| 1 |
//@HEADER |
| 2 |
// ************************************************************************ |
| 3 |
// |
| 4 |
// Kokkos v. 4.0 |
| 5 |
// Copyright (2022) National Technology & Engineering |
| 6 |
// Solutions of Sandia, LLC (NTESS). |
| 7 |
// |
| 8 |
// Under the terms of Contract DE-NA0003525 with NTESS, |
| 9 |
// the U.S. Government retains certain rights in this software. |
| 10 |
// |
| 11 |
// Part of Kokkos, under the Apache License v2.0 with LLVM Exceptions. |
| 12 |
// See https://kokkos.org/LICENSE for license information. |
| 13 |
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 14 |
// |
| 15 |
//@HEADER |
| 16 |
#ifndef KOKKOSBATCHED_GESV_HPP |
| 17 |
#define KOKKOSBATCHED_GESV_HPP |
| 18 |
|
| 19 |
/// \author Kim Liegeois ([email protected]) |
| 20 |
|
| 21 |
#include "KokkosBatched_Util.hpp" |
| 22 |
#include "KokkosBatched_Vector.hpp" |
| 23 |
|
| 24 |
namespace KokkosBatched { |
| 25 |
|
| 26 |
struct Gesv { |
| 27 |
struct StaticPivoting {}; |
| 28 |
struct NoPivoting {}; |
| 29 |
|
| 30 |
using Default = StaticPivoting; |
| 31 |
}; |
| 32 |
|
| 33 |
/// \brief Serial Batched GESV: |
| 34 |
/// |
| 35 |
/// Solve A_l x_l = b_l for all l = 0, ..., N |
| 36 |
/// using a batched LU decomposition, 2 batched triangular solves, and a batched |
| 37 |
/// static pivoting. |
| 38 |
/// |
| 39 |
/// \tparam MatrixType: Input type for the matrix, needs to be a 2D view |
| 40 |
/// \tparam VectorType: Input type for the right-hand side and the solution, |
| 41 |
/// needs to be a 1D view |
| 42 |
/// |
| 43 |
/// \param A [in]: matrix, a rank 2 view |
| 44 |
/// \param X [out]: solution, a rank 1 view |
| 45 |
/// \param B [in]: right-hand side, a rank 1 view |
| 46 |
/// \param tmp [in]: a rank 2 view used to store temporary variable; dimension |
| 47 |
/// must be n x (n+4) where n is the number of rows. |
| 48 |
/// |
| 49 |
/// |
| 50 |
/// Two versions are available (those are chosen based on ArgAlgo): |
| 51 |
/// |
| 52 |
/// 1. NoPivoting: the solver does not use a pivoting strategy, |
| 53 |
/// 2. StaticPivoting: the solver uses a static pivoting strategy that relies |
| 54 |
/// on using |
| 55 |
/// maximal absolute value of row and column to choose pivots and apply |
| 56 |
/// them before calling the LU decomposition. Known limitation: the |
| 57 |
/// currently implemented strategy would not work with some matrices such |
| 58 |
/// as [[2, 1], [1, 0]], when this is the case, the Gesv (if used with |
| 59 |
/// pivoting), will return 1 and print an error message. |
| 60 |
/// |
| 61 |
/// No nested parallel_for is used inside of the function. |
| 62 |
/// |
| 63 |
|
| 64 |
template <typename ArgAlgo> |
| 65 |
struct SerialGesv { |
| 66 |
template <typename MatrixType, typename XVectorType, typename YVectorType> |
| 67 |
KOKKOS_INLINE_FUNCTION static int invoke(const MatrixType A, const XVectorType X, const YVectorType Y, |
| 68 |
const MatrixType tmp); |
| 69 |
|
| 70 |
template <typename MatrixType, typename VectorType> |
| 71 |
[[deprecated]] KOKKOS_INLINE_FUNCTION static int invoke(const MatrixType A, const VectorType X, const VectorType Y, |
| 72 |
const MatrixType tmp) { |
| 73 |
return invoke(A, X, Y, tmp); |
| 74 |
} |
| 75 |
}; |
| 76 |
|
| 77 |
/// \brief Team Batched GESV: |
| 78 |
/// |
| 79 |
/// Solve A_l x_l = b_l for all l = 0, ..., N |
| 80 |
/// using a batched LU decomposition, 2 batched triangular solves, and a batched |
| 81 |
/// static pivoting. |
| 82 |
/// |
| 83 |
/// \tparam MatrixType: Input type for the matrix, needs to be a 2D view |
| 84 |
/// \tparam VectorType: Input type for the right-hand side and the solution, |
| 85 |
/// needs to be a 1D view |
| 86 |
/// |
| 87 |
/// \param member [in]: TeamPolicy member |
| 88 |
/// \param A [in]: matrix, a rank 2 view |
| 89 |
/// \param X [out]: solution, a rank 1 view |
| 90 |
/// \param B [in]: right-hand side, a rank 1 view |
| 91 |
/// |
| 92 |
/// Two versions are available (those are chosen based on ArgAlgo): |
| 93 |
/// |
| 94 |
/// 1. NoPivoting: the solver does not use a pivoting strategy, |
| 95 |
/// 2. StaticPivoting: the solver uses a static pivoting strategy that relies |
| 96 |
/// on using |
| 97 |
/// maximal absolute value of row and column to choose pivots and apply |
| 98 |
/// them before calling the LU decomposition. Known limitation: the |
| 99 |
/// currently implemented strategy would not work with some matrices such |
| 100 |
/// as [[2, 1], [1, 0]], when this is the case, the Gesv (if used with |
| 101 |
/// pivoting), will return 1 and print an error message. |
| 102 |
/// |
| 103 |
/// A nested parallel_for with TeamThreadRange is used. |
| 104 |
/// |
| 105 |
|
| 106 |
template <typename MemberType, typename ArgAlgo> |
| 107 |
struct TeamGesv { |
| 108 |
template <typename MatrixType, typename VectorType> |
| 109 |
KOKKOS_INLINE_FUNCTION static int invoke(const MemberType &member, const MatrixType A, const VectorType X, |
| 110 |
const VectorType Y); |
| 111 |
}; |
| 112 |
|
| 113 |
/// \brief Team Vector Batched GESV: |
| 114 |
/// |
| 115 |
/// Solve A_l x_l = b_l for all l = 0, ..., N |
| 116 |
/// using a batched LU decomposition, 2 batched triangular solves, and a batched |
| 117 |
/// static pivoting. |
| 118 |
/// |
| 119 |
/// \tparam MatrixType: Input type for the matrix, needs to be a 2D view |
| 120 |
/// \tparam VectorType: Input type for the right-hand side and the solution, |
| 121 |
/// needs to be a 1D view |
| 122 |
/// |
| 123 |
/// \param member [in]: TeamPolicy member |
| 124 |
/// \param A [in]: matrix, a rank 2 view |
| 125 |
/// \param X [out]: solution, a rank 1 view |
| 126 |
/// \param B [in]: right-hand side, a rank 1 view |
| 127 |
/// |
| 128 |
/// Two versions are available (those are chosen based on ArgAlgo): |
| 129 |
/// |
| 130 |
/// 1. NoPivoting: the solver does not use a pivoting strategy, |
| 131 |
/// 2. StaticPivoting: the solver uses a static pivoting strategy that relies |
| 132 |
/// on using |
| 133 |
/// maximal absolute value of row and column to choose pivots and apply |
| 134 |
/// them before calling the LU decomposition. Known limitation: the |
| 135 |
/// currently implemented strategy would not work with some matrices such |
| 136 |
/// as [[2, 1], [1, 0]], when this is the case, the Gesv (if used with |
| 137 |
/// pivoting), will return 1 and print an error message. |
| 138 |
/// |
| 139 |
/// Two nested parallel_for with both TeamVectorRange and ThreadVectorRange |
| 140 |
/// (or one with TeamVectorRange) are used inside. |
| 141 |
/// |
| 142 |
|
| 143 |
template <typename MemberType, typename ArgAlgo> |
| 144 |
struct TeamVectorGesv { |
| 145 |
template <typename MatrixType, typename VectorType> |
| 146 |
KOKKOS_INLINE_FUNCTION static int invoke(const MemberType &member, const MatrixType A, const VectorType X, |
| 147 |
const VectorType Y); |
| 148 |
}; |
| 149 |
|
| 150 |
} // namespace KokkosBatched |
| 151 |
|
| 152 |
#include "KokkosBatched_Gesv_Impl.hpp" |
| 153 |
|
| 154 |
#endif |