Mercurial > hg > js-dsp-test
comparison fft/fftw/fftw-3.3.4/rdft/scalar/r2cb/r2cbIII_25.c @ 19:26056e866c29
Add FFTW to comparison table
author | Chris Cannam |
---|---|
date | Tue, 06 Oct 2015 13:08:39 +0100 |
parents | |
children |
comparison
equal
deleted
inserted
replaced
18:8db794ca3e0b | 19:26056e866c29 |
---|---|
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 Tue Mar 4 13:50:36 EST 2014 */ | |
23 | |
24 #include "codelet-rdft.h" | |
25 | |
26 #ifdef HAVE_FMA | |
27 | |
28 /* Generated by: ../../../genfft/gen_r2cb.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -sign 1 -n 25 -name r2cbIII_25 -dft-III -include r2cbIII.h */ | |
29 | |
30 /* | |
31 * This function contains 152 FP additions, 120 FP multiplications, | |
32 * (or, 32 additions, 0 multiplications, 120 fused multiply/add), | |
33 * 115 stack variables, 44 constants, and 50 memory accesses | |
34 */ | |
35 #include "r2cbIII.h" | |
36 | |
37 static void r2cbIII_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(KP438153340, +0.438153340021931793654057951961031291699532119); | |
41 DK(KP1_752613360, +1.752613360087727174616231807844125166798128477); | |
42 DK(KP963507348, +0.963507348203430549974383005744259307057084020); | |
43 DK(KP1_721083328, +1.721083328735889354196523361841037632825608373); | |
44 DK(KP1_606007150, +1.606007150877320829666881187140752009270929701); | |
45 DK(KP1_011627398, +1.011627398597394192215998921771049272931807941); | |
46 DK(KP641441904, +0.641441904830606407298806329068862424939687989); | |
47 DK(KP595480289, +0.595480289600000014706716770488118292997907308); | |
48 DK(KP452413526, +0.452413526233009763856834323966348796985206956); | |
49 DK(KP1_809654104, +1.809654104932039055427337295865395187940827822); | |
50 DK(KP933137358, +0.933137358350283770603023973254446451924190884); | |
51 DK(KP1_666834356, +1.666834356657377354817925100486477686277992119); | |
52 DK(KP1_842354653, +1.842354653930286640500894870830132058718564461); | |
53 DK(KP1_082908895, +1.082908895072625554092571180165639018104066379); | |
54 DK(KP576710603, +0.576710603632765877371579268136471017090111488); | |
55 DK(KP662318342, +0.662318342759882818626911127577439236802190210); | |
56 DK(KP484291580, +0.484291580564315559745084187732367906918006201); | |
57 DK(KP1_937166322, +1.937166322257262238980336750929471627672024806); | |
58 DK(KP1_898359647, +1.898359647016882523151110931686726543423167685); | |
59 DK(KP1_386580726, +1.386580726567734802700860150804827247498955921); | |
60 DK(KP904730450, +0.904730450839922351881287709692877908104763647); | |
61 DK(KP1_115827804, +1.115827804063668528375399296931134075984874304); | |
62 DK(KP470564281, +0.470564281212251493087595091036643380879947982); | |
63 DK(KP634619297, +0.634619297544148100711287640319130485732531031); | |
64 DK(KP499013364, +0.499013364214135780976168403431725276668452610); | |
65 DK(KP1_996053456, +1.996053456856543123904673613726901106673810439); | |
66 DK(KP559154169, +0.559154169276087864842202529084232643714075927); | |
67 DK(KP683113946, +0.683113946453479238701949862233725244439656928); | |
68 DK(KP730409924, +0.730409924561256563751459444999838399157094302); | |
69 DK(KP549754652, +0.549754652192770074288023275540779861653779767); | |
70 DK(KP256756360, +0.256756360367726783319498520922669048172391148); | |
71 DK(KP451418159, +0.451418159099103183892477933432151804893354132); | |
72 DK(KP846146756, +0.846146756728608505452954290121135880883743802); | |
73 DK(KP1_902113032, +1.902113032590307144232878666758764286811397268); | |
74 DK(KP062914667, +0.062914667253649757225485955897349402364686947); | |
75 DK(KP939062505, +0.939062505817492352556001843133229685779824606); | |
76 DK(KP951056516, +0.951056516295153572116439333379382143405698634); | |
77 DK(KP559016994, +0.559016994374947424102293417182819058860154590); | |
78 DK(KP250000000, +0.250000000000000000000000000000000000000000000); | |
79 DK(KP1_118033988, +1.118033988749894848204586834365638117720309180); | |
80 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000); | |
81 DK(KP500000000, +0.500000000000000000000000000000000000000000000); | |
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 T1P, T2c, T2a, T24, T26, T25, T27, T2b; | |
87 { | |
88 E T1O, TS, T5, T1N, TP, Te, TA, T2i, T1V, T17, T1B, T2h, T1S, T10, T1C; | |
89 E T1a, T19, Tn, T1h, T1l, T1Y, T1e, T21, TJ, T1g; | |
90 { | |
91 E T1, T2, T3, TQ, TR; | |
92 TQ = Ci[WS(csi, 7)]; | |
93 TR = Ci[WS(csi, 2)]; | |
94 T1 = Cr[WS(csr, 12)]; | |
95 T2 = Cr[WS(csr, 7)]; | |
96 T3 = Cr[WS(csr, 2)]; | |
97 T1O = FNMS(KP618033988, TQ, TR); | |
98 TS = FMA(KP618033988, TR, TQ); | |
99 { | |
100 E TV, TU, T1U, T16, T12, T1R, TZ, T11; | |
101 { | |
102 E T6, Tz, T14, T15, TX, Tu, Td, Tx, TY, T4, TO, Ty; | |
103 T6 = Cr[WS(csr, 11)]; | |
104 T4 = T2 + T3; | |
105 TO = T3 - T2; | |
106 Tz = Ci[WS(csi, 11)]; | |
107 { | |
108 E Ta, T9, Tb, T7, T8, TN; | |
109 T7 = Cr[WS(csr, 6)]; | |
110 T8 = Cr[WS(csr, 8)]; | |
111 TN = FNMS(KP500000000, T4, T1); | |
112 T5 = FMA(KP2_000000000, T4, T1); | |
113 Ta = Cr[WS(csr, 1)]; | |
114 T14 = T8 - T7; | |
115 T9 = T7 + T8; | |
116 T1N = FMA(KP1_118033988, TO, TN); | |
117 TP = FNMS(KP1_118033988, TO, TN); | |
118 Tb = Cr[WS(csr, 3)]; | |
119 { | |
120 E Tv, Tw, Ts, Tt, Tc; | |
121 Ts = Ci[WS(csi, 8)]; | |
122 Tt = Ci[WS(csi, 6)]; | |
123 T15 = Tb - Ta; | |
124 Tc = Ta + Tb; | |
125 Tv = Ci[WS(csi, 3)]; | |
126 TX = Tt + Ts; | |
127 Tu = Ts - Tt; | |
128 Tw = Ci[WS(csi, 1)]; | |
129 Td = T9 + Tc; | |
130 TV = Tc - T9; | |
131 Tx = Tv - Tw; | |
132 TY = Tw + Tv; | |
133 } | |
134 } | |
135 Te = T6 + Td; | |
136 TU = FMS(KP250000000, Td, T6); | |
137 T1U = FNMS(KP618033988, T14, T15); | |
138 T16 = FMA(KP618033988, T15, T14); | |
139 T12 = Tx - Tu; | |
140 Ty = Tu + Tx; | |
141 T1R = FNMS(KP618033988, TX, TY); | |
142 TZ = FMA(KP618033988, TY, TX); | |
143 TA = Ty - Tz; | |
144 T11 = FMA(KP250000000, Ty, Tz); | |
145 } | |
146 { | |
147 E Tf, TI, T1j, T1k, Tm, T1c, TD, TG, T1d, TH; | |
148 Tf = Cr[WS(csr, 10)]; | |
149 TI = Ci[WS(csi, 10)]; | |
150 { | |
151 E T13, T1T, TW, T1Q; | |
152 T13 = FMA(KP559016994, T12, T11); | |
153 T1T = FNMS(KP559016994, T12, T11); | |
154 TW = FMA(KP559016994, TV, TU); | |
155 T1Q = FNMS(KP559016994, TV, TU); | |
156 T2i = FMA(KP951056516, T1U, T1T); | |
157 T1V = FNMS(KP951056516, T1U, T1T); | |
158 T17 = FMA(KP951056516, T16, T13); | |
159 T1B = FNMS(KP951056516, T16, T13); | |
160 T2h = FNMS(KP951056516, T1R, T1Q); | |
161 T1S = FMA(KP951056516, T1R, T1Q); | |
162 T10 = FNMS(KP951056516, TZ, TW); | |
163 T1C = FMA(KP951056516, TZ, TW); | |
164 { | |
165 E Tg, Th, Tj, Tk; | |
166 Tg = Cr[WS(csr, 5)]; | |
167 Th = Cr[WS(csr, 9)]; | |
168 Tj = Cr[0]; | |
169 Tk = Cr[WS(csr, 4)]; | |
170 { | |
171 E TB, Ti, Tl, TC, TE, TF; | |
172 TB = Ci[WS(csi, 9)]; | |
173 T1j = Tg - Th; | |
174 Ti = Tg + Th; | |
175 T1k = Tk - Tj; | |
176 Tl = Tj + Tk; | |
177 TC = Ci[WS(csi, 5)]; | |
178 TE = Ci[WS(csi, 4)]; | |
179 TF = Ci[0]; | |
180 Tm = Ti + Tl; | |
181 T1a = Ti - Tl; | |
182 T1c = TC + TB; | |
183 TD = TB - TC; | |
184 TG = TE - TF; | |
185 T1d = TF + TE; | |
186 } | |
187 } | |
188 } | |
189 T19 = FMS(KP250000000, Tm, Tf); | |
190 Tn = Tf + Tm; | |
191 T1h = TD - TG; | |
192 TH = TD + TG; | |
193 T1l = FNMS(KP618033988, T1k, T1j); | |
194 T1Y = FMA(KP618033988, T1j, T1k); | |
195 T1e = FMA(KP618033988, T1d, T1c); | |
196 T21 = FNMS(KP618033988, T1c, T1d); | |
197 TJ = TH - TI; | |
198 T1g = FMA(KP250000000, TH, TI); | |
199 } | |
200 } | |
201 } | |
202 { | |
203 E T1Z, T1m, T1y, T22, T1f, T1z, T2j, T2g, T2d, T2q, T2s; | |
204 { | |
205 E Tq, To, T2e, T2f; | |
206 Tq = Tn - Te; | |
207 To = Te + Tn; | |
208 { | |
209 E T1i, T1X, T1b, T20; | |
210 T1i = FNMS(KP559016994, T1h, T1g); | |
211 T1X = FMA(KP559016994, T1h, T1g); | |
212 T1b = FNMS(KP559016994, T1a, T19); | |
213 T20 = FMA(KP559016994, T1a, T19); | |
214 T2e = FMA(KP951056516, T1Y, T1X); | |
215 T1Z = FNMS(KP951056516, T1Y, T1X); | |
216 T1m = FNMS(KP951056516, T1l, T1i); | |
217 T1y = FMA(KP951056516, T1l, T1i); | |
218 T2f = FNMS(KP951056516, T21, T20); | |
219 T22 = FMA(KP951056516, T21, T20); | |
220 T1f = FNMS(KP951056516, T1e, T1b); | |
221 T1z = FMA(KP951056516, T1e, T1b); | |
222 } | |
223 { | |
224 E T2o, TK, TM, T2p, Tr, TL, Tp; | |
225 T2o = FMA(KP939062505, T2h, T2i); | |
226 T2j = FNMS(KP939062505, T2i, T2h); | |
227 R0[0] = FMA(KP2_000000000, To, T5); | |
228 Tp = FNMS(KP500000000, To, T5); | |
229 TK = FMA(KP618033988, TJ, TA); | |
230 TM = FNMS(KP618033988, TA, TJ); | |
231 T2g = FNMS(KP062914667, T2f, T2e); | |
232 T2p = FMA(KP062914667, T2e, T2f); | |
233 Tr = FNMS(KP1_118033988, Tq, Tp); | |
234 TL = FMA(KP1_118033988, Tq, Tp); | |
235 T2d = FMA(KP1_902113032, T1O, T1N); | |
236 T1P = FNMS(KP1_902113032, T1O, T1N); | |
237 T2q = FMA(KP846146756, T2p, T2o); | |
238 T2s = FNMS(KP451418159, T2o, T2p); | |
239 R0[WS(rs, 10)] = FMA(KP1_902113032, TK, Tr); | |
240 R1[WS(rs, 2)] = FMS(KP1_902113032, TK, Tr); | |
241 R1[WS(rs, 7)] = FMS(KP1_902113032, TM, TL); | |
242 R0[WS(rs, 5)] = FMA(KP1_902113032, TM, TL); | |
243 } | |
244 } | |
245 { | |
246 E T18, T1n, T1x, TT, T2m, T1w, T1u, T2l, T1s, T1t, T2k; | |
247 T18 = FNMS(KP256756360, T17, T10); | |
248 T1s = FMA(KP256756360, T10, T17); | |
249 T1t = FMA(KP549754652, T1f, T1m); | |
250 T1n = FNMS(KP549754652, T1m, T1f); | |
251 T1x = FNMS(KP1_902113032, TS, TP); | |
252 TT = FMA(KP1_902113032, TS, TP); | |
253 T2m = FMA(KP730409924, T2j, T2g); | |
254 T2k = FNMS(KP730409924, T2j, T2g); | |
255 T1w = FNMS(KP683113946, T1s, T1t); | |
256 T1u = FMA(KP559154169, T1t, T1s); | |
257 R1[WS(rs, 1)] = -(FMA(KP1_996053456, T2k, T2d)); | |
258 T2l = FNMS(KP499013364, T2k, T2d); | |
259 { | |
260 E T1K, T1M, T1G, T1E; | |
261 { | |
262 E T1D, T1A, T1q, T1p, T1v, T1r; | |
263 { | |
264 E T1I, T1J, T2n, T2r, T1o; | |
265 T1I = FMA(KP634619297, T1B, T1C); | |
266 T1D = FNMS(KP634619297, T1C, T1B); | |
267 T1A = FMA(KP470564281, T1z, T1y); | |
268 T1J = FNMS(KP470564281, T1y, T1z); | |
269 T2n = FNMS(KP1_115827804, T2m, T2l); | |
270 T2r = FMA(KP1_115827804, T2m, T2l); | |
271 T1q = FNMS(KP904730450, T1n, T18); | |
272 T1o = FMA(KP904730450, T1n, T18); | |
273 R1[WS(rs, 11)] = FMS(KP1_386580726, T2q, T2n); | |
274 R0[WS(rs, 4)] = FMA(KP1_386580726, T2q, T2n); | |
275 R0[WS(rs, 9)] = FMA(KP1_898359647, T2s, T2r); | |
276 R1[WS(rs, 6)] = FMS(KP1_898359647, T2s, T2r); | |
277 R1[0] = FMS(KP1_937166322, T1o, TT); | |
278 T1p = FMA(KP484291580, T1o, TT); | |
279 T1K = FMA(KP662318342, T1J, T1I); | |
280 T1M = FNMS(KP576710603, T1I, T1J); | |
281 } | |
282 T1v = FMA(KP1_082908895, T1q, T1p); | |
283 T1r = FNMS(KP1_082908895, T1q, T1p); | |
284 R1[WS(rs, 10)] = FMS(KP1_842354653, T1u, T1r); | |
285 R0[WS(rs, 3)] = FMA(KP1_842354653, T1u, T1r); | |
286 R0[WS(rs, 8)] = FMA(KP1_666834356, T1w, T1v); | |
287 R1[WS(rs, 5)] = FMS(KP1_666834356, T1w, T1v); | |
288 T1G = FNMS(KP933137358, T1D, T1A); | |
289 T1E = FMA(KP933137358, T1D, T1A); | |
290 } | |
291 { | |
292 E T23, T28, T29, T1W, T1F, T1H, T1L; | |
293 T23 = FNMS(KP634619297, T22, T1Z); | |
294 T28 = FMA(KP634619297, T1Z, T22); | |
295 T29 = FMA(KP549754652, T1S, T1V); | |
296 T1W = FNMS(KP549754652, T1V, T1S); | |
297 R0[WS(rs, 2)] = FMA(KP1_809654104, T1E, T1x); | |
298 T1F = FNMS(KP452413526, T1E, T1x); | |
299 T2c = FMA(KP595480289, T28, T29); | |
300 T2a = FNMS(KP641441904, T29, T28); | |
301 T1H = FNMS(KP1_011627398, T1G, T1F); | |
302 T1L = FMA(KP1_011627398, T1G, T1F); | |
303 R0[WS(rs, 12)] = FNMS(KP1_606007150, T1K, T1H); | |
304 R1[WS(rs, 4)] = -(FMA(KP1_606007150, T1K, T1H)); | |
305 R1[WS(rs, 9)] = -(FMA(KP1_721083328, T1M, T1L)); | |
306 R0[WS(rs, 7)] = FNMS(KP1_721083328, T1M, T1L); | |
307 T24 = FNMS(KP963507348, T23, T1W); | |
308 T26 = FMA(KP963507348, T23, T1W); | |
309 } | |
310 } | |
311 } | |
312 } | |
313 } | |
314 R0[WS(rs, 1)] = FNMS(KP1_752613360, T24, T1P); | |
315 T25 = FMA(KP438153340, T24, T1P); | |
316 T27 = FMA(KP979740652, T26, T25); | |
317 T2b = FNMS(KP979740652, T26, T25); | |
318 R1[WS(rs, 8)] = -(FMA(KP1_606007150, T2a, T27)); | |
319 R0[WS(rs, 6)] = FNMS(KP1_606007150, T2a, T27); | |
320 R1[WS(rs, 3)] = -(FMA(KP1_666834356, T2c, T2b)); | |
321 R0[WS(rs, 11)] = FNMS(KP1_666834356, T2c, T2b); | |
322 } | |
323 } | |
324 } | |
325 | |
326 static const kr2c_desc desc = { 25, "r2cbIII_25", {32, 0, 120, 0}, &GENUS }; | |
327 | |
328 void X(codelet_r2cbIII_25) (planner *p) { | |
329 X(kr2c_register) (p, r2cbIII_25, &desc); | |
330 } | |
331 | |
332 #else /* HAVE_FMA */ | |
333 | |
334 /* Generated by: ../../../genfft/gen_r2cb.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 25 -name r2cbIII_25 -dft-III -include r2cbIII.h */ | |
335 | |
336 /* | |
337 * This function contains 152 FP additions, 98 FP multiplications, | |
338 * (or, 100 additions, 46 multiplications, 52 fused multiply/add), | |
339 * 65 stack variables, 21 constants, and 50 memory accesses | |
340 */ | |
341 #include "r2cbIII.h" | |
342 | |
343 static void r2cbIII_25(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs) | |
344 { | |
345 DK(KP968583161, +0.968583161128631119490168375464735813836012403); | |
346 DK(KP248689887, +0.248689887164854788242283746006447968417567406); | |
347 DK(KP684547105, +0.684547105928688673732283357621209269889519233); | |
348 DK(KP728968627, +0.728968627421411523146730319055259111372571664); | |
349 DK(KP062790519, +0.062790519529313376076178224565631133122484832); | |
350 DK(KP998026728, +0.998026728428271561952336806863450553336905220); | |
351 DK(KP876306680, +0.876306680043863587308115903922062583399064238); | |
352 DK(KP481753674, +0.481753674101715274987191502872129653528542010); | |
353 DK(KP535826794, +0.535826794978996618271308767867639978063575346); | |
354 DK(KP844327925, +0.844327925502015078548558063966681505381659241); | |
355 DK(KP904827052, +0.904827052466019527713668647932697593970413911); | |
356 DK(KP425779291, +0.425779291565072648862502445744251703979973042); | |
357 DK(KP250000000, +0.250000000000000000000000000000000000000000000); | |
358 DK(KP951056516, +0.951056516295153572116439333379382143405698634); | |
359 DK(KP587785252, +0.587785252292473129168705954639072768597652438); | |
360 DK(KP559016994, +0.559016994374947424102293417182819058860154590); | |
361 DK(KP500000000, +0.500000000000000000000000000000000000000000000); | |
362 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000); | |
363 DK(KP1_118033988, +1.118033988749894848204586834365638117720309180); | |
364 DK(KP1_175570504, +1.175570504584946258337411909278145537195304875); | |
365 DK(KP1_902113032, +1.902113032590307144232878666758764286811397268); | |
366 { | |
367 INT i; | |
368 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)) { | |
369 E TS, T1O, T5, TP, T1N, TI, TH, Te, T17, T2h, T1y, T1V, T10, T2g, T1x; | |
370 E T1S, Tz, Ty, Tn, T1m, T2e, T1B, T22, T1f, T2d, T1A, T1Z, TQ, TR; | |
371 TQ = Ci[WS(csi, 2)]; | |
372 TR = Ci[WS(csi, 7)]; | |
373 TS = FNMS(KP1_175570504, TR, KP1_902113032 * TQ); | |
374 T1O = FMA(KP1_902113032, TR, KP1_175570504 * TQ); | |
375 { | |
376 E T1, T4, TN, T2, T3, TO; | |
377 T1 = Cr[WS(csr, 12)]; | |
378 T2 = Cr[WS(csr, 7)]; | |
379 T3 = Cr[WS(csr, 2)]; | |
380 T4 = T2 + T3; | |
381 TN = KP1_118033988 * (T3 - T2); | |
382 T5 = FMA(KP2_000000000, T4, T1); | |
383 TO = FMS(KP500000000, T4, T1); | |
384 TP = TN - TO; | |
385 T1N = TO + TN; | |
386 } | |
387 { | |
388 E T6, Td, T15, TU, T14, T11, TX, TY; | |
389 T6 = Cr[WS(csr, 11)]; | |
390 TI = Ci[WS(csi, 11)]; | |
391 { | |
392 E T7, T8, T9, Ta, Tb, Tc; | |
393 T7 = Cr[WS(csr, 6)]; | |
394 T8 = Cr[WS(csr, 8)]; | |
395 T9 = T7 + T8; | |
396 Ta = Cr[WS(csr, 1)]; | |
397 Tb = Cr[WS(csr, 3)]; | |
398 Tc = Ta + Tb; | |
399 Td = T9 + Tc; | |
400 T15 = Ta - Tb; | |
401 TU = KP559016994 * (Tc - T9); | |
402 T14 = T8 - T7; | |
403 } | |
404 { | |
405 E TB, TC, TD, TE, TF, TG; | |
406 TB = Ci[WS(csi, 6)]; | |
407 TC = Ci[WS(csi, 8)]; | |
408 TD = TB - TC; | |
409 TE = Ci[WS(csi, 1)]; | |
410 TF = Ci[WS(csi, 3)]; | |
411 TG = TE - TF; | |
412 TH = TD + TG; | |
413 T11 = KP559016994 * (TD - TG); | |
414 TX = TB + TC; | |
415 TY = TE + TF; | |
416 } | |
417 Te = T6 + Td; | |
418 { | |
419 E T16, T1T, T13, T1U, T12; | |
420 T16 = FMA(KP587785252, T14, KP951056516 * T15); | |
421 T1T = FNMS(KP587785252, T15, KP951056516 * T14); | |
422 T12 = FNMS(KP250000000, TH, TI); | |
423 T13 = T11 - T12; | |
424 T1U = T11 + T12; | |
425 T17 = T13 - T16; | |
426 T2h = T1T - T1U; | |
427 T1y = T16 + T13; | |
428 T1V = T1T + T1U; | |
429 } | |
430 { | |
431 E TZ, T1R, TW, T1Q, TV; | |
432 TZ = FNMS(KP951056516, TY, KP587785252 * TX); | |
433 T1R = FMA(KP951056516, TX, KP587785252 * TY); | |
434 TV = FMS(KP250000000, Td, T6); | |
435 TW = TU - TV; | |
436 T1Q = TV + TU; | |
437 T10 = TW + TZ; | |
438 T2g = T1Q + T1R; | |
439 T1x = TZ - TW; | |
440 T1S = T1Q - T1R; | |
441 } | |
442 } | |
443 { | |
444 E Tf, Tm, T1k, T19, T1j, T1g, T1c, T1d; | |
445 Tf = Cr[WS(csr, 10)]; | |
446 Tz = Ci[WS(csi, 10)]; | |
447 { | |
448 E Tg, Th, Ti, Tj, Tk, Tl; | |
449 Tg = Cr[WS(csr, 5)]; | |
450 Th = Cr[WS(csr, 9)]; | |
451 Ti = Tg + Th; | |
452 Tj = Cr[0]; | |
453 Tk = Cr[WS(csr, 4)]; | |
454 Tl = Tj + Tk; | |
455 Tm = Ti + Tl; | |
456 T1k = Tj - Tk; | |
457 T19 = KP559016994 * (Tl - Ti); | |
458 T1j = Th - Tg; | |
459 } | |
460 { | |
461 E Ts, Tt, Tu, Tv, Tw, Tx; | |
462 Ts = Ci[WS(csi, 4)]; | |
463 Tt = Ci[0]; | |
464 Tu = Ts - Tt; | |
465 Tv = Ci[WS(csi, 5)]; | |
466 Tw = Ci[WS(csi, 9)]; | |
467 Tx = Tv - Tw; | |
468 Ty = Tu - Tx; | |
469 T1g = KP559016994 * (Tx + Tu); | |
470 T1c = Tv + Tw; | |
471 T1d = Tt + Ts; | |
472 } | |
473 Tn = Tf + Tm; | |
474 { | |
475 E T1l, T20, T1i, T21, T1h; | |
476 T1l = FMA(KP587785252, T1j, KP951056516 * T1k); | |
477 T20 = FNMS(KP587785252, T1k, KP951056516 * T1j); | |
478 T1h = FMA(KP250000000, Ty, Tz); | |
479 T1i = T1g - T1h; | |
480 T21 = T1g + T1h; | |
481 T1m = T1i - T1l; | |
482 T2e = T21 - T20; | |
483 T1B = T1l + T1i; | |
484 T22 = T20 + T21; | |
485 } | |
486 { | |
487 E T1e, T1Y, T1b, T1X, T1a; | |
488 T1e = FNMS(KP951056516, T1d, KP587785252 * T1c); | |
489 T1Y = FMA(KP951056516, T1c, KP587785252 * T1d); | |
490 T1a = FMS(KP250000000, Tm, Tf); | |
491 T1b = T19 - T1a; | |
492 T1X = T1a + T19; | |
493 T1f = T1b + T1e; | |
494 T2d = T1X + T1Y; | |
495 T1A = T1e - T1b; | |
496 T1Z = T1X - T1Y; | |
497 } | |
498 } | |
499 { | |
500 E Tq, To, Tp, TK, TM, TA, TJ, TL, Tr; | |
501 Tq = KP1_118033988 * (Tn - Te); | |
502 To = Te + Tn; | |
503 Tp = FMS(KP500000000, To, T5); | |
504 TA = Ty - Tz; | |
505 TJ = TH + TI; | |
506 TK = FNMS(KP1_902113032, TJ, KP1_175570504 * TA); | |
507 TM = FMA(KP1_175570504, TJ, KP1_902113032 * TA); | |
508 R0[0] = FMA(KP2_000000000, To, T5); | |
509 TL = Tq - Tp; | |
510 R0[WS(rs, 5)] = TL + TM; | |
511 R1[WS(rs, 7)] = TM - TL; | |
512 Tr = Tp + Tq; | |
513 R1[WS(rs, 2)] = Tr + TK; | |
514 R0[WS(rs, 10)] = TK - Tr; | |
515 } | |
516 { | |
517 E T2q, T2s, T2k, T2j, T2l, T2m, T2r, T2n; | |
518 { | |
519 E T2o, T2p, T2f, T2i; | |
520 T2o = FNMS(KP904827052, T2d, KP425779291 * T2e); | |
521 T2p = FNMS(KP535826794, T2h, KP844327925 * T2g); | |
522 T2q = FNMS(KP1_902113032, T2p, KP1_175570504 * T2o); | |
523 T2s = FMA(KP1_175570504, T2p, KP1_902113032 * T2o); | |
524 T2k = T1N + T1O; | |
525 T2f = FMA(KP425779291, T2d, KP904827052 * T2e); | |
526 T2i = FMA(KP535826794, T2g, KP844327925 * T2h); | |
527 T2j = T2f - T2i; | |
528 T2l = FMA(KP500000000, T2j, T2k); | |
529 T2m = KP1_118033988 * (T2i + T2f); | |
530 } | |
531 R0[WS(rs, 2)] = FMS(KP2_000000000, T2j, T2k); | |
532 T2r = T2m - T2l; | |
533 R0[WS(rs, 7)] = T2r + T2s; | |
534 R1[WS(rs, 9)] = T2s - T2r; | |
535 T2n = T2l + T2m; | |
536 R1[WS(rs, 4)] = T2n + T2q; | |
537 R0[WS(rs, 12)] = T2q - T2n; | |
538 } | |
539 { | |
540 E T1u, T1w, TT, T1o, T1p, T1q, T1v, T1r; | |
541 { | |
542 E T1s, T1t, T18, T1n; | |
543 T1s = FMA(KP481753674, T10, KP876306680 * T17); | |
544 T1t = FMA(KP844327925, T1f, KP535826794 * T1m); | |
545 T1u = FMA(KP1_902113032, T1s, KP1_175570504 * T1t); | |
546 T1w = FNMS(KP1_175570504, T1s, KP1_902113032 * T1t); | |
547 TT = TP - TS; | |
548 T18 = FNMS(KP481753674, T17, KP876306680 * T10); | |
549 T1n = FNMS(KP844327925, T1m, KP535826794 * T1f); | |
550 T1o = T18 + T1n; | |
551 T1p = FMS(KP500000000, T1o, TT); | |
552 T1q = KP1_118033988 * (T1n - T18); | |
553 } | |
554 R0[WS(rs, 1)] = FMA(KP2_000000000, T1o, TT); | |
555 T1v = T1q - T1p; | |
556 R0[WS(rs, 6)] = T1v + T1w; | |
557 R1[WS(rs, 8)] = T1w - T1v; | |
558 T1r = T1p + T1q; | |
559 R1[WS(rs, 3)] = T1r + T1u; | |
560 R0[WS(rs, 11)] = T1u - T1r; | |
561 } | |
562 { | |
563 E T1H, T1L, T1E, T1D, T1I, T1J, T1M, T1K; | |
564 { | |
565 E T1F, T1G, T1z, T1C; | |
566 T1F = FNMS(KP062790519, T1B, KP998026728 * T1A); | |
567 T1G = FNMS(KP684547105, T1x, KP728968627 * T1y); | |
568 T1H = FNMS(KP1_902113032, T1G, KP1_175570504 * T1F); | |
569 T1L = FMA(KP1_175570504, T1G, KP1_902113032 * T1F); | |
570 T1E = TP + TS; | |
571 T1z = FMA(KP728968627, T1x, KP684547105 * T1y); | |
572 T1C = FMA(KP062790519, T1A, KP998026728 * T1B); | |
573 T1D = T1z + T1C; | |
574 T1I = FMA(KP500000000, T1D, T1E); | |
575 T1J = KP1_118033988 * (T1C - T1z); | |
576 } | |
577 R1[WS(rs, 1)] = FMS(KP2_000000000, T1D, T1E); | |
578 T1M = T1J - T1I; | |
579 R0[WS(rs, 9)] = T1L - T1M; | |
580 R1[WS(rs, 6)] = T1L + T1M; | |
581 T1K = T1I + T1J; | |
582 R1[WS(rs, 11)] = T1H - T1K; | |
583 R0[WS(rs, 4)] = T1H + T1K; | |
584 } | |
585 { | |
586 E T2a, T2c, T1P, T24, T25, T26, T2b, T27; | |
587 { | |
588 E T28, T29, T1W, T23; | |
589 T28 = FMA(KP248689887, T1S, KP968583161 * T1V); | |
590 T29 = FMA(KP481753674, T1Z, KP876306680 * T22); | |
591 T2a = FMA(KP1_902113032, T28, KP1_175570504 * T29); | |
592 T2c = FNMS(KP1_175570504, T28, KP1_902113032 * T29); | |
593 T1P = T1N - T1O; | |
594 T1W = FNMS(KP248689887, T1V, KP968583161 * T1S); | |
595 T23 = FNMS(KP481753674, T22, KP876306680 * T1Z); | |
596 T24 = T1W + T23; | |
597 T25 = FMS(KP500000000, T24, T1P); | |
598 T26 = KP1_118033988 * (T23 - T1W); | |
599 } | |
600 R1[0] = FMA(KP2_000000000, T24, T1P); | |
601 T2b = T26 - T25; | |
602 R1[WS(rs, 5)] = T2b + T2c; | |
603 R0[WS(rs, 8)] = T2c - T2b; | |
604 T27 = T25 + T26; | |
605 R0[WS(rs, 3)] = T27 + T2a; | |
606 R1[WS(rs, 10)] = T2a - T27; | |
607 } | |
608 } | |
609 } | |
610 } | |
611 | |
612 static const kr2c_desc desc = { 25, "r2cbIII_25", {100, 46, 52, 0}, &GENUS }; | |
613 | |
614 void X(codelet_r2cbIII_25) (planner *p) { | |
615 X(kr2c_register) (p, r2cbIII_25, &desc); | |
616 } | |
617 | |
618 #endif /* HAVE_FMA */ |