Shuto Naruse / Mbed 2 deprecated Eurobot2012_Secondary

Eurobot2012_Secondary

Fork of Eurobot_2012_Secondary by Shuto Naruse

tvmet/loop/Gemm.h@1:cc2a9eb0bd55, 2012-10-17 (annotated)

Committer:
narshu
Date:
Wed Oct 17 22:25:31 2012 +0000
Revision:
1:cc2a9eb0bd55
Commit before publishing

Who changed what in which revision?

UserRevisionLine numberNew contents of line
narshu 1:cc2a9eb0bd55 1 /*
narshu 1:cc2a9eb0bd55 2 * Tiny Vector Matrix Library
narshu 1:cc2a9eb0bd55 3 * Dense Vector Matrix Libary of Tiny size using Expression Templates
narshu 1:cc2a9eb0bd55 4 *
narshu 1:cc2a9eb0bd55 5 * Copyright (C) 2001 - 2007 Olaf Petzold <opetzold@users.sourceforge.net>
narshu 1:cc2a9eb0bd55 6 *
narshu 1:cc2a9eb0bd55 7 * This library is free software; you can redistribute it and/or
narshu 1:cc2a9eb0bd55 8 * modify it under the terms of the GNU Lesser General Public
narshu 1:cc2a9eb0bd55 9 * License as published by the Free Software Foundation; either
narshu 1:cc2a9eb0bd55 10 * version 2.1 of the License, or (at your option) any later version.
narshu 1:cc2a9eb0bd55 11 *
narshu 1:cc2a9eb0bd55 12 * This library is distributed in the hope that it will be useful,
narshu 1:cc2a9eb0bd55 13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
narshu 1:cc2a9eb0bd55 14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
narshu 1:cc2a9eb0bd55 15 * Lesser General Public License for more details.
narshu 1:cc2a9eb0bd55 16 *
narshu 1:cc2a9eb0bd55 17 * You should have received a copy of the GNU Lesser General Public
narshu 1:cc2a9eb0bd55 18 * License along with this library; if not, write to the Free Software
narshu 1:cc2a9eb0bd55 19 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
narshu 1:cc2a9eb0bd55 20 *
narshu 1:cc2a9eb0bd55 21 * $Id: Gemm.h,v 1.12 2007-06-23 15:58:59 opetzold Exp $
narshu 1:cc2a9eb0bd55 22 */
narshu 1:cc2a9eb0bd55 23
narshu 1:cc2a9eb0bd55 24 #ifndef TVMET_LOOP_GEMM_H
narshu 1:cc2a9eb0bd55 25 #define TVMET_LOOP_GEMM_H
narshu 1:cc2a9eb0bd55 26
narshu 1:cc2a9eb0bd55 27 namespace tvmet {
narshu 1:cc2a9eb0bd55 28
narshu 1:cc2a9eb0bd55 29 namespace loop {
narshu 1:cc2a9eb0bd55 30
narshu 1:cc2a9eb0bd55 31
narshu 1:cc2a9eb0bd55 32 /**
narshu 1:cc2a9eb0bd55 33 * \class gemm Gemm.h "tvmet/loop/Gemm.h"
narshu 1:cc2a9eb0bd55 34 * \brief class for matrix-matrix product using loop unrolling.
narshu 1:cc2a9eb0bd55 35 * using formula
narshu 1:cc2a9eb0bd55 36 * \f[
narshu 1:cc2a9eb0bd55 37 * M_1\,M_2
narshu 1:cc2a9eb0bd55 38 * \f]
narshu 1:cc2a9eb0bd55 39 * \par Example:
narshu 1:cc2a9eb0bd55 40 * \code
narshu 1:cc2a9eb0bd55 41 * template<class T, std::size_t Rows1, std::size_t Cols1, std::size_t Cols2>
narshu 1:cc2a9eb0bd55 42 * inline
narshu 1:cc2a9eb0bd55 43 * void
narshu 1:cc2a9eb0bd55 44 * prod(const Matrix<T, Rows1, Cols1>& lhs, const Matrix<T, Cols1, Cols2>& rhs,
narshu 1:cc2a9eb0bd55 45 * Matrix<T, Rows1, Cols2>& dest)
narshu 1:cc2a9eb0bd55 46 * {
narshu 1:cc2a9eb0bd55 47 * for (std::size_t i = 0; i != Rows1; ++i) {
narshu 1:cc2a9eb0bd55 48 * for (std::size_t j = 0; j != Cols2; ++j) {
narshu 1:cc2a9eb0bd55 49 * dest(i, j) = tvmet::loop::gemm<Rows1, Cols1, Cols2>().prod(lhs, rhs, i, j);
narshu 1:cc2a9eb0bd55 50 * }
narshu 1:cc2a9eb0bd55 51 * }
narshu 1:cc2a9eb0bd55 52 * }
narshu 1:cc2a9eb0bd55 53 * \endcode
narshu 1:cc2a9eb0bd55 54 * \note The number of rows of rhs matrix have to be equal to cols of lhs matrix.
narshu 1:cc2a9eb0bd55 55 * The result is a (Rows1 x Cols2) matrix.
narshu 1:cc2a9eb0bd55 56 */
narshu 1:cc2a9eb0bd55 57 template<std::size_t Rows1, std::size_t Cols1,
narshu 1:cc2a9eb0bd55 58 std::size_t Cols2>
narshu 1:cc2a9eb0bd55 59 class gemm
narshu 1:cc2a9eb0bd55 60 {
narshu 1:cc2a9eb0bd55 61 gemm(const gemm&);
narshu 1:cc2a9eb0bd55 62 gemm& operator=(const gemm&);
narshu 1:cc2a9eb0bd55 63
narshu 1:cc2a9eb0bd55 64 private:
narshu 1:cc2a9eb0bd55 65 enum {
narshu 1:cc2a9eb0bd55 66 count = Cols1,
narshu 1:cc2a9eb0bd55 67 N = (count+7)/8
narshu 1:cc2a9eb0bd55 68 };
narshu 1:cc2a9eb0bd55 69
narshu 1:cc2a9eb0bd55 70 public:
narshu 1:cc2a9eb0bd55 71 gemm() { }
narshu 1:cc2a9eb0bd55 72
narshu 1:cc2a9eb0bd55 73 public:
narshu 1:cc2a9eb0bd55 74 template<class E1, class E2>
narshu 1:cc2a9eb0bd55 75 static inline
narshu 1:cc2a9eb0bd55 76 typename PromoteTraits<
narshu 1:cc2a9eb0bd55 77 typename E1::value_type,
narshu 1:cc2a9eb0bd55 78 typename E2::value_type
narshu 1:cc2a9eb0bd55 79 >::value_type
narshu 1:cc2a9eb0bd55 80 prod(const E1& lhs, const E2& rhs, std::size_t i, std::size_t j) {
narshu 1:cc2a9eb0bd55 81 typename PromoteTraits<
narshu 1:cc2a9eb0bd55 82 typename E1::value_type,
narshu 1:cc2a9eb0bd55 83 typename E2::value_type
narshu 1:cc2a9eb0bd55 84 >::value_type sum(0);
narshu 1:cc2a9eb0bd55 85 std::size_t k(0);
narshu 1:cc2a9eb0bd55 86 std::size_t n(N);
narshu 1:cc2a9eb0bd55 87
narshu 1:cc2a9eb0bd55 88 // Duff's device
narshu 1:cc2a9eb0bd55 89 switch(count % 8) {
narshu 1:cc2a9eb0bd55 90 case 0: do { sum += lhs(i, k) * rhs(k, j); ++k;
narshu 1:cc2a9eb0bd55 91 case 7: sum += lhs(i, k) * rhs(k, j); ++k;
narshu 1:cc2a9eb0bd55 92 case 6: sum += lhs(i, k) * rhs(k, j); ++k;
narshu 1:cc2a9eb0bd55 93 case 5: sum += lhs(i, k) * rhs(k, j); ++k;
narshu 1:cc2a9eb0bd55 94 case 4: sum += lhs(i, k) * rhs(k, j); ++k;
narshu 1:cc2a9eb0bd55 95 case 3: sum += lhs(i, k) * rhs(k, j); ++k;
narshu 1:cc2a9eb0bd55 96 case 2: sum += lhs(i, k) * rhs(k, j); ++k;
narshu 1:cc2a9eb0bd55 97 case 1: sum += lhs(i, k) * rhs(k, j); ++k;
narshu 1:cc2a9eb0bd55 98 } while(--n != 0);
narshu 1:cc2a9eb0bd55 99 }
narshu 1:cc2a9eb0bd55 100
narshu 1:cc2a9eb0bd55 101 return sum;
narshu 1:cc2a9eb0bd55 102 }
narshu 1:cc2a9eb0bd55 103 };
narshu 1:cc2a9eb0bd55 104
narshu 1:cc2a9eb0bd55 105
narshu 1:cc2a9eb0bd55 106 } // namespace loop
narshu 1:cc2a9eb0bd55 107
narshu 1:cc2a9eb0bd55 108 } // namespace tvmet
narshu 1:cc2a9eb0bd55 109
narshu 1:cc2a9eb0bd55 110 #endif /* TVMET_LOOP_GEMM_H */
narshu 1:cc2a9eb0bd55 111
narshu 1:cc2a9eb0bd55 112 // Local Variables:
narshu 1:cc2a9eb0bd55 113 // mode:C++
narshu 1:cc2a9eb0bd55 114 // tab-width:8
narshu 1:cc2a9eb0bd55 115 // End: