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
|
/*
-- MAGMA (version 2.9.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
@date January 2025
@generated from testing/testing_ztrtri_diag.cpp, normal z -> d, Wed Jan 22 14:40:25 2025
*/
// includes, system
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <math.h>
// includes, project
#include "flops.h"
#include "magma_v2.h"
#include "magma_lapack.h"
#include "testings.h"
#define h_A(i_, j_) (h_A + (i_) + (j_)*lda)
#define h_dinvA(i_, j_) (h_dinvA + (i_) + (j_)*nb)
/* ////////////////////////////////////////////////////////////////////////////
-- like axpy for matrices: B += alpha*A.
*/
void dgeadd(
magma_int_t m, magma_int_t n,
double alpha,
const double *A, magma_int_t lda,
double *B, magma_int_t ldb )
{
#define A(i_, j_) (A + (i_) + (j_)*lda)
#define B(i_, j_) (B + (i_) + (j_)*ldb)
const magma_int_t ione = 1;
magma_int_t j;
for( j=0; j < n; ++j ) {
blasf77_daxpy( &m, &alpha, A(0,j), &ione, B(0,j), &ione );
}
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing dtrtri
*/
int main( int argc, char** argv )
{
TESTING_CHECK( magma_init() );
magma_print_environment();
real_Double_t gflops, magma_perf, magma_time=0; //, cpu_perf=0, cpu_time=0;
double magma_error, norm_invA, work[1];
magma_int_t i, j, N, lda, ldda, info;
magma_int_t jb, nb, nblock, size_inv;
magma_int_t *ipiv;
double *h_A, *h_dinvA;
magmaDouble_ptr d_A, d_dinvA;
double c_neg_one = MAGMA_D_NEG_ONE;
int status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
opts.lapack |= opts.check; // check (-c) implies lapack (-l)
double tol = opts.tolerance * lapackf77_dlamch("E");
const char *uplo_ = lapack_uplo_const(opts.uplo);
// this is the NB hard coded into dtrtri_diag.
nb = 128;
printf("%% uplo = %s, diag = %s\n",
lapack_uplo_const(opts.uplo), lapack_diag_const(opts.diag) );
printf("%% N MAGMA Gflop/s (ms) MAGMA error\n");
printf("%%======================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
ldda = magma_roundup( lda, opts.align ); // multiple of 32 by default
nblock = magma_ceildiv( N, nb );
gflops = nblock * FLOPS_DTRTRI( nb ) / 1e9;
TESTING_CHECK( magma_dmalloc_cpu( &h_A, lda*N ));
TESTING_CHECK( magma_imalloc_cpu( &ipiv, N ));
size_inv = nblock*nb*nb;
TESTING_CHECK( magma_dmalloc( &d_A, ldda*N ));
TESTING_CHECK( magma_dmalloc( &d_dinvA, size_inv ));
TESTING_CHECK( magma_dmalloc_cpu( &h_dinvA, size_inv ));
/* Initialize the matrices */
/* Factor A into LU to get well-conditioned triangular matrix.
* Copy L to U, since L seems okay when used with non-unit diagonal
* (i.e., from U), while U fails when used with unit diagonal. */
//sizeA = lda*N;
magma_generate_matrix( opts, N, N, h_A, lda );
lapackf77_dgetrf( &N, &N, h_A, &lda, ipiv, &info );
for( j = 0; j < N; ++j ) {
for( i = 0; i < j; ++i ) {
*h_A(i,j) = *h_A(j,i);
}
}
/* =====================================================================
Performs operation using MAGMABLAS
=================================================================== */
magma_dsetmatrix( N, N, h_A, lda, d_A, ldda, opts.queue );
magma_time = magma_sync_wtime( opts.queue );
magmablas_dtrtri_diag( opts.uplo, opts.diag, N, d_A, ldda, d_dinvA, opts.queue );
magma_time = magma_sync_wtime( opts.queue ) - magma_time;
magma_perf = gflops / magma_time;
magma_dgetvector( size_inv, d_dinvA, 1, h_dinvA, 1, opts.queue );
if ( opts.verbose ) {
printf( "A%lld=", (long long) N );
magma_dprint( N, N, h_A, lda );
printf( "d_dinvA%lld=", (long long) N );
magma_dprint( min(N+4, nb), min(N+4, nblock*nb), h_dinvA, nb );
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
//cpu_time = magma_wtime();
lapackf77_dtrtri(
lapack_uplo_const(opts.uplo), lapack_diag_const(opts.diag),
&N, h_A, &lda, &info );
//cpu_time = magma_wtime() - cpu_time;
//cpu_perf = gflops / cpu_time;
}
/* =====================================================================
Check the result
=================================================================== */
if ( opts.check ) {
// |invA - invA_magma| / |invA|, accumulated over all diagonal blocks
magma_error = 0;
norm_invA = 0;
for( i=0; i < N; i += nb ) {
jb = min( nb, N-i );
dgeadd( jb, jb, c_neg_one, h_A(i, i), lda, h_dinvA(0, i), nb );
magma_error = max( magma_error, lapackf77_dlantr( "M", uplo_, MagmaNonUnitStr, &jb, &jb, h_dinvA(0, i), &nb, work ));
norm_invA = max( norm_invA, lapackf77_dlantr( "M", uplo_, MagmaNonUnitStr, &jb, &jb, h_A(i, i), &lda, work ));
}
magma_error /= norm_invA;
// CPU is doing N-by-N inverse, while GPU is doing (N/NB) NB-by-NB inverses.
// So don't compare performance.
printf("%5lld %7.2f (%7.2f) %8.2e %s\n",
(long long) N,
magma_perf, 1000.*magma_time,
//cpu_perf, 1000.*cpu_time,
magma_error,
(magma_error < tol ? "ok" : "failed"));
status += ! (magma_error < tol);
}
else {
printf("%5lld %7.2f (%7.2f) ---\n",
(long long) N,
magma_perf, 1000.*magma_time );
}
magma_free_cpu( h_A );
magma_free_cpu( ipiv );
magma_free( d_A );
magma_free( d_dinvA );
magma_free_cpu( h_dinvA );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_CHECK( magma_finalize() );
return status;
}
|
