Logo ROOT  
Reference Guide
 
Loading...
Searching...
No Matches
RooHeterogeneousMath.h
Go to the documentation of this file.
1/*
2 * Project: RooFit
3 *
4 * Copyright (c) 2023, CERN
5 *
6 * Redistribution and use in source and binary forms,
7 * with or without modification, are permitted according to the terms
8 * listed in LICENSE (http://roofit.sourceforge.net/license.txt)
9 */
10
11#ifndef ROOFIT_BATCHCOMPUTE_ROOHETEROGENEOUSMATH_H
12#define ROOFIT_BATCHCOMPUTE_ROOHETEROGENEOUSMATH_H
13
15#include <RooBatchCompute.h>
16
17#include <algorithm>
18#include <cmath>
19#include <complex>
20#include <iostream>
21
22#if defined(__CUDACC__)
23#include <cuda/std/complex>
24#endif
25
27
28// The C++ std::complex type operators don't work on the GPU (silently gives
29// wrong results). But if we use the cuda::std:: namespace, all the math
30// operations work again.
31#if defined(__CUDACC__)
32namespace STD = cuda::std;
33#else
34namespace STD = std;
35#endif
36
37__roodevice__ __roohost__ static inline void cexp(double &re, double &im)
38{
39 // with gcc on unix machines and on x86_64, we can gain by hand-coding
40 // exp(z) for the x87 coprocessor; other platforms have the default
41 // routines as fallback implementation, and compilers other than gcc on
42 // x86_64 generate better code with the default routines; also avoid
43 // the inline assembly code when the compiler is not optimising code, or
44 // is optimising for code size
45 // (we insist on __unix__ here, since the assemblers on other OSs
46 // typically do not speak AT&T syntax as gas does...)
47#if defined(__CUDACC__) || !(defined(__GNUC__) || defined(__clang__)) || !defined(__unix__) || !defined(__x86_64__) || \
48 !defined(__OPTIMIZE__) || defined(__OPTIMIZE_SIZE__) || defined(__INTEL_COMPILER) || defined(__OPEN64__) || \
49 defined(__PATHSCALE__)
50 const double e = std::exp(re);
51 re = e * std::cos(im);
52 im = e * std::sin(im);
53#else
54 // clang-format off
55 __asm__ (
56 "fxam\n\t" // examine st(0): NaN? Inf?
57 "fstsw %%ax\n\t"
58 "movb $0x45,%%dh\n\t"
59 "andb %%ah,%%dh\n\t"
60 "cmpb $0x05,%%dh\n\t"
61 "jz 1f\n\t" // have NaN or infinity, handle below
62 "fldl2e\n\t" // load log2(e)
63 "fmulp\n\t" // re * log2(e)
64 "fld %%st(0)\n\t" // duplicate re * log2(e)
65 "frndint\n\t" // int(re * log2(e))
66 "fsubr %%st,%%st(1)\n\t" // st(1) = x = frac(re * log2(e))
67 "fxch\n\t" // swap st(0), st(1)
68 "f2xm1\n\t" // 2^x - 1
69 "fld1\n\t" // st(0) = 1
70 "faddp\n\t" // st(0) = 2^x
71 "fscale\n\t" // 2 ^ (int(re * log2(e)) + x)
72 "fstp %%st(1)\n\t" // pop st(1)
73 "jmp 2f\n\t"
74 "1:\n\t" // handle NaN, Inf...
75 "testl $0x200, %%eax\n\t"// -infinity?
76 "jz 2f\n\t"
77 "fstp %%st\n\t" // -Inf, so pop st(0)
78 "fldz\n\t" // st(0) = 0
79 "2:\n\t" // here. we have st(0) == exp(re)
80 "fxch\n\t" // st(0) = im, st(1) = exp(re)
81 "fsincos\n\t" // st(0) = cos(im), st(1) = sin(im)
82 "fnstsw %%ax\n\t"
83 "testl $0x400,%%eax\n\t"
84 "jz 4f\n\t" // |im| too large for fsincos?
85 "fldpi\n\t" // st(0) = pi
86 "fadd %%st(0)\n\t" // st(0) *= 2;
87 "fxch %%st(1)\n\t" // st(0) = im, st(1) = 2 * pi
88 "3:\n\t"
89 "fprem1\n\t" // st(0) = fmod(im, 2 * pi)
90 "fnstsw %%ax\n\t"
91 "testl $0x400,%%eax\n\t"
92 "jnz 3b\n\t" // fmod done?
93 "fstp %%st(1)\n\t" // yes, pop st(1) == 2 * pi
94 "fsincos\n\t" // st(0) = cos(im), st(1) = sin(im)
95 "4:\n\t" // all fine, fsincos succeeded
96 "fmul %%st(2)\n\t" // st(0) *= st(2)
97 "fxch %%st(2)\n\t" // st(2)=exp(re)*cos(im),st(0)=exp(im)
98 "fmulp %%st(1)\n\t" // st(1)=exp(re)*sin(im), pop st(0)
99 : "=t" (im), "=u" (re): "0" (re), "1" (im) :
100 "eax", "dh", "cc"
101#ifndef __clang__
102 // normal compilers (like gcc) want to be told that we
103 // clobber x87 registers, even if we pop them afterwards
104 // (so they can make sure they don't save anything there)
105 , "st(5)", "st(6)", "st(7)"
106#else // __clang__
107 // clang produces an error message with the clobber list
108 // from above - not sure why; it seems harmless to leave
109 // the popped x87 registers out of the clobber list for
110 // clang, and that is in fact what seems to be recommended
111 // here:
112 // http://lists.cs.uiuc.edu/pipermail/cfe-dev/2012-May/021715.html
113#endif // __clang__
114 );
115 // clang-format on
116#endif
117}
118
119template <class T, unsigned N, unsigned NTAYLOR, unsigned NCF>
120__roodevice__ __roohost__ static inline STD::complex<T>
121faddeeva_smabmq_impl(T zre, T zim, const T tm, const T (&a)[N], const T (&npi)[N],
122 const T (&taylorarr)[N * NTAYLOR * 2])
123{
124 // catch singularities in the Fourier representation At
125 // z = n pi / tm, and provide a Taylor series expansion in those
126 // points, and only use it when we're close enough to the real axis
127 // that there is a chance we need it
128 const T zim2 = zim * zim;
129 const T maxnorm = T(9) / T(1000000);
130 if (zim2 < maxnorm) {
131 // we're close enough to the real axis that we need to worry about
132 // singularities
133 const T dnsing = tm * zre / npi[1];
134 const T dnsingmax2 = (T(N) - T(1) / T(2)) * (T(N) - T(1) / T(2));
135 if (dnsing * dnsing < dnsingmax2) {
136 // we're in the interesting range of the real axis as well...
137 // deal with Re(z) < 0 so we only need N different Taylor
138 // expansions; use w(-x+iy) = conj(w(x+iy))
139 const bool negrez = zre < T(0);
140 // figure out closest singularity
141 const int nsing = int(std::abs(dnsing) + T(1) / T(2));
142 // and calculate just how far we are from it
143 const T zmnpire = std::abs(zre) - npi[nsing];
144 const T zmnpinorm = zmnpire * zmnpire + zim2;
145 // close enough to one of the singularities?
146 if (zmnpinorm < maxnorm) {
147 const T *coeffs = &taylorarr[nsing * NTAYLOR * 2];
148 // calculate value of taylor expansion...
149 // (note: there's no chance to vectorize this one, since
150 // the value of the next iteration depend on the ones from
151 // the previous iteration)
152 T sumre = coeffs[0];
153 T sumim = coeffs[1];
154 for (unsigned i = 1; i < NTAYLOR; ++i) {
155 const T re = sumre * zmnpire - sumim * zim;
156 const T im = sumim * zmnpire + sumre * zim;
157 sumre = re + coeffs[2 * i + 0];
158 sumim = im + coeffs[2 * i + 1];
159 }
160 // undo the flip in real part of z if needed
161 if (negrez) {
162 return STD::complex<T>(sumre, -sumim);
163 } else {
164 return STD::complex<T>(sumre, sumim);
165 }
166 }
167 }
168 }
169 // negative Im(z) is treated by calculating for -z, and using the
170 // symmetry properties of erfc(z)
171 const bool negimz = zim < T(0);
172 if (negimz) {
173 zre = -zre;
174 zim = -zim;
175 }
176 const T znorm = zre * zre + zim2;
177 if (znorm > tm * tm) {
178 // use continued fraction approximation for |z| large
179 const T isqrtpi = 5.64189583547756287e-01;
180 const T z2re = (zre + zim) * (zre - zim);
181 const T z2im = T(2) * zre * zim;
182 T cfre = T(1);
183 T cfim = T(0);
184 T cfnorm = T(1);
185 for (unsigned k = NCF; k; --k) {
186 cfre = +(T(k) / T(2)) * cfre / cfnorm;
187 cfim = -(T(k) / T(2)) * cfim / cfnorm;
188 if (k & 1) {
189 cfre -= z2re, cfim -= z2im;
190 } else {
191 cfre += T(1);
192 }
193 cfnorm = cfre * cfre + cfim * cfim;
194 }
195 T sumre = (zim * cfre - zre * cfim) * isqrtpi / cfnorm;
196 T sumim = -(zre * cfre + zim * cfim) * isqrtpi / cfnorm;
197 if (negimz) {
198 // use erfc(-z) = 2 - erfc(z) to get good accuracy for
199 // Im(z) < 0: 2 / exp(z^2) - w(z)
200 T ez2re = -z2re;
201 T ez2im = -z2im;
202 RooHeterogeneousMath::cexp(ez2re, ez2im);
203 return STD::complex<T>(T(2) * ez2re - sumre, T(2) * ez2im - sumim);
204 } else {
205 return STD::complex<T>(sumre, sumim);
206 }
207 }
208 const T twosqrtpi = 3.54490770181103205e+00;
209 const T tmzre = tm * zre;
210 const T tmzim = tm * zim;
211 // calculate exp(i tm z)
212 T eitmzre = -tmzim;
213 T eitmzim = tmzre;
214 RooHeterogeneousMath::cexp(eitmzre, eitmzim);
215 // form 1 +/- exp (i tm z)
216 const T numerarr[4] = {T(1) - eitmzre, -eitmzim, T(1) + eitmzre, +eitmzim};
217 // form tm z * (1 +/- exp(i tm z))
218 const T numertmz[4] = {tmzre * numerarr[0] - tmzim * numerarr[1], tmzre * numerarr[1] + tmzim * numerarr[0],
219 tmzre * numerarr[2] - tmzim * numerarr[3], tmzre * numerarr[3] + tmzim * numerarr[2]};
220 // common subexpressions for use inside the loop
221 const T reimtmzm2 = T(-2) * tmzre * tmzim;
222 const T imtmz2 = tmzim * tmzim;
223 const T reimtmzm22 = reimtmzm2 * reimtmzm2;
224 // on non-x86_64 architectures, when the compiler is producing
225 // unoptimised code and when optimising for code size, we use the
226 // straightforward implementation, but for x86_64, we use the
227 // brainf*cked code below that the gcc vectorizer likes to gain a few
228 // clock cycles; non-gcc compilers also get the normal code, since they
229 // usually do a better job with the default code (and yes, it's a pain
230 // that they're all pretending to be gcc)
231#if (defined(__CUDACC__) || !defined(__x86_64__)) || !defined(__OPTIMIZE__) || defined(__OPTIMIZE_SIZE__) || \
232 defined(__INTEL_COMPILER) || defined(__clang__) || defined(__OPEN64__) || defined(__PATHSCALE__) || \
233 !defined(__GNUC__)
234 T sumre = (-a[0] / znorm) * (numerarr[0] * zre + numerarr[1] * zim);
235 T sumim = (-a[0] / znorm) * (numerarr[1] * zre - numerarr[0] * zim);
236 for (unsigned i = 0; i < N; ++i) {
237 const unsigned j = (i << 1) & 2;
238 // denominator
239 const T wk = imtmz2 + (npi[i] + tmzre) * (npi[i] - tmzre);
240 // norm of denominator
241 const T norm = wk * wk + reimtmzm22;
242 const T f = T(2) * tm * a[i] / norm;
243 // sum += a[i] * numerator / wk
244 sumre -= f * (numertmz[j] * wk + numertmz[j + 1] * reimtmzm2);
245 sumim -= f * (numertmz[j + 1] * wk - numertmz[j] * reimtmzm2);
246 }
247#else
248 // BEGIN fully vectorisable code - enjoy reading... ;)
249 T tmp[2 * N];
250 for (unsigned i = 0; i < N; ++i) {
251 const T wk = imtmz2 + (npi[i] + tmzre) * (npi[i] - tmzre);
252 tmp[2 * i + 0] = wk;
253 tmp[2 * i + 1] = T(2) * tm * a[i] / (wk * wk + reimtmzm22);
254 }
255 for (unsigned i = 0; i < N / 2; ++i) {
256 T wk = tmp[4 * i + 0], f = tmp[4 * i + 1];
257 tmp[4 * i + 0] = -f * (numertmz[0] * wk + numertmz[1] * reimtmzm2);
258 tmp[4 * i + 1] = -f * (numertmz[1] * wk - numertmz[0] * reimtmzm2);
259 wk = tmp[4 * i + 2], f = tmp[4 * i + 3];
260 tmp[4 * i + 2] = -f * (numertmz[2] * wk + numertmz[3] * reimtmzm2);
261 tmp[4 * i + 3] = -f * (numertmz[3] * wk - numertmz[2] * reimtmzm2);
262 }
263 if (N & 1) {
264 // we may have missed one element in the last loop; if so, process
265 // it now...
266 const T wk = tmp[2 * N - 2], f = tmp[2 * N - 1];
267 tmp[2 * (N - 1) + 0] = -f * (numertmz[0] * wk + numertmz[1] * reimtmzm2);
268 tmp[2 * (N - 1) + 1] = -f * (numertmz[1] * wk - numertmz[0] * reimtmzm2);
269 }
270 T sumre = (-a[0] / znorm) * (numerarr[0] * zre + numerarr[1] * zim);
271 T sumim = (-a[0] / znorm) * (numerarr[1] * zre - numerarr[0] * zim);
272 for (unsigned i = 0; i < N; ++i) {
273 sumre += tmp[2 * i + 0];
274 sumim += tmp[2 * i + 1];
275 }
276 // END fully vectorisable code
277#endif
278 // prepare the result
279 if (negimz) {
280 // use erfc(-z) = 2 - erfc(z) to get good accuracy for
281 // Im(z) < 0: 2 / exp(z^2) - w(z)
282 const T z2im = -T(2) * zre * zim;
283 const T z2re = -(zre + zim) * (zre - zim);
284 T ez2re = z2re;
285 T ez2im = z2im;
286 RooHeterogeneousMath::cexp(ez2re, ez2im);
287 return STD::complex<T>(T(2) * ez2re + sumim / twosqrtpi, T(2) * ez2im - sumre / twosqrtpi);
288 } else {
289 return STD::complex<T>(-sumim / twosqrtpi, sumre / twosqrtpi);
290 }
291}
292
293// clang-format off
294__roodevice__ static const double npi24[24] = { // precomputed values n * pi
295 0.00000000000000000e+00, 3.14159265358979324e+00, 6.28318530717958648e+00,
296 9.42477796076937972e+00, 1.25663706143591730e+01, 1.57079632679489662e+01,
297 1.88495559215387594e+01, 2.19911485751285527e+01, 2.51327412287183459e+01,
298 2.82743338823081391e+01, 3.14159265358979324e+01, 3.45575191894877256e+01,
299 3.76991118430775189e+01, 4.08407044966673121e+01, 4.39822971502571053e+01,
300 4.71238898038468986e+01, 5.02654824574366918e+01, 5.34070751110264851e+01,
301 5.65486677646162783e+01, 5.96902604182060715e+01, 6.28318530717958648e+01,
302 6.59734457253856580e+01, 6.91150383789754512e+01, 7.22566310325652445e+01,
303 };
304__roodevice__ static const double a24[24] = { // precomputed Fourier coefficient prefactors
305 2.95408975150919338e-01, 2.75840233292177084e-01, 2.24573955224615866e-01,
306 1.59414938273911723e-01, 9.86657664154541891e-02, 5.32441407876394120e-02,
307 2.50521500053936484e-02, 1.02774656705395362e-02, 3.67616433284484706e-03,
308 1.14649364124223317e-03, 3.11757015046197600e-04, 7.39143342960301488e-05,
309 1.52794934280083635e-05, 2.75395660822107093e-06, 4.32785878190124505e-07,
310 5.93003040874588103e-08, 7.08449030774820423e-09, 7.37952063581678038e-10,
311 6.70217160600200763e-11, 5.30726516347079017e-12, 3.66432411346763916e-13,
312 2.20589494494103134e-14, 1.15782686262855879e-15, 5.29871142946730482e-17,
313 };
314__roodevice__ static const double taylorarr24[24 * 12] = {
315 // real part imaginary part, low order coefficients last
316 // nsing = 0
317 0.00000000000000000e-00, 3.00901111225470020e-01,
318 5.00000000000000000e-01, 0.00000000000000000e-00,
319 0.00000000000000000e-00, -7.52252778063675049e-01,
320 -1.00000000000000000e-00, 0.00000000000000000e-00,
321 0.00000000000000000e-00, 1.12837916709551257e+00,
322 1.00000000000000000e-00, 0.00000000000000000e-00,
323 // nsing = 1
324 -2.22423508493755319e-01, 1.87966717746229718e-01,
325 3.41805419240637628e-01, 3.42752593807919263e-01,
326 4.66574321730757753e-01, -5.59649213591058097e-01,
327 -8.05759710273191021e-01, -5.38989366115424093e-01,
328 -4.88914083733395200e-01, 9.80580906465856792e-01,
329 9.33757118080975970e-01, 2.82273885115127769e-01,
330 // nsing = 2
331 -2.60522586513312894e-01, -4.26259455096092786e-02,
332 1.36549702008863349e-03, 4.39243227763478846e-01,
333 6.50591493715480700e-01, -1.23422352472779046e-01,
334 -3.43379903564271318e-01, -8.13862662890748911e-01,
335 -7.96093943501906645e-01, 6.11271022503935772e-01,
336 7.60213717643090957e-01, 4.93801903948967945e-01,
337 // nsing = 3
338 -1.18249853727020186e-01, -1.90471659765411376e-01,
339 -2.59044664869706839e-01, 2.69333898502392004e-01,
340 4.99077838344125714e-01, 2.64644800189075006e-01,
341 1.26114512111568737e-01, -7.46519337025968199e-01,
342 -8.47666863706379907e-01, 1.89347715957263646e-01,
343 5.39641485816297176e-01, 5.97805988669631615e-01,
344 // nsing = 4
345 4.94825297066481491e-02, -1.71428212158876197e-01,
346 -2.97766677111471585e-01, 1.60773286596649656e-02,
347 1.88114210832460682e-01, 4.11734391195006462e-01,
348 3.98540613293909842e-01, -4.63321903522162715e-01,
349 -6.99522070542463639e-01, -1.32412024008354582e-01,
350 3.33997185986131785e-01, 6.01983450812696742e-01,
351 // nsing = 5
352 1.18367078448232332e-01, -6.09533063579086850e-02,
353 -1.74762998833038991e-01, -1.39098099222000187e-01,
354 -6.71534655984154549e-02, 3.34462251996496680e-01,
355 4.37429678577360024e-01, -1.59613865629038012e-01,
356 -4.71863911886034656e-01, -2.92759316465055762e-01,
357 1.80238737704018306e-01, 5.42834914744283253e-01,
358 // nsing = 6
359 8.87698096005701290e-02, 2.84339354980994902e-02,
360 -3.18943083830766399e-02, -1.53946887977045862e-01,
361 -1.71825061547624858e-01, 1.70734367410600348e-01,
362 3.33690792296469441e-01, 3.97048587678703930e-02,
363 -2.66422678503135697e-01, -3.18469797424381480e-01,
364 8.48049724711137773e-02, 4.60546329221462864e-01,
365 // nsing = 7
366 2.99767046276705077e-02, 5.34659695701718247e-02,
367 4.53131030251822568e-02, -9.37915401977138648e-02,
368 -1.57982359988083777e-01, 3.82170507060760740e-02,
369 1.98891589845251706e-01, 1.17546677047049354e-01,
370 -1.27514335237079297e-01, -2.72741112680307074e-01,
371 3.47906344595283767e-02, 3.82277517244493224e-01,
372 // nsing = 8
373 -7.35922494437203395e-03, 3.72011290318534610e-02,
374 5.66783220847204687e-02, -3.21015398169199501e-02,
375 -1.00308737825172555e-01, -2.57695148077963515e-02,
376 9.67294850588435368e-02, 1.18174625238337507e-01,
377 -5.21266530264988508e-02, -2.08850084114630861e-01,
378 1.24443217440050976e-02, 3.19239968065752286e-01,
379 // nsing = 9
380 -1.66126772808035320e-02, 1.46180329587665321e-02,
381 3.85927576915247303e-02, 1.18910471133003227e-03,
382 -4.94003498320899806e-02, -3.93468443660139110e-02,
383 3.92113167048952835e-02, 9.03306084789976219e-02,
384 -1.82889636251263500e-02, -1.53816215444915245e-01,
385 3.88103861995563741e-03, 2.72090310854550347e-01,
386 // nsing = 10
387 -1.21245068916826880e-02, 1.59080224420074489e-03,
388 1.91116222508366035e-02, 1.05879549199053302e-02,
389 -1.97228428219695318e-02, -3.16962067712639397e-02,
390 1.34110372628315158e-02, 6.18045654429108837e-02,
391 -5.52574921865441838e-03, -1.14259663804569455e-01,
392 1.05534036292203489e-03, 2.37326534898818288e-01,
393 // nsing = 11
394 -5.96835002183177493e-03, -2.42594931567031205e-03,
395 7.44753817476594184e-03, 9.33450807578394386e-03,
396 -6.52649522783026481e-03, -2.08165802069352019e-02,
397 3.89988065678848650e-03, 4.12784313451549132e-02,
398 -1.44110721106127920e-03, -8.76484782997757425e-02,
399 2.50210184908121337e-04, 2.11131066219336647e-01,
400 // nsing = 12
401 -2.24505212235034193e-03, -2.38114524227619446e-03,
402 2.36375918970809340e-03, 5.97324040603806266e-03,
403 -1.81333819936645381e-03, -1.28126250720444051e-02,
404 9.69251586187208358e-04, 2.83055679874589732e-02,
405 -3.24986363596307374e-04, -6.97056268370209313e-02,
406 5.17231862038123061e-05, 1.90681117197597520e-01,
407 // nsing = 13
408 -6.76887607549779069e-04, -1.48589685249767064e-03,
409 6.22548369472046953e-04, 3.43871156746448680e-03,
410 -4.26557147166379929e-04, -7.98854145009655400e-03,
411 2.06644460919535524e-04, 2.03107152586353217e-02,
412 -6.34563929410856987e-05, -5.71425144910115832e-02,
413 9.32252179140502456e-06, 1.74167663785025829e-01,
414 // nsing = 14
415 -1.67596437777156162e-04, -8.05384193869903178e-04,
416 1.37627277777023791e-04, 1.97652692602724093e-03,
417 -8.54392244879459717e-05, -5.23088906415977167e-03,
418 3.78965577556493513e-05, 1.52191559129376333e-02,
419 -1.07393019498185646e-05, -4.79347862153366295e-02,
420 1.46503970628861795e-06, 1.60471011683477685e-01,
421 // nsing = 15
422 -3.45715760630978778e-05, -4.31089554210205493e-04,
423 2.57350138106549737e-05, 1.19449262097417514e-03,
424 -1.46322227517372253e-05, -3.61303766799909378e-03,
425 5.99057675687392260e-06, 1.17993805017130890e-02,
426 -1.57660578509526722e-06, -4.09165023743669707e-02,
427 2.00739683204152177e-07, 1.48879348585662670e-01,
428 // nsing = 16
429 -5.99735188857573424e-06, -2.42949218855805052e-04,
430 4.09249090936269722e-06, 7.67400152727128171e-04,
431 -2.14920611287648034e-06, -2.60710519575546230e-03,
432 8.17591694958640978e-07, 9.38581640137393053e-03,
433 -2.00910914042737743e-07, -3.54045580123653803e-02,
434 2.39819738182594508e-08, 1.38916449405613711e-01,
435 // nsing = 17
436 -8.80708505155966658e-07, -1.46479474515521504e-04,
437 5.55693207391871904e-07, 5.19165587844615415e-04,
438 -2.71391142598826750e-07, -1.94439427580099576e-03,
439 9.64641799864928425e-08, 7.61536975207357980e-03,
440 -2.22357616069432967e-08, -3.09762939485679078e-02,
441 2.49806920458212581e-09, 1.30247401712293206e-01,
442 // nsing = 18
443 -1.10007111030476390e-07, -9.35886150886691786e-05,
444 6.46244096997824390e-08, 3.65267193418479043e-04,
445 -2.95175785569292542e-08, -1.48730955943961081e-03,
446 9.84949251974795537e-09, 6.27824679148707177e-03,
447 -2.13827217704781576e-09, -2.73545766571797965e-02,
448 2.26877724435352177e-10, 1.22627158810895267e-01,
449 // nsing = 19
450 -1.17302439957657553e-08, -6.24890956722053332e-05,
451 6.45231881609786173e-09, 2.64799907072561543e-04,
452 -2.76943921343331654e-09, -1.16094187847598385e-03,
453 8.71074689656480749e-10, 5.24514377390761210e-03,
454 -1.78730768958639407e-10, -2.43489203319091538e-02,
455 1.79658223341365988e-11, 1.15870972518909888e-01,
456 // nsing = 20
457 -1.07084502471985403e-09, -4.31515421260633319e-05,
458 5.54152563270547927e-10, 1.96606443937168357e-04,
459 -2.24423474431542338e-10, -9.21550077887211094e-04,
460 6.67734377376211580e-11, 4.43201203646827019e-03,
461 -1.29896907717633162e-11, -2.18236356404862774e-02,
462 1.24042409733678516e-12, 1.09836276968151848e-01,
463 // nsing = 21
464 -8.38816525569060600e-11, -3.06091807093959821e-05,
465 4.10033961556230842e-11, 1.48895624771753491e-04,
466 -1.57238128435253905e-11, -7.42073499862065649e-04,
467 4.43938379112418832e-12, 3.78197089773957382e-03,
468 -8.21067867869285873e-13, -1.96793607299577220e-02,
469 7.46725770201828754e-14, 1.04410965521273064e-01,
470 // nsing = 22
471 -5.64848922712870507e-12, -2.22021942382507691e-05,
472 2.61729537775838587e-12, 1.14683068921649992e-04,
473 -9.53316139085394895e-13, -6.05021573565916914e-04,
474 2.56116039498542220e-13, 3.25530796858307225e-03,
475 -4.51482829896525004e-14, -1.78416955716514289e-02,
476 3.91940313268087086e-15, 9.95054815464739996e-02,
477 // nsing = 23
478 -3.27482357793897640e-13, -1.64138890390689871e-05,
479 1.44278798346454523e-13, 8.96362542918265398e-05,
480 -5.00524303437266481e-14, -4.98699756861136127e-04,
481 1.28274026095767213e-14, 2.82359118537843949e-03,
482 -2.16009593993917109e-15, -1.62538825704327487e-02,
483 1.79368667683853708e-16, 9.50473084594884184e-02
484 };
485
486__roodevice__ const double npi11[11] = { // precomputed values n * pi
487 0.00000000000000000e+00, 3.14159265358979324e+00, 6.28318530717958648e+00,
488 9.42477796076937972e+00, 1.25663706143591730e+01, 1.57079632679489662e+01,
489 1.88495559215387594e+01, 2.19911485751285527e+01, 2.51327412287183459e+01,
490 2.82743338823081391e+01, 3.14159265358979324e+01
491 };
492__roodevice__ const double a11[11] = { // precomputed Fourier coefficient prefactors
493 4.43113462726379007e-01, 3.79788034073635143e-01, 2.39122407410867584e-01,
494 1.10599187402169792e-01, 3.75782250080904725e-02, 9.37936104296856288e-03,
495 1.71974046186334976e-03, 2.31635559000523461e-04, 2.29192401420125452e-05,
496 1.66589592139340077e-06, 8.89504561311882155e-08
497 };
498__roodevice__ const double taylorarr11[11 * 6] = {
499 // real part imaginary part, low order coefficients last
500 // nsing = 0
501 -1.00000000000000000e+00, 0.00000000000000000e+00,
502 0.00000000000000000e-01, 1.12837916709551257e+00,
503 1.00000000000000000e+00, 0.00000000000000000e+00,
504 // nsing = 1
505 -5.92741768247463996e-01, -7.19914991991294310e-01,
506 -6.73156763521649944e-01, 8.14025039279059577e-01,
507 8.57089811121701143e-01, 4.00248106586639754e-01,
508 // nsing = 2
509 1.26114512111568737e-01, -7.46519337025968199e-01,
510 -8.47666863706379907e-01, 1.89347715957263646e-01,
511 5.39641485816297176e-01, 5.97805988669631615e-01,
512 // nsing = 3
513 4.43238482668529408e-01, -3.03563167310638372e-01,
514 -5.88095866853990048e-01, -2.32638360700858412e-01,
515 2.49595637924601714e-01, 5.77633779156009340e-01,
516 // nsing = 4
517 3.33690792296469441e-01, 3.97048587678703930e-02,
518 -2.66422678503135697e-01, -3.18469797424381480e-01,
519 8.48049724711137773e-02, 4.60546329221462864e-01,
520 // nsing = 5
521 1.42043544696751869e-01, 1.24094227867032671e-01,
522 -8.31224229982140323e-02, -2.40766729258442100e-01,
523 2.11669512031059302e-02, 3.48650139549945097e-01,
524 // nsing = 6
525 3.92113167048952835e-02, 9.03306084789976219e-02,
526 -1.82889636251263500e-02, -1.53816215444915245e-01,
527 3.88103861995563741e-03, 2.72090310854550347e-01,
528 // nsing = 7
529 7.37741897722738503e-03, 5.04625223970221539e-02,
530 -2.87394336989990770e-03, -9.96122819257496929e-02,
531 5.22745478269428248e-04, 2.23361039070072101e-01,
532 // nsing = 8
533 9.69251586187208358e-04, 2.83055679874589732e-02,
534 -3.24986363596307374e-04, -6.97056268370209313e-02,
535 5.17231862038123061e-05, 1.90681117197597520e-01,
536 // nsing = 9
537 9.01625563468897100e-05, 1.74961124275657019e-02,
538 -2.65745127697337342e-05, -5.22070356354932341e-02,
539 3.75952450449939411e-06, 1.67018782142871146e-01,
540 // nsing = 10
541 5.99057675687392260e-06, 1.17993805017130890e-02,
542 -1.57660578509526722e-06, -4.09165023743669707e-02,
543 2.00739683204152177e-07, 1.48879348585662670e-01
544 };
545// clang-format on
546
547__roodevice__ __roohost__ inline STD::complex<double> faddeeva(STD::complex<double> z)
548{
549 return RooHeterogeneousMath::faddeeva_smabmq_impl<double, 24, 6, 9>(
552}
553
554__roodevice__ __roohost__ inline STD::complex<double> faddeeva_fast(STD::complex<double> z)
555{
556 return RooHeterogeneousMath::faddeeva_smabmq_impl<double, 11, 3, 3>(
559}
560
561////////////////////////////////////////////////////////////////////////////////
562/// use the approximation: erf(z) = exp(-z*z)/(STD::sqrt(pi)*z)
563/// to explicitly cancel the divergent exp(y*y) behaviour of
564/// CWERF for z = x + i y with large negative y
565
566__roohost__ __roodevice__ inline STD::complex<double> evalCerfApprox(double _x, double u, double c)
567{
568 const double rootpi = STD::sqrt(STD::atan2(0., -1.));
569 const STD::complex<double> z(_x * c, u + c);
570 const STD::complex<double> zc(u + c, -_x * c);
571 const STD::complex<double> zsq((z.real() + z.imag()) * (z.real() - z.imag()), 2. * z.real() * z.imag());
572 const STD::complex<double> v(-zsq.real() - u * u, -zsq.imag());
573 const STD::complex<double> ev = STD::exp(v);
574 const STD::complex<double> mez2zcrootpi = -STD::exp(zsq) / (zc * rootpi);
575
576 return 2. * (ev * (mez2zcrootpi + 1.));
577}
578
579// Calculate exp(-u^2) cwerf(swt*c + i(u+c)), taking care of numerical instabilities
580__roohost__ __roodevice__ inline STD::complex<double> evalCerf(double swt, double u, double c)
581{
582 if (swt == 0.0) {
583 // For a purely complex argument z, the faddeeva function equals to
584 // exp(z*z) * erfc(z). Together with coefficient exp(-u*u), this means the
585 // function can be simplified to:
586 const double z = u + c;
587 return z > -4.0 ? (STD::exp(c * (c + 2. * u)) * STD::erfc(z)) : evalCerfApprox(0., u, c);
588 // This version with STD::erfc is about twice as fast as the faddeeva_fast
589 // code path, speeding up in particular the analytical convolution of an
590 // exponential decay with a Gaussian (like in RooDecay).
591 }
592 STD::complex<double> z(swt * c, u + c);
593 return (z.imag() > -4.0) ? (STD::exp(-u * u) * faddeeva_fast(z)) : evalCerfApprox(swt, u, c);
594}
595
596} // namespace RooHeterogeneousMath
597
598#endif
#define f(i)
Definition RSha256.hxx:104
#define c(i)
Definition RSha256.hxx:101
#define a(i)
Definition RSha256.hxx:99
#define e(i)
Definition RSha256.hxx:103
#define __roohost__
#define __roodevice__
#define N
#define NCF(TN, I, C)
Definition cfortran.h:897
__roodevice__ const double npi11[11]
__roodevice__ static __roohost__ STD::complex< T > faddeeva_smabmq_impl(T zre, T zim, const T tm, const T(&a)[N], const T(&npi)[N], const T(&taylorarr)[N *NTAYLOR *2])
__roodevice__ static __roohost__ void cexp(double &re, double &im)
static __roodevice__ const double a24[24]
__roodevice__ const double a11[11]
__roodevice__ const double taylorarr11[11 *6]
__roodevice__ __roohost__ STD::complex< double > faddeeva(STD::complex< double > z)
static __roodevice__ const double taylorarr24[24 *12]
__roohost__ __roodevice__ STD::complex< double > evalCerfApprox(double _x, double u, double c)
use the approximation: erf(z) = exp(-z*z)/(STD::sqrt(pi)*z) to explicitly cancel the divergent exp(y*...
__roohost__ __roodevice__ STD::complex< double > evalCerf(double swt, double u, double c)
__roodevice__ __roohost__ STD::complex< double > faddeeva_fast(STD::complex< double > z)
static __roodevice__ const double npi24[24]