| 1 |
/* |
| 2 |
* (C) Copyright 1996- 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 |
* |
| 7 |
* In applying this licence, ECMWF does not waive the privileges and immunities |
| 8 |
* granted to it by virtue of its status as an intergovernmental organisation nor |
| 9 |
* does it submit to any jurisdiction. |
| 10 |
*/ |
| 11 |
|
| 12 |
|
| 13 |
#include "atlas/interpolation/nonlinear/Missing.h" |
| 14 |
|
| 15 |
#include "eckit/types/FloatCompare.h" |
| 16 |
|
| 17 |
#include "atlas/field/MissingValue.h" |
| 18 |
#include "atlas/util/DataType.h" |
| 19 |
#include "atlas/util/Metadata.h" |
| 20 |
|
| 21 |
|
| 22 |
namespace atlas { |
| 23 |
namespace interpolation { |
| 24 |
namespace nonlinear { |
| 25 |
|
| 26 |
// Factory builders |
| 27 |
static NonLinearFactoryBuilder<MissingIfAllMissing> __nl1(MissingIfAllMissing::static_type()); |
| 28 |
static NonLinearFactoryBuilder<MissingIfAnyMissing> __nl2(MissingIfAnyMissing::static_type()); |
| 29 |
static NonLinearFactoryBuilder<MissingIfHeaviestMissing> __nl3(MissingIfHeaviestMissing::static_type()); |
| 30 |
|
| 31 |
// Deprecated factory entries with "-real32" and "-real64" suffix for backwards compatibility. |
| 32 |
// static NonLinearFactoryBuilder<MissingIfAllMissing> __nl1_real32(MissingIfAllMissing::static_type()+"-real32"); |
| 33 |
// static NonLinearFactoryBuilder<MissingIfAnyMissing> __nl2_real32(MissingIfAnyMissing::static_type()+"-real32"); |
| 34 |
// static NonLinearFactoryBuilder<MissingIfHeaviestMissing> __nl3_real32(MissingIfHeaviestMissing::static_type()+"-real32"); |
| 35 |
// static NonLinearFactoryBuilder<MissingIfAllMissing> __nl1_real64(MissingIfAllMissing::static_type()+"-real64"); |
| 36 |
// static NonLinearFactoryBuilder<MissingIfAnyMissing> __nl2_real64(MissingIfAnyMissing::static_type()+"-real64"); |
| 37 |
// static NonLinearFactoryBuilder<MissingIfHeaviestMissing> __nl3_real64(MissingIfHeaviestMissing::static_type()+"-real64"); |
| 38 |
|
| 39 |
namespace { |
| 40 |
struct force_link { |
| 41 |
template <typename M> |
| 42 |
void load_builder() { |
| 43 |
NonLinearFactoryBuilder<M>("tmp"); |
| 44 |
} |
| 45 |
force_link() { |
| 46 |
load_builder<MissingIfAllMissing>(); |
| 47 |
load_builder<MissingIfAnyMissing>(); |
| 48 |
load_builder<MissingIfHeaviestMissing>(); |
| 49 |
} |
| 50 |
}; |
| 51 |
|
| 52 |
std::string missing_value_type(const DataType& datatype, const field::MissingValue::Config& c) { |
| 53 |
std::string result; |
| 54 |
if (c.get("missing_value_type", result)) { |
| 55 |
result += "-" + datatype.str(); |
| 56 |
} |
| 57 |
return result; |
| 58 |
} |
| 59 |
|
| 60 |
} // namespace |
| 61 |
void force_link_missing() { |
| 62 |
static force_link static_linking; |
| 63 |
} |
| 64 |
|
| 65 |
bool Missing::applicable(const Field& f) const { |
| 66 |
return field::MissingValue(f); |
| 67 |
} |
| 68 |
|
| 69 |
bool MissingIfAllMissing::execute(NonLinear::Matrix& W, const Field& field) const { |
| 70 |
return execute(W, field.array(), field.metadata()); |
| 71 |
} |
| 72 |
|
| 73 |
bool MissingIfAllMissing::execute(NonLinear::Matrix& W, const array::Array& array, const Config& config) const { |
| 74 |
switch(array.datatype().kind()) { |
| 75 |
case (DataType::kind<double>()): return executeT<double>(W, array, config); |
| 76 |
case (DataType::kind<float>()): return executeT<float>(W, array, config); |
| 77 |
case (DataType::kind<int>()): return executeT<int>(W, array, config); |
| 78 |
case (DataType::kind<long>()): return executeT<long>(W, array, config); |
| 79 |
case (DataType::kind<unsigned long>()): return executeT<unsigned long>(W, array, config); |
| 80 |
default: ATLAS_NOTIMPLEMENTED; |
| 81 |
} |
| 82 |
} |
| 83 |
|
| 84 |
template<typename T> |
| 85 |
bool MissingIfAllMissing::executeT(NonLinear::Matrix& W, const array::Array& array, const Config& config) const { |
| 86 |
field::MissingValue mv(missing_value_type(array.datatype(), config), config); |
| 87 |
|
| 88 |
auto& missingValue = mv.ref(); |
| 89 |
|
| 90 |
ATLAS_ASSERT(array.rank() == 1); |
| 91 |
|
| 92 |
// NOTE only for scalars (for now) |
| 93 |
auto values = make_view_array_values<T, 1>(array); |
| 94 |
ATLAS_ASSERT(idx_t(W.cols()) == values.shape(0)); |
| 95 |
|
| 96 |
auto data = const_cast<Scalar*>(W.data()); |
| 97 |
bool modif = false; |
| 98 |
bool zeros = false; |
| 99 |
|
| 100 |
Size i = 0; |
| 101 |
Matrix::iterator it(W); |
| 102 |
for (Size r = 0; r < W.rows(); ++r) { |
| 103 |
const Matrix::iterator end = W.end(r); |
| 104 |
|
| 105 |
// count missing values, accumulate weights (disregarding missing values) |
| 106 |
size_t i_missing = i; |
| 107 |
size_t N_missing = 0; |
| 108 |
size_t N_entries = 0; |
| 109 |
Scalar sum = 0.; |
| 110 |
|
| 111 |
Matrix::iterator kt(it); |
| 112 |
Size k = i; |
| 113 |
for (; it != end; ++it, ++i, ++N_entries) { |
| 114 |
const bool miss = missingValue(values[it.col()]); |
| 115 |
|
| 116 |
if (miss) { |
| 117 |
++N_missing; |
| 118 |
i_missing = i; |
| 119 |
} |
| 120 |
else { |
| 121 |
sum += *it; |
| 122 |
} |
| 123 |
} |
| 124 |
|
| 125 |
// weights redistribution: zero-weight all missing values, linear re-weighting for the others; |
| 126 |
// the result is missing value if all values in row are missing |
| 127 |
if (N_missing > 0) { |
| 128 |
if (N_missing == N_entries || eckit::types::is_approximately_equal(sum, 0.)) { |
| 129 |
for (Size j = k; j < k + N_entries; ++j) { |
| 130 |
data[j] = j == i_missing ? 1. : 0.; |
| 131 |
} |
| 132 |
} |
| 133 |
else { |
| 134 |
const Scalar factor = 1. / sum; |
| 135 |
for (Size j = k; j < k + N_entries; ++j, ++kt) { |
| 136 |
if (missingValue(values[kt.col()])) { |
| 137 |
data[j] = 0.; |
| 138 |
zeros = true; |
| 139 |
} |
| 140 |
else { |
| 141 |
data[j] *= factor; |
| 142 |
} |
| 143 |
} |
| 144 |
} |
| 145 |
modif = true; |
| 146 |
} |
| 147 |
} |
| 148 |
|
| 149 |
if (zeros && missingValue.isnan()) { |
| 150 |
W.prune(0.); |
| 151 |
} |
| 152 |
|
| 153 |
return modif; |
| 154 |
} |
| 155 |
|
| 156 |
bool MissingIfAnyMissing::execute(NonLinear::Matrix& W, const Field& field) const { |
| 157 |
return execute(W, field.array(), field.metadata()); |
| 158 |
} |
| 159 |
|
| 160 |
bool MissingIfAnyMissing::execute(NonLinear::Matrix& W, const array::Array& array, const Config& config) const { |
| 161 |
switch(array.datatype().kind()) { |
| 162 |
case (DataType::kind<double>()): return executeT<double>(W, array, config); |
| 163 |
case (DataType::kind<float>()): return executeT<float>(W, array, config); |
| 164 |
case (DataType::kind<int>()): return executeT<int>(W, array, config); |
| 165 |
case (DataType::kind<long>()): return executeT<long>(W, array, config); |
| 166 |
case (DataType::kind<unsigned long>()): return executeT<unsigned long>(W, array, config); |
| 167 |
default: ATLAS_NOTIMPLEMENTED; |
| 168 |
} |
| 169 |
} |
| 170 |
|
| 171 |
|
| 172 |
template<typename T> |
| 173 |
bool MissingIfAnyMissing::executeT(NonLinear::Matrix& W, const array::Array& array, const Config& config) const { |
| 174 |
field::MissingValue mv(missing_value_type(array.datatype(), config), config); |
| 175 |
|
| 176 |
auto& missingValue = mv.ref(); |
| 177 |
|
| 178 |
// NOTE only for scalars (for now) |
| 179 |
auto values = make_view_array_values<T, 1>(array); |
| 180 |
ATLAS_ASSERT(idx_t(W.cols()) == values.size()); |
| 181 |
|
| 182 |
auto data = const_cast<Scalar*>(W.data()); |
| 183 |
bool modif = false; |
| 184 |
bool zeros = false; |
| 185 |
|
| 186 |
Size i = 0; |
| 187 |
Matrix::iterator it(W); |
| 188 |
for (Size r = 0; r < W.rows(); ++r) { |
| 189 |
const Matrix::iterator end = W.end(r); |
| 190 |
|
| 191 |
// count missing values, accumulate weights (disregarding missing values) |
| 192 |
size_t i_missing = i; |
| 193 |
size_t N_missing = 0; |
| 194 |
size_t N_entries = 0; |
| 195 |
|
| 196 |
Matrix::iterator kt(it); |
| 197 |
Size k = i; |
| 198 |
for (; it != end; ++it, ++i, ++N_entries) { |
| 199 |
const bool miss = missingValue(values[it.col()]); |
| 200 |
|
| 201 |
if (miss) { |
| 202 |
++N_missing; |
| 203 |
i_missing = i; |
| 204 |
} |
| 205 |
} |
| 206 |
|
| 207 |
// if any values in row are missing, force missing value |
| 208 |
if (N_missing > 0) { |
| 209 |
for (Size j = k; j < k + N_entries; ++j) { |
| 210 |
if (j == i_missing) { |
| 211 |
data[j] = 1.; |
| 212 |
} |
| 213 |
else { |
| 214 |
data[j] = 0.; |
| 215 |
zeros = true; |
| 216 |
} |
| 217 |
} |
| 218 |
modif = true; |
| 219 |
} |
| 220 |
} |
| 221 |
|
| 222 |
if (zeros && missingValue.isnan()) { |
| 223 |
W.prune(0.); |
| 224 |
} |
| 225 |
|
| 226 |
return modif; |
| 227 |
} |
| 228 |
|
| 229 |
bool MissingIfHeaviestMissing::execute(NonLinear::Matrix& W, const Field& field) const { |
| 230 |
return execute(W, field.array(), field.metadata()); |
| 231 |
} |
| 232 |
|
| 233 |
bool MissingIfHeaviestMissing::execute(NonLinear::Matrix& W, const array::Array& array, const Config& config) const { |
| 234 |
switch(array.datatype().kind()) { |
| 235 |
case (DataType::kind<double>()): return executeT<double>(W, array, config); |
| 236 |
case (DataType::kind<float>()): return executeT<float>(W, array, config); |
| 237 |
case (DataType::kind<int>()): return executeT<int>(W, array, config); |
| 238 |
case (DataType::kind<long>()): return executeT<long>(W, array, config); |
| 239 |
case (DataType::kind<unsigned long>()): return executeT<unsigned long>(W, array, config); |
| 240 |
default: ATLAS_NOTIMPLEMENTED; |
| 241 |
} |
| 242 |
} |
| 243 |
|
| 244 |
template<typename T> |
| 245 |
bool MissingIfHeaviestMissing::executeT(NonLinear::Matrix& W, const array::Array& array, const Config& config) const { |
| 246 |
field::MissingValue mv(missing_value_type(array.datatype(), config), config); |
| 247 |
auto& missingValue = mv.ref(); |
| 248 |
|
| 249 |
// NOTE only for scalars (for now) |
| 250 |
auto values = make_view_array_values<T, 1>(array); |
| 251 |
ATLAS_ASSERT(idx_t(W.cols()) == values.size()); |
| 252 |
|
| 253 |
auto data = const_cast<Scalar*>(W.data()); |
| 254 |
bool modif = false; |
| 255 |
bool zeros = false; |
| 256 |
|
| 257 |
Size i = 0; |
| 258 |
Matrix::iterator it(W); |
| 259 |
for (Size r = 0; r < W.rows(); ++r) { |
| 260 |
const Matrix::iterator end = W.end(r); |
| 261 |
|
| 262 |
// count missing values, accumulate weights (disregarding missing values) and find maximum weight in row |
| 263 |
size_t i_missing = i; |
| 264 |
size_t N_missing = 0; |
| 265 |
size_t N_entries = 0; |
| 266 |
Scalar sum = 0.; |
| 267 |
Scalar heaviest = -1.; |
| 268 |
bool heaviest_is_missing = false; |
| 269 |
|
| 270 |
Matrix::iterator kt(it); |
| 271 |
Size k = i; |
| 272 |
for (; it != end; ++it, ++i, ++N_entries) { |
| 273 |
const bool miss = missingValue(values[it.col()]); |
| 274 |
|
| 275 |
if (miss) { |
| 276 |
++N_missing; |
| 277 |
i_missing = i; |
| 278 |
} |
| 279 |
else { |
| 280 |
sum += *it; |
| 281 |
} |
| 282 |
|
| 283 |
if (heaviest < data[i]) { |
| 284 |
heaviest = data[i]; |
| 285 |
heaviest_is_missing = miss; |
| 286 |
} |
| 287 |
} |
| 288 |
|
| 289 |
// weights redistribution: zero-weight all missing values, linear re-weighting for the others; |
| 290 |
// if all values are missing, or the closest value is missing, force missing value |
| 291 |
if (N_missing > 0) { |
| 292 |
if (N_missing == N_entries || heaviest_is_missing || eckit::types::is_approximately_equal(sum, 0.)) { |
| 293 |
for (Size j = k; j < k + N_entries; ++j) { |
| 294 |
data[j] = j == i_missing ? 1. : 0.; |
| 295 |
} |
| 296 |
} |
| 297 |
else { |
| 298 |
const Scalar factor = 1. / sum; |
| 299 |
for (Size j = k; j < k + N_entries; ++j, ++kt) { |
| 300 |
if (missingValue(values[kt.col()])) { |
| 301 |
data[j] = 0.; |
| 302 |
zeros = true; |
| 303 |
} |
| 304 |
else { |
| 305 |
data[j] *= factor; |
| 306 |
} |
| 307 |
} |
| 308 |
} |
| 309 |
modif = true; |
| 310 |
} |
| 311 |
} |
| 312 |
|
| 313 |
if (zeros && missingValue.isnan()) { |
| 314 |
W.prune(0.); |
| 315 |
} |
| 316 |
|
| 317 |
return modif; |
| 318 |
} |
| 319 |
|
| 320 |
|
| 321 |
} // namespace nonlinear |
| 322 |
} // namespace interpolation |
| 323 |
} // namespace atlas |