| 1 | // |
| 2 | // Copyright (c) 2000-2002 |
| 3 | // Joerg Walter, Mathias Koch |
| 4 | // |
| 5 | // Distributed under the Boost Software License, Version 1.0. (See |
| 6 | // accompanying file LICENSE_1_0.txt or copy at |
| 7 | // http://www.boost.org/LICENSE_1_0.txt) |
| 8 | // |
| 9 | // The authors gratefully acknowledge the support of |
| 10 | // GeNeSys mbH & Co. KG in producing this work. |
| 11 | // |
| 12 | |
| 13 | #ifndef _BOOST_UBLAS_OPERATION_ |
| 14 | #define _BOOST_UBLAS_OPERATION_ |
| 15 | |
| 16 | #include <boost/numeric/ublas/matrix_proxy.hpp> |
| 17 | |
| 18 | /** \file operation.hpp |
| 19 | * \brief This file contains some specialized products. |
| 20 | */ |
| 21 | |
| 22 | // axpy-based products |
| 23 | // Alexei Novakov had a lot of ideas to improve these. Thanks. |
| 24 | // Hendrik Kueck proposed some new kernel. Thanks again. |
| 25 | |
| 26 | namespace boost { namespace numeric { namespace ublas { |
| 27 | |
| 28 | template<class V, class T1, class L1, class IA1, class TA1, class E2> |
| 29 | BOOST_UBLAS_INLINE |
| 30 | V & |
| 31 | axpy_prod (const compressed_matrix<T1, L1, 0, IA1, TA1> &e1, |
| 32 | const vector_expression<E2> &e2, |
| 33 | V &v, row_major_tag) { |
| 34 | typedef typename V::size_type size_type; |
| 35 | typedef typename V::value_type value_type; |
| 36 | |
| 37 | for (size_type i = 0; i < e1.filled1 () -1; ++ i) { |
| 38 | size_type begin = e1.index1_data () [i]; |
| 39 | size_type end = e1.index1_data () [i + 1]; |
| 40 | value_type t (v (i)); |
| 41 | for (size_type j = begin; j < end; ++ j) |
| 42 | t += e1.value_data () [j] * e2 () (e1.index2_data () [j]); |
| 43 | v (i) = t; |
| 44 | } |
| 45 | return v; |
| 46 | } |
| 47 | |
| 48 | template<class V, class T1, class L1, class IA1, class TA1, class E2> |
| 49 | BOOST_UBLAS_INLINE |
| 50 | V & |
| 51 | axpy_prod (const compressed_matrix<T1, L1, 0, IA1, TA1> &e1, |
| 52 | const vector_expression<E2> &e2, |
| 53 | V &v, column_major_tag) { |
| 54 | typedef typename V::size_type size_type; |
| 55 | |
| 56 | for (size_type j = 0; j < e1.filled1 () -1; ++ j) { |
| 57 | size_type begin = e1.index1_data () [j]; |
| 58 | size_type end = e1.index1_data () [j + 1]; |
| 59 | for (size_type i = begin; i < end; ++ i) |
| 60 | v (e1.index2_data () [i]) += e1.value_data () [i] * e2 () (j); |
| 61 | } |
| 62 | return v; |
| 63 | } |
| 64 | |
| 65 | // Dispatcher |
| 66 | template<class V, class T1, class L1, class IA1, class TA1, class E2> |
| 67 | BOOST_UBLAS_INLINE |
| 68 | V & |
| 69 | axpy_prod (const compressed_matrix<T1, L1, 0, IA1, TA1> &e1, |
| 70 | const vector_expression<E2> &e2, |
| 71 | V &v, bool init = true) { |
| 72 | typedef typename V::value_type value_type; |
| 73 | typedef typename L1::orientation_category orientation_category; |
| 74 | |
| 75 | if (init) |
| 76 | v.assign (zero_vector<value_type> (e1.size1 ())); |
| 77 | #if BOOST_UBLAS_TYPE_CHECK |
| 78 | vector<value_type> cv (v); |
| 79 | typedef typename type_traits<value_type>::real_type real_type; |
| 80 | real_type verrorbound (norm_1 (v) + norm_1 (e1) * norm_1 (e2)); |
| 81 | indexing_vector_assign<scalar_plus_assign> (cv, prod (e1, e2)); |
| 82 | #endif |
| 83 | axpy_prod (e1, e2, v, orientation_category ()); |
| 84 | #if BOOST_UBLAS_TYPE_CHECK |
| 85 | BOOST_UBLAS_CHECK (norm_1 (v - cv) <= 2 * std::numeric_limits<real_type>::epsilon () * verrorbound, internal_logic ()); |
| 86 | #endif |
| 87 | return v; |
| 88 | } |
| 89 | template<class V, class T1, class L1, class IA1, class TA1, class E2> |
| 90 | BOOST_UBLAS_INLINE |
| 91 | V |
| 92 | axpy_prod (const compressed_matrix<T1, L1, 0, IA1, TA1> &e1, |
| 93 | const vector_expression<E2> &e2) { |
| 94 | typedef V vector_type; |
| 95 | |
| 96 | vector_type v (e1.size1 ()); |
| 97 | return axpy_prod (e1, e2, v, true); |
| 98 | } |
| 99 | |
| 100 | template<class V, class T1, class L1, class IA1, class TA1, class E2> |
| 101 | BOOST_UBLAS_INLINE |
| 102 | V & |
| 103 | axpy_prod (const coordinate_matrix<T1, L1, 0, IA1, TA1> &e1, |
| 104 | const vector_expression<E2> &e2, |
| 105 | V &v, bool init = true) { |
| 106 | typedef typename V::size_type size_type; |
| 107 | typedef typename V::value_type value_type; |
| 108 | typedef L1 layout_type; |
| 109 | |
| 110 | size_type size1 = e1.size1(); |
| 111 | size_type size2 = e1.size2(); |
| 112 | |
| 113 | if (init) { |
| 114 | noalias(v) = zero_vector<value_type>(size1); |
| 115 | } |
| 116 | |
| 117 | for (size_type i = 0; i < e1.nnz(); ++i) { |
| 118 | size_type row_index = layout_type::index_M( e1.index1_data () [i], e1.index2_data () [i] ); |
| 119 | size_type col_index = layout_type::index_m( e1.index1_data () [i], e1.index2_data () [i] ); |
| 120 | v( row_index ) += e1.value_data () [i] * e2 () (col_index); |
| 121 | } |
| 122 | return v; |
| 123 | } |
| 124 | |
| 125 | template<class V, class E1, class E2> |
| 126 | BOOST_UBLAS_INLINE |
| 127 | V & |
| 128 | axpy_prod (const matrix_expression<E1> &e1, |
| 129 | const vector_expression<E2> &e2, |
| 130 | V &v, packed_random_access_iterator_tag, row_major_tag) { |
| 131 | typedef const E1 expression1_type; |
| 132 | typedef typename V::size_type size_type; |
| 133 | |
| 134 | typename expression1_type::const_iterator1 it1 (e1 ().begin1 ()); |
| 135 | typename expression1_type::const_iterator1 it1_end (e1 ().end1 ()); |
| 136 | while (it1 != it1_end) { |
| 137 | size_type index1 (it1.index1 ()); |
| 138 | #ifndef BOOST_UBLAS_NO_NESTED_CLASS_RELATION |
| 139 | typename expression1_type::const_iterator2 it2 (it1.begin ()); |
| 140 | typename expression1_type::const_iterator2 it2_end (it1.end ()); |
| 141 | #else |
| 142 | typename expression1_type::const_iterator2 it2 (boost::numeric::ublas::begin (it1, iterator1_tag ())); |
| 143 | typename expression1_type::const_iterator2 it2_end (boost::numeric::ublas::end (it1, iterator1_tag ())); |
| 144 | #endif |
| 145 | while (it2 != it2_end) { |
| 146 | v (index1) += *it2 * e2 () (it2.index2 ()); |
| 147 | ++ it2; |
| 148 | } |
| 149 | ++ it1; |
| 150 | } |
| 151 | return v; |
| 152 | } |
| 153 | |
| 154 | template<class V, class E1, class E2> |
| 155 | BOOST_UBLAS_INLINE |
| 156 | V & |
| 157 | axpy_prod (const matrix_expression<E1> &e1, |
| 158 | const vector_expression<E2> &e2, |
| 159 | V &v, packed_random_access_iterator_tag, column_major_tag) { |
| 160 | typedef const E1 expression1_type; |
| 161 | typedef typename V::size_type size_type; |
| 162 | |
| 163 | typename expression1_type::const_iterator2 it2 (e1 ().begin2 ()); |
| 164 | typename expression1_type::const_iterator2 it2_end (e1 ().end2 ()); |
| 165 | while (it2 != it2_end) { |
| 166 | size_type index2 (it2.index2 ()); |
| 167 | #ifndef BOOST_UBLAS_NO_NESTED_CLASS_RELATION |
| 168 | typename expression1_type::const_iterator1 it1 (it2.begin ()); |
| 169 | typename expression1_type::const_iterator1 it1_end (it2.end ()); |
| 170 | #else |
| 171 | typename expression1_type::const_iterator1 it1 (boost::numeric::ublas::begin (it2, iterator2_tag ())); |
| 172 | typename expression1_type::const_iterator1 it1_end (boost::numeric::ublas::end (it2, iterator2_tag ())); |
| 173 | #endif |
| 174 | while (it1 != it1_end) { |
| 175 | v (it1.index1 ()) += *it1 * e2 () (index2); |
| 176 | ++ it1; |
| 177 | } |
| 178 | ++ it2; |
| 179 | } |
| 180 | return v; |
| 181 | } |
| 182 | |
| 183 | template<class V, class E1, class E2> |
| 184 | BOOST_UBLAS_INLINE |
| 185 | V & |
| 186 | axpy_prod (const matrix_expression<E1> &e1, |
| 187 | const vector_expression<E2> &e2, |
| 188 | V &v, sparse_bidirectional_iterator_tag) { |
| 189 | typedef const E2 expression2_type; |
| 190 | |
| 191 | typename expression2_type::const_iterator it (e2 ().begin ()); |
| 192 | typename expression2_type::const_iterator it_end (e2 ().end ()); |
| 193 | while (it != it_end) { |
| 194 | v.plus_assign (column (e1 (), it.index ()) * *it); |
| 195 | ++ it; |
| 196 | } |
| 197 | return v; |
| 198 | } |
| 199 | |
| 200 | // Dispatcher |
| 201 | template<class V, class E1, class E2> |
| 202 | BOOST_UBLAS_INLINE |
| 203 | V & |
| 204 | axpy_prod (const matrix_expression<E1> &e1, |
| 205 | const vector_expression<E2> &e2, |
| 206 | V &v, packed_random_access_iterator_tag) { |
| 207 | typedef typename E1::orientation_category orientation_category; |
| 208 | return axpy_prod (e1, e2, v, packed_random_access_iterator_tag (), orientation_category ()); |
| 209 | } |
| 210 | |
| 211 | |
| 212 | /** \brief computes <tt>v += A x</tt> or <tt>v = A x</tt> in an |
| 213 | optimized fashion. |
| 214 | |
| 215 | \param e1 the matrix expression \c A |
| 216 | \param e2 the vector expression \c x |
| 217 | \param v the result vector \c v |
| 218 | \param init a boolean parameter |
| 219 | |
| 220 | <tt>axpy_prod(A, x, v, init)</tt> implements the well known |
| 221 | axpy-product. Setting \a init to \c true is equivalent to call |
| 222 | <tt>v.clear()</tt> before <tt>axpy_prod</tt>. Currently \a init |
| 223 | defaults to \c true, but this may change in the future. |
| 224 | |
| 225 | Up to now there are some specialisation for compressed |
| 226 | matrices that give a large speed up compared to prod. |
| 227 | |
| 228 | \ingroup blas2 |
| 229 | |
| 230 | \internal |
| 231 | |
| 232 | template parameters: |
| 233 | \param V type of the result vector \c v |
| 234 | \param E1 type of a matrix expression \c A |
| 235 | \param E2 type of a vector expression \c x |
| 236 | */ |
| 237 | template<class V, class E1, class E2> |
| 238 | BOOST_UBLAS_INLINE |
| 239 | V & |
| 240 | axpy_prod (const matrix_expression<E1> &e1, |
| 241 | const vector_expression<E2> &e2, |
| 242 | V &v, bool init = true) { |
| 243 | typedef typename V::value_type value_type; |
| 244 | typedef typename E2::const_iterator::iterator_category iterator_category; |
| 245 | |
| 246 | if (init) |
| 247 | v.assign (zero_vector<value_type> (e1 ().size1 ())); |
| 248 | #if BOOST_UBLAS_TYPE_CHECK |
| 249 | vector<value_type> cv (v); |
| 250 | typedef typename type_traits<value_type>::real_type real_type; |
| 251 | real_type verrorbound (norm_1 (v) + norm_1 (e1) * norm_1 (e2)); |
| 252 | indexing_vector_assign<scalar_plus_assign> (cv, prod (e1, e2)); |
| 253 | #endif |
| 254 | axpy_prod (e1, e2, v, iterator_category ()); |
| 255 | #if BOOST_UBLAS_TYPE_CHECK |
| 256 | BOOST_UBLAS_CHECK (norm_1 (v - cv) <= 2 * std::numeric_limits<real_type>::epsilon () * verrorbound, internal_logic ()); |
| 257 | #endif |
| 258 | return v; |
| 259 | } |
| 260 | template<class V, class E1, class E2> |
| 261 | BOOST_UBLAS_INLINE |
| 262 | V |
| 263 | axpy_prod (const matrix_expression<E1> &e1, |
| 264 | const vector_expression<E2> &e2) { |
| 265 | typedef V vector_type; |
| 266 | |
| 267 | vector_type v (e1 ().size1 ()); |
| 268 | return axpy_prod (e1, e2, v, true); |
| 269 | } |
| 270 | |
| 271 | template<class V, class E1, class T2, class IA2, class TA2> |
| 272 | BOOST_UBLAS_INLINE |
| 273 | V & |
| 274 | axpy_prod (const vector_expression<E1> &e1, |
| 275 | const compressed_matrix<T2, column_major, 0, IA2, TA2> &e2, |
| 276 | V &v, column_major_tag) { |
| 277 | typedef typename V::size_type size_type; |
| 278 | typedef typename V::value_type value_type; |
| 279 | |
| 280 | for (size_type j = 0; j < e2.filled1 () -1; ++ j) { |
| 281 | size_type begin = e2.index1_data () [j]; |
| 282 | size_type end = e2.index1_data () [j + 1]; |
| 283 | value_type t (v (j)); |
| 284 | for (size_type i = begin; i < end; ++ i) |
| 285 | t += e2.value_data () [i] * e1 () (e2.index2_data () [i]); |
| 286 | v (j) = t; |
| 287 | } |
| 288 | return v; |
| 289 | } |
| 290 | |
| 291 | template<class V, class E1, class T2, class IA2, class TA2> |
| 292 | BOOST_UBLAS_INLINE |
| 293 | V & |
| 294 | axpy_prod (const vector_expression<E1> &e1, |
| 295 | const compressed_matrix<T2, row_major, 0, IA2, TA2> &e2, |
| 296 | V &v, row_major_tag) { |
| 297 | typedef typename V::size_type size_type; |
| 298 | |
| 299 | for (size_type i = 0; i < e2.filled1 () -1; ++ i) { |
| 300 | size_type begin = e2.index1_data () [i]; |
| 301 | size_type end = e2.index1_data () [i + 1]; |
| 302 | for (size_type j = begin; j < end; ++ j) |
| 303 | v (e2.index2_data () [j]) += e2.value_data () [j] * e1 () (i); |
| 304 | } |
| 305 | return v; |
| 306 | } |
| 307 | |
| 308 | // Dispatcher |
| 309 | template<class V, class E1, class T2, class L2, class IA2, class TA2> |
| 310 | BOOST_UBLAS_INLINE |
| 311 | V & |
| 312 | axpy_prod (const vector_expression<E1> &e1, |
| 313 | const compressed_matrix<T2, L2, 0, IA2, TA2> &e2, |
| 314 | V &v, bool init = true) { |
| 315 | typedef typename V::value_type value_type; |
| 316 | typedef typename L2::orientation_category orientation_category; |
| 317 | |
| 318 | if (init) |
| 319 | v.assign (zero_vector<value_type> (e2.size2 ())); |
| 320 | #if BOOST_UBLAS_TYPE_CHECK |
| 321 | vector<value_type> cv (v); |
| 322 | typedef typename type_traits<value_type>::real_type real_type; |
| 323 | real_type verrorbound (norm_1 (v) + norm_1 (e1) * norm_1 (e2)); |
| 324 | indexing_vector_assign<scalar_plus_assign> (cv, prod (e1, e2)); |
| 325 | #endif |
| 326 | axpy_prod (e1, e2, v, orientation_category ()); |
| 327 | #if BOOST_UBLAS_TYPE_CHECK |
| 328 | BOOST_UBLAS_CHECK (norm_1 (v - cv) <= 2 * std::numeric_limits<real_type>::epsilon () * verrorbound, internal_logic ()); |
| 329 | #endif |
| 330 | return v; |
| 331 | } |
| 332 | template<class V, class E1, class T2, class L2, class IA2, class TA2> |
| 333 | BOOST_UBLAS_INLINE |
| 334 | V |
| 335 | axpy_prod (const vector_expression<E1> &e1, |
| 336 | const compressed_matrix<T2, L2, 0, IA2, TA2> &e2) { |
| 337 | typedef V vector_type; |
| 338 | |
| 339 | vector_type v (e2.size2 ()); |
| 340 | return axpy_prod (e1, e2, v, true); |
| 341 | } |
| 342 | |
| 343 | template<class V, class E1, class E2> |
| 344 | BOOST_UBLAS_INLINE |
| 345 | V & |
| 346 | axpy_prod (const vector_expression<E1> &e1, |
| 347 | const matrix_expression<E2> &e2, |
| 348 | V &v, packed_random_access_iterator_tag, column_major_tag) { |
| 349 | typedef const E2 expression2_type; |
| 350 | typedef typename V::size_type size_type; |
| 351 | |
| 352 | typename expression2_type::const_iterator2 it2 (e2 ().begin2 ()); |
| 353 | typename expression2_type::const_iterator2 it2_end (e2 ().end2 ()); |
| 354 | while (it2 != it2_end) { |
| 355 | size_type index2 (it2.index2 ()); |
| 356 | #ifndef BOOST_UBLAS_NO_NESTED_CLASS_RELATION |
| 357 | typename expression2_type::const_iterator1 it1 (it2.begin ()); |
| 358 | typename expression2_type::const_iterator1 it1_end (it2.end ()); |
| 359 | #else |
| 360 | typename expression2_type::const_iterator1 it1 (boost::numeric::ublas::begin (it2, iterator2_tag ())); |
| 361 | typename expression2_type::const_iterator1 it1_end (boost::numeric::ublas::end (it2, iterator2_tag ())); |
| 362 | #endif |
| 363 | while (it1 != it1_end) { |
| 364 | v (index2) += *it1 * e1 () (it1.index1 ()); |
| 365 | ++ it1; |
| 366 | } |
| 367 | ++ it2; |
| 368 | } |
| 369 | return v; |
| 370 | } |
| 371 | |
| 372 | template<class V, class E1, class E2> |
| 373 | BOOST_UBLAS_INLINE |
| 374 | V & |
| 375 | axpy_prod (const vector_expression<E1> &e1, |
| 376 | const matrix_expression<E2> &e2, |
| 377 | V &v, packed_random_access_iterator_tag, row_major_tag) { |
| 378 | typedef const E2 expression2_type; |
| 379 | typedef typename V::size_type size_type; |
| 380 | |
| 381 | typename expression2_type::const_iterator1 it1 (e2 ().begin1 ()); |
| 382 | typename expression2_type::const_iterator1 it1_end (e2 ().end1 ()); |
| 383 | while (it1 != it1_end) { |
| 384 | size_type index1 (it1.index1 ()); |
| 385 | #ifndef BOOST_UBLAS_NO_NESTED_CLASS_RELATION |
| 386 | typename expression2_type::const_iterator2 it2 (it1.begin ()); |
| 387 | typename expression2_type::const_iterator2 it2_end (it1.end ()); |
| 388 | #else |
| 389 | typename expression2_type::const_iterator2 it2 (boost::numeric::ublas::begin (it1, iterator1_tag ())); |
| 390 | typename expression2_type::const_iterator2 it2_end (boost::numeric::ublas::end (it1, iterator1_tag ())); |
| 391 | #endif |
| 392 | while (it2 != it2_end) { |
| 393 | v (it2.index2 ()) += *it2 * e1 () (index1); |
| 394 | ++ it2; |
| 395 | } |
| 396 | ++ it1; |
| 397 | } |
| 398 | return v; |
| 399 | } |
| 400 | |
| 401 | template<class V, class E1, class E2> |
| 402 | BOOST_UBLAS_INLINE |
| 403 | V & |
| 404 | axpy_prod (const vector_expression<E1> &e1, |
| 405 | const matrix_expression<E2> &e2, |
| 406 | V &v, sparse_bidirectional_iterator_tag) { |
| 407 | typedef const E1 expression1_type; |
| 408 | |
| 409 | typename expression1_type::const_iterator it (e1 ().begin ()); |
| 410 | typename expression1_type::const_iterator it_end (e1 ().end ()); |
| 411 | while (it != it_end) { |
| 412 | v.plus_assign (*it * row (e2 (), it.index ())); |
| 413 | ++ it; |
| 414 | } |
| 415 | return v; |
| 416 | } |
| 417 | |
| 418 | // Dispatcher |
| 419 | template<class V, class E1, class E2> |
| 420 | BOOST_UBLAS_INLINE |
| 421 | V & |
| 422 | axpy_prod (const vector_expression<E1> &e1, |
| 423 | const matrix_expression<E2> &e2, |
| 424 | V &v, packed_random_access_iterator_tag) { |
| 425 | typedef typename E2::orientation_category orientation_category; |
| 426 | return axpy_prod (e1, e2, v, packed_random_access_iterator_tag (), orientation_category ()); |
| 427 | } |
| 428 | |
| 429 | |
| 430 | /** \brief computes <tt>v += A<sup>T</sup> x</tt> or <tt>v = A<sup>T</sup> x</tt> in an |
| 431 | optimized fashion. |
| 432 | |
| 433 | \param e1 the vector expression \c x |
| 434 | \param e2 the matrix expression \c A |
| 435 | \param v the result vector \c v |
| 436 | \param init a boolean parameter |
| 437 | |
| 438 | <tt>axpy_prod(x, A, v, init)</tt> implements the well known |
| 439 | axpy-product. Setting \a init to \c true is equivalent to call |
| 440 | <tt>v.clear()</tt> before <tt>axpy_prod</tt>. Currently \a init |
| 441 | defaults to \c true, but this may change in the future. |
| 442 | |
| 443 | Up to now there are some specialisation for compressed |
| 444 | matrices that give a large speed up compared to prod. |
| 445 | |
| 446 | \ingroup blas2 |
| 447 | |
| 448 | \internal |
| 449 | |
| 450 | template parameters: |
| 451 | \param V type of the result vector \c v |
| 452 | \param E1 type of a vector expression \c x |
| 453 | \param E2 type of a matrix expression \c A |
| 454 | */ |
| 455 | template<class V, class E1, class E2> |
| 456 | BOOST_UBLAS_INLINE |
| 457 | V & |
| 458 | axpy_prod (const vector_expression<E1> &e1, |
| 459 | const matrix_expression<E2> &e2, |
| 460 | V &v, bool init = true) { |
| 461 | typedef typename V::value_type value_type; |
| 462 | typedef typename E1::const_iterator::iterator_category iterator_category; |
| 463 | |
| 464 | if (init) |
| 465 | v.assign (zero_vector<value_type> (e2 ().size2 ())); |
| 466 | #if BOOST_UBLAS_TYPE_CHECK |
| 467 | vector<value_type> cv (v); |
| 468 | typedef typename type_traits<value_type>::real_type real_type; |
| 469 | real_type verrorbound (norm_1 (v) + norm_1 (e1) * norm_1 (e2)); |
| 470 | indexing_vector_assign<scalar_plus_assign> (cv, prod (e1, e2)); |
| 471 | #endif |
| 472 | axpy_prod (e1, e2, v, iterator_category ()); |
| 473 | #if BOOST_UBLAS_TYPE_CHECK |
| 474 | BOOST_UBLAS_CHECK (norm_1 (v - cv) <= 2 * std::numeric_limits<real_type>::epsilon () * verrorbound, internal_logic ()); |
| 475 | #endif |
| 476 | return v; |
| 477 | } |
| 478 | template<class V, class E1, class E2> |
| 479 | BOOST_UBLAS_INLINE |
| 480 | V |
| 481 | axpy_prod (const vector_expression<E1> &e1, |
| 482 | const matrix_expression<E2> &e2) { |
| 483 | typedef V vector_type; |
| 484 | |
| 485 | vector_type v (e2 ().size2 ()); |
| 486 | return axpy_prod (e1, e2, v, true); |
| 487 | } |
| 488 | |
| 489 | template<class M, class E1, class E2, class TRI> |
| 490 | BOOST_UBLAS_INLINE |
| 491 | M & |
| 492 | axpy_prod (const matrix_expression<E1> &e1, |
| 493 | const matrix_expression<E2> &e2, |
| 494 | M &m, TRI, |
| 495 | dense_proxy_tag, row_major_tag) { |
| 496 | |
| 497 | typedef typename M::size_type size_type; |
| 498 | |
| 499 | #if BOOST_UBLAS_TYPE_CHECK |
| 500 | typedef typename M::value_type value_type; |
| 501 | matrix<value_type, row_major> cm (m); |
| 502 | typedef typename type_traits<value_type>::real_type real_type; |
| 503 | real_type merrorbound (norm_1 (m) + norm_1 (e1) * norm_1 (e2)); |
| 504 | indexing_matrix_assign<scalar_plus_assign> (cm, prod (e1, e2), row_major_tag ()); |
| 505 | #endif |
| 506 | size_type size1 (e1 ().size1 ()); |
| 507 | size_type size2 (e1 ().size2 ()); |
| 508 | for (size_type i = 0; i < size1; ++ i) |
| 509 | for (size_type j = 0; j < size2; ++ j) |
| 510 | row (m, i).plus_assign (e1 () (i, j) * row (e2 (), j)); |
| 511 | #if BOOST_UBLAS_TYPE_CHECK |
| 512 | BOOST_UBLAS_CHECK (norm_1 (m - cm) <= 2 * std::numeric_limits<real_type>::epsilon () * merrorbound, internal_logic ()); |
| 513 | #endif |
| 514 | return m; |
| 515 | } |
| 516 | template<class M, class E1, class E2, class TRI> |
| 517 | BOOST_UBLAS_INLINE |
| 518 | M & |
| 519 | axpy_prod (const matrix_expression<E1> &e1, |
| 520 | const matrix_expression<E2> &e2, |
| 521 | M &m, TRI, |
| 522 | sparse_proxy_tag, row_major_tag) { |
| 523 | |
| 524 | typedef TRI triangular_restriction; |
| 525 | typedef const E1 expression1_type; |
| 526 | typedef const E2 expression2_type; |
| 527 | |
| 528 | #if BOOST_UBLAS_TYPE_CHECK |
| 529 | typedef typename M::value_type value_type; |
| 530 | matrix<value_type, row_major> cm (m); |
| 531 | typedef typename type_traits<value_type>::real_type real_type; |
| 532 | real_type merrorbound (norm_1 (m) + norm_1 (e1) * norm_1 (e2)); |
| 533 | indexing_matrix_assign<scalar_plus_assign> (cm, prod (e1, e2), row_major_tag ()); |
| 534 | #endif |
| 535 | typename expression1_type::const_iterator1 it1 (e1 ().begin1 ()); |
| 536 | typename expression1_type::const_iterator1 it1_end (e1 ().end1 ()); |
| 537 | while (it1 != it1_end) { |
| 538 | #ifndef BOOST_UBLAS_NO_NESTED_CLASS_RELATION |
| 539 | typename expression1_type::const_iterator2 it2 (it1.begin ()); |
| 540 | typename expression1_type::const_iterator2 it2_end (it1.end ()); |
| 541 | #else |
| 542 | typename expression1_type::const_iterator2 it2 (boost::numeric::ublas::begin (it1, iterator1_tag ())); |
| 543 | typename expression1_type::const_iterator2 it2_end (boost::numeric::ublas::end (it1, iterator1_tag ())); |
| 544 | #endif |
| 545 | while (it2 != it2_end) { |
| 546 | // row (m, it1.index1 ()).plus_assign (*it2 * row (e2 (), it2.index2 ())); |
| 547 | matrix_row<expression2_type> mr (e2 (), it2.index2 ()); |
| 548 | typename matrix_row<expression2_type>::const_iterator itr (mr.begin ()); |
| 549 | typename matrix_row<expression2_type>::const_iterator itr_end (mr.end ()); |
| 550 | while (itr != itr_end) { |
| 551 | if (triangular_restriction::other (it1.index1 (), itr.index ())) |
| 552 | m (it1.index1 (), itr.index ()) += *it2 * *itr; |
| 553 | ++ itr; |
| 554 | } |
| 555 | ++ it2; |
| 556 | } |
| 557 | ++ it1; |
| 558 | } |
| 559 | #if BOOST_UBLAS_TYPE_CHECK |
| 560 | BOOST_UBLAS_CHECK (norm_1 (m - cm) <= 2 * std::numeric_limits<real_type>::epsilon () * merrorbound, internal_logic ()); |
| 561 | #endif |
| 562 | return m; |
| 563 | } |
| 564 | |
| 565 | template<class M, class E1, class E2, class TRI> |
| 566 | BOOST_UBLAS_INLINE |
| 567 | M & |
| 568 | axpy_prod (const matrix_expression<E1> &e1, |
| 569 | const matrix_expression<E2> &e2, |
| 570 | M &m, TRI, |
| 571 | dense_proxy_tag, column_major_tag) { |
| 572 | typedef typename M::size_type size_type; |
| 573 | |
| 574 | #if BOOST_UBLAS_TYPE_CHECK |
| 575 | typedef typename M::value_type value_type; |
| 576 | matrix<value_type, column_major> cm (m); |
| 577 | typedef typename type_traits<value_type>::real_type real_type; |
| 578 | real_type merrorbound (norm_1 (m) + norm_1 (e1) * norm_1 (e2)); |
| 579 | indexing_matrix_assign<scalar_plus_assign> (cm, prod (e1, e2), column_major_tag ()); |
| 580 | #endif |
| 581 | size_type size1 (e2 ().size1 ()); |
| 582 | size_type size2 (e2 ().size2 ()); |
| 583 | for (size_type j = 0; j < size2; ++ j) |
| 584 | for (size_type i = 0; i < size1; ++ i) |
| 585 | column (m, j).plus_assign (e2 () (i, j) * column (e1 (), i)); |
| 586 | #if BOOST_UBLAS_TYPE_CHECK |
| 587 | BOOST_UBLAS_CHECK (norm_1 (m - cm) <= 2 * std::numeric_limits<real_type>::epsilon () * merrorbound, internal_logic ()); |
| 588 | #endif |
| 589 | return m; |
| 590 | } |
| 591 | template<class M, class E1, class E2, class TRI> |
| 592 | BOOST_UBLAS_INLINE |
| 593 | M & |
| 594 | axpy_prod (const matrix_expression<E1> &e1, |
| 595 | const matrix_expression<E2> &e2, |
| 596 | M &m, TRI, |
| 597 | sparse_proxy_tag, column_major_tag) { |
| 598 | typedef TRI triangular_restriction; |
| 599 | typedef const E1 expression1_type; |
| 600 | typedef const E2 expression2_type; |
| 601 | |
| 602 | |
| 603 | #if BOOST_UBLAS_TYPE_CHECK |
| 604 | typedef typename M::value_type value_type; |
| 605 | matrix<value_type, column_major> cm (m); |
| 606 | typedef typename type_traits<value_type>::real_type real_type; |
| 607 | real_type merrorbound (norm_1 (m) + norm_1 (e1) * norm_1 (e2)); |
| 608 | indexing_matrix_assign<scalar_plus_assign> (cm, prod (e1, e2), column_major_tag ()); |
| 609 | #endif |
| 610 | typename expression2_type::const_iterator2 it2 (e2 ().begin2 ()); |
| 611 | typename expression2_type::const_iterator2 it2_end (e2 ().end2 ()); |
| 612 | while (it2 != it2_end) { |
| 613 | #ifndef BOOST_UBLAS_NO_NESTED_CLASS_RELATION |
| 614 | typename expression2_type::const_iterator1 it1 (it2.begin ()); |
| 615 | typename expression2_type::const_iterator1 it1_end (it2.end ()); |
| 616 | #else |
| 617 | typename expression2_type::const_iterator1 it1 (boost::numeric::ublas::begin (it2, iterator2_tag ())); |
| 618 | typename expression2_type::const_iterator1 it1_end (boost::numeric::ublas::end (it2, iterator2_tag ())); |
| 619 | #endif |
| 620 | while (it1 != it1_end) { |
| 621 | // column (m, it2.index2 ()).plus_assign (*it1 * column (e1 (), it1.index1 ())); |
| 622 | matrix_column<expression1_type> mc (e1 (), it1.index1 ()); |
| 623 | typename matrix_column<expression1_type>::const_iterator itc (mc.begin ()); |
| 624 | typename matrix_column<expression1_type>::const_iterator itc_end (mc.end ()); |
| 625 | while (itc != itc_end) { |
| 626 | if(triangular_restriction::other (itc.index (), it2.index2 ())) |
| 627 | m (itc.index (), it2.index2 ()) += *it1 * *itc; |
| 628 | ++ itc; |
| 629 | } |
| 630 | ++ it1; |
| 631 | } |
| 632 | ++ it2; |
| 633 | } |
| 634 | #if BOOST_UBLAS_TYPE_CHECK |
| 635 | BOOST_UBLAS_CHECK (norm_1 (m - cm) <= 2 * std::numeric_limits<real_type>::epsilon () * merrorbound, internal_logic ()); |
| 636 | #endif |
| 637 | return m; |
| 638 | } |
| 639 | |
| 640 | // Dispatcher |
| 641 | template<class M, class E1, class E2, class TRI> |
| 642 | BOOST_UBLAS_INLINE |
| 643 | M & |
| 644 | axpy_prod (const matrix_expression<E1> &e1, |
| 645 | const matrix_expression<E2> &e2, |
| 646 | M &m, TRI, bool init = true) { |
| 647 | typedef typename M::value_type value_type; |
| 648 | typedef typename M::storage_category storage_category; |
| 649 | typedef typename M::orientation_category orientation_category; |
| 650 | typedef TRI triangular_restriction; |
| 651 | |
| 652 | if (init) |
| 653 | m.assign (zero_matrix<value_type> (e1 ().size1 (), e2 ().size2 ())); |
| 654 | return axpy_prod (e1, e2, m, triangular_restriction (), storage_category (), orientation_category ()); |
| 655 | } |
| 656 | template<class M, class E1, class E2, class TRI> |
| 657 | BOOST_UBLAS_INLINE |
| 658 | M |
| 659 | axpy_prod (const matrix_expression<E1> &e1, |
| 660 | const matrix_expression<E2> &e2, |
| 661 | TRI) { |
| 662 | typedef M matrix_type; |
| 663 | typedef TRI triangular_restriction; |
| 664 | |
| 665 | matrix_type m (e1 ().size1 (), e2 ().size2 ()); |
| 666 | return axpy_prod (e1, e2, m, triangular_restriction (), true); |
| 667 | } |
| 668 | |
| 669 | /** \brief computes <tt>M += A X</tt> or <tt>M = A X</tt> in an |
| 670 | optimized fashion. |
| 671 | |
| 672 | \param e1 the matrix expression \c A |
| 673 | \param e2 the matrix expression \c X |
| 674 | \param m the result matrix \c M |
| 675 | \param init a boolean parameter |
| 676 | |
| 677 | <tt>axpy_prod(A, X, M, init)</tt> implements the well known |
| 678 | axpy-product. Setting \a init to \c true is equivalent to call |
| 679 | <tt>M.clear()</tt> before <tt>axpy_prod</tt>. Currently \a init |
| 680 | defaults to \c true, but this may change in the future. |
| 681 | |
| 682 | Up to now there are no specialisations. |
| 683 | |
| 684 | \ingroup blas3 |
| 685 | |
| 686 | \internal |
| 687 | |
| 688 | template parameters: |
| 689 | \param M type of the result matrix \c M |
| 690 | \param E1 type of a matrix expression \c A |
| 691 | \param E2 type of a matrix expression \c X |
| 692 | */ |
| 693 | template<class M, class E1, class E2> |
| 694 | BOOST_UBLAS_INLINE |
| 695 | M & |
| 696 | axpy_prod (const matrix_expression<E1> &e1, |
| 697 | const matrix_expression<E2> &e2, |
| 698 | M &m, bool init = true) { |
| 699 | typedef typename M::value_type value_type; |
| 700 | typedef typename M::storage_category storage_category; |
| 701 | typedef typename M::orientation_category orientation_category; |
| 702 | |
| 703 | if (init) |
| 704 | m.assign (zero_matrix<value_type> (e1 ().size1 (), e2 ().size2 ())); |
| 705 | return axpy_prod (e1, e2, m, full (), storage_category (), orientation_category ()); |
| 706 | } |
| 707 | template<class M, class E1, class E2> |
| 708 | BOOST_UBLAS_INLINE |
| 709 | M |
| 710 | axpy_prod (const matrix_expression<E1> &e1, |
| 711 | const matrix_expression<E2> &e2) { |
| 712 | typedef M matrix_type; |
| 713 | |
| 714 | matrix_type m (e1 ().size1 (), e2 ().size2 ()); |
| 715 | return axpy_prod (e1, e2, m, full (), true); |
| 716 | } |
| 717 | |
| 718 | |
| 719 | template<class M, class E1, class E2> |
| 720 | BOOST_UBLAS_INLINE |
| 721 | M & |
| 722 | opb_prod (const matrix_expression<E1> &e1, |
| 723 | const matrix_expression<E2> &e2, |
| 724 | M &m, |
| 725 | dense_proxy_tag, row_major_tag) { |
| 726 | typedef typename M::size_type size_type; |
| 727 | typedef typename M::value_type value_type; |
| 728 | |
| 729 | #if BOOST_UBLAS_TYPE_CHECK |
| 730 | matrix<value_type, row_major> cm (m); |
| 731 | typedef typename type_traits<value_type>::real_type real_type; |
| 732 | real_type merrorbound (norm_1 (m) + norm_1 (e1) * norm_1 (e2)); |
| 733 | indexing_matrix_assign<scalar_plus_assign> (cm, prod (e1, e2), row_major_tag ()); |
| 734 | #endif |
| 735 | size_type size (BOOST_UBLAS_SAME (e1 ().size2 (), e2 ().size1 ())); |
| 736 | for (size_type k = 0; k < size; ++ k) { |
| 737 | vector<value_type> ce1 (column (e1 (), k)); |
| 738 | vector<value_type> re2 (row (e2 (), k)); |
| 739 | m.plus_assign (outer_prod (ce1, re2)); |
| 740 | } |
| 741 | #if BOOST_UBLAS_TYPE_CHECK |
| 742 | BOOST_UBLAS_CHECK (norm_1 (m - cm) <= 2 * std::numeric_limits<real_type>::epsilon () * merrorbound, internal_logic ()); |
| 743 | #endif |
| 744 | return m; |
| 745 | } |
| 746 | |
| 747 | template<class M, class E1, class E2> |
| 748 | BOOST_UBLAS_INLINE |
| 749 | M & |
| 750 | opb_prod (const matrix_expression<E1> &e1, |
| 751 | const matrix_expression<E2> &e2, |
| 752 | M &m, |
| 753 | dense_proxy_tag, column_major_tag) { |
| 754 | typedef typename M::size_type size_type; |
| 755 | typedef typename M::value_type value_type; |
| 756 | |
| 757 | #if BOOST_UBLAS_TYPE_CHECK |
| 758 | matrix<value_type, column_major> cm (m); |
| 759 | typedef typename type_traits<value_type>::real_type real_type; |
| 760 | real_type merrorbound (norm_1 (m) + norm_1 (e1) * norm_1 (e2)); |
| 761 | indexing_matrix_assign<scalar_plus_assign> (cm, prod (e1, e2), column_major_tag ()); |
| 762 | #endif |
| 763 | size_type size (BOOST_UBLAS_SAME (e1 ().size2 (), e2 ().size1 ())); |
| 764 | for (size_type k = 0; k < size; ++ k) { |
| 765 | vector<value_type> ce1 (column (e1 (), k)); |
| 766 | vector<value_type> re2 (row (e2 (), k)); |
| 767 | m.plus_assign (outer_prod (ce1, re2)); |
| 768 | } |
| 769 | #if BOOST_UBLAS_TYPE_CHECK |
| 770 | BOOST_UBLAS_CHECK (norm_1 (m - cm) <= 2 * std::numeric_limits<real_type>::epsilon () * merrorbound, internal_logic ()); |
| 771 | #endif |
| 772 | return m; |
| 773 | } |
| 774 | |
| 775 | // Dispatcher |
| 776 | |
| 777 | /** \brief computes <tt>M += A X</tt> or <tt>M = A X</tt> in an |
| 778 | optimized fashion. |
| 779 | |
| 780 | \param e1 the matrix expression \c A |
| 781 | \param e2 the matrix expression \c X |
| 782 | \param m the result matrix \c M |
| 783 | \param init a boolean parameter |
| 784 | |
| 785 | <tt>opb_prod(A, X, M, init)</tt> implements the well known |
| 786 | axpy-product. Setting \a init to \c true is equivalent to call |
| 787 | <tt>M.clear()</tt> before <tt>opb_prod</tt>. Currently \a init |
| 788 | defaults to \c true, but this may change in the future. |
| 789 | |
| 790 | This function may give a speedup if \c A has less columns than |
| 791 | rows, because the product is computed as a sum of outer |
| 792 | products. |
| 793 | |
| 794 | \ingroup blas3 |
| 795 | |
| 796 | \internal |
| 797 | |
| 798 | template parameters: |
| 799 | \param M type of the result matrix \c M |
| 800 | \param E1 type of a matrix expression \c A |
| 801 | \param E2 type of a matrix expression \c X |
| 802 | */ |
| 803 | template<class M, class E1, class E2> |
| 804 | BOOST_UBLAS_INLINE |
| 805 | M & |
| 806 | opb_prod (const matrix_expression<E1> &e1, |
| 807 | const matrix_expression<E2> &e2, |
| 808 | M &m, bool init = true) { |
| 809 | typedef typename M::value_type value_type; |
| 810 | typedef typename M::storage_category storage_category; |
| 811 | typedef typename M::orientation_category orientation_category; |
| 812 | |
| 813 | if (init) |
| 814 | m.assign (zero_matrix<value_type> (e1 ().size1 (), e2 ().size2 ())); |
| 815 | return opb_prod (e1, e2, m, storage_category (), orientation_category ()); |
| 816 | } |
| 817 | template<class M, class E1, class E2> |
| 818 | BOOST_UBLAS_INLINE |
| 819 | M |
| 820 | opb_prod (const matrix_expression<E1> &e1, |
| 821 | const matrix_expression<E2> &e2) { |
| 822 | typedef M matrix_type; |
| 823 | |
| 824 | matrix_type m (e1 ().size1 (), e2 ().size2 ()); |
| 825 | return opb_prod (e1, e2, m, true); |
| 826 | } |
| 827 | |
| 828 | }}} |
| 829 | |
| 830 | #endif |
| 831 | |