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
|
/*
-- MAGMA (version 2.9.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
@date January 2025
@generated from testing/testing_zgemv.cpp, normal z -> s, Wed Jan 22 14:40:19 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 "magma_operators.h"
#include "testings.h"
/* ////////////////////////////////////////////////////////////////////////////
-- Testing sgemv
*/
int main(int argc, char **argv)
{
#ifdef MAGMA_HAVE_OPENCL
#define dA(i_, j_) dA, ((i_) + (j_)*ldda)
#define dX(i_) dX, ((i_))
#define dY(i_) dY, ((i_))
#else
#define dA(i_, j_) (dA + (i_) + (j_)*ldda)
#define dX(i_) (dX + (i_))
#define dY(i_) (dY + (i_))
#endif
TESTING_CHECK( magma_init() );
magma_print_environment();
real_Double_t gflops, magma_perf, magma_time, dev_perf, dev_time, cpu_perf, cpu_time;
float magma_error, dev_error, work[1];
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t M, N, Xm, Ym, lda, ldda, sizeA, sizeX, sizeY;
magma_int_t incx = 1;
magma_int_t incy = 1;
float c_neg_one = MAGMA_S_NEG_ONE;
float alpha = MAGMA_S_MAKE( 1.5, -2.3 );
float beta = MAGMA_S_MAKE( -0.6, 0.8 );
float *A, *X, *Y, *Ydev, *Ymagma;
magmaFloat_ptr dA, dX, dY;
int status = 0;
// used only with CUDA
MAGMA_UNUSED( magma_perf );
MAGMA_UNUSED( magma_time );
MAGMA_UNUSED( magma_error );
magma_opts opts;
opts.parse_opts( argc, argv );
// Allow 3*eps; real needs 2*sqrt(2) factor; see Higham, 2002, sec. 3.6.
float eps = lapackf77_slamch("E");
float tol = 3*eps;
printf("%% trans = %s\n", lapack_trans_const(opts.transA) );
#if defined(MAGMA_HAVE_CUDA) || defined(MAGMA_HAVE_HIP)
printf("%% M N MAGMA Gflop/s (ms) %s Gflop/s (ms) CPU Gflop/s (ms) MAGMA error %s error\n",
g_platform_str, g_platform_str );
#else
printf("%% M N %s Gflop/s (ms) CPU Gflop/s (ms) %s error\n",
g_platform_str, g_platform_str );
#endif
printf("%%==================================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
lda = M;
ldda = magma_roundup( M, opts.align ); // multiple of 32 by default
gflops = FLOPS_SGEMV( M, N ) / 1e9;
if ( opts.transA == MagmaNoTrans ) {
Xm = N;
Ym = M;
} else {
Xm = M;
Ym = N;
}
sizeA = lda*N;
sizeX = incx*Xm;
sizeY = incy*Ym;
TESTING_CHECK( magma_smalloc_cpu( &A, sizeA ));
TESTING_CHECK( magma_smalloc_cpu( &X, sizeX ));
TESTING_CHECK( magma_smalloc_cpu( &Y, sizeY ));
TESTING_CHECK( magma_smalloc_cpu( &Ydev, sizeY ));
TESTING_CHECK( magma_smalloc_cpu( &Ymagma, sizeY ));
TESTING_CHECK( magma_smalloc( &dA, ldda*N ));
TESTING_CHECK( magma_smalloc( &dX, sizeX ));
TESTING_CHECK( magma_smalloc( &dY, sizeY ));
/* Initialize the matrix */
lapackf77_slarnv( &ione, ISEED, &sizeA, A );
lapackf77_slarnv( &ione, ISEED, &sizeX, X );
lapackf77_slarnv( &ione, ISEED, &sizeY, Y );
// for error checks
float Anorm = lapackf77_slange( "F", &M, &N, A, &lda, work );
float Xnorm = lapackf77_slange( "F", &Xm, &ione, X, &Xm, work );
float Ynorm = lapackf77_slange( "F", &Ym, &ione, Y, &Ym, work );
/* =====================================================================
Performs operation using cuBLAS / clBLAS
=================================================================== */
magma_ssetmatrix( M, N, A, lda, dA(0,0), ldda, opts.queue );
magma_ssetvector( Xm, X, incx, dX(0), incx, opts.queue );
magma_ssetvector( Ym, Y, incy, dY(0), incy, opts.queue );
magma_flush_cache( opts.cache );
dev_time = magma_sync_wtime( opts.queue );
magma_sgemv( opts.transA, M, N,
alpha, dA(0,0), ldda,
dX(0), incx,
beta, dY(0), incy, opts.queue );
dev_time = magma_sync_wtime( opts.queue ) - dev_time;
dev_perf = gflops / dev_time;
magma_sgetvector( Ym, dY(0), incy, Ydev, incy, opts.queue );
/* =====================================================================
Performs operation using MAGMABLAS (currently only with CUDA)
=================================================================== */
#if defined(MAGMA_HAVE_CUDA) || defined(MAGMA_HAVE_HIP)
magma_ssetvector( Ym, Y, incy, dY(0), incy, opts.queue );
magma_flush_cache( opts.cache );
magma_time = magma_sync_wtime( opts.queue );
magmablas_sgemv( opts.transA, M, N,
alpha, dA(0,0), ldda,
dX(0), incx,
beta, dY(0), incy, opts.queue );
magma_time = magma_sync_wtime( opts.queue ) - magma_time;
magma_perf = gflops / magma_time;
magma_sgetvector( Ym, dY(0), incy, Ymagma, incy, opts.queue );
#endif
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
magma_flush_cache( opts.cache );
cpu_time = magma_wtime();
blasf77_sgemv( lapack_trans_const(opts.transA), &M, &N,
&alpha, A, &lda,
X, &incx,
&beta, Y, &incy );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
/* =====================================================================
Check the result
=================================================================== */
// See testing_sgemm for formula. Here K = Xm.
blasf77_saxpy( &Ym, &c_neg_one, Y, &incy, Ydev, &incy );
dev_error = lapackf77_slange( "F", &Ym, &ione, Ydev, &Ym, work )
/ (sqrt(float(Xm+2))*fabs(alpha)*Anorm*Xnorm + 2*fabs(beta)*Ynorm);
// Really tall or wide (e.g., 200000 x 10) matrices need looser bound.
// TODO: investigate why.
tol = (M < 20000 && N < 20000 ? 3*eps : opts.tolerance*eps);
#if defined(MAGMA_HAVE_CUDA) || defined(MAGMA_HAVE_HIP)
blasf77_saxpy( &Ym, &c_neg_one, Y, &incy, Ymagma, &incy );
magma_error = lapackf77_slange( "F", &Ym, &ione, Ymagma, &Ym, work )
/ (sqrt(float(N+2))*fabs(alpha)*Anorm*Xnorm + 2*fabs(beta)*Ynorm);
bool okay = (magma_error < tol) && (dev_error < tol);
status += ! okay;
printf("%5lld %5lld %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %s\n",
(long long) M, (long long) N,
magma_perf, 1000.*magma_time,
dev_perf, 1000.*dev_time,
cpu_perf, 1000.*cpu_time,
magma_error, dev_error,
(okay ? "ok" : "failed"));
#else
bool okay = (dev_error < tol);
status += ! okay;
printf("%5lld %5lld %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(long long) M, (long long) N,
dev_perf, 1000.*dev_time,
cpu_perf, 1000.*cpu_time,
dev_error,
(okay ? "ok" : "failed"));
#endif
magma_free_cpu( A );
magma_free_cpu( X );
magma_free_cpu( Y );
magma_free_cpu( Ydev );
magma_free_cpu( Ymagma );
magma_free( dA );
magma_free( dX );
magma_free( dY );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_CHECK( magma_finalize() );
return status;
}
|
