File: testing_cgbsv_gpu.cpp

package info (click to toggle)
magma 2.9.0%2Bds-3
  • links: PTS, VCS
  • area: contrib
  • in suites: forky, sid
  • size: 83,556 kB
  • sloc: cpp: 709,115; fortran: 121,916; ansic: 32,343; python: 25,603; f90: 15,208; makefile: 945; xml: 253; csh: 232; sh: 203; perl: 104
file content (220 lines) | stat: -rw-r--r-- 8,086 bytes parent folder | download | duplicates (3)
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
/*
   -- MAGMA (version 2.9.0) --
   Univ. of Tennessee, Knoxville
   Univ. of California, Berkeley
   Univ. of Colorado, Denver
   @date January 2025

   @author Mark gates
   @author Ahmad Abdelfattah

   @generated from testing/testing_zgbsv_gpu.cpp, normal z -> c, Wed Jan 22 14:40:30 2025
 */

// includes, system
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <math.h>

// includes, project
#include "flops.h"
#include "magma_v2.h"
#include "magma_lapack.h"
#include "testings.h"

#if defined(_OPENMP)
#include <omp.h>
#include "../control/magma_threadsetting.h"  // internal header
#endif

/* ////////////////////////////////////////////////////////////////////////////
   -- Testing cgesv_batched
*/
int main(int argc, char **argv)
{
    TESTING_CHECK( magma_init() );
    magma_print_environment();

    real_Double_t   gflops, cpu_perf, cpu_time, gpu_perf, gpu_time = 0;
    float          error, Rnorm, Anorm, Xnorm, *work;
    magmaFloatComplex c_one     = MAGMA_C_ONE;
    magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
    magmaFloatComplex *h_A, *h_B, *h_X;
    magmaFloatComplex_ptr d_A, d_B;
    magma_int_t *dipiv;
    magma_int_t *ipiv;
    magma_int_t N, Nband, KL, KU, KV, nrhs, lda, ldb, ldda, lddb, info = 0, sizeA, sizeB;
    magma_int_t ione     = 1;
    magma_int_t ISEED[4] = {0,0,0,1};
    int status = 0;

    magma_opts opts;
    opts.parse_opts( argc, argv );

    float tol = opts.tolerance * lapackf77_slamch("E");

    nrhs       = opts.nrhs;
    KL         = opts.kl;
    KU         = opts.ku;
    KV         = KL + KU;

    printf("%% ## INFO ##: Gflop/s calculation is not available\n");
    printf("%% Lower bandwidth (KL) = %lld\n", (long long)KL);
    printf("%% Upper bandwidth (KU) = %lld\n", (long long)KU);
    printf("%% N  NRHS   CPU Gflop/s (ms)   GPU Gflop/s (ms)   ||B - AX|| / N*||A||*||X||\n");
    printf("%%============================================================================================\n");
    for( int itest = 0; itest < opts.ntest; ++itest ) {
        for( int iter = 0; iter < opts.niter; ++iter ) {
            N = opts.nsize[itest];
            Nband  = KL + 1 + KV; // need extra KL superdiagonals for the upper factor
            lda    = Nband;
            ldb    = N;
            ldda   = magma_roundup( lda, opts.align );  // multiple of 32 by default
            lddb   = magma_roundup( ldb, opts.align );
            gflops = 0.;  // TODO: gflop formula for gbsv?

            sizeA = lda*N;
            sizeB = ldb*nrhs;

            TESTING_CHECK( magma_cmalloc_cpu( &h_A, sizeA ));
            TESTING_CHECK( magma_cmalloc_cpu( &h_B, sizeB ));
            TESTING_CHECK( magma_cmalloc_cpu( &h_X, sizeB ));
            TESTING_CHECK( magma_smalloc_cpu( &work, N ));
            TESTING_CHECK( magma_imalloc_cpu( &ipiv, N ));

            TESTING_CHECK( magma_cmalloc( &d_A, ldda*N    ));
            TESTING_CHECK( magma_cmalloc( &d_B, lddb*nrhs ));
            TESTING_CHECK( magma_imalloc( &dipiv, N ));

            /* Initialize the matrices */
            lapackf77_clarnv( &ione, ISEED, &sizeA, h_A );
            lapackf77_clarnv( &ione, ISEED, &sizeB, h_B );

            // random initialization of h_A seems to produce
            // some matrices that are singular, the additive statements below
            // seem to avoid that
            #pragma omp parallel for schedule(dynamic)
            for(int j = 0; j < lda*N; j++) {
                MAGMA_C_REAL( h_A[j] ) += 20.;
                #if defined(PRECISION_c) || defined(PRECISION_z)
                MAGMA_C_IMAG( h_A[j] ) += 20.;
                #endif
            }


            magma_csetmatrix( Nband, N,    h_A, lda, d_A, ldda, opts.queue );
            magma_csetmatrix( N,     nrhs, h_B, ldb, d_B, lddb, opts.queue );

            /* ====================================================================
               Performs operation using MAGMA
               =================================================================== */

            if(opts.version == 1) {
                // sync. interface
                gpu_time = magma_wtime();
                magma_cgbsv_native(
                    N, KL, KU, nrhs,
                    d_A, ldda, dipiv,
                    d_B, lddb, &info);
                gpu_time = magma_wtime() - gpu_time;
            }
            else {
                // async. interface

                // query wrokspace
                magma_int_t lwork[1] = {-1};
                magma_cgbsv_native_work(
                    N, KL, KU, nrhs,
                    NULL, ldda, NULL,
                    NULL, lddb, &info, NULL, lwork, opts.queue);

                void* device_work = NULL;
                TESTING_CHECK( magma_malloc(&device_work, lwork[0]) );

                // time async. call only
                gpu_time = magma_sync_wtime( opts.queue );
                magma_cgbsv_native_work(
                    N, KL, KU, nrhs,
                    d_A, ldda, dipiv,
                    d_B, lddb, &info, device_work, lwork, opts.queue);
                gpu_time = magma_sync_wtime( opts.queue ) - gpu_time;

                magma_free( device_work );
            }
            gpu_perf = gflops / gpu_time;

            if (info != 0) {
                printf("magma_cgbsv_gpu returned error %lld: %s.\n",
                        (long long) info, magma_strerror( info ));
            }

            //=====================================================================
            // Residual
            //=====================================================================
            magma_cgetmatrix( N, nrhs, d_B, lddb, h_X, ldb, opts.queue );

            error = 0;
            Anorm = lapackf77_clangb("I", &N, &KL, &KU, h_A + KL, &lda, work);
            Xnorm = lapackf77_clange("I", &N, &nrhs, h_X, &ldb, work);

            for(magma_int_t j = 0; j < nrhs; j++) {
                blasf77_cgbmv( MagmaNoTransStr, &N, &N, &KL, &KU,
                               &c_one, h_A + KL     , &lda,
                                       h_X + j * ldb, &ione,
                           &c_neg_one, h_B + j * ldb, &ione);
            }

            Rnorm = lapackf77_clange("I", &N, &nrhs, h_B, &ldb, work);

            error = Rnorm/(N*Anorm*Xnorm);
            bool okay = (error < tol);
            status += ! okay;

            /* ====================================================================
               Performs operation using LAPACK
               =================================================================== */
            if ( opts.lapack ) {

                cpu_time = magma_wtime();
                lapackf77_cgbsv(&N, &KL, &KU, &nrhs, h_A, &lda, ipiv, h_B, &ldb, &info );
                cpu_time = magma_wtime() - cpu_time;
                cpu_perf = gflops / cpu_time;

                if (info != 0) {
                    printf("lapackf77_cgesv returned error %lld: %s.\n",
                            (long long)info, magma_strerror( info ));
                }
                printf( "%5lld %5lld   %7.2f (%7.2f)   %7.2f (%7.2f)   %8.2e   %s\n",
                        (long long) N, (long long) nrhs,
                        cpu_perf, 1000.*cpu_time, gpu_perf, 1000.*gpu_time,
                        error, (okay ? "ok" : "failed"));
            }
            else {
                printf( "%5lld %5lld     ---   (  ---  )   %7.2f (%7.2f)   %8.2e   %s\n",
                        (long long) N, (long long) nrhs,
                        gpu_perf, 1000.*gpu_time,
                        error, (okay ? "ok" : "failed"));
            }

            magma_free_cpu( h_A );
            magma_free_cpu( h_B );
            magma_free_cpu( h_X );
            magma_free_cpu( work );
            magma_free_cpu( ipiv );

            magma_free( d_A );
            magma_free( d_B );
            magma_free( dipiv );

            fflush( stdout );
        }
        if ( opts.niter > 1 ) {
            printf( "\n" );
        }
    }

    opts.cleanup();
    TESTING_CHECK( magma_finalize() );
    return status;
}