Mercurial > hg > sv-dependency-builds
comparison src/fftw-3.3.8/rdft/scalar/r2cb/r2cb_25.c @ 167:bd3cc4d1df30
Add FFTW 3.3.8 source, and a Linux build
author | Chris Cannam <cannam@all-day-breakfast.com> |
---|---|
date | Tue, 19 Nov 2019 14:52:55 +0000 |
parents | |
children |
comparison
equal
deleted
inserted
replaced
166:cbd6d7e562c7 | 167:bd3cc4d1df30 |
---|---|
1 /* | |
2 * Copyright (c) 2003, 2007-14 Matteo Frigo | |
3 * Copyright (c) 2003, 2007-14 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 Thu May 24 08:07:30 EDT 2018 */ | |
23 | |
24 #include "rdft/codelet-rdft.h" | |
25 | |
26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA) | |
27 | |
28 /* Generated by: ../../../genfft/gen_r2cb.native -fma -compact -variables 4 -pipeline-latency 4 -sign 1 -n 25 -name r2cb_25 -include rdft/scalar/r2cb.h */ | |
29 | |
30 /* | |
31 * This function contains 152 FP additions, 120 FP multiplications, | |
32 * (or, 32 additions, 0 multiplications, 120 fused multiply/add), | |
33 * 88 stack variables, 44 constants, and 50 memory accesses | |
34 */ | |
35 #include "rdft/scalar/r2cb.h" | |
36 | |
37 static void r2cb_25(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs) | |
38 { | |
39 DK(KP979740652, +0.979740652857618686258237536568998933733477632); | |
40 DK(KP1_752613360, +1.752613360087727174616231807844125166798128477); | |
41 DK(KP438153340, +0.438153340021931793654057951961031291699532119); | |
42 DK(KP963507348, +0.963507348203430549974383005744259307057084020); | |
43 DK(KP641441904, +0.641441904830606407298806329068862424939687989); | |
44 DK(KP595480289, +0.595480289600000014706716770488118292997907308); | |
45 DK(KP1_721083328, +1.721083328735889354196523361841037632825608373); | |
46 DK(KP1_606007150, +1.606007150877320829666881187140752009270929701); | |
47 DK(KP1_011627398, +1.011627398597394192215998921771049272931807941); | |
48 DK(KP1_809654104, +1.809654104932039055427337295865395187940827822); | |
49 DK(KP452413526, +0.452413526233009763856834323966348796985206956); | |
50 DK(KP933137358, +0.933137358350283770603023973254446451924190884); | |
51 DK(KP662318342, +0.662318342759882818626911127577439236802190210); | |
52 DK(KP576710603, +0.576710603632765877371579268136471017090111488); | |
53 DK(KP634619297, +0.634619297544148100711287640319130485732531031); | |
54 DK(KP470564281, +0.470564281212251493087595091036643380879947982); | |
55 DK(KP1_842354653, +1.842354653930286640500894870830132058718564461); | |
56 DK(KP1_666834356, +1.666834356657377354817925100486477686277992119); | |
57 DK(KP1_082908895, +1.082908895072625554092571180165639018104066379); | |
58 DK(KP1_937166322, +1.937166322257262238980336750929471627672024806); | |
59 DK(KP484291580, +0.484291580564315559745084187732367906918006201); | |
60 DK(KP904730450, +0.904730450839922351881287709692877908104763647); | |
61 DK(KP683113946, +0.683113946453479238701949862233725244439656928); | |
62 DK(KP559154169, +0.559154169276087864842202529084232643714075927); | |
63 DK(KP549754652, +0.549754652192770074288023275540779861653779767); | |
64 DK(KP256756360, +0.256756360367726783319498520922669048172391148); | |
65 DK(KP1_386580726, +1.386580726567734802700860150804827247498955921); | |
66 DK(KP1_898359647, +1.898359647016882523151110931686726543423167685); | |
67 DK(KP1_115827804, +1.115827804063668528375399296931134075984874304); | |
68 DK(KP1_996053456, +1.996053456856543123904673613726901106673810439); | |
69 DK(KP499013364, +0.499013364214135780976168403431725276668452610); | |
70 DK(KP730409924, +0.730409924561256563751459444999838399157094302); | |
71 DK(KP451418159, +0.451418159099103183892477933432151804893354132); | |
72 DK(KP846146756, +0.846146756728608505452954290121135880883743802); | |
73 DK(KP062914667, +0.062914667253649757225485955897349402364686947); | |
74 DK(KP939062505, +0.939062505817492352556001843133229685779824606); | |
75 DK(KP1_902113032, +1.902113032590307144232878666758764286811397268); | |
76 DK(KP951056516, +0.951056516295153572116439333379382143405698634); | |
77 DK(KP559016994, +0.559016994374947424102293417182819058860154590); | |
78 DK(KP250000000, +0.250000000000000000000000000000000000000000000); | |
79 DK(KP1_118033988, +1.118033988749894848204586834365638117720309180); | |
80 DK(KP500000000, +0.500000000000000000000000000000000000000000000); | |
81 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000); | |
82 DK(KP618033988, +0.618033988749894848204586834365638117720309180); | |
83 { | |
84 INT i; | |
85 for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(100, rs), MAKE_VOLATILE_STRIDE(100, csr), MAKE_VOLATILE_STRIDE(100, csi)) { | |
86 E Tu, T1G, T5, Tr, T1F, TH, TK, Te, TR, T2a, T1t, T1N, TG, T29, T1u; | |
87 E T1K, T14, T17, Tn, T1e, T26, T1q, T1R, T13, T27, T1r, T1U, Ts, Tt; | |
88 Ts = Ci[WS(csi, 5)]; | |
89 Tt = Ci[WS(csi, 10)]; | |
90 Tu = FMA(KP618033988, Tt, Ts); | |
91 T1G = FMS(KP618033988, Ts, Tt); | |
92 { | |
93 E T1, T4, Tq, T2, T3, Tp; | |
94 T1 = Cr[0]; | |
95 T2 = Cr[WS(csr, 5)]; | |
96 T3 = Cr[WS(csr, 10)]; | |
97 T4 = T2 + T3; | |
98 Tq = T2 - T3; | |
99 T5 = FMA(KP2_000000000, T4, T1); | |
100 Tp = FNMS(KP500000000, T4, T1); | |
101 Tr = FMA(KP1_118033988, Tq, Tp); | |
102 T1F = FNMS(KP1_118033988, Tq, Tp); | |
103 } | |
104 { | |
105 E T6, Td, TP, Tx, TO, TB, TM, TE; | |
106 T6 = Cr[WS(csr, 1)]; | |
107 TH = Ci[WS(csi, 1)]; | |
108 { | |
109 E T7, T8, T9, Ta, Tb, Tc; | |
110 T7 = Cr[WS(csr, 6)]; | |
111 T8 = Cr[WS(csr, 4)]; | |
112 T9 = T7 + T8; | |
113 Ta = Cr[WS(csr, 11)]; | |
114 Tb = Cr[WS(csr, 9)]; | |
115 Tc = Ta + Tb; | |
116 Td = T9 + Tc; | |
117 TP = Tb - Ta; | |
118 Tx = T9 - Tc; | |
119 TO = T7 - T8; | |
120 } | |
121 { | |
122 E Tz, TA, TI, TC, TD, TJ; | |
123 Tz = Ci[WS(csi, 6)]; | |
124 TA = Ci[WS(csi, 4)]; | |
125 TI = Tz - TA; | |
126 TC = Ci[WS(csi, 11)]; | |
127 TD = Ci[WS(csi, 9)]; | |
128 TJ = TC - TD; | |
129 TB = Tz + TA; | |
130 TM = TI - TJ; | |
131 TE = TC + TD; | |
132 TK = TI + TJ; | |
133 } | |
134 Te = T6 + Td; | |
135 { | |
136 E TQ, T1M, TN, T1L, TL; | |
137 TQ = FNMS(KP618033988, TP, TO); | |
138 T1M = FMA(KP618033988, TO, TP); | |
139 TL = FNMS(KP250000000, TK, TH); | |
140 TN = FMA(KP559016994, TM, TL); | |
141 T1L = FNMS(KP559016994, TM, TL); | |
142 TR = FMA(KP951056516, TQ, TN); | |
143 T2a = FNMS(KP951056516, T1M, T1L); | |
144 T1t = FNMS(KP951056516, TQ, TN); | |
145 T1N = FMA(KP951056516, T1M, T1L); | |
146 } | |
147 { | |
148 E TF, T1J, Ty, T1I, Tw; | |
149 TF = FMA(KP618033988, TE, TB); | |
150 T1J = FNMS(KP618033988, TB, TE); | |
151 Tw = FNMS(KP250000000, Td, T6); | |
152 Ty = FMA(KP559016994, Tx, Tw); | |
153 T1I = FNMS(KP559016994, Tx, Tw); | |
154 TG = FNMS(KP951056516, TF, Ty); | |
155 T29 = FNMS(KP951056516, T1J, T1I); | |
156 T1u = FMA(KP951056516, TF, Ty); | |
157 T1K = FMA(KP951056516, T1J, T1I); | |
158 } | |
159 } | |
160 { | |
161 E Tf, Tm, T1c, TU, T1b, TY, T19, T11; | |
162 Tf = Cr[WS(csr, 2)]; | |
163 T14 = Ci[WS(csi, 2)]; | |
164 { | |
165 E Tg, Th, Ti, Tj, Tk, Tl; | |
166 Tg = Cr[WS(csr, 7)]; | |
167 Th = Cr[WS(csr, 3)]; | |
168 Ti = Tg + Th; | |
169 Tj = Cr[WS(csr, 12)]; | |
170 Tk = Cr[WS(csr, 8)]; | |
171 Tl = Tj + Tk; | |
172 Tm = Ti + Tl; | |
173 T1c = Tj - Tk; | |
174 TU = Tl - Ti; | |
175 T1b = Th - Tg; | |
176 } | |
177 { | |
178 E TW, TX, T15, TZ, T10, T16; | |
179 TW = Ci[WS(csi, 7)]; | |
180 TX = Ci[WS(csi, 3)]; | |
181 T15 = TW - TX; | |
182 TZ = Ci[WS(csi, 12)]; | |
183 T10 = Ci[WS(csi, 8)]; | |
184 T16 = TZ - T10; | |
185 TY = TW + TX; | |
186 T19 = T16 - T15; | |
187 T11 = TZ + T10; | |
188 T17 = T15 + T16; | |
189 } | |
190 Tn = Tf + Tm; | |
191 { | |
192 E T1d, T1Q, T1a, T1P, T18; | |
193 T1d = FNMS(KP618033988, T1c, T1b); | |
194 T1Q = FMA(KP618033988, T1b, T1c); | |
195 T18 = FNMS(KP250000000, T17, T14); | |
196 T1a = FNMS(KP559016994, T19, T18); | |
197 T1P = FMA(KP559016994, T19, T18); | |
198 T1e = FNMS(KP951056516, T1d, T1a); | |
199 T26 = FMA(KP951056516, T1Q, T1P); | |
200 T1q = FMA(KP951056516, T1d, T1a); | |
201 T1R = FNMS(KP951056516, T1Q, T1P); | |
202 } | |
203 { | |
204 E T12, T1T, TV, T1S, TT; | |
205 T12 = FMA(KP618033988, T11, TY); | |
206 T1T = FNMS(KP618033988, TY, T11); | |
207 TT = FNMS(KP250000000, Tm, Tf); | |
208 TV = FNMS(KP559016994, TU, TT); | |
209 T1S = FMA(KP559016994, TU, TT); | |
210 T13 = FNMS(KP951056516, T12, TV); | |
211 T27 = FNMS(KP951056516, T1T, T1S); | |
212 T1r = FMA(KP951056516, T12, TV); | |
213 T1U = FMA(KP951056516, T1T, T1S); | |
214 } | |
215 } | |
216 { | |
217 E T2m, To, T2l, T2q, T2s, T2o, T2p, T2r, T2n; | |
218 T2m = Te - Tn; | |
219 To = Te + Tn; | |
220 T2l = FNMS(KP500000000, To, T5); | |
221 T2o = TK + TH; | |
222 T2p = T17 + T14; | |
223 T2q = FMA(KP618033988, T2p, T2o); | |
224 T2s = FNMS(KP618033988, T2o, T2p); | |
225 R0[0] = FMA(KP2_000000000, To, T5); | |
226 T2r = FNMS(KP1_118033988, T2m, T2l); | |
227 R1[WS(rs, 7)] = FNMS(KP1_902113032, T2s, T2r); | |
228 R0[WS(rs, 5)] = FMA(KP1_902113032, T2s, T2r); | |
229 T2n = FMA(KP1_118033988, T2m, T2l); | |
230 R1[WS(rs, 2)] = FNMS(KP1_902113032, T2q, T2n); | |
231 R0[WS(rs, 10)] = FMA(KP1_902113032, T2q, T2n); | |
232 } | |
233 { | |
234 E T2i, T2k, T25, T2c, T2d, T2e, T2j, T2f; | |
235 { | |
236 E T2g, T2h, T28, T2b; | |
237 T2g = FMA(KP939062505, T29, T2a); | |
238 T2h = FMA(KP062914667, T26, T27); | |
239 T2i = FMA(KP846146756, T2h, T2g); | |
240 T2k = FNMS(KP451418159, T2g, T2h); | |
241 T25 = FMA(KP1_902113032, T1G, T1F); | |
242 T28 = FNMS(KP062914667, T27, T26); | |
243 T2b = FNMS(KP939062505, T2a, T29); | |
244 T2c = FNMS(KP730409924, T2b, T28); | |
245 T2d = FMA(KP499013364, T2c, T25); | |
246 T2e = FMA(KP730409924, T2b, T28); | |
247 } | |
248 R1[WS(rs, 1)] = FNMS(KP1_996053456, T2c, T25); | |
249 T2j = FNMS(KP1_115827804, T2e, T2d); | |
250 R0[WS(rs, 9)] = FNMS(KP1_898359647, T2k, T2j); | |
251 R1[WS(rs, 6)] = FMA(KP1_898359647, T2k, T2j); | |
252 T2f = FMA(KP1_115827804, T2e, T2d); | |
253 R0[WS(rs, 4)] = FNMS(KP1_386580726, T2i, T2f); | |
254 R1[WS(rs, 11)] = FMA(KP1_386580726, T2i, T2f); | |
255 } | |
256 { | |
257 E T1m, T1o, Tv, T1g, T1h, T1i, T1n, T1j; | |
258 { | |
259 E T1k, T1l, TS, T1f; | |
260 T1k = FMA(KP256756360, TG, TR); | |
261 T1l = FMA(KP549754652, T13, T1e); | |
262 T1m = FMA(KP559154169, T1l, T1k); | |
263 T1o = FNMS(KP683113946, T1k, T1l); | |
264 Tv = FNMS(KP1_902113032, Tu, Tr); | |
265 TS = FNMS(KP256756360, TR, TG); | |
266 T1f = FNMS(KP549754652, T1e, T13); | |
267 T1g = FMA(KP904730450, T1f, TS); | |
268 T1h = FNMS(KP484291580, T1g, Tv); | |
269 T1i = FNMS(KP904730450, T1f, TS); | |
270 } | |
271 R1[0] = FMA(KP1_937166322, T1g, Tv); | |
272 T1n = FNMS(KP1_082908895, T1i, T1h); | |
273 R0[WS(rs, 8)] = FNMS(KP1_666834356, T1o, T1n); | |
274 R1[WS(rs, 5)] = FMA(KP1_666834356, T1o, T1n); | |
275 T1j = FMA(KP1_082908895, T1i, T1h); | |
276 R0[WS(rs, 3)] = FNMS(KP1_842354653, T1m, T1j); | |
277 R1[WS(rs, 10)] = FMA(KP1_842354653, T1m, T1j); | |
278 } | |
279 { | |
280 E T1C, T1E, T1p, T1w, T1x, T1y, T1D, T1z; | |
281 { | |
282 E T1A, T1B, T1s, T1v; | |
283 T1A = FNMS(KP470564281, T1q, T1r); | |
284 T1B = FMA(KP634619297, T1t, T1u); | |
285 T1C = FNMS(KP576710603, T1B, T1A); | |
286 T1E = FMA(KP662318342, T1A, T1B); | |
287 T1p = FMA(KP1_902113032, Tu, Tr); | |
288 T1s = FMA(KP470564281, T1r, T1q); | |
289 T1v = FNMS(KP634619297, T1u, T1t); | |
290 T1w = FMA(KP933137358, T1v, T1s); | |
291 T1x = FMA(KP452413526, T1w, T1p); | |
292 T1y = FNMS(KP933137358, T1v, T1s); | |
293 } | |
294 R0[WS(rs, 2)] = FNMS(KP1_809654104, T1w, T1p); | |
295 T1D = FMA(KP1_011627398, T1y, T1x); | |
296 R1[WS(rs, 4)] = FNMS(KP1_606007150, T1E, T1D); | |
297 R0[WS(rs, 12)] = FMA(KP1_606007150, T1E, T1D); | |
298 T1z = FNMS(KP1_011627398, T1y, T1x); | |
299 R0[WS(rs, 7)] = FMA(KP1_721083328, T1C, T1z); | |
300 R1[WS(rs, 9)] = FNMS(KP1_721083328, T1C, T1z); | |
301 } | |
302 { | |
303 E T22, T24, T1H, T1W, T1X, T1Y, T23, T1Z; | |
304 { | |
305 E T20, T21, T1O, T1V; | |
306 T20 = FMA(KP549754652, T1K, T1N); | |
307 T21 = FMA(KP634619297, T1R, T1U); | |
308 T22 = FMA(KP595480289, T21, T20); | |
309 T24 = FNMS(KP641441904, T20, T21); | |
310 T1H = FNMS(KP1_902113032, T1G, T1F); | |
311 T1O = FNMS(KP549754652, T1N, T1K); | |
312 T1V = FNMS(KP634619297, T1U, T1R); | |
313 T1W = FNMS(KP963507348, T1V, T1O); | |
314 T1X = FNMS(KP438153340, T1W, T1H); | |
315 T1Y = FMA(KP963507348, T1V, T1O); | |
316 } | |
317 R0[WS(rs, 1)] = FMA(KP1_752613360, T1W, T1H); | |
318 T23 = FNMS(KP979740652, T1Y, T1X); | |
319 R0[WS(rs, 6)] = FMA(KP1_606007150, T24, T23); | |
320 R1[WS(rs, 8)] = FNMS(KP1_606007150, T24, T23); | |
321 T1Z = FMA(KP979740652, T1Y, T1X); | |
322 R1[WS(rs, 3)] = FNMS(KP1_666834356, T22, T1Z); | |
323 R0[WS(rs, 11)] = FMA(KP1_666834356, T22, T1Z); | |
324 } | |
325 } | |
326 } | |
327 } | |
328 | |
329 static const kr2c_desc desc = { 25, "r2cb_25", {32, 0, 120, 0}, &GENUS }; | |
330 | |
331 void X(codelet_r2cb_25) (planner *p) { | |
332 X(kr2c_register) (p, r2cb_25, &desc); | |
333 } | |
334 | |
335 #else | |
336 | |
337 /* Generated by: ../../../genfft/gen_r2cb.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 25 -name r2cb_25 -include rdft/scalar/r2cb.h */ | |
338 | |
339 /* | |
340 * This function contains 152 FP additions, 98 FP multiplications, | |
341 * (or, 100 additions, 46 multiplications, 52 fused multiply/add), | |
342 * 65 stack variables, 21 constants, and 50 memory accesses | |
343 */ | |
344 #include "rdft/scalar/r2cb.h" | |
345 | |
346 static void r2cb_25(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs) | |
347 { | |
348 DK(KP425779291, +0.425779291565072648862502445744251703979973042); | |
349 DK(KP904827052, +0.904827052466019527713668647932697593970413911); | |
350 DK(KP535826794, +0.535826794978996618271308767867639978063575346); | |
351 DK(KP844327925, +0.844327925502015078548558063966681505381659241); | |
352 DK(KP876306680, +0.876306680043863587308115903922062583399064238); | |
353 DK(KP481753674, +0.481753674101715274987191502872129653528542010); | |
354 DK(KP968583161, +0.968583161128631119490168375464735813836012403); | |
355 DK(KP248689887, +0.248689887164854788242283746006447968417567406); | |
356 DK(KP062790519, +0.062790519529313376076178224565631133122484832); | |
357 DK(KP998026728, +0.998026728428271561952336806863450553336905220); | |
358 DK(KP728968627, +0.728968627421411523146730319055259111372571664); | |
359 DK(KP684547105, +0.684547105928688673732283357621209269889519233); | |
360 DK(KP250000000, +0.250000000000000000000000000000000000000000000); | |
361 DK(KP587785252, +0.587785252292473129168705954639072768597652438); | |
362 DK(KP951056516, +0.951056516295153572116439333379382143405698634); | |
363 DK(KP559016994, +0.559016994374947424102293417182819058860154590); | |
364 DK(KP500000000, +0.500000000000000000000000000000000000000000000); | |
365 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000); | |
366 DK(KP1_118033988, +1.118033988749894848204586834365638117720309180); | |
367 DK(KP1_175570504, +1.175570504584946258337411909278145537195304875); | |
368 DK(KP1_902113032, +1.902113032590307144232878666758764286811397268); | |
369 { | |
370 INT i; | |
371 for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(100, rs), MAKE_VOLATILE_STRIDE(100, csr), MAKE_VOLATILE_STRIDE(100, csi)) { | |
372 E Tu, T1G, T5, Tr, T1F, TN, TO, Te, TR, T27, T1r, T1N, TG, T26, T1q; | |
373 E T1K, T1a, T1b, Tn, T1e, T2a, T1u, T1U, T13, T29, T1t, T1R, Ts, Tt; | |
374 Ts = Ci[WS(csi, 5)]; | |
375 Tt = Ci[WS(csi, 10)]; | |
376 Tu = FMA(KP1_902113032, Ts, KP1_175570504 * Tt); | |
377 T1G = FNMS(KP1_902113032, Tt, KP1_175570504 * Ts); | |
378 { | |
379 E T1, T4, Tp, T2, T3, Tq; | |
380 T1 = Cr[0]; | |
381 T2 = Cr[WS(csr, 5)]; | |
382 T3 = Cr[WS(csr, 10)]; | |
383 T4 = T2 + T3; | |
384 Tp = KP1_118033988 * (T2 - T3); | |
385 T5 = FMA(KP2_000000000, T4, T1); | |
386 Tq = FNMS(KP500000000, T4, T1); | |
387 Tr = Tp + Tq; | |
388 T1F = Tq - Tp; | |
389 } | |
390 { | |
391 E T6, Td, TI, Tw, TH, TB, TE, TM; | |
392 T6 = Cr[WS(csr, 1)]; | |
393 TN = Ci[WS(csi, 1)]; | |
394 { | |
395 E T7, T8, T9, Ta, Tb, Tc; | |
396 T7 = Cr[WS(csr, 6)]; | |
397 T8 = Cr[WS(csr, 4)]; | |
398 T9 = T7 + T8; | |
399 Ta = Cr[WS(csr, 11)]; | |
400 Tb = Cr[WS(csr, 9)]; | |
401 Tc = Ta + Tb; | |
402 Td = T9 + Tc; | |
403 TI = Ta - Tb; | |
404 Tw = KP559016994 * (T9 - Tc); | |
405 TH = T7 - T8; | |
406 } | |
407 { | |
408 E Tz, TA, TK, TC, TD, TL; | |
409 Tz = Ci[WS(csi, 6)]; | |
410 TA = Ci[WS(csi, 4)]; | |
411 TK = Tz - TA; | |
412 TC = Ci[WS(csi, 11)]; | |
413 TD = Ci[WS(csi, 9)]; | |
414 TL = TC - TD; | |
415 TB = Tz + TA; | |
416 TO = TK + TL; | |
417 TE = TC + TD; | |
418 TM = KP559016994 * (TK - TL); | |
419 } | |
420 Te = T6 + Td; | |
421 { | |
422 E TJ, T1L, TQ, T1M, TP; | |
423 TJ = FMA(KP951056516, TH, KP587785252 * TI); | |
424 T1L = FNMS(KP951056516, TI, KP587785252 * TH); | |
425 TP = FNMS(KP250000000, TO, TN); | |
426 TQ = TM + TP; | |
427 T1M = TP - TM; | |
428 TR = TJ + TQ; | |
429 T27 = T1M - T1L; | |
430 T1r = TQ - TJ; | |
431 T1N = T1L + T1M; | |
432 } | |
433 { | |
434 E TF, T1J, Ty, T1I, Tx; | |
435 TF = FMA(KP951056516, TB, KP587785252 * TE); | |
436 T1J = FNMS(KP951056516, TE, KP587785252 * TB); | |
437 Tx = FNMS(KP250000000, Td, T6); | |
438 Ty = Tw + Tx; | |
439 T1I = Tx - Tw; | |
440 TG = Ty - TF; | |
441 T26 = T1I + T1J; | |
442 T1q = Ty + TF; | |
443 T1K = T1I - T1J; | |
444 } | |
445 } | |
446 { | |
447 E Tf, Tm, T15, TT, T14, TY, T11, T19; | |
448 Tf = Cr[WS(csr, 2)]; | |
449 T1a = Ci[WS(csi, 2)]; | |
450 { | |
451 E Tg, Th, Ti, Tj, Tk, Tl; | |
452 Tg = Cr[WS(csr, 7)]; | |
453 Th = Cr[WS(csr, 3)]; | |
454 Ti = Tg + Th; | |
455 Tj = Cr[WS(csr, 12)]; | |
456 Tk = Cr[WS(csr, 8)]; | |
457 Tl = Tj + Tk; | |
458 Tm = Ti + Tl; | |
459 T15 = Tj - Tk; | |
460 TT = KP559016994 * (Ti - Tl); | |
461 T14 = Tg - Th; | |
462 } | |
463 { | |
464 E TW, TX, T17, TZ, T10, T18; | |
465 TW = Ci[WS(csi, 7)]; | |
466 TX = Ci[WS(csi, 3)]; | |
467 T17 = TW - TX; | |
468 TZ = Ci[WS(csi, 12)]; | |
469 T10 = Ci[WS(csi, 8)]; | |
470 T18 = TZ - T10; | |
471 TY = TW + TX; | |
472 T1b = T17 + T18; | |
473 T11 = TZ + T10; | |
474 T19 = KP559016994 * (T17 - T18); | |
475 } | |
476 Tn = Tf + Tm; | |
477 { | |
478 E T16, T1S, T1d, T1T, T1c; | |
479 T16 = FMA(KP951056516, T14, KP587785252 * T15); | |
480 T1S = FNMS(KP951056516, T15, KP587785252 * T14); | |
481 T1c = FNMS(KP250000000, T1b, T1a); | |
482 T1d = T19 + T1c; | |
483 T1T = T1c - T19; | |
484 T1e = T16 + T1d; | |
485 T2a = T1T - T1S; | |
486 T1u = T1d - T16; | |
487 T1U = T1S + T1T; | |
488 } | |
489 { | |
490 E T12, T1Q, TV, T1P, TU; | |
491 T12 = FMA(KP951056516, TY, KP587785252 * T11); | |
492 T1Q = FNMS(KP951056516, T11, KP587785252 * TY); | |
493 TU = FNMS(KP250000000, Tm, Tf); | |
494 TV = TT + TU; | |
495 T1P = TU - TT; | |
496 T13 = TV - T12; | |
497 T29 = T1P + T1Q; | |
498 T1t = TV + T12; | |
499 T1R = T1P - T1Q; | |
500 } | |
501 } | |
502 { | |
503 E T2m, To, T2l, T2q, T2s, T2o, T2p, T2r, T2n; | |
504 T2m = KP1_118033988 * (Te - Tn); | |
505 To = Te + Tn; | |
506 T2l = FNMS(KP500000000, To, T5); | |
507 T2o = TO + TN; | |
508 T2p = T1b + T1a; | |
509 T2q = FNMS(KP1_902113032, T2p, KP1_175570504 * T2o); | |
510 T2s = FMA(KP1_902113032, T2o, KP1_175570504 * T2p); | |
511 R0[0] = FMA(KP2_000000000, To, T5); | |
512 T2r = T2m + T2l; | |
513 R1[WS(rs, 2)] = T2r - T2s; | |
514 R0[WS(rs, 10)] = T2r + T2s; | |
515 T2n = T2l - T2m; | |
516 R0[WS(rs, 5)] = T2n - T2q; | |
517 R1[WS(rs, 7)] = T2n + T2q; | |
518 } | |
519 { | |
520 E T2i, T2k, T25, T2c, T2d, T2e, T2j, T2f; | |
521 { | |
522 E T2g, T2h, T28, T2b; | |
523 T2g = FMA(KP684547105, T26, KP728968627 * T27); | |
524 T2h = FMA(KP998026728, T29, KP062790519 * T2a); | |
525 T2i = FNMS(KP1_902113032, T2h, KP1_175570504 * T2g); | |
526 T2k = FMA(KP1_902113032, T2g, KP1_175570504 * T2h); | |
527 T25 = T1F + T1G; | |
528 T28 = FNMS(KP684547105, T27, KP728968627 * T26); | |
529 T2b = FNMS(KP998026728, T2a, KP062790519 * T29); | |
530 T2c = T28 + T2b; | |
531 T2d = FNMS(KP500000000, T2c, T25); | |
532 T2e = KP1_118033988 * (T28 - T2b); | |
533 } | |
534 R1[WS(rs, 1)] = FMA(KP2_000000000, T2c, T25); | |
535 T2j = T2e + T2d; | |
536 R0[WS(rs, 4)] = T2j - T2k; | |
537 R1[WS(rs, 11)] = T2j + T2k; | |
538 T2f = T2d - T2e; | |
539 R1[WS(rs, 6)] = T2f - T2i; | |
540 R0[WS(rs, 9)] = T2f + T2i; | |
541 } | |
542 { | |
543 E T1m, T1o, Tv, T1g, T1h, T1i, T1n, T1j; | |
544 { | |
545 E T1k, T1l, TS, T1f; | |
546 T1k = FMA(KP248689887, TG, KP968583161 * TR); | |
547 T1l = FMA(KP481753674, T13, KP876306680 * T1e); | |
548 T1m = FNMS(KP1_902113032, T1l, KP1_175570504 * T1k); | |
549 T1o = FMA(KP1_902113032, T1k, KP1_175570504 * T1l); | |
550 Tv = Tr - Tu; | |
551 TS = FNMS(KP248689887, TR, KP968583161 * TG); | |
552 T1f = FNMS(KP481753674, T1e, KP876306680 * T13); | |
553 T1g = TS + T1f; | |
554 T1h = FNMS(KP500000000, T1g, Tv); | |
555 T1i = KP1_118033988 * (TS - T1f); | |
556 } | |
557 R1[0] = FMA(KP2_000000000, T1g, Tv); | |
558 T1n = T1i + T1h; | |
559 R0[WS(rs, 3)] = T1n - T1o; | |
560 R1[WS(rs, 10)] = T1n + T1o; | |
561 T1j = T1h - T1i; | |
562 R1[WS(rs, 5)] = T1j - T1m; | |
563 R0[WS(rs, 8)] = T1j + T1m; | |
564 } | |
565 { | |
566 E T1C, T1E, T1p, T1w, T1x, T1y, T1D, T1z; | |
567 { | |
568 E T1A, T1B, T1s, T1v; | |
569 T1A = FMA(KP844327925, T1q, KP535826794 * T1r); | |
570 T1B = FNMS(KP425779291, T1u, KP904827052 * T1t); | |
571 T1C = FNMS(KP1_902113032, T1B, KP1_175570504 * T1A); | |
572 T1E = FMA(KP1_902113032, T1A, KP1_175570504 * T1B); | |
573 T1p = Tr + Tu; | |
574 T1s = FNMS(KP844327925, T1r, KP535826794 * T1q); | |
575 T1v = FMA(KP425779291, T1t, KP904827052 * T1u); | |
576 T1w = T1s - T1v; | |
577 T1x = FNMS(KP500000000, T1w, T1p); | |
578 T1y = KP1_118033988 * (T1s + T1v); | |
579 } | |
580 R0[WS(rs, 2)] = FMA(KP2_000000000, T1w, T1p); | |
581 T1D = T1x + T1y; | |
582 R1[WS(rs, 4)] = T1D - T1E; | |
583 R0[WS(rs, 12)] = T1E + T1D; | |
584 T1z = T1x - T1y; | |
585 R0[WS(rs, 7)] = T1z - T1C; | |
586 R1[WS(rs, 9)] = T1C + T1z; | |
587 } | |
588 { | |
589 E T22, T24, T1H, T1W, T1X, T1Y, T23, T1Z; | |
590 { | |
591 E T20, T21, T1O, T1V; | |
592 T20 = FMA(KP481753674, T1K, KP876306680 * T1N); | |
593 T21 = FMA(KP844327925, T1R, KP535826794 * T1U); | |
594 T22 = FNMS(KP1_902113032, T21, KP1_175570504 * T20); | |
595 T24 = FMA(KP1_902113032, T20, KP1_175570504 * T21); | |
596 T1H = T1F - T1G; | |
597 T1O = FNMS(KP481753674, T1N, KP876306680 * T1K); | |
598 T1V = FNMS(KP844327925, T1U, KP535826794 * T1R); | |
599 T1W = T1O + T1V; | |
600 T1X = FNMS(KP500000000, T1W, T1H); | |
601 T1Y = KP1_118033988 * (T1O - T1V); | |
602 } | |
603 R0[WS(rs, 1)] = FMA(KP2_000000000, T1W, T1H); | |
604 T23 = T1Y + T1X; | |
605 R1[WS(rs, 3)] = T23 - T24; | |
606 R0[WS(rs, 11)] = T23 + T24; | |
607 T1Z = T1X - T1Y; | |
608 R0[WS(rs, 6)] = T1Z - T22; | |
609 R1[WS(rs, 8)] = T1Z + T22; | |
610 } | |
611 } | |
612 } | |
613 } | |
614 | |
615 static const kr2c_desc desc = { 25, "r2cb_25", {100, 46, 52, 0}, &GENUS }; | |
616 | |
617 void X(codelet_r2cb_25) (planner *p) { | |
618 X(kr2c_register) (p, r2cb_25, &desc); | |
619 } | |
620 | |
621 #endif |