1 | //
|
---|
2 | // Copyright (c) 2000-2002
|
---|
3 | // Joerg Walter, Mathias Koch
|
---|
4 | //
|
---|
5 | // Permission to use, copy, modify, distribute and sell this software
|
---|
6 | // and its documentation for any purpose is hereby granted without fee,
|
---|
7 | // provided that the above copyright notice appear in all copies and
|
---|
8 | // that both that copyright notice and this permission notice appear
|
---|
9 | // in supporting documentation. The authors make no representations
|
---|
10 | // about the suitability of this software for any purpose.
|
---|
11 | // It is provided "as is" without express or implied warranty.
|
---|
12 | //
|
---|
13 | // The authors gratefully acknowledge the support of
|
---|
14 | // GeNeSys mbH & Co. KG in producing this work.
|
---|
15 | //
|
---|
16 |
|
---|
17 | #ifndef _BOOST_UBLAS_OPERATION_
|
---|
18 | #define _BOOST_UBLAS_OPERATION_
|
---|
19 |
|
---|
20 | #include <boost/numeric/ublas/matrix_proxy.hpp>
|
---|
21 |
|
---|
22 | /** \file operation.hpp
|
---|
23 | * \brief This file contains some specialized products.
|
---|
24 | */
|
---|
25 |
|
---|
26 | // axpy-based products
|
---|
27 | // Alexei Novakov had a lot of ideas to improve these. Thanks.
|
---|
28 | // Hendrik Kueck proposed some new kernel. Thanks again.
|
---|
29 |
|
---|
30 | namespace boost { namespace numeric { namespace ublas {
|
---|
31 |
|
---|
32 | template<class V, class T1, class IA1, class TA1, class E2>
|
---|
33 | BOOST_UBLAS_INLINE
|
---|
34 | V &
|
---|
35 | axpy_prod (const compressed_matrix<T1, row_major, 0, IA1, TA1> &e1,
|
---|
36 | const vector_expression<E2> &e2,
|
---|
37 | V &v, row_major_tag) {
|
---|
38 | typedef typename V::size_type size_type;
|
---|
39 | typedef typename V::value_type value_type;
|
---|
40 |
|
---|
41 | for (size_type i = 0; i < e1.filled1 () -1; ++ i) {
|
---|
42 | size_type begin = e1.index1_data () [i];
|
---|
43 | size_type end = e1.index1_data () [i + 1];
|
---|
44 | value_type t (v (i));
|
---|
45 | for (size_type j = begin; j < end; ++ j)
|
---|
46 | t += e1.value_data () [j] * e2 () (e1.index2_data () [j]);
|
---|
47 | v (i) = t;
|
---|
48 | }
|
---|
49 | return v;
|
---|
50 | }
|
---|
51 |
|
---|
52 | template<class V, class T1, class IA1, class TA1, class E2>
|
---|
53 | BOOST_UBLAS_INLINE
|
---|
54 | V &
|
---|
55 | axpy_prod (const compressed_matrix<T1, column_major, 0, IA1, TA1> &e1,
|
---|
56 | const vector_expression<E2> &e2,
|
---|
57 | V &v, column_major_tag) {
|
---|
58 | typedef typename V::size_type size_type;
|
---|
59 |
|
---|
60 | for (size_type j = 0; j < e1.filled1 () -1; ++ j) {
|
---|
61 | size_type begin = e1.index1_data () [j];
|
---|
62 | size_type end = e1.index1_data () [j + 1];
|
---|
63 | for (size_type i = begin; i < end; ++ i)
|
---|
64 | v (e1.index2_data () [i]) += e1.value_data () [i] * e2 () (j);
|
---|
65 | }
|
---|
66 | return v;
|
---|
67 | }
|
---|
68 |
|
---|
69 | // Dispatcher
|
---|
70 | template<class V, class T1, class L1, class IA1, class TA1, class E2>
|
---|
71 | BOOST_UBLAS_INLINE
|
---|
72 | V &
|
---|
73 | axpy_prod (const compressed_matrix<T1, L1, 0, IA1, TA1> &e1,
|
---|
74 | const vector_expression<E2> &e2,
|
---|
75 | V &v, bool init = true) {
|
---|
76 | typedef typename V::value_type value_type;
|
---|
77 | typedef typename L1::orientation_category orientation_category;
|
---|
78 |
|
---|
79 | if (init)
|
---|
80 | v.assign (zero_vector<value_type> (e1.size1 ()));
|
---|
81 | #if BOOST_UBLAS_TYPE_CHECK
|
---|
82 | vector<value_type> cv (v);
|
---|
83 | typedef typename type_traits<value_type>::real_type real_type;
|
---|
84 | real_type verrorbound (norm_1 (v) + norm_1 (e1) * norm_1 (e2));
|
---|
85 | indexing_vector_assign<scalar_plus_assign> (cv, prod (e1, e2));
|
---|
86 | #endif
|
---|
87 | axpy_prod (e1, e2, v, orientation_category ());
|
---|
88 | #if BOOST_UBLAS_TYPE_CHECK
|
---|
89 | BOOST_UBLAS_CHECK (norm_1 (v - cv) <= 2 * std::numeric_limits<real_type>::epsilon () * verrorbound, internal_logic ());
|
---|
90 | #endif
|
---|
91 | return v;
|
---|
92 | }
|
---|
93 | template<class V, class T1, class L1, class IA1, class TA1, class E2>
|
---|
94 | BOOST_UBLAS_INLINE
|
---|
95 | V
|
---|
96 | axpy_prod (const compressed_matrix<T1, L1, 0, IA1, TA1> &e1,
|
---|
97 | const vector_expression<E2> &e2) {
|
---|
98 | typedef V vector_type;
|
---|
99 |
|
---|
100 | vector_type v (e1.size1 ());
|
---|
101 | return axpy_prod (e1, e2, v, true);
|
---|
102 | }
|
---|
103 |
|
---|
104 | template<class V, class T1, class L1, class IA1, class TA1, class E2>
|
---|
105 | BOOST_UBLAS_INLINE
|
---|
106 | V &
|
---|
107 | axpy_prod (const coordinate_matrix<T1, L1, 0, IA1, TA1> &e1,
|
---|
108 | const vector_expression<E2> &e2,
|
---|
109 | V &v, bool init = true) {
|
---|
110 | typedef typename V::size_type size_type;
|
---|
111 | typedef typename V::value_type value_type;
|
---|
112 | typedef L1 layout_type;
|
---|
113 |
|
---|
114 | size_type size1 = e1.size1();
|
---|
115 | size_type size2 = e1.size2();
|
---|
116 |
|
---|
117 | if (init) {
|
---|
118 | noalias(v) = zero_vector<value_type>(size1);
|
---|
119 | }
|
---|
120 |
|
---|
121 | for (size_type i = 0; i < e1.nnz(); ++i) {
|
---|
122 | size_type row_index = layout_type::element1( e1.index1_data () [i], size1, e1.index2_data () [i], size2 );
|
---|
123 | size_type col_index = layout_type::element2( e1.index1_data () [i], size1, e1.index2_data () [i], size2 );
|
---|
124 | v( row_index ) += e1.value_data () [i] * e2 () (col_index);
|
---|
125 | }
|
---|
126 | return v;
|
---|
127 | }
|
---|
128 |
|
---|
129 | template<class V, class E1, class E2>
|
---|
130 | BOOST_UBLAS_INLINE
|
---|
131 | V &
|
---|
132 | axpy_prod (const matrix_expression<E1> &e1,
|
---|
133 | const vector_expression<E2> &e2,
|
---|
134 | V &v, packed_random_access_iterator_tag, row_major_tag) {
|
---|
135 | typedef const E1 expression1_type;
|
---|
136 | typedef const E2 expression2_type;
|
---|
137 | typedef typename V::size_type size_type;
|
---|
138 |
|
---|
139 | typename expression1_type::const_iterator1 it1 (e1 ().begin1 ());
|
---|
140 | typename expression1_type::const_iterator1 it1_end (e1 ().end1 ());
|
---|
141 | while (it1 != it1_end) {
|
---|
142 | size_type index1 (it1.index1 ());
|
---|
143 | #ifndef BOOST_UBLAS_NO_NESTED_CLASS_RELATION
|
---|
144 | typename expression1_type::const_iterator2 it2 (it1.begin ());
|
---|
145 | typename expression1_type::const_iterator2 it2_end (it1.end ());
|
---|
146 | #else
|
---|
147 | typename expression1_type::const_iterator2 it2 (boost::numeric::ublas::begin (it1, iterator1_tag ()));
|
---|
148 | typename expression1_type::const_iterator2 it2_end (boost::numeric::ublas::end (it1, iterator1_tag ()));
|
---|
149 | #endif
|
---|
150 | while (it2 != it2_end) {
|
---|
151 | v (index1) += *it2 * e2 () (it2.index2 ());
|
---|
152 | ++ it2;
|
---|
153 | }
|
---|
154 | ++ it1;
|
---|
155 | }
|
---|
156 | return v;
|
---|
157 | }
|
---|
158 |
|
---|
159 | template<class V, class E1, class E2>
|
---|
160 | BOOST_UBLAS_INLINE
|
---|
161 | V &
|
---|
162 | axpy_prod (const matrix_expression<E1> &e1,
|
---|
163 | const vector_expression<E2> &e2,
|
---|
164 | V &v, packed_random_access_iterator_tag, column_major_tag) {
|
---|
165 | typedef const E1 expression1_type;
|
---|
166 | typedef const E2 expression2_type;
|
---|
167 | typedef typename V::size_type size_type;
|
---|
168 |
|
---|
169 | typename expression1_type::const_iterator2 it2 (e1 ().begin2 ());
|
---|
170 | typename expression1_type::const_iterator2 it2_end (e1 ().end2 ());
|
---|
171 | while (it2 != it2_end) {
|
---|
172 | size_type index2 (it2.index2 ());
|
---|
173 | #ifndef BOOST_UBLAS_NO_NESTED_CLASS_RELATION
|
---|
174 | typename expression1_type::const_iterator1 it1 (it2.begin ());
|
---|
175 | typename expression1_type::const_iterator1 it1_end (it2.end ());
|
---|
176 | #else
|
---|
177 | typename expression1_type::const_iterator1 it1 (boost::numeric::ublas::begin (it2, iterator2_tag ()));
|
---|
178 | typename expression1_type::const_iterator1 it1_end (boost::numeric::ublas::end (it2, iterator2_tag ()));
|
---|
179 | #endif
|
---|
180 | while (it1 != it1_end) {
|
---|
181 | v (it1.index1 ()) += *it1 * e2 () (index2);
|
---|
182 | ++ it1;
|
---|
183 | }
|
---|
184 | ++ it2;
|
---|
185 | }
|
---|
186 | return v;
|
---|
187 | }
|
---|
188 |
|
---|
189 | template<class V, class E1, class E2>
|
---|
190 | BOOST_UBLAS_INLINE
|
---|
191 | V &
|
---|
192 | axpy_prod (const matrix_expression<E1> &e1,
|
---|
193 | const vector_expression<E2> &e2,
|
---|
194 | V &v, sparse_bidirectional_iterator_tag) {
|
---|
195 | typedef const E1 expression1_type;
|
---|
196 | typedef const E2 expression2_type;
|
---|
197 | typedef typename V::size_type size_type;
|
---|
198 |
|
---|
199 | typename expression2_type::const_iterator it (e2 ().begin ());
|
---|
200 | typename expression2_type::const_iterator it_end (e2 ().end ());
|
---|
201 | while (it != it_end) {
|
---|
202 | v.plus_assign (column (e1 (), it.index ()) * *it);
|
---|
203 | ++ it;
|
---|
204 | }
|
---|
205 | return v;
|
---|
206 | }
|
---|
207 |
|
---|
208 | // Dispatcher
|
---|
209 | template<class V, class E1, class E2>
|
---|
210 | BOOST_UBLAS_INLINE
|
---|
211 | V &
|
---|
212 | axpy_prod (const matrix_expression<E1> &e1,
|
---|
213 | const vector_expression<E2> &e2,
|
---|
214 | V &v, packed_random_access_iterator_tag) {
|
---|
215 | typedef typename E1::orientation_category orientation_category;
|
---|
216 | return axpy_prod (e1, e2, v, packed_random_access_iterator_tag (), orientation_category ());
|
---|
217 | }
|
---|
218 |
|
---|
219 |
|
---|
220 | /** \brief computes <tt>v += A x</tt> or <tt>v = A x</tt> in an
|
---|
221 | optimized fashion.
|
---|
222 |
|
---|
223 | \param e1 the matrix expression \c A
|
---|
224 | \param e2 the vector expression \c x
|
---|
225 | \param v the result vector \c v
|
---|
226 | \param init a boolean parameter
|
---|
227 |
|
---|
228 | <tt>axpy_prod(A, x, v, init)</tt> implements the well known
|
---|
229 | axpy-product. Setting \a init to \c true is equivalent to call
|
---|
230 | <tt>v.clear()</tt> before <tt>axpy_prod</tt>. Currently \a init
|
---|
231 | defaults to \c true, but this may change in the future.
|
---|
232 |
|
---|
233 | Up to now there are some specialisation for compressed
|
---|
234 | matrices that give a large speed up compared to prod.
|
---|
235 |
|
---|
236 | \ingroup blas2
|
---|
237 |
|
---|
238 | \internal
|
---|
239 |
|
---|
240 | template parameters:
|
---|
241 | \param V type of the result vector \c v
|
---|
242 | \param E1 type of a matrix expression \c A
|
---|
243 | \param E2 type of a vector expression \c x
|
---|
244 | */
|
---|
245 | template<class V, class E1, class E2>
|
---|
246 | BOOST_UBLAS_INLINE
|
---|
247 | V &
|
---|
248 | axpy_prod (const matrix_expression<E1> &e1,
|
---|
249 | const vector_expression<E2> &e2,
|
---|
250 | V &v, bool init = true) {
|
---|
251 | typedef typename V::value_type value_type;
|
---|
252 | typedef typename E2::const_iterator::iterator_category iterator_category;
|
---|
253 |
|
---|
254 | if (init)
|
---|
255 | v.assign (zero_vector<value_type> (e1 ().size1 ()));
|
---|
256 | #if BOOST_UBLAS_TYPE_CHECK
|
---|
257 | vector<value_type> cv (v);
|
---|
258 | typedef typename type_traits<value_type>::real_type real_type;
|
---|
259 | real_type verrorbound (norm_1 (v) + norm_1 (e1) * norm_1 (e2));
|
---|
260 | indexing_vector_assign<scalar_plus_assign> (cv, prod (e1, e2));
|
---|
261 | #endif
|
---|
262 | axpy_prod (e1, e2, v, iterator_category ());
|
---|
263 | #if BOOST_UBLAS_TYPE_CHECK
|
---|
264 | BOOST_UBLAS_CHECK (norm_1 (v - cv) <= 2 * std::numeric_limits<real_type>::epsilon () * verrorbound, internal_logic ());
|
---|
265 | #endif
|
---|
266 | return v;
|
---|
267 | }
|
---|
268 | template<class V, class E1, class E2>
|
---|
269 | BOOST_UBLAS_INLINE
|
---|
270 | V
|
---|
271 | axpy_prod (const matrix_expression<E1> &e1,
|
---|
272 | const vector_expression<E2> &e2) {
|
---|
273 | typedef V vector_type;
|
---|
274 |
|
---|
275 | vector_type v (e1 ().size1 ());
|
---|
276 | return axpy_prod (e1, e2, v, true);
|
---|
277 | }
|
---|
278 |
|
---|
279 | template<class V, class E1, class T2, class IA2, class TA2>
|
---|
280 | BOOST_UBLAS_INLINE
|
---|
281 | V &
|
---|
282 | axpy_prod (const vector_expression<E1> &e1,
|
---|
283 | const compressed_matrix<T2, column_major, 0, IA2, TA2> &e2,
|
---|
284 | V &v, column_major_tag) {
|
---|
285 | typedef typename V::size_type size_type;
|
---|
286 | typedef typename V::value_type value_type;
|
---|
287 |
|
---|
288 | for (size_type j = 0; j < e2.filled1 () -1; ++ j) {
|
---|
289 | size_type begin = e2.index1_data () [j];
|
---|
290 | size_type end = e2.index1_data () [j + 1];
|
---|
291 | value_type t (v (j));
|
---|
292 | for (size_type i = begin; i < end; ++ i)
|
---|
293 | t += e2.value_data () [i] * e1 () (e2.index2_data () [i]);
|
---|
294 | v (j) = t;
|
---|
295 | }
|
---|
296 | return v;
|
---|
297 | }
|
---|
298 |
|
---|
299 | template<class V, class E1, class T2, class IA2, class TA2>
|
---|
300 | BOOST_UBLAS_INLINE
|
---|
301 | V &
|
---|
302 | axpy_prod (const vector_expression<E1> &e1,
|
---|
303 | const compressed_matrix<T2, row_major, 0, IA2, TA2> &e2,
|
---|
304 | V &v, row_major_tag) {
|
---|
305 | typedef typename V::size_type size_type;
|
---|
306 |
|
---|
307 | for (size_type i = 0; i < e2.filled1 () -1; ++ i) {
|
---|
308 | size_type begin = e2.index1_data () [i];
|
---|
309 | size_type end = e2.index1_data () [i + 1];
|
---|
310 | for (size_type j = begin; j < end; ++ j)
|
---|
311 | v (e2.index2_data () [j]) += e2.value_data () [j] * e1 () (i);
|
---|
312 | }
|
---|
313 | return v;
|
---|
314 | }
|
---|
315 |
|
---|
316 | // Dispatcher
|
---|
317 | template<class V, class E1, class T2, class L2, class IA2, class TA2>
|
---|
318 | BOOST_UBLAS_INLINE
|
---|
319 | V &
|
---|
320 | axpy_prod (const vector_expression<E1> &e1,
|
---|
321 | const compressed_matrix<T2, L2, 0, IA2, TA2> &e2,
|
---|
322 | V &v, bool init = true) {
|
---|
323 | typedef typename V::value_type value_type;
|
---|
324 | typedef typename L2::orientation_category orientation_category;
|
---|
325 |
|
---|
326 | if (init)
|
---|
327 | v.assign (zero_vector<value_type> (e2.size2 ()));
|
---|
328 | #if BOOST_UBLAS_TYPE_CHECK
|
---|
329 | vector<value_type> cv (v);
|
---|
330 | typedef typename type_traits<value_type>::real_type real_type;
|
---|
331 | real_type verrorbound (norm_1 (v) + norm_1 (e1) * norm_1 (e2));
|
---|
332 | indexing_vector_assign<scalar_plus_assign> (cv, prod (e1, e2));
|
---|
333 | #endif
|
---|
334 | axpy_prod (e1, e2, v, orientation_category ());
|
---|
335 | #if BOOST_UBLAS_TYPE_CHECK
|
---|
336 | BOOST_UBLAS_CHECK (norm_1 (v - cv) <= 2 * std::numeric_limits<real_type>::epsilon () * verrorbound, internal_logic ());
|
---|
337 | #endif
|
---|
338 | return v;
|
---|
339 | }
|
---|
340 | template<class V, class E1, class T2, class L2, class IA2, class TA2>
|
---|
341 | BOOST_UBLAS_INLINE
|
---|
342 | V
|
---|
343 | axpy_prod (const vector_expression<E1> &e1,
|
---|
344 | const compressed_matrix<T2, L2, 0, IA2, TA2> &e2) {
|
---|
345 | typedef V vector_type;
|
---|
346 |
|
---|
347 | vector_type v (e2.size2 ());
|
---|
348 | return axpy_prod (e1, e2, v, true);
|
---|
349 | }
|
---|
350 |
|
---|
351 | template<class V, class E1, class E2>
|
---|
352 | BOOST_UBLAS_INLINE
|
---|
353 | V &
|
---|
354 | axpy_prod (const vector_expression<E1> &e1,
|
---|
355 | const matrix_expression<E2> &e2,
|
---|
356 | V &v, packed_random_access_iterator_tag, column_major_tag) {
|
---|
357 | typedef const E1 expression1_type;
|
---|
358 | typedef const E2 expression2_type;
|
---|
359 | typedef typename V::size_type size_type;
|
---|
360 |
|
---|
361 | typename expression2_type::const_iterator2 it2 (e2 ().begin2 ());
|
---|
362 | typename expression2_type::const_iterator2 it2_end (e2 ().end2 ());
|
---|
363 | while (it2 != it2_end) {
|
---|
364 | size_type index2 (it2.index2 ());
|
---|
365 | #ifndef BOOST_UBLAS_NO_NESTED_CLASS_RELATION
|
---|
366 | typename expression2_type::const_iterator1 it1 (it2.begin ());
|
---|
367 | typename expression2_type::const_iterator1 it1_end (it2.end ());
|
---|
368 | #else
|
---|
369 | typename expression2_type::const_iterator1 it1 (boost::numeric::ublas::begin (it2, iterator2_tag ()));
|
---|
370 | typename expression2_type::const_iterator1 it1_end (boost::numeric::ublas::end (it2, iterator2_tag ()));
|
---|
371 | #endif
|
---|
372 | while (it1 != it1_end) {
|
---|
373 | v (index2) += *it1 * e1 () (it1.index1 ());
|
---|
374 | ++ it1;
|
---|
375 | }
|
---|
376 | ++ it2;
|
---|
377 | }
|
---|
378 | return v;
|
---|
379 | }
|
---|
380 |
|
---|
381 | template<class V, class E1, class E2>
|
---|
382 | BOOST_UBLAS_INLINE
|
---|
383 | V &
|
---|
384 | axpy_prod (const vector_expression<E1> &e1,
|
---|
385 | const matrix_expression<E2> &e2,
|
---|
386 | V &v, packed_random_access_iterator_tag, row_major_tag) {
|
---|
387 | typedef const E1 expression1_type;
|
---|
388 | typedef const E2 expression2_type;
|
---|
389 | typedef typename V::size_type size_type;
|
---|
390 |
|
---|
391 | typename expression2_type::const_iterator1 it1 (e2 ().begin1 ());
|
---|
392 | typename expression2_type::const_iterator1 it1_end (e2 ().end1 ());
|
---|
393 | while (it1 != it1_end) {
|
---|
394 | size_type index1 (it1.index1 ());
|
---|
395 | #ifndef BOOST_UBLAS_NO_NESTED_CLASS_RELATION
|
---|
396 | typename expression2_type::const_iterator2 it2 (it1.begin ());
|
---|
397 | typename expression2_type::const_iterator2 it2_end (it1.end ());
|
---|
398 | #else
|
---|
399 | typename expression2_type::const_iterator2 it2 (boost::numeric::ublas::begin (it1, iterator1_tag ()));
|
---|
400 | typename expression2_type::const_iterator2 it2_end (boost::numeric::ublas::end (it1, iterator1_tag ()));
|
---|
401 | #endif
|
---|
402 | while (it2 != it2_end) {
|
---|
403 | v (it2.index2 ()) += *it2 * e1 () (index1);
|
---|
404 | ++ it2;
|
---|
405 | }
|
---|
406 | ++ it1;
|
---|
407 | }
|
---|
408 | return v;
|
---|
409 | }
|
---|
410 |
|
---|
411 | template<class V, class E1, class E2>
|
---|
412 | BOOST_UBLAS_INLINE
|
---|
413 | V &
|
---|
414 | axpy_prod (const vector_expression<E1> &e1,
|
---|
415 | const matrix_expression<E2> &e2,
|
---|
416 | V &v, sparse_bidirectional_iterator_tag) {
|
---|
417 | typedef const E1 expression1_type;
|
---|
418 | typedef const E2 expression2_type;
|
---|
419 | typedef typename V::size_type size_type;
|
---|
420 |
|
---|
421 | typename expression1_type::const_iterator it (e1 ().begin ());
|
---|
422 | typename expression1_type::const_iterator it_end (e1 ().end ());
|
---|
423 | while (it != it_end) {
|
---|
424 | v.plus_assign (*it * row (e2 (), it.index ()));
|
---|
425 | ++ it;
|
---|
426 | }
|
---|
427 | return v;
|
---|
428 | }
|
---|
429 |
|
---|
430 | // Dispatcher
|
---|
431 | template<class V, class E1, class E2>
|
---|
432 | BOOST_UBLAS_INLINE
|
---|
433 | V &
|
---|
434 | axpy_prod (const vector_expression<E1> &e1,
|
---|
435 | const matrix_expression<E2> &e2,
|
---|
436 | V &v, packed_random_access_iterator_tag) {
|
---|
437 | typedef typename E2::orientation_category orientation_category;
|
---|
438 | return axpy_prod (e1, e2, v, packed_random_access_iterator_tag (), orientation_category ());
|
---|
439 | }
|
---|
440 |
|
---|
441 |
|
---|
442 | /** \brief computes <tt>v += A<sup>T</sup> x</tt> or <tt>v = A<sup>T</sup> x</tt> in an
|
---|
443 | optimized fashion.
|
---|
444 |
|
---|
445 | \param e1 the vector expression \c x
|
---|
446 | \param e2 the matrix expression \c A
|
---|
447 | \param v the result vector \c v
|
---|
448 | \param init a boolean parameter
|
---|
449 |
|
---|
450 | <tt>axpy_prod(x, A, v, init)</tt> implements the well known
|
---|
451 | axpy-product. Setting \a init to \c true is equivalent to call
|
---|
452 | <tt>v.clear()</tt> before <tt>axpy_prod</tt>. Currently \a init
|
---|
453 | defaults to \c true, but this may change in the future.
|
---|
454 |
|
---|
455 | Up to now there are some specialisation for compressed
|
---|
456 | matrices that give a large speed up compared to prod.
|
---|
457 |
|
---|
458 | \ingroup blas2
|
---|
459 |
|
---|
460 | \internal
|
---|
461 |
|
---|
462 | template parameters:
|
---|
463 | \param V type of the result vector \c v
|
---|
464 | \param E1 type of a vector expression \c x
|
---|
465 | \param E2 type of a matrix expression \c A
|
---|
466 | */
|
---|
467 | template<class V, class E1, class E2>
|
---|
468 | BOOST_UBLAS_INLINE
|
---|
469 | V &
|
---|
470 | axpy_prod (const vector_expression<E1> &e1,
|
---|
471 | const matrix_expression<E2> &e2,
|
---|
472 | V &v, bool init = true) {
|
---|
473 | typedef typename V::value_type value_type;
|
---|
474 | typedef typename E1::const_iterator::iterator_category iterator_category;
|
---|
475 |
|
---|
476 | if (init)
|
---|
477 | v.assign (zero_vector<value_type> (e2 ().size2 ()));
|
---|
478 | #if BOOST_UBLAS_TYPE_CHECK
|
---|
479 | vector<value_type> cv (v);
|
---|
480 | typedef typename type_traits<value_type>::real_type real_type;
|
---|
481 | real_type verrorbound (norm_1 (v) + norm_1 (e1) * norm_1 (e2));
|
---|
482 | indexing_vector_assign<scalar_plus_assign> (cv, prod (e1, e2));
|
---|
483 | #endif
|
---|
484 | axpy_prod (e1, e2, v, iterator_category ());
|
---|
485 | #if BOOST_UBLAS_TYPE_CHECK
|
---|
486 | BOOST_UBLAS_CHECK (norm_1 (v - cv) <= 2 * std::numeric_limits<real_type>::epsilon () * verrorbound, internal_logic ());
|
---|
487 | #endif
|
---|
488 | return v;
|
---|
489 | }
|
---|
490 | template<class V, class E1, class E2>
|
---|
491 | BOOST_UBLAS_INLINE
|
---|
492 | V
|
---|
493 | axpy_prod (const vector_expression<E1> &e1,
|
---|
494 | const matrix_expression<E2> &e2) {
|
---|
495 | typedef V vector_type;
|
---|
496 |
|
---|
497 | vector_type v (e2 ().size2 ());
|
---|
498 | return axpy_prod (e1, e2, v, true);
|
---|
499 | }
|
---|
500 |
|
---|
501 | template<class M, class E1, class E2, class TRI>
|
---|
502 | BOOST_UBLAS_INLINE
|
---|
503 | M &
|
---|
504 | axpy_prod (const matrix_expression<E1> &e1,
|
---|
505 | const matrix_expression<E2> &e2,
|
---|
506 | M &m, TRI,
|
---|
507 | dense_proxy_tag, row_major_tag) {
|
---|
508 | typedef M matrix_type;
|
---|
509 | typedef const E1 expression1_type;
|
---|
510 | typedef const E2 expression2_type;
|
---|
511 | typedef typename M::size_type size_type;
|
---|
512 | typedef typename M::value_type value_type;
|
---|
513 |
|
---|
514 | #if BOOST_UBLAS_TYPE_CHECK
|
---|
515 | matrix<value_type, row_major> cm (m);
|
---|
516 | typedef typename type_traits<value_type>::real_type real_type;
|
---|
517 | real_type merrorbound (norm_1 (m) + norm_1 (e1) * norm_1 (e2));
|
---|
518 | indexing_matrix_assign<scalar_plus_assign> (cm, prod (e1, e2), row_major_tag ());
|
---|
519 | #endif
|
---|
520 | size_type size1 (e1 ().size1 ());
|
---|
521 | size_type size2 (e1 ().size2 ());
|
---|
522 | for (size_type i = 0; i < size1; ++ i)
|
---|
523 | for (size_type j = 0; j < size2; ++ j)
|
---|
524 | row (m, i).plus_assign (e1 () (i, j) * row (e2 (), j));
|
---|
525 | #if BOOST_UBLAS_TYPE_CHECK
|
---|
526 | BOOST_UBLAS_CHECK (norm_1 (m - cm) <= 2 * std::numeric_limits<real_type>::epsilon () * merrorbound, internal_logic ());
|
---|
527 | #endif
|
---|
528 | return m;
|
---|
529 | }
|
---|
530 | template<class M, class E1, class E2, class TRI>
|
---|
531 | BOOST_UBLAS_INLINE
|
---|
532 | M &
|
---|
533 | axpy_prod (const matrix_expression<E1> &e1,
|
---|
534 | const matrix_expression<E2> &e2,
|
---|
535 | M &m, TRI,
|
---|
536 | sparse_proxy_tag, row_major_tag) {
|
---|
537 | typedef M matrix_type;
|
---|
538 | typedef TRI triangular_restriction;
|
---|
539 | typedef const E1 expression1_type;
|
---|
540 | typedef const E2 expression2_type;
|
---|
541 | typedef typename M::size_type size_type;
|
---|
542 | typedef typename M::value_type value_type;
|
---|
543 |
|
---|
544 | #if BOOST_UBLAS_TYPE_CHECK
|
---|
545 | matrix<value_type, row_major> cm (m);
|
---|
546 | typedef typename type_traits<value_type>::real_type real_type;
|
---|
547 | real_type merrorbound (norm_1 (m) + norm_1 (e1) * norm_1 (e2));
|
---|
548 | indexing_matrix_assign<scalar_plus_assign> (cm, prod (e1, e2), row_major_tag ());
|
---|
549 | #endif
|
---|
550 | typename expression1_type::const_iterator1 it1 (e1 ().begin1 ());
|
---|
551 | typename expression1_type::const_iterator1 it1_end (e1 ().end1 ());
|
---|
552 | while (it1 != it1_end) {
|
---|
553 | #ifndef BOOST_UBLAS_NO_NESTED_CLASS_RELATION
|
---|
554 | typename expression1_type::const_iterator2 it2 (it1.begin ());
|
---|
555 | typename expression1_type::const_iterator2 it2_end (it1.end ());
|
---|
556 | #else
|
---|
557 | typename expression1_type::const_iterator2 it2 (boost::numeric::ublas::begin (it1, iterator1_tag ()));
|
---|
558 | typename expression1_type::const_iterator2 it2_end (boost::numeric::ublas::end (it1, iterator1_tag ()));
|
---|
559 | #endif
|
---|
560 | while (it2 != it2_end) {
|
---|
561 | // row (m, it1.index1 ()).plus_assign (*it2 * row (e2 (), it2.index2 ()));
|
---|
562 | matrix_row<expression2_type> mr (e2 (), it2.index2 ());
|
---|
563 | typename matrix_row<expression2_type>::const_iterator itr (mr.begin ());
|
---|
564 | typename matrix_row<expression2_type>::const_iterator itr_end (mr.end ());
|
---|
565 | while (itr != itr_end) {
|
---|
566 | if (triangular_restriction::other (it1.index1 (), itr.index ()))
|
---|
567 | m (it1.index1 (), itr.index ()) += *it2 * *itr;
|
---|
568 | ++ itr;
|
---|
569 | }
|
---|
570 | ++ it2;
|
---|
571 | }
|
---|
572 | ++ it1;
|
---|
573 | }
|
---|
574 | #if BOOST_UBLAS_TYPE_CHECK
|
---|
575 | BOOST_UBLAS_CHECK (norm_1 (m - cm) <= 2 * std::numeric_limits<real_type>::epsilon () * merrorbound, internal_logic ());
|
---|
576 | #endif
|
---|
577 | return m;
|
---|
578 | }
|
---|
579 |
|
---|
580 | template<class M, class E1, class E2, class TRI>
|
---|
581 | BOOST_UBLAS_INLINE
|
---|
582 | M &
|
---|
583 | axpy_prod (const matrix_expression<E1> &e1,
|
---|
584 | const matrix_expression<E2> &e2,
|
---|
585 | M &m, TRI,
|
---|
586 | dense_proxy_tag, column_major_tag) {
|
---|
587 | typedef M matrix_type;
|
---|
588 | typedef const E1 expression1_type;
|
---|
589 | typedef const E2 expression2_type;
|
---|
590 | typedef typename M::size_type size_type;
|
---|
591 | typedef typename M::value_type value_type;
|
---|
592 |
|
---|
593 | #if BOOST_UBLAS_TYPE_CHECK
|
---|
594 | matrix<value_type, column_major> cm (m);
|
---|
595 | typedef typename type_traits<value_type>::real_type real_type;
|
---|
596 | real_type merrorbound (norm_1 (m) + norm_1 (e1) * norm_1 (e2));
|
---|
597 | indexing_matrix_assign<scalar_plus_assign> (cm, prod (e1, e2), column_major_tag ());
|
---|
598 | #endif
|
---|
599 | size_type size1 (e2 ().size1 ());
|
---|
600 | size_type size2 (e2 ().size2 ());
|
---|
601 | for (size_type j = 0; j < size2; ++ j)
|
---|
602 | for (size_type i = 0; i < size1; ++ i)
|
---|
603 | column (m, j).plus_assign (e2 () (i, j) * column (e1 (), i));
|
---|
604 | #if BOOST_UBLAS_TYPE_CHECK
|
---|
605 | BOOST_UBLAS_CHECK (norm_1 (m - cm) <= 2 * std::numeric_limits<real_type>::epsilon () * merrorbound, internal_logic ());
|
---|
606 | #endif
|
---|
607 | return m;
|
---|
608 | }
|
---|
609 | template<class M, class E1, class E2, class TRI>
|
---|
610 | BOOST_UBLAS_INLINE
|
---|
611 | M &
|
---|
612 | axpy_prod (const matrix_expression<E1> &e1,
|
---|
613 | const matrix_expression<E2> &e2,
|
---|
614 | M &m, TRI,
|
---|
615 | sparse_proxy_tag, column_major_tag) {
|
---|
616 | typedef M matrix_type;
|
---|
617 | typedef TRI triangular_restriction;
|
---|
618 | typedef const E1 expression1_type;
|
---|
619 | typedef const E2 expression2_type;
|
---|
620 | typedef typename M::size_type size_type;
|
---|
621 | typedef typename M::value_type value_type;
|
---|
622 |
|
---|
623 | #if BOOST_UBLAS_TYPE_CHECK
|
---|
624 | matrix<value_type, column_major> cm (m);
|
---|
625 | typedef typename type_traits<value_type>::real_type real_type;
|
---|
626 | real_type merrorbound (norm_1 (m) + norm_1 (e1) * norm_1 (e2));
|
---|
627 | indexing_matrix_assign<scalar_plus_assign> (cm, prod (e1, e2), column_major_tag ());
|
---|
628 | #endif
|
---|
629 | typename expression2_type::const_iterator2 it2 (e2 ().begin2 ());
|
---|
630 | typename expression2_type::const_iterator2 it2_end (e2 ().end2 ());
|
---|
631 | while (it2 != it2_end) {
|
---|
632 | #ifndef BOOST_UBLAS_NO_NESTED_CLASS_RELATION
|
---|
633 | typename expression2_type::const_iterator1 it1 (it2.begin ());
|
---|
634 | typename expression2_type::const_iterator1 it1_end (it2.end ());
|
---|
635 | #else
|
---|
636 | typename expression2_type::const_iterator1 it1 (boost::numeric::ublas::begin (it2, iterator2_tag ()));
|
---|
637 | typename expression2_type::const_iterator1 it1_end (boost::numeric::ublas::end (it2, iterator2_tag ()));
|
---|
638 | #endif
|
---|
639 | while (it1 != it1_end) {
|
---|
640 | // column (m, it2.index2 ()).plus_assign (*it1 * column (e1 (), it1.index1 ()));
|
---|
641 | matrix_column<expression1_type> mc (e1 (), it1.index1 ());
|
---|
642 | typename matrix_column<expression1_type>::const_iterator itc (mc.begin ());
|
---|
643 | typename matrix_column<expression1_type>::const_iterator itc_end (mc.end ());
|
---|
644 | while (itc != itc_end) {
|
---|
645 | if (triangular_restriction::functor_type ().other (itc.index (), it2.index2 ()))
|
---|
646 | m (itc.index (), it2.index2 ()) += *it1 * *itc;
|
---|
647 | ++ itc;
|
---|
648 | }
|
---|
649 | ++ it1;
|
---|
650 | }
|
---|
651 | ++ it2;
|
---|
652 | }
|
---|
653 | #if BOOST_UBLAS_TYPE_CHECK
|
---|
654 | BOOST_UBLAS_CHECK (norm_1 (m - cm) <= 2 * std::numeric_limits<real_type>::epsilon () * merrorbound, internal_logic ());
|
---|
655 | #endif
|
---|
656 | return m;
|
---|
657 | }
|
---|
658 |
|
---|
659 | // Dispatcher
|
---|
660 | template<class M, class E1, class E2, class TRI>
|
---|
661 | BOOST_UBLAS_INLINE
|
---|
662 | M &
|
---|
663 | axpy_prod (const matrix_expression<E1> &e1,
|
---|
664 | const matrix_expression<E2> &e2,
|
---|
665 | M &m, TRI, bool init = true) {
|
---|
666 | typedef typename M::value_type value_type;
|
---|
667 | typedef typename M::storage_category storage_category;
|
---|
668 | typedef typename M::orientation_category orientation_category;
|
---|
669 | typedef TRI triangular_restriction;
|
---|
670 |
|
---|
671 | if (init)
|
---|
672 | m.assign (zero_matrix<value_type> (e1 ().size1 (), e2 ().size2 ()));
|
---|
673 | return axpy_prod (e1, e2, m, triangular_restriction (), storage_category (), orientation_category ());
|
---|
674 | }
|
---|
675 | template<class M, class E1, class E2, class TRI>
|
---|
676 | BOOST_UBLAS_INLINE
|
---|
677 | M
|
---|
678 | axpy_prod (const matrix_expression<E1> &e1,
|
---|
679 | const matrix_expression<E2> &e2,
|
---|
680 | TRI) {
|
---|
681 | typedef M matrix_type;
|
---|
682 | typedef TRI triangular_restriction;
|
---|
683 |
|
---|
684 | matrix_type m (e1 ().size1 (), e2 ().size2 ());
|
---|
685 | return axpy_prod (e1, e2, m, triangular_restriction (), true);
|
---|
686 | }
|
---|
687 |
|
---|
688 | /** \brief computes <tt>M += A X</tt> or <tt>M = A X</tt> in an
|
---|
689 | optimized fashion.
|
---|
690 |
|
---|
691 | \param e1 the matrix expression \c A
|
---|
692 | \param e2 the matrix expression \c X
|
---|
693 | \param m the result matrix \c M
|
---|
694 | \param init a boolean parameter
|
---|
695 |
|
---|
696 | <tt>axpy_prod(A, X, M, init)</tt> implements the well known
|
---|
697 | axpy-product. Setting \a init to \c true is equivalent to call
|
---|
698 | <tt>M.clear()</tt> before <tt>axpy_prod</tt>. Currently \a init
|
---|
699 | defaults to \c true, but this may change in the future.
|
---|
700 |
|
---|
701 | Up to now there are no specialisations.
|
---|
702 |
|
---|
703 | \ingroup blas3
|
---|
704 |
|
---|
705 | \internal
|
---|
706 |
|
---|
707 | template parameters:
|
---|
708 | \param M type of the result matrix \c M
|
---|
709 | \param E1 type of a matrix expression \c A
|
---|
710 | \param E2 type of a matrix expression \c X
|
---|
711 | */
|
---|
712 | template<class M, class E1, class E2>
|
---|
713 | BOOST_UBLAS_INLINE
|
---|
714 | M &
|
---|
715 | axpy_prod (const matrix_expression<E1> &e1,
|
---|
716 | const matrix_expression<E2> &e2,
|
---|
717 | M &m, bool init = true) {
|
---|
718 | typedef typename M::value_type value_type;
|
---|
719 | typedef typename M::storage_category storage_category;
|
---|
720 | typedef typename M::orientation_category orientation_category;
|
---|
721 |
|
---|
722 | if (init)
|
---|
723 | m.assign (zero_matrix<value_type> (e1 ().size1 (), e2 ().size2 ()));
|
---|
724 | return axpy_prod (e1, e2, m, full (), storage_category (), orientation_category ());
|
---|
725 | }
|
---|
726 | template<class M, class E1, class E2>
|
---|
727 | BOOST_UBLAS_INLINE
|
---|
728 | M
|
---|
729 | axpy_prod (const matrix_expression<E1> &e1,
|
---|
730 | const matrix_expression<E2> &e2) {
|
---|
731 | typedef M matrix_type;
|
---|
732 |
|
---|
733 | matrix_type m (e1 ().size1 (), e2 ().size2 ());
|
---|
734 | return axpy_prod (e1, e2, m, full (), true);
|
---|
735 | }
|
---|
736 |
|
---|
737 |
|
---|
738 | template<class M, class E1, class E2>
|
---|
739 | BOOST_UBLAS_INLINE
|
---|
740 | M &
|
---|
741 | opb_prod (const matrix_expression<E1> &e1,
|
---|
742 | const matrix_expression<E2> &e2,
|
---|
743 | M &m,
|
---|
744 | dense_proxy_tag, row_major_tag) {
|
---|
745 | typedef M matrix_type;
|
---|
746 | typedef const E1 expression1_type;
|
---|
747 | typedef const E2 expression2_type;
|
---|
748 | typedef typename M::size_type size_type;
|
---|
749 | typedef typename M::value_type value_type;
|
---|
750 |
|
---|
751 | #if BOOST_UBLAS_TYPE_CHECK
|
---|
752 | matrix<value_type, row_major> cm (m);
|
---|
753 | typedef typename type_traits<value_type>::real_type real_type;
|
---|
754 | real_type merrorbound (norm_1 (m) + norm_1 (e1) * norm_1 (e2));
|
---|
755 | indexing_matrix_assign<scalar_plus_assign> (cm, prod (e1, e2), row_major_tag ());
|
---|
756 | #endif
|
---|
757 | size_type size (BOOST_UBLAS_SAME (e1 ().size2 (), e2 ().size1 ()));
|
---|
758 | for (size_type k = 0; k < size; ++ k) {
|
---|
759 | vector<value_type> ce1 (column (e1 (), k));
|
---|
760 | vector<value_type> re2 (row (e2 (), k));
|
---|
761 | m.plus_assign (outer_prod (ce1, re2));
|
---|
762 | }
|
---|
763 | #if BOOST_UBLAS_TYPE_CHECK
|
---|
764 | BOOST_UBLAS_CHECK (norm_1 (m - cm) <= 2 * std::numeric_limits<real_type>::epsilon () * merrorbound, internal_logic ());
|
---|
765 | #endif
|
---|
766 | return m;
|
---|
767 | }
|
---|
768 |
|
---|
769 | template<class M, class E1, class E2>
|
---|
770 | BOOST_UBLAS_INLINE
|
---|
771 | M &
|
---|
772 | opb_prod (const matrix_expression<E1> &e1,
|
---|
773 | const matrix_expression<E2> &e2,
|
---|
774 | M &m,
|
---|
775 | dense_proxy_tag, column_major_tag) {
|
---|
776 | typedef M matrix_type;
|
---|
777 | typedef const E1 expression1_type;
|
---|
778 | typedef const E2 expression2_type;
|
---|
779 | typedef typename M::size_type size_type;
|
---|
780 | typedef typename M::value_type value_type;
|
---|
781 |
|
---|
782 | #if BOOST_UBLAS_TYPE_CHECK
|
---|
783 | matrix<value_type, column_major> cm (m);
|
---|
784 | typedef typename type_traits<value_type>::real_type real_type;
|
---|
785 | real_type merrorbound (norm_1 (m) + norm_1 (e1) * norm_1 (e2));
|
---|
786 | indexing_matrix_assign<scalar_plus_assign> (cm, prod (e1, e2), column_major_tag ());
|
---|
787 | #endif
|
---|
788 | size_type size (BOOST_UBLAS_SAME (e1 ().size2 (), e2 ().size1 ()));
|
---|
789 | for (size_type k = 0; k < size; ++ k) {
|
---|
790 | vector<value_type> ce1 (column (e1 (), k));
|
---|
791 | vector<value_type> re2 (row (e2 (), k));
|
---|
792 | m.plus_assign (outer_prod (ce1, re2));
|
---|
793 | }
|
---|
794 | #if BOOST_UBLAS_TYPE_CHECK
|
---|
795 | BOOST_UBLAS_CHECK (norm_1 (m - cm) <= 2 * std::numeric_limits<real_type>::epsilon () * merrorbound, internal_logic ());
|
---|
796 | #endif
|
---|
797 | return m;
|
---|
798 | }
|
---|
799 |
|
---|
800 | // Dispatcher
|
---|
801 |
|
---|
802 | /** \brief computes <tt>M += A X</tt> or <tt>M = A X</tt> in an
|
---|
803 | optimized fashion.
|
---|
804 |
|
---|
805 | \param e1 the matrix expression \c A
|
---|
806 | \param e2 the matrix expression \c X
|
---|
807 | \param m the result matrix \c M
|
---|
808 | \param init a boolean parameter
|
---|
809 |
|
---|
810 | <tt>opb_prod(A, X, M, init)</tt> implements the well known
|
---|
811 | axpy-product. Setting \a init to \c true is equivalent to call
|
---|
812 | <tt>M.clear()</tt> before <tt>opb_prod</tt>. Currently \a init
|
---|
813 | defaults to \c true, but this may change in the future.
|
---|
814 |
|
---|
815 | This function may give a speedup if \c A has less columns than
|
---|
816 | rows, because the product is computed as a sum of outer
|
---|
817 | products.
|
---|
818 |
|
---|
819 | \ingroup blas3
|
---|
820 |
|
---|
821 | \internal
|
---|
822 |
|
---|
823 | template parameters:
|
---|
824 | \param M type of the result matrix \c M
|
---|
825 | \param E1 type of a matrix expression \c A
|
---|
826 | \param E2 type of a matrix expression \c X
|
---|
827 | */
|
---|
828 | template<class M, class E1, class E2>
|
---|
829 | BOOST_UBLAS_INLINE
|
---|
830 | M &
|
---|
831 | opb_prod (const matrix_expression<E1> &e1,
|
---|
832 | const matrix_expression<E2> &e2,
|
---|
833 | M &m, bool init = true) {
|
---|
834 | typedef typename M::value_type value_type;
|
---|
835 | typedef typename M::storage_category storage_category;
|
---|
836 | typedef typename M::orientation_category orientation_category;
|
---|
837 |
|
---|
838 | if (init)
|
---|
839 | m.assign (zero_matrix<value_type> (e1 ().size1 (), e2 ().size2 ()));
|
---|
840 | return opb_prod (e1, e2, m, storage_category (), orientation_category ());
|
---|
841 | }
|
---|
842 | template<class M, class E1, class E2>
|
---|
843 | BOOST_UBLAS_INLINE
|
---|
844 | M
|
---|
845 | opb_prod (const matrix_expression<E1> &e1,
|
---|
846 | const matrix_expression<E2> &e2) {
|
---|
847 | typedef M matrix_type;
|
---|
848 |
|
---|
849 | matrix_type m (e1 ().size1 (), e2 ().size2 ());
|
---|
850 | return opb_prod (e1, e2, m, true);
|
---|
851 | }
|
---|
852 |
|
---|
853 | }}}
|
---|
854 |
|
---|
855 | #endif
|
---|