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
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
|
/*
-- MAGMA (version 2.9.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
@date January 2025
@generated from testing/testing_zswap.cpp, normal z -> c, Wed Jan 22 14:40:25 2025
@author Mark Gates
*/
// includes, system
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <math.h>
#include <assert.h>
// includes, project
#include "magma_v2.h"
#include "magma_lapack.h"
#include "testings.h"
// if ( A == B ) return 0, else return 1
static magma_int_t diff_matrix(
magma_int_t m, magma_int_t n,
magmaFloatComplex *A, magma_int_t lda,
magmaFloatComplex *B, magma_int_t ldb )
{
for( magma_int_t j = 0; j < n; j++ ) {
for( magma_int_t i = 0; i < m; i++ ) {
if ( ! MAGMA_C_EQUAL( A[lda*j+i], B[ldb*j+i] ) )
return 1;
}
}
return 0;
}
// fill matrix with entries Aij = offset + (i+1) + (j+1)/10000,
// which makes it easy to identify which rows & cols have been swapped.
static void init_matrix(
magma_int_t m, magma_int_t n,
magmaFloatComplex *A, magma_int_t lda, magma_int_t offset )
{
assert( lda >= m );
for( magma_int_t j = 0; j < n; ++j ) {
for( magma_int_t i=0; i < m; ++i ) {
A[i + j*lda] = MAGMA_C_MAKE( offset + (i+1) + (j+1)/10000., 0 );
}
}
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing cswap, cswapblk, claswp, claswpx
*/
int main( int argc, char** argv)
{
TESTING_CHECK( magma_init() );
magma_print_environment();
// OpenCL use: cl_mem , offset (two arguments);
// else use: pointer + offset (one argument).
#ifdef MAGMA_HAVE_OPENCL
#define d_A1(i_, j_) d_A1 , (i_) + (j_)*ldda
#define d_A2(i_, j_) d_A2 , (i_) + (j_)*ldda
#define d_ipiv(i_) d_ipiv , (i_)
#else
#define d_A1(i_, j_) (d_A1 + (i_) + (j_)*ldda)
#define d_A2(i_, j_) (d_A2 + (i_) + (j_)*ldda)
#define d_ipiv(i_) (d_ipiv + (i_))
#endif
#define h_A1(i_, j_) (h_A1 + (i_) + (j_)*lda)
#define h_A2(i_, j_) (h_A2 + (i_) + (j_)*lda)
magmaFloatComplex *h_A1, *h_A2;
magmaFloatComplex *h_R1, *h_R2;
magmaFloatComplex_ptr d_A1, d_A2;
// row-major and column-major performance
real_Double_t row_perf0 = MAGMA_D_NAN, col_perf0 = MAGMA_D_NAN;
real_Double_t row_perf1 = MAGMA_D_NAN, col_perf1 = MAGMA_D_NAN;
real_Double_t row_perf2 = MAGMA_D_NAN, col_perf2 = MAGMA_D_NAN;
real_Double_t row_perf4 = MAGMA_D_NAN;
real_Double_t row_perf5 = MAGMA_D_NAN, col_perf5 = MAGMA_D_NAN;
real_Double_t row_perf6 = MAGMA_D_NAN, col_perf6 = MAGMA_D_NAN;
real_Double_t row_perf7 = MAGMA_D_NAN;
real_Double_t cpu_perf = MAGMA_D_NAN;
real_Double_t time, gbytes;
magma_int_t N, lda, ldda, nb, j;
magma_int_t ione = 1;
magma_int_t *ipiv, *ipiv2;
magmaInt_ptr d_ipiv;
int status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
printf("%% %8s cswap cswap cswapblk claswp claswp2 claswpx ccopymatrix CPU (all in )\n", g_platform_str );
printf("%% N nb row-maj/col-maj row-maj/col-maj row-maj/col-maj row-maj row-maj row-maj/col-maj row-blk/col-blk claswp (GByte/s)\n");
printf("%%========================================================================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
// For an N x N matrix, swap nb rows or nb columns using various methods.
// Each test is assigned one bit in the 'check' bitmask; bit=1 indicates failure.
// The variable 'shift' keeps track of which bit is for current test
magma_int_t shift = 1;
magma_int_t check = 0;
N = opts.nsize[itest];
lda = N;
ldda = magma_roundup( N, opts.align ); // multiple of 32 by default
nb = (opts.nb > 0 ? opts.nb : magma_get_cgetrf_nb( N, N ));
nb = min( N, nb );
// each swap does 2N loads and 2N stores, for nb swaps
gbytes = sizeof(magmaFloatComplex) * 4.*N*nb / 1e9;
TESTING_CHECK( magma_cmalloc_pinned( &h_A1, lda*N ));
TESTING_CHECK( magma_cmalloc_pinned( &h_A2, lda*N ));
TESTING_CHECK( magma_cmalloc_pinned( &h_R1, lda*N ));
TESTING_CHECK( magma_cmalloc_pinned( &h_R2, lda*N ));
TESTING_CHECK( magma_imalloc_cpu( &ipiv, nb ));
TESTING_CHECK( magma_imalloc_cpu( &ipiv2, nb ));
TESTING_CHECK( magma_imalloc( &d_ipiv, nb ));
TESTING_CHECK( magma_cmalloc( &d_A1, ldda*N ));
TESTING_CHECK( magma_cmalloc( &d_A2, ldda*N ));
// getrf always makes ipiv[j] >= j+1, where ipiv is one based and j is zero based
// some implementations (e.g., MacOS dlaswp) assume this
for( j=0; j < nb; j++ ) {
ipiv[j] = (rand() % (N-j)) + j + 1;
assert( ipiv[j] >= j+1 );
assert( ipiv[j] <= N );
}
/* =====================================================================
* cublas / clBLAS / Xeon Phi cswap, row-by-row (2 matrices)
*/
/* Row Major */
init_matrix( N, N, h_A1, lda, 0 );
init_matrix( N, N, h_A2, lda, 100 );
magma_csetmatrix( N, N, h_A1, lda, d_A1(0,0), ldda, opts.queue );
magma_csetmatrix( N, N, h_A2, lda, d_A2(0,0), ldda, opts.queue );
time = magma_sync_wtime( opts.queue );
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
magma_cswap( N, d_A1(0,j), 1, d_A2(0,ipiv[j]-1), 1, opts.queue );
}
}
time = magma_sync_wtime( opts.queue ) - time;
row_perf0 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_cswap( &N, h_A1(0,j), &ione, h_A2(0,ipiv[j]-1), &ione);
}
}
magma_cgetmatrix( N, N, d_A1(0,0), ldda, h_R1, lda, opts.queue );
magma_cgetmatrix( N, N, d_A2(0,0), ldda, h_R2, lda, opts.queue );
check += (diff_matrix( N, N, h_A1, lda, h_R1, lda ) ||
diff_matrix( N, N, h_A2, lda, h_R2, lda ))*shift;
shift *= 2;
/* Column Major */
init_matrix( N, N, h_A1, lda, 0 );
init_matrix( N, N, h_A2, lda, 100 );
magma_csetmatrix( N, N, h_A1, lda, d_A1(0,0), ldda, opts.queue );
magma_csetmatrix( N, N, h_A2, lda, d_A2(0,0), ldda, opts.queue );
time = magma_sync_wtime( opts.queue );
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
magma_cswap( N, d_A1(j,0), ldda, d_A2(ipiv[j]-1,0), ldda, opts.queue );
}
}
time = magma_sync_wtime( opts.queue ) - time;
col_perf0 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_cswap( &N, h_A1+j, &lda, h_A2+(ipiv[j]-1), &lda);
}
}
magma_cgetmatrix( N, N, d_A1(0,0), ldda, h_R1, lda, opts.queue );
magma_cgetmatrix( N, N, d_A2(0,0), ldda, h_R2, lda, opts.queue );
check += (diff_matrix( N, N, h_A1, lda, h_R1, lda ) ||
diff_matrix( N, N, h_A2, lda, h_R2, lda ))*shift;
shift *= 2;
/* =====================================================================
* cswap, row-by-row (2 matrices)
*/
/* Row Major */
init_matrix( N, N, h_A1, lda, 0 );
init_matrix( N, N, h_A2, lda, 100 );
magma_csetmatrix( N, N, h_A1, lda, d_A1(0,0), ldda, opts.queue );
magma_csetmatrix( N, N, h_A2, lda, d_A2(0,0), ldda, opts.queue );
time = magma_sync_wtime( opts.queue );
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
magmablas_cswap( N, d_A1(0,j), 1, d_A2(0,ipiv[j]-1), 1, opts.queue );
}
}
time = magma_sync_wtime( opts.queue ) - time;
row_perf1 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_cswap( &N, h_A1(0,j), &ione, h_A2(0,ipiv[j]-1), &ione);
}
}
magma_cgetmatrix( N, N, d_A1(0,0), ldda, h_R1, lda, opts.queue );
magma_cgetmatrix( N, N, d_A2(0,0), ldda, h_R2, lda, opts.queue );
check += (diff_matrix( N, N, h_A1, lda, h_R1, lda ) ||
diff_matrix( N, N, h_A2, lda, h_R2, lda ))*shift;
shift *= 2;
/* Column Major */
init_matrix( N, N, h_A1, lda, 0 );
init_matrix( N, N, h_A2, lda, 100 );
magma_csetmatrix( N, N, h_A1, lda, d_A1(0,0), ldda, opts.queue );
magma_csetmatrix( N, N, h_A2, lda, d_A2(0,0), ldda, opts.queue );
time = magma_sync_wtime( opts.queue );
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
magmablas_cswap( N, d_A1(j,0), ldda, d_A2(ipiv[j]-1,0), ldda, opts.queue );
}
}
time = magma_sync_wtime( opts.queue ) - time;
col_perf1 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_cswap( &N, h_A1+j, &lda, h_A2+(ipiv[j]-1), &lda);
}
}
magma_cgetmatrix( N, N, d_A1(0,0), ldda, h_R1, lda, opts.queue );
magma_cgetmatrix( N, N, d_A2(0,0), ldda, h_R2, lda, opts.queue );
check += (diff_matrix( N, N, h_A1, lda, h_R1, lda ) ||
diff_matrix( N, N, h_A2, lda, h_R2, lda ))*shift;
shift *= 2;
/* =====================================================================
* cswapblk, blocked version (2 matrices)
*/
#if defined(MAGMA_HAVE_CUDA) || defined(MAGMA_HAVE_HIP)
/* Row Major */
init_matrix( N, N, h_A1, lda, 0 );
init_matrix( N, N, h_A2, lda, 100 );
magma_csetmatrix( N, N, h_A1, lda, d_A1(0,0), ldda, opts.queue );
magma_csetmatrix( N, N, h_A2, lda, d_A2(0,0), ldda, opts.queue );
time = magma_sync_wtime( opts.queue );
magmablas_cswapblk( MagmaRowMajor, N, d_A1(0,0), ldda, d_A2(0,0), ldda, 1, nb, ipiv, 1, 0, opts.queue );
time = magma_sync_wtime( opts.queue ) - time;
row_perf2 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_cswap( &N, h_A1(0,j), &ione, h_A2(0,ipiv[j]-1), &ione );
}
}
magma_cgetmatrix( N, N, d_A1(0,0), ldda, h_R1, lda, opts.queue );
magma_cgetmatrix( N, N, d_A2(0,0), ldda, h_R2, lda, opts.queue );
check += (diff_matrix( N, N, h_A1, lda, h_R1, lda ) ||
diff_matrix( N, N, h_A2, lda, h_R2, lda ))*shift;
shift *= 2;
/* Column Major */
init_matrix( N, N, h_A1, lda, 0 );
init_matrix( N, N, h_A2, lda, 100 );
magma_csetmatrix( N, N, h_A1, lda, d_A1(0,0), ldda, opts.queue );
magma_csetmatrix( N, N, h_A2, lda, d_A2(0,0), ldda, opts.queue );
time = magma_sync_wtime( opts.queue );
magmablas_cswapblk( MagmaColMajor, N, d_A1(0,0), ldda, d_A2(0,0), ldda, 1, nb, ipiv, 1, 0, opts.queue );
time = magma_sync_wtime( opts.queue ) - time;
col_perf2 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_cswap( &N, h_A1(j,0), &lda, h_A2(ipiv[j]-1,0), &lda );
}
}
magma_cgetmatrix( N, N, d_A1(0,0), ldda, h_R1, lda, opts.queue );
magma_cgetmatrix( N, N, d_A2(0,0), ldda, h_R2, lda, opts.queue );
check += (diff_matrix( N, N, h_A1, lda, h_R1, lda ) ||
diff_matrix( N, N, h_A2, lda, h_R2, lda ))*shift;
shift *= 2;
#endif
/* =====================================================================
* LAPACK-style claswp (1 matrix)
*/
/* Row Major */
init_matrix( N, N, h_A1, lda, 0 );
magma_csetmatrix( N, N, h_A1, lda, d_A1(0,0), ldda, opts.queue );
time = magma_sync_wtime( opts.queue );
magmablas_claswp( N, d_A1(0,0), ldda, 1, nb, ipiv, 1, opts.queue );
time = magma_sync_wtime( opts.queue ) - time;
row_perf4 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_cswap( &N, h_A1(0,j), &ione, h_A1(0,ipiv[j]-1), &ione );
}
}
magma_cgetmatrix( N, N, d_A1(0,0), ldda, h_R1, lda, opts.queue );
check += diff_matrix( N, N, h_A1, lda, h_R1, lda )*shift;
shift *= 2;
/* =====================================================================
* LAPACK-style claswp (1 matrix) - d_ipiv on GPU
*/
/* Row Major */
init_matrix( N, N, h_A1, lda, 0 );
magma_csetmatrix( N, N, h_A1, lda, d_A1(0,0), ldda, opts.queue );
time = magma_sync_wtime( opts.queue );
magma_isetvector( nb, ipiv, 1, d_ipiv(0), 1, opts.queue );
magmablas_claswp2( N, d_A1(0,0), ldda, 1, nb, d_ipiv(0), 1, opts.queue );
time = magma_sync_wtime( opts.queue ) - time;
row_perf7 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_cswap( &N, h_A1(0,j), &ione, h_A1(0,ipiv[j]-1), &ione );
}
}
magma_cgetmatrix( N, N, d_A1(0,0), ldda, h_R1, lda, opts.queue );
check += diff_matrix( N, N, h_A1, lda, h_R1, lda )*shift;
shift *= 2;
/* =====================================================================
* LAPACK-style claswpx (extended for row- and col-major) (1 matrix)
*/
/* Row Major */
init_matrix( N, N, h_A1, lda, 0 );
magma_csetmatrix( N, N, h_A1, lda, d_A1(0,0), ldda, opts.queue );
time = magma_sync_wtime( opts.queue );
magmablas_claswpx( N, d_A1(0,0), ldda, 1, 1, nb, ipiv, 1, opts.queue );
time = magma_sync_wtime( opts.queue ) - time;
row_perf5 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_cswap( &N, h_A1(0,j), &ione, h_A1(0,ipiv[j]-1), &ione );
}
}
magma_cgetmatrix( N, N, d_A1(0,0), ldda, h_R1, lda, opts.queue );
check += diff_matrix( N, N, h_A1, lda, h_R1, lda )*shift;
shift *= 2;
/* Col Major */
init_matrix( N, N, h_A1, lda, 0 );
magma_csetmatrix( N, N, h_A1, lda, d_A1(0,0), ldda, opts.queue );
time = magma_sync_wtime( opts.queue );
magmablas_claswpx( N, d_A1(0,0), 1, ldda, 1, nb, ipiv, 1, opts.queue );
time = magma_sync_wtime( opts.queue ) - time;
col_perf5 = gbytes / time;
/* LAPACK swap on CPU for comparison */
time = magma_wtime();
lapackf77_claswp( &N, h_A1, &lda, &ione, &nb, ipiv, &ione );
time = magma_wtime() - time;
cpu_perf = gbytes / time;
magma_cgetmatrix( N, N, d_A1(0,0), ldda, h_R1, lda, opts.queue );
check += diff_matrix( N, N, h_A1, lda, h_R1, lda )*shift;
shift *= 2;
/* =====================================================================
* Copy matrix.
*/
time = magma_sync_wtime( opts.queue );
magma_ccopymatrix( N, nb, d_A1(0,0), ldda, d_A2(0,0), ldda, opts.queue );
time = magma_sync_wtime( opts.queue ) - time;
// copy reads 1 matrix and writes 1 matrix, so has half gbytes of swap
col_perf6 = 0.5 * gbytes / time;
time = magma_sync_wtime( opts.queue );
magma_ccopymatrix( nb, N, d_A1(0,0), ldda, d_A2(0,0), ldda, opts.queue );
time = magma_sync_wtime( opts.queue ) - time;
// copy reads 1 matrix and writes 1 matrix, so has half gbytes of swap
row_perf6 = 0.5 * gbytes / time;
printf("%5lld %3lld %6.2f%c/ %6.2f%c %6.2f%c/ %6.2f%c %6.2f%c/ %6.2f%c %6.2f%c %6.2f%c %6.2f%c/ %6.2f%c %6.2f / %6.2f %6.2f %10s\n",
(long long) N, (long long) nb,
row_perf0, ((check & 0x001) != 0 ? '*' : ' '),
col_perf0, ((check & 0x002) != 0 ? '*' : ' '),
row_perf1, ((check & 0x004) != 0 ? '*' : ' '),
col_perf1, ((check & 0x008) != 0 ? '*' : ' '),
row_perf2, ((check & 0x010) != 0 ? '*' : ' '),
col_perf2, ((check & 0x020) != 0 ? '*' : ' '),
row_perf4, ((check & 0x040) != 0 ? '*' : ' '),
row_perf7, ((check & 0x080) != 0 ? '*' : ' '),
row_perf5, ((check & 0x100) != 0 ? '*' : ' '),
col_perf5, ((check & 0x200) != 0 ? '*' : ' '),
row_perf6,
col_perf6,
cpu_perf,
(check == 0 ? "ok" : "* failed") );
status += ! (check == 0);
magma_free_pinned( h_A1 );
magma_free_pinned( h_A2 );
magma_free_pinned( h_R1 );
magma_free_pinned( h_R2 );
magma_free_cpu( ipiv );
magma_free_cpu( ipiv2 );
magma_free( d_ipiv );
magma_free( d_A1 );
magma_free( d_A2 );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_CHECK( magma_finalize() );
return status;
}
|
