2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2019,2020,2021, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
37 * Implements Gaussian function evaluations on lattices and related functionality
39 * \author Christian Blau <blau@kth.se>
41 * \ingroup module_math
45 #include "gausstransform.h"
52 #include "gromacs/math/functions.h"
53 #include "gromacs/math/multidimarray.h"
54 #include "gromacs/math/units.h"
55 #include "gromacs/math/utilities.h"
60 /********************************************************************
61 * GaussianOn1DLattice::Impl
64 class GaussianOn1DLattice::Impl
67 Impl(int numGridPointsForSpreadingHalfWidth, real sigma);
69 Impl(const Impl& other) = default;
70 Impl& operator=(const Impl& other) = default;
72 /*! \brief evaluate Gaussian function at all lattice points
73 * \param[in] amplitude the amplitude of the Gaussian
74 * \param[in] dx distance from the center
76 void spread(double amplitude, real dx);
77 //! Largest distance in number of gridpoints from 0
78 int numGridPointsForSpreadingHalfWidth_;
79 /*! \brief Avoid overflow for E2^offset and underflow for E3(i).
81 * Occurs when sigma is much smaller than numGridPointsForSpreadingHalfWidth_.
83 * E2^offset smaller than maximum float requires
84 * \f$exp(dx / (2*square(sigma))^numGridPointsForSpreadingHalfWidth_ \leq max_float \f$
85 * The maximum expected distance of the Gaussian center to the next lattice point is dx = 0.5,
86 * thus the maximum spread distance here is \f$4 * sigma^2 * \log(\mathrm{maxfloat})\f$ .
88 * E3(i) larger than minmium float requires
89 * exp(i^2 / 2*(sigma)^2) > min_float
90 * Thus the maximum spread distance here is \f$\sigma \sqrt(-2\log(\mathrm{minfloat}))\f$
92 int maxEvaluatedSpreadDistance_;
93 //! Width of the Gaussian function
95 //! The result of the spreading calculation
96 std::vector<float> spreadingResult_;
97 //! Pre-calculated exp(-gridIndex^2/2 * (sigma^2)) named as in Greengard2004
98 std::vector<float> e3_;
99 /*! \brief Equal to std::floor(std::log(std::numeric_limits<float>::max())).
100 * Above expression is not constexpr and a const variable would implicitly delete default copy
101 * assignment. Therefore resorting to setting number manually.
103 static constexpr double c_logMaxFloat = 88.72284;
104 static constexpr double c_logMinFloat = -87.33654;
107 GaussianOn1DLattice::Impl::Impl(int numGridPointsForSpreadingHalfWidth, real sigma) :
108 numGridPointsForSpreadingHalfWidth_(numGridPointsForSpreadingHalfWidth),
110 spreadingResult_(2 * numGridPointsForSpreadingHalfWidth + 1)
112 maxEvaluatedSpreadDistance_ =
113 std::min(numGridPointsForSpreadingHalfWidth_,
114 static_cast<int>(std::floor(4 * square(sigma) * c_logMaxFloat)) - 1);
115 maxEvaluatedSpreadDistance_ =
116 std::min(maxEvaluatedSpreadDistance_,
117 static_cast<int>(std::floor(sigma * sqrt(-2.0 * c_logMinFloat))) - 1);
120 std::back_inserter(e3_), maxEvaluatedSpreadDistance_ + 1, [sigma, latticeIndex = 0]() mutable {
121 return std::exp(-0.5 * square(latticeIndex++ / sigma));
124 std::fill(std::begin(spreadingResult_), std::end(spreadingResult_), 0.);
127 void GaussianOn1DLattice::Impl::spread(double amplitude, real dx)
129 /* The spreading routine implements the fast gaussian gridding as in
131 * Leslie Greengard and June-Yub Lee,
132 * "Accelerating the Nonuniform Fast Fourier Transform"
133 * SIAM REV 2004 Vol. 46, No. 3, pp. 443-454 DOI. 10.1137/S003614450343200X
135 * Following the naming conventions for e1, e2 and e3, nu = 1, m = numGridPointsForSpreadingHalfWidth_.
137 * Speed up is achieved by factorization of the exponential that is evaluted
138 * at regular lattice points i, where the distance from the
139 * Gaussian center is \f$x-i\f$:
142 * a * \exp(-(x^2-2*i*x+ i^2)/(2*\sigma^2)) =
143 * a * \exp(-x^2/2*\sigma^2) * \exp(x/\sigma^2)^i * \exp(i/2*sigma^2) =
144 * e_1(x) * e_2(x)^i * e_3(i)
147 * Requiring only two exp evaluations per spreading operation.
150 const double e1 = amplitude * exp(-0.5 * dx * dx / square(sigma_)) / (sqrt(2 * M_PI) * sigma_);
151 spreadingResult_[numGridPointsForSpreadingHalfWidth_] = e1;
153 const double e2 = exp(dx / square(sigma_));
155 double e2pow = e2; //< powers of e2, e2^offset
157 // Move outwards from mid-point, using e2pow value for both points simultaneously
158 // o o o<----O---->o o o
159 for (int offset = 1; offset < maxEvaluatedSpreadDistance_; offset++)
161 const double e1_3 = e1 * e3_[offset];
162 spreadingResult_[numGridPointsForSpreadingHalfWidth_ + offset] = e1_3 * e2pow;
163 spreadingResult_[numGridPointsForSpreadingHalfWidth_ - offset] = e1_3 / e2pow;
166 // separate statement for gridpoints at the end of the range avoids
167 // overflow for large sigma and saves one e2 multiplication operation
168 spreadingResult_[numGridPointsForSpreadingHalfWidth_ - maxEvaluatedSpreadDistance_] =
169 (e1 / e2pow) * e3_[maxEvaluatedSpreadDistance_];
170 spreadingResult_[numGridPointsForSpreadingHalfWidth_ + maxEvaluatedSpreadDistance_] =
171 (e1 * e2pow) * e3_[maxEvaluatedSpreadDistance_];
174 /********************************************************************
175 * GaussianOn1DLattice
178 GaussianOn1DLattice::GaussianOn1DLattice(int numGridPointsForSpreadingHalfWidth_, real sigma) :
179 impl_(new Impl(numGridPointsForSpreadingHalfWidth_, sigma))
183 GaussianOn1DLattice::~GaussianOn1DLattice() {}
185 void GaussianOn1DLattice::spread(double amplitude, real dx)
187 impl_->spread(amplitude, dx);
190 ArrayRef<const float> GaussianOn1DLattice::view()
192 return impl_->spreadingResult_;
195 GaussianOn1DLattice::GaussianOn1DLattice(const GaussianOn1DLattice& other) :
196 impl_(new Impl(*other.impl_))
200 GaussianOn1DLattice& GaussianOn1DLattice::operator=(const GaussianOn1DLattice& other)
202 *impl_ = *other.impl_;
206 GaussianOn1DLattice::GaussianOn1DLattice(GaussianOn1DLattice&&) noexcept = default;
208 GaussianOn1DLattice& GaussianOn1DLattice::operator=(GaussianOn1DLattice&&) noexcept = default;
213 //! rounds real-valued coordinate to the closest integer values
214 IVec closestIntegerPoint(const RVec& coordinate)
216 return { roundToInt(coordinate[XX]), roundToInt(coordinate[YY]), roundToInt(coordinate[ZZ]) };
219 /*! \brief Substracts a range from a three-dimensional integer coordinate and ensures
220 * the resulting coordinate is within a lattice.
221 * \param[in] index point in lattice
222 * \param[in] range to be shifted
223 * \returns Shifted index or zero if shifted index is smaller than zero.
225 IVec rangeBeginWithinLattice(const IVec& index, const IVec& range)
227 return elementWiseMax({ 0, 0, 0 }, index - range);
230 /*! \brief Adds a range from a three-dimensional integer coordinate and ensures
231 * the resulting coordinate is within a lattice.
232 * \param[in] index point in lattice
233 * \param[in] extents extent of the lattice
234 * \param[in] range to be shifted
235 * \returns Shifted index or the lattice extent if shifted index is larger than the extent
237 IVec rangeEndWithinLattice(const IVec& index, const dynamicExtents3D& extents, const IVec& range)
239 IVec extentAsIvec(static_cast<int>(extents.extent(ZZ)),
240 static_cast<int>(extents.extent(YY)),
241 static_cast<int>(extents.extent(XX)));
242 return elementWiseMin(extentAsIvec, index + range);
248 /********************************************************************
249 * OuterProductEvaluator
252 mdspan<const float, dynamic_extent, dynamic_extent> OuterProductEvaluator::
253 operator()(ArrayRef<const float> x, ArrayRef<const float> y)
255 data_.resize(ssize(x), ssize(y));
256 for (gmx::index xIndex = 0; xIndex < ssize(x); ++xIndex)
258 const auto xValue = x[xIndex];
259 std::transform(std::begin(y), std::end(y), begin(data_.asView()[xIndex]), [xValue](float yValue) {
260 return xValue * yValue;
263 return data_.asConstView();
266 /********************************************************************
270 IntegerBox::IntegerBox(const IVec& begin, const IVec& end) : begin_{ begin }, end_{ end } {}
272 const IVec& IntegerBox::begin() const
276 const IVec& IntegerBox::end() const
281 bool IntegerBox::empty() const
283 return !((begin_[XX] < end_[XX]) && (begin_[YY] < end_[YY]) && (begin_[ZZ] < end_[ZZ]));
286 IntegerBox spreadRangeWithinLattice(const IVec& center, dynamicExtents3D extent, IVec range)
288 const IVec begin = rangeBeginWithinLattice(center, range);
289 const IVec end = rangeEndWithinLattice(center, extent, range);
290 return { begin, end };
292 /********************************************************************
293 * GaussianSpreadKernel
296 IVec GaussianSpreadKernelParameters::Shape::latticeSpreadRange() const
298 DVec range(std::ceil(sigma_[XX] * spreadWidthMultiplesOfSigma_),
299 std::ceil(sigma_[YY] * spreadWidthMultiplesOfSigma_),
300 std::ceil(sigma_[ZZ] * spreadWidthMultiplesOfSigma_));
301 return range.toIVec();
304 /********************************************************************
305 * GaussTransform3D::Impl
309 * Private implementation class for GaussTransform3D.
311 class GaussTransform3D::Impl
314 //! Construct from extent and spreading width and range
315 Impl(const dynamicExtents3D& extent, const GaussianSpreadKernelParameters::Shape& kernelShapeParameters);
318 Impl(const Impl& other) = default;
320 Impl& operator=(const Impl& other) = default;
321 //! Add another gaussian
322 void add(const GaussianSpreadKernelParameters::PositionAndAmplitude& localParamters);
323 //! The width of the Gaussian in lattice spacing units
324 BasicVector<double> sigma_;
325 //! The spread range in lattice points
327 //! The result of the Gauss transform
328 MultiDimArray<std::vector<float>, dynamicExtents3D> data_;
329 //! The outer product of a Gaussian along the z and y dimension
330 OuterProductEvaluator outerProductZY_;
331 //! The three one-dimensional Gaussians, whose outer product is added to the Gauss transform
332 std::array<GaussianOn1DLattice, DIM> gauss1d_;
335 GaussTransform3D::Impl::Impl(const dynamicExtents3D& extent,
336 const GaussianSpreadKernelParameters::Shape& kernelShapeParameters) :
337 sigma_{ kernelShapeParameters.sigma_ },
338 spreadRange_{ kernelShapeParameters.latticeSpreadRange() },
340 gauss1d_({ GaussianOn1DLattice(spreadRange_[XX], sigma_[XX]),
341 GaussianOn1DLattice(spreadRange_[YY], sigma_[YY]),
342 GaussianOn1DLattice(spreadRange_[ZZ], sigma_[ZZ]) })
346 void GaussTransform3D::Impl::add(const GaussianSpreadKernelParameters::PositionAndAmplitude& localParameters)
348 const IVec closestLatticePoint = closestIntegerPoint(localParameters.coordinate_);
349 const auto spreadRange =
350 spreadRangeWithinLattice(closestLatticePoint, data_.asView().extents(), spreadRange_);
352 // do nothing if the added Gaussian will never reach the lattice
353 if (spreadRange.empty())
358 for (int dimension = XX; dimension <= ZZ; ++dimension)
360 // multiply with amplitude so that Gauss3D = (amplitude * Gauss_x) * Gauss_y * Gauss_z
361 const float gauss1DAmplitude = dimension > XX ? 1.0 : localParameters.amplitude_;
362 gauss1d_[dimension].spread(
363 gauss1DAmplitude, localParameters.coordinate_[dimension] - closestLatticePoint[dimension]);
366 const auto spreadZY = outerProductZY_(gauss1d_[ZZ].view(), gauss1d_[YY].view());
367 const auto spreadX = gauss1d_[XX].view();
368 const IVec spreadGridOffset = spreadRange_ - closestLatticePoint;
370 // \todo optimize these loops if performance critical
371 // The looping strategy uses that the last, x-dimension is contiguous in the memory layout
372 for (int zLatticeIndex = spreadRange.begin()[ZZ]; zLatticeIndex < spreadRange.end()[ZZ]; ++zLatticeIndex)
374 const auto zSlice = data_.asView()[zLatticeIndex];
376 for (int yLatticeIndex = spreadRange.begin()[YY]; yLatticeIndex < spreadRange.end()[YY]; ++yLatticeIndex)
378 const auto ySlice = zSlice[yLatticeIndex];
379 const float zyPrefactor = spreadZY(zLatticeIndex + spreadGridOffset[ZZ],
380 yLatticeIndex + spreadGridOffset[YY]);
382 for (int xLatticeIndex = spreadRange.begin()[XX]; xLatticeIndex < spreadRange.end()[XX];
385 const float xPrefactor = spreadX[xLatticeIndex + spreadGridOffset[XX]];
386 ySlice[xLatticeIndex] += zyPrefactor * xPrefactor;
392 /********************************************************************
396 GaussTransform3D::GaussTransform3D(const dynamicExtents3D& extent,
397 const GaussianSpreadKernelParameters::Shape& kernelShapeParameters) :
398 impl_(new Impl(extent, kernelShapeParameters))
402 void GaussTransform3D::add(const GaussianSpreadKernelParameters::PositionAndAmplitude& localParameters)
404 impl_->add(localParameters);
407 void GaussTransform3D::setZero()
409 std::fill(begin(impl_->data_), end(impl_->data_), 0.);
412 basic_mdspan<float, dynamicExtents3D> GaussTransform3D::view()
414 return impl_->data_.asView();
417 basic_mdspan<const float, dynamicExtents3D> GaussTransform3D::constView() const
419 return impl_->data_.asConstView();
422 GaussTransform3D::~GaussTransform3D() {}
424 GaussTransform3D::GaussTransform3D(const GaussTransform3D& other) : impl_(new Impl(*other.impl_)) {}
426 GaussTransform3D& GaussTransform3D::operator=(const GaussTransform3D& other)
428 *impl_ = *other.impl_;
432 GaussTransform3D::GaussTransform3D(GaussTransform3D&&) noexcept = default;
434 GaussTransform3D& GaussTransform3D::operator=(GaussTransform3D&&) noexcept = default;