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
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
|
/*
-- MAGMA (version 2.9.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
@date January 2025
@generated from testing/testing_zhemm_mgpu.cpp, normal z -> c, Wed Jan 22 14:40:21 2025
@author Mark Gates
@author Azzam Haidar
*/
// 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 magma_chemm_mgpu
*/
int main( int argc, char** argv)
{
TESTING_CHECK( magma_init() );
magma_print_environment();
const magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
const magmaFloatComplex alpha = MAGMA_C_MAKE( 3.456, 5.678 );
const magmaFloatComplex beta = MAGMA_C_MAKE( 1.234, 2.456 );
const magma_int_t ione = 1;
real_Double_t gflops, gpu_perf=0., cpu_perf=0., gpu_time=0., cpu_time=0.;
real_Double_t gpu_perf2=0., gpu_time2=0.;
float error, work[1];
magmaFloatComplex *hA, *hB, *hC, *hR;
magmaFloatComplex_ptr dA[MagmaMaxGPUs], dB[MagmaMaxGPUs], dC[MagmaMaxGPUs], dwork[MagmaMaxGPUs];
magmaFloatComplex_ptr dA2;
magma_int_t i, j, dev, M, N, K, Moff, Noff, Koff, size, lda, ldb, ldc, ldda, lddb, lddc, nb;
magma_int_t iseed[4] = {0,0,0,1};
int status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
opts.ngpu = abs( opts.ngpu ); // always uses multi-GPU code
// See testing_cgemm about tolerance.
float eps = lapackf77_slamch("E");
float tol = 3*eps;
// default values
nb = (opts.nb > 0 ? opts.nb : 64);
magma_int_t nodes[MagmaMaxGPUs][MagmaMaxGPUs+2];
magma_int_t nnode = 0;
magma_buildconnection_mgpu( nodes, &nnode, opts.ngpu );
printf("%% GPU communication pattern: nnode %lld\n", (long long) nnode);
for (i = 0; i < nnode; ++i) {
magma_int_t my_ndev = nodes[i][MagmaMaxGPUs];
printf("%% node %lld has %lld GPUs:", (long long) i, (long long) my_ndev );
for (j=0; j < my_ndev; ++j) {
printf(" %lld", (long long) nodes[i][j] );
if (j < my_ndev-1) {
printf(",");
}
}
printf("\n");
}
// number of queues per GPU. Requires ngpu.
magma_int_t nqueue = opts.ngpu;
// number of events per GPU. Requires ngpu.
magma_int_t nevents = opts.ngpu;
magma_queue_t queues[MagmaMaxGPUs][20], queues0[MagmaMaxGPUs];
magma_event_t events[MagmaMaxGPUs][MagmaMaxGPUs*MagmaMaxGPUs + 10];
for( dev = 0; dev < opts.ngpu; ++dev ) {
magma_setdevice( dev );
for( i = 0; i < nqueue; ++i ) {
magma_queue_create( dev, &queues[dev][i] );
}
queues0[dev] = queues[dev][0];
for( i = 0; i < nevents; ++i ) {
magma_event_create( &events[dev][i] );
}
}
printf("%% nb %lld, ngpu %lld, version %lld\n", (long long) nb, (long long) opts.ngpu, (long long) opts.version );
printf("%% M N nb offset CPU Gflop/s (sec) GPU Gflop/s (sec) %s hemm (sec) ||R|| / ||A||*||B||\n", g_platform_str);
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];
K = (opts.side == MagmaLeft ? M : N);
lda = K;
ldb = M;
ldc = M;
ldda = magma_roundup( lda, opts.align ); // multiple of 32 by default
lddb = magma_roundup( ldb, opts.align ); // multiple of 32 by default
lddc = magma_roundup( ldc, opts.align ); // multiple of 32 by default
magma_int_t lwork = lddc*N + (M*N)*opts.ngpu;
TESTING_CHECK( magma_cmalloc_cpu( &hA, lda*K ));
TESTING_CHECK( magma_cmalloc_cpu( &hB, ldb*N ));
TESTING_CHECK( magma_cmalloc_cpu( &hC, ldc*N ));
TESTING_CHECK( magma_cmalloc_pinned( &hR, ldc*N ));
for( dev = 0; dev < opts.ngpu; ++dev ) {
magma_int_t Klocal = ((K / nb) / opts.ngpu + 1) * nb; // over estimate
magma_setdevice( dev );
TESTING_CHECK( magma_cmalloc( &dA[dev], ldda*Klocal ));
TESTING_CHECK( magma_cmalloc( &dB[dev], lddb*N ));
TESTING_CHECK( magma_cmalloc( &dC[dev], lddc*N ));
TESTING_CHECK( magma_cmalloc( &dwork[dev], lwork ));
}
if ( opts.check ) {
// for running cuBLAS
magma_setdevice( 0 );
TESTING_CHECK( magma_cmalloc( &dA2, ldda*K ));
}
for( int offset = 0; offset < N; offset += nb/2 ) {
if (opts.side == MagmaLeft) {
Moff = M - offset;
Noff = N;
}
else {
Moff = M;
Noff = N - offset;
}
Koff = K - offset;
gflops = FLOPS_CHEMM( MagmaLeft, Moff, Noff ) / 1e9;
size = lda*K;
lapackf77_clarnv( &ione, iseed, &size, hA );
// set opposite triangle to NAN to ensure we don't use it.
magma_int_t K1 = K - 1;
if (opts.uplo == MagmaLower)
lapackf77_claset( "Upper", &K1, &K1, &MAGMA_C_NAN, &MAGMA_C_NAN, hA + lda, &lda );
else
lapackf77_claset( "Lower", &K1, &K1, &MAGMA_C_NAN, &MAGMA_C_NAN, hA + 1, &lda );
size = ldb*N;
lapackf77_clarnv( &ione, iseed, &size, hB );
size = ldc*N;
lapackf77_clarnv( &ione, iseed, &size, hC );
lapackf77_clacpy( "Full", &M, &N, hC, &ldc, hR, &ldc );
// for error checks
float Anorm = safe_lapackf77_clanhe( "F", lapack_uplo_const(opts.uplo), &Koff, &hA[offset + offset*lda], &lda, work );
float Bnorm = lapackf77_clange( "F", &Moff, &Noff, hB, &ldb, work );
float Cnorm = lapackf77_clange( "F", &Moff, &Noff, hC, &ldc, work );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_csetmatrix_1D_col_bcyclic( opts.ngpu, K, K, nb, hA, lda, dA, ldda, queues0 );
for( dev = 0; dev < opts.ngpu; ++dev ) {
magma_setdevice( dev );
magma_csetmatrix( M, N, hB, ldb, dB[dev], lddb, opts.queue );
// send C to GPU corresponding to offset; not needed everywhere
magma_int_t start_dev = (offset / nb) % opts.ngpu;
if (dev == start_dev) {
magma_csetmatrix( M, N, hC, ldc, dC[dev], lddc, opts.queue );
}
}
// chemm_mgpu is synchronous, unlike async, single GPU BLAS
gpu_time = magma_wtime();
// only Left, Lower supported
magmablas_chemm_mgpu(
opts.side, opts.uplo, Moff, Noff,
alpha, dA, ldda, offset,
dB, lddb,
beta, dC, lddc, dwork, lwork,
opts.ngpu, nb, queues, nqueue, events, nevents, nodes, nnode );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
/* ====================================================================
Performs operation using CUBLAS
=================================================================== */
if ( opts.check && iter == 0 ) {
magma_setdevice( 0 );
magma_csetmatrix( K, K, hA, lda, dA2, ldda, opts.queue );
magma_csetmatrix( M, N, hB, ldb, dB[0], lddb, opts.queue );
magma_csetmatrix( M, N, hC, ldc, dwork[0], lddc, opts.queue );
gpu_time2 = magma_sync_wtime(0);
magma_chemm(
opts.side, opts.uplo, Moff, Noff,
alpha, dA2 + offset + offset*ldda, ldda,
dB[0], lddb,
beta, dwork[0], lddc, opts.queue );
gpu_time2 = magma_sync_wtime(0) - gpu_time2;
gpu_perf2 = gflops / gpu_time2;
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.check ) {
cpu_time = magma_wtime();
blasf77_chemm( lapack_side_const(opts.side),
lapack_uplo_const(opts.uplo), &Moff, &Noff,
&alpha, hA + offset + offset*lda, &lda,
hB, &ldb,
&beta, hC, &ldc );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
for (dev=0; dev < opts.ngpu; ++dev) {
magma_setdevice( dev );
magma_cgetmatrix( M, N, dC[dev], lddc, hR, ldc, opts.queue );
// See testing_cgemm for formula.
size = ldc*N;
blasf77_caxpy( &size, &c_neg_one, hC, &ione, hR, &ione );
error = lapackf77_clange( "F", &Moff, &Noff, hR, &ldc, work )
/ (sqrt(float(Koff+2))*fabs(alpha)*Anorm*Bnorm + 2*fabs(beta)*Cnorm);
bool okay = (error < tol);
status += ! okay;
if (dev == 0) {
printf( "%5lld %5lld %5lld %5lld %7.1f (%7.4f) %7.1f (%7.4f) %7.1f (%7.4f) %8.2e %s\n",
(long long) M, (long long) N, (long long) nb, (long long) offset,
cpu_perf, cpu_time,
gpu_perf, gpu_time,
gpu_perf2, gpu_time2,
error, (okay ? "ok" : "failed") );
}
else {
printf( " dev %lld %74s %8.2e %s\n",
(long long) dev, "",
error, (okay ? "ok" : "failed") );
}
}
}
else {
printf( "%5lld %5lld %5lld %5lld --- ( --- ) %7.1f (%7.4f) --- ( --- ) ---\n",
(long long) M, (long long) N, (long long) nb, (long long) offset,
gpu_perf, gpu_time );
}
} // offset
printf( "\n" );
magma_free_cpu( hA );
magma_free_cpu( hB );
magma_free_cpu( hC );
magma_free_pinned( hR );
for( dev = 0; dev < opts.ngpu; ++dev ) {
magma_setdevice( dev );
magma_free( dA[dev] );
magma_free( dB[dev] );
magma_free( dC[dev] );
magma_free( dwork[dev] );
}
if ( opts.check ) {
magma_setdevice( 0 );
magma_free( dA2 );
}
fflush( stdout );
} // iter
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
for( dev = 0; dev < opts.ngpu; ++dev ) {
magma_setdevice( dev );
for( i = 0; i < nqueue; ++i ) {
magma_queue_destroy( queues[dev][i] );
}
for( i = 0; i < nevents; ++i ) {
magma_event_destroy( events[dev][i] );
}
}
opts.cleanup();
TESTING_CHECK( magma_finalize() );
return status;
}
|
