| 1 |
/* |
| 2 |
* (C) Copyright 2013 ECMWF. |
| 3 |
* |
| 4 |
* This software is licensed under the terms of the Apache Licence Version 2.0 |
| 5 |
* which can be obtained at http://www.apache.org/licenses/LICENSE-2.0. |
| 6 |
* In applying this licence, ECMWF does not waive the privileges and immunities |
| 7 |
* granted to it by virtue of its status as an intergovernmental organisation |
| 8 |
* nor does it submit to any jurisdiction. and Interpolation |
| 9 |
*/ |
| 10 |
|
| 11 |
#pragma once |
| 12 |
|
| 13 |
#include <cmath> |
| 14 |
#include <limits> |
| 15 |
|
| 16 |
#include "CubicHorizontalLimiter.h" |
| 17 |
|
| 18 |
#include "eckit/linalg/Triplet.h" |
| 19 |
|
| 20 |
#include "atlas/array/ArrayView.h" |
| 21 |
#include "atlas/functionspace/StructuredColumns.h" |
| 22 |
#include "atlas/grid/Stencil.h" |
| 23 |
#include "atlas/grid/StencilComputer.h" |
| 24 |
#include "atlas/library/config.h" |
| 25 |
#include "atlas/runtime/Exception.h" |
| 26 |
#include "atlas/util/CoordinateEnums.h" |
| 27 |
#include "atlas/util/Point.h" |
| 28 |
|
| 29 |
namespace atlas { |
| 30 |
namespace interpolation { |
| 31 |
namespace method { |
| 32 |
|
| 33 |
class QuasiCubicHorizontalKernel { |
| 34 |
using Limiter = CubicHorizontalLimiter; |
| 35 |
using Triplet = eckit::linalg::Triplet; |
| 36 |
using Triplets = std::vector<Triplet>; |
| 37 |
|
| 38 |
public: |
| 39 |
QuasiCubicHorizontalKernel() = default; |
| 40 |
|
| 41 |
QuasiCubicHorizontalKernel(const functionspace::StructuredColumns& fs, |
| 42 |
const util::Config& config = util::NoConfig()) { |
| 43 |
src_ = fs; |
| 44 |
ATLAS_ASSERT(src_); |
| 45 |
ATLAS_ASSERT(src_.halo() >= 2); |
| 46 |
compute_horizontal_stencil_ = grid::ComputeHorizontalStencil(src_.grid(), stencil_width()); |
| 47 |
limiter_ = config.getBool("limiter", false); |
| 48 |
} |
| 49 |
|
| 50 |
protected: |
| 51 |
functionspace::StructuredColumns src_; |
| 52 |
grid::ComputeHorizontalStencil compute_horizontal_stencil_; |
| 53 |
bool limiter_{false}; |
| 54 |
|
| 55 |
public: |
| 56 |
static std::string className() { return "CubicHorizontalKernel"; } |
| 57 |
static constexpr idx_t stencil_width() { return 4; } |
| 58 |
static constexpr idx_t stencil_size() { return 12; } |
| 59 |
static constexpr idx_t stencil_halo() { |
| 60 |
return static_cast<idx_t>(static_cast<double>(stencil_width()) / 2. + 0.5); |
| 61 |
} |
| 62 |
|
| 63 |
public: |
| 64 |
using Stencil = grid::HorizontalStencil<4>; |
| 65 |
struct Weights { |
| 66 |
std::array<std::array<double, 4>, 4> weights_i; |
| 67 |
std::array<double, 4> weights_j; |
| 68 |
}; |
| 69 |
|
| 70 |
public: |
| 71 |
struct WorkSpace { |
| 72 |
Stencil stencil; |
| 73 |
Weights weights; |
| 74 |
}; |
| 75 |
|
| 76 |
template <typename stencil_t> |
| 77 |
void compute_stencil(const double x, const double y, stencil_t& stencil) const { |
| 78 |
compute_horizontal_stencil_(x, y, stencil); |
| 79 |
} |
| 80 |
|
| 81 |
template <typename stencil_t> |
| 82 |
void make_valid_stencil(double& x, double y, stencil_t& stencil, bool retry = true) const { |
| 83 |
for (idx_t j = 0; j < stencil_width(); ++j) { |
| 84 |
idx_t imin = stencil.i(0, j); |
| 85 |
idx_t imax = stencil.i(stencil_width()-1, j); |
| 86 |
if (imin < src_.i_begin_halo(stencil.j(j))) { |
| 87 |
if (retry ) { |
| 88 |
x += 360.; |
| 89 |
compute_stencil(x, y, stencil); |
| 90 |
return make_valid_stencil(x, y, stencil, false); |
| 91 |
} |
| 92 |
else { |
| 93 |
Log::error() << "Stencil out of bounds" << std::endl; |
| 94 |
ATLAS_THROW_EXCEPTION("stencil out of bounds"); |
| 95 |
} |
| 96 |
} |
| 97 |
if (imax >= src_.i_end_halo(stencil.j(j))) { |
| 98 |
if (retry ) { |
| 99 |
x -= 360.; |
| 100 |
compute_stencil(x, y, stencil); |
| 101 |
return make_valid_stencil(x, y, stencil, false); |
| 102 |
} |
| 103 |
else { |
| 104 |
Log::error() << "Stencil out of bounds" << std::endl; |
| 105 |
ATLAS_THROW_EXCEPTION("Stencil out of bounds"); |
| 106 |
} |
| 107 |
} |
| 108 |
} |
| 109 |
|
| 110 |
} |
| 111 |
|
| 112 |
template <typename weights_t> |
| 113 |
void compute_weights(double x, double y, weights_t& weights) const { |
| 114 |
Stencil stencil; |
| 115 |
compute_stencil(x, y, stencil); |
| 116 |
compute_weights(x, y, stencil, weights); |
| 117 |
} |
| 118 |
|
| 119 |
template <typename stencil_t, typename weights_t> |
| 120 |
void compute_weights(const double x, const double y, const stencil_t& stencil, weights_t& weights) const { |
| 121 |
PointXY P1, P2; |
| 122 |
std::array<double, 4> yvec; |
| 123 |
|
| 124 |
// Compute x-direction weights LINEAR for outer rows ( j = {0,3} ) |
| 125 |
for (idx_t j = 0; j < 4; j += 3) { |
| 126 |
auto& weights_i = weights.weights_i[j]; |
| 127 |
src_.compute_xy(stencil.i(1, j), stencil.j(j), P1); |
| 128 |
src_.compute_xy(stencil.i(2, j), stencil.j(j), P2); |
| 129 |
const double alpha = (P2.x() - x) / (P2.x() - P1.x()); |
| 130 |
weights_i[1] = alpha; |
| 131 |
weights_i[2] = 1. - alpha; |
| 132 |
yvec[j] = P1.y(); |
| 133 |
} |
| 134 |
|
| 135 |
// Compute x-direction weights CUBIC for inner rows ( j = {1,2} ) |
| 136 |
for (idx_t j = 1; j < 3; ++j) { |
| 137 |
auto& weights_i = weights.weights_i[j]; |
| 138 |
src_.compute_xy(stencil.i(1, j), stencil.j(j), P1); |
| 139 |
src_.compute_xy(stencil.i(2, j), stencil.j(j), P2); |
| 140 |
const double alpha = (P2.x() - x) / (P2.x() - P1.x()); |
| 141 |
const double alpha_sqr = alpha * alpha; |
| 142 |
const double two_minus_alpha = 2. - alpha; |
| 143 |
const double one_minus_alpha_sqr = 1. - alpha_sqr; |
| 144 |
weights_i[0] = -alpha * one_minus_alpha_sqr / 6.; |
| 145 |
weights_i[1] = 0.5 * alpha * (1. + alpha) * two_minus_alpha; |
| 146 |
weights_i[2] = 0.5 * one_minus_alpha_sqr * two_minus_alpha; |
| 147 |
weights_i[3] = 1. - weights_i[0] - weights_i[1] - weights_i[2]; |
| 148 |
yvec[j] = P1.y(); |
| 149 |
} |
| 150 |
|
| 151 |
// Compute weights in y-direction |
| 152 |
const double dl12 = yvec[0] - yvec[1]; |
| 153 |
const double dl13 = yvec[0] - yvec[2]; |
| 154 |
const double dl14 = yvec[0] - yvec[3]; |
| 155 |
const double dl23 = yvec[1] - yvec[2]; |
| 156 |
const double dl24 = yvec[1] - yvec[3]; |
| 157 |
const double dl34 = yvec[2] - yvec[3]; |
| 158 |
const double dcl1 = dl12 * dl13 * dl14; |
| 159 |
const double dcl2 = -dl12 * dl23 * dl24; |
| 160 |
const double dcl3 = dl13 * dl23 * dl34; |
| 161 |
|
| 162 |
const double dl1 = y - yvec[0]; |
| 163 |
const double dl2 = y - yvec[1]; |
| 164 |
const double dl3 = y - yvec[2]; |
| 165 |
const double dl4 = y - yvec[3]; |
| 166 |
|
| 167 |
auto& weights_j = weights.weights_j; |
| 168 |
weights_j[0] = (dl2 * dl3 * dl4) / dcl1; |
| 169 |
#if defined(_CRAYC) && ATLAS_BUILD_TYPE_RELEASE |
| 170 |
// prevents FE_INVALID somehow (tested with Cray 8.7) |
| 171 |
ATLAS_ASSERT(!std::isnan(weights_j[0])); |
| 172 |
#endif |
| 173 |
weights_j[1] = (dl1 * dl3 * dl4) / dcl2; |
| 174 |
weights_j[2] = (dl1 * dl2 * dl4) / dcl3; |
| 175 |
weights_j[3] = 1. - weights_j[0] - weights_j[1] - weights_j[2]; |
| 176 |
} |
| 177 |
|
| 178 |
template <typename stencil_t, typename weights_t, typename array_t> |
| 179 |
typename array_t::value_type interpolate(const stencil_t& stencil, const weights_t& weights, |
| 180 |
const array_t& input) const { |
| 181 |
using Value = typename array_t::value_type; |
| 182 |
|
| 183 |
std::array<std::array<idx_t, stencil_width()>, stencil_width()> index; |
| 184 |
const auto& weights_j = weights.weights_j; |
| 185 |
Value output = 0.; |
| 186 |
// LINEAR for outer rows ( j = {0,3} ) |
| 187 |
for (idx_t j = 0; j < 4; j += 3) { |
| 188 |
const auto& weights_i = weights.weights_i[j]; |
| 189 |
for (idx_t i : {1, 2}) { |
| 190 |
idx_t n = src_.index(stencil.i(i, j), stencil.j(j)); |
| 191 |
Value w = weights_i[i] * weights_j[j]; |
| 192 |
output += w * input[n]; |
| 193 |
index[j][i] = n; |
| 194 |
} |
| 195 |
} |
| 196 |
// CUBIC for inner rows ( j = {1,2} ) |
| 197 |
for (idx_t j = 1; j < 3; ++j) { |
| 198 |
const auto& weights_i = weights.weights_i[j]; |
| 199 |
for (idx_t i = 0; i < stencil_width(); ++i) { |
| 200 |
idx_t n = src_.index(stencil.i(i, j), stencil.j(j)); |
| 201 |
Value w = weights_i[i] * weights_j[j]; |
| 202 |
output += w * input[n]; |
| 203 |
index[j][i] = n; |
| 204 |
} |
| 205 |
} |
| 206 |
|
| 207 |
if (limiter_) { |
| 208 |
Limiter::limit(output, index, input); |
| 209 |
} |
| 210 |
return output; |
| 211 |
} |
| 212 |
|
| 213 |
template <typename stencil_t, typename weights_t, typename Value, int Rank> |
| 214 |
typename std::enable_if<(Rank == 1), void>::type interpolate(const stencil_t& stencil, const weights_t& weights, |
| 215 |
const array::ArrayView<const Value, Rank>& input, |
| 216 |
array::ArrayView<Value, Rank>& output, idx_t r) const { |
| 217 |
std::array<std::array<idx_t, stencil_width()>, stencil_width()> index; |
| 218 |
const auto& weights_j = weights.weights_j; |
| 219 |
output(r) = 0.; |
| 220 |
|
| 221 |
// LINEAR for outer rows ( j = {0,3} ) |
| 222 |
for (idx_t j = 0; j < 4; j += 3) { |
| 223 |
const auto& weights_i = weights.weights_i[j]; |
| 224 |
for (idx_t i = 1; i < 3; ++i) { // i = {1,2} |
| 225 |
idx_t n = src_.index(stencil.i(i, j), stencil.j(j)); |
| 226 |
Value w = weights_i[i] * weights_j[j]; |
| 227 |
output(r) += w * input[n]; |
| 228 |
index[j][i] = n; |
| 229 |
} |
| 230 |
} |
| 231 |
// CUBIC for inner rows ( j = {1,2} ) |
| 232 |
for (idx_t j = 1; j < 3; ++j) { |
| 233 |
const auto& weights_i = weights.weights_i[j]; |
| 234 |
for (idx_t i = 0; i < stencil_width(); ++i) { |
| 235 |
idx_t n = src_.index(stencil.i(i, j), stencil.j(j)); |
| 236 |
Value w = weights_i[i] * weights_j[j]; |
| 237 |
output(r) += w * input[n]; |
| 238 |
index[j][i] = n; |
| 239 |
} |
| 240 |
} |
| 241 |
|
| 242 |
if (limiter_) { |
| 243 |
Limiter::limit(index, input, output, r); |
| 244 |
} |
| 245 |
} |
| 246 |
|
| 247 |
template <typename stencil_t, typename weights_t, typename Value, int Rank> |
| 248 |
typename std::enable_if<(Rank == 2), void>::type interpolate(const stencil_t& stencil, const weights_t& weights, |
| 249 |
const array::ArrayView<const Value, Rank>& input, |
| 250 |
array::ArrayView<Value, Rank>& output, idx_t r) const { |
| 251 |
std::array<std::array<idx_t, stencil_width()>, stencil_width()> index; |
| 252 |
const auto& weights_j = weights.weights_j; |
| 253 |
const idx_t Nk = output.shape(1); |
| 254 |
for (idx_t k = 0; k < Nk; ++k) { |
| 255 |
output(r, k) = 0.; |
| 256 |
} |
| 257 |
|
| 258 |
// LINEAR for outer rows ( j = {0,3} ) |
| 259 |
for (idx_t j = 0; j < 4; j += 3) { |
| 260 |
const auto& weights_i = weights.weights_i[j]; |
| 261 |
for (idx_t i = 1; i < 3; ++i) { // i = {1,2} |
| 262 |
idx_t n = src_.index(stencil.i(i, j), stencil.j(j)); |
| 263 |
Value w = weights_i[i] * weights_j[j]; |
| 264 |
for (idx_t k = 0; k < Nk; ++k) { |
| 265 |
output(r, k) += w * input(n, k); |
| 266 |
} |
| 267 |
index[j][i] = n; |
| 268 |
} |
| 269 |
} |
| 270 |
// CUBIC for inner rows ( j = {1,2} ) |
| 271 |
for (idx_t j = 1; j < 3; ++j) { |
| 272 |
const auto& weights_i = weights.weights_i[j]; |
| 273 |
for (idx_t i = 0; i < stencil_width(); ++i) { |
| 274 |
idx_t n = src_.index(stencil.i(i, j), stencil.j(j)); |
| 275 |
Value w = weights_i[i] * weights_j[j]; |
| 276 |
for (idx_t k = 0; k < Nk; ++k) { |
| 277 |
output(r, k) += w * input(n, k); |
| 278 |
} |
| 279 |
index[j][i] = n; |
| 280 |
} |
| 281 |
} |
| 282 |
|
| 283 |
if (limiter_) { |
| 284 |
Limiter::limit(index, input, output, r); |
| 285 |
} |
| 286 |
} |
| 287 |
|
| 288 |
|
| 289 |
template <typename stencil_t, typename weights_t, typename Value, int Rank> |
| 290 |
typename std::enable_if<(Rank == 3), void>::type interpolate(const stencil_t& stencil, const weights_t& weights, |
| 291 |
const array::ArrayView<const Value, Rank>& input, |
| 292 |
array::ArrayView<Value, Rank>& output, idx_t r) const { |
| 293 |
std::array<std::array<idx_t, stencil_width()>, stencil_width()> index; |
| 294 |
const auto& weights_j = weights.weights_j; |
| 295 |
const idx_t Nk = output.shape(1); |
| 296 |
const idx_t Nl = output.shape(2); |
| 297 |
for (idx_t k = 0; k < Nk; ++k) { |
| 298 |
for (idx_t l = 0; l < Nl; ++l) { |
| 299 |
output(r, k, l) = 0.; |
| 300 |
} |
| 301 |
} |
| 302 |
|
| 303 |
// LINEAR for outer rows ( j = {0,3} ) |
| 304 |
for (idx_t j = 0; j < 4; j += 3) { |
| 305 |
const auto& weights_i = weights.weights_i[j]; |
| 306 |
for (idx_t i = 1; i < 3; ++i) { // i = {1,2} |
| 307 |
idx_t n = src_.index(stencil.i(i, j), stencil.j(j)); |
| 308 |
Value w = weights_i[i] * weights_j[j]; |
| 309 |
for (idx_t k = 0; k < Nk; ++k) { |
| 310 |
for (idx_t l = 0; l < Nl; ++l) { |
| 311 |
output(r, k, l) += w * input(n, k, l); |
| 312 |
} |
| 313 |
} |
| 314 |
index[j][i] = n; |
| 315 |
} |
| 316 |
} |
| 317 |
// CUBIC for inner rows ( j = {1,2} ) |
| 318 |
for (idx_t j = 1; j < 3; ++j) { |
| 319 |
const auto& weights_i = weights.weights_i[j]; |
| 320 |
for (idx_t i = 0; i < stencil_width(); ++i) { |
| 321 |
idx_t n = src_.index(stencil.i(i, j), stencil.j(j)); |
| 322 |
Value w = weights_i[i] * weights_j[j]; |
| 323 |
for (idx_t k = 0; k < Nk; ++k) { |
| 324 |
for (idx_t l = 0; l < Nl; ++l) { |
| 325 |
output(r, k, l) += w * input(n, k, l); |
| 326 |
} |
| 327 |
} |
| 328 |
index[j][i] = n; |
| 329 |
} |
| 330 |
} |
| 331 |
|
| 332 |
if (limiter_) { |
| 333 |
Limiter::limit(index, input, output, r); |
| 334 |
} |
| 335 |
} |
| 336 |
|
| 337 |
|
| 338 |
|
| 339 |
template <typename array_t> |
| 340 |
typename array_t::value_type operator()(double x, double y, const array_t& input) const { |
| 341 |
Stencil stencil; |
| 342 |
compute_stencil(x, y, stencil); |
| 343 |
Weights weights; |
| 344 |
compute_weights(x, y, stencil, weights); |
| 345 |
make_valid_stencil(x, y, stencil); |
| 346 |
return interpolate(stencil, weights, input); |
| 347 |
} |
| 348 |
|
| 349 |
template <typename array_t> |
| 350 |
typename array_t::value_type interpolate(PointXY p, const array_t& input, WorkSpace& ws) const { |
| 351 |
compute_stencil(p.x(), p.y(), ws.stencil); |
| 352 |
compute_weights(p.x(), p.y(), ws.stencil, ws.weights); |
| 353 |
make_valid_stencil(p.x(), p.y(), ws.stencil); |
| 354 |
return interpolate(ws.stencil, ws.weights, input); |
| 355 |
} |
| 356 |
|
| 357 |
// Thread private workspace |
| 358 |
Triplets compute_triplets(const idx_t row, const double x, const double y, WorkSpace& ws) const { |
| 359 |
Triplets triplets; |
| 360 |
triplets.reserve(stencil_size()); |
| 361 |
insert_triplets(row, x, y, triplets, ws); |
| 362 |
return triplets; |
| 363 |
} |
| 364 |
|
| 365 |
Triplets reserve_triplets(size_t N) { |
| 366 |
Triplets triplets; |
| 367 |
triplets.reserve(N * stencil_size()); |
| 368 |
return triplets; |
| 369 |
} |
| 370 |
|
| 371 |
Triplets allocate_triplets(size_t N) { return Triplets(N * stencil_size()); } |
| 372 |
|
| 373 |
void insert_triplets(const idx_t row, const PointXY& p, Triplets& triplets, WorkSpace& ws) const { |
| 374 |
insert_triplets(row, p.x(), p.y(), triplets, ws); |
| 375 |
} |
| 376 |
|
| 377 |
void insert_triplets(const idx_t row, double x, double y, Triplets& triplets, WorkSpace& ws) const { |
| 378 |
compute_stencil(x, y, ws.stencil); |
| 379 |
compute_weights(x, y, ws.stencil, ws.weights); |
| 380 |
|
| 381 |
make_valid_stencil(x, y, ws.stencil); |
| 382 |
const auto& wj = ws.weights.weights_j; |
| 383 |
|
| 384 |
idx_t pos = row * stencil_size(); |
| 385 |
|
| 386 |
// LINEAR for outer rows ( j = {0,3} ) |
| 387 |
for (idx_t j = 0; j < 4; j += 3) { |
| 388 |
const auto& wi = ws.weights.weights_i[j]; |
| 389 |
for (idx_t i = 1; i < 3; ++i) { // i = {1,2} |
| 390 |
idx_t col = src_.index(ws.stencil.i(i, j), ws.stencil.j(j)); |
| 391 |
double w = wi[i] * wj[j]; |
| 392 |
triplets[pos++] = Triplet(row, col, w); |
| 393 |
} |
| 394 |
} |
| 395 |
|
| 396 |
// CUBIC for inner rows ( j = {1,2} ) |
| 397 |
for (idx_t j = 1; j < 3; ++j) { |
| 398 |
const auto& wi = ws.weights.weights_i[j]; |
| 399 |
for (idx_t i = 0; i < stencil_width(); ++i) { |
| 400 |
idx_t col = src_.index(ws.stencil.i(i, j), ws.stencil.j(j)); |
| 401 |
double w = wi[i] * wj[j]; |
| 402 |
triplets[pos++] = Triplet(row, col, w); |
| 403 |
} |
| 404 |
} |
| 405 |
} |
| 406 |
}; |
| 407 |
|
| 408 |
} // namespace method |
| 409 |
} // namespace interpolation |
| 410 |
} // namespace atlas |