Mercurial > hg > sv-dependency-builds
comparison src/fftw-3.3.3/rdft/simd/common/hc2cfdftv_32.c @ 10:37bf6b4a2645
Add FFTW3
author | Chris Cannam |
---|---|
date | Wed, 20 Mar 2013 15:35:50 +0000 |
parents | |
children |
comparison
equal
deleted
inserted
replaced
9:c0fb53affa76 | 10:37bf6b4a2645 |
---|---|
1 /* | |
2 * Copyright (c) 2003, 2007-11 Matteo Frigo | |
3 * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology | |
4 * | |
5 * This program is free software; you can redistribute it and/or modify | |
6 * it under the terms of the GNU General Public License as published by | |
7 * the Free Software Foundation; either version 2 of the License, or | |
8 * (at your option) any later version. | |
9 * | |
10 * This program is distributed in the hope that it will be useful, | |
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
13 * GNU General Public License for more details. | |
14 * | |
15 * You should have received a copy of the GNU General Public License | |
16 * along with this program; if not, write to the Free Software | |
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
18 * | |
19 */ | |
20 | |
21 /* This file was automatically generated --- DO NOT EDIT */ | |
22 /* Generated on Sun Nov 25 07:42:29 EST 2012 */ | |
23 | |
24 #include "codelet-rdft.h" | |
25 | |
26 #ifdef HAVE_FMA | |
27 | |
28 /* Generated by: ../../../genfft/gen_hc2cdft_c.native -fma -reorder-insns -schedule-for-pipeline -simd -compact -variables 4 -pipeline-latency 8 -trivial-stores -variables 32 -no-generate-bytw -n 32 -dit -name hc2cfdftv_32 -include hc2cfv.h */ | |
29 | |
30 /* | |
31 * This function contains 249 FP additions, 224 FP multiplications, | |
32 * (or, 119 additions, 94 multiplications, 130 fused multiply/add), | |
33 * 167 stack variables, 8 constants, and 64 memory accesses | |
34 */ | |
35 #include "hc2cfv.h" | |
36 | |
37 static void hc2cfdftv_32(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms) | |
38 { | |
39 DVK(KP831469612, +0.831469612302545237078788377617905756738560812); | |
40 DVK(KP980785280, +0.980785280403230449126182236134239036973933731); | |
41 DVK(KP668178637, +0.668178637919298919997757686523080761552472251); | |
42 DVK(KP198912367, +0.198912367379658006911597622644676228597850501); | |
43 DVK(KP500000000, +0.500000000000000000000000000000000000000000000); | |
44 DVK(KP923879532, +0.923879532511286756128183189396788286822416626); | |
45 DVK(KP414213562, +0.414213562373095048801688724209698078569671875); | |
46 DVK(KP707106781, +0.707106781186547524400844362104849039284835938); | |
47 { | |
48 INT m; | |
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)) { | |
50 V T2m, T2b, T2c, T2d, T2v, T2r, T20, T2i, T2n, T2e, T2o, T2u, T2j, T2f, T2t; | |
51 V T2s, T2x, T2w, T2l, T2k, T2h, T2g; | |
52 { | |
53 V T41, T3B, T40, T3a, T2J, T27, T2y, Ts, T2C, T1X, T2B, T1Q, T3F, T3w, T4l; | |
54 V T49, T1b, T1s, T3c, TB, T1f, T3g, T44, T1l, T3k, T3o, T4b, T28, T14, T1d; | |
55 V T3b, TK; | |
56 { | |
57 V T1V, T1E, T3A, Th, T3v, T47, T1J, T3q, T8, T38, T25, T39, T3z, Tq, T1O; | |
58 V T3r, T3, T7, T3u, T24, T22, T3t, T1I, Tn, T1G, To, Tm, T1K, Tl, T1N; | |
59 V Tp, T1L, TU, T3f, T3m, T13, T3e, T3n, T1i, TH, TI, T1k, TG, TF, T1c; | |
60 V TJ; | |
61 { | |
62 V T1x, T1y, T1U, T1B, T1S, T1C, T1A, T23, T21, T1z, T1, T2, T1T, T5, T6; | |
63 V T1R, T4, T1w, Ta, Tb, T1H, Te, Tf, Td, Tc, T1F, T9, T1D, Tj, Tk; | |
64 V Ti, Tg, T1M; | |
65 T1 = LD(&(Rp[0]), ms, &(Rp[0])); | |
66 T2 = LD(&(Rm[0]), -ms, &(Rm[0])); | |
67 T1T = LDW(&(W[0])); | |
68 T5 = LD(&(Rp[WS(rs, 8)]), ms, &(Rp[0])); | |
69 T6 = LD(&(Rm[WS(rs, 8)]), -ms, &(Rm[0])); | |
70 T1R = LDW(&(W[TWVL * 32])); | |
71 T4 = LDW(&(W[TWVL * 30])); | |
72 T1x = LD(&(Rp[WS(rs, 12)]), ms, &(Rp[0])); | |
73 T1y = LD(&(Rm[WS(rs, 12)]), -ms, &(Rm[0])); | |
74 T3 = VFMACONJ(T2, T1); | |
75 T1U = VZMULIJ(T1T, VFNMSCONJ(T2, T1)); | |
76 T1w = LDW(&(W[TWVL * 48])); | |
77 T1B = LD(&(Rp[WS(rs, 4)]), ms, &(Rp[0])); | |
78 T1S = VZMULIJ(T1R, VFNMSCONJ(T6, T5)); | |
79 T7 = VZMULJ(T4, VFMACONJ(T6, T5)); | |
80 T1C = LD(&(Rm[WS(rs, 4)]), -ms, &(Rm[0])); | |
81 T1A = LDW(&(W[TWVL * 16])); | |
82 T23 = LDW(&(W[TWVL * 46])); | |
83 T21 = LDW(&(W[TWVL * 14])); | |
84 T1z = VZMULIJ(T1w, VFNMSCONJ(T1y, T1x)); | |
85 Ta = LD(&(Rp[WS(rs, 2)]), ms, &(Rp[0])); | |
86 T3u = VADD(T1U, T1S); | |
87 T1V = VSUB(T1S, T1U); | |
88 Tb = LD(&(Rm[WS(rs, 2)]), -ms, &(Rm[0])); | |
89 T9 = LDW(&(W[TWVL * 6])); | |
90 T1D = VZMULIJ(T1A, VFNMSCONJ(T1C, T1B)); | |
91 T24 = VZMULJ(T23, VFMACONJ(T1y, T1x)); | |
92 T22 = VZMULJ(T21, VFMACONJ(T1C, T1B)); | |
93 T1H = LDW(&(W[TWVL * 8])); | |
94 Te = LD(&(Rp[WS(rs, 10)]), ms, &(Rp[0])); | |
95 Tf = LD(&(Rm[WS(rs, 10)]), -ms, &(Rm[0])); | |
96 Td = LDW(&(W[TWVL * 38])); | |
97 Tc = VZMULJ(T9, VFMACONJ(Tb, Ta)); | |
98 T1E = VSUB(T1z, T1D); | |
99 T3t = VADD(T1D, T1z); | |
100 T1F = LDW(&(W[TWVL * 40])); | |
101 Tj = LD(&(Rp[WS(rs, 14)]), ms, &(Rp[0])); | |
102 T1I = VZMULIJ(T1H, VFNMSCONJ(Tb, Ta)); | |
103 Tk = LD(&(Rm[WS(rs, 14)]), -ms, &(Rm[0])); | |
104 Ti = LDW(&(W[TWVL * 54])); | |
105 Tg = VZMULJ(Td, VFMACONJ(Tf, Te)); | |
106 T1M = LDW(&(W[TWVL * 56])); | |
107 Tn = LD(&(Rp[WS(rs, 6)]), ms, &(Rp[0])); | |
108 T1G = VZMULIJ(T1F, VFNMSCONJ(Tf, Te)); | |
109 To = LD(&(Rm[WS(rs, 6)]), -ms, &(Rm[0])); | |
110 Tm = LDW(&(W[TWVL * 22])); | |
111 T1K = LDW(&(W[TWVL * 24])); | |
112 Tl = VZMULJ(Ti, VFMACONJ(Tk, Tj)); | |
113 T3A = VADD(Tc, Tg); | |
114 Th = VSUB(Tc, Tg); | |
115 T1N = VZMULIJ(T1M, VFNMSCONJ(Tk, Tj)); | |
116 } | |
117 T3v = VSUB(T3t, T3u); | |
118 T47 = VADD(T3u, T3t); | |
119 T1J = VSUB(T1G, T1I); | |
120 T3q = VADD(T1I, T1G); | |
121 Tp = VZMULJ(Tm, VFMACONJ(To, Tn)); | |
122 T1L = VZMULIJ(T1K, VFNMSCONJ(To, Tn)); | |
123 T8 = VSUB(T3, T7); | |
124 T38 = VADD(T3, T7); | |
125 T25 = VSUB(T22, T24); | |
126 T39 = VADD(T22, T24); | |
127 T3z = VADD(Tl, Tp); | |
128 Tq = VSUB(Tl, Tp); | |
129 T1O = VSUB(T1L, T1N); | |
130 T3r = VADD(T1N, T1L); | |
131 { | |
132 V T10, T11, TZ, T1o, TY, T1r, TN, TO, TM, T19, TR, TS, TQ, T17, T26; | |
133 V Tr, T1W, T1P, T3s, T48, TW, TX, TP, T1a, TV, T1q, TT, T18, Ty, Tz; | |
134 V Tx, Tw, T1j, Tu, T12, T1p, Tv, Tt, T1h, TD, TA, TE, TC, T1e; | |
135 TN = LD(&(Rp[WS(rs, 3)]), ms, &(Rp[WS(rs, 1)])); | |
136 TO = LD(&(Rm[WS(rs, 3)]), -ms, &(Rm[WS(rs, 1)])); | |
137 T41 = VADD(T3A, T3z); | |
138 T3B = VSUB(T3z, T3A); | |
139 T26 = VSUB(Tq, Th); | |
140 Tr = VADD(Th, Tq); | |
141 T1W = VADD(T1J, T1O); | |
142 T1P = VSUB(T1J, T1O); | |
143 T3s = VSUB(T3q, T3r); | |
144 T48 = VADD(T3q, T3r); | |
145 T40 = VADD(T38, T39); | |
146 T3a = VSUB(T38, T39); | |
147 T2J = VFNMS(LDK(KP707106781), T26, T25); | |
148 T27 = VFMA(LDK(KP707106781), T26, T25); | |
149 T2y = VFMA(LDK(KP707106781), Tr, T8); | |
150 Ts = VFNMS(LDK(KP707106781), Tr, T8); | |
151 T2C = VFMA(LDK(KP707106781), T1W, T1V); | |
152 T1X = VFNMS(LDK(KP707106781), T1W, T1V); | |
153 T2B = VFMA(LDK(KP707106781), T1P, T1E); | |
154 T1Q = VFNMS(LDK(KP707106781), T1P, T1E); | |
155 T3F = VFMA(LDK(KP414213562), T3s, T3v); | |
156 T3w = VFNMS(LDK(KP414213562), T3v, T3s); | |
157 T4l = VSUB(T48, T47); | |
158 T49 = VADD(T47, T48); | |
159 TM = LDW(&(W[TWVL * 10])); | |
160 T19 = LDW(&(W[TWVL * 12])); | |
161 TR = LD(&(Rp[WS(rs, 11)]), ms, &(Rp[WS(rs, 1)])); | |
162 TS = LD(&(Rm[WS(rs, 11)]), -ms, &(Rm[WS(rs, 1)])); | |
163 TQ = LDW(&(W[TWVL * 42])); | |
164 T17 = LDW(&(W[TWVL * 44])); | |
165 TW = LD(&(Rp[WS(rs, 15)]), ms, &(Rp[WS(rs, 1)])); | |
166 TX = LD(&(Rm[WS(rs, 15)]), -ms, &(Rm[WS(rs, 1)])); | |
167 TP = VZMULJ(TM, VFMACONJ(TO, TN)); | |
168 T1a = VZMULIJ(T19, VFNMSCONJ(TO, TN)); | |
169 TV = LDW(&(W[TWVL * 58])); | |
170 T1q = LDW(&(W[TWVL * 60])); | |
171 TT = VZMULJ(TQ, VFMACONJ(TS, TR)); | |
172 T18 = VZMULIJ(T17, VFNMSCONJ(TS, TR)); | |
173 T10 = LD(&(Rp[WS(rs, 7)]), ms, &(Rp[WS(rs, 1)])); | |
174 T11 = LD(&(Rm[WS(rs, 7)]), -ms, &(Rm[WS(rs, 1)])); | |
175 TZ = LDW(&(W[TWVL * 26])); | |
176 T1o = LDW(&(W[TWVL * 28])); | |
177 TY = VZMULJ(TV, VFMACONJ(TX, TW)); | |
178 T1r = VZMULIJ(T1q, VFNMSCONJ(TX, TW)); | |
179 TU = VSUB(TP, TT); | |
180 T3f = VADD(TP, TT); | |
181 T1b = VSUB(T18, T1a); | |
182 T3m = VADD(T1a, T18); | |
183 Tu = LD(&(Rp[WS(rs, 5)]), ms, &(Rp[WS(rs, 1)])); | |
184 T12 = VZMULJ(TZ, VFMACONJ(T11, T10)); | |
185 T1p = VZMULIJ(T1o, VFNMSCONJ(T11, T10)); | |
186 Tv = LD(&(Rm[WS(rs, 5)]), -ms, &(Rm[WS(rs, 1)])); | |
187 Tt = LDW(&(W[TWVL * 18])); | |
188 T1h = LDW(&(W[TWVL * 20])); | |
189 Ty = LD(&(Rp[WS(rs, 13)]), ms, &(Rp[WS(rs, 1)])); | |
190 Tz = LD(&(Rm[WS(rs, 13)]), -ms, &(Rm[WS(rs, 1)])); | |
191 Tx = LDW(&(W[TWVL * 50])); | |
192 T13 = VSUB(TY, T12); | |
193 T3e = VADD(TY, T12); | |
194 T1s = VSUB(T1p, T1r); | |
195 T3n = VADD(T1r, T1p); | |
196 Tw = VZMULJ(Tt, VFMACONJ(Tv, Tu)); | |
197 T1i = VZMULIJ(T1h, VFNMSCONJ(Tv, Tu)); | |
198 T1j = LDW(&(W[TWVL * 52])); | |
199 TD = LD(&(Rp[WS(rs, 1)]), ms, &(Rp[WS(rs, 1)])); | |
200 TA = VZMULJ(Tx, VFMACONJ(Tz, Ty)); | |
201 TE = LD(&(Rm[WS(rs, 1)]), -ms, &(Rm[WS(rs, 1)])); | |
202 TC = LDW(&(W[TWVL * 2])); | |
203 T1e = LDW(&(W[TWVL * 4])); | |
204 TH = LD(&(Rp[WS(rs, 9)]), ms, &(Rp[WS(rs, 1)])); | |
205 TI = LD(&(Rm[WS(rs, 9)]), -ms, &(Rm[WS(rs, 1)])); | |
206 T1k = VZMULIJ(T1j, VFNMSCONJ(Tz, Ty)); | |
207 TG = LDW(&(W[TWVL * 34])); | |
208 T3c = VADD(Tw, TA); | |
209 TB = VSUB(Tw, TA); | |
210 TF = VZMULJ(TC, VFMACONJ(TE, TD)); | |
211 T1f = VZMULIJ(T1e, VFNMSCONJ(TE, TD)); | |
212 T1c = LDW(&(W[TWVL * 36])); | |
213 } | |
214 T3g = VSUB(T3e, T3f); | |
215 T44 = VADD(T3e, T3f); | |
216 T1l = VSUB(T1i, T1k); | |
217 T3k = VADD(T1i, T1k); | |
218 TJ = VZMULJ(TG, VFMACONJ(TI, TH)); | |
219 T3o = VSUB(T3m, T3n); | |
220 T4b = VADD(T3n, T3m); | |
221 T28 = VFMA(LDK(KP414213562), TU, T13); | |
222 T14 = VFNMS(LDK(KP414213562), T13, TU); | |
223 T1d = VZMULIJ(T1c, VFNMSCONJ(TI, TH)); | |
224 T3b = VADD(TF, TJ); | |
225 TK = VSUB(TF, TJ); | |
226 } | |
227 { | |
228 V T4k, T4p, T2z, T2a, T2K, T15, T2E, T1n, T2F, T1u, T4c, T3R, T3D, T3i, T3O; | |
229 V T46, T4g, T3G, T3P, T3S, T3x, T4q, T4n, T42, T1g, T3j, T3E, T3p, T4m, T3d; | |
230 V T43, T29, TL, T1m, T1t, T3l, T4a, T3C, T3h, T45, T3Q, T3W, T4d, T4h, T3H; | |
231 V T3L, T3y, T3K, T4r, T4v, T4o, T4u, T4j, T4i, T4e, T4f, T3N, T3M, T3I, T3J; | |
232 V T4x, T4w, T4s, T4t; | |
233 T42 = VADD(T40, T41); | |
234 T4k = VSUB(T40, T41); | |
235 T1g = VSUB(T1d, T1f); | |
236 T3j = VADD(T1f, T1d); | |
237 T3d = VSUB(T3b, T3c); | |
238 T43 = VADD(T3b, T3c); | |
239 T29 = VFNMS(LDK(KP414213562), TB, TK); | |
240 TL = VFMA(LDK(KP414213562), TK, TB); | |
241 T1m = VSUB(T1g, T1l); | |
242 T1t = VADD(T1g, T1l); | |
243 T3l = VSUB(T3j, T3k); | |
244 T4a = VADD(T3j, T3k); | |
245 T3C = VSUB(T3g, T3d); | |
246 T3h = VADD(T3d, T3g); | |
247 T45 = VADD(T43, T44); | |
248 T4p = VSUB(T44, T43); | |
249 T2z = VADD(T29, T28); | |
250 T2a = VSUB(T28, T29); | |
251 T2K = VADD(TL, T14); | |
252 T15 = VSUB(TL, T14); | |
253 T2E = VFMA(LDK(KP707106781), T1m, T1b); | |
254 T1n = VFNMS(LDK(KP707106781), T1m, T1b); | |
255 T2F = VFMA(LDK(KP707106781), T1t, T1s); | |
256 T1u = VFNMS(LDK(KP707106781), T1t, T1s); | |
257 T3E = VFNMS(LDK(KP414213562), T3l, T3o); | |
258 T3p = VFMA(LDK(KP414213562), T3o, T3l); | |
259 T4m = VSUB(T4a, T4b); | |
260 T4c = VADD(T4a, T4b); | |
261 T3R = VFMA(LDK(KP707106781), T3C, T3B); | |
262 T3D = VFNMS(LDK(KP707106781), T3C, T3B); | |
263 T3i = VFNMS(LDK(KP707106781), T3h, T3a); | |
264 T3O = VFMA(LDK(KP707106781), T3h, T3a); | |
265 T46 = VSUB(T42, T45); | |
266 T4g = VADD(T42, T45); | |
267 T3G = VSUB(T3E, T3F); | |
268 T3P = VADD(T3F, T3E); | |
269 T3S = VADD(T3w, T3p); | |
270 T3x = VSUB(T3p, T3w); | |
271 T4q = VSUB(T4m, T4l); | |
272 T4n = VADD(T4l, T4m); | |
273 T4d = VSUB(T49, T4c); | |
274 T4h = VADD(T49, T4c); | |
275 T3H = VFNMS(LDK(KP923879532), T3G, T3D); | |
276 T3L = VFMA(LDK(KP923879532), T3G, T3D); | |
277 T3y = VFMA(LDK(KP923879532), T3x, T3i); | |
278 T3K = VFNMS(LDK(KP923879532), T3x, T3i); | |
279 T4r = VFMA(LDK(KP707106781), T4q, T4p); | |
280 T4v = VFNMS(LDK(KP707106781), T4q, T4p); | |
281 T4o = VFMA(LDK(KP707106781), T4n, T4k); | |
282 T4u = VFNMS(LDK(KP707106781), T4n, T4k); | |
283 T3Q = VFMA(LDK(KP923879532), T3P, T3O); | |
284 T3W = VFNMS(LDK(KP923879532), T3P, T3O); | |
285 T4j = VCONJ(VMUL(LDK(KP500000000), VADD(T4h, T4g))); | |
286 T4i = VMUL(LDK(KP500000000), VSUB(T4g, T4h)); | |
287 T4e = VMUL(LDK(KP500000000), VFMAI(T4d, T46)); | |
288 T4f = VCONJ(VMUL(LDK(KP500000000), VFNMSI(T4d, T46))); | |
289 T3N = VMUL(LDK(KP500000000), VFMAI(T3L, T3K)); | |
290 T3M = VCONJ(VMUL(LDK(KP500000000), VFNMSI(T3L, T3K))); | |
291 T3I = VMUL(LDK(KP500000000), VFNMSI(T3H, T3y)); | |
292 T3J = VCONJ(VMUL(LDK(KP500000000), VFMAI(T3H, T3y))); | |
293 T4x = VCONJ(VMUL(LDK(KP500000000), VFMAI(T4v, T4u))); | |
294 T4w = VMUL(LDK(KP500000000), VFNMSI(T4v, T4u)); | |
295 T4s = VCONJ(VMUL(LDK(KP500000000), VFNMSI(T4r, T4o))); | |
296 T4t = VMUL(LDK(KP500000000), VFMAI(T4r, T4o)); | |
297 ST(&(Rp[0]), T4i, ms, &(Rp[0])); | |
298 ST(&(Rm[WS(rs, 15)]), T4j, -ms, &(Rm[WS(rs, 1)])); | |
299 ST(&(Rm[WS(rs, 7)]), T4f, -ms, &(Rm[WS(rs, 1)])); | |
300 ST(&(Rp[WS(rs, 8)]), T4e, ms, &(Rp[0])); | |
301 ST(&(Rm[WS(rs, 9)]), T3M, -ms, &(Rm[WS(rs, 1)])); | |
302 ST(&(Rp[WS(rs, 10)]), T3N, ms, &(Rp[0])); | |
303 ST(&(Rm[WS(rs, 5)]), T3J, -ms, &(Rm[WS(rs, 1)])); | |
304 ST(&(Rp[WS(rs, 6)]), T3I, ms, &(Rp[0])); | |
305 ST(&(Rp[WS(rs, 12)]), T4w, ms, &(Rp[0])); | |
306 ST(&(Rm[WS(rs, 11)]), T4x, -ms, &(Rm[WS(rs, 1)])); | |
307 ST(&(Rp[WS(rs, 4)]), T4t, ms, &(Rp[0])); | |
308 ST(&(Rm[WS(rs, 3)]), T4s, -ms, &(Rm[WS(rs, 1)])); | |
309 { | |
310 V T2A, T2W, T2L, T2Z, T2D, T2N, T2M, T2G, T3T, T3X, T16, T2p, T1v, T35, T31; | |
311 V T2I, T2S, T34, T2Y, T2P, T2T, T1Y, T2H, T30, T3Z, T3Y, T3U, T3V, T2O, T2X; | |
312 V T32, T33, T36, T37, T2U, T2V, T2Q, T2R, T1Z, T2q; | |
313 T2A = VFNMS(LDK(KP923879532), T2z, T2y); | |
314 T2W = VFMA(LDK(KP923879532), T2z, T2y); | |
315 T2L = VFNMS(LDK(KP923879532), T2K, T2J); | |
316 T2Z = VFMA(LDK(KP923879532), T2K, T2J); | |
317 T2D = VFMA(LDK(KP198912367), T2C, T2B); | |
318 T2N = VFNMS(LDK(KP198912367), T2B, T2C); | |
319 T2M = VFMA(LDK(KP198912367), T2E, T2F); | |
320 T2G = VFNMS(LDK(KP198912367), T2F, T2E); | |
321 T3T = VFMA(LDK(KP923879532), T3S, T3R); | |
322 T3X = VFNMS(LDK(KP923879532), T3S, T3R); | |
323 T16 = VFNMS(LDK(KP923879532), T15, Ts); | |
324 T2m = VFMA(LDK(KP923879532), T15, Ts); | |
325 T2H = VSUB(T2D, T2G); | |
326 T30 = VADD(T2D, T2G); | |
327 T2b = VFNMS(LDK(KP923879532), T2a, T27); | |
328 T2p = VFMA(LDK(KP923879532), T2a, T27); | |
329 T1v = VFMA(LDK(KP668178637), T1u, T1n); | |
330 T2c = VFNMS(LDK(KP668178637), T1n, T1u); | |
331 T3Z = VCONJ(VMUL(LDK(KP500000000), VFMAI(T3X, T3W))); | |
332 T3Y = VMUL(LDK(KP500000000), VFNMSI(T3X, T3W)); | |
333 T3U = VCONJ(VMUL(LDK(KP500000000), VFNMSI(T3T, T3Q))); | |
334 T3V = VMUL(LDK(KP500000000), VFMAI(T3T, T3Q)); | |
335 T2O = VSUB(T2M, T2N); | |
336 T2X = VADD(T2N, T2M); | |
337 T35 = VFNMS(LDK(KP980785280), T30, T2Z); | |
338 T31 = VFMA(LDK(KP980785280), T30, T2Z); | |
339 T2I = VFMA(LDK(KP980785280), T2H, T2A); | |
340 T2S = VFNMS(LDK(KP980785280), T2H, T2A); | |
341 ST(&(Rp[WS(rs, 14)]), T3Y, ms, &(Rp[0])); | |
342 ST(&(Rm[WS(rs, 13)]), T3Z, -ms, &(Rm[WS(rs, 1)])); | |
343 ST(&(Rp[WS(rs, 2)]), T3V, ms, &(Rp[0])); | |
344 ST(&(Rm[WS(rs, 1)]), T3U, -ms, &(Rm[WS(rs, 1)])); | |
345 T34 = VFNMS(LDK(KP980785280), T2X, T2W); | |
346 T2Y = VFMA(LDK(KP980785280), T2X, T2W); | |
347 T2P = VFMA(LDK(KP980785280), T2O, T2L); | |
348 T2T = VFNMS(LDK(KP980785280), T2O, T2L); | |
349 T2d = VFMA(LDK(KP668178637), T1Q, T1X); | |
350 T1Y = VFNMS(LDK(KP668178637), T1X, T1Q); | |
351 T32 = VMUL(LDK(KP500000000), VFNMSI(T31, T2Y)); | |
352 T33 = VCONJ(VMUL(LDK(KP500000000), VFMAI(T31, T2Y))); | |
353 T36 = VCONJ(VMUL(LDK(KP500000000), VFNMSI(T35, T34))); | |
354 T37 = VMUL(LDK(KP500000000), VFMAI(T35, T34)); | |
355 T2U = VMUL(LDK(KP500000000), VFNMSI(T2T, T2S)); | |
356 T2V = VCONJ(VMUL(LDK(KP500000000), VFMAI(T2T, T2S))); | |
357 T2Q = VCONJ(VMUL(LDK(KP500000000), VFNMSI(T2P, T2I))); | |
358 T2R = VMUL(LDK(KP500000000), VFMAI(T2P, T2I)); | |
359 T1Z = VSUB(T1v, T1Y); | |
360 T2q = VADD(T1Y, T1v); | |
361 ST(&(Rm[0]), T33, -ms, &(Rm[0])); | |
362 ST(&(Rp[WS(rs, 1)]), T32, ms, &(Rp[WS(rs, 1)])); | |
363 ST(&(Rp[WS(rs, 15)]), T37, ms, &(Rp[WS(rs, 1)])); | |
364 ST(&(Rm[WS(rs, 14)]), T36, -ms, &(Rm[0])); | |
365 ST(&(Rm[WS(rs, 8)]), T2V, -ms, &(Rm[0])); | |
366 ST(&(Rp[WS(rs, 9)]), T2U, ms, &(Rp[WS(rs, 1)])); | |
367 ST(&(Rp[WS(rs, 7)]), T2R, ms, &(Rp[WS(rs, 1)])); | |
368 ST(&(Rm[WS(rs, 6)]), T2Q, -ms, &(Rm[0])); | |
369 T2v = VFNMS(LDK(KP831469612), T2q, T2p); | |
370 T2r = VFMA(LDK(KP831469612), T2q, T2p); | |
371 T20 = VFMA(LDK(KP831469612), T1Z, T16); | |
372 T2i = VFNMS(LDK(KP831469612), T1Z, T16); | |
373 } | |
374 } | |
375 } | |
376 T2n = VADD(T2d, T2c); | |
377 T2e = VSUB(T2c, T2d); | |
378 T2o = VFMA(LDK(KP831469612), T2n, T2m); | |
379 T2u = VFNMS(LDK(KP831469612), T2n, T2m); | |
380 T2j = VFMA(LDK(KP831469612), T2e, T2b); | |
381 T2f = VFNMS(LDK(KP831469612), T2e, T2b); | |
382 T2t = VCONJ(VMUL(LDK(KP500000000), VFNMSI(T2r, T2o))); | |
383 T2s = VMUL(LDK(KP500000000), VFMAI(T2r, T2o)); | |
384 T2x = VCONJ(VMUL(LDK(KP500000000), VFMAI(T2v, T2u))); | |
385 T2w = VMUL(LDK(KP500000000), VFNMSI(T2v, T2u)); | |
386 T2l = VCONJ(VMUL(LDK(KP500000000), VFNMSI(T2j, T2i))); | |
387 T2k = VMUL(LDK(KP500000000), VFMAI(T2j, T2i)); | |
388 T2h = VCONJ(VMUL(LDK(KP500000000), VFMAI(T2f, T20))); | |
389 T2g = VMUL(LDK(KP500000000), VFNMSI(T2f, T20)); | |
390 ST(&(Rm[WS(rs, 2)]), T2t, -ms, &(Rm[0])); | |
391 ST(&(Rp[WS(rs, 3)]), T2s, ms, &(Rp[WS(rs, 1)])); | |
392 ST(&(Rm[WS(rs, 12)]), T2x, -ms, &(Rm[0])); | |
393 ST(&(Rp[WS(rs, 13)]), T2w, ms, &(Rp[WS(rs, 1)])); | |
394 ST(&(Rm[WS(rs, 10)]), T2l, -ms, &(Rm[0])); | |
395 ST(&(Rp[WS(rs, 11)]), T2k, ms, &(Rp[WS(rs, 1)])); | |
396 ST(&(Rm[WS(rs, 4)]), T2h, -ms, &(Rm[0])); | |
397 ST(&(Rp[WS(rs, 5)]), T2g, ms, &(Rp[WS(rs, 1)])); | |
398 } | |
399 } | |
400 VLEAVE(); | |
401 } | |
402 | |
403 static const tw_instr twinstr[] = { | |
404 VTW(1, 1), | |
405 VTW(1, 2), | |
406 VTW(1, 3), | |
407 VTW(1, 4), | |
408 VTW(1, 5), | |
409 VTW(1, 6), | |
410 VTW(1, 7), | |
411 VTW(1, 8), | |
412 VTW(1, 9), | |
413 VTW(1, 10), | |
414 VTW(1, 11), | |
415 VTW(1, 12), | |
416 VTW(1, 13), | |
417 VTW(1, 14), | |
418 VTW(1, 15), | |
419 VTW(1, 16), | |
420 VTW(1, 17), | |
421 VTW(1, 18), | |
422 VTW(1, 19), | |
423 VTW(1, 20), | |
424 VTW(1, 21), | |
425 VTW(1, 22), | |
426 VTW(1, 23), | |
427 VTW(1, 24), | |
428 VTW(1, 25), | |
429 VTW(1, 26), | |
430 VTW(1, 27), | |
431 VTW(1, 28), | |
432 VTW(1, 29), | |
433 VTW(1, 30), | |
434 VTW(1, 31), | |
435 {TW_NEXT, VL, 0} | |
436 }; | |
437 | |
438 static const hc2c_desc desc = { 32, XSIMD_STRING("hc2cfdftv_32"), twinstr, &GENUS, {119, 94, 130, 0} }; | |
439 | |
440 void XSIMD(codelet_hc2cfdftv_32) (planner *p) { | |
441 X(khc2c_register) (p, hc2cfdftv_32, &desc, HC2C_VIA_DFT); | |
442 } | |
443 #else /* HAVE_FMA */ | |
444 | |
445 /* 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 hc2cfv.h */ | |
446 | |
447 /* | |
448 * This function contains 249 FP additions, 133 FP multiplications, | |
449 * (or, 233 additions, 117 multiplications, 16 fused multiply/add), | |
450 * 130 stack variables, 9 constants, and 64 memory accesses | |
451 */ | |
452 #include "hc2cfv.h" | |
453 | |
454 static void hc2cfdftv_32(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms) | |
455 { | |
456 DVK(KP555570233, +0.555570233019602224742830813948532874374937191); | |
457 DVK(KP831469612, +0.831469612302545237078788377617905756738560812); | |
458 DVK(KP195090322, +0.195090322016128267848284868477022240927691618); | |
459 DVK(KP980785280, +0.980785280403230449126182236134239036973933731); | |
460 DVK(KP382683432, +0.382683432365089771728459984030398866761344562); | |
461 DVK(KP923879532, +0.923879532511286756128183189396788286822416626); | |
462 DVK(KP707106781, +0.707106781186547524400844362104849039284835938); | |
463 DVK(KP353553390, +0.353553390593273762200422181052424519642417969); | |
464 DVK(KP500000000, +0.500000000000000000000000000000000000000000000); | |
465 { | |
466 INT m; | |
467 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)) { | |
468 V Ta, T2m, Tx, T2h, T3R, T4h, T3q, T4g, T3B, T4n, T3E, T4o, T1B, T2S, T1O; | |
469 V T2R, TV, T2p, T1i, T2o, T3L, T4q, T3I, T4r, T3w, T4k, T3t, T4j, T26, T2V; | |
470 V T2d, T2U; | |
471 { | |
472 V T4, T1m, T1H, T2j, T1M, T2l, T9, T1o, Tf, T1r, Tq, T1w, Tv, T1y, Tk; | |
473 V T1t, Tl, Tw, T3P, T3Q, T3o, T3p, T3z, T3A, T3C, T3D, T1p, T1N, T1A, T1C; | |
474 V T1u, T1z; | |
475 { | |
476 V T1, T3, T2, T1l, T1G, T1F, T1E, T1D, T2i, T1L, T1K, T1J, T1I, T2k, T6; | |
477 V T8, T7, T5, T1n, Tc, Te, Td, Tb, T1q, Tn, Tp, To, Tm, T1v, Ts; | |
478 V Tu, Tt, Tr, T1x, Th, Tj, Ti, Tg, T1s; | |
479 T1 = LD(&(Rp[0]), ms, &(Rp[0])); | |
480 T2 = LD(&(Rm[0]), -ms, &(Rm[0])); | |
481 T3 = VCONJ(T2); | |
482 T4 = VADD(T1, T3); | |
483 T1l = LDW(&(W[0])); | |
484 T1m = VZMULIJ(T1l, VSUB(T3, T1)); | |
485 T1G = LD(&(Rp[WS(rs, 4)]), ms, &(Rp[0])); | |
486 T1E = LD(&(Rm[WS(rs, 4)]), -ms, &(Rm[0])); | |
487 T1F = VCONJ(T1E); | |
488 T1D = LDW(&(W[TWVL * 16])); | |
489 T1H = VZMULIJ(T1D, VSUB(T1F, T1G)); | |
490 T2i = LDW(&(W[TWVL * 14])); | |
491 T2j = VZMULJ(T2i, VADD(T1G, T1F)); | |
492 T1L = LD(&(Rp[WS(rs, 12)]), ms, &(Rp[0])); | |
493 T1J = LD(&(Rm[WS(rs, 12)]), -ms, &(Rm[0])); | |
494 T1K = VCONJ(T1J); | |
495 T1I = LDW(&(W[TWVL * 48])); | |
496 T1M = VZMULIJ(T1I, VSUB(T1K, T1L)); | |
497 T2k = LDW(&(W[TWVL * 46])); | |
498 T2l = VZMULJ(T2k, VADD(T1L, T1K)); | |
499 T6 = LD(&(Rp[WS(rs, 8)]), ms, &(Rp[0])); | |
500 T7 = LD(&(Rm[WS(rs, 8)]), -ms, &(Rm[0])); | |
501 T8 = VCONJ(T7); | |
502 T5 = LDW(&(W[TWVL * 30])); | |
503 T9 = VZMULJ(T5, VADD(T6, T8)); | |
504 T1n = LDW(&(W[TWVL * 32])); | |
505 T1o = VZMULIJ(T1n, VSUB(T8, T6)); | |
506 Tc = LD(&(Rp[WS(rs, 2)]), ms, &(Rp[0])); | |
507 Td = LD(&(Rm[WS(rs, 2)]), -ms, &(Rm[0])); | |
508 Te = VCONJ(Td); | |
509 Tb = LDW(&(W[TWVL * 6])); | |
510 Tf = VZMULJ(Tb, VADD(Tc, Te)); | |
511 T1q = LDW(&(W[TWVL * 8])); | |
512 T1r = VZMULIJ(T1q, VSUB(Te, Tc)); | |
513 Tn = LD(&(Rp[WS(rs, 14)]), ms, &(Rp[0])); | |
514 To = LD(&(Rm[WS(rs, 14)]), -ms, &(Rm[0])); | |
515 Tp = VCONJ(To); | |
516 Tm = LDW(&(W[TWVL * 54])); | |
517 Tq = VZMULJ(Tm, VADD(Tn, Tp)); | |
518 T1v = LDW(&(W[TWVL * 56])); | |
519 T1w = VZMULIJ(T1v, VSUB(Tp, Tn)); | |
520 Ts = LD(&(Rp[WS(rs, 6)]), ms, &(Rp[0])); | |
521 Tt = LD(&(Rm[WS(rs, 6)]), -ms, &(Rm[0])); | |
522 Tu = VCONJ(Tt); | |
523 Tr = LDW(&(W[TWVL * 22])); | |
524 Tv = VZMULJ(Tr, VADD(Ts, Tu)); | |
525 T1x = LDW(&(W[TWVL * 24])); | |
526 T1y = VZMULIJ(T1x, VSUB(Tu, Ts)); | |
527 Th = LD(&(Rp[WS(rs, 10)]), ms, &(Rp[0])); | |
528 Ti = LD(&(Rm[WS(rs, 10)]), -ms, &(Rm[0])); | |
529 Tj = VCONJ(Ti); | |
530 Tg = LDW(&(W[TWVL * 38])); | |
531 Tk = VZMULJ(Tg, VADD(Th, Tj)); | |
532 T1s = LDW(&(W[TWVL * 40])); | |
533 T1t = VZMULIJ(T1s, VSUB(Tj, Th)); | |
534 } | |
535 Ta = VMUL(LDK(KP500000000), VSUB(T4, T9)); | |
536 T2m = VSUB(T2j, T2l); | |
537 Tl = VSUB(Tf, Tk); | |
538 Tw = VSUB(Tq, Tv); | |
539 Tx = VMUL(LDK(KP353553390), VADD(Tl, Tw)); | |
540 T2h = VMUL(LDK(KP707106781), VSUB(Tw, Tl)); | |
541 T3P = VADD(Tq, Tv); | |
542 T3Q = VADD(Tf, Tk); | |
543 T3R = VSUB(T3P, T3Q); | |
544 T4h = VADD(T3Q, T3P); | |
545 T3o = VADD(T4, T9); | |
546 T3p = VADD(T2j, T2l); | |
547 T3q = VMUL(LDK(KP500000000), VSUB(T3o, T3p)); | |
548 T4g = VADD(T3o, T3p); | |
549 T3z = VADD(T1m, T1o); | |
550 T3A = VADD(T1H, T1M); | |
551 T3B = VSUB(T3z, T3A); | |
552 T4n = VADD(T3z, T3A); | |
553 T3C = VADD(T1w, T1y); | |
554 T3D = VADD(T1r, T1t); | |
555 T3E = VSUB(T3C, T3D); | |
556 T4o = VADD(T3D, T3C); | |
557 T1p = VSUB(T1m, T1o); | |
558 T1N = VSUB(T1H, T1M); | |
559 T1u = VSUB(T1r, T1t); | |
560 T1z = VSUB(T1w, T1y); | |
561 T1A = VMUL(LDK(KP707106781), VADD(T1u, T1z)); | |
562 T1C = VMUL(LDK(KP707106781), VSUB(T1z, T1u)); | |
563 T1B = VADD(T1p, T1A); | |
564 T2S = VADD(T1N, T1C); | |
565 T1O = VSUB(T1C, T1N); | |
566 T2R = VSUB(T1p, T1A); | |
567 } | |
568 { | |
569 V TD, T1R, T1b, T29, T1g, T2b, TI, T1T, TO, T1Y, T10, T22, T15, T24, TT; | |
570 V T1W, TJ, TU, T16, T1h, T3J, T3K, T3G, T3H, T3u, T3v, T3r, T3s, T25, T2c; | |
571 V T20, T27, T1U, T1Z; | |
572 { | |
573 V TA, TC, TB, Tz, T1Q, T18, T1a, T19, T17, T28, T1d, T1f, T1e, T1c, T2a; | |
574 V TF, TH, TG, TE, T1S, TL, TN, TM, TK, T1X, TX, TZ, TY, TW, T21; | |
575 V T12, T14, T13, T11, T23, TQ, TS, TR, TP, T1V; | |
576 TA = LD(&(Rp[WS(rs, 1)]), ms, &(Rp[WS(rs, 1)])); | |
577 TB = LD(&(Rm[WS(rs, 1)]), -ms, &(Rm[WS(rs, 1)])); | |
578 TC = VCONJ(TB); | |
579 Tz = LDW(&(W[TWVL * 2])); | |
580 TD = VZMULJ(Tz, VADD(TA, TC)); | |
581 T1Q = LDW(&(W[TWVL * 4])); | |
582 T1R = VZMULIJ(T1Q, VSUB(TC, TA)); | |
583 T18 = LD(&(Rp[WS(rs, 3)]), ms, &(Rp[WS(rs, 1)])); | |
584 T19 = LD(&(Rm[WS(rs, 3)]), -ms, &(Rm[WS(rs, 1)])); | |
585 T1a = VCONJ(T19); | |
586 T17 = LDW(&(W[TWVL * 10])); | |
587 T1b = VZMULJ(T17, VADD(T18, T1a)); | |
588 T28 = LDW(&(W[TWVL * 12])); | |
589 T29 = VZMULIJ(T28, VSUB(T1a, T18)); | |
590 T1d = LD(&(Rp[WS(rs, 11)]), ms, &(Rp[WS(rs, 1)])); | |
591 T1e = LD(&(Rm[WS(rs, 11)]), -ms, &(Rm[WS(rs, 1)])); | |
592 T1f = VCONJ(T1e); | |
593 T1c = LDW(&(W[TWVL * 42])); | |
594 T1g = VZMULJ(T1c, VADD(T1d, T1f)); | |
595 T2a = LDW(&(W[TWVL * 44])); | |
596 T2b = VZMULIJ(T2a, VSUB(T1f, T1d)); | |
597 TF = LD(&(Rp[WS(rs, 9)]), ms, &(Rp[WS(rs, 1)])); | |
598 TG = LD(&(Rm[WS(rs, 9)]), -ms, &(Rm[WS(rs, 1)])); | |
599 TH = VCONJ(TG); | |
600 TE = LDW(&(W[TWVL * 34])); | |
601 TI = VZMULJ(TE, VADD(TF, TH)); | |
602 T1S = LDW(&(W[TWVL * 36])); | |
603 T1T = VZMULIJ(T1S, VSUB(TH, TF)); | |
604 TL = LD(&(Rp[WS(rs, 5)]), ms, &(Rp[WS(rs, 1)])); | |
605 TM = LD(&(Rm[WS(rs, 5)]), -ms, &(Rm[WS(rs, 1)])); | |
606 TN = VCONJ(TM); | |
607 TK = LDW(&(W[TWVL * 18])); | |
608 TO = VZMULJ(TK, VADD(TL, TN)); | |
609 T1X = LDW(&(W[TWVL * 20])); | |
610 T1Y = VZMULIJ(T1X, VSUB(TN, TL)); | |
611 TX = LD(&(Rp[WS(rs, 15)]), ms, &(Rp[WS(rs, 1)])); | |
612 TY = LD(&(Rm[WS(rs, 15)]), -ms, &(Rm[WS(rs, 1)])); | |
613 TZ = VCONJ(TY); | |
614 TW = LDW(&(W[TWVL * 58])); | |
615 T10 = VZMULJ(TW, VADD(TX, TZ)); | |
616 T21 = LDW(&(W[TWVL * 60])); | |
617 T22 = VZMULIJ(T21, VSUB(TZ, TX)); | |
618 T12 = LD(&(Rp[WS(rs, 7)]), ms, &(Rp[WS(rs, 1)])); | |
619 T13 = LD(&(Rm[WS(rs, 7)]), -ms, &(Rm[WS(rs, 1)])); | |
620 T14 = VCONJ(T13); | |
621 T11 = LDW(&(W[TWVL * 26])); | |
622 T15 = VZMULJ(T11, VADD(T12, T14)); | |
623 T23 = LDW(&(W[TWVL * 28])); | |
624 T24 = VZMULIJ(T23, VSUB(T14, T12)); | |
625 TQ = LD(&(Rp[WS(rs, 13)]), ms, &(Rp[WS(rs, 1)])); | |
626 TR = LD(&(Rm[WS(rs, 13)]), -ms, &(Rm[WS(rs, 1)])); | |
627 TS = VCONJ(TR); | |
628 TP = LDW(&(W[TWVL * 50])); | |
629 TT = VZMULJ(TP, VADD(TQ, TS)); | |
630 T1V = LDW(&(W[TWVL * 52])); | |
631 T1W = VZMULIJ(T1V, VSUB(TS, TQ)); | |
632 } | |
633 TJ = VSUB(TD, TI); | |
634 TU = VSUB(TO, TT); | |
635 TV = VFNMS(LDK(KP382683432), TU, VMUL(LDK(KP923879532), TJ)); | |
636 T2p = VFMA(LDK(KP382683432), TJ, VMUL(LDK(KP923879532), TU)); | |
637 T16 = VSUB(T10, T15); | |
638 T1h = VSUB(T1b, T1g); | |
639 T1i = VFMA(LDK(KP923879532), T16, VMUL(LDK(KP382683432), T1h)); | |
640 T2o = VFNMS(LDK(KP923879532), T1h, VMUL(LDK(KP382683432), T16)); | |
641 T3J = VADD(T1Y, T1W); | |
642 T3K = VADD(T1R, T1T); | |
643 T3L = VSUB(T3J, T3K); | |
644 T4q = VADD(T3K, T3J); | |
645 T3G = VADD(T22, T24); | |
646 T3H = VADD(T29, T2b); | |
647 T3I = VSUB(T3G, T3H); | |
648 T4r = VADD(T3G, T3H); | |
649 T3u = VADD(T10, T15); | |
650 T3v = VADD(T1b, T1g); | |
651 T3w = VSUB(T3u, T3v); | |
652 T4k = VADD(T3u, T3v); | |
653 T3r = VADD(TD, TI); | |
654 T3s = VADD(TO, TT); | |
655 T3t = VSUB(T3r, T3s); | |
656 T4j = VADD(T3r, T3s); | |
657 T25 = VSUB(T22, T24); | |
658 T2c = VSUB(T29, T2b); | |
659 T1U = VSUB(T1R, T1T); | |
660 T1Z = VSUB(T1W, T1Y); | |
661 T20 = VMUL(LDK(KP707106781), VADD(T1U, T1Z)); | |
662 T27 = VMUL(LDK(KP707106781), VSUB(T1Z, T1U)); | |
663 T26 = VADD(T20, T25); | |
664 T2V = VADD(T27, T2c); | |
665 T2d = VSUB(T27, T2c); | |
666 T2U = VSUB(T25, T20); | |
667 } | |
668 { | |
669 V T4m, T4w, T4t, T4x, T4i, T4l, T4p, T4s, T4u, T4z, T4v, T4y, T4E, T4L, T4H; | |
670 V T4K, T4A, T4F, T4D, T4G, T4B, T4C, T4I, T4N, T4J, T4M, T3O, T4c, T4d, T3X; | |
671 V T40, T46, T49, T41, T3y, T47, T3T, T45, T3N, T44, T3W, T48, T3x, T3S, T3F; | |
672 V T3M, T3U, T3V, T3Y, T4e, T4f, T3Z, T42, T4a, T4b, T43; | |
673 T4i = VADD(T4g, T4h); | |
674 T4l = VADD(T4j, T4k); | |
675 T4m = VADD(T4i, T4l); | |
676 T4w = VSUB(T4i, T4l); | |
677 T4p = VADD(T4n, T4o); | |
678 T4s = VADD(T4q, T4r); | |
679 T4t = VADD(T4p, T4s); | |
680 T4x = VBYI(VSUB(T4s, T4p)); | |
681 T4u = VCONJ(VMUL(LDK(KP500000000), VSUB(T4m, T4t))); | |
682 ST(&(Rm[WS(rs, 15)]), T4u, -ms, &(Rm[WS(rs, 1)])); | |
683 T4z = VMUL(LDK(KP500000000), VADD(T4w, T4x)); | |
684 ST(&(Rp[WS(rs, 8)]), T4z, ms, &(Rp[0])); | |
685 T4v = VMUL(LDK(KP500000000), VADD(T4m, T4t)); | |
686 ST(&(Rp[0]), T4v, ms, &(Rp[0])); | |
687 T4y = VCONJ(VMUL(LDK(KP500000000), VSUB(T4w, T4x))); | |
688 ST(&(Rm[WS(rs, 7)]), T4y, -ms, &(Rm[WS(rs, 1)])); | |
689 T4A = VMUL(LDK(KP500000000), VSUB(T4g, T4h)); | |
690 T4F = VSUB(T4k, T4j); | |
691 T4B = VSUB(T4n, T4o); | |
692 T4C = VSUB(T4r, T4q); | |
693 T4D = VMUL(LDK(KP353553390), VADD(T4B, T4C)); | |
694 T4G = VMUL(LDK(KP707106781), VSUB(T4C, T4B)); | |
695 T4E = VADD(T4A, T4D); | |
696 T4L = VMUL(LDK(KP500000000), VBYI(VSUB(T4G, T4F))); | |
697 T4H = VMUL(LDK(KP500000000), VBYI(VADD(T4F, T4G))); | |
698 T4K = VSUB(T4A, T4D); | |
699 T4I = VCONJ(VSUB(T4E, T4H)); | |
700 ST(&(Rm[WS(rs, 3)]), T4I, -ms, &(Rm[WS(rs, 1)])); | |
701 T4N = VADD(T4K, T4L); | |
702 ST(&(Rp[WS(rs, 12)]), T4N, ms, &(Rp[0])); | |
703 T4J = VADD(T4E, T4H); | |
704 ST(&(Rp[WS(rs, 4)]), T4J, ms, &(Rp[0])); | |
705 T4M = VCONJ(VSUB(T4K, T4L)); | |
706 ST(&(Rm[WS(rs, 11)]), T4M, -ms, &(Rm[WS(rs, 1)])); | |
707 T3x = VMUL(LDK(KP353553390), VADD(T3t, T3w)); | |
708 T3y = VADD(T3q, T3x); | |
709 T47 = VSUB(T3q, T3x); | |
710 T3S = VMUL(LDK(KP707106781), VSUB(T3w, T3t)); | |
711 T3T = VADD(T3R, T3S); | |
712 T45 = VSUB(T3S, T3R); | |
713 T3F = VFMA(LDK(KP923879532), T3B, VMUL(LDK(KP382683432), T3E)); | |
714 T3M = VFNMS(LDK(KP382683432), T3L, VMUL(LDK(KP923879532), T3I)); | |
715 T3N = VMUL(LDK(KP500000000), VADD(T3F, T3M)); | |
716 T44 = VSUB(T3M, T3F); | |
717 T3U = VFNMS(LDK(KP382683432), T3B, VMUL(LDK(KP923879532), T3E)); | |
718 T3V = VFMA(LDK(KP923879532), T3L, VMUL(LDK(KP382683432), T3I)); | |
719 T3W = VADD(T3U, T3V); | |
720 T48 = VMUL(LDK(KP500000000), VSUB(T3V, T3U)); | |
721 T3O = VADD(T3y, T3N); | |
722 T4c = VMUL(LDK(KP500000000), VBYI(VADD(T45, T44))); | |
723 T4d = VADD(T47, T48); | |
724 T3X = VMUL(LDK(KP500000000), VBYI(VADD(T3T, T3W))); | |
725 T40 = VSUB(T3y, T3N); | |
726 T46 = VMUL(LDK(KP500000000), VBYI(VSUB(T44, T45))); | |
727 T49 = VSUB(T47, T48); | |
728 T41 = VMUL(LDK(KP500000000), VBYI(VSUB(T3W, T3T))); | |
729 T3Y = VCONJ(VSUB(T3O, T3X)); | |
730 ST(&(Rm[WS(rs, 1)]), T3Y, -ms, &(Rm[WS(rs, 1)])); | |
731 T4e = VADD(T4c, T4d); | |
732 ST(&(Rp[WS(rs, 6)]), T4e, ms, &(Rp[0])); | |
733 T4f = VCONJ(VSUB(T4d, T4c)); | |
734 ST(&(Rm[WS(rs, 5)]), T4f, -ms, &(Rm[WS(rs, 1)])); | |
735 T3Z = VADD(T3O, T3X); | |
736 ST(&(Rp[WS(rs, 2)]), T3Z, ms, &(Rp[0])); | |
737 T42 = VCONJ(VSUB(T40, T41)); | |
738 ST(&(Rm[WS(rs, 13)]), T42, -ms, &(Rm[WS(rs, 1)])); | |
739 T4a = VADD(T46, T49); | |
740 ST(&(Rp[WS(rs, 10)]), T4a, ms, &(Rp[0])); | |
741 T4b = VCONJ(VSUB(T49, T46)); | |
742 ST(&(Rm[WS(rs, 9)]), T4b, -ms, &(Rm[WS(rs, 1)])); | |
743 T43 = VADD(T40, T41); | |
744 ST(&(Rp[WS(rs, 14)]), T43, ms, &(Rp[0])); | |
745 { | |
746 V T2g, T2K, T2L, T2v, T2y, T2E, T2H, T2z, T1k, T2F, T2u, T2G, T2f, T2C, T2r; | |
747 V T2D, Ty, T1j, T2s, T2t, T1P, T2e, T2n, T2q, T2w, T2M, T2N, T2x, T2A, T2I; | |
748 V T2J, T2B; | |
749 Ty = VADD(Ta, Tx); | |
750 T1j = VMUL(LDK(KP500000000), VADD(TV, T1i)); | |
751 T1k = VADD(Ty, T1j); | |
752 T2F = VSUB(Ty, T1j); | |
753 T2s = VFNMS(LDK(KP195090322), T1B, VMUL(LDK(KP980785280), T1O)); | |
754 T2t = VFMA(LDK(KP195090322), T26, VMUL(LDK(KP980785280), T2d)); | |
755 T2u = VADD(T2s, T2t); | |
756 T2G = VMUL(LDK(KP500000000), VSUB(T2t, T2s)); | |
757 T1P = VFMA(LDK(KP980785280), T1B, VMUL(LDK(KP195090322), T1O)); | |
758 T2e = VFNMS(LDK(KP195090322), T2d, VMUL(LDK(KP980785280), T26)); | |
759 T2f = VMUL(LDK(KP500000000), VADD(T1P, T2e)); | |
760 T2C = VSUB(T2e, T1P); | |
761 T2n = VSUB(T2h, T2m); | |
762 T2q = VSUB(T2o, T2p); | |
763 T2r = VADD(T2n, T2q); | |
764 T2D = VSUB(T2q, T2n); | |
765 T2g = VADD(T1k, T2f); | |
766 T2K = VMUL(LDK(KP500000000), VBYI(VADD(T2D, T2C))); | |
767 T2L = VADD(T2F, T2G); | |
768 T2v = VMUL(LDK(KP500000000), VBYI(VADD(T2r, T2u))); | |
769 T2y = VSUB(T1k, T2f); | |
770 T2E = VMUL(LDK(KP500000000), VBYI(VSUB(T2C, T2D))); | |
771 T2H = VSUB(T2F, T2G); | |
772 T2z = VMUL(LDK(KP500000000), VBYI(VSUB(T2u, T2r))); | |
773 T2w = VCONJ(VSUB(T2g, T2v)); | |
774 ST(&(Rm[0]), T2w, -ms, &(Rm[0])); | |
775 T2M = VADD(T2K, T2L); | |
776 ST(&(Rp[WS(rs, 7)]), T2M, ms, &(Rp[WS(rs, 1)])); | |
777 T2N = VCONJ(VSUB(T2L, T2K)); | |
778 ST(&(Rm[WS(rs, 6)]), T2N, -ms, &(Rm[0])); | |
779 T2x = VADD(T2g, T2v); | |
780 ST(&(Rp[WS(rs, 1)]), T2x, ms, &(Rp[WS(rs, 1)])); | |
781 T2A = VCONJ(VSUB(T2y, T2z)); | |
782 ST(&(Rm[WS(rs, 14)]), T2A, -ms, &(Rm[0])); | |
783 T2I = VADD(T2E, T2H); | |
784 ST(&(Rp[WS(rs, 9)]), T2I, ms, &(Rp[WS(rs, 1)])); | |
785 T2J = VCONJ(VSUB(T2H, T2E)); | |
786 ST(&(Rm[WS(rs, 8)]), T2J, -ms, &(Rm[0])); | |
787 T2B = VADD(T2y, T2z); | |
788 ST(&(Rp[WS(rs, 15)]), T2B, ms, &(Rp[WS(rs, 1)])); | |
789 } | |
790 { | |
791 V T2Y, T3k, T3l, T35, T38, T3e, T3h, T39, T2Q, T3f, T34, T3g, T2X, T3c, T31; | |
792 V T3d, T2O, T2P, T32, T33, T2T, T2W, T2Z, T30, T36, T3m, T3n, T37, T3a, T3i; | |
793 V T3j, T3b; | |
794 T2O = VSUB(Ta, Tx); | |
795 T2P = VMUL(LDK(KP500000000), VADD(T2p, T2o)); | |
796 T2Q = VADD(T2O, T2P); | |
797 T3f = VSUB(T2O, T2P); | |
798 T32 = VFNMS(LDK(KP555570233), T2R, VMUL(LDK(KP831469612), T2S)); | |
799 T33 = VFMA(LDK(KP555570233), T2U, VMUL(LDK(KP831469612), T2V)); | |
800 T34 = VADD(T32, T33); | |
801 T3g = VMUL(LDK(KP500000000), VSUB(T33, T32)); | |
802 T2T = VFMA(LDK(KP831469612), T2R, VMUL(LDK(KP555570233), T2S)); | |
803 T2W = VFNMS(LDK(KP555570233), T2V, VMUL(LDK(KP831469612), T2U)); | |
804 T2X = VMUL(LDK(KP500000000), VADD(T2T, T2W)); | |
805 T3c = VSUB(T2W, T2T); | |
806 T2Z = VADD(T2m, T2h); | |
807 T30 = VSUB(T1i, TV); | |
808 T31 = VADD(T2Z, T30); | |
809 T3d = VSUB(T30, T2Z); | |
810 T2Y = VADD(T2Q, T2X); | |
811 T3k = VMUL(LDK(KP500000000), VBYI(VADD(T3d, T3c))); | |
812 T3l = VADD(T3f, T3g); | |
813 T35 = VMUL(LDK(KP500000000), VBYI(VADD(T31, T34))); | |
814 T38 = VSUB(T2Q, T2X); | |
815 T3e = VMUL(LDK(KP500000000), VBYI(VSUB(T3c, T3d))); | |
816 T3h = VSUB(T3f, T3g); | |
817 T39 = VMUL(LDK(KP500000000), VBYI(VSUB(T34, T31))); | |
818 T36 = VCONJ(VSUB(T2Y, T35)); | |
819 ST(&(Rm[WS(rs, 2)]), T36, -ms, &(Rm[0])); | |
820 T3m = VADD(T3k, T3l); | |
821 ST(&(Rp[WS(rs, 5)]), T3m, ms, &(Rp[WS(rs, 1)])); | |
822 T3n = VCONJ(VSUB(T3l, T3k)); | |
823 ST(&(Rm[WS(rs, 4)]), T3n, -ms, &(Rm[0])); | |
824 T37 = VADD(T2Y, T35); | |
825 ST(&(Rp[WS(rs, 3)]), T37, ms, &(Rp[WS(rs, 1)])); | |
826 T3a = VCONJ(VSUB(T38, T39)); | |
827 ST(&(Rm[WS(rs, 12)]), T3a, -ms, &(Rm[0])); | |
828 T3i = VADD(T3e, T3h); | |
829 ST(&(Rp[WS(rs, 11)]), T3i, ms, &(Rp[WS(rs, 1)])); | |
830 T3j = VCONJ(VSUB(T3h, T3e)); | |
831 ST(&(Rm[WS(rs, 10)]), T3j, -ms, &(Rm[0])); | |
832 T3b = VADD(T38, T39); | |
833 ST(&(Rp[WS(rs, 13)]), T3b, ms, &(Rp[WS(rs, 1)])); | |
834 } | |
835 } | |
836 } | |
837 } | |
838 VLEAVE(); | |
839 } | |
840 | |
841 static const tw_instr twinstr[] = { | |
842 VTW(1, 1), | |
843 VTW(1, 2), | |
844 VTW(1, 3), | |
845 VTW(1, 4), | |
846 VTW(1, 5), | |
847 VTW(1, 6), | |
848 VTW(1, 7), | |
849 VTW(1, 8), | |
850 VTW(1, 9), | |
851 VTW(1, 10), | |
852 VTW(1, 11), | |
853 VTW(1, 12), | |
854 VTW(1, 13), | |
855 VTW(1, 14), | |
856 VTW(1, 15), | |
857 VTW(1, 16), | |
858 VTW(1, 17), | |
859 VTW(1, 18), | |
860 VTW(1, 19), | |
861 VTW(1, 20), | |
862 VTW(1, 21), | |
863 VTW(1, 22), | |
864 VTW(1, 23), | |
865 VTW(1, 24), | |
866 VTW(1, 25), | |
867 VTW(1, 26), | |
868 VTW(1, 27), | |
869 VTW(1, 28), | |
870 VTW(1, 29), | |
871 VTW(1, 30), | |
872 VTW(1, 31), | |
873 {TW_NEXT, VL, 0} | |
874 }; | |
875 | |
876 static const hc2c_desc desc = { 32, XSIMD_STRING("hc2cfdftv_32"), twinstr, &GENUS, {233, 117, 16, 0} }; | |
877 | |
878 void XSIMD(codelet_hc2cfdftv_32) (planner *p) { | |
879 X(khc2c_register) (p, hc2cfdftv_32, &desc, HC2C_VIA_DFT); | |
880 } | |
881 #endif /* HAVE_FMA */ |