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
|
/*
-- MAGMA (version 2.9.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
@date January 2025
@author Mark Gates
@precisions normal z -> c d s
*/
// includes, system
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <math.h>
#include <assert.h>
// includes, project
#include "flops.h"
#include "magma_v2.h"
#include "magma_lapack.h"
#include "magma_operators.h"
#include "testings.h"
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zunmqr
*/
int main( int argc, char** argv )
{
TESTING_CHECK( magma_init() );
magma_print_environment();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
double Cnorm, error, work[1];
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magma_int_t ione = 1;
magma_int_t mm, m, n, k, size, info;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t nb, ldc, lda, lwork, lwork_max;
magmaDoubleComplex *C, *R, *A, *W, *tau;
int status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
// need slightly looser bound (60*eps instead of 30*eps) for some tests
opts.tolerance = max( 60., opts.tolerance );
double tol = opts.tolerance * lapackf77_dlamch("E");
// pass ngpu = -1 to test multi-GPU code using 1 gpu
magma_int_t abs_ngpu = abs( opts.ngpu );
// test all combinations of input parameters
magma_side_t side [] = { MagmaLeft, MagmaRight };
magma_trans_t trans[] = { Magma_ConjTrans, MagmaNoTrans };
printf("%% M N K side trans CPU Gflop/s (sec) GPU Gflop/s (sec) ||R||_F / ||QC||_F\n");
printf("%%==============================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iside = 0; iside < 2; ++iside ) {
for( int itran = 0; itran < 2; ++itran ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
m = opts.msize[itest];
n = opts.nsize[itest];
k = opts.ksize[itest];
nb = magma_get_zgeqrf_nb( m, n );
ldc = m;
// A is mm x k == m x k (left) or n x k (right)
mm = (side[iside] == MagmaLeft ? m : n);
lda = mm;
gflops = FLOPS_ZUNMQR( m, n, k, side[iside] ) / 1e9;
if ( side[iside] == MagmaLeft && m < k ) {
printf( "%5lld %5lld %5lld %4c %5c skipping because side=left and m < k\n",
(long long) m, (long long) n, (long long) k,
lapacke_side_const( side[iside] ),
lapacke_trans_const( trans[itran] ) );
continue;
}
if ( side[iside] == MagmaRight && n < k ) {
printf( "%5lld %5lld %5lld %4c %5c skipping because side=right and n < k\n",
(long long) m, (long long) n, (long long) k,
lapacke_side_const( side[iside] ),
lapacke_trans_const( trans[itran] ) );
continue;
}
// need at least 2*nb*nb for geqrf
lwork_max = max( max( m*nb, n*nb ), 2*nb*nb );
// this rounds it up slightly if needed to agree with lwork query below
lwork_max = magma_int_t( real( magma_zmake_lwork( lwork_max )));
TESTING_CHECK( magma_zmalloc_cpu( &C, ldc*n ));
TESTING_CHECK( magma_zmalloc_cpu( &R, ldc*n ));
TESTING_CHECK( magma_zmalloc_cpu( &A, lda*k ));
TESTING_CHECK( magma_zmalloc_cpu( &W, lwork_max ));
TESTING_CHECK( magma_zmalloc_cpu( &tau, k ));
// C is full, m x n
size = ldc*n;
lapackf77_zlarnv( &ione, ISEED, &size, C );
lapackf77_zlacpy( "Full", &m, &n, C, &ldc, R, &ldc );
// A is mm x k
magma_generate_matrix( opts, mm, k, A, lda );
// compute QR factorization to get Householder vectors in A, tau
magma_zgeqrf( mm, k, A, lda, tau, W, lwork_max, &info );
if (info != 0) {
printf("magma_zgeqrf returned error %lld: %s.\n",
(long long) info, magma_strerror( info ));
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
lapackf77_zunmqr( lapack_side_const( side[iside] ), lapack_trans_const( trans[itran] ),
&m, &n, &k,
A, &lda, tau, C, &ldc, W, &lwork_max, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0) {
printf("lapackf77_zunmqr returned error %lld: %s.\n",
(long long) info, magma_strerror( info ));
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
// query for workspace size
lwork = -1;
magma_zunmqr( side[iside], trans[itran],
m, n, k,
A, lda, tau, R, ldc, W, lwork, &info );
if (info != 0) {
printf("magma_zunmqr (lwork query) returned error %lld: %s.\n",
(long long) info, magma_strerror( info ));
}
lwork = (magma_int_t) MAGMA_Z_REAL( W[0] );
if ( lwork < 0 || lwork > lwork_max ) {
printf("Warning: optimal lwork %lld > allocated lwork_max %lld\n", (long long) lwork, (long long) lwork_max );
lwork = lwork_max;
}
gpu_time = magma_wtime();
if ( opts.ngpu == 1 ) {
magma_zunmqr( side[iside], trans[itran],
m, n, k,
A, lda, tau, R, ldc, W, lwork, &info );
}
else {
if ( side[iside] == MagmaLeft ) {
magma_zunmqr_m( abs_ngpu, side[iside], trans[itran],
m, n, k,
A, lda, tau, R, ldc, W, lwork, &info );
}
else {
printf( "%5lld %5lld %5lld %4c %5c skipping because magma_zunmqr_m doesn't support MagmaRight\n",
(long long) m, (long long) n, (long long) k,
lapacke_side_const( side[iside] ),
lapacke_trans_const( trans[itran] ) );
goto cleanup;
}
}
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0) {
printf("magma_zunmqr returned error %lld: %s.\n",
(long long) info, magma_strerror( info ));
}
/* =====================================================================
compute relative error |QC_magma - QC_lapack| / |QC_lapack|
=================================================================== */
size = ldc*n;
blasf77_zaxpy( &size, &c_neg_one, C, &ione, R, &ione );
Cnorm = lapackf77_zlange( "Fro", &m, &n, C, &ldc, work );
error = lapackf77_zlange( "Fro", &m, &n, R, &ldc, work ) / (magma_dsqrt(m*n) * Cnorm);
printf( "%5lld %5lld %5lld %4c %5c %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(long long) m, (long long) n, (long long) k,
lapacke_side_const( side[iside] ),
lapacke_trans_const( trans[itran] ),
cpu_perf, cpu_time, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed") );
status += ! (error < tol);
cleanup:
magma_free_cpu( C );
magma_free_cpu( R );
magma_free_cpu( A );
magma_free_cpu( W );
magma_free_cpu( tau );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}} // end iside, itran
printf( "\n" );
}
opts.cleanup();
TESTING_CHECK( magma_finalize() );
return status;
}
|
