Repository navigation
Expand file tree
/
Copy pathcg_solve.hpp
More file actions
222 lines (185 loc) · 6.66 KB
/
Copy pathcg_solve.hpp
File metadata and controls
222 lines (185 loc) · 6.66 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
#ifndef _cg_solve_hpp_
#define _cg_solve_hpp_
//@HEADER
// ************************************************************************
//
// HPCCG: Simple Conjugate Gradient Benchmark Code
// Copyright (2006) Sandia Corporation
//
// Under terms of Contract DE-AC04-94AL85000, there is a non-exclusive
// license for use of this work by or on behalf of the U.S. Government.
//
// This library is free software; you can redistribute it and/or modify
// it under the terms of the GNU Lesser General Public License as
// published by the Free Software Foundation; either version 2.1 of the
// License, or (at your option) any later version.
//
// This library is distributed in the hope that it will be useful, but
// WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
// Lesser General Public License for more details.
//
// You should have received a copy of the GNU Lesser General Public
// License along with this library; if not, write to the Free Software
// Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307
// USA
// Questions? Contact Michael A. Heroux (maherou@sandia.gov)
//
// ************************************************************************
//@HEADER
#include <cmath>
#include <limits>
#include <Vector_functions.hpp>
#include <mytimer.hpp>
#include <outstream.hpp>
//#include <cuda_profiler_api.h>
namespace miniFE {
template<typename Scalar>
void print_vec(const std::vector<Scalar>& vec, const std::string& name)
{
for(size_t i=0; i<vec.size(); ++i) {
std::cout << name << "["<<i<<"]: " << vec[i] << std::endl;
}
}
template<typename VectorType>
bool breakdown(typename VectorType::ScalarType inner,
const VectorType& v,
const VectorType& w)
{
typedef typename VectorType::ScalarType Scalar;
typedef typename TypeTraits<Scalar>::magnitude_type magnitude;
//This is code that was copied from Aztec, and originally written
//by my hero, Ray Tuminaro.
//
//Assuming that inner = <v,w> (inner product of v and w),
//v and w are considered orthogonal if
// |inner| < 100 * ||v||_2 * ||w||_2 * epsilon
magnitude vnorm = std::sqrt(dot(v,v));
magnitude wnorm = std::sqrt(dot(w,w));
return std::abs(inner) <= 100*vnorm*wnorm*std::numeric_limits<magnitude>::epsilon();
}
template<typename OperatorType,
typename VectorType>
void
cg_solve(OperatorType& A,
const VectorType& b,
VectorType& x,
typename OperatorType::LocalOrdinalType max_iter,
typename TypeTraits<typename OperatorType::ScalarType>::magnitude_type& tolerance,
typename OperatorType::LocalOrdinalType& num_iters,
typename TypeTraits<typename OperatorType::ScalarType>::magnitude_type& normr,
timer_type* my_cg_times)
{
typedef typename OperatorType::ScalarType ScalarType;
typedef typename OperatorType::GlobalOrdinalType GlobalOrdinalType;
typedef typename OperatorType::LocalOrdinalType LocalOrdinalType;
typedef typename TypeTraits<ScalarType>::magnitude_type magnitude_type;
if(A.mat.numRows() ==0)
A.init_kokkos_matrix();
A.host_to_device(); A.on_device();
x.coefs.host_to_device(); x.coefs.on_device();
b.coefs.host_to_device(); //b.coefs.on_device();
timer_type t0 = 0, tWAXPY = 0, tDOT = 0, tMATVEC = 0, tMATVECDOT = 0;
int myproc = 0;
#ifdef HAVE_MPI
MPI_Comm_rank(MPI_COMM_WORLD, &myproc);
#endif
if (!A.has_local_indices) {
std::cerr << "miniFE::cg_solve ERROR, A.has_local_indices is false, needs to be true. This probably means "
<< "miniFE::make_local_matrix(A) was not called prior to calling miniFE::cg_solve."
<< std::endl;
return;
}
size_t nrows = A.rows.size();
LocalOrdinalType ncols = A.num_cols;
VectorType r(b.startIndex, nrows);
VectorType p(0, ncols);
VectorType Ap(b.startIndex, nrows);
normr = 0;
magnitude_type rtrans = 0;
magnitude_type oldrtrans = 0;
LocalOrdinalType print_freq = max_iter/10;
if (print_freq>50) print_freq = 50;
if (print_freq<1) print_freq = 1;
ScalarType >
ScalarType zero = 0.0;
// Copy data from CPU (on Host assembly) to GPU
Ap.coefs.host_to_device(); A.on_device();
r.coefs.host_to_device(); r.coefs.on_device();
p.coefs.host_to_device(); p.coefs.on_device();
timer_type total_time = mytimer();
TICK(); waxpby(one, x, zero, x, p); TOCK(tWAXPY);
TICK();
matvec(A, p, Ap);
TOCK(tMATVEC);
TICK(); waxpby(one, b, -one, Ap, r); TOCK(tWAXPY);
TICK(); rtrans = dot(r, r); TOCK(tDOT);
//std::cout << "rtrans="<<rtrans<<std::endl;
normr = std::sqrt(rtrans);
if (myproc == 0) {
std::cout << "Initial Residual = "<< normr << std::endl;
}
magnitude_type brkdown_tol = std::numeric_limits<magnitude_type>::epsilon();
#ifdef MINIFE_DEBUG
std::ostream& os = outstream();
os << "brkdown_tol = " << brkdown_tol << std::endl;
#endif
for(LocalOrdinalType k=1; k <= max_iter && normr > tolerance; ++k) {
if (k == 1) {
TICK(); waxpby(one, r, zero, r, p); TOCK(tWAXPY);
}
else {
oldrtrans = rtrans;
TICK(); rtrans = dot(r, r); TOCK(tDOT);
magnitude_type beta = rtrans/oldrtrans;
TICK(); waxpby(beta, p, one, r, p); TOCK(tWAXPY);
}
normr = std::sqrt(rtrans);
if (myproc == 0 && (k%print_freq==0 || k==max_iter)) {
std::cout << "Iteration = "<<k<<" Residual = "<<normr<<std::endl;
}
magnitude_type alpha = 0;
magnitude_type p_ap_dot = 0;
#ifdef MINIFE_FUSED
TICK();
p_ap_dot = matvec_and_dot(A, p, Ap);
TOCK(tMATVECDOT);
#else
TICK(); matvec(A, p, Ap); TOCK(tMATVEC);
TICK(); p_ap_dot = dot(Ap, p); TOCK(tDOT);
#endif
#ifdef MINIFE_DEBUG
os << "iter " << k << ", p_ap_dot = " << p_ap_dot;
os.flush();
#endif
if (p_ap_dot < brkdown_tol) {
if (p_ap_dot < 0 || breakdown(p_ap_dot, Ap, p)) {
std::cerr << "miniFE::cg_solve ERROR, numerical breakdown!"<<std::endl;
#ifdef MINIFE_DEBUG
os << "ERROR, numerical breakdown!"<<std::endl;
#endif
//update the timers before jumping out.
my_cg_times[WAXPY] = tWAXPY;
my_cg_times[DOT] = tDOT;
my_cg_times[MATVEC] = tMATVEC;
my_cg_times[TOTAL] = mytimer() - total_time;
return;
}
else brkdown_tol = 0.1 * p_ap_dot;
}
alpha = rtrans/p_ap_dot;
#ifdef MINIFE_DEBUG
os << ", rtrans = " << rtrans << ", alpha = " << alpha << std::endl;
#endif
TICK(); waxpby(one, x, alpha, p, x);
waxpby(one, r, -alpha, Ap, r); TOCK(tWAXPY);
num_iters = k;
}
my_cg_times[WAXPY] = tWAXPY;
my_cg_times[DOT] = tDOT;
my_cg_times[MATVEC] = tMATVEC;
my_cg_times[MATVECDOT] = tMATVECDOT;
my_cg_times[TOTAL] = mytimer() - total_time;
}
}//namespace miniFE
#endif