Chris@16: //
Chris@16: //  Copyright (c) 2000-2002
Chris@16: //  Joerg Walter, Mathias Koch
Chris@16: //
Chris@16: //  Distributed under the Boost Software License, Version 1.0. (See
Chris@16: //  accompanying file LICENSE_1_0.txt or copy at
Chris@16: //  http://www.boost.org/LICENSE_1_0.txt)
Chris@16: //
Chris@16: //  The authors gratefully acknowledge the support of
Chris@16: //  GeNeSys mbH & Co. KG in producing this work.
Chris@16: //
Chris@16: 
Chris@16: #ifndef _BOOST_UBLAS_OPERATION_BLOCKED_
Chris@16: #define _BOOST_UBLAS_OPERATION_BLOCKED_
Chris@16: 
Chris@16: #include <boost/numeric/ublas/traits.hpp>
Chris@16: #include <boost/numeric/ublas/detail/vector_assign.hpp> // indexing_vector_assign
Chris@16: #include <boost/numeric/ublas/detail/matrix_assign.hpp> // indexing_matrix_assign
Chris@16: 
Chris@16: 
Chris@16: namespace boost { namespace numeric { namespace ublas {
Chris@16: 
Chris@16:     template<class V, typename V::size_type BS, class E1, class E2>
Chris@16:     BOOST_UBLAS_INLINE
Chris@16:     V
Chris@16:     block_prod (const matrix_expression<E1> &e1,
Chris@16:                 const vector_expression<E2> &e2) {
Chris@16:         typedef V vector_type;
Chris@16:         typedef const E1 expression1_type;
Chris@16:         typedef const E2 expression2_type;
Chris@16:         typedef typename V::size_type size_type;
Chris@16:         typedef typename V::value_type value_type;
Chris@16:         const size_type block_size = BS;
Chris@16: 
Chris@16:         V v (e1 ().size1 ());
Chris@16: #if BOOST_UBLAS_TYPE_CHECK
Chris@16:         vector<value_type> cv (v.size ());
Chris@16:         typedef typename type_traits<value_type>::real_type real_type;
Chris@16:         real_type verrorbound (norm_1 (v) + norm_1 (e1) * norm_1 (e2));
Chris@16:         indexing_vector_assign<scalar_assign> (cv, prod (e1, e2));
Chris@16: #endif
Chris@16:         size_type i_size = e1 ().size1 ();
Chris@16:         size_type j_size = BOOST_UBLAS_SAME (e1 ().size2 (), e2 ().size ());
Chris@16:         for (size_type i_begin = 0; i_begin < i_size; i_begin += block_size) {
Chris@16:             size_type i_end = i_begin + (std::min) (i_size - i_begin, block_size);
Chris@16:             // FIX: never ignore Martin Weiser's advice ;-(
Chris@16: #ifdef BOOST_UBLAS_NO_CACHE
Chris@16:             vector_range<vector_type> v_range (v, range (i_begin, i_end));
Chris@16: #else
Chris@16:             // vector<value_type, bounded_array<value_type, block_size> > v_range (i_end - i_begin);
Chris@16:             vector<value_type> v_range (i_end - i_begin);
Chris@16: #endif
Chris@16:             v_range.assign (zero_vector<value_type> (i_end - i_begin));
Chris@16:             for (size_type j_begin = 0; j_begin < j_size; j_begin += block_size) {
Chris@16:                 size_type j_end = j_begin + (std::min) (j_size - j_begin, block_size);
Chris@16: #ifdef BOOST_UBLAS_NO_CACHE
Chris@16:                 const matrix_range<expression1_type> e1_range (e1 (), range (i_begin, i_end), range (j_begin, j_end));
Chris@16:                 const vector_range<expression2_type> e2_range (e2 (), range (j_begin, j_end));
Chris@16:                 v_range.plus_assign (prod (e1_range, e2_range));
Chris@16: #else
Chris@16:                 // const matrix<value_type, row_major, bounded_array<value_type, block_size * block_size> > e1_range (project (e1 (), range (i_begin, i_end), range (j_begin, j_end)));
Chris@16:                 // const vector<value_type, bounded_array<value_type, block_size> > e2_range (project (e2 (), range (j_begin, j_end)));
Chris@16:                 const matrix<value_type, row_major> e1_range (project (e1 (), range (i_begin, i_end), range (j_begin, j_end)));
Chris@16:                 const vector<value_type> e2_range (project (e2 (), range (j_begin, j_end)));
Chris@16:                 v_range.plus_assign (prod (e1_range, e2_range));
Chris@16: #endif
Chris@16:             }
Chris@16: #ifndef BOOST_UBLAS_NO_CACHE
Chris@16:             project (v, range (i_begin, i_end)).assign (v_range);
Chris@16: #endif
Chris@16:         }
Chris@16: #if BOOST_UBLAS_TYPE_CHECK
Chris@16:         BOOST_UBLAS_CHECK (norm_1 (v - cv) <= 2 * std::numeric_limits<real_type>::epsilon () * verrorbound, internal_logic ());
Chris@16: #endif
Chris@16:         return v;
Chris@16:     }
Chris@16: 
Chris@16:     template<class V, typename V::size_type BS, class E1, class E2>
Chris@16:     BOOST_UBLAS_INLINE
Chris@16:     V
Chris@16:     block_prod (const vector_expression<E1> &e1,
Chris@16:                 const matrix_expression<E2> &e2) {
Chris@16:         typedef V vector_type;
Chris@16:         typedef const E1 expression1_type;
Chris@16:         typedef const E2 expression2_type;
Chris@16:         typedef typename V::size_type size_type;
Chris@16:         typedef typename V::value_type value_type;
Chris@16:         const size_type block_size = BS;
Chris@16: 
Chris@16:         V v (e2 ().size2 ());
Chris@16: #if BOOST_UBLAS_TYPE_CHECK
Chris@16:         vector<value_type> cv (v.size ());
Chris@16:         typedef typename type_traits<value_type>::real_type real_type;
Chris@16:         real_type verrorbound (norm_1 (v) + norm_1 (e1) * norm_1 (e2));
Chris@16:         indexing_vector_assign<scalar_assign> (cv, prod (e1, e2));
Chris@16: #endif
Chris@16:         size_type i_size = BOOST_UBLAS_SAME (e1 ().size (), e2 ().size1 ());
Chris@16:         size_type j_size = e2 ().size2 ();
Chris@16:         for (size_type j_begin = 0; j_begin < j_size; j_begin += block_size) {
Chris@16:             size_type j_end = j_begin + (std::min) (j_size - j_begin, block_size);
Chris@16:             // FIX: never ignore Martin Weiser's advice ;-(
Chris@16: #ifdef BOOST_UBLAS_NO_CACHE
Chris@16:             vector_range<vector_type> v_range (v, range (j_begin, j_end));
Chris@16: #else
Chris@16:             // vector<value_type, bounded_array<value_type, block_size> > v_range (j_end - j_begin);
Chris@16:             vector<value_type> v_range (j_end - j_begin);
Chris@16: #endif
Chris@16:             v_range.assign (zero_vector<value_type> (j_end - j_begin));
Chris@16:             for (size_type i_begin = 0; i_begin < i_size; i_begin += block_size) {
Chris@16:                 size_type i_end = i_begin + (std::min) (i_size - i_begin, block_size);
Chris@16: #ifdef BOOST_UBLAS_NO_CACHE
Chris@16:                 const vector_range<expression1_type> e1_range (e1 (), range (i_begin, i_end));
Chris@16:                 const matrix_range<expression2_type> e2_range (e2 (), range (i_begin, i_end), range (j_begin, j_end));
Chris@16: #else
Chris@16:                 // const vector<value_type, bounded_array<value_type, block_size> > e1_range (project (e1 (), range (i_begin, i_end)));
Chris@16:                 // const matrix<value_type, column_major, bounded_array<value_type, block_size * block_size> > e2_range (project (e2 (), range (i_begin, i_end), range (j_begin, j_end)));
Chris@16:                 const vector<value_type> e1_range (project (e1 (), range (i_begin, i_end)));
Chris@16:                 const matrix<value_type, column_major> e2_range (project (e2 (), range (i_begin, i_end), range (j_begin, j_end)));
Chris@16: #endif
Chris@16:                 v_range.plus_assign (prod (e1_range, e2_range));
Chris@16:             }
Chris@16: #ifndef BOOST_UBLAS_NO_CACHE
Chris@16:             project (v, range (j_begin, j_end)).assign (v_range);
Chris@16: #endif
Chris@16:         }
Chris@16: #if BOOST_UBLAS_TYPE_CHECK
Chris@16:         BOOST_UBLAS_CHECK (norm_1 (v - cv) <= 2 * std::numeric_limits<real_type>::epsilon () * verrorbound, internal_logic ());
Chris@16: #endif
Chris@16:         return v;
Chris@16:     }
Chris@16: 
Chris@16:     template<class M, typename M::size_type BS, class E1, class E2>
Chris@16:     BOOST_UBLAS_INLINE
Chris@16:     M
Chris@16:     block_prod (const matrix_expression<E1> &e1,
Chris@16:                 const matrix_expression<E2> &e2,
Chris@16:                 row_major_tag) {
Chris@16:         typedef M matrix_type;
Chris@16:         typedef const E1 expression1_type;
Chris@16:         typedef const E2 expression2_type;
Chris@16:         typedef typename M::size_type size_type;
Chris@16:         typedef typename M::value_type value_type;
Chris@16:         const size_type block_size = BS;
Chris@16: 
Chris@16:         M m (e1 ().size1 (), e2 ().size2 ());
Chris@16: #if BOOST_UBLAS_TYPE_CHECK
Chris@16:         matrix<value_type, row_major> cm (m.size1 (), m.size2 ());
Chris@16:         typedef typename type_traits<value_type>::real_type real_type;
Chris@16:         real_type merrorbound (norm_1 (m) + norm_1 (e1) * norm_1 (e2));
Chris@16:         indexing_matrix_assign<scalar_assign> (cm, prod (e1, e2), row_major_tag ());
Chris@16:         disable_type_check<bool>::value = true;
Chris@16: #endif
Chris@16:         size_type i_size = e1 ().size1 ();
Chris@16:         size_type j_size = e2 ().size2 ();
Chris@16:         size_type k_size = BOOST_UBLAS_SAME (e1 ().size2 (), e2 ().size1 ());
Chris@16:         for (size_type i_begin = 0; i_begin < i_size; i_begin += block_size) {
Chris@16:             size_type i_end = i_begin + (std::min) (i_size - i_begin, block_size);
Chris@16:             for (size_type j_begin = 0; j_begin < j_size; j_begin += block_size) {
Chris@16:                 size_type j_end = j_begin + (std::min) (j_size - j_begin, block_size);
Chris@16:                 // FIX: never ignore Martin Weiser's advice ;-(
Chris@16: #ifdef BOOST_UBLAS_NO_CACHE
Chris@16:                 matrix_range<matrix_type> m_range (m, range (i_begin, i_end), range (j_begin, j_end));
Chris@16: #else
Chris@16:                 // matrix<value_type, row_major, bounded_array<value_type, block_size * block_size> > m_range (i_end - i_begin, j_end - j_begin);
Chris@16:                 matrix<value_type, row_major> m_range (i_end - i_begin, j_end - j_begin);
Chris@16: #endif
Chris@16:                 m_range.assign (zero_matrix<value_type> (i_end - i_begin, j_end - j_begin));
Chris@16:                 for (size_type k_begin = 0; k_begin < k_size; k_begin += block_size) {
Chris@16:                     size_type k_end = k_begin + (std::min) (k_size - k_begin, block_size);
Chris@16: #ifdef BOOST_UBLAS_NO_CACHE
Chris@16:                     const matrix_range<expression1_type> e1_range (e1 (), range (i_begin, i_end), range (k_begin, k_end));
Chris@16:                     const matrix_range<expression2_type> e2_range (e2 (), range (k_begin, k_end), range (j_begin, j_end));
Chris@16: #else
Chris@16:                     // const matrix<value_type, row_major, bounded_array<value_type, block_size * block_size> > e1_range (project (e1 (), range (i_begin, i_end), range (k_begin, k_end)));
Chris@16:                     // const matrix<value_type, column_major, bounded_array<value_type, block_size * block_size> > e2_range (project (e2 (), range (k_begin, k_end), range (j_begin, j_end)));
Chris@16:                     const matrix<value_type, row_major> e1_range (project (e1 (), range (i_begin, i_end), range (k_begin, k_end)));
Chris@16:                     const matrix<value_type, column_major> e2_range (project (e2 (), range (k_begin, k_end), range (j_begin, j_end)));
Chris@16: #endif
Chris@16:                     m_range.plus_assign (prod (e1_range, e2_range));
Chris@16:                 }
Chris@16: #ifndef BOOST_UBLAS_NO_CACHE
Chris@16:                 project (m, range (i_begin, i_end), range (j_begin, j_end)).assign (m_range);
Chris@16: #endif
Chris@16:             }
Chris@16:         }
Chris@16: #if BOOST_UBLAS_TYPE_CHECK
Chris@16:         disable_type_check<bool>::value = false;
Chris@16:         BOOST_UBLAS_CHECK (norm_1 (m - cm) <= 2 * std::numeric_limits<real_type>::epsilon () * merrorbound, internal_logic ());
Chris@16: #endif
Chris@16:         return m;
Chris@16:     }
Chris@16: 
Chris@16:     template<class M, typename M::size_type BS, class E1, class E2>
Chris@16:     BOOST_UBLAS_INLINE
Chris@16:     M
Chris@16:     block_prod (const matrix_expression<E1> &e1,
Chris@16:                 const matrix_expression<E2> &e2,
Chris@16:                 column_major_tag) {
Chris@16:         typedef M matrix_type;
Chris@16:         typedef const E1 expression1_type;
Chris@16:         typedef const E2 expression2_type;
Chris@16:         typedef typename M::size_type size_type;
Chris@16:         typedef typename M::value_type value_type;
Chris@16:         const size_type block_size = BS;
Chris@16: 
Chris@16:         M m (e1 ().size1 (), e2 ().size2 ());
Chris@16: #if BOOST_UBLAS_TYPE_CHECK
Chris@16:         matrix<value_type, column_major> cm (m.size1 (), m.size2 ());
Chris@16:         typedef typename type_traits<value_type>::real_type real_type;
Chris@16:         real_type merrorbound (norm_1 (m) + norm_1 (e1) * norm_1 (e2));
Chris@16:         indexing_matrix_assign<scalar_assign> (cm, prod (e1, e2), column_major_tag ());
Chris@16:         disable_type_check<bool>::value = true;
Chris@16: #endif
Chris@16:         size_type i_size = e1 ().size1 ();
Chris@16:         size_type j_size = e2 ().size2 ();
Chris@16:         size_type k_size = BOOST_UBLAS_SAME (e1 ().size2 (), e2 ().size1 ());
Chris@16:         for (size_type j_begin = 0; j_begin < j_size; j_begin += block_size) {
Chris@16:             size_type j_end = j_begin + (std::min) (j_size - j_begin, block_size);
Chris@16:             for (size_type i_begin = 0; i_begin < i_size; i_begin += block_size) {
Chris@16:                 size_type i_end = i_begin + (std::min) (i_size - i_begin, block_size);
Chris@16:                 // FIX: never ignore Martin Weiser's advice ;-(
Chris@16: #ifdef BOOST_UBLAS_NO_CACHE
Chris@16:                 matrix_range<matrix_type> m_range (m, range (i_begin, i_end), range (j_begin, j_end));
Chris@16: #else
Chris@16:                 // matrix<value_type, column_major, bounded_array<value_type, block_size * block_size> > m_range (i_end - i_begin, j_end - j_begin);
Chris@16:                 matrix<value_type, column_major> m_range (i_end - i_begin, j_end - j_begin);
Chris@16: #endif
Chris@16:                 m_range.assign (zero_matrix<value_type> (i_end - i_begin, j_end - j_begin));
Chris@16:                 for (size_type k_begin = 0; k_begin < k_size; k_begin += block_size) {
Chris@16:                     size_type k_end = k_begin + (std::min) (k_size - k_begin, block_size);
Chris@16: #ifdef BOOST_UBLAS_NO_CACHE
Chris@16:                     const matrix_range<expression1_type> e1_range (e1 (), range (i_begin, i_end), range (k_begin, k_end));
Chris@16:                     const matrix_range<expression2_type> e2_range (e2 (), range (k_begin, k_end), range (j_begin, j_end));
Chris@16: #else
Chris@16:                     // const matrix<value_type, row_major, bounded_array<value_type, block_size * block_size> > e1_range (project (e1 (), range (i_begin, i_end), range (k_begin, k_end)));
Chris@16:                     // const matrix<value_type, column_major, bounded_array<value_type, block_size * block_size> > e2_range (project (e2 (), range (k_begin, k_end), range (j_begin, j_end)));
Chris@16:                     const matrix<value_type, row_major> e1_range (project (e1 (), range (i_begin, i_end), range (k_begin, k_end)));
Chris@16:                     const matrix<value_type, column_major> e2_range (project (e2 (), range (k_begin, k_end), range (j_begin, j_end)));
Chris@16: #endif
Chris@16:                     m_range.plus_assign (prod (e1_range, e2_range));
Chris@16:                 }
Chris@16: #ifndef BOOST_UBLAS_NO_CACHE
Chris@16:                 project (m, range (i_begin, i_end), range (j_begin, j_end)).assign (m_range);
Chris@16: #endif
Chris@16:             }
Chris@16:         }
Chris@16: #if BOOST_UBLAS_TYPE_CHECK
Chris@16:         disable_type_check<bool>::value = false;
Chris@16:         BOOST_UBLAS_CHECK (norm_1 (m - cm) <= 2 * std::numeric_limits<real_type>::epsilon () * merrorbound, internal_logic ());
Chris@16: #endif
Chris@16:         return m;
Chris@16:     }
Chris@16: 
Chris@16:     // Dispatcher
Chris@16:     template<class M, typename M::size_type BS, class E1, class E2>
Chris@16:     BOOST_UBLAS_INLINE
Chris@16:     M
Chris@16:     block_prod (const matrix_expression<E1> &e1,
Chris@16:                 const matrix_expression<E2> &e2) {
Chris@16:         typedef typename M::orientation_category orientation_category;
Chris@16:         return block_prod<M, BS> (e1, e2, orientation_category ());
Chris@16:     }
Chris@16: 
Chris@16: }}}
Chris@16: 
Chris@16: #endif