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 08:08:11 EDT 2018 */
|
cannam@167
|
23
|
cannam@167
|
24 #include "rdft/codelet-rdft.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_hc2cdft_c.native -fma -simd -compact -variables 4 -pipeline-latency 8 -trivial-stores -variables 32 -no-generate-bytw -n 32 -dit -name hc2cfdftv_32 -include rdft/simd/hc2cfv.h */
|
cannam@167
|
29
|
cannam@167
|
30 /*
|
cannam@167
|
31 * This function contains 249 FP additions, 224 FP multiplications,
|
cannam@167
|
32 * (or, 119 additions, 94 multiplications, 130 fused multiply/add),
|
cannam@167
|
33 * 154 stack variables, 8 constants, and 64 memory accesses
|
cannam@167
|
34 */
|
cannam@167
|
35 #include "rdft/simd/hc2cfv.h"
|
cannam@167
|
36
|
cannam@167
|
37 static void hc2cfdftv_32(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
38 {
|
cannam@167
|
39 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
cannam@167
|
40 DVK(KP668178637, +0.668178637919298919997757686523080761552472251);
|
cannam@167
|
41 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
cannam@167
|
42 DVK(KP198912367, +0.198912367379658006911597622644676228597850501);
|
cannam@167
|
43 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
44 DVK(KP500000000, +0.500000000000000000000000000000000000000000000);
|
cannam@167
|
45 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
46 DVK(KP414213562, +0.414213562373095048801688724209698078569671875);
|
cannam@167
|
47 {
|
cannam@167
|
48 INT m;
|
cannam@167
|
49 for (m = mb, W = W + ((mb - 1) * ((TWVL / VL) * 62)); m < me; m = m + VL, Rp = Rp + (VL * ms), Ip = Ip + (VL * ms), Rm = Rm - (VL * ms), Im = Im - (VL * ms), W = W + (TWVL * 62), MAKE_VOLATILE_STRIDE(128, rs)) {
|
cannam@167
|
50 V T47, T48, T4l, T3w, T3F, T3B, T41, Ts, T2y, T1Q, T2B, T27, T2J, T3a, T40;
|
cannam@167
|
51 V T1X, T2C, T43, T44, T4a, T4b, T4m, T3p, T3E, T15, T2K, T1u, T2F, T3h, T3C;
|
cannam@167
|
52 V T1n, T2E, T2a, T2z, T1a, T18, TU, T3m, T3f, T1r, T1p, T13, T3n, T3e, TB;
|
cannam@167
|
53 V T3k, T1l, T3c, TK, T3j, T1g, T3b, T3l, T3o, TL, T14, T1s, T1t, T3d, T3g;
|
cannam@167
|
54 V T1b, T1m, T28, T29, T3Q, T3W, T3T, T3X, T3O, T3P, T3R, T3S, T3U, T3Z, T3V;
|
cannam@167
|
55 V T3Y;
|
cannam@167
|
56 {
|
cannam@167
|
57 V T1U, T1S, T3, T3u, T7, T1z, T1D, T3t, T24, T22, Tc, Tg, Th, T3q, T1J;
|
cannam@167
|
58 V Tl, Tp, Tq, T3r, T1O, T3s, T3v, T3z, T3A, T8, Tr, T1E, T1P, T25, T26;
|
cannam@167
|
59 V T38, T39, T1V, T1W;
|
cannam@167
|
60 {
|
cannam@167
|
61 V T1, T2, T5, T6, T1T, T1R, T4, T1x, T1y, T1B, T1C, T1w, T1A, T23, T21;
|
cannam@167
|
62 V T1I, T1G, Ta, Tb, T9, T1H, Te, Tf, Td, T1F, T1N, T1L, Tj, Tk, Ti;
|
cannam@167
|
63 V T1M, Tn, To, Tm, T1K;
|
cannam@167
|
64 T1 = LD(&(Rp[0]), ms, &(Rp[0]));
|
cannam@167
|
65 T2 = LD(&(Rm[0]), -ms, &(Rm[0]));
|
cannam@167
|
66 T1T = LDW(&(W[0]));
|
cannam@167
|
67 T1U = VZMULIJ(T1T, VFNMSCONJ(T2, T1));
|
cannam@167
|
68 T5 = LD(&(Rp[WS(rs, 8)]), ms, &(Rp[0]));
|
cannam@167
|
69 T6 = LD(&(Rm[WS(rs, 8)]), -ms, &(Rm[0]));
|
cannam@167
|
70 T1R = LDW(&(W[TWVL * 32]));
|
cannam@167
|
71 T1S = VZMULIJ(T1R, VFNMSCONJ(T6, T5));
|
cannam@167
|
72 T3 = VFMACONJ(T2, T1);
|
cannam@167
|
73 T3u = VADD(T1U, T1S);
|
cannam@167
|
74 T4 = LDW(&(W[TWVL * 30]));
|
cannam@167
|
75 T7 = VZMULJ(T4, VFMACONJ(T6, T5));
|
cannam@167
|
76 T1x = LD(&(Rp[WS(rs, 12)]), ms, &(Rp[0]));
|
cannam@167
|
77 T1y = LD(&(Rm[WS(rs, 12)]), -ms, &(Rm[0]));
|
cannam@167
|
78 T1w = LDW(&(W[TWVL * 48]));
|
cannam@167
|
79 T1z = VZMULIJ(T1w, VFNMSCONJ(T1y, T1x));
|
cannam@167
|
80 T1B = LD(&(Rp[WS(rs, 4)]), ms, &(Rp[0]));
|
cannam@167
|
81 T1C = LD(&(Rm[WS(rs, 4)]), -ms, &(Rm[0]));
|
cannam@167
|
82 T1A = LDW(&(W[TWVL * 16]));
|
cannam@167
|
83 T1D = VZMULIJ(T1A, VFNMSCONJ(T1C, T1B));
|
cannam@167
|
84 T3t = VADD(T1D, T1z);
|
cannam@167
|
85 T23 = LDW(&(W[TWVL * 46]));
|
cannam@167
|
86 T24 = VZMULJ(T23, VFMACONJ(T1y, T1x));
|
cannam@167
|
87 T21 = LDW(&(W[TWVL * 14]));
|
cannam@167
|
88 T22 = VZMULJ(T21, VFMACONJ(T1C, T1B));
|
cannam@167
|
89 Ta = LD(&(Rp[WS(rs, 2)]), ms, &(Rp[0]));
|
cannam@167
|
90 Tb = LD(&(Rm[WS(rs, 2)]), -ms, &(Rm[0]));
|
cannam@167
|
91 T9 = LDW(&(W[TWVL * 6]));
|
cannam@167
|
92 Tc = VZMULJ(T9, VFMACONJ(Tb, Ta));
|
cannam@167
|
93 T1H = LDW(&(W[TWVL * 8]));
|
cannam@167
|
94 T1I = VZMULIJ(T1H, VFNMSCONJ(Tb, Ta));
|
cannam@167
|
95 Te = LD(&(Rp[WS(rs, 10)]), ms, &(Rp[0]));
|
cannam@167
|
96 Tf = LD(&(Rm[WS(rs, 10)]), -ms, &(Rm[0]));
|
cannam@167
|
97 Td = LDW(&(W[TWVL * 38]));
|
cannam@167
|
98 Tg = VZMULJ(Td, VFMACONJ(Tf, Te));
|
cannam@167
|
99 T1F = LDW(&(W[TWVL * 40]));
|
cannam@167
|
100 T1G = VZMULIJ(T1F, VFNMSCONJ(Tf, Te));
|
cannam@167
|
101 Th = VSUB(Tc, Tg);
|
cannam@167
|
102 T3q = VADD(T1I, T1G);
|
cannam@167
|
103 T1J = VSUB(T1G, T1I);
|
cannam@167
|
104 Tj = LD(&(Rp[WS(rs, 14)]), ms, &(Rp[0]));
|
cannam@167
|
105 Tk = LD(&(Rm[WS(rs, 14)]), -ms, &(Rm[0]));
|
cannam@167
|
106 Ti = LDW(&(W[TWVL * 54]));
|
cannam@167
|
107 Tl = VZMULJ(Ti, VFMACONJ(Tk, Tj));
|
cannam@167
|
108 T1M = LDW(&(W[TWVL * 56]));
|
cannam@167
|
109 T1N = VZMULIJ(T1M, VFNMSCONJ(Tk, Tj));
|
cannam@167
|
110 Tn = LD(&(Rp[WS(rs, 6)]), ms, &(Rp[0]));
|
cannam@167
|
111 To = LD(&(Rm[WS(rs, 6)]), -ms, &(Rm[0]));
|
cannam@167
|
112 Tm = LDW(&(W[TWVL * 22]));
|
cannam@167
|
113 Tp = VZMULJ(Tm, VFMACONJ(To, Tn));
|
cannam@167
|
114 T1K = LDW(&(W[TWVL * 24]));
|
cannam@167
|
115 T1L = VZMULIJ(T1K, VFNMSCONJ(To, Tn));
|
cannam@167
|
116 Tq = VSUB(Tl, Tp);
|
cannam@167
|
117 T3r = VADD(T1N, T1L);
|
cannam@167
|
118 T1O = VSUB(T1L, T1N);
|
cannam@167
|
119 }
|
cannam@167
|
120 T47 = VADD(T3u, T3t);
|
cannam@167
|
121 T48 = VADD(T3q, T3r);
|
cannam@167
|
122 T4l = VSUB(T48, T47);
|
cannam@167
|
123 T3s = VSUB(T3q, T3r);
|
cannam@167
|
124 T3v = VSUB(T3t, T3u);
|
cannam@167
|
125 T3w = VFNMS(LDK(KP414213562), T3v, T3s);
|
cannam@167
|
126 T3F = VFMA(LDK(KP414213562), T3s, T3v);
|
cannam@167
|
127 T3z = VADD(Tl, Tp);
|
cannam@167
|
128 T3A = VADD(Tc, Tg);
|
cannam@167
|
129 T3B = VSUB(T3z, T3A);
|
cannam@167
|
130 T41 = VADD(T3A, T3z);
|
cannam@167
|
131 T8 = VSUB(T3, T7);
|
cannam@167
|
132 Tr = VADD(Th, Tq);
|
cannam@167
|
133 Ts = VFNMS(LDK(KP707106781), Tr, T8);
|
cannam@167
|
134 T2y = VFMA(LDK(KP707106781), Tr, T8);
|
cannam@167
|
135 T1E = VSUB(T1z, T1D);
|
cannam@167
|
136 T1P = VSUB(T1J, T1O);
|
cannam@167
|
137 T1Q = VFNMS(LDK(KP707106781), T1P, T1E);
|
cannam@167
|
138 T2B = VFMA(LDK(KP707106781), T1P, T1E);
|
cannam@167
|
139 T25 = VSUB(T22, T24);
|
cannam@167
|
140 T26 = VSUB(Tq, Th);
|
cannam@167
|
141 T27 = VFMA(LDK(KP707106781), T26, T25);
|
cannam@167
|
142 T2J = VFNMS(LDK(KP707106781), T26, T25);
|
cannam@167
|
143 T38 = VADD(T3, T7);
|
cannam@167
|
144 T39 = VADD(T22, T24);
|
cannam@167
|
145 T3a = VSUB(T38, T39);
|
cannam@167
|
146 T40 = VADD(T38, T39);
|
cannam@167
|
147 T1V = VSUB(T1S, T1U);
|
cannam@167
|
148 T1W = VADD(T1J, T1O);
|
cannam@167
|
149 T1X = VFNMS(LDK(KP707106781), T1W, T1V);
|
cannam@167
|
150 T2C = VFMA(LDK(KP707106781), T1W, T1V);
|
cannam@167
|
151 }
|
cannam@167
|
152 {
|
cannam@167
|
153 V TP, TT, TN, TO, TM, T19, TR, TS, TQ, T17, TY, T12, TW, TX, TV;
|
cannam@167
|
154 V T1q, T10, T11, TZ, T1o, Tw, T1i, TA, T1k, Tu, Tv, Tt, T1h, Ty, Tz;
|
cannam@167
|
155 V Tx, T1j, TF, T1f, TJ, T1d, TD, TE, TC, T1e, TH, TI, TG, T1c;
|
cannam@167
|
156 TN = LD(&(Rp[WS(rs, 3)]), ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
157 TO = LD(&(Rm[WS(rs, 3)]), -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
158 TM = LDW(&(W[TWVL * 10]));
|
cannam@167
|
159 TP = VZMULJ(TM, VFMACONJ(TO, TN));
|
cannam@167
|
160 T19 = LDW(&(W[TWVL * 12]));
|
cannam@167
|
161 T1a = VZMULIJ(T19, VFNMSCONJ(TO, TN));
|
cannam@167
|
162 TR = LD(&(Rp[WS(rs, 11)]), ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
163 TS = LD(&(Rm[WS(rs, 11)]), -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
164 TQ = LDW(&(W[TWVL * 42]));
|
cannam@167
|
165 TT = VZMULJ(TQ, VFMACONJ(TS, TR));
|
cannam@167
|
166 T17 = LDW(&(W[TWVL * 44]));
|
cannam@167
|
167 T18 = VZMULIJ(T17, VFNMSCONJ(TS, TR));
|
cannam@167
|
168 TU = VSUB(TP, TT);
|
cannam@167
|
169 T3m = VADD(T1a, T18);
|
cannam@167
|
170 T3f = VADD(TP, TT);
|
cannam@167
|
171 TW = LD(&(Rp[WS(rs, 15)]), ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
172 TX = LD(&(Rm[WS(rs, 15)]), -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
173 TV = LDW(&(W[TWVL * 58]));
|
cannam@167
|
174 TY = VZMULJ(TV, VFMACONJ(TX, TW));
|
cannam@167
|
175 T1q = LDW(&(W[TWVL * 60]));
|
cannam@167
|
176 T1r = VZMULIJ(T1q, VFNMSCONJ(TX, TW));
|
cannam@167
|
177 T10 = LD(&(Rp[WS(rs, 7)]), ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
178 T11 = LD(&(Rm[WS(rs, 7)]), -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
179 TZ = LDW(&(W[TWVL * 26]));
|
cannam@167
|
180 T12 = VZMULJ(TZ, VFMACONJ(T11, T10));
|
cannam@167
|
181 T1o = LDW(&(W[TWVL * 28]));
|
cannam@167
|
182 T1p = VZMULIJ(T1o, VFNMSCONJ(T11, T10));
|
cannam@167
|
183 T13 = VSUB(TY, T12);
|
cannam@167
|
184 T3n = VADD(T1r, T1p);
|
cannam@167
|
185 T3e = VADD(TY, T12);
|
cannam@167
|
186 Tu = LD(&(Rp[WS(rs, 5)]), ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
187 Tv = LD(&(Rm[WS(rs, 5)]), -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
188 Tt = LDW(&(W[TWVL * 18]));
|
cannam@167
|
189 Tw = VZMULJ(Tt, VFMACONJ(Tv, Tu));
|
cannam@167
|
190 T1h = LDW(&(W[TWVL * 20]));
|
cannam@167
|
191 T1i = VZMULIJ(T1h, VFNMSCONJ(Tv, Tu));
|
cannam@167
|
192 Ty = LD(&(Rp[WS(rs, 13)]), ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
193 Tz = LD(&(Rm[WS(rs, 13)]), -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
194 Tx = LDW(&(W[TWVL * 50]));
|
cannam@167
|
195 TA = VZMULJ(Tx, VFMACONJ(Tz, Ty));
|
cannam@167
|
196 T1j = LDW(&(W[TWVL * 52]));
|
cannam@167
|
197 T1k = VZMULIJ(T1j, VFNMSCONJ(Tz, Ty));
|
cannam@167
|
198 TB = VSUB(Tw, TA);
|
cannam@167
|
199 T3k = VADD(T1i, T1k);
|
cannam@167
|
200 T1l = VSUB(T1i, T1k);
|
cannam@167
|
201 T3c = VADD(Tw, TA);
|
cannam@167
|
202 TD = LD(&(Rp[WS(rs, 1)]), ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
203 TE = LD(&(Rm[WS(rs, 1)]), -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
204 TC = LDW(&(W[TWVL * 2]));
|
cannam@167
|
205 TF = VZMULJ(TC, VFMACONJ(TE, TD));
|
cannam@167
|
206 T1e = LDW(&(W[TWVL * 4]));
|
cannam@167
|
207 T1f = VZMULIJ(T1e, VFNMSCONJ(TE, TD));
|
cannam@167
|
208 TH = LD(&(Rp[WS(rs, 9)]), ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
209 TI = LD(&(Rm[WS(rs, 9)]), -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
210 TG = LDW(&(W[TWVL * 34]));
|
cannam@167
|
211 TJ = VZMULJ(TG, VFMACONJ(TI, TH));
|
cannam@167
|
212 T1c = LDW(&(W[TWVL * 36]));
|
cannam@167
|
213 T1d = VZMULIJ(T1c, VFNMSCONJ(TI, TH));
|
cannam@167
|
214 TK = VSUB(TF, TJ);
|
cannam@167
|
215 T3j = VADD(T1f, T1d);
|
cannam@167
|
216 T1g = VSUB(T1d, T1f);
|
cannam@167
|
217 T3b = VADD(TF, TJ);
|
cannam@167
|
218 }
|
cannam@167
|
219 T43 = VADD(T3b, T3c);
|
cannam@167
|
220 T44 = VADD(T3e, T3f);
|
cannam@167
|
221 T4a = VADD(T3j, T3k);
|
cannam@167
|
222 T4b = VADD(T3n, T3m);
|
cannam@167
|
223 T4m = VSUB(T4a, T4b);
|
cannam@167
|
224 T3l = VSUB(T3j, T3k);
|
cannam@167
|
225 T3o = VSUB(T3m, T3n);
|
cannam@167
|
226 T3p = VFMA(LDK(KP414213562), T3o, T3l);
|
cannam@167
|
227 T3E = VFNMS(LDK(KP414213562), T3l, T3o);
|
cannam@167
|
228 TL = VFMA(LDK(KP414213562), TK, TB);
|
cannam@167
|
229 T14 = VFNMS(LDK(KP414213562), T13, TU);
|
cannam@167
|
230 T15 = VSUB(TL, T14);
|
cannam@167
|
231 T2K = VADD(TL, T14);
|
cannam@167
|
232 T1s = VSUB(T1p, T1r);
|
cannam@167
|
233 T1t = VADD(T1g, T1l);
|
cannam@167
|
234 T1u = VFNMS(LDK(KP707106781), T1t, T1s);
|
cannam@167
|
235 T2F = VFMA(LDK(KP707106781), T1t, T1s);
|
cannam@167
|
236 T3d = VSUB(T3b, T3c);
|
cannam@167
|
237 T3g = VSUB(T3e, T3f);
|
cannam@167
|
238 T3h = VADD(T3d, T3g);
|
cannam@167
|
239 T3C = VSUB(T3g, T3d);
|
cannam@167
|
240 T1b = VSUB(T18, T1a);
|
cannam@167
|
241 T1m = VSUB(T1g, T1l);
|
cannam@167
|
242 T1n = VFNMS(LDK(KP707106781), T1m, T1b);
|
cannam@167
|
243 T2E = VFMA(LDK(KP707106781), T1m, T1b);
|
cannam@167
|
244 T28 = VFMA(LDK(KP414213562), TU, T13);
|
cannam@167
|
245 T29 = VFNMS(LDK(KP414213562), TB, TK);
|
cannam@167
|
246 T2a = VSUB(T28, T29);
|
cannam@167
|
247 T2z = VADD(T29, T28);
|
cannam@167
|
248 {
|
cannam@167
|
249 V T4o, T4u, T4r, T4v, T4k, T4n, T4p, T4q, T4s, T4x, T4t, T4w, T3y, T3K, T3H;
|
cannam@167
|
250 V T3L, T3i, T3x, T3D, T3G, T3I, T3N, T3J, T3M, T46, T4g, T4d, T4h, T42, T45;
|
cannam@167
|
251 V T49, T4c, T4e, T4j, T4f, T4i;
|
cannam@167
|
252 T4k = VSUB(T40, T41);
|
cannam@167
|
253 T4n = VADD(T4l, T4m);
|
cannam@167
|
254 T4o = VFMA(LDK(KP707106781), T4n, T4k);
|
cannam@167
|
255 T4u = VFNMS(LDK(KP707106781), T4n, T4k);
|
cannam@167
|
256 T4p = VSUB(T44, T43);
|
cannam@167
|
257 T4q = VSUB(T4m, T4l);
|
cannam@167
|
258 T4r = VFMA(LDK(KP707106781), T4q, T4p);
|
cannam@167
|
259 T4v = VFNMS(LDK(KP707106781), T4q, T4p);
|
cannam@167
|
260 T4s = VCONJ(VMUL(LDK(KP500000000), VFNMSI(T4r, T4o)));
|
cannam@167
|
261 ST(&(Rm[WS(rs, 3)]), T4s, -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
262 T4x = VCONJ(VMUL(LDK(KP500000000), VFMAI(T4v, T4u)));
|
cannam@167
|
263 ST(&(Rm[WS(rs, 11)]), T4x, -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
264 T4t = VMUL(LDK(KP500000000), VFMAI(T4r, T4o));
|
cannam@167
|
265 ST(&(Rp[WS(rs, 4)]), T4t, ms, &(Rp[0]));
|
cannam@167
|
266 T4w = VMUL(LDK(KP500000000), VFNMSI(T4v, T4u));
|
cannam@167
|
267 ST(&(Rp[WS(rs, 12)]), T4w, ms, &(Rp[0]));
|
cannam@167
|
268 T3i = VFNMS(LDK(KP707106781), T3h, T3a);
|
cannam@167
|
269 T3x = VSUB(T3p, T3w);
|
cannam@167
|
270 T3y = VFMA(LDK(KP923879532), T3x, T3i);
|
cannam@167
|
271 T3K = VFNMS(LDK(KP923879532), T3x, T3i);
|
cannam@167
|
272 T3D = VFNMS(LDK(KP707106781), T3C, T3B);
|
cannam@167
|
273 T3G = VSUB(T3E, T3F);
|
cannam@167
|
274 T3H = VFNMS(LDK(KP923879532), T3G, T3D);
|
cannam@167
|
275 T3L = VFMA(LDK(KP923879532), T3G, T3D);
|
cannam@167
|
276 T3I = VMUL(LDK(KP500000000), VFNMSI(T3H, T3y));
|
cannam@167
|
277 ST(&(Rp[WS(rs, 6)]), T3I, ms, &(Rp[0]));
|
cannam@167
|
278 T3N = VMUL(LDK(KP500000000), VFMAI(T3L, T3K));
|
cannam@167
|
279 ST(&(Rp[WS(rs, 10)]), T3N, ms, &(Rp[0]));
|
cannam@167
|
280 T3J = VCONJ(VMUL(LDK(KP500000000), VFMAI(T3H, T3y)));
|
cannam@167
|
281 ST(&(Rm[WS(rs, 5)]), T3J, -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
282 T3M = VCONJ(VMUL(LDK(KP500000000), VFNMSI(T3L, T3K)));
|
cannam@167
|
283 ST(&(Rm[WS(rs, 9)]), T3M, -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
284 T42 = VADD(T40, T41);
|
cannam@167
|
285 T45 = VADD(T43, T44);
|
cannam@167
|
286 T46 = VSUB(T42, T45);
|
cannam@167
|
287 T4g = VADD(T42, T45);
|
cannam@167
|
288 T49 = VADD(T47, T48);
|
cannam@167
|
289 T4c = VADD(T4a, T4b);
|
cannam@167
|
290 T4d = VSUB(T49, T4c);
|
cannam@167
|
291 T4h = VADD(T49, T4c);
|
cannam@167
|
292 T4e = VMUL(LDK(KP500000000), VFMAI(T4d, T46));
|
cannam@167
|
293 ST(&(Rp[WS(rs, 8)]), T4e, ms, &(Rp[0]));
|
cannam@167
|
294 T4j = VCONJ(VMUL(LDK(KP500000000), VADD(T4h, T4g)));
|
cannam@167
|
295 ST(&(Rm[WS(rs, 15)]), T4j, -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
296 T4f = VCONJ(VMUL(LDK(KP500000000), VFNMSI(T4d, T46)));
|
cannam@167
|
297 ST(&(Rm[WS(rs, 7)]), T4f, -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
298 T4i = VMUL(LDK(KP500000000), VSUB(T4g, T4h));
|
cannam@167
|
299 ST(&(Rp[0]), T4i, ms, &(Rp[0]));
|
cannam@167
|
300 }
|
cannam@167
|
301 T3O = VFMA(LDK(KP707106781), T3h, T3a);
|
cannam@167
|
302 T3P = VADD(T3F, T3E);
|
cannam@167
|
303 T3Q = VFMA(LDK(KP923879532), T3P, T3O);
|
cannam@167
|
304 T3W = VFNMS(LDK(KP923879532), T3P, T3O);
|
cannam@167
|
305 T3R = VFMA(LDK(KP707106781), T3C, T3B);
|
cannam@167
|
306 T3S = VADD(T3w, T3p);
|
cannam@167
|
307 T3T = VFMA(LDK(KP923879532), T3S, T3R);
|
cannam@167
|
308 T3X = VFNMS(LDK(KP923879532), T3S, T3R);
|
cannam@167
|
309 T3U = VCONJ(VMUL(LDK(KP500000000), VFNMSI(T3T, T3Q)));
|
cannam@167
|
310 ST(&(Rm[WS(rs, 1)]), T3U, -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
311 T3Z = VCONJ(VMUL(LDK(KP500000000), VFMAI(T3X, T3W)));
|
cannam@167
|
312 ST(&(Rm[WS(rs, 13)]), T3Z, -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
313 T3V = VMUL(LDK(KP500000000), VFMAI(T3T, T3Q));
|
cannam@167
|
314 ST(&(Rp[WS(rs, 2)]), T3V, ms, &(Rp[0]));
|
cannam@167
|
315 T3Y = VMUL(LDK(KP500000000), VFNMSI(T3X, T3W));
|
cannam@167
|
316 ST(&(Rp[WS(rs, 14)]), T3Y, ms, &(Rp[0]));
|
cannam@167
|
317 {
|
cannam@167
|
318 V T2I, T35, T2S, T31, T2P, T34, T2T, T2Y, T2A, T2Z, T2H, T30, T2D, T2G, T2L;
|
cannam@167
|
319 V T2W, T2O, T2X, T2M, T2N, T2Q, T36, T37, T2R, T2U, T32, T33, T2V, T20, T2v;
|
cannam@167
|
320 V T2i, T2r, T2f, T2u, T2j, T2o, T16, T2p, T1Z, T2q, T1v, T1Y, T2b, T2m, T2e;
|
cannam@167
|
321 V T2n, T2c, T2d, T2g, T2w, T2x, T2h, T2k, T2s, T2t, T2l;
|
cannam@167
|
322 T2A = VFNMS(LDK(KP923879532), T2z, T2y);
|
cannam@167
|
323 T2Z = VFMA(LDK(KP923879532), T2K, T2J);
|
cannam@167
|
324 T2D = VFMA(LDK(KP198912367), T2C, T2B);
|
cannam@167
|
325 T2G = VFNMS(LDK(KP198912367), T2F, T2E);
|
cannam@167
|
326 T2H = VSUB(T2D, T2G);
|
cannam@167
|
327 T30 = VADD(T2D, T2G);
|
cannam@167
|
328 T2I = VFMA(LDK(KP980785280), T2H, T2A);
|
cannam@167
|
329 T35 = VFNMS(LDK(KP980785280), T30, T2Z);
|
cannam@167
|
330 T2S = VFNMS(LDK(KP980785280), T2H, T2A);
|
cannam@167
|
331 T31 = VFMA(LDK(KP980785280), T30, T2Z);
|
cannam@167
|
332 T2L = VFNMS(LDK(KP923879532), T2K, T2J);
|
cannam@167
|
333 T2W = VFMA(LDK(KP923879532), T2z, T2y);
|
cannam@167
|
334 T2M = VFMA(LDK(KP198912367), T2E, T2F);
|
cannam@167
|
335 T2N = VFNMS(LDK(KP198912367), T2B, T2C);
|
cannam@167
|
336 T2O = VSUB(T2M, T2N);
|
cannam@167
|
337 T2X = VADD(T2N, T2M);
|
cannam@167
|
338 T2P = VFMA(LDK(KP980785280), T2O, T2L);
|
cannam@167
|
339 T34 = VFNMS(LDK(KP980785280), T2X, T2W);
|
cannam@167
|
340 T2T = VFNMS(LDK(KP980785280), T2O, T2L);
|
cannam@167
|
341 T2Y = VFMA(LDK(KP980785280), T2X, T2W);
|
cannam@167
|
342 T2Q = VCONJ(VMUL(LDK(KP500000000), VFNMSI(T2P, T2I)));
|
cannam@167
|
343 ST(&(Rm[WS(rs, 6)]), T2Q, -ms, &(Rm[0]));
|
cannam@167
|
344 T36 = VCONJ(VMUL(LDK(KP500000000), VFNMSI(T35, T34)));
|
cannam@167
|
345 ST(&(Rm[WS(rs, 14)]), T36, -ms, &(Rm[0]));
|
cannam@167
|
346 T37 = VMUL(LDK(KP500000000), VFMAI(T35, T34));
|
cannam@167
|
347 ST(&(Rp[WS(rs, 15)]), T37, ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
348 T2R = VMUL(LDK(KP500000000), VFMAI(T2P, T2I));
|
cannam@167
|
349 ST(&(Rp[WS(rs, 7)]), T2R, ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
350 T2U = VMUL(LDK(KP500000000), VFNMSI(T2T, T2S));
|
cannam@167
|
351 ST(&(Rp[WS(rs, 9)]), T2U, ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
352 T32 = VMUL(LDK(KP500000000), VFNMSI(T31, T2Y));
|
cannam@167
|
353 ST(&(Rp[WS(rs, 1)]), T32, ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
354 T33 = VCONJ(VMUL(LDK(KP500000000), VFMAI(T31, T2Y)));
|
cannam@167
|
355 ST(&(Rm[0]), T33, -ms, &(Rm[0]));
|
cannam@167
|
356 T2V = VCONJ(VMUL(LDK(KP500000000), VFMAI(T2T, T2S)));
|
cannam@167
|
357 ST(&(Rm[WS(rs, 8)]), T2V, -ms, &(Rm[0]));
|
cannam@167
|
358 T16 = VFNMS(LDK(KP923879532), T15, Ts);
|
cannam@167
|
359 T2p = VFMA(LDK(KP923879532), T2a, T27);
|
cannam@167
|
360 T1v = VFMA(LDK(KP668178637), T1u, T1n);
|
cannam@167
|
361 T1Y = VFNMS(LDK(KP668178637), T1X, T1Q);
|
cannam@167
|
362 T1Z = VSUB(T1v, T1Y);
|
cannam@167
|
363 T2q = VADD(T1Y, T1v);
|
cannam@167
|
364 T20 = VFMA(LDK(KP831469612), T1Z, T16);
|
cannam@167
|
365 T2v = VFNMS(LDK(KP831469612), T2q, T2p);
|
cannam@167
|
366 T2i = VFNMS(LDK(KP831469612), T1Z, T16);
|
cannam@167
|
367 T2r = VFMA(LDK(KP831469612), T2q, T2p);
|
cannam@167
|
368 T2b = VFNMS(LDK(KP923879532), T2a, T27);
|
cannam@167
|
369 T2m = VFMA(LDK(KP923879532), T15, Ts);
|
cannam@167
|
370 T2c = VFNMS(LDK(KP668178637), T1n, T1u);
|
cannam@167
|
371 T2d = VFMA(LDK(KP668178637), T1Q, T1X);
|
cannam@167
|
372 T2e = VSUB(T2c, T2d);
|
cannam@167
|
373 T2n = VADD(T2d, T2c);
|
cannam@167
|
374 T2f = VFNMS(LDK(KP831469612), T2e, T2b);
|
cannam@167
|
375 T2u = VFNMS(LDK(KP831469612), T2n, T2m);
|
cannam@167
|
376 T2j = VFMA(LDK(KP831469612), T2e, T2b);
|
cannam@167
|
377 T2o = VFMA(LDK(KP831469612), T2n, T2m);
|
cannam@167
|
378 T2g = VMUL(LDK(KP500000000), VFNMSI(T2f, T20));
|
cannam@167
|
379 ST(&(Rp[WS(rs, 5)]), T2g, ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
380 T2w = VMUL(LDK(KP500000000), VFNMSI(T2v, T2u));
|
cannam@167
|
381 ST(&(Rp[WS(rs, 13)]), T2w, ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
382 T2x = VCONJ(VMUL(LDK(KP500000000), VFMAI(T2v, T2u)));
|
cannam@167
|
383 ST(&(Rm[WS(rs, 12)]), T2x, -ms, &(Rm[0]));
|
cannam@167
|
384 T2h = VCONJ(VMUL(LDK(KP500000000), VFMAI(T2f, T20)));
|
cannam@167
|
385 ST(&(Rm[WS(rs, 4)]), T2h, -ms, &(Rm[0]));
|
cannam@167
|
386 T2k = VMUL(LDK(KP500000000), VFMAI(T2j, T2i));
|
cannam@167
|
387 ST(&(Rp[WS(rs, 11)]), T2k, ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
388 T2s = VMUL(LDK(KP500000000), VFMAI(T2r, T2o));
|
cannam@167
|
389 ST(&(Rp[WS(rs, 3)]), T2s, ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
390 T2t = VCONJ(VMUL(LDK(KP500000000), VFNMSI(T2r, T2o)));
|
cannam@167
|
391 ST(&(Rm[WS(rs, 2)]), T2t, -ms, &(Rm[0]));
|
cannam@167
|
392 T2l = VCONJ(VMUL(LDK(KP500000000), VFNMSI(T2j, T2i)));
|
cannam@167
|
393 ST(&(Rm[WS(rs, 10)]), T2l, -ms, &(Rm[0]));
|
cannam@167
|
394 }
|
cannam@167
|
395 }
|
cannam@167
|
396 }
|
cannam@167
|
397 VLEAVE();
|
cannam@167
|
398 }
|
cannam@167
|
399
|
cannam@167
|
400 static const tw_instr twinstr[] = {
|
cannam@167
|
401 VTW(1, 1),
|
cannam@167
|
402 VTW(1, 2),
|
cannam@167
|
403 VTW(1, 3),
|
cannam@167
|
404 VTW(1, 4),
|
cannam@167
|
405 VTW(1, 5),
|
cannam@167
|
406 VTW(1, 6),
|
cannam@167
|
407 VTW(1, 7),
|
cannam@167
|
408 VTW(1, 8),
|
cannam@167
|
409 VTW(1, 9),
|
cannam@167
|
410 VTW(1, 10),
|
cannam@167
|
411 VTW(1, 11),
|
cannam@167
|
412 VTW(1, 12),
|
cannam@167
|
413 VTW(1, 13),
|
cannam@167
|
414 VTW(1, 14),
|
cannam@167
|
415 VTW(1, 15),
|
cannam@167
|
416 VTW(1, 16),
|
cannam@167
|
417 VTW(1, 17),
|
cannam@167
|
418 VTW(1, 18),
|
cannam@167
|
419 VTW(1, 19),
|
cannam@167
|
420 VTW(1, 20),
|
cannam@167
|
421 VTW(1, 21),
|
cannam@167
|
422 VTW(1, 22),
|
cannam@167
|
423 VTW(1, 23),
|
cannam@167
|
424 VTW(1, 24),
|
cannam@167
|
425 VTW(1, 25),
|
cannam@167
|
426 VTW(1, 26),
|
cannam@167
|
427 VTW(1, 27),
|
cannam@167
|
428 VTW(1, 28),
|
cannam@167
|
429 VTW(1, 29),
|
cannam@167
|
430 VTW(1, 30),
|
cannam@167
|
431 VTW(1, 31),
|
cannam@167
|
432 {TW_NEXT, VL, 0}
|
cannam@167
|
433 };
|
cannam@167
|
434
|
cannam@167
|
435 static const hc2c_desc desc = { 32, XSIMD_STRING("hc2cfdftv_32"), twinstr, &GENUS, {119, 94, 130, 0} };
|
cannam@167
|
436
|
cannam@167
|
437 void XSIMD(codelet_hc2cfdftv_32) (planner *p) {
|
cannam@167
|
438 X(khc2c_register) (p, hc2cfdftv_32, &desc, HC2C_VIA_DFT);
|
cannam@167
|
439 }
|
cannam@167
|
440 #else
|
cannam@167
|
441
|
cannam@167
|
442 /* Generated by: ../../../genfft/gen_hc2cdft_c.native -simd -compact -variables 4 -pipeline-latency 8 -trivial-stores -variables 32 -no-generate-bytw -n 32 -dit -name hc2cfdftv_32 -include rdft/simd/hc2cfv.h */
|
cannam@167
|
443
|
cannam@167
|
444 /*
|
cannam@167
|
445 * This function contains 249 FP additions, 133 FP multiplications,
|
cannam@167
|
446 * (or, 233 additions, 117 multiplications, 16 fused multiply/add),
|
cannam@167
|
447 * 130 stack variables, 9 constants, and 64 memory accesses
|
cannam@167
|
448 */
|
cannam@167
|
449 #include "rdft/simd/hc2cfv.h"
|
cannam@167
|
450
|
cannam@167
|
451 static void hc2cfdftv_32(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
|
cannam@167
|
452 {
|
cannam@167
|
453 DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
|
cannam@167
|
454 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
|
cannam@167
|
455 DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
|
cannam@167
|
456 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
|
cannam@167
|
457 DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
|
cannam@167
|
458 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
|
cannam@167
|
459 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
|
cannam@167
|
460 DVK(KP353553390, +0.353553390593273762200422181052424519642417969);
|
cannam@167
|
461 DVK(KP500000000, +0.500000000000000000000000000000000000000000000);
|
cannam@167
|
462 {
|
cannam@167
|
463 INT m;
|
cannam@167
|
464 for (m = mb, W = W + ((mb - 1) * ((TWVL / VL) * 62)); m < me; m = m + VL, Rp = Rp + (VL * ms), Ip = Ip + (VL * ms), Rm = Rm - (VL * ms), Im = Im - (VL * ms), W = W + (TWVL * 62), MAKE_VOLATILE_STRIDE(128, rs)) {
|
cannam@167
|
465 V Ta, T2m, Tx, T2h, T3R, T4h, T3q, T4g, T3B, T4n, T3E, T4o, T1B, T2S, T1O;
|
cannam@167
|
466 V T2R, TV, T2p, T1i, T2o, T3L, T4q, T3I, T4r, T3w, T4k, T3t, T4j, T26, T2V;
|
cannam@167
|
467 V T2d, T2U;
|
cannam@167
|
468 {
|
cannam@167
|
469 V T4, T1m, T1H, T2j, T1M, T2l, T9, T1o, Tf, T1r, Tq, T1w, Tv, T1y, Tk;
|
cannam@167
|
470 V T1t, Tl, Tw, T3P, T3Q, T3o, T3p, T3z, T3A, T3C, T3D, T1p, T1N, T1A, T1C;
|
cannam@167
|
471 V T1u, T1z;
|
cannam@167
|
472 {
|
cannam@167
|
473 V T1, T3, T2, T1l, T1G, T1F, T1E, T1D, T2i, T1L, T1K, T1J, T1I, T2k, T6;
|
cannam@167
|
474 V T8, T7, T5, T1n, Tc, Te, Td, Tb, T1q, Tn, Tp, To, Tm, T1v, Ts;
|
cannam@167
|
475 V Tu, Tt, Tr, T1x, Th, Tj, Ti, Tg, T1s;
|
cannam@167
|
476 T1 = LD(&(Rp[0]), ms, &(Rp[0]));
|
cannam@167
|
477 T2 = LD(&(Rm[0]), -ms, &(Rm[0]));
|
cannam@167
|
478 T3 = VCONJ(T2);
|
cannam@167
|
479 T4 = VADD(T1, T3);
|
cannam@167
|
480 T1l = LDW(&(W[0]));
|
cannam@167
|
481 T1m = VZMULIJ(T1l, VSUB(T3, T1));
|
cannam@167
|
482 T1G = LD(&(Rp[WS(rs, 4)]), ms, &(Rp[0]));
|
cannam@167
|
483 T1E = LD(&(Rm[WS(rs, 4)]), -ms, &(Rm[0]));
|
cannam@167
|
484 T1F = VCONJ(T1E);
|
cannam@167
|
485 T1D = LDW(&(W[TWVL * 16]));
|
cannam@167
|
486 T1H = VZMULIJ(T1D, VSUB(T1F, T1G));
|
cannam@167
|
487 T2i = LDW(&(W[TWVL * 14]));
|
cannam@167
|
488 T2j = VZMULJ(T2i, VADD(T1G, T1F));
|
cannam@167
|
489 T1L = LD(&(Rp[WS(rs, 12)]), ms, &(Rp[0]));
|
cannam@167
|
490 T1J = LD(&(Rm[WS(rs, 12)]), -ms, &(Rm[0]));
|
cannam@167
|
491 T1K = VCONJ(T1J);
|
cannam@167
|
492 T1I = LDW(&(W[TWVL * 48]));
|
cannam@167
|
493 T1M = VZMULIJ(T1I, VSUB(T1K, T1L));
|
cannam@167
|
494 T2k = LDW(&(W[TWVL * 46]));
|
cannam@167
|
495 T2l = VZMULJ(T2k, VADD(T1L, T1K));
|
cannam@167
|
496 T6 = LD(&(Rp[WS(rs, 8)]), ms, &(Rp[0]));
|
cannam@167
|
497 T7 = LD(&(Rm[WS(rs, 8)]), -ms, &(Rm[0]));
|
cannam@167
|
498 T8 = VCONJ(T7);
|
cannam@167
|
499 T5 = LDW(&(W[TWVL * 30]));
|
cannam@167
|
500 T9 = VZMULJ(T5, VADD(T6, T8));
|
cannam@167
|
501 T1n = LDW(&(W[TWVL * 32]));
|
cannam@167
|
502 T1o = VZMULIJ(T1n, VSUB(T8, T6));
|
cannam@167
|
503 Tc = LD(&(Rp[WS(rs, 2)]), ms, &(Rp[0]));
|
cannam@167
|
504 Td = LD(&(Rm[WS(rs, 2)]), -ms, &(Rm[0]));
|
cannam@167
|
505 Te = VCONJ(Td);
|
cannam@167
|
506 Tb = LDW(&(W[TWVL * 6]));
|
cannam@167
|
507 Tf = VZMULJ(Tb, VADD(Tc, Te));
|
cannam@167
|
508 T1q = LDW(&(W[TWVL * 8]));
|
cannam@167
|
509 T1r = VZMULIJ(T1q, VSUB(Te, Tc));
|
cannam@167
|
510 Tn = LD(&(Rp[WS(rs, 14)]), ms, &(Rp[0]));
|
cannam@167
|
511 To = LD(&(Rm[WS(rs, 14)]), -ms, &(Rm[0]));
|
cannam@167
|
512 Tp = VCONJ(To);
|
cannam@167
|
513 Tm = LDW(&(W[TWVL * 54]));
|
cannam@167
|
514 Tq = VZMULJ(Tm, VADD(Tn, Tp));
|
cannam@167
|
515 T1v = LDW(&(W[TWVL * 56]));
|
cannam@167
|
516 T1w = VZMULIJ(T1v, VSUB(Tp, Tn));
|
cannam@167
|
517 Ts = LD(&(Rp[WS(rs, 6)]), ms, &(Rp[0]));
|
cannam@167
|
518 Tt = LD(&(Rm[WS(rs, 6)]), -ms, &(Rm[0]));
|
cannam@167
|
519 Tu = VCONJ(Tt);
|
cannam@167
|
520 Tr = LDW(&(W[TWVL * 22]));
|
cannam@167
|
521 Tv = VZMULJ(Tr, VADD(Ts, Tu));
|
cannam@167
|
522 T1x = LDW(&(W[TWVL * 24]));
|
cannam@167
|
523 T1y = VZMULIJ(T1x, VSUB(Tu, Ts));
|
cannam@167
|
524 Th = LD(&(Rp[WS(rs, 10)]), ms, &(Rp[0]));
|
cannam@167
|
525 Ti = LD(&(Rm[WS(rs, 10)]), -ms, &(Rm[0]));
|
cannam@167
|
526 Tj = VCONJ(Ti);
|
cannam@167
|
527 Tg = LDW(&(W[TWVL * 38]));
|
cannam@167
|
528 Tk = VZMULJ(Tg, VADD(Th, Tj));
|
cannam@167
|
529 T1s = LDW(&(W[TWVL * 40]));
|
cannam@167
|
530 T1t = VZMULIJ(T1s, VSUB(Tj, Th));
|
cannam@167
|
531 }
|
cannam@167
|
532 Ta = VMUL(LDK(KP500000000), VSUB(T4, T9));
|
cannam@167
|
533 T2m = VSUB(T2j, T2l);
|
cannam@167
|
534 Tl = VSUB(Tf, Tk);
|
cannam@167
|
535 Tw = VSUB(Tq, Tv);
|
cannam@167
|
536 Tx = VMUL(LDK(KP353553390), VADD(Tl, Tw));
|
cannam@167
|
537 T2h = VMUL(LDK(KP707106781), VSUB(Tw, Tl));
|
cannam@167
|
538 T3P = VADD(Tq, Tv);
|
cannam@167
|
539 T3Q = VADD(Tf, Tk);
|
cannam@167
|
540 T3R = VSUB(T3P, T3Q);
|
cannam@167
|
541 T4h = VADD(T3Q, T3P);
|
cannam@167
|
542 T3o = VADD(T4, T9);
|
cannam@167
|
543 T3p = VADD(T2j, T2l);
|
cannam@167
|
544 T3q = VMUL(LDK(KP500000000), VSUB(T3o, T3p));
|
cannam@167
|
545 T4g = VADD(T3o, T3p);
|
cannam@167
|
546 T3z = VADD(T1m, T1o);
|
cannam@167
|
547 T3A = VADD(T1H, T1M);
|
cannam@167
|
548 T3B = VSUB(T3z, T3A);
|
cannam@167
|
549 T4n = VADD(T3z, T3A);
|
cannam@167
|
550 T3C = VADD(T1w, T1y);
|
cannam@167
|
551 T3D = VADD(T1r, T1t);
|
cannam@167
|
552 T3E = VSUB(T3C, T3D);
|
cannam@167
|
553 T4o = VADD(T3D, T3C);
|
cannam@167
|
554 T1p = VSUB(T1m, T1o);
|
cannam@167
|
555 T1N = VSUB(T1H, T1M);
|
cannam@167
|
556 T1u = VSUB(T1r, T1t);
|
cannam@167
|
557 T1z = VSUB(T1w, T1y);
|
cannam@167
|
558 T1A = VMUL(LDK(KP707106781), VADD(T1u, T1z));
|
cannam@167
|
559 T1C = VMUL(LDK(KP707106781), VSUB(T1z, T1u));
|
cannam@167
|
560 T1B = VADD(T1p, T1A);
|
cannam@167
|
561 T2S = VADD(T1N, T1C);
|
cannam@167
|
562 T1O = VSUB(T1C, T1N);
|
cannam@167
|
563 T2R = VSUB(T1p, T1A);
|
cannam@167
|
564 }
|
cannam@167
|
565 {
|
cannam@167
|
566 V TD, T1R, T1b, T29, T1g, T2b, TI, T1T, TO, T1Y, T10, T22, T15, T24, TT;
|
cannam@167
|
567 V T1W, TJ, TU, T16, T1h, T3J, T3K, T3G, T3H, T3u, T3v, T3r, T3s, T25, T2c;
|
cannam@167
|
568 V T20, T27, T1U, T1Z;
|
cannam@167
|
569 {
|
cannam@167
|
570 V TA, TC, TB, Tz, T1Q, T18, T1a, T19, T17, T28, T1d, T1f, T1e, T1c, T2a;
|
cannam@167
|
571 V TF, TH, TG, TE, T1S, TL, TN, TM, TK, T1X, TX, TZ, TY, TW, T21;
|
cannam@167
|
572 V T12, T14, T13, T11, T23, TQ, TS, TR, TP, T1V;
|
cannam@167
|
573 TA = LD(&(Rp[WS(rs, 1)]), ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
574 TB = LD(&(Rm[WS(rs, 1)]), -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
575 TC = VCONJ(TB);
|
cannam@167
|
576 Tz = LDW(&(W[TWVL * 2]));
|
cannam@167
|
577 TD = VZMULJ(Tz, VADD(TA, TC));
|
cannam@167
|
578 T1Q = LDW(&(W[TWVL * 4]));
|
cannam@167
|
579 T1R = VZMULIJ(T1Q, VSUB(TC, TA));
|
cannam@167
|
580 T18 = LD(&(Rp[WS(rs, 3)]), ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
581 T19 = LD(&(Rm[WS(rs, 3)]), -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
582 T1a = VCONJ(T19);
|
cannam@167
|
583 T17 = LDW(&(W[TWVL * 10]));
|
cannam@167
|
584 T1b = VZMULJ(T17, VADD(T18, T1a));
|
cannam@167
|
585 T28 = LDW(&(W[TWVL * 12]));
|
cannam@167
|
586 T29 = VZMULIJ(T28, VSUB(T1a, T18));
|
cannam@167
|
587 T1d = LD(&(Rp[WS(rs, 11)]), ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
588 T1e = LD(&(Rm[WS(rs, 11)]), -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
589 T1f = VCONJ(T1e);
|
cannam@167
|
590 T1c = LDW(&(W[TWVL * 42]));
|
cannam@167
|
591 T1g = VZMULJ(T1c, VADD(T1d, T1f));
|
cannam@167
|
592 T2a = LDW(&(W[TWVL * 44]));
|
cannam@167
|
593 T2b = VZMULIJ(T2a, VSUB(T1f, T1d));
|
cannam@167
|
594 TF = LD(&(Rp[WS(rs, 9)]), ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
595 TG = LD(&(Rm[WS(rs, 9)]), -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
596 TH = VCONJ(TG);
|
cannam@167
|
597 TE = LDW(&(W[TWVL * 34]));
|
cannam@167
|
598 TI = VZMULJ(TE, VADD(TF, TH));
|
cannam@167
|
599 T1S = LDW(&(W[TWVL * 36]));
|
cannam@167
|
600 T1T = VZMULIJ(T1S, VSUB(TH, TF));
|
cannam@167
|
601 TL = LD(&(Rp[WS(rs, 5)]), ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
602 TM = LD(&(Rm[WS(rs, 5)]), -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
603 TN = VCONJ(TM);
|
cannam@167
|
604 TK = LDW(&(W[TWVL * 18]));
|
cannam@167
|
605 TO = VZMULJ(TK, VADD(TL, TN));
|
cannam@167
|
606 T1X = LDW(&(W[TWVL * 20]));
|
cannam@167
|
607 T1Y = VZMULIJ(T1X, VSUB(TN, TL));
|
cannam@167
|
608 TX = LD(&(Rp[WS(rs, 15)]), ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
609 TY = LD(&(Rm[WS(rs, 15)]), -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
610 TZ = VCONJ(TY);
|
cannam@167
|
611 TW = LDW(&(W[TWVL * 58]));
|
cannam@167
|
612 T10 = VZMULJ(TW, VADD(TX, TZ));
|
cannam@167
|
613 T21 = LDW(&(W[TWVL * 60]));
|
cannam@167
|
614 T22 = VZMULIJ(T21, VSUB(TZ, TX));
|
cannam@167
|
615 T12 = LD(&(Rp[WS(rs, 7)]), ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
616 T13 = LD(&(Rm[WS(rs, 7)]), -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
617 T14 = VCONJ(T13);
|
cannam@167
|
618 T11 = LDW(&(W[TWVL * 26]));
|
cannam@167
|
619 T15 = VZMULJ(T11, VADD(T12, T14));
|
cannam@167
|
620 T23 = LDW(&(W[TWVL * 28]));
|
cannam@167
|
621 T24 = VZMULIJ(T23, VSUB(T14, T12));
|
cannam@167
|
622 TQ = LD(&(Rp[WS(rs, 13)]), ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
623 TR = LD(&(Rm[WS(rs, 13)]), -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
624 TS = VCONJ(TR);
|
cannam@167
|
625 TP = LDW(&(W[TWVL * 50]));
|
cannam@167
|
626 TT = VZMULJ(TP, VADD(TQ, TS));
|
cannam@167
|
627 T1V = LDW(&(W[TWVL * 52]));
|
cannam@167
|
628 T1W = VZMULIJ(T1V, VSUB(TS, TQ));
|
cannam@167
|
629 }
|
cannam@167
|
630 TJ = VSUB(TD, TI);
|
cannam@167
|
631 TU = VSUB(TO, TT);
|
cannam@167
|
632 TV = VFNMS(LDK(KP382683432), TU, VMUL(LDK(KP923879532), TJ));
|
cannam@167
|
633 T2p = VFMA(LDK(KP382683432), TJ, VMUL(LDK(KP923879532), TU));
|
cannam@167
|
634 T16 = VSUB(T10, T15);
|
cannam@167
|
635 T1h = VSUB(T1b, T1g);
|
cannam@167
|
636 T1i = VFMA(LDK(KP923879532), T16, VMUL(LDK(KP382683432), T1h));
|
cannam@167
|
637 T2o = VFNMS(LDK(KP923879532), T1h, VMUL(LDK(KP382683432), T16));
|
cannam@167
|
638 T3J = VADD(T1Y, T1W);
|
cannam@167
|
639 T3K = VADD(T1R, T1T);
|
cannam@167
|
640 T3L = VSUB(T3J, T3K);
|
cannam@167
|
641 T4q = VADD(T3K, T3J);
|
cannam@167
|
642 T3G = VADD(T22, T24);
|
cannam@167
|
643 T3H = VADD(T29, T2b);
|
cannam@167
|
644 T3I = VSUB(T3G, T3H);
|
cannam@167
|
645 T4r = VADD(T3G, T3H);
|
cannam@167
|
646 T3u = VADD(T10, T15);
|
cannam@167
|
647 T3v = VADD(T1b, T1g);
|
cannam@167
|
648 T3w = VSUB(T3u, T3v);
|
cannam@167
|
649 T4k = VADD(T3u, T3v);
|
cannam@167
|
650 T3r = VADD(TD, TI);
|
cannam@167
|
651 T3s = VADD(TO, TT);
|
cannam@167
|
652 T3t = VSUB(T3r, T3s);
|
cannam@167
|
653 T4j = VADD(T3r, T3s);
|
cannam@167
|
654 T25 = VSUB(T22, T24);
|
cannam@167
|
655 T2c = VSUB(T29, T2b);
|
cannam@167
|
656 T1U = VSUB(T1R, T1T);
|
cannam@167
|
657 T1Z = VSUB(T1W, T1Y);
|
cannam@167
|
658 T20 = VMUL(LDK(KP707106781), VADD(T1U, T1Z));
|
cannam@167
|
659 T27 = VMUL(LDK(KP707106781), VSUB(T1Z, T1U));
|
cannam@167
|
660 T26 = VADD(T20, T25);
|
cannam@167
|
661 T2V = VADD(T27, T2c);
|
cannam@167
|
662 T2d = VSUB(T27, T2c);
|
cannam@167
|
663 T2U = VSUB(T25, T20);
|
cannam@167
|
664 }
|
cannam@167
|
665 {
|
cannam@167
|
666 V T4m, T4w, T4t, T4x, T4i, T4l, T4p, T4s, T4u, T4z, T4v, T4y, T4E, T4L, T4H;
|
cannam@167
|
667 V T4K, T4A, T4F, T4D, T4G, T4B, T4C, T4I, T4N, T4J, T4M, T3O, T4c, T4d, T3X;
|
cannam@167
|
668 V T40, T46, T49, T41, T3y, T47, T3T, T45, T3N, T44, T3W, T48, T3x, T3S, T3F;
|
cannam@167
|
669 V T3M, T3U, T3V, T3Y, T4e, T4f, T3Z, T42, T4a, T4b, T43;
|
cannam@167
|
670 T4i = VADD(T4g, T4h);
|
cannam@167
|
671 T4l = VADD(T4j, T4k);
|
cannam@167
|
672 T4m = VADD(T4i, T4l);
|
cannam@167
|
673 T4w = VSUB(T4i, T4l);
|
cannam@167
|
674 T4p = VADD(T4n, T4o);
|
cannam@167
|
675 T4s = VADD(T4q, T4r);
|
cannam@167
|
676 T4t = VADD(T4p, T4s);
|
cannam@167
|
677 T4x = VBYI(VSUB(T4s, T4p));
|
cannam@167
|
678 T4u = VCONJ(VMUL(LDK(KP500000000), VSUB(T4m, T4t)));
|
cannam@167
|
679 ST(&(Rm[WS(rs, 15)]), T4u, -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
680 T4z = VMUL(LDK(KP500000000), VADD(T4w, T4x));
|
cannam@167
|
681 ST(&(Rp[WS(rs, 8)]), T4z, ms, &(Rp[0]));
|
cannam@167
|
682 T4v = VMUL(LDK(KP500000000), VADD(T4m, T4t));
|
cannam@167
|
683 ST(&(Rp[0]), T4v, ms, &(Rp[0]));
|
cannam@167
|
684 T4y = VCONJ(VMUL(LDK(KP500000000), VSUB(T4w, T4x)));
|
cannam@167
|
685 ST(&(Rm[WS(rs, 7)]), T4y, -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
686 T4A = VMUL(LDK(KP500000000), VSUB(T4g, T4h));
|
cannam@167
|
687 T4F = VSUB(T4k, T4j);
|
cannam@167
|
688 T4B = VSUB(T4n, T4o);
|
cannam@167
|
689 T4C = VSUB(T4r, T4q);
|
cannam@167
|
690 T4D = VMUL(LDK(KP353553390), VADD(T4B, T4C));
|
cannam@167
|
691 T4G = VMUL(LDK(KP707106781), VSUB(T4C, T4B));
|
cannam@167
|
692 T4E = VADD(T4A, T4D);
|
cannam@167
|
693 T4L = VMUL(LDK(KP500000000), VBYI(VSUB(T4G, T4F)));
|
cannam@167
|
694 T4H = VMUL(LDK(KP500000000), VBYI(VADD(T4F, T4G)));
|
cannam@167
|
695 T4K = VSUB(T4A, T4D);
|
cannam@167
|
696 T4I = VCONJ(VSUB(T4E, T4H));
|
cannam@167
|
697 ST(&(Rm[WS(rs, 3)]), T4I, -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
698 T4N = VADD(T4K, T4L);
|
cannam@167
|
699 ST(&(Rp[WS(rs, 12)]), T4N, ms, &(Rp[0]));
|
cannam@167
|
700 T4J = VADD(T4E, T4H);
|
cannam@167
|
701 ST(&(Rp[WS(rs, 4)]), T4J, ms, &(Rp[0]));
|
cannam@167
|
702 T4M = VCONJ(VSUB(T4K, T4L));
|
cannam@167
|
703 ST(&(Rm[WS(rs, 11)]), T4M, -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
704 T3x = VMUL(LDK(KP353553390), VADD(T3t, T3w));
|
cannam@167
|
705 T3y = VADD(T3q, T3x);
|
cannam@167
|
706 T47 = VSUB(T3q, T3x);
|
cannam@167
|
707 T3S = VMUL(LDK(KP707106781), VSUB(T3w, T3t));
|
cannam@167
|
708 T3T = VADD(T3R, T3S);
|
cannam@167
|
709 T45 = VSUB(T3S, T3R);
|
cannam@167
|
710 T3F = VFMA(LDK(KP923879532), T3B, VMUL(LDK(KP382683432), T3E));
|
cannam@167
|
711 T3M = VFNMS(LDK(KP382683432), T3L, VMUL(LDK(KP923879532), T3I));
|
cannam@167
|
712 T3N = VMUL(LDK(KP500000000), VADD(T3F, T3M));
|
cannam@167
|
713 T44 = VSUB(T3M, T3F);
|
cannam@167
|
714 T3U = VFNMS(LDK(KP382683432), T3B, VMUL(LDK(KP923879532), T3E));
|
cannam@167
|
715 T3V = VFMA(LDK(KP923879532), T3L, VMUL(LDK(KP382683432), T3I));
|
cannam@167
|
716 T3W = VADD(T3U, T3V);
|
cannam@167
|
717 T48 = VMUL(LDK(KP500000000), VSUB(T3V, T3U));
|
cannam@167
|
718 T3O = VADD(T3y, T3N);
|
cannam@167
|
719 T4c = VMUL(LDK(KP500000000), VBYI(VADD(T45, T44)));
|
cannam@167
|
720 T4d = VADD(T47, T48);
|
cannam@167
|
721 T3X = VMUL(LDK(KP500000000), VBYI(VADD(T3T, T3W)));
|
cannam@167
|
722 T40 = VSUB(T3y, T3N);
|
cannam@167
|
723 T46 = VMUL(LDK(KP500000000), VBYI(VSUB(T44, T45)));
|
cannam@167
|
724 T49 = VSUB(T47, T48);
|
cannam@167
|
725 T41 = VMUL(LDK(KP500000000), VBYI(VSUB(T3W, T3T)));
|
cannam@167
|
726 T3Y = VCONJ(VSUB(T3O, T3X));
|
cannam@167
|
727 ST(&(Rm[WS(rs, 1)]), T3Y, -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
728 T4e = VADD(T4c, T4d);
|
cannam@167
|
729 ST(&(Rp[WS(rs, 6)]), T4e, ms, &(Rp[0]));
|
cannam@167
|
730 T4f = VCONJ(VSUB(T4d, T4c));
|
cannam@167
|
731 ST(&(Rm[WS(rs, 5)]), T4f, -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
732 T3Z = VADD(T3O, T3X);
|
cannam@167
|
733 ST(&(Rp[WS(rs, 2)]), T3Z, ms, &(Rp[0]));
|
cannam@167
|
734 T42 = VCONJ(VSUB(T40, T41));
|
cannam@167
|
735 ST(&(Rm[WS(rs, 13)]), T42, -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
736 T4a = VADD(T46, T49);
|
cannam@167
|
737 ST(&(Rp[WS(rs, 10)]), T4a, ms, &(Rp[0]));
|
cannam@167
|
738 T4b = VCONJ(VSUB(T49, T46));
|
cannam@167
|
739 ST(&(Rm[WS(rs, 9)]), T4b, -ms, &(Rm[WS(rs, 1)]));
|
cannam@167
|
740 T43 = VADD(T40, T41);
|
cannam@167
|
741 ST(&(Rp[WS(rs, 14)]), T43, ms, &(Rp[0]));
|
cannam@167
|
742 {
|
cannam@167
|
743 V T2g, T2K, T2L, T2v, T2y, T2E, T2H, T2z, T1k, T2F, T2u, T2G, T2f, T2C, T2r;
|
cannam@167
|
744 V T2D, Ty, T1j, T2s, T2t, T1P, T2e, T2n, T2q, T2w, T2M, T2N, T2x, T2A, T2I;
|
cannam@167
|
745 V T2J, T2B;
|
cannam@167
|
746 Ty = VADD(Ta, Tx);
|
cannam@167
|
747 T1j = VMUL(LDK(KP500000000), VADD(TV, T1i));
|
cannam@167
|
748 T1k = VADD(Ty, T1j);
|
cannam@167
|
749 T2F = VSUB(Ty, T1j);
|
cannam@167
|
750 T2s = VFNMS(LDK(KP195090322), T1B, VMUL(LDK(KP980785280), T1O));
|
cannam@167
|
751 T2t = VFMA(LDK(KP195090322), T26, VMUL(LDK(KP980785280), T2d));
|
cannam@167
|
752 T2u = VADD(T2s, T2t);
|
cannam@167
|
753 T2G = VMUL(LDK(KP500000000), VSUB(T2t, T2s));
|
cannam@167
|
754 T1P = VFMA(LDK(KP980785280), T1B, VMUL(LDK(KP195090322), T1O));
|
cannam@167
|
755 T2e = VFNMS(LDK(KP195090322), T2d, VMUL(LDK(KP980785280), T26));
|
cannam@167
|
756 T2f = VMUL(LDK(KP500000000), VADD(T1P, T2e));
|
cannam@167
|
757 T2C = VSUB(T2e, T1P);
|
cannam@167
|
758 T2n = VSUB(T2h, T2m);
|
cannam@167
|
759 T2q = VSUB(T2o, T2p);
|
cannam@167
|
760 T2r = VADD(T2n, T2q);
|
cannam@167
|
761 T2D = VSUB(T2q, T2n);
|
cannam@167
|
762 T2g = VADD(T1k, T2f);
|
cannam@167
|
763 T2K = VMUL(LDK(KP500000000), VBYI(VADD(T2D, T2C)));
|
cannam@167
|
764 T2L = VADD(T2F, T2G);
|
cannam@167
|
765 T2v = VMUL(LDK(KP500000000), VBYI(VADD(T2r, T2u)));
|
cannam@167
|
766 T2y = VSUB(T1k, T2f);
|
cannam@167
|
767 T2E = VMUL(LDK(KP500000000), VBYI(VSUB(T2C, T2D)));
|
cannam@167
|
768 T2H = VSUB(T2F, T2G);
|
cannam@167
|
769 T2z = VMUL(LDK(KP500000000), VBYI(VSUB(T2u, T2r)));
|
cannam@167
|
770 T2w = VCONJ(VSUB(T2g, T2v));
|
cannam@167
|
771 ST(&(Rm[0]), T2w, -ms, &(Rm[0]));
|
cannam@167
|
772 T2M = VADD(T2K, T2L);
|
cannam@167
|
773 ST(&(Rp[WS(rs, 7)]), T2M, ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
774 T2N = VCONJ(VSUB(T2L, T2K));
|
cannam@167
|
775 ST(&(Rm[WS(rs, 6)]), T2N, -ms, &(Rm[0]));
|
cannam@167
|
776 T2x = VADD(T2g, T2v);
|
cannam@167
|
777 ST(&(Rp[WS(rs, 1)]), T2x, ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
778 T2A = VCONJ(VSUB(T2y, T2z));
|
cannam@167
|
779 ST(&(Rm[WS(rs, 14)]), T2A, -ms, &(Rm[0]));
|
cannam@167
|
780 T2I = VADD(T2E, T2H);
|
cannam@167
|
781 ST(&(Rp[WS(rs, 9)]), T2I, ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
782 T2J = VCONJ(VSUB(T2H, T2E));
|
cannam@167
|
783 ST(&(Rm[WS(rs, 8)]), T2J, -ms, &(Rm[0]));
|
cannam@167
|
784 T2B = VADD(T2y, T2z);
|
cannam@167
|
785 ST(&(Rp[WS(rs, 15)]), T2B, ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
786 }
|
cannam@167
|
787 {
|
cannam@167
|
788 V T2Y, T3k, T3l, T35, T38, T3e, T3h, T39, T2Q, T3f, T34, T3g, T2X, T3c, T31;
|
cannam@167
|
789 V T3d, T2O, T2P, T32, T33, T2T, T2W, T2Z, T30, T36, T3m, T3n, T37, T3a, T3i;
|
cannam@167
|
790 V T3j, T3b;
|
cannam@167
|
791 T2O = VSUB(Ta, Tx);
|
cannam@167
|
792 T2P = VMUL(LDK(KP500000000), VADD(T2p, T2o));
|
cannam@167
|
793 T2Q = VADD(T2O, T2P);
|
cannam@167
|
794 T3f = VSUB(T2O, T2P);
|
cannam@167
|
795 T32 = VFNMS(LDK(KP555570233), T2R, VMUL(LDK(KP831469612), T2S));
|
cannam@167
|
796 T33 = VFMA(LDK(KP555570233), T2U, VMUL(LDK(KP831469612), T2V));
|
cannam@167
|
797 T34 = VADD(T32, T33);
|
cannam@167
|
798 T3g = VMUL(LDK(KP500000000), VSUB(T33, T32));
|
cannam@167
|
799 T2T = VFMA(LDK(KP831469612), T2R, VMUL(LDK(KP555570233), T2S));
|
cannam@167
|
800 T2W = VFNMS(LDK(KP555570233), T2V, VMUL(LDK(KP831469612), T2U));
|
cannam@167
|
801 T2X = VMUL(LDK(KP500000000), VADD(T2T, T2W));
|
cannam@167
|
802 T3c = VSUB(T2W, T2T);
|
cannam@167
|
803 T2Z = VADD(T2m, T2h);
|
cannam@167
|
804 T30 = VSUB(T1i, TV);
|
cannam@167
|
805 T31 = VADD(T2Z, T30);
|
cannam@167
|
806 T3d = VSUB(T30, T2Z);
|
cannam@167
|
807 T2Y = VADD(T2Q, T2X);
|
cannam@167
|
808 T3k = VMUL(LDK(KP500000000), VBYI(VADD(T3d, T3c)));
|
cannam@167
|
809 T3l = VADD(T3f, T3g);
|
cannam@167
|
810 T35 = VMUL(LDK(KP500000000), VBYI(VADD(T31, T34)));
|
cannam@167
|
811 T38 = VSUB(T2Q, T2X);
|
cannam@167
|
812 T3e = VMUL(LDK(KP500000000), VBYI(VSUB(T3c, T3d)));
|
cannam@167
|
813 T3h = VSUB(T3f, T3g);
|
cannam@167
|
814 T39 = VMUL(LDK(KP500000000), VBYI(VSUB(T34, T31)));
|
cannam@167
|
815 T36 = VCONJ(VSUB(T2Y, T35));
|
cannam@167
|
816 ST(&(Rm[WS(rs, 2)]), T36, -ms, &(Rm[0]));
|
cannam@167
|
817 T3m = VADD(T3k, T3l);
|
cannam@167
|
818 ST(&(Rp[WS(rs, 5)]), T3m, ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
819 T3n = VCONJ(VSUB(T3l, T3k));
|
cannam@167
|
820 ST(&(Rm[WS(rs, 4)]), T3n, -ms, &(Rm[0]));
|
cannam@167
|
821 T37 = VADD(T2Y, T35);
|
cannam@167
|
822 ST(&(Rp[WS(rs, 3)]), T37, ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
823 T3a = VCONJ(VSUB(T38, T39));
|
cannam@167
|
824 ST(&(Rm[WS(rs, 12)]), T3a, -ms, &(Rm[0]));
|
cannam@167
|
825 T3i = VADD(T3e, T3h);
|
cannam@167
|
826 ST(&(Rp[WS(rs, 11)]), T3i, ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
827 T3j = VCONJ(VSUB(T3h, T3e));
|
cannam@167
|
828 ST(&(Rm[WS(rs, 10)]), T3j, -ms, &(Rm[0]));
|
cannam@167
|
829 T3b = VADD(T38, T39);
|
cannam@167
|
830 ST(&(Rp[WS(rs, 13)]), T3b, ms, &(Rp[WS(rs, 1)]));
|
cannam@167
|
831 }
|
cannam@167
|
832 }
|
cannam@167
|
833 }
|
cannam@167
|
834 }
|
cannam@167
|
835 VLEAVE();
|
cannam@167
|
836 }
|
cannam@167
|
837
|
cannam@167
|
838 static const tw_instr twinstr[] = {
|
cannam@167
|
839 VTW(1, 1),
|
cannam@167
|
840 VTW(1, 2),
|
cannam@167
|
841 VTW(1, 3),
|
cannam@167
|
842 VTW(1, 4),
|
cannam@167
|
843 VTW(1, 5),
|
cannam@167
|
844 VTW(1, 6),
|
cannam@167
|
845 VTW(1, 7),
|
cannam@167
|
846 VTW(1, 8),
|
cannam@167
|
847 VTW(1, 9),
|
cannam@167
|
848 VTW(1, 10),
|
cannam@167
|
849 VTW(1, 11),
|
cannam@167
|
850 VTW(1, 12),
|
cannam@167
|
851 VTW(1, 13),
|
cannam@167
|
852 VTW(1, 14),
|
cannam@167
|
853 VTW(1, 15),
|
cannam@167
|
854 VTW(1, 16),
|
cannam@167
|
855 VTW(1, 17),
|
cannam@167
|
856 VTW(1, 18),
|
cannam@167
|
857 VTW(1, 19),
|
cannam@167
|
858 VTW(1, 20),
|
cannam@167
|
859 VTW(1, 21),
|
cannam@167
|
860 VTW(1, 22),
|
cannam@167
|
861 VTW(1, 23),
|
cannam@167
|
862 VTW(1, 24),
|
cannam@167
|
863 VTW(1, 25),
|
cannam@167
|
864 VTW(1, 26),
|
cannam@167
|
865 VTW(1, 27),
|
cannam@167
|
866 VTW(1, 28),
|
cannam@167
|
867 VTW(1, 29),
|
cannam@167
|
868 VTW(1, 30),
|
cannam@167
|
869 VTW(1, 31),
|
cannam@167
|
870 {TW_NEXT, VL, 0}
|
cannam@167
|
871 };
|
cannam@167
|
872
|
cannam@167
|
873 static const hc2c_desc desc = { 32, XSIMD_STRING("hc2cfdftv_32"), twinstr, &GENUS, {233, 117, 16, 0} };
|
cannam@167
|
874
|
cannam@167
|
875 void XSIMD(codelet_hc2cfdftv_32) (planner *p) {
|
cannam@167
|
876 X(khc2c_register) (p, hc2cfdftv_32, &desc, HC2C_VIA_DFT);
|
cannam@167
|
877 }
|
cannam@167
|
878 #endif
|