| 1 |
/* |
| 2 |
* Copyright Nick Thompson, John Maddock 2020 |
| 3 |
* Use, modification and distribution are subject to the |
| 4 |
* Boost Software License, Version 1.0. (See accompanying file |
| 5 |
* LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt) |
| 6 |
*/ |
| 7 |
|
| 8 |
#ifndef BOOST_MATH_SPECIAL_DAUBECHIES_SCALING_HPP |
| 9 |
#define BOOST_MATH_SPECIAL_DAUBECHIES_SCALING_HPP |
| 10 |
|
| 11 |
#include <cstdint> |
| 12 |
#include <cstring> |
| 13 |
#include <cmath> |
| 14 |
#include <vector> |
| 15 |
#include <array> |
| 16 |
#include <thread> |
| 17 |
#include <future> |
| 18 |
#include <iostream> |
| 19 |
#include <memory> |
| 20 |
#include <boost/math/special_functions/detail/daubechies_scaling_integer_grid.hpp> |
| 21 |
#include <boost/math/filters/daubechies.hpp> |
| 22 |
#include <boost/math/interpolators/detail/cubic_hermite_detail.hpp> |
| 23 |
#include <boost/math/interpolators/detail/quintic_hermite_detail.hpp> |
| 24 |
#include <boost/math/interpolators/detail/septic_hermite_detail.hpp> |
| 25 |
|
| 26 |
#include <boost/math/tools/is_standalone.hpp> |
| 27 |
#ifndef BOOST_MATH_STANDALONE |
| 28 |
# include <boost/config.hpp> |
| 29 |
# ifdef BOOST_MATH_NO_CXX17_IF_CONSTEXPR |
| 30 |
# error "The header <boost/special_functions/daubechies_scaling.hpp> can only be used in C++17 and later." |
| 31 |
# endif |
| 32 |
#endif |
| 33 |
|
| 34 |
namespace boost::math { |
| 35 |
|
| 36 |
template<class Real, int p, int order> |
| 37 |
std::vector<Real> daubechies_scaling_dyadic_grid(int64_t j_max) |
| 38 |
{ |
| 39 |
using std::isnan; |
| 40 |
using std::sqrt; |
| 41 |
auto c = boost::math::filters::daubechies_scaling_filter<Real, p>(); |
| 42 |
Real scale = sqrt(static_cast<Real>(2))*(1 << order); |
| 43 |
for (auto & x : c) |
| 44 |
{ |
| 45 |
x *= scale; |
| 46 |
} |
| 47 |
|
| 48 |
auto phik = detail::daubechies_scaling_integer_grid<Real, p, order>(); |
| 49 |
|
| 50 |
// Maximum sensible j for 32 bit floats is j_max = 22: |
| 51 |
if constexpr (std::is_same_v<Real, float>) |
| 52 |
{ |
| 53 |
if (j_max > 23) |
| 54 |
{ |
| 55 |
throw std::logic_error("Requested dyadic grid more dense than number of representables on the interval."); |
| 56 |
} |
| 57 |
} |
| 58 |
std::vector<Real> v(2*p + (2*p-1)*((1<<j_max) -1), std::numeric_limits<Real>::quiet_NaN()); |
| 59 |
v[0] = 0; |
| 60 |
v[v.size()-1] = 0; |
| 61 |
for (int64_t i = 0; i < static_cast<int64_t>(phik.size()); ++i) { |
| 62 |
v[i*(1uLL<<j_max)] = phik[i]; |
| 63 |
} |
| 64 |
|
| 65 |
for (int64_t j = 1; j <= j_max; ++j) |
| 66 |
{ |
| 67 |
int64_t k_max = v.size()/(int64_t(1) << (j_max-j)); |
| 68 |
for (int64_t k = 1; k < k_max; k += 2) |
| 69 |
{ |
| 70 |
// Where this value will go: |
| 71 |
int64_t delivery_idx = k*(1uLL << (j_max-j)); |
| 72 |
// This is a nice check, but we've tested this exhaustively, and it's an expensive check: |
| 73 |
//if (delivery_idx >= static_cast<int64_t>(v.size())) { |
| 74 |
// std::cerr << "Delivery index out of range!\n"; |
| 75 |
// continue; |
| 76 |
//} |
| 77 |
Real term = 0; |
| 78 |
for (int64_t l = 0; l < static_cast<int64_t>(c.size()); ++l) |
| 79 |
{ |
| 80 |
int64_t idx = k*(int64_t(1) << (j_max - j + 1)) - l*(int64_t(1) << j_max); |
| 81 |
if (idx < 0) |
| 82 |
{ |
| 83 |
break; |
| 84 |
} |
| 85 |
if (idx < static_cast<int64_t>(v.size())) |
| 86 |
{ |
| 87 |
term += c[l]*v[idx]; |
| 88 |
} |
| 89 |
} |
| 90 |
// Again, another nice check: |
| 91 |
//if (!isnan(v[delivery_idx])) { |
| 92 |
// std::cerr << "Delivery index already populated!, = " << v[delivery_idx] << "\n"; |
| 93 |
// std::cerr << "would overwrite with " << term << "\n"; |
| 94 |
//} |
| 95 |
v[delivery_idx] = term; |
| 96 |
} |
| 97 |
} |
| 98 |
return v; |
| 99 |
} |
| 100 |
|
| 101 |
namespace detail { |
| 102 |
|
| 103 |
template<class RandomAccessContainer> |
| 104 |
class matched_holder { |
| 105 |
public: |
| 106 |
using Real = typename RandomAccessContainer::value_type; |
| 107 |
|
| 108 |
matched_holder(RandomAccessContainer && y, RandomAccessContainer && dydx, int grid_refinements, Real x0) : x0_{x0}, y_{std::move(y)}, dy_{std::move(dydx)} |
| 109 |
{ |
| 110 |
inv_h_ = (1 << grid_refinements); |
| 111 |
Real h = 1/inv_h_; |
| 112 |
for (auto & dy : dy_) |
| 113 |
{ |
| 114 |
dy *= h; |
| 115 |
} |
| 116 |
} |
| 117 |
|
| 118 |
inline Real operator()(Real x) const |
| 119 |
{ |
| 120 |
using std::floor; |
| 121 |
using std::sqrt; |
| 122 |
// This is the exact Holder exponent, but it's pessimistic almost everywhere! |
| 123 |
// It's only exactly right at dyadic rationals. |
| 124 |
//Real const alpha = 2 - log(1+sqrt(Real(3)))/log(Real(2)); |
| 125 |
// We're gonna use alpha = 1/2, rather than 0.5500... |
| 126 |
Real s = (x-x0_)*inv_h_; |
| 127 |
Real ii = floor(s); |
| 128 |
auto i = static_cast<decltype(y_.size())>(ii); |
| 129 |
Real t = s - ii; |
| 130 |
Real dphi = dy_[i+1]; |
| 131 |
Real diff = y_[i+1] - y_[i]; |
| 132 |
return y_[i] + (2*dphi - diff)*t + 2*sqrt(t)*(diff-dphi); |
| 133 |
} |
| 134 |
|
| 135 |
int64_t bytes() const |
| 136 |
{ |
| 137 |
return 2*y_.size()*sizeof(Real) + sizeof(*this); |
| 138 |
} |
| 139 |
|
| 140 |
private: |
| 141 |
Real x0_; |
| 142 |
Real inv_h_; |
| 143 |
RandomAccessContainer y_; |
| 144 |
RandomAccessContainer dy_; |
| 145 |
}; |
| 146 |
|
| 147 |
template<class RandomAccessContainer> |
| 148 |
class matched_holder_aos { |
| 149 |
public: |
| 150 |
using Point = typename RandomAccessContainer::value_type; |
| 151 |
using Real = typename Point::value_type; |
| 152 |
|
| 153 |
matched_holder_aos(RandomAccessContainer && data, int grid_refinements, Real x0) : x0_{x0}, data_{std::move(data)} |
| 154 |
{ |
| 155 |
inv_h_ = Real(1uLL << grid_refinements); |
| 156 |
Real h = 1/inv_h_; |
| 157 |
for (auto & datum : data_) |
| 158 |
{ |
| 159 |
datum[1] *= h; |
| 160 |
} |
| 161 |
} |
| 162 |
|
| 163 |
inline Real operator()(Real x) const |
| 164 |
{ |
| 165 |
using std::floor; |
| 166 |
using std::sqrt; |
| 167 |
Real s = (x-x0_)*inv_h_; |
| 168 |
Real ii = floor(s); |
| 169 |
auto i = static_cast<decltype(data_.size())>(ii); |
| 170 |
Real t = s - ii; |
| 171 |
Real y0 = data_[i][0]; |
| 172 |
Real y1 = data_[i+1][0]; |
| 173 |
Real dphi = data_[i+1][1]; |
| 174 |
Real diff = y1 - y0; |
| 175 |
return y0 + (2*dphi - diff)*t + 2*sqrt(t)*(diff-dphi); |
| 176 |
} |
| 177 |
|
| 178 |
int64_t bytes() const |
| 179 |
{ |
| 180 |
return data_.size()*data_[0].size()*sizeof(Real) + sizeof(*this); |
| 181 |
} |
| 182 |
|
| 183 |
private: |
| 184 |
Real x0_; |
| 185 |
Real inv_h_; |
| 186 |
RandomAccessContainer data_; |
| 187 |
}; |
| 188 |
|
| 189 |
|
| 190 |
template<class RandomAccessContainer> |
| 191 |
class linear_interpolation { |
| 192 |
public: |
| 193 |
using Real = typename RandomAccessContainer::value_type; |
| 194 |
|
| 195 |
linear_interpolation(RandomAccessContainer && y, RandomAccessContainer && dydx, int grid_refinements) : y_{std::move(y)}, dydx_{std::move(dydx)} |
| 196 |
{ |
| 197 |
s_ = (1 << grid_refinements); |
| 198 |
} |
| 199 |
|
| 200 |
inline Real operator()(Real x) const |
| 201 |
{ |
| 202 |
using std::floor; |
| 203 |
Real y = x*s_; |
| 204 |
Real k = floor(y); |
| 205 |
|
| 206 |
int64_t kk = static_cast<int64_t>(k); |
| 207 |
Real t = y - k; |
| 208 |
return (1-t)*y_[kk] + t*y_[kk+1]; |
| 209 |
} |
| 210 |
|
| 211 |
inline Real prime(Real x) const |
| 212 |
{ |
| 213 |
using std::floor; |
| 214 |
|
| 215 |
Real y = x*s_; |
| 216 |
Real k = floor(y); |
| 217 |
|
| 218 |
int64_t kk = static_cast<int64_t>(k); |
| 219 |
Real t = y - k; |
| 220 |
return static_cast<Real>((Real(1)-t)*dydx_[kk] + t*dydx_[kk+1]); |
| 221 |
} |
| 222 |
|
| 223 |
int64_t bytes() const |
| 224 |
{ |
| 225 |
return (1 + y_.size() + dydx_.size())*sizeof(Real) + sizeof(y_) + sizeof(dydx_); |
| 226 |
} |
| 227 |
|
| 228 |
private: |
| 229 |
Real s_; |
| 230 |
RandomAccessContainer y_; |
| 231 |
RandomAccessContainer dydx_; |
| 232 |
}; |
| 233 |
|
| 234 |
template<class RandomAccessContainer> |
| 235 |
class linear_interpolation_aos { |
| 236 |
public: |
| 237 |
using Point = typename RandomAccessContainer::value_type; |
| 238 |
using Real = typename Point::value_type; |
| 239 |
|
| 240 |
linear_interpolation_aos(RandomAccessContainer && data, int grid_refinements, Real x0) : x0_{x0}, data_{std::move(data)} |
| 241 |
{ |
| 242 |
s_ = Real(1uLL << grid_refinements); |
| 243 |
} |
| 244 |
|
| 245 |
inline Real operator()(Real x) const |
| 246 |
{ |
| 247 |
using std::floor; |
| 248 |
Real y = (x-x0_)*s_; |
| 249 |
Real k = floor(y); |
| 250 |
|
| 251 |
int64_t kk = static_cast<int64_t>(k); |
| 252 |
Real t = y - k; |
| 253 |
return (t != 0) ? (1-t)*data_[kk][0] + t*data_[kk+1][0] : data_[kk][0]; |
| 254 |
} |
| 255 |
|
| 256 |
inline Real prime(Real x) const |
| 257 |
{ |
| 258 |
using std::floor; |
| 259 |
Real y = (x-x0_)*s_; |
| 260 |
Real k = floor(y); |
| 261 |
|
| 262 |
int64_t kk = static_cast<int64_t>(k); |
| 263 |
Real t = y - k; |
| 264 |
return t != 0 ? (1-t)*data_[kk][1] + t*data_[kk+1][1] : data_[kk][1]; |
| 265 |
} |
| 266 |
|
| 267 |
int64_t bytes() const |
| 268 |
{ |
| 269 |
return sizeof(*this) + data_.size()*data_[0].size()*sizeof(Real); |
| 270 |
} |
| 271 |
|
| 272 |
private: |
| 273 |
Real x0_; |
| 274 |
Real s_; |
| 275 |
RandomAccessContainer data_; |
| 276 |
}; |
| 277 |
|
| 278 |
|
| 279 |
template <class T> |
| 280 |
struct daubechies_eval_type |
| 281 |
{ |
| 282 |
using type = T; |
| 283 |
|
| 284 |
static const std::vector<T>& vector_cast(const std::vector<T>& v) { return v; } |
| 285 |
|
| 286 |
}; |
| 287 |
template <> |
| 288 |
struct daubechies_eval_type<float> |
| 289 |
{ |
| 290 |
using type = double; |
| 291 |
|
| 292 |
inline static std::vector<float> vector_cast(const std::vector<double>& v) |
| 293 |
{ |
| 294 |
std::vector<float> result(v.size()); |
| 295 |
for (unsigned i = 0; i < v.size(); ++i) |
| 296 |
result[i] = static_cast<float>(v[i]); |
| 297 |
return result; |
| 298 |
} |
| 299 |
}; |
| 300 |
template <> |
| 301 |
struct daubechies_eval_type<double> |
| 302 |
{ |
| 303 |
using type = long double; |
| 304 |
|
| 305 |
inline static std::vector<double> vector_cast(const std::vector<long double>& v) |
| 306 |
{ |
| 307 |
std::vector<double> result(v.size()); |
| 308 |
for (unsigned i = 0; i < v.size(); ++i) |
| 309 |
result[i] = static_cast<double>(v[i]); |
| 310 |
return result; |
| 311 |
} |
| 312 |
}; |
| 313 |
|
| 314 |
struct null_interpolator |
| 315 |
{ |
| 316 |
template <class T> |
| 317 |
T operator()(const T&) |
| 318 |
{ |
| 319 |
return 1; |
| 320 |
} |
| 321 |
}; |
| 322 |
|
| 323 |
} // namespace detail |
| 324 |
|
| 325 |
template<class Real, int p> |
| 326 |
class daubechies_scaling { |
| 327 |
// |
| 328 |
// Some type manipulation so we know the type of the interpolator, and the vector type it requires: |
| 329 |
// |
| 330 |
using vector_type = std::vector<std::array<Real, p < 6 ? 2 : p < 10 ? 3 : 4>>; |
| 331 |
// |
| 332 |
// List our interpolators: |
| 333 |
// |
| 334 |
using interpolator_list = std::tuple< |
| 335 |
detail::null_interpolator, detail::matched_holder_aos<vector_type>, detail::linear_interpolation_aos<vector_type>, |
| 336 |
interpolators::detail::cardinal_cubic_hermite_detail_aos<vector_type>, interpolators::detail::cardinal_quintic_hermite_detail_aos<vector_type>, |
| 337 |
interpolators::detail::cardinal_septic_hermite_detail_aos<vector_type> >; |
| 338 |
// |
| 339 |
// Select the one we need: |
| 340 |
// |
| 341 |
using interpolator_type = std::tuple_element_t< |
| 342 |
p == 1 ? 0 : |
| 343 |
p == 2 ? 1 : |
| 344 |
p == 3 ? 2 : |
| 345 |
p <= 5 ? 3 : |
| 346 |
p <= 9 ? 4 : 5, interpolator_list>; |
| 347 |
|
| 348 |
public: |
| 349 |
daubechies_scaling(int grid_refinements = -1) |
| 350 |
{ |
| 351 |
static_assert(p < 20, "Daubechies scaling functions are only implemented for p < 20."); |
| 352 |
static_assert(p > 0, "Daubechies scaling functions must have at least 1 vanishing moment."); |
| 353 |
if constexpr (p == 1) |
| 354 |
{ |
| 355 |
return; |
| 356 |
} |
| 357 |
else { |
| 358 |
if (grid_refinements < 0) |
| 359 |
{ |
| 360 |
if constexpr (std::is_same_v<Real, float>) |
| 361 |
{ |
| 362 |
if (grid_refinements == -2) |
| 363 |
{ |
| 364 |
// Control absolute error: |
| 365 |
// p= 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19 |
| 366 |
constexpr std::array<int, 20> r{ -1, -1, 18, 19, 16, 11, 8, 7, 7, 7, 5, 5, 4, 4, 4, 4, 3, 3, 3, 3 }; |
| 367 |
grid_refinements = r[p]; |
| 368 |
} |
| 369 |
else |
| 370 |
{ |
| 371 |
// Control relative error: |
| 372 |
// p= 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19 |
| 373 |
constexpr std::array<int, 20> r{ -1, -1, 21, 21, 21, 17, 16, 15, 14, 13, 12, 11, 11, 11, 11, 11, 11, 11, 11, 11 }; |
| 374 |
grid_refinements = r[p]; |
| 375 |
} |
| 376 |
} |
| 377 |
else if constexpr (std::is_same_v<Real, double>) |
| 378 |
{ |
| 379 |
// p= 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19 |
| 380 |
constexpr std::array<int, 20> r{ -1, -1, 21, 21, 21, 21, 21, 21, 21, 21, 20, 20, 19, 19, 18, 18, 18, 18, 18, 18 }; |
| 381 |
grid_refinements = r[p]; |
| 382 |
} |
| 383 |
else |
| 384 |
{ |
| 385 |
grid_refinements = 21; |
| 386 |
} |
| 387 |
} |
| 388 |
|
| 389 |
// Compute the refined grid: |
| 390 |
// In fact for float precision I know the grid must be computed in double precision and then cast back down, or else parts of the support are systematically inaccurate. |
| 391 |
std::future<std::vector<Real>> t0 = std::async(std::launch::async, [&grid_refinements]() { |
| 392 |
// Computing in higher precision and downcasting is essential for 1ULP evaluation in float precision: |
| 393 |
auto v = daubechies_scaling_dyadic_grid<typename detail::daubechies_eval_type<Real>::type, p, 0>(grid_refinements); |
| 394 |
return detail::daubechies_eval_type<Real>::vector_cast(v); |
| 395 |
}); |
| 396 |
// Compute the derivative of the refined grid: |
| 397 |
std::future<std::vector<Real>> t1 = std::async(std::launch::async, [&grid_refinements]() { |
| 398 |
auto v = daubechies_scaling_dyadic_grid<typename detail::daubechies_eval_type<Real>::type, p, 1>(grid_refinements); |
| 399 |
return detail::daubechies_eval_type<Real>::vector_cast(v); |
| 400 |
}); |
| 401 |
|
| 402 |
// if necessary, compute the second and third derivative: |
| 403 |
std::vector<Real> d2ydx2; |
| 404 |
std::vector<Real> d3ydx3; |
| 405 |
if constexpr (p >= 6) { |
| 406 |
std::future<std::vector<Real>> t3 = std::async(std::launch::async, [&grid_refinements]() { |
| 407 |
auto v = daubechies_scaling_dyadic_grid<typename detail::daubechies_eval_type<Real>::type, p, 2>(grid_refinements); |
| 408 |
return detail::daubechies_eval_type<Real>::vector_cast(v); |
| 409 |
}); |
| 410 |
|
| 411 |
if constexpr (p >= 10) { |
| 412 |
std::future<std::vector<Real>> t4 = std::async(std::launch::async, [&grid_refinements]() { |
| 413 |
auto v = daubechies_scaling_dyadic_grid<typename detail::daubechies_eval_type<Real>::type, p, 3>(grid_refinements); |
| 414 |
return detail::daubechies_eval_type<Real>::vector_cast(v); |
| 415 |
}); |
| 416 |
d3ydx3 = t4.get(); |
| 417 |
} |
| 418 |
d2ydx2 = t3.get(); |
| 419 |
} |
| 420 |
|
| 421 |
|
| 422 |
auto y = t0.get(); |
| 423 |
auto dydx = t1.get(); |
| 424 |
|
| 425 |
if constexpr (p >= 2) |
| 426 |
{ |
| 427 |
vector_type data(y.size()); |
| 428 |
for (size_t i = 0; i < y.size(); ++i) |
| 429 |
{ |
| 430 |
data[i][0] = y[i]; |
| 431 |
data[i][1] = dydx[i]; |
| 432 |
if constexpr (p >= 6) |
| 433 |
data[i][2] = d2ydx2[i]; |
| 434 |
if constexpr (p >= 10) |
| 435 |
data[i][3] = d3ydx3[i]; |
| 436 |
} |
| 437 |
if constexpr (p <= 3) |
| 438 |
m_interpolator = std::make_shared<interpolator_type>(std::move(data), grid_refinements, Real(0)); |
| 439 |
else |
| 440 |
m_interpolator = std::make_shared<interpolator_type>(std::move(data), Real(0), Real(1) / (1 << grid_refinements)); |
| 441 |
} |
| 442 |
else |
| 443 |
m_interpolator = std::make_shared<detail::null_interpolator>(); |
| 444 |
} |
| 445 |
} |
| 446 |
|
| 447 |
inline Real operator()(Real x) const |
| 448 |
{ |
| 449 |
if (x <= 0 || x >= 2*p-1) |
| 450 |
{ |
| 451 |
return 0; |
| 452 |
} |
| 453 |
return (*m_interpolator)(x); |
| 454 |
} |
| 455 |
|
| 456 |
inline Real prime(Real x) const |
| 457 |
{ |
| 458 |
static_assert(p > 2, "The 3-vanishing moment Daubechies scaling function is the first which is continuously differentiable."); |
| 459 |
if (x <= Real(0) || x >= 2*p-1) |
| 460 |
{ |
| 461 |
return 0; |
| 462 |
} |
| 463 |
return m_interpolator->prime(x); |
| 464 |
} |
| 465 |
|
| 466 |
inline Real double_prime(Real x) const |
| 467 |
{ |
| 468 |
static_assert(p >= 6, "Second derivatives require at least 6 vanishing moments."); |
| 469 |
if (x <= 0 || x >= 2*p - 1) |
| 470 |
{ |
| 471 |
return Real(0); |
| 472 |
} |
| 473 |
return m_interpolator->double_prime(x); |
| 474 |
} |
| 475 |
|
| 476 |
std::pair<Real, Real> support() const |
| 477 |
{ |
| 478 |
return {Real(0), Real(2*p-1)}; |
| 479 |
} |
| 480 |
|
| 481 |
int64_t bytes() const |
| 482 |
{ |
| 483 |
return m_interpolator->bytes() + sizeof(m_interpolator); |
| 484 |
} |
| 485 |
|
| 486 |
private: |
| 487 |
std::shared_ptr<interpolator_type> m_interpolator; |
| 488 |
}; |
| 489 |
|
| 490 |
} |
| 491 |
#endif |