@167
|
1517 }
|
cannam@167
|
1518 ST(&(ri[WS(rs, 16)]), VSUB(T1t, T30), ms, &(ri[0]));
|
cannam@167
|
1519 ST(&(ii[WS(rs, 16)]), VSUB(T7a, T71), ms, &(ii[0]));
|
cannam@167
|
1520 ST(&(ri[0]), VADD(T1t, T30), ms, &(ri[0]));
|
cannam@167
|
1521 ST(&(ii[0]), VADD(T71, T7a), ms, &(ii[0]));
|
cannam@167
|
1522 ST(&(ri[WS(rs, 24)]), VSUB(T6X, T70), ms, &(ri[0]));
|
cannam@167
|
1523 ST(&(ii[WS(rs, 24)]), VSUB(T7c, T7b), ms, &(ii[0]));
|
cannam@167
|
1524 ST(&(ri[WS(rs, 8)]), VADD(T6X, T70), ms, &(ri[0]));
|
cannam@167
|
1525 ST(&(ii[WS(rs, 8)]), VADD(T7b, T7c), ms, &(ii[0]));
|
cannam@167
|
1526 }
|
cannam@167
|
1527 {
|
cannam@167
|
1528 V T6H, T6T, T7g, T7i, T6M, T6U, T6R, T6V;
|
cannam@167
|
1529 {
|
cannam@167
|
1530 V T6D, T6G, T7e, T7f;
|
cannam@167
|
1531 T6D = VSUB(Tj, TG);
|
cannam@167
|
1532 T6G = VSUB(T6E, T6F);
|
cannam@167
|
1533 T6H = VADD(T6D, T6G);
|
cannam@167
|
1534 T6T = VSUB(T6D, T6G);
|
cannam@167
|
1535 T7e = VSUB(T1r, T14);
|
cannam@167
|
1536 T7f = VSUB(T78, T73);
|
cannam@167
|
1537 T7g = VADD(T7e, T7f);
|
cannam@167
|
1538 T7i = VSUB(T7f, T7e);
|
cannam@167
|
1539 }
|
cannam@167
|
1540 {
|
cannam@167
|
1541 V T6I, T6L, T6N, T6Q;
|
cannam@167
|
1542 T6I = VSUB(T1Q, T2d);
|
cannam@167
|
1543 T6L = VSUB(T6J, T6K);
|
cannam@167
|
1544 T6M = VADD(T6I, T6L);
|
cannam@167
|
1545 T6U = VSUB(T6L, T6I);
|
cannam@167
|
1546 T6N = VSUB(T2B, T2Y);
|
cannam@167
|
1547 T6Q = VSUB(T6O, T6P);
|
cannam@167
|
1548 T6R = VSUB(T6N, T6Q);
|
cannam@167
|
1549 T6V = VADD(T6N, T6Q);
|
cannam@167
|
1550 }
|
cannam@167
|
1551 {
|
cannam@167
|
1552 V T6S, T7d, T6W, T7h;
|
cannam@167
|
1553 T6S = VMUL(LDK(KP707106781), VADD(T6M, T6R));
|
cannam@167
|
1554 ST(&(ri[WS(rs, 20)]), VSUB(T6H, T6S), ms, &(ri[0]));
|
cannam@167
|
1555 ST(&(ri[WS(rs, 4)]), VADD(T6H, T6S), ms, &(ri[0]));
|
cannam@167
|
1556 T7d = VMUL(LDK(KP707106781), VADD(T6U, T6V));
|
cannam@167
|
1557 ST(&(ii[WS(rs, 4)]), VADD(T7d, T7g), ms, &(ii[0]));
|
cannam@167
|
1558 ST(&(ii[WS(rs, 20)]), VSUB(T7g, T7d), ms, &(ii[0]));
|
cannam@167
|
1559 T6W = VMUL(LDK(KP707106781), VSUB(T6U, T6V));
|
cannam@167
|
1560 ST(&(ri[WS(rs, 28)]), VSUB(T6T, T6W), ms, &(ri[0]));
|
cannam@167
|
1561 ST(&(ri[WS(rs, 12)]), VADD(T6T, T6W), ms, &(ri[0]));
|
cannam@167
|
1562 T7h = VMUL(LDK(KP707106781), VSUB(T6R, T6M));
|
cannam@167
|
1563 ST(&(ii[WS(rs, 12)]), VADD(T7h, T7i), ms, &(ii[0]));
|
cannam@167
|
1564 ST(&(ii[WS(rs, 28)]), VSUB(T7i, T7h), ms, &(ii[0]));
|
cannam@167
|
1565 }
|
cannam@167
|
1566 }
|
cannam@167
|
1567 {
|
cannam@167
|
1568 V T5J, T7n, T7t, T6n, T5U, T7k, T6x, T6B, T6q, T7s, T66, T6k, T6u, T6A, T6h;
|
cannam@167
|
1569 V T6l;
|
cannam@167
|
1570 {
|
cannam@167
|
1571 V T5O, T5T, T60, T65;
|
cannam@167
|
1572 T5J = VSUB(T5F, T5I);
|
cannam@167
|
1573 T7n = VADD(T7l, T7m);
|
cannam@167
|
1574 T7t = VSUB(T7m, T7l);
|
cannam@167
|
1575 T6n = VADD(T5F, T5I);
|
cannam@167
|
1576 T5O = VSUB(T5M, T5N);
|
cannam@167
|
1577 T5T = VADD(T5P, T5S);
|
cannam@167
|
1578 T5U = VMUL(LDK(KP707106781), VSUB(T5O, T5T));
|
cannam@167
|
1579 T7k = VMUL(LDK(KP707106781), VADD(T5O, T5T));
|
cannam@167
|
1580 {
|
cannam@167
|
1581 V T6v, T6w, T6o, T6p;
|
cannam@167
|
1582 T6v = VADD(T67, T6a);
|
cannam@167
|
1583 T6w = VADD(T6e, T6f);
|
cannam@167
|
1584 T6x = VFNMS(LDK(KP382683432), T6w, VMUL(LDK(KP923879532), T6v));
|
cannam@167
|
1585 T6B = VFMA(LDK(KP923879532), T6w, VMUL(LDK(KP382683432), T6v));
|
cannam@167
|
1586 T6o = VADD(T5N, T5M);
|
cannam@167
|
1587 T6p = VSUB(T5P, T5S);
|
cannam@167
|
1588 T6q = VMUL(LDK(KP707106781), VADD(T6o, T6p));
|
cannam@167
|
1589 T7s = VMUL(LDK(KP707106781), VSUB(T6p, T6o));
|
cannam@167
|
1590 }
|
cannam@167
|
1591 T60 = VSUB(T5Y, T5Z);
|
cannam@167
|
1592 T65 = VSUB(T61, T64);
|
cannam@167
|
1593 T66 = VFMA(LDK(KP923879532), T60, VMUL(LDK(KP382683432), T65));
|
cannam@167
|
1594 T6k = VFNMS(LDK(KP923879532), T65, VMUL(LDK(KP382683432), T60));
|
cannam@167
|
1595 {
|
cannam@167
|
1596 V T6s, T6t, T6b, T6g;
|
cannam@167
|
1597 T6s = VADD(T5Y, T5Z);
|
cannam@167
|
1598 T6t = VADD(T61, T64);
|
cannam@167
|
1599 T6u = VFMA(LDK(KP382683432), T6s, VMUL(LDK(KP923879532), T6t));
|
cannam@167
|
1600 T6A = VFNMS(LDK(KP382683432), T6t, VMUL(LDK(KP923879532), T6s));
|
cannam@167
|
1601 T6b = VSUB(T67, T6a);
|
cannam@167
|
1602 T6g = VSUB(T6e, T6f);
|
cannam@167
|
1603 T6h = VFNMS(LDK(KP923879532), T6g, VMUL(LDK(KP382683432), T6b));
|
cannam@167
|
1604 T6l = VFMA(LDK(KP382683432), T6g, VMUL(LDK(KP923879532), T6b));
|
cannam@167
|
1605 }
|
cannam@167
|
1606 }
|
cannam@167
|
1607 {
|
cannam@167
|
1608 V T5V, T6i, T7r, T7u;
|
cannam@167
|
1609 T5V = VADD(T5J, T5U);
|
cannam@167
|
1610 T6i = VADD(T66, T6h);
|
cannam@167
|
1611 ST(&(ri[WS(rs, 22)]), VSUB(T5V, T6i), ms, &(ri[0]));
|
cannam@167
|
1612 ST(&(ri[WS(rs, 6)]), VADD(T5V, T6i), ms, &(ri[0]));
|
cannam@167
|
1613 T7r = VADD(T6k, T6l);
|
cannam@167
|
1614 T7u = VADD(T7s, T7t);
|
cannam@167
|
1615 ST(&(ii[WS(rs, 6)]), VADD(T7r, T7u), ms, &(ii[0]));
|
cannam@167
|
1616 ST(&(ii[WS(rs, 22)]), VSUB(T7u, T7r), ms, &(ii[0]));
|
cannam@167
|
1617 }
|
cannam@167
|
1618 {
|
cannam@167
|
1619 V T6j, T6m, T7v, T7w;
|
cannam@167
|
1620 T6j = VSUB(T5J, T5U);
|
cannam@167
|
1621 T6m = VSUB(T6k, T6l);
|
cannam@167
|
1622 ST(&(ri[WS(rs, 30)]), VSUB(T6j, T6m), ms, &(ri[0]));
|
cannam@167
|
1623 ST(&(ri[WS(rs, 14)]), VADD(T6j, T6m), ms, &(ri[0]));
|
cannam@167
|
1624 T7v = VSUB(T6h, T66);
|
cannam@167
|
1625 T7w = VSUB(T7t, T7s);
|
cannam@167
|
1626 ST(&(ii[WS(rs, 14)]), VADD(T7v, T7w), ms, &(ii[0]));
|
cannam@167
|
1627 ST(&(ii[WS(rs, 30)]), VSUB(T7w, T7v), ms, &(ii[0]));
|
cannam@167
|
1628 }
|
cannam@167
|
1629 {
|
cannam@167
|
1630 V T6r, T6y, T7j, T7o;
|
cannam@167
|
1631 T6r = VADD(T6n, T6q);
|
cannam@167
|
1632 T6y = VADD(T6u, T6x);
|
cannam@167
|
1633 ST(&(ri[WS(rs, 18)]), VSUB(T6r, T6y), ms, &(ri[0]));
|
cannam@167
|
1634 ST(&(ri[WS(rs, 2)]), VADD(T6r, T6y), ms, &(ri[0]));
|
cannam@167
|
1635 T7j = VADD(T6A, T6B);
|
cannam@167
|
1636 T7o = VADD(T7k, T7n);
|
cannam@167
|
1637 ST(&(ii[WS(rs, 2)]), VADD(T7j, T7o), ms, &(ii[0]));
|
cannam@167
|
1638 ST(&(ii[WS(rs, 18)]), VSUB(T7o, T7j), ms, &(ii[0]));
|
cannam@167
|
1639 }
|
cannam@167
|
1640 {
|
cannam@167
|
1641 V T6z, T6C, T7p, T7q;
|
cannam@167
|
1642 T6z = VSUB(T6n, T6q);
|
cannam@167
|
1643 T6C = VSUB(T6A, T6B);
|
cannam@167
|
1644 ST(&(ri[WS(rs, 26)]), VSUB(T6z, T6C), ms, &(ri[0]));
|
cannam@167
|
1645 ST(&(ri[WS(rs, 10)]), VADD(T6z, T6C), ms, &(ri[0]));
|
cannam@167
|
1646 T7p = VSUB(T6x, T6u);
|
cannam@167
|
1647 T7q = VSUB(T7n, T7k);
|
cannam@167
|
1648 ST(&(ii[WS(rs, 10)]), VADD(T7p, T7q), ms, &(ii[0]));
|
cannam@167
|
1649 ST(&(ii[WS(rs, 26)]), VSUB(T7q, T7p), ms, &(ii[0]));
|
cannam@167
|
1650 }
|
cannam@167
|
1651 }
|
cannam@167
|
1652 {
|
cannam@167
|
1653 V T3h, T4D, T7R, T7X, T3E, T7O, T4N, T4R, T46, T4A, T4G, T7W, T4K, T4Q, T4x;
|
cannam@167
|
1654 V T4B, T3g, T7P;
|
cannam@167
|
1655 T3g = VMUL(LDK(KP707106781), VSUB(T3a, T3f));
|
cannam@167
|
1656 T3h = VSUB(T35, T3g);
|
cannam@167
|
1657 T4D = VADD(T35, T3g);
|
cannam@167
|
1658 T7P = VMUL(LDK(KP707106781), VSUB(T4V, T4U));
|
cannam@167
|
1659 T7R = VADD(T7P, T7Q);
|
cannam@167
|
1660 T7X = VSUB(T7Q, T7P);
|
cannam@167
|
1661 {
|
cannam@167
|
1662 V T3s, T3D, T4L, T4M;
|
cannam@167
|
1663 T3s = VFNMS(LDK(KP923879532), T3r, VMUL(LDK(KP382683432), T3m));
|
cannam@167
|
1664 T3D = VFMA(LDK(KP382683432), T3x, VMUL(LDK(KP923879532), T3C));
|
cannam@167
|
1665 T3E = VSUB(T3s, T3D);
|
cannam@167
|
1666 T7O = VADD(T3s, T3D);
|
cannam@167
|
1667 T4L = VADD(T4b, T4m);
|
cannam@167
|
1668 T4M = VADD(T4s, T4v);
|
cannam@167
|
1669 T4N = VFNMS(LDK(KP555570233), T4M, VMUL(LDK(KP831469612), T4L));
|
cannam@167
|
1670 T4R = VFMA(LDK(KP831469612), T4M, VMUL(LDK(KP555570233), T4L));
|
cannam@167
|
1671 }
|
cannam@167
|
1672 {
|
cannam@167
|
1673 V T3W, T45, T4E, T4F;
|
cannam@167
|
1674 T3W = VSUB(T3K, T3V);
|
cannam@167
|
1675 T45 = VSUB(T41, T44);
|
cannam@167
|
1676 T46 = VFMA(LDK(KP980785280), T3W, VMUL(LDK(KP195090322), T45));
|
cannam@167
|
1677 T4A = VFNMS(LDK(KP980785280), T45, VMUL(LDK(KP195090322), T3W));
|
cannam@167
|
1678 T4E = VFMA(LDK(KP923879532), T3m, VMUL(LDK(KP382683432), T3r));
|
cannam@167
|
1679 T4F = VFNMS(LDK(KP923879532), T3x, VMUL(LDK(KP382683432), T3C));
|
cannam@167
|
1680 T4G = VADD(T4E, T4F);
|
cannam@167
|
1681 T7W = VSUB(T4F, T4E);
|
cannam@167
|
1682 }
|
cannam@167
|
1683 {
|
cannam@167
|
1684 V T4I, T4J, T4n, T4w;
|
cannam@167
|
1685 T4I = VADD(T3K, T3V);
|
cannam@167
|
1686 T4J = VADD(T41, T44);
|
cannam@167
|
1687 T4K = VFMA(LDK(KP555570233), T4I, VMUL(LDK(KP831469612), T4J));
|
cannam@167
|
1688 T4Q = VFNMS(LDK(KP555570233), T4J, VMUL(LDK(KP831469612), T4I));
|
cannam@167
|
1689 T4n = VSUB(T4b, T4m);
|
cannam@167
|
1690 T4w = VSUB(T4s, T4v);
|
cannam@167
|
1691 T4x = VFNMS(LDK(KP980785280), T4w, VMUL(LDK(KP195090322), T4n));
|
cannam@167
|
1692 T4B = VFMA(LDK(KP195090322), T4w, VMUL(LDK(KP980785280), T4n));
|
cannam@167
|
1693 }
|
cannam@167
|
1694 {
|
cannam@167
|
1695 V T3F, T4y, T7V, T7Y;
|
cannam@167
|
1696 T3F = VADD(T3h, T3E);
|
cannam@167
|
1697 T4y = VADD(T46, T4x);
|
cannam@167
|
1698 ST(&(ri[WS(rs, 23)]), VSUB(T3F, T4y), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1699 ST(&(ri[WS(rs, 7)]), VADD(T3F, T4y), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1700 T7V = VADD(T4A, T4B);
|
cannam@167
|
1701 T7Y = VADD(T7W, T7X);
|
cannam@167
|
1702 ST(&(ii[WS(rs, 7)]), VADD(T7V, T7Y), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1703 ST(&(ii[WS(rs, 23)]), VSUB(T7Y, T7V), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1704 }
|
cannam@167
|
1705 {
|
cannam@167
|
1706 V T4z, T4C, T7Z, T80;
|
cannam@167
|
1707 T4z = VSUB(T3h, T3E);
|
cannam@167
|
1708 T4C = VSUB(T4A, T4B);
|
cannam@167
|
1709 ST(&(ri[WS(rs, 31)]), VSUB(T4z, T4C), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1710 ST(&(ri[WS(rs, 15)]), VADD(T4z, T4C), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1711 T7Z = VSUB(T4x, T46);
|
cannam@167
|
1712 T80 = VSUB(T7X, T7W);
|
cannam@167
|
1713 ST(&(ii[WS(rs, 15)]), VADD(T7Z, T80), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1714 ST(&(ii[WS(rs, 31)]), VSUB(T80, T7Z), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1715 }
|
cannam@167
|
1716 {
|
cannam@167
|
1717 V T4H, T4O, T7N, T7S;
|
cannam@167
|
1718 T4H = VADD(T4D, T4G);
|
cannam@167
|
1719 T4O = VADD(T4K, T4N);
|
cannam@167
|
1720 ST(&(ri[WS(rs, 19)]), VSUB(T4H, T4O), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1721 ST(&(ri[WS(rs, 3)]), VADD(T4H, T4O), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1722 T7N = VADD(T4Q, T4R);
|
cannam@167
|
1723 T7S = VADD(T7O, T7R);
|
cannam@167
|
1724 ST(&(ii[WS(rs, 3)]), VADD(T7N, T7S), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1725 ST(&(ii[WS(rs, 19)]), VSUB(T7S, T7N), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1726 }
|
cannam@167
|
1727 {
|
cannam@167
|
1728 V T4P, T4S, T7T, T7U;
|
cannam@167
|
1729 T4P = VSUB(T4D, T4G);
|
cannam@167
|
1730 T4S = VSUB(T4Q, T4R);
|
cannam@167
|
1731 ST(&(ri[WS(rs, 27)]), VSUB(T4P, T4S), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1732 ST(&(ri[WS(rs, 11)]), VADD(T4P, T4S), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1733 T7T = VSUB(T4N, T4K);
|
cannam@167
|
1734 T7U = VSUB(T7R, T7O);
|
cannam@167
|
1735 ST(&(ii[WS(rs, 11)]), VADD(T7T, T7U), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1736 ST(&(ii[WS(rs, 27)]), VSUB(T7U, T7T), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1737 }
|
cannam@167
|
1738 }
|
cannam@167
|
1739 {
|
cannam@167
|
1740 V T4X, T5p, T7D, T7J, T54, T7y, T5z, T5D, T5c, T5m, T5s, T7I, T5w, T5C, T5j;
|
cannam@167
|
1741 V T5n, T4W, T7z;
|
cannam@167
|
1742 T4W = VMUL(LDK(KP707106781), VADD(T4U, T4V));
|
cannam@167
|
1743 T4X = VSUB(T4T, T4W);
|
cannam@167
|
1744 T5p = VADD(T4T, T4W);
|
cannam@167
|
1745 T7z = VMUL(LDK(KP707106781), VADD(T3a, T3f));
|
cannam@167
|
1746 T7D = VADD(T7z, T7C);
|
cannam@167
|
1747 T7J = VSUB(T7C, T7z);
|
cannam@167
|
1748 {
|
cannam@167
|
1749 V T50, T53, T5x, T5y;
|
cannam@167
|
1750 T50 = VFNMS(LDK(KP382683432), T4Z, VMUL(LDK(KP923879532), T4Y));
|
cannam@167
|
1751 T53 = VFMA(LDK(KP923879532), T51, VMUL(LDK(KP382683432), T52));
|
cannam@167
|
1752 T54 = VSUB(T50, T53);
|
cannam@167
|
1753 T7y = VADD(T50, T53);
|
cannam@167
|
1754 T5x = VADD(T5d, T5e);
|
cannam@167
|
1755 T5y = VADD(T5g, T5h);
|
cannam@167
|
1756 T5z = VFNMS(LDK(KP195090322), T5y, VMUL(LDK(KP980785280), T5x));
|
cannam@167
|
1757 T5D = VFMA(LDK(KP195090322), T5x, VMUL(LDK(KP980785280), T5y));
|
cannam@167
|
1758 }
|
cannam@167
|
1759 {
|
cannam@167
|
1760 V T58, T5b, T5q, T5r;
|
cannam@167
|
1761 T58 = VSUB(T56, T57);
|
cannam@167
|
1762 T5b = VSUB(T59, T5a);
|
cannam@167
|
1763 T5c = VFMA(LDK(KP555570233), T58, VMUL(LDK(KP831469612), T5b));
|
cannam@167
|
1764 T5m = VFNMS(LDK(KP831469612), T58, VMUL(LDK(KP555570233), T5b));
|
cannam@167
|
1765 T5q = VFMA(LDK(KP382683432), T4Y, VMUL(LDK(KP923879532), T4Z));
|
cannam@167
|
1766 T5r = VFNMS(LDK(KP382683432), T51, VMUL(LDK(KP923879532), T52));
|
cannam@167
|
1767 T5s = VADD(T5q, T5r);
|
cannam@167
|
1768 T7I = VSUB(T5r, T5q);
|
cannam@167
|
1769 }
|
cannam@167
|
1770 {
|
cannam@167
|
1771 V T5u, T5v, T5f, T5i;
|
cannam@167
|
1772 T5u = VADD(T56, T57);
|
cannam@167
|
1773 T5v = VADD(T59, T5a);
|
cannam@167
|
1774 T5w = VFMA(LDK(KP980785280), T5u, VMUL(LDK(KP195090322), T5v));
|
cannam@167
|
1775 T5C = VFNMS(LDK(KP195090322), T5u, VMUL(LDK(KP980785280), T5v));
|
cannam@167
|
1776 T5f = VSUB(T5d, T5e);
|
cannam@167
|
1777 T5i = VSUB(T5g, T5h);
|
cannam@167
|
1778 T5j = VFNMS(LDK(KP831469612), T5i, VMUL(LDK(KP555570233), T5f));
|
cannam@167
|
1779 T5n = VFMA(LDK(KP831469612), T5f, VMUL(LDK(KP555570233), T5i));
|
cannam@167
|
1780 }
|
cannam@167
|
1781 {
|
cannam@167
|
1782 V T55, T5k, T7H, T7K;
|
cannam@167
|
1783 T55 = VADD(T4X, T54);
|
cannam@167
|
1784 T5k = VADD(T5c, T5j);
|
cannam@167
|
1785 ST(&(ri[WS(rs, 21)]), VSUB(T55, T5k), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1786 ST(&(ri[WS(rs, 5)]), VADD(T55, T5k), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1787 T7H = VADD(T5m, T5n);
|
cannam@167
|
1788 T7K = VADD(T7I, T7J);
|
cannam@167
|
1789 ST(&(ii[WS(rs, 5)]), VADD(T7H, T7K), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1790 ST(&(ii[WS(rs, 21)]), VSUB(T7K, T7H), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1791 }
|
cannam@167
|
1792 {
|
cannam@167
|
1793 V T5l, T5o, T7L, T7M;
|
cannam@167
|
1794 T5l = VSUB(T4X, T54);
|
cannam@167
|
1795 T5o = VSUB(T5m, T5n);
|
cannam@167
|
1796 ST(&(ri[WS(rs, 29)]), VSUB(T5l, T5o), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1797 ST(&(ri[WS(rs, 13)]), VADD(T5l, T5o), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1798 T7L = VSUB(T5j, T5c);
|
cannam@167
|
1799 T7M = VSUB(T7J, T7I);
|
cannam@167
|
1800 ST(&(ii[WS(rs, 13)]), VADD(T7L, T7M), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1801 ST(&(ii[WS(rs, 29)]), VSUB(T7M, T7L), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1802 }
|
cannam@167
|
1803 {
|
cannam@167
|
1804 V T5t, T5A, T7x, T7E;
|
cannam@167
|
1805 T5t = VADD(T5p, T5s);
|
cannam@167
|
1806 T5A = VADD(T5w, T5z);
|
cannam@167
|
1807 ST(&(ri[WS(rs, 17)]), VSUB(T5t, T5A), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1808 ST(&(ri[WS(rs, 1)]), VADD(T5t, T5A), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1809 T7x = VADD(T5C, T5D);
|
cannam@167
|
1810 T7E = VADD(T7y, T7D);
|
cannam@167
|
1811 ST(&(ii[WS(rs, 1)]), VADD(T7x, T7E), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1812 ST(&(ii[WS(rs, 17)]), VSUB(T7E, T7x), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1813 }
|
cannam@167
|
1814 {
|
cannam@167
|
1815 V T5B, T5E, T7F, T7G;
|
cannam@167
|
1816 T5B = VSUB(T5p, T5s);
|
cannam@167
|
1817 T5E = VSUB(T5C, T5D);
|
cannam@167
|
1818 ST(&(ri[WS(rs, 25)]), VSUB(T5B, T5E), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1819 ST(&(ri[WS(rs, 9)]), VADD(T5B, T5E), ms, &(ri[WS(rs, 1)]));
|
cannam@167
|
1820 T7F = VSUB(T5z, T5w);
|
cannam@167
|
1821 T7G = VSUB(T7D, T7y);
|
cannam@167
|
1822 ST(&(ii[WS(rs, 9)]), VADD(T7F, T7G), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1823 ST(&(ii[WS(rs, 25)]), VSUB(T7G, T7F), ms, &(ii[WS(rs, 1)]));
|
cannam@167
|
1824 }
|
cannam@167
|
1825 }
|
cannam@167
|
1826 }
|
cannam@167
|
1827 }
|
cannam@167
|
1828 VLEAVE();
|
cannam@167
|
1829 }
|
cannam@167
|
1830
|
cannam@167
|
1831 static const tw_instr twinstr[] = {
|
cannam@167
|
1832 VTW(0, 1),
|
cannam@167
|
1833 VTW(0, 2),
|
cannam@167
|
1834 VTW(0, 3),
|
cannam@167
|
1835 VTW(0, 4),
|
cannam@167
|
1836 VTW(0, 5),
|
cannam@167
|
1837 VTW(0, 6),
|
cannam@167
|
1838 VTW(0, 7),
|
cannam@167
|
1839 VTW(0, 8),
|
cannam@167
|
1840 VTW(0, 9),
|
cannam@167
|
1841 VTW(0, 10),
|
cannam@167
|
1842 VTW(0, 11),
|
cannam@167
|
1843 VTW(0, 12),
|
cannam@167
|
1844 VTW(0, 13),
|
cannam@167
|
1845 VTW(0, 14),
|
cannam@167
|
1846 VTW(0, 15),
|
cannam@167
|
1847 VTW(0, 16),
|
cannam@167
|
1848 VTW(0, 17),
|
cannam@167
|
1849 VTW(0, 18),
|
cannam@167
|
1850 VTW(0, 19),
|
cannam@167
|
1851 VTW(0, 20),
|
cannam@167
|
1852 VTW(0, 21),
|
cannam@167
|
1853 VTW(0, 22),
|
cannam@167
|
1854 VTW(0, 23),
|
cannam@167
|
1855 VTW(0, 24),
|
cannam@167
|
1856 VTW(0, 25),
|
cannam@167
|
1857 VTW(0, 26),
|
cannam@167
|
1858 VTW(0, 27),
|
cannam@167
|
1859 VTW(0, 28),
|
cannam@167
|
1860 VTW(0, 29),
|
cannam@167
|
1861 VTW(0, 30),
|
cannam@167
|
1862 VTW(0, 31),
|
cannam@167
|
1863 {TW_NEXT, (2 * VL), 0}
|
cannam@167
|
1864 };
|
cannam@167
|
1865
|
cannam@167
|
1866 static const ct_desc desc = { 32, XSIMD_STRING("t1sv_32"), twinstr, &GENUS, {340, 114, 94, 0}, 0, 0, 0 };
|
cannam@167
|
1867
|
cannam@167
|
1868 void XSIMD(codelet_t1sv_32) (planner *p) {
|
cannam@167
|
1869 X(kdft_dit_register) (p, t1sv_32, &desc);
|
cannam@167
|
1870 }
|
cannam@167
|
1871 #endif
|