/* ////////////////////////////////////////////////////////////////////////////
-- Testing csymmetrize
Code is very similar to testing_ctranspose.cpp
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gbytes, gpu_perf, gpu_time, cpu_perf, cpu_time;
float error, work[1];
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magmaFloatComplex *h_A, *h_R;
magmaFloatComplex *d_A;
magma_int_t N, size, lda, ldda;
magma_int_t ione = 1;
magma_opts opts;
parse_opts( argc, argv, &opts );
printf(" N CPU GByte/s (sec) GPU GByte/s (sec) check\n");
printf("=====================================================\n");
for( int i = 0; i < opts.ntest; ++i ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[i];
lda = N;
ldda = ((N+31)/32)*32;
size = lda*N;
// load strictly lower triangle, save strictly upper triangle
gbytes = sizeof(magmaFloatComplex) * 1.*N*(N-1) / 1e9;
TESTING_MALLOC( h_A, magmaFloatComplex, size );
TESTING_MALLOC( h_R, magmaFloatComplex, size );
TESTING_DEVALLOC( d_A, magmaFloatComplex, ldda*N );
/* Initialize the matrix */
for( int j = 0; j < N; ++j ) {
for( int i = 0; i < N; ++i ) {
h_A[i + j*lda] = MAGMA_C_MAKE( i + j/10000., j );
}
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_csetmatrix( N, N, h_A, lda, d_A, ldda );
gpu_time = magma_sync_wtime( 0 );
//magmablas_csymmetrize( opts.uplo, N-2, d_A+1+ldda, ldda ); // inset by 1 row & col
magmablas_csymmetrize( opts.uplo, N, d_A, ldda );
gpu_time = magma_sync_wtime( 0 ) - gpu_time;
gpu_perf = gbytes / gpu_time;
/* =====================================================================
Performs operation using naive in-place algorithm
(LAPACK doesn't implement symmetrize)
=================================================================== */
cpu_time = magma_wtime();
//for( int j = 1; j < N-1; ++j ) { // inset by 1 row & col
// for( int i = 1; i < j; ++i ) {
for( int j = 0; j < N; ++j ) {
for( int i = 0; i < j; ++i ) {
if ( opts.uplo == MagmaLower ) {
h_A[i + j*lda] = MAGMA_C_CNJG( h_A[j + i*lda] );
}
else {
h_A[j + i*lda] = MAGMA_C_CNJG( h_A[i + j*lda] );
}
}
}
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gbytes / cpu_time;
/* =====================================================================
Check the result
=================================================================== */
magma_cgetmatrix( N, N, d_A, ldda, h_R, lda );
blasf77_caxpy(&size, &c_neg_one, h_A, &ione, h_R, &ione);
error = lapackf77_clange("f", &N, &N, h_R, &lda, work);
printf("%5d %7.2f (%7.2f) %7.2f (%7.2f) %s\n",
(int) N, cpu_perf, cpu_time, gpu_perf, gpu_time,
(error == 0. ? "ok" : "failed") );
TESTING_FREE( h_A );
TESTING_FREE( h_R );
TESTING_DEVFREE( d_A );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return 0;
}
int main(int argc, char *argv[])
{
size_t memsize = 0; /* at first in MB, limited to 4Gb on 32 bits */
int pagesize;
int i;
int lc;
char *msg;
char *p, *bigmalloc;
int loop_count; /* limited to 16Go on 32 bits systems */
pagesize = sysconf(_SC_PAGESIZE);
if ((msg = parse_opts(argc, argv, options, help)) != (char *)NULL)
tst_brkm(TBROK, cleanup, "OPTION PARSING ERROR - %s", msg);
if (m_opt) {
memsize = (size_t) atoi(m_copt) * 1024 * 1024;
if (memsize < 1) {
tst_brkm(TBROK, cleanup, "Invalid arg for -m: %s",
m_copt);
}
}
if (r_opt)
srand(time(NULL));
setup();
for (lc = 0; TEST_LOOPING(lc); lc++) {
Tst_count = 0;
if (!m_opt) {
/* find out by ourselves! */
memsize = get_memsize();
if (memsize < 1) {
tst_brkm(TBROK, cleanup,
"Unable to guess maxmemsize from /proc/meminfo");
}
}
/* Allocate (virtual) memory */
bigmalloc = p = malloc(memsize);
if (!p) {
tst_resm(TFAIL, "malloc - alloc of %zuMB failed",
memsize / 1024 / 1024);
cleanup();
}
/*
* Dirty all the pages, to force physical RAM allocation
* and exercise eventually the swapper
*/
tst_resm(TINFO, "touching %zuMB of malloc'ed memory (%s)",
memsize / 1024 / 1024, r_opt ? "random" : "linear");
loop_count = memsize / pagesize;
for (i = 0; i < loop_count; i++) {
if (v_opt
&& (i % (PROGRESS_LEAP * 1024 / pagesize) == 0)) {
printf(".");
fflush(stdout);
}
/*
* Make the page dirty,
* and make sure compiler won't optimize it away
* Touching more than one word per page is useless
* because of cache.
*/
*(int *)p = 0xdeadbeef ^ i;
if (r_opt) {
p = bigmalloc +
(size_t) ((double)(memsize - sizeof(int)) *
rand() / (RAND_MAX + 1.0));
} else {
p += pagesize;
}
}
if (v_opt)
printf("\n");
/* This is not mandatory (except in a loop), but it exercise mm again */
free(bigmalloc);
/*
* seems that if the malloc'ed area was bad, we'd get SEGV (or kicked
* somehow by the OOM killer?), hence we can indicate a PASS.
*/
tst_resm(TPASS, "malloc - alloc of %zuMB succeeded",
memsize / 1024 / 1024);
}
cleanup();
return 0;
}
int main(int ac, char **av)
{
int lc;
char *msg;
void check_functionality(void);
int r_val;
if ((msg = parse_opts(ac, av, NULL, NULL)) != NULL)
tst_brkm(TBROK, NULL, "OPTION PARSING ERROR - %s", msg);
setup(); /* global setup */
for (lc = 0; TEST_LOOPING(lc); lc++) {
Tst_count = 0;
/* get the name of the test dirctory */
if ((temp_dir = (getcwd(temp_dir, 0))) == NULL)
tst_brkm(TBROK, cleanup, "getcwd failed");
/*
* create a new directory and open it
*/
if ((r_val = mkdir(TEST_DIR, MODES)) == -1)
tst_brkm(TBROK, cleanup, "mkdir failed");
if ((fd = open(TEST_DIR, O_RDONLY)) == -1)
tst_brkm(TBROK, cleanup, "open of directory failed");
TEST(fchdir(fd));
if (TEST_RETURN == -1)
tst_brkm(TFAIL | TTERRNO, cleanup,
"fchdir call failed");
else {
if (STD_FUNCTIONAL_TEST)
check_functionality();
else
tst_resm(TPASS, "call succeeded");
}
/*
* clean up things in case we are looping
*/
/*
* NOTE: in case of failure here, we need to use "tst_resm()"
* and not "tst_brkm()". This is because if we get to this
* point, we have already set a PASS or FAIL for the test
* and "tst_brkm()" won't report as we might expect.
*/
/* chdir back to our temporary work directory */
if ((r_val = chdir("..")) == -1)
tst_resm(TBROK | TERRNO, "chdir failed");
if ((r_val = rmdir(TEST_DIR)) == -1)
tst_resm(TBROK | TERRNO, "rmdir failed");
free(temp_dir);
temp_dir = NULL;
}
cleanup();
tst_exit();
}
int main(int ac, char **av)
{
int lc;
char *msg;
int ind;
int offset;
/***************************************************************
* parse standard options
***************************************************************/
if ((msg = parse_opts(ac, av, NULL, NULL)) != NULL) {
tst_brkm(TBROK, NULL, "OPTION PARSING ERROR - %s", msg);
}
/***************************************************************
* perform global setup for test
***************************************************************/
setup();
/* set the expected errnos... */
TEST_EXP_ENOS(exp_enos);
/***************************************************************
* check looping state if -c option given
***************************************************************/
for (lc = 0; TEST_LOOPING(lc); lc++) {
tst_count = 0;
offset = (lc % 100) * 4096; /* max size is 100 blocks */
for (ind = 0; Whence[ind] >= 0; ind++) {
/*
* Call lseek(2)
*/
TEST(lseek(Fd, (long)offset, Whence[ind]));
/* check return code */
if (TEST_RETURN == -1) {
TEST_ERROR_LOG(TEST_ERRNO);
tst_resm(TFAIL,
"lseek(%s, %d, 0) Failed, errno=%d : %s",
Fname, offset, TEST_ERRNO,
strerror(TEST_ERRNO));
} else {
/***************************************************************
* only perform functional verification if flag set (-f not given)
***************************************************************/
if (STD_FUNCTIONAL_TEST) {
/* No Verification test, yet... */
tst_resm(TPASS,
"lseek(%s, %d, %d) returned %ld",
Fname, offset, Whence[ind],
TEST_RETURN);
} else
tst_count++;
}
}
}
/***************************************************************
* cleanup and exit
***************************************************************/
cleanup();
tst_exit();
}
// ----------------------------------------
int main( int argc, char** argv )
{
TESTING_INIT();
//real_Double_t t_m, t_c, t_f;
magma_int_t ione = 1;
magmaDoubleComplex *A, *B;
double diff, error;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t m, n, k, size, maxn, ld;
magmaDoubleComplex x2_m, x2_c; // complex x for magma, cblas/fortran blas respectively
double x_m, x_c; // x for magma, cblas/fortran blas respectively
magma_opts opts;
parse_opts( argc, argv, &opts );
opts.tolerance = max( 100., opts.tolerance );
double tol = opts.tolerance * lapackf77_dlamch("E");
gTol = tol;
printf( "!! Calling these CBLAS and Fortran BLAS sometimes crashes (segfault), which !!\n"
"!! is why we use wrappers. It does not necesarily indicate a bug in MAGMA. !!\n"
"\n"
"Diff compares MAGMA wrapper to CBLAS and BLAS function; should be exactly 0.\n"
"Error compares MAGMA implementation to CBLAS and BLAS function; should be ~ machine epsilon.\n"
"\n" );
double total_diff = 0.;
double total_error = 0.;
int inc[] = { 1 }; //{ -2, -1, 1, 2 }; //{ 1 }; //{ -1, 1 };
int ninc = sizeof(inc)/sizeof(*inc);
for( int itest = 0; itest < opts.ntest; ++itest ) {
m = opts.msize[itest];
n = opts.nsize[itest];
k = opts.ksize[itest];
for( int iincx = 0; iincx < ninc; ++iincx ) {
magma_int_t incx = inc[iincx];
for( int iincy = 0; iincy < ninc; ++iincy ) {
magma_int_t incy = inc[iincy];
printf("=========================================================================\n");
printf( "m=%d, n=%d, k=%d, incx = %d, incy = %d\n",
(int) m, (int) n, (int) k, (int) incx, (int) incy );
printf( "Function MAGMA CBLAS BLAS Diff Error\n"
" msec msec msec\n" );
// allocate matrices
// over-allocate so they can be any combination of
// {m,n,k} * {abs(incx), abs(incy)} by
// {m,n,k} * {abs(incx), abs(incy)}
maxn = max( max( m, n ), k ) * max( abs(incx), abs(incy) );
ld = max( 1, maxn );
size = ld*maxn;
magma_zmalloc_pinned( &A, size ); assert( A != NULL );
magma_zmalloc_pinned( &B, size ); assert( B != NULL );
// initialize matrices
lapackf77_zlarnv( &ione, ISEED, &size, A );
lapackf77_zlarnv( &ione, ISEED, &size, B );
printf( "Level 1 BLAS ----------------------------------------------------------\n" );
// ----- test DZASUM
// get one-norm of column j of A
if ( incx > 0 && incx == incy ) { // positive, no incy
diff = 0;
error = 0;
for( int j = 0; j < k; ++j ) {
x_m = magma_cblas_dzasum( m, A(0,j), incx );
x_c = cblas_dzasum( m, A(0,j), incx );
diff += fabs( x_m - x_c );
x_c = blasf77_dzasum( &m, A(0,j), &incx );
error += fabs( (x_m - x_c) / (m*x_c) );
}
output( "dzasum", diff, error );
total_diff += diff;
total_error += error;
}
// ----- test DZNRM2
// get two-norm of column j of A
if ( incx > 0 && incx == incy ) { // positive, no incy
diff = 0;
error = 0;
for( int j = 0; j < k; ++j ) {
x_m = magma_cblas_dznrm2( m, A(0,j), incx );
x_c = cblas_dznrm2( m, A(0,j), incx );
diff += fabs( x_m - x_c );
x_c = blasf77_dznrm2( &m, A(0,j), &incx );
error += fabs( (x_m - x_c) / (m*x_c) );
}
output( "dznrm2", diff, error );
total_diff += diff;
total_error += error;
}
// ----- test ZDOTC
// dot columns, Aj^H Bj
diff = 0;
error = 0;
for( int j = 0; j < k; ++j ) {
// MAGMA implementation, not just wrapper
x2_m = magma_cblas_zdotc( m, A(0,j), incx, B(0,j), incy );
// crashes on MKL 11.1.2, ILP64
#if ! defined( MAGMA_WITH_MKL )
#ifdef COMPLEX
cblas_zdotc_sub( m, A(0,j), incx, B(0,j), incy, &x2_c );
#else
x2_c = cblas_zdotc( m, A(0,j), incx, B(0,j), incy );
#endif
error += fabs( x2_m - x2_c ) / fabs( m*x2_c );
#endif
// crashes on MacOS 10.9
#if ! defined( __APPLE__ )
x2_c = blasf77_zdotc( &m, A(0,j), &incx, B(0,j), &incy );
error += fabs( x2_m - x2_c ) / fabs( m*x2_c );
#endif
}
output( "zdotc", diff, error );
total_diff += diff;
total_error += error;
total_error += error;
// ----- test ZDOTU
// dot columns, Aj^T * Bj
diff = 0;
error = 0;
for( int j = 0; j < k; ++j ) {
// MAGMA implementation, not just wrapper
x2_m = magma_cblas_zdotu( m, A(0,j), incx, B(0,j), incy );
// crashes on MKL 11.1.2, ILP64
#if ! defined( MAGMA_WITH_MKL )
#ifdef COMPLEX
cblas_zdotu_sub( m, A(0,j), incx, B(0,j), incy, &x2_c );
#else
x2_c = cblas_zdotu( m, A(0,j), incx, B(0,j), incy );
#endif
error += fabs( x2_m - x2_c ) / fabs( m*x2_c );
#endif
// crashes on MacOS 10.9
#if ! defined( __APPLE__ )
x2_c = blasf77_zdotu( &m, A(0,j), &incx, B(0,j), &incy );
error += fabs( x2_m - x2_c ) / fabs( m*x2_c );
#endif
}
output( "zdotu", diff, error );
total_diff += diff;
total_error += error;
// tell user about disabled functions
#if defined( MAGMA_WITH_MKL )
printf( "cblas_zdotc and cblas_zdotu disabled with MKL (segfaults)\n" );
#endif
#if defined( __APPLE__ )
printf( "blasf77_zdotc and blasf77_zdotu disabled on MacOS (segfaults)\n" );
#endif
// cleanup
magma_free_pinned( A );
magma_free_pinned( B );
fflush( stdout );
}}} // itest, incx, incy
// TODO use average error?
printf( "sum diffs = %8.2g, MAGMA wrapper compared to CBLAS and Fortran BLAS; should be exactly 0.\n"
"sum errors = %8.2e, MAGMA implementation compared to CBLAS and Fortran BLAS; should be ~ machine epsilon.\n\n",
total_diff, total_error );
if ( total_diff != 0. ) {
printf( "some tests failed diff == 0.; see above.\n" );
}
else {
printf( "all tests passed diff == 0.\n" );
}
TESTING_FINALIZE();
int status = (total_diff != 0.);
return status;
}
static void report_to_kerneloops(
const char *dump_dir_name,
map_string_t *settings)
{
problem_data_t *problem_data = create_problem_data_for_reporting(dump_dir_name);
if (!problem_data)
xfunc_die(); /* create_problem_data_for_reporting already emitted error msg */
const char *backtrace = problem_data_get_content_or_NULL(problem_data, FILENAME_BACKTRACE);
if (!backtrace)
error_msg_and_die("Error sending kernel oops due to missing backtrace");
const char *env = getenv("KerneloopsReporter_SubmitURL");
const char *submitURL = (env ? env : get_map_string_item_or_empty(settings, "SubmitURL"));
if (!submitURL[0])
submitURL = "http://oops.kernel.org/submitoops.php";
log(_("Submitting oops report to %s"), submitURL);
CURLcode ret = http_post_to_kerneloops_site(submitURL, backtrace);
if (ret != CURLE_OK)
error_msg_and_die("Kernel oops has not been sent due to %s", curl_easy_strerror(ret));
problem_data_free(problem_data);
/* Server replies with:
* 200 thank you for submitting the kernel oops information
* RemoteIP: 34192fd15e34bf60fac6a5f01bba04ddbd3f0558
* - no URL or bug ID apparently...
*/
struct dump_dir *dd = dd_opendir(dump_dir_name, /*flags:*/ 0);
if (dd)
{
report_result_t rr = { .label = (char *)"kerneloops" };
rr.url = (char *)submitURL;
add_reported_to_entry(dd, &rr);
dd_close(dd);
}
log("Kernel oops report was uploaded");
}
int main(int argc, char **argv)
{
abrt_init(argv);
/* I18n */
setlocale(LC_ALL, "");
#if ENABLE_NLS
bindtextdomain(PACKAGE, LOCALEDIR);
textdomain(PACKAGE);
#endif
map_string_t *settings = new_map_string();
const char *dump_dir_name = ".";
GList *conf_file = NULL;
/* Can't keep these strings/structs static: _() doesn't support that */
const char *program_usage_string = _(
"& [-v] [-c CONFFILE]... -d DIR\n"
"\n"
"Reports kernel oops to kerneloops.org (or similar) site.\n"
"\n"
"Files with names listed in $EXCLUDE_FROM_REPORT are not included\n"
"into the tarball.\n"
"\n"
"CONFFILE lines should have 'PARAM = VALUE' format.\n"
"Recognized string parameter: SubmitURL.\n"
"Parameter can be overridden via $KerneloopsReporter_SubmitURL."
);
enum {
OPT_v = 1 << 0,
OPT_d = 1 << 1,
OPT_c = 1 << 2,
};
/* Keep enum above and order of options below in sync! */
struct options program_options[] = {
OPT__VERBOSE(&g_verbose),
OPT_STRING('d', NULL, &dump_dir_name, "DIR" , _("Problem directory")),
OPT_LIST( 'c', NULL, &conf_file , "FILE", _("Configuration file")),
OPT_END()
};
/*unsigned opts =*/ parse_opts(argc, argv, program_options, program_usage_string);
export_abrt_envvars(0);
while (conf_file)
{
char *fn = (char *)conf_file->data;
log_notice("Loading settings from '%s'", fn);
load_conf_file(fn, settings, /*skip key w/o values:*/ false);
log_debug("Loaded '%s'", fn);
conf_file = g_list_remove(conf_file, fn);
}
report_to_kerneloops(dump_dir_name, settings);
free_map_string(settings);
return 0;
}
/*ARGSUSED*/
int
main(int argc, char **argv)
{
register int i, j, ok, pid;
int count, child, status, nwait;
#ifdef UCLINUX
char *msg;
if ((msg = parse_opts(argc, argv, (option_t *)NULL, NULL)) != (char *)NULL){
tst_brkm(TBROK, cleanup, "OPTION PARSING ERROR - %s", msg);
}
argv0 = argv[0];
maybe_run_child(&do_child, "dS", &id_uclinux, &maxsemstring);
#endif
prog = argv[0];
nwait = 0;
setup();
/*--------------------------------------------------------------*/
srand(getpid());
tid = -1;
for (i = 0; i < NPROCS; i++) {
do {
keyarray[i] = (key_t)rand();
if (keyarray[i] == IPC_PRIVATE) {
ok = 0;
continue;
}
ok = 1;
for (j = 0; j < i; j++) {
if (keyarray[j] == keyarray[i]) {
ok = 0;
break;
}
}
} while (ok == 0);
}
if ((signal(SIGTERM, term)) == SIG_ERR) {
tst_resm(TFAIL, "\tsignal failed. errno = %d", errno);
tst_exit();
}
for (i = 0; i < NPROCS; i++) {
if ((pid = FORK_OR_VFORK()) < 0) {
tst_resm(TFAIL, "\tFork failed (may be OK if under stress)");
tst_exit();
}
if (pid == 0) {
procstat = 1;
dotest(keyarray[i]);
exit(0);
}
pidarray[i] = pid;
nwait++;
}
/*
* Wait for children to finish.
*/
count = 0;
while((child = wait(&status)) > 0) {
if (status) {
tst_resm(TFAIL, "%s[%d] Test failed. exit=0x%x", prog, child, status);
local_flag = FAILED;
}
++count;
}
/*
* Should have collected all children.
*/
if (count != nwait) {
tst_resm(TFAIL, "\tWrong # children waited on, count = %d", count);
local_flag = FAILED;
}
if (local_flag != FAILED)
tst_resm(TPASS, "semctl06 ran successfully!");
else tst_resm(TFAIL, "semctl06 failed");
/*--------------------------------------------------------------*/
/* Clean up any files created by test before call to anyfail. */
cleanup ();
return (0); /* shut lint up */
}
int main(int ac, char **av)
{
int lc; /* loop counter */
char *msg; /* message returned from parse_opts */
int incr; /* increment */
long nbrkpt; /* new brk point value */
long cur_brk_val; /* current size returned by sbrk */
long aft_brk_val; /* current size returned by sbrk */
if ((msg = parse_opts(ac, av, NULL, NULL)) != NULL)
tst_brkm(TBROK, NULL, "OPTION PARSING ERROR - %s", msg);
setup();
/*
* Attempt to control how fast we get to test max size.
* Every MAX_SIZE_LC'th lc will be fastest test will reach max size.
*/
incr = (Max_brk_byte_size - Beg_brk_val) / (MAX_SIZE_LC / 2);
if ((incr * 2) < 4096) /* make sure that process will grow */
incr += 4096 / 2;
for (lc = 0; TEST_LOOPING(lc); lc++) {
Tst_count = 0;
/*
* Determine new value to give brk
* Every even lc value, grow by 2 incr and
* every odd lc value, strink by one incr.
* If lc is equal to 3, no change, special case.
*/
cur_brk_val = (long)sbrk(0);
if (lc == 3) {
nbrkpt = cur_brk_val; /* no change, special one time case */
} else if ((lc % 2) == 0) {
/*
* grow
*/
nbrkpt = cur_brk_val + (2 * incr);
if (nbrkpt > Max_brk_byte_size)
nbrkpt = Beg_brk_val; /* start over */
} else {
/*
* shrink
*/
nbrkpt = cur_brk_val - incr;
}
/****
printf("cur_brk_val = %d, nbrkpt = %d, incr = %d, lc = %d\n",
cur_brk_val, nbrkpt, incr, lc);
****/
/*
* Call brk(2)
*/
TEST(brk((char *)nbrkpt));
/* check return code */
if (TEST_RETURN == -1) {
aft_brk_val = (long)sbrk(0);
tst_resm(TFAIL|TTERRNO,
"brk(%ld) failed (size before %ld, after %ld)",
nbrkpt, cur_brk_val, aft_brk_val);
} else {
if (STD_FUNCTIONAL_TEST) {
aft_brk_val = (long)sbrk(0);
if (aft_brk_val == nbrkpt) {
tst_resm(TPASS,
"brk(%ld) returned %ld, new size verified by sbrk",
nbrkpt, TEST_RETURN);
} else {
tst_resm(TFAIL,
"brk(%ld) returned %ld, sbrk before %ld, after %ld",
nbrkpt, TEST_RETURN,
cur_brk_val, aft_brk_val);
}
}
}
}
cleanup();
tst_exit();
}
int main(int ac, char **av)
{
int lc;
char *msg;
int results;
/* Disable test if the version of the kernel is less than 2.6.17 */
if (((results = tst_kvercmp(2, 6, 17)) < 0)) {
tst_resm(TWARN, "This test can only run on kernels that are ");
tst_resm(TWARN, "2.6.17 and higher");
exit(0);
}
/*
* parse standard options
*/
if ((msg = parse_opts(ac, av, NULL, NULL)) != NULL)
tst_brkm(TBROK, NULL, "OPTION PARSING ERROR - %s", msg);
/*
* perform global setup for test
*/
setup();
/*
* check if the current filesystem is nfs
*/
if (tst_is_cwd_nfs()) {
tst_brkm(TCONF, cleanup,
"Cannot do tee on a file located on an NFS filesystem");
}
/*
* check looping state if -c option given
*/
for (lc = 0; TEST_LOOPING(lc); lc++) {
Tst_count = 0;
/*
* Call tee_test
*/
TEST(tee_test());
/* check return code */
if (TEST_RETURN < 0) {
if (TEST_RETURN != -1) {
TEST_ERRNO = -TEST_RETURN;
}
TEST_ERROR_LOG(TEST_ERRNO);
tst_resm(TFAIL, "tee() Failed, errno=%d : %s",
TEST_ERRNO, strerror(TEST_ERRNO));
} else {
/*
* only perform functional verification if flag set (-f not given)
*/
if (STD_FUNCTIONAL_TEST) {
/* No Verification test, yet... */
tst_resm(TPASS, "tee() returned %ld",
TEST_RETURN);
}
}
}
/*
* cleanup and exit
*/
cleanup();
return (0);
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zgesv
*/
int main(int argc, char **argv)
{
TESTING_INIT();
real_Double_t gflops, cpu_perf, cpu_time, gpu_perf, gpu_time;
double error, Rnorm, Anorm, Xnorm, *work;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex *h_A, *h_LU, *h_B, *h_X;
magma_int_t *ipiv;
magma_int_t N, nrhs, lda, ldb, info, sizeA, sizeB;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
double tol = opts.tolerance * lapackf77_dlamch("E");
nrhs = opts.nrhs;
printf("ngpu %d\n", (int) opts.ngpu );
printf(" N NRHS CPU Gflop/s (sec) GPU GFlop/s (sec) ||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];
lda = N;
ldb = lda;
gflops = ( FLOPS_ZGETRF( N, N ) + FLOPS_ZGETRS( N, nrhs ) ) / 1e9;
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, lda*N );
TESTING_MALLOC_CPU( h_LU, magmaDoubleComplex, lda*N );
TESTING_MALLOC_CPU( h_B, magmaDoubleComplex, ldb*nrhs );
TESTING_MALLOC_CPU( h_X, magmaDoubleComplex, ldb*nrhs );
TESTING_MALLOC_CPU( work, double, N );
TESTING_MALLOC_CPU( ipiv, magma_int_t, N );
/* Initialize the matrices */
sizeA = lda*N;
sizeB = ldb*nrhs;
lapackf77_zlarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_zlarnv( &ione, ISEED, &sizeB, h_B );
// copy A to LU and B to X; save A and B for residual
lapackf77_zlacpy( "F", &N, &N, h_A, &lda, h_LU, &lda );
lapackf77_zlacpy( "F", &N, &nrhs, h_B, &ldb, h_X, &ldb );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_zgesv( N, nrhs, h_LU, lda, ipiv, h_X, ldb, opts.queues2, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_zgesv returned error %d: %s.\n",
(int) info, magma_strerror( info ));
//=====================================================================
// Residual
//=====================================================================
Anorm = lapackf77_zlange("I", &N, &N, h_A, &lda, work);
Xnorm = lapackf77_zlange("I", &N, &nrhs, h_X, &ldb, work);
blasf77_zgemm( MagmaNoTransStr, MagmaNoTransStr, &N, &nrhs, &N,
&c_one, h_A, &lda,
h_X, &ldb,
&c_neg_one, h_B, &ldb);
Rnorm = lapackf77_zlange("I", &N, &nrhs, h_B, &ldb, work);
error = Rnorm/(N*Anorm*Xnorm);
status += ! (error < tol);
/* ====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_zgesv( &N, &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_zgesv returned error %d: %s.\n",
(int) info, magma_strerror( info ));
printf( "%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) N, (int) nrhs, cpu_perf, cpu_time, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed"));
}
else {
printf( "%5d %5d --- ( --- ) %7.2f (%7.2f) %8.2e %s\n",
(int) N, (int) nrhs, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed"));
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_LU );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_X );
TESTING_FREE_CPU( work );
TESTING_FREE_CPU( ipiv );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing cunmqr
*/
int main( int argc, char** argv )
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
float error, work[1];
magmaFloatComplex c_neg_one = MAGMA_C_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;
magmaFloatComplex *C, *R, *A, *W, *tau;
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
// need slightly looser bound (60*eps instead of 30*eps) for some tests
opts.tolerance = max( 60., opts.tolerance );
float tol = opts.tolerance * lapackf77_slamch("E");
// test all combinations of input parameters
magma_side_t side [] = { MagmaLeft, MagmaRight };
magma_trans_t trans[] = { MagmaConjTrans, 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_cgeqrf_nb( m );
ldc = m;
// A is m x k (left) or n x k (right)
mm = (side[iside] == MagmaLeft ? m : n);
lda = mm;
gflops = FLOPS_CUNMQR( m, n, k, side[iside] ) / 1e9;
if ( side[iside] == MagmaLeft && m < k ) {
printf( "%5d %5d %5d %4c %5c skipping because side=left and m < k\n",
(int) m, (int) n, (int) k,
lapacke_side_const( side[iside] ),
lapacke_trans_const( trans[itran] ) );
continue;
}
if ( side[iside] == MagmaRight && n < k ) {
printf( "%5d %5d %5d %4c %5c skipping because side=right and n < k\n",
(int) m, (int) n, (int) 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 );
TESTING_MALLOC_CPU( C, magmaFloatComplex, ldc*n );
TESTING_MALLOC_CPU( R, magmaFloatComplex, ldc*n );
TESTING_MALLOC_CPU( A, magmaFloatComplex, lda*k );
TESTING_MALLOC_CPU( W, magmaFloatComplex, lwork_max );
TESTING_MALLOC_CPU( tau, magmaFloatComplex, k );
// C is full, m x n
size = ldc*n;
lapackf77_clarnv( &ione, ISEED, &size, C );
lapackf77_clacpy( "Full", &m, &n, C, &ldc, R, &ldc );
size = lda*k;
lapackf77_clarnv( &ione, ISEED, &size, A );
// compute QR factorization to get Householder vectors in A, tau
magma_cgeqrf( mm, k, A, lda, tau, W, lwork_max, &info );
if (info != 0)
printf("magma_cgeqrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
lapackf77_cunmqr( 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_cunmqr returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
// query for workspace size
lwork = -1;
magma_cunmqr( side[iside], trans[itran],
m, n, k,
A, lda, tau, R, ldc, W, lwork, &info );
if (info != 0)
printf("magma_cunmqr (lwork query) returned error %d: %s.\n",
(int) info, magma_strerror( info ));
lwork = (magma_int_t) MAGMA_C_REAL( W[0] );
if ( lwork < 0 || lwork > lwork_max ) {
printf("optimal lwork %d > lwork_max %d\n", (int) lwork, (int) lwork_max );
lwork = lwork_max;
}
gpu_time = magma_wtime();
magma_cunmqr( side[iside], trans[itran],
m, n, k,
A, lda, tau, R, ldc, W, lwork, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_cunmqr returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
compute relative error |QC_magma - QC_lapack| / |QC_lapack|
=================================================================== */
error = lapackf77_clange( "Fro", &m, &n, C, &ldc, work );
size = ldc*n;
blasf77_caxpy( &size, &c_neg_one, C, &ione, R, &ione );
error = lapackf77_clange( "Fro", &m, &n, R, &ldc, work ) / error;
printf( "%5d %5d %5d %4c %5c %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) m, (int) n, (int) 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);
TESTING_FREE_CPU( C );
TESTING_FREE_CPU( R );
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( W );
TESTING_FREE_CPU( tau );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}} // end iside, itran
printf( "\n" );
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zhetrd_he2hb
*/
int main( int argc, char** argv)
{
TESTING_INIT_MGPU();
real_Double_t gflops, gpu_time, gpu_perf;
magmaDoubleComplex *h_A, *h_R, *h_work;
magmaDoubleComplex *tau;
double *D, *E;
magma_int_t N, n2, lda, ldda, lwork, ldt, info, nstream;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
// TODO add these options to parse_opts
magma_int_t NE = 0;
magma_int_t distblk = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
magma_int_t WANTZ = (opts.jobz == MagmaVec);
double tol = opts.tolerance * lapackf77_dlamch("E");
if (opts.nb == 0)
opts.nb = 64; //magma_get_zhetrd_he2hb_nb(N);
if (NE < 1)
NE = N; //64; //magma_get_zhetrd_he2hb_nb(N);
nstream = max(3, opts.ngpu+2);
magma_queue_t streams[MagmaMaxGPUs][20];
magmaDoubleComplex *da[MagmaMaxGPUs], *dT1[MagmaMaxGPUs];
if ((distblk == 0) || (distblk < opts.nb))
distblk = max(256, opts.nb);
printf("voici ngpu %d distblk %d NB %d nstream %d\n ",
(int) opts.ngpu, (int) distblk, (int) opts.nb, (int) nstream);
for( magma_int_t dev = 0; dev < opts.ngpu; ++dev ) {
magma_setdevice( dev );
for( int i = 0; i < nstream; ++i ) {
magma_queue_create( &streams[dev][i] );
}
}
magma_setdevice( 0 );
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
ldt = N;
ldda = ((N+31)/32)*32;
n2 = N*lda;
/* We suppose the magma NB is bigger than lapack NB */
lwork = N*opts.nb;
//gflops = ....?
/* Allocate host memory for the matrix */
TESTING_MALLOC_CPU( tau, magmaDoubleComplex, N-1 );
TESTING_MALLOC_PIN( h_A, magmaDoubleComplex, lda*N );
TESTING_MALLOC_PIN( h_R, magmaDoubleComplex, lda*N );
TESTING_MALLOC_PIN( h_work, magmaDoubleComplex, lwork );
TESTING_MALLOC_PIN( D, double, N );
TESTING_MALLOC_PIN( E, double, N );
for( magma_int_t dev = 0; dev < opts.ngpu; ++dev ) {
magma_int_t mlocal = ((N / distblk) / opts.ngpu + 1) * distblk;
magma_setdevice( dev );
TESTING_MALLOC_DEV( da[dev], magmaDoubleComplex, ldda*mlocal );
TESTING_MALLOC_DEV( dT1[dev], magmaDoubleComplex, N*opts.nb );
}
/* ====================================================================
Initialize the matrix
=================================================================== */
lapackf77_zlarnv( &ione, ISEED, &n2, h_A );
magma_zmake_hermitian( N, h_A, lda );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
/* Copy the matrix to the GPU */
magma_zsetmatrix_1D_col_bcyclic( N, N, h_R, lda, da, ldda, opts.ngpu, distblk);
//magmaDoubleComplex *dabis;
//TESTING_MALLOC_DEV( dabis, magmaDoubleComplex, ldda*N );
//magma_zsetmatrix(N, N, h_R, lda, dabis, ldda);
for (int count=0; count < 1; ++count) {
magma_setdevice(0);
gpu_time = magma_wtime();
if (opts.version == 30) {
magma_zhetrd_he2hb_mgpu_spec(
opts.uplo, N, opts.nb, h_R, lda, tau, h_work, lwork,
da, ldda, dT1, opts.nb, opts.ngpu, distblk,
streams, nstream, opts.nthread, &info);
} else {
nstream = 3;
magma_zhetrd_he2hb_mgpu(
opts.uplo, N, opts.nb, h_R, lda, tau, h_work, lwork,
da, ldda, dT1, opts.nb, opts.ngpu, distblk,
streams, nstream, opts.nthread, &info);
}
// magma_zhetrd_he2hb(opts.uplo, N, opts.nb, h_R, lda, tau, h_work, lwork, dT1[0], &info);
gpu_time = magma_wtime() - gpu_time;
printf(" Finish BAND N %d NB %d dist %d ngpu %d version %d timing= %f\n",
N, opts.nb, distblk, opts.ngpu, opts.version, gpu_time);
}
magma_setdevice(0);
for( magma_int_t dev = 0; dev < opts.ngpu; ++dev ) {
magma_setdevice(dev);
cudaDeviceSynchronize();
}
magma_setdevice(0);
magmablasSetKernelStream( NULL );
// todo neither of these is declared in headers
// magma_zhetrd_bhe2trc_v5(opts.nthread, WANTZ, opts.uplo, NE, N, opts.nb, h_R, lda, D, E, dT1[0], ldt);
// magma_zhetrd_bhe2trc(opts.nthread, WANTZ, opts.uplo, NE, N, opts.nb, h_R, lda, D, E, dT1[0], ldt);
// todo where is this timer started?
// gpu_time = magma_wtime() - gpu_time;
// todo what are the gflops?
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_zhetrd_he2hb returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Print performance and error.
=================================================================== */
#if defined(CHECKEIG)
#if defined(PRECISION_z) || defined(PRECISION_d)
if ( opts.check ) {
printf(" Total N %5d flops %6.2f timing %6.2f seconds\n", (int) N, gpu_perf, gpu_time );
char JOBZ;
if (WANTZ == 0)
JOBZ = 'N';
else
JOBZ = 'V';
double nrmI=0.0, nrm1=0.0, nrm2=0.0;
int lwork2 = 256*N;
magmaDoubleComplex *work2, *AINIT;
double *rwork2, *D2;
// TODO free this memory !
magma_zmalloc_cpu( &work2, lwork2 );
magma_dmalloc_cpu( &rwork2, N );
magma_dmalloc_cpu( &D2, N );
magma_zmalloc_cpu( &AINIT, N*lda );
memcpy(AINIT, h_A, N*lda*sizeof(magmaDoubleComplex));
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
int nt = min(12, opts.nthread);
#if defined(USEMKL)
mkl_set_num_threads(nt);
#endif
#if defined(USEACML)
omp_set_num_threads(nt);
#endif
#if defined(PRECISION_z) || defined (PRECISION_c)
lapackf77_zheev( "N", "L", &N, h_A, &lda, D2, work2, &lwork2, rwork2, &info );
#else
lapackf77_dsyev( "N", "L", &N, h_A, &lda, D2, work2, &lwork2, &info );
#endif
///* call eigensolver for our resulting tridiag [D E] and for Q */
//dstedc_withZ('V', N, D, E, h_R, lda);
////dsterf_( &N, D, E, &info);
cpu_time = magma_wtime() - cpu_time;
printf(" Finish CHECK - EIGEN timing= %f threads %d\n", cpu_time, nt);
/* compare result */
cmp_vals(N, D2, D, &nrmI, &nrm1, &nrm2);
magmaDoubleComplex *WORKAJETER;
double *RWORKAJETER, *RESU;
// TODO free this memory !
magma_zmalloc_cpu( &WORKAJETER, (2* N * N + N) );
magma_dmalloc_cpu( &RWORKAJETER, N );
magma_dmalloc_cpu( &RESU, 10 );
int MATYPE;
memset(RESU, 0, 10*sizeof(double));
MATYPE=3;
double NOTHING=0.0;
cpu_time = magma_wtime();
// check results
zcheck_eig_( lapack_vec_const(opts.jobz), &MATYPE, &N, &opts.nb,
AINIT, &lda, &NOTHING, &NOTHING, D2, D,
h_R, &lda, WORKAJETER, RWORKAJETER, RESU );
cpu_time = magma_wtime() - cpu_time;
printf(" Finish CHECK - results timing= %f\n", cpu_time);
#if defined(USEMKL)
mkl_set_num_threads(1);
#endif
#if defined(USEACML)
omp_set_num_threads(1);
#endif
printf("\n");
printf(" ================================================================================================================\n");
printf(" ==> INFO voici threads=%d N=%d NB=%d WANTZ=%d\n", (int) opts.nthread, (int) N, (int) opts.nb, (int) WANTZ);
printf(" ================================================================================================================\n");
printf(" DSBTRD : %15s \n", "STATblgv9withQ ");
printf(" ================================================================================================================\n");
if (WANTZ > 0)
printf(" | A - U S U' | / ( |A| n ulp ) : %15.3E \n", RESU[0]);
if (WANTZ > 0)
printf(" | I - U U' | / ( n ulp ) : %15.3E \n", RESU[1]);
printf(" | D1 - EVEIGS | / (|D| ulp) : %15.3E \n", RESU[2]);
printf(" max | D1 - EVEIGS | : %15.3E \n", RESU[6]);
printf(" ================================================================================================================\n\n\n");
printf(" ****************************************************************************************************************\n");
printf(" * Hello here are the norm Infinite (max)=%8.2e norm one (sum)=%8.2e norm2(sqrt)=%8.2e *\n", nrmI, nrm1, nrm2);
printf(" ****************************************************************************************************************\n\n");
}
#endif // PRECISION_z || PRECISION_d
#endif // CHECKEIG
printf(" Total N %5d flops %6.2f timing %6.2f seconds\n", (int) N, 0.0, gpu_time );
printf("============================================================================\n\n\n");
TESTING_FREE_CPU( tau );
TESTING_FREE_PIN( h_A );
TESTING_FREE_PIN( h_R );
TESTING_FREE_PIN( h_work );
TESTING_FREE_PIN( D );
TESTING_FREE_PIN( E );
for( magma_int_t dev = 0; dev < opts.ngpu; ++dev ) {
magma_setdevice( dev );
TESTING_FREE_DEV( da[dev] );
TESTING_FREE_DEV( dT1[dev] );
}
magma_setdevice( 0 );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
for( magma_int_t dev = 0; dev < opts.ngpu; ++dev ) {
for( int i = 0; i < nstream; ++i ) {
magma_queue_destroy( streams[dev][i] );
}
}
TESTING_FINALIZE_MGPU();
return status;
}
int main(int ac, char **av)
{
int lc; /* loop counter */
char *msg; /* message returned from parse_opts */
int status;
/***************************************************************
* parse standard options
***************************************************************/
if ((msg = parse_opts(ac, av, (option_t *) NULL, NULL)) != (char *)NULL) {
tst_brkm(TBROK, NULL, "OPTION PARSING ERROR - %s", msg);
tst_exit();
}
#ifdef UCLINUX
maybe_run_child(&do_child, "");
#endif
/***************************************************************
* perform global setup for test
***************************************************************/
setup();
/***************************************************************
* check looping state if -c option given
***************************************************************/
for (lc = 0; TEST_LOOPING(lc); lc++) {
/* reset Tst_count in case we are looping. */
Tst_count = 0;
/* make a child process so we can kill it */
/* If we cannot fork => we cannot test kill, so break and exit */
if ((fork_pid = FORK_OR_VFORK()) == -1) {
tst_brkm(TBROK, cleanup,
"fork() Failure. errno=%d : %s",
errno, strerror(errno));
}
if (fork_pid == 0) {
/* CHILD */
#ifdef UCLINUX
if (self_exec(av[0], "") < 0) {
tst_brkm(TBROK, cleanup,
"self_exec of child failed");
}
#else
do_child();
#endif
}
/* PARENT */
/*
* Call kill(2)
*/
TEST(kill(fork_pid, SIGKILL));
/* check return code */
if (TEST_RETURN == -1) {
TEST_ERROR_LOG(TEST_ERRNO);
tst_resm(TFAIL,
"kill(%d, SIGKILL) Failed, errno=%d : %s",
fork_pid, TEST_ERRNO, strerror(TEST_ERRNO));
} else {
/***************************************************************
* only perform functional verification if flag set (-f not given)
***************************************************************/
if (STD_FUNCTIONAL_TEST) {
/* No Verification test, yet... */
tst_resm(TPASS, "kill(%d, SIGKILL) returned %ld",
fork_pid, TEST_RETURN);
}
}
/*
* wait for process to cleanup zombies.
*
*/
waitpid(0, &status, WNOHANG);
} /* End for TEST_LOOPING */
/***************************************************************
* cleanup and exit
***************************************************************/
cleanup();
return 0;
} /* End main */
/* ////////////////////////////////////////////////////////////////////////////
-- Testing cungqr
*/
int main( int argc, char** argv )
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
float Anorm, error, work[1];
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magmaFloatComplex *hA, *hR, *tau, *h_work;
magmaFloatComplex_ptr dA, dT;
magma_int_t m, n, k;
magma_int_t n2, lda, ldda, lwork, min_mn, nb, info;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
float tol = opts.tolerance * lapackf77_slamch("E");
opts.lapack |= opts.check; // check (-c) implies lapack (-l)
printf("Running version %d; available are (specified through --version num):\n",
(int) opts.version);
printf("1 - uses precomputed clarft matrices (default)\n");
printf("2 - recomputes the clarft matrices on the fly\n\n");
printf(" m n k CPU GFlop/s (sec) GPU GFlop/s (sec) ||R|| / ||A||\n");
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.ksize[itest];
if ( m < n || n < k ) {
printf( "%5d %5d %5d skipping because m < n or n < k\n", (int) m, (int) n, (int) k );
continue;
}
lda = m;
ldda = ((m + 31)/32)*32;
n2 = lda*n;
min_mn = min(m, n);
nb = magma_get_cgeqrf_nb( m );
lwork = (m + 2*n+nb)*nb;
gflops = FLOPS_CUNGQR( m, n, k ) / 1e9;
TESTING_MALLOC_PIN( h_work, magmaFloatComplex, lwork );
TESTING_MALLOC_PIN( hR, magmaFloatComplex, lda*n );
TESTING_MALLOC_CPU( hA, magmaFloatComplex, lda*n );
TESTING_MALLOC_CPU( tau, magmaFloatComplex, min_mn );
TESTING_MALLOC_DEV( dA, magmaFloatComplex, ldda*n );
TESTING_MALLOC_DEV( dT, magmaFloatComplex, ( 2*min_mn + ((n + 31)/32)*32 )*nb );
lapackf77_clarnv( &ione, ISEED, &n2, hA );
lapackf77_clacpy( MagmaFullStr, &m, &n, hA, &lda, hR, &lda );
Anorm = lapackf77_clange("f", &m, &n, hA, &lda, work );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
// first, get QR factors in both hA and hR
// okay that magma_cgeqrf_gpu has special structure for R; R isn't used here.
magma_csetmatrix( m, n, hA, lda, dA, ldda );
magma_cgeqrf_gpu( m, n, dA, ldda, tau, dT, &info );
if (info != 0)
printf("magma_cgeqrf_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
magma_cgetmatrix( m, n, dA, ldda, hA, lda );
lapackf77_clacpy( MagmaFullStr, &m, &n, hA, &lda, hR, &lda );
gpu_time = magma_wtime();
if (opts.version == 1)
magma_cungqr( m, n, k, hR, lda, tau, dT, nb, &info );
else
magma_cungqr2(m, n, k, hR, lda, tau, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_cungqr_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_cungqr( &m, &n, &k, hA, &lda, tau, h_work, &lwork, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_cungqr returned error %d: %s.\n",
(int) info, magma_strerror( info ));
// compute relative error |R|/|A| := |Q_magma - Q_lapack|/|A|
blasf77_caxpy( &n2, &c_neg_one, hA, &ione, hR, &ione );
error = lapackf77_clange("f", &m, &n, hR, &lda, work) / Anorm;
bool okay = (error < tol);
status += ! okay;
printf("%5d %5d %5d %7.1f (%7.2f) %7.1f (%7.2f) %8.2e %s\n",
(int) m, (int) n, (int) k,
cpu_perf, cpu_time, gpu_perf, gpu_time,
error, (okay ? "ok" : "failed"));
}
else {
printf("%5d %5d %5d --- ( --- ) %7.1f (%7.2f) --- \n",
(int) m, (int) n, (int) k,
gpu_perf, gpu_time );
}
TESTING_FREE_PIN( h_work );
TESTING_FREE_PIN( hR );
TESTING_FREE_CPU( hA );
TESTING_FREE_CPU( tau );
TESTING_FREE_DEV( dA );
TESTING_FREE_DEV( dT );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing magma_ssymm_mgpu
*/
int main( int argc, char** argv)
{
TESTING_INIT();
float c_neg_one = MAGMA_S_NEG_ONE;
float calpha = MAGMA_S_MAKE( 3.456, 5.678 );
float cbeta = MAGMA_S_MAKE( 1.234, 2.456 );
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=0., errorbis=0., work[1];
float *hA, *hX, *hB, *hR;
float *dA[MagmaMaxGPUs], *dX[MagmaMaxGPUs], *dB[MagmaMaxGPUs], *dwork[MagmaMaxGPUs], *hwork[MagmaMaxGPUs+1];
float *dA2;
magma_int_t M, N, size, lda, ldda, msize, nb, nstream;
magma_int_t ione = 1;
magma_int_t iseed[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
float tol = opts.tolerance * lapackf77_slamch("E");
// default values
nb = (opts.nb > 0 ? opts.nb : 64);
nstream = (opts.nstream > 0 ? opts.nstream : 2);
magma_int_t gnode[MagmaMaxGPUs][MagmaMaxGPUs+2];
magma_int_t nbcmplx = 0;
magma_buildconnection_mgpu(gnode, &nbcmplx, opts.ngpu);
printf("Initializing communication pattern... GPU-ncmplx %d\n\n", (int) nbcmplx);
for (int i=0; i < nbcmplx; ++i) {
int myngpu = gnode[i][MagmaMaxGPUs];
printf("cmplx %d has %d gpu ", i, myngpu);
for(int j=0; j < myngpu; ++j)
printf(" %d", (int) gnode[i][j]);
printf("\n");
}
magma_int_t nbevents = 2;
magma_queue_t streams[MagmaMaxGPUs][20];
magma_event_t redevents[MagmaMaxGPUs][20];
magma_event_t redevents2[MagmaMaxGPUs][MagmaMaxGPUs*MagmaMaxGPUs+10];
for( int d = 0; d < opts.ngpu; ++d ) {
for( magma_int_t i = 0; i < nstream; ++i ) {
magma_queue_create( &streams[d][i] );
}
for( magma_int_t i = 0; i < nbevents; ++i ) {
cudaEventCreateWithFlags(&redevents[d][i], cudaEventDisableTiming);
cudaEventCreateWithFlags(&redevents2[d][i], cudaEventDisableTiming);
}
}
printf( "nb %d, ngpu %d, nstream %d version %d\n", (int) nb, (int) opts.ngpu, (int) nstream, (int) opts.version );
printf(" M N nb offset CPU GFlop/s (sec) GPU GFlop/s (sec) CUBLAS hemm (sec) ||R|| / ||A||*||X||\n");
printf("=========================================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
M = opts.msize[itest];
N = opts.nsize[itest];
for( int offset = 0; offset < N; offset += min(N,nb) ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
msize = M - offset;
lda = M;
ldda = ((M + 31)/32)*32;
size = lda*M;
gflops = FLOPS_SSYMM( MagmaLeft, (float)msize, (float)N ) / 1e9;
magma_int_t dworksiz = ldda*N*3;
magma_int_t hworksiz = lda*N;
TESTING_MALLOC_CPU( hA, float, lda*M );
TESTING_MALLOC_CPU( hX, float, lda*N );
TESTING_MALLOC_CPU( hB, float, lda*N );
TESTING_MALLOC_PIN( hR, float, lda*N );
for( int d = 0; d < opts.ngpu; ++d ) {
magma_int_t mlocal = ((M / nb) / opts.ngpu + 1) * nb;
magma_setdevice( d );
TESTING_MALLOC_DEV( dA[d], float, ldda*mlocal );
TESTING_MALLOC_DEV( dX[d], float, ldda*N );
TESTING_MALLOC_DEV( dB[d], float, ldda*N );
TESTING_MALLOC_DEV( dwork[d], float, dworksiz );
TESTING_MALLOC_PIN( hwork[d], float, hworksiz );
}
TESTING_MALLOC_PIN( hwork[opts.ngpu], float, lda*N );
if ( opts.check ) {
magma_setdevice( 0 );
TESTING_MALLOC_DEV( dA2, float, ldda*M );
}
lapackf77_slarnv( &ione, iseed, &size, hA );
magma_smake_symmetric( M, hA, lda );
size = lda*N;
lapackf77_slarnv( &ione, iseed, &size, hX );
lapackf77_slarnv( &ione, iseed, &size, hB );
lapackf77_slacpy( "Full", &M, &N, hB, &lda, hR, &lda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_ssetmatrix_1D_col_bcyclic( M, M, hA, lda, dA, ldda, opts.ngpu, nb );
for( int d = 0; d < opts.ngpu; ++d ) {
magma_setdevice( d );
//magmablasSetKernelStream( streams[ d ][ 0 ] );
magma_ssetmatrix( M, N, hX, lda, dX[d], ldda );
//if (d == 0) magma_ssetmatrix( M, N, hB, lda, dB[d], ldda ); // this is wrong coz when offset != 0 the gpu who do the beta*C may be not 0 so this should be related to stdev(starting device who own i=0 first col)
magma_ssetmatrix( M, N, hB, lda, dB[d], ldda );
}
//memset(hR, 0, lda*N*sizeof(float));
//trace_init( 1, opts.ngpu, nstream, (magma_queue_t*) streams );
//magma_int_t offset = 0; //nb;
gpu_time = magma_sync_wtime(0);
magmablas_ssymm_mgpu_com(
MagmaLeft, MagmaLower, msize, N,
calpha, dA, ldda, offset,
dX, ldda,
cbeta, dB, ldda, dwork, dworksiz, hR, lda, hwork, hworksiz,
opts.ngpu, nb, streams, nstream, redevents2, nbevents, gnode, nbcmplx);
gpu_time = magma_sync_wtime(0) - gpu_time;
gpu_perf = gflops / gpu_time;
#ifdef TRACING
char buf[80];
snprintf( buf, sizeof(buf), "ssymm-m%d-n%d-nb%d-stream%d-ngpu%d-run%d.svg",
(int) M, (int) N, (int) nb, (int) nstream, (int) opts.ngpu, (int) j );
trace_finalize( buf, "trace.css" );
#endif
/* ====================================================================
Performs operation using CUBLAS
=================================================================== */
if ( opts.check && iter == 0 ) {
magma_setdevice( 0 );
magmablasSetKernelStream( 0 );
magma_ssetmatrix( M, M, hA, lda, dA2, ldda );
magma_ssetmatrix( M, N, hX, lda, dX[0], ldda );
magma_ssetmatrix( M, N, hB, lda, dwork[0], ldda );
gpu_time2 = magma_sync_wtime(0);
magma_ssymm(
MagmaLeft, MagmaLower, msize, N,
calpha, dA2+offset*ldda+offset, ldda,
dX[0], ldda,
cbeta, dwork[0], ldda );
gpu_time2 = magma_sync_wtime(0) - gpu_time2;
gpu_perf2 = gflops / gpu_time2;
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.check ) {
// store ||A||*||X||
errorbis = lapackf77_slange("fro", &msize, &msize, hA+offset*lda+offset, &lda, work );
errorbis *= lapackf77_slange("fro", &msize, &N, hX, &lda, work );
//printf( "A =" ); magma_sprint( M, M, hA, lda );
//printf( "X =" ); magma_sprint( M, N, hX, lda );
//printf( "B =" ); magma_sprint( M, N, hB, lda );
cpu_time = magma_wtime();
blasf77_ssymm( "Left", "Lower", &msize, &N,
&calpha, hA+offset*lda+offset, &lda,
hX, &lda,
&cbeta, hB, &lda );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
/*
trace_file = fopen("AJETE/C", "w");
for (int j = 0; j < N; j++)
for (int i = 0; i < siz; i++)
fprintf(trace_file, "%10d%10d%40.30e\n", i+1, j+1, hB[j*lda+i]);
fclose(trace_file);
*/
magma_int_t firstprint=0;
for(magma_int_t dev=0; dev < opts.ngpu; ++dev) {
magma_setdevice( dev );
magma_sgetmatrix( M, N, dB[dev], ldda, hR, lda );
// compute relative error ||R||/||A||*||X||, where R := B_magma - B_lapack = R - B
size = lda*N;
blasf77_saxpy( &size, &c_neg_one, hB, &ione, hR, &ione );
error = lapackf77_slange("fro", &msize, &N, hR, &lda, work) / errorbis;
//printf( "R =" ); magma_sprint( M, N, hR, lda );
if (firstprint == 0) {
printf( "%5d %5d %5d %5d %7.1f (%7.4f) %7.1f (%7.4f) %7.1f (%7.4f) %8.2e %s\n",
(int) M, (int) N, (int) nb, (int) offset,
cpu_perf, cpu_time,
gpu_perf, gpu_time,
gpu_perf2, gpu_time2,
error, (error < tol ? "ok" : "failed") );
}
else {
printf( "%89s %8.2e %s\n", " ",
error, (error < tol ? "ok" : "failed") );
}
status += ! (error < tol);
firstprint =1;
}
} else {
printf( "%5d %5d %5d %5d --- ( --- ) %7.1f (%7.4f) --- ( --- ) ---\n",
(int) M, (int) N, (int) nb, (int) offset,
gpu_perf, gpu_time );
}
TESTING_FREE_CPU( hA );
TESTING_FREE_CPU( hX );
TESTING_FREE_CPU( hB );
TESTING_FREE_PIN( hR );
for( int d = 0; d < opts.ngpu; ++d ) {
magma_setdevice( d );
TESTING_FREE_DEV( dA[d] );
TESTING_FREE_DEV( dX[d] );
TESTING_FREE_DEV( dB[d] );
TESTING_FREE_DEV( dwork[d] );
TESTING_FREE_PIN( hwork[d] );
}
TESTING_FREE_PIN( hwork[opts.ngpu] );
if ( opts.check ) {
magma_setdevice( 0 );
TESTING_FREE_DEV( dA2 );
}
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
} // offset
printf( "\n" );
}
for( int d = 0; d < opts.ngpu; ++d ) {
magma_setdevice( d );
for( magma_int_t i = 0; i < nstream; ++i ) {
magma_queue_destroy( streams[d][i] );
}
for( magma_int_t i = 0; i < nbevents; ++i ) {
magma_event_destroy( redevents[d][i] );
magma_event_destroy( redevents2[d][i] );
}
}
TESTING_FINALIZE();
return status;
}
int main(int ac, char **av)
{
int lc, i; /* loop counter */
char *msg; /* message returned from parse_opts */
/* parse standard options */
if ((msg = parse_opts(ac, av, NULL, NULL)) != NULL)
tst_brkm(TBROK, NULL, "OPTION PARSING ERROR - %s", msg);
setup();
for (lc = 0; TEST_LOOPING(lc); lc++) {
Tst_count = 0;
for (i = 0; i < TST_TOTAL; i++) {
if (testcase[i].setupfunc &&
testcase[i].setupfunc() == -1) {
tst_resm(TWARN, "Failed to setup test %d."
" Skipping test", i);
continue;
} else {
TEST(syscall(__NR_swapoff, testcase[i].path));
}
if (testcase[i].cleanfunc &&
testcase[i].cleanfunc() == -1) {
tst_brkm(TBROK, cleanup, "cleanup failed,"
" quitting the test");
}
/* check return code */
if ((TEST_RETURN == -1) && (TEST_ERRNO == testcase[i].
exp_errno)) {
tst_resm(TPASS, "swapoff(2) expected failure;"
" Got errno - %s : %s",
testcase[i].exp_errval,
testcase[i].err_desc);
} else {
tst_resm(TFAIL, "swapoff(2) failed to produce"
" expected error; %d, errno"
": %s and got %d",
testcase[i].exp_errno,
testcase[i].exp_errval, TEST_ERRNO);
if ((TEST_RETURN == 0) && (i == 2)) {
if (syscall(__NR_swapon, "./swapfile01", 0) != 0) {
tst_brkm(TBROK, cleanup,
" Failed to turn on"
" swap file");
}
}
}
TEST_ERROR_LOG(TEST_ERRNO);
} /*End of TEST LOOPS */
}
/*Clean up and exit */
cleanup();
tst_exit();
} /*End of main */
int main(int ac, char **av)
{
struct stat stat_buf; /* stat structure buffer */
int lc;
const char *msg;
char *file_name; /* ptr. for file name whose mode is modified */
char *test_desc; /* test specific error message */
int ind; /* counter to test different test conditions */
if ((msg = parse_opts(ac, av, NULL, NULL)) != NULL)
tst_brkm(TBROK, NULL, "OPTION PARSING ERROR - %s", msg);
/*
* Invoke setup function to call individual test setup functions
* to simulate test conditions.
*/
setup();
/* set the expected errnos... */
TEST_EXP_ENOS(exp_enos);
for (lc = 0; TEST_LOOPING(lc); lc++) {
tst_count = 0;
for (ind = 0; Test_cases[ind].desc != NULL; ind++) {
file_name = Test_cases[ind].pathname;
test_desc = Test_cases[ind].desc;
#if !defined(UCLINUX)
if (file_name == High_address_node) {
file_name = (char *)get_high_address();
}
#endif
/*
* Call stat(2) to test different test conditions.
* verify that it fails with -1 return value and
* sets appropriate errno.
*/
TEST(stat(file_name, &stat_buf));
/* Check return code from stat(2) */
if (TEST_RETURN == -1) {
TEST_ERROR_LOG(TEST_ERRNO);
if (TEST_ERRNO == Test_cases[ind].exp_errno) {
tst_resm(TPASS,
"stat() fails, %s, errno:%d",
test_desc, TEST_ERRNO);
} else {
tst_resm(TFAIL,
"stat() fails, %s, errno:%d, expected errno:%d",
test_desc, TEST_ERRNO,
Test_cases[ind].exp_errno);
}
} else {
tst_resm(TFAIL,
"stat(2) returned %ld, expected -1, errno:%d",
TEST_RETURN,
Test_cases[ind].exp_errno);
}
}
tst_count++; /* incr TEST_LOOP counter */
}
/*
* Invoke cleanup() to delete the test directory/file(s) created
* in the setup().
*/
cleanup();
tst_exit();
}
int main(int ac, char *av[])
{
int *fildes;
int ifile;
char pfilname[40];
int min;
int serrno;
int lc;
const char *msg;
ifile = -1;
if ((msg = parse_opts(ac, av, NULL, NULL)) != NULL)
tst_brkm(TBROK, NULL, "OPTION PARSING ERROR - %s", msg);
local_flag = PASSED;
setup();
for (lc = 0; TEST_LOOPING(lc); lc++) {
min = getdtablesize(); /* get number of files allowed open */
fildes = malloc((min + 10) * sizeof(int));
if (fildes == NULL)
tst_brkm(TBROK | TERRNO, cleanup, "malloc error");
sprintf(pfilname, "./dup205.%d\n", getpid());
unlink(pfilname);
serrno = 0;
if ((fildes[0] = creat(pfilname, 0666)) == -1)
tst_brkm(TBROK | TERRNO, cleanup, "creat failed");
else {
fildes[fildes[0]] = fildes[0];
for (ifile = fildes[0] + 1; ifile < min + 10; ifile++) {
if ((fildes[ifile] = dup2(fildes[ifile - 1],
ifile)) == -1) {
serrno = errno;
break;
} else {
if (fildes[ifile] != ifile) {
tst_brkm(TFAIL, cleanup,
"got wrong descriptor "
"number back (%d != %d)",
fildes[ifile], ifile);
}
}
} /* end for */
if (ifile < min) {
tst_resm(TFAIL, "Not enough files duped");
local_flag = FAILED;
} else if (ifile > min) {
tst_resm(TFAIL, "Too many files duped");
local_flag = FAILED;
}
if (serrno != EBADF && serrno != EMFILE &&
serrno != EINVAL) {
tst_resm(TFAIL, "bad errno on dup2 failure");
local_flag = FAILED;
}
}
unlink(pfilname);
for (ifile = fildes[0]; ifile < min + 10; ifile++)
close(fildes[ifile]);
if (local_flag == PASSED) {
tst_resm(TPASS, "Test passed.");
} else {
tst_resm(TFAIL, "Test failed.");
}
}
cleanup();
tst_exit();
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing ssygvdx
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gpu_time;
float *h_A, *h_R, *h_work;
#if defined(PRECISION_z) || defined(PRECISION_c)
float *rwork;
magma_int_t lrwork;
#endif
/* Matrix size */
float *w1, *w2;
magma_int_t *iwork;
magma_int_t N, n2, info, lwork, liwork;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};;
magma_int_t info_ortho = 0;
magma_int_t info_solution = 0;
magma_int_t info_reduction = 0;
magma_timestr_t start, end;
magma_opts opts;
parse_opts( argc, argv, &opts );
magma_int_t ngpu = opts.ngpu;
char jobz = opts.jobz;
magma_int_t checkres = opts.check;
char range = 'A';
char uplo = opts.uplo;
magma_int_t itype = opts.itype;
float f = opts.fraction;
if (f != 1)
range='I';
if ( checkres && jobz == MagmaNoVec ) {
fprintf( stderr, "checking results requires vectors; setting jobz=V (option -JV)\n" );
jobz = MagmaVec;
}
printf("using: itype = %d, jobz = %c, range = %c, uplo = %c, checkres = %d, fraction = %6.4f\n",
(int) itype, jobz, range, uplo, (int) checkres, f);
printf(" N M GPU Time(s) \n");
printf("==========================\n");
magma_int_t threads = magma_get_numthreads();
for( magma_int_t i = 0; i < opts.ntest; ++i ) {
for( magma_int_t iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[i];
n2 = N*N;
#if defined(PRECISION_z) || defined(PRECISION_c)
lwork = magma_sbulge_get_lq2(N, threads) + 2*N + N*N;
lrwork = 1 + 5*N +2*N*N;
#else
lwork = magma_sbulge_get_lq2(N, threads) + 1 + 6*N + 2*N*N;
#endif
liwork = 3 + 5*N;
/* Allocate host memory for the matrix */
TESTING_MALLOC( h_A, float, n2);
TESTING_MALLOC( w1, float , N);
TESTING_MALLOC( w2, float , N);
TESTING_HOSTALLOC(h_R, float, n2);
TESTING_HOSTALLOC(h_work, float, lwork);
#if defined(PRECISION_z) || defined(PRECISION_c)
TESTING_HOSTALLOC( rwork, float, lrwork);
#endif
TESTING_MALLOC( iwork, magma_int_t, liwork);
/* Initialize the matrix */
lapackf77_slarnv( &ione, ISEED, &n2, h_A );
/* Make diagonal real */
for(int i=0; i<N; i++) {
MAGMA_S_SET2REAL( h_A[i*N+i], MAGMA_S_REAL(h_A[i*N+i]) );
}
magma_int_t m1 = 0;
float vl = 0;
float vu = 0;
magma_int_t il = 0;
magma_int_t iu = 0;
if (range == 'I'){
il = 1;
iu = (int) (f*N);
}
if(opts.warmup){
// ==================================================================
// Warmup using MAGMA
// ==================================================================
lapackf77_slacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
if(ngpu==1){
printf("calling ssyevdx_2stage 1 GPU\n");
magma_ssyevdx_2stage(jobz, range, uplo, N,
h_R, N,
vl, vu, il, iu,
&m1, w1,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info);
}else{
printf("calling ssyevdx_2stage_m %d GPU\n", (int) ngpu);
magma_ssyevdx_2stage_m(ngpu, jobz, range, uplo, N,
h_R, N,
vl, vu, il, iu,
&m1, w1,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info);
}
}
// ===================================================================
// Performs operation using MAGMA
// ===================================================================
lapackf77_slacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
start = get_current_time();
if(ngpu==1){
printf("calling ssyevdx_2stage 1 GPU\n");
magma_ssyevdx_2stage(jobz, range, uplo, N,
h_R, N,
vl, vu, il, iu,
&m1, w1,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info);
}else{
printf("calling ssyevdx_2stage_m %d GPU\n", (int) ngpu);
magma_ssyevdx_2stage_m(ngpu, jobz, range, uplo, N,
h_R, N,
vl, vu, il, iu,
&m1, w1,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info);
}
end = get_current_time();
gpu_time = GetTimerValue(start,end)/1000.;
if ( checkres ) {
float eps = lapackf77_slamch("E");
printf("\n");
printf("------ TESTS FOR MAGMA SSYEVD ROUTINE ------- \n");
printf(" Size of the Matrix %d by %d\n", (int) N, (int) N);
printf("\n");
printf(" The matrix A is randomly generated for each test.\n");
printf("============\n");
printf(" The relative machine precision (eps) is %8.2e\n",eps);
printf(" Computational tests pass if scaled residuals are less than 60.\n");
/* Check the orthogonality, reduction and the eigen solutions */
if (jobz == MagmaVec) {
info_ortho = check_orthogonality(N, N, h_R, N, eps);
info_reduction = check_reduction(uplo, N, 1, h_A, w1, N, h_R, eps);
}
printf("------ CALLING LAPACK SSYEVD TO COMPUTE only eigenvalue and verify elementswise ------- \n");
lapackf77_ssyevd("N", "L", &N,
h_A, &N, w2,
h_work, &lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, &lrwork,
#endif
iwork, &liwork,
&info);
info_solution = check_solution(N, w2, w1, eps);
if ( (info_solution == 0) & (info_ortho == 0) & (info_reduction == 0) ) {
printf("***************************************************\n");
printf(" ---- TESTING SSYEVD ...................... PASSED !\n");
printf("***************************************************\n");
}
else {
printf("************************************************\n");
printf(" - TESTING SSYEVD ... FAILED !\n");
printf("************************************************\n");
}
}
/* =====================================================================
Print execution time
=================================================================== */
printf("%5d %5d %6.2f\n",
(int) N, (int) m1, gpu_time);
TESTING_FREE( h_A);
TESTING_FREE( w1);
TESTING_FREE( w2);
#if defined(PRECISION_z) || defined(PRECISION_c)
TESTING_HOSTFREE( rwork);
#endif
TESTING_FREE( iwork);
TESTING_HOSTFREE(h_work);
TESTING_HOSTFREE( h_R);
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
/* Shutdown */
TESTING_FINALIZE();
return 0;
}
int main(int ac, char **av)
{
int lc; /* loop counter */
char *msg; /* message returned from parse_opts */
pid_t cpid; /* Child process id */
int status; /* child exit status */
/* Parse standard options given to run the test. */
msg = parse_opts(ac, av, (option_t *) NULL, NULL);
if (msg != (char *)NULL) {
tst_brkm(TBROK, tst_exit, "OPTION PARSING ERROR - %s", msg);
}
#ifdef UCLINUX
maybe_run_child(&do_child, "dddd", &timereq.tv_sec, &timereq.tv_nsec,
&timerem.tv_sec, &timerem.tv_nsec);
#endif
/* Perform global setup for test */
setup();
/* Check looping state if -i option given */
for (lc = 0; TEST_LOOPING(lc); lc++) {
/* Reset Tst_count in case we are looping. */
Tst_count = 0;
/*
* Creat a child process and suspend its
* execution using nanosleep()
*/
if ((cpid = FORK_OR_VFORK()) == -1) {
tst_brkm(TBROK, cleanup,
"fork() failed to create child process");
}
if (cpid == 0) { /* Child process */
#ifdef UCLINUX
if (self_exec(av[0], "dddd",
timereq.tv_sec, timereq.tv_nsec,
timerem.tv_sec, timerem.tv_nsec) < 0) {
tst_brkm(TBROK, cleanup, "self_exec failed");
}
#else
do_child();
#endif
}
/* wait for child to time slot for execution */
sleep(1);
/* Now send signal to child */
if (kill(cpid, SIGINT) < 0) {
tst_brkm(TBROK, cleanup,
"kill() fails send signal to child");
}
/* Wait for child to execute */
wait(&status);
if (WIFEXITED(status) && WEXITSTATUS(status) == 0) {
tst_resm(TPASS, "Functionality of nanosleep() is "
"correct");
} else {
tst_resm(TFAIL, "child process exited abnormally; "
"status = %d", status);
}
} /* End for TEST_LOOPING */
/* Call cleanup() to undo setup done for the test. */
cleanup();
/*NOTREACHED*/ return 0;
} /* End main */
int main(int ac, char **av)
{
int lc; /* loop counter */
char *msg; /* message returned from parse_opts */
struct flock flock;
/* parse standard options */
if ((msg = parse_opts(ac, av, NULL, NULL)) != NULL) {
tst_brkm(TBROK, NULL, "OPTION PARSING ERROR - %s", msg);
}
setup();
/* check for looping state if -i option is given */
for (lc = 0; TEST_LOOPING(lc); lc++) {
/* reset Tst_count in case we are looping */
Tst_count = 0;
/* //block1: */
tst_resm(TINFO, "Enter block 1");
tst_resm(TINFO, "Test for errno EINVAL");
fail = 0;
if (fcntl(1, F_BADCMD, 1) != -1) {
tst_resm(TFAIL, "fcntl(2) failed to FAIL");
fail = 1;
} else if (errno != EINVAL) {
tst_resm(TFAIL, "Expected EINVAL got %d", errno);
fail = 1;
}
if (fail) {
tst_resm(TINFO, "block 1 FAILED");
} else {
tst_resm(TINFO, "block 1 PASSED");
}
tst_resm(TINFO, "Exit block 1");
/* //block2: */
#ifndef UCLINUX
/* Skip since uClinux does not implement memory protection */
tst_resm(TINFO, "Enter block 2");
tst_resm(TINFO, "Test for errno EFAULT");
fail = 0;
/* case 1: F_SETLK */
if (fcntl(1, F_SETLK, (void *)-1) != -1) {
tst_resm(TFAIL, "F_SETLK: fcntl(2) failed to FAIL");
fail = 1;
} else if (errno != EFAULT) {
tst_resm(TFAIL, "F_SETLK: Expected EFAULT got %d",
errno);
fail = 1;
}
/* case 2: F_SETLKW */
if (fcntl(1, F_SETLKW, (void *)-1) != -1) {
tst_resm(TFAIL, "F_SETLKW: fcntl(2) failed to FAIL");
fail = 1;
} else if (errno != EFAULT) {
tst_resm(TFAIL, "F_SETLKW: Expected EFAULT got %d",
errno);
fail = 1;
}
/* case 3: F_GETLK */
if (fcntl(1, F_GETLK, (void *)-1) != -1) {
tst_resm(TFAIL, "F_GETLK: fcntl(2) failed to FAIL");
fail = 1;
} else if (errno != EFAULT) {
tst_resm(TFAIL, "F_GETLK: Expected EFAULT got %d",
errno);
fail = 1;
}
if (fail) {
tst_resm(TINFO, "blcok 2 FAILED");
} else {
tst_resm(TINFO, "block 2 PASSED");
}
tst_resm(TINFO, "Exit block 2");
#else
tst_resm(TINFO, "Skip block 2 on uClinux");
#endif
/* //block3: */
tst_resm(TINFO, "Enter block 3");
tst_resm(TINFO, "Test for errno EINVAL");
fail = 0;
flock.l_whence = -1;
flock.l_type = F_WRLCK;
flock.l_start = 0L;
flock.l_len = 0L;
if (fcntl(1, F_SETLK, &flock) != -1) {
tst_resm(TFAIL, "fcntl(2) failed to FAIL");
fail = 1;
} else if (errno != EINVAL) {
tst_resm(TFAIL, "Expected EINVAL, got %d", errno);
fail = 1;
}
if (fail) {
tst_resm(TINFO, "block 3 FAILED");
} else {
tst_resm(TINFO, "block 3 PASSED");
}
tst_resm(TINFO, "Exit block 3");
/* //block4: */
tst_resm(TINFO, "Enter block 4");
tst_resm(TINFO, "Test for errno EBADF");
fail = 0;
if (fcntl(-1, F_GETLK, &flock) != -1) {
tst_resm(TFAIL, "fcntl(2) failed to FAIL");
fail = 1;
} else if (errno != EBADF) {
tst_resm(TFAIL, "Expected EBADF, got %d", errno);
fail = 1;
}
if (fail) {
tst_resm(TINFO, "block 4 FAILED");
} else {
tst_resm(TINFO, "block 4 PASSED");
}
tst_resm(TINFO, "Exit block 4");
}
cleanup();
tst_exit();
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zgemm
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, magma_perf, magma_time, dev_perf, dev_time, cpu_perf, cpu_time;
double magma_error, dev_error, Cnorm, work[1];
magma_int_t M, N, K;
magma_int_t Am, An, Bm, Bn;
magma_int_t sizeA, sizeB, sizeC;
magma_int_t lda, ldb, ldc, ldda, lddb, lddc;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magmaDoubleComplex *h_A, *h_B, *h_C, *h_Cmagma, *h_Cdev;
magmaDoubleComplex_ptr d_A, d_B, d_C;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex alpha = MAGMA_Z_MAKE( 0.29, -0.86 );
magmaDoubleComplex beta = MAGMA_Z_MAKE( -0.48, 0.38 );
magma_opts opts;
parse_opts( argc, argv, &opts );
double tol = opts.tolerance * lapackf77_dlamch("E");
#ifdef HAVE_CUBLAS
// for CUDA, we can check MAGMA vs. CUBLAS, without running LAPACK
printf("If running lapack (option --lapack), MAGMA and %s error are both computed\n"
"relative to CPU BLAS result. Else, MAGMA error is computed relative to %s result.\n\n",
g_platform_str, g_platform_str );
printf("transA = %s, transB = %s\n",
lapack_trans_const(opts.transA),
lapack_trans_const(opts.transB) );
printf(" M N K MAGMA Gflop/s (ms) %s Gflop/s (ms) CPU Gflop/s (ms) MAGMA error %s error\n",
g_platform_str, g_platform_str );
#else
// for others, we need LAPACK for check
opts.lapack |= opts.check; // check (-c) implies lapack (-l)
printf("transA = %s, transB = %s\n",
lapack_trans_const(opts.transA),
lapack_trans_const(opts.transB) );
printf(" M N K %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];
K = opts.ksize[itest];
gflops = FLOPS_ZGEMM( M, N, K ) / 1e9;
if ( opts.transA == MagmaNoTrans ) {
lda = Am = M;
An = K;
} else {
lda = Am = K;
An = M;
}
if ( opts.transB == MagmaNoTrans ) {
ldb = Bm = K;
Bn = N;
} else {
ldb = Bm = N;
Bn = K;
}
ldc = M;
ldda = ((lda+31)/32)*32;
lddb = ((ldb+31)/32)*32;
lddc = ((ldc+31)/32)*32;
sizeA = lda*An;
sizeB = ldb*Bn;
sizeC = ldc*N;
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, lda*An );
TESTING_MALLOC_CPU( h_B, magmaDoubleComplex, ldb*Bn );
TESTING_MALLOC_CPU( h_C, magmaDoubleComplex, ldc*N );
TESTING_MALLOC_CPU( h_Cmagma, magmaDoubleComplex, ldc*N );
TESTING_MALLOC_CPU( h_Cdev, magmaDoubleComplex, ldc*N );
TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, ldda*An );
TESTING_MALLOC_DEV( d_B, magmaDoubleComplex, lddb*Bn );
TESTING_MALLOC_DEV( d_C, magmaDoubleComplex, lddc*N );
/* Initialize the matrices */
lapackf77_zlarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_zlarnv( &ione, ISEED, &sizeB, h_B );
lapackf77_zlarnv( &ione, ISEED, &sizeC, h_C );
magma_zsetmatrix( Am, An, h_A, lda, d_A, ldda );
magma_zsetmatrix( Bm, Bn, h_B, ldb, d_B, lddb );
/* =====================================================================
Performs operation using MAGMABLAS (currently only with CUDA)
=================================================================== */
#ifdef HAVE_CUBLAS
magma_zsetmatrix( M, N, h_C, ldc, d_C, lddc );
magma_time = magma_sync_wtime( NULL );
magmablas_zgemm( opts.transA, opts.transB, M, N, K,
alpha, d_A, ldda,
d_B, lddb,
beta, d_C, lddc );
magma_time = magma_sync_wtime( NULL ) - magma_time;
magma_perf = gflops / magma_time;
magma_zgetmatrix( M, N, d_C, lddc, h_Cmagma, ldc );
#endif
/* =====================================================================
Performs operation using CUBLAS / clBLAS / Xeon Phi MKL
=================================================================== */
magma_zsetmatrix( M, N, h_C, ldc, d_C, lddc );
#ifdef HAVE_CUBLAS
dev_time = magma_sync_wtime( NULL );
cublasZgemm( opts.handle, cublas_trans_const(opts.transA), cublas_trans_const(opts.transB), M, N, K,
&alpha, d_A, ldda,
d_B, lddb,
&beta, d_C, lddc );
dev_time = magma_sync_wtime( NULL ) - dev_time;
#else
dev_time = magma_sync_wtime( opts.queue );
magma_zgemm( opts.transA, opts.transB, M, N, K,
alpha, d_A, 0, ldda,
d_B, 0, lddb,
beta, d_C, 0, lddc, opts.queue );
dev_time = magma_sync_wtime( opts.queue ) - dev_time;
#endif
dev_perf = gflops / dev_time;
magma_zgetmatrix( M, N, d_C, lddc, h_Cdev, ldc );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
blasf77_zgemm( lapack_trans_const(opts.transA), lapack_trans_const(opts.transB), &M, &N, &K,
&alpha, h_A, &lda,
h_B, &ldb,
&beta, h_C, &ldc );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
}
/* =====================================================================
Check the result
=================================================================== */
if ( opts.lapack ) {
// compute relative error for both magma & dev, relative to lapack,
// |C_magma - C_lapack| / |C_lapack|
Cnorm = lapackf77_zlange( "F", &M, &N, h_C, &ldc, work );
blasf77_zaxpy( &sizeC, &c_neg_one, h_C, &ione, h_Cdev, &ione );
dev_error = lapackf77_zlange( "F", &M, &N, h_Cdev, &ldc, work ) / Cnorm;
#ifdef HAVE_CUBLAS
blasf77_zaxpy( &sizeC, &c_neg_one, h_C, &ione, h_Cmagma, &ione );
magma_error = lapackf77_zlange( "F", &M, &N, h_Cmagma, &ldc, work ) / Cnorm;
printf("%5d %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %s\n",
(int) M, (int) N, (int) K,
magma_perf, 1000.*magma_time,
dev_perf, 1000.*dev_time,
cpu_perf, 1000.*cpu_time,
magma_error, dev_error,
(magma_error < tol && dev_error < tol ? "ok" : "failed"));
status += ! (magma_error < tol && dev_error < tol);
#else
printf("%5d %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) M, (int) N, (int) K,
dev_perf, 1000.*dev_time,
cpu_perf, 1000.*cpu_time,
dev_error,
(dev_error < tol ? "ok" : "failed"));
status += ! (dev_error < tol);
#endif
}
else {
#ifdef HAVE_CUBLAS
// compute relative error for magma, relative to dev (currently only with CUDA)
Cnorm = lapackf77_zlange( "F", &M, &N, h_Cdev, &ldc, work );
blasf77_zaxpy( &sizeC, &c_neg_one, h_Cdev, &ione, h_Cmagma, &ione );
magma_error = lapackf77_zlange( "F", &M, &N, h_Cmagma, &ldc, work ) / Cnorm;
printf("%5d %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) --- ( --- ) %8.2e --- %s\n",
(int) M, (int) N, (int) K,
magma_perf, 1000.*magma_time,
dev_perf, 1000.*dev_time,
magma_error,
(magma_error < tol ? "ok" : "failed"));
status += ! (magma_error < tol);
#else
printf("%5d %5d %5d %7.2f (%7.2f) --- ( --- ) ---\n",
(int) M, (int) N, (int) K,
dev_perf, 1000.*dev_time );
#endif
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_C );
TESTING_FREE_CPU( h_Cmagma );
TESTING_FREE_CPU( h_Cdev );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
TESTING_FREE_DEV( d_C );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
int main(int ac, char **av)
{
int lc;
const char *msg;
int rval;
pid_t pid, pid1;
int status;
struct passwd *ltpuser1, *ltpuser2;
/*
* parse standard options
*/
if ((msg = parse_opts(ac, av, NULL, NULL)) != NULL) {
tst_brkm(TBROK, NULL, "OPTION PARSING ERROR - %s", msg);
}
/*
* perform global setup for test
*/
setup();
/* set the expected errnos... */
TEST_EXP_ENOS(exp_enos);
/*
* check looping state if -i option given
*/
for (lc = 0; TEST_LOOPING(lc); lc++) {
tst_count = 0;
/* Initialize the test directories name */
sprintf(tstdir1, "tstdir1.%d", getpid());
ltpuser1 = my_getpwnam(user1name);
if ((pid = FORK_OR_VFORK()) < 0) {
tst_brkm(TBROK, cleanup, "fork #1 failed");
}
if (pid == 0) { /* first child */
/* set to ltpuser1 */
rval = setreuid(ltpuser1->pw_uid, ltpuser1->pw_uid);
if (rval < 0) {
tst_resm(TFAIL, "setreuid failed to "
"to set the real uid to %d and "
"effective uid to %d",
ltpuser1->pw_uid, ltpuser1->pw_uid);
perror("setreuid");
exit(1);
}
/* create the parent directory with 0700 permits */
if (mkdir(tstdir1, PERMS) == -1) {
tst_resm(TFAIL, "mkdir(%s, %#o) Failed",
tstdir1, PERMS);
exit(1);
}
/* create tstdir1 succeeded */
exit(0);
}
wait(&status);
if (WEXITSTATUS(status) != 0) {
tst_brkm(TFAIL, cleanup,
"Test to check mkdir EACCES failed"
"in create parent directory");
}
sprintf(tstdir2, "%s/tst", tstdir1);
ltpuser2 = my_getpwnam(user2name);
if ((pid1 = FORK_OR_VFORK()) < 0) {
tst_brkm(TBROK, cleanup, "fork #2 failed");
}
if (pid1 == 0) { /* second child */
/* set to ltpuser2 */
rval = setreuid(ltpuser2->pw_uid, ltpuser2->pw_uid);
if (rval < 0) {
tst_resm(TFAIL, "setreuid failed to "
"to set the real uid to %d and "
"effective uid to %d",
ltpuser2->pw_uid, ltpuser2->pw_uid);
perror("setreuid");
exit(1);
}
if (mkdir(tstdir2, PERMS) != -1) {
tst_resm(TFAIL, "mkdir(%s, %#o) unexpected "
"succeeded", tstdir2, PERMS);
exit(1);
}
if (errno != EACCES) {
tst_resm(TFAIL, "Expected EACCES got %d",
errno);
exit(1);
}
/* PASS */
exit(0);
}
waitpid(pid1, &status, 0);
if (WEXITSTATUS(status) == 0) {
tst_resm(TPASS, "Test to attempt to creat a directory "
"in a directory having no permissions "
"SUCCEEDED in setting errno to EACCES");
} else {
tst_resm(TFAIL, "Test to attempt to creat a directory "
"in a directory having no permissions FAILED");
cleanup();
}
}
/*
* cleanup and exit
*/
cleanup();
tst_exit();
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zhetrd_he2hb
*/
int main( int argc, char** argv)
{
TESTING_INIT_MGPU();
real_Double_t gpu_time, gpu_perf, gflops;
magmaDoubleComplex *h_A, *h_R, *h_work, *dT1;
magmaDoubleComplex *tau;
double *D, *E;
/* Matrix size */
magma_int_t N, n2, lda, lwork, ldt, lwork0;
magma_int_t info;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
#if defined(CHECKEIG)
#if defined(PRECISION_z) || defined(PRECISION_d)
magma_int_t WANTZ=0;
magma_int_t THREADS=1;
#endif
#endif
magma_int_t NE = 0;
magma_int_t NB = 0;
magma_int_t ngpu = 1;
magma_opts opts;
parse_opts( argc, argv, &opts );
NB = opts.nb;
if (NB < 1)
NB = 64; //64; //magma_get_zhetrd_he2hb_nb(N);
// what is NE ?
if (NE < 1)
NE = 64; //N; //magma_get_zhetrd_he2hb_nb(N); // N not yet initialized
printf(" N GPU GFlop/s \n");
printf("=====================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
ldt = N;
n2 = N*lda;
gflops = FLOPS_ZHETRD( N ) / 1e9;
/* We suppose the magma NB is bigger than lapack NB */
lwork0 = N*NB;
/* Allocate host memory for the matrix */
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, lda*N );
TESTING_MALLOC_CPU( tau, magmaDoubleComplex, N-1 );
TESTING_MALLOC_PIN( h_R, magmaDoubleComplex, lda*N );
TESTING_MALLOC_PIN( h_work, magmaDoubleComplex, lwork0 );
TESTING_MALLOC_PIN( D, double, N );
TESTING_MALLOC_PIN( E, double, N );
//TESTING_MALLOC_DEV( dT1, magmaDoubleComplex, (2*min(N,N)+(N+31)/32*32)*NB );
TESTING_MALLOC_DEV( dT1, magmaDoubleComplex, (N*NB) );
// if (WANTZ) gflops = 2.0*gflops;
/* ====================================================================
Initialize the matrix
=================================================================== */
lapackf77_zlarnv( &ione, ISEED, &n2, h_A );
magma_zmake_hermitian( N, h_A, lda );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_device_t cdev;
magma_getdevice( &cdev );
gpu_time = magma_wtime();
/*
magma_zhetrd_he2hb( opts.uplo, N, NB, h_R, lda, tau, h_work, lwork0, dT1, THREADS, &info);
tband = magma_wtime - gpu_time();
printf(" Finish BAND N %d NB %d ngpu %d timing= %f\n", N, NB, ngpu, tband);
magma_zhetrd_bhe2trc_v5(THREADS, WANTZ, opts.uplo, NE, N, NB, h_R, lda, D, E, dT1, ldt);
*/
/*
magma_zhetrd_he2hb( opts.uplo, N, NB, h_R, lda, tau, h_work, lwork, dT1, THREADS, &info);
tband = magma_wtime - gpu_time();
printf(" Finish BAND N %d NB %d ngpu %d timing= %f\n", N, NB, ngpu, tband);
magma_zhetrd_bhe2trc(THREADS, WANTZ, opts.uplo, NE, N, NB, h_R, lda, D, E, dT1, ldt);
*/
magma_range_t range = MagmaRangeAll;
magma_int_t fraction_ev = 100;
magma_int_t il, iu, m1;
double vl=0., vu=0.;
if (fraction_ev == 0) {
il = N / 10;
iu = N / 5+il;
}
else {
il = 1;
iu = (int)(fraction_ev*N);
if (iu < 1) iu = 1;
}
magmaDoubleComplex *hh_work;
magma_int_t *iwork;
magma_int_t nb, /*lwork,*/ liwork;
magma_int_t threads = magma_get_parallel_numthreads();
#if defined(PRECISION_z) || defined(PRECISION_c)
double *rwork;
magma_int_t lrwork;
lwork = magma_zbulge_get_lq2(N, threads) + 2*N + N*N;
lrwork = 1 + 5*N +2*N*N;
TESTING_MALLOC_PIN( rwork, double, lrwork );
#else
lwork = magma_zbulge_get_lq2(N, threads) + 1 + 6*N + 2*N*N;
#endif
liwork = 3 + 5*N;
nb = magma_get_zhetrd_nb(N);
TESTING_MALLOC_PIN( hh_work, magmaDoubleComplex, lwork );
TESTING_MALLOC_CPU( iwork, magma_int_t, liwork );
if (ngpu == 1) {
printf("calling zheevdx_2stage 1 GPU\n");
magma_zheevdx_2stage( opts.jobz, range, opts.uplo, N,
h_R, lda,
vl, vu, il, iu,
&m1, D,
hh_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info);
} else {
printf("calling zheevdx_2stage_m %d GPU\n", (int) ngpu);
magma_zheevdx_2stage_m(ngpu, opts.jobz, range, opts.uplo, N,
h_R, lda,
vl, vu, il, iu,
&m1, D,
hh_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info);
}
magma_setdevice( cdev );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
/* =====================================================================
Check the factorization
=================================================================== */
/*
if ( opts.check ) {
FILE *fp ;
printf("Writing input matrix in matlab_i_mat.txt ...\n");
fp = fopen ("matlab_i_mat.txt", "w") ;
if ( fp == NULL ) { printf("Couldn't open output file\n"); exit(1); }
for (j=0; j < N; j++) {
for (k=0; k < N; k++) {
#if defined(PRECISION_z) || defined(PRECISION_c)
fprintf(fp, "%5d %5d %11.8f %11.8f\n", k+1, j+1,
h_A[k+j*lda].x, h_A[k+j*lda].y);
#else
fprintf(fp, "%5d %5d %11.8f\n", k+1, j+1, h_A[k+j*lda]);
#endif
}
}
fclose( fp ) ;
printf("Writing output matrix in matlab_o_mat.txt ...\n");
fp = fopen ("matlab_o_mat.txt", "w") ;
if ( fp == NULL ) { printf("Couldn't open output file\n"); exit(1); }
for (j=0; j < N; j++) {
for (k=0; k < N; k++) {
#if defined(PRECISION_z) || defined(PRECISION_c)
fprintf(fp, "%5d %5d %11.8f %11.8f\n", k+1, j+1,
h_R[k+j*lda].x, h_R[k+j*lda].y);
#else
fprintf(fp, "%5d %5d %11.8f\n", k+1, j+1, h_R[k+j*lda]);
#endif
}
}
fclose( fp ) ;
}
*/
/* =====================================================================
Print performance and error.
=================================================================== */
#if defined(CHECKEIG)
#if defined(PRECISION_z) || defined(PRECISION_d)
if ( opts.check ) {
printf(" Total N %5d gflops %6.2f timing %6.2f seconds\n", (int) N, gpu_perf, gpu_time );
char JOBZ;
if (WANTZ == 0)
JOBZ = 'N';
else
JOBZ = 'V';
double nrmI=0.0, nrm1=0.0, nrm2=0.0;
int lwork2 = 256*N;
magmaDoubleComplex *work2, *AINIT;
double *rwork2, *D2;
// TODO free this memory !
magma_zmalloc_cpu( &work2, lwork2 );
magma_dmalloc_cpu( &rwork2, N );
magma_dmalloc_cpu( &D2, N );
magma_zmalloc_cpu( &AINIT, N*lda );
memcpy(AINIT, h_A, N*lda*sizeof(magmaDoubleComplex));
/* compute the eigenvalues using lapack routine to be able to compare to it and used as ref */
cpu_time = magma_wtime();
i= min(12, THREADS);
#if defined(USEMKL)
mkl_set_num_threads( i );
#endif
#if defined(USEACML)
omp_set_num_threads(i);
#endif
lapackf77_zheev( "N", "L", &N, h_A, &lda, D2, work2, &lwork2,
#if defined(PRECISION_z) || defined (PRECISION_c)
rwork2,
#endif
&info );
///* call eigensolver for our resulting tridiag [D E] and for Q */
//dstedc_withZ('V', N, D, E, h_R, lda);
////dsterf_( &N, D, E, &info);
////
cpu_time = magma_wtime() - cpu_time;
printf(" Finish CHECK - EIGEN timing= %f threads %d\n", cpu_time, i);
/*
for (i=0; i < 10; i++)
printf(" voici lpk D[%d] %8.2e\n", i, D2[i]);
*/
//magmaDoubleComplex mydz=0.0, mydo=1.0;
//magmaDoubleComplex *Z;
// magma_zmalloc_cpu( &Z, N*lda );
// dgemm_("N", "N", &N, &N, &N, &mydo, h_R, &lda, h_A, &lda, &mydz, Z, &lda);
/* compare result */
cmp_vals(N, D2, D, &nrmI, &nrm1, &nrm2);
magmaDoubleComplex *WORKAJETER;
double *RWORKAJETER, *RESU;
// TODO free this memory !
magma_zmalloc_cpu( &WORKAJETER, (2* N * N + N) );
magma_dmalloc_cpu( &RWORKAJETER, N );
magma_dmalloc_cpu( &RESU, 10 );
int MATYPE;
memset(RESU, 0, 10*sizeof(double));
MATYPE=3;
double NOTHING=0.0;
cpu_time = magma_wtime();
// check results
zcheck_eig_(&JOBZ, &MATYPE, &N, &NB, AINIT, &lda, &NOTHING, &NOTHING, D2, D, h_R, &lda, WORKAJETER, RWORKAJETER, RESU );
cpu_time = magma_wtime() - cpu_time;
printf(" Finish CHECK - results timing= %f\n", cpu_time);
#if defined(USEMKL)
mkl_set_num_threads( 1 );
#endif
#if defined(USEACML)
omp_set_num_threads(1);
#endif
printf("\n");
printf(" ================================================================================================================\n");
printf(" ==> INFO voici threads=%d N=%d NB=%d WANTZ=%d\n", (int) THREADS, (int) N, (int) NB, (int) WANTZ);
printf(" ================================================================================================================\n");
printf(" DSBTRD : %15s \n", "STATblgv9withQ ");
printf(" ================================================================================================================\n");
if (WANTZ > 0)
printf(" | A - U S U' | / ( |A| n ulp ) : %15.3E \n", RESU[0]);
if (WANTZ > 0)
printf(" | I - U U' | / ( n ulp ) : %15.3E \n", RESU[1]);
printf(" | D1 - EVEIGS | / (|D| ulp) : %15.3E \n", RESU[2]);
printf(" max | D1 - EVEIGS | : %15.3E \n", RESU[6]);
printf(" ================================================================================================================\n\n\n");
printf(" ****************************************************************************************************************\n");
printf(" * Hello here are the norm Infinite (max)=%8.2e norm one (sum)=%8.2e norm2(sqrt)=%8.2e *\n", nrmI, nrm1, nrm2);
printf(" ****************************************************************************************************************\n\n");
}
#endif
#endif
printf(" Total N %5d gflops %6.2f timing %6.2f seconds\n", (int) N, gpu_perf, gpu_time );
printf("============================================================================\n\n\n");
/* Memory clean up */
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( tau );
TESTING_FREE_PIN( h_R );
TESTING_FREE_PIN( h_work );
TESTING_FREE_PIN( D );
TESTING_FREE_PIN( E );
TESTING_FREE_DEV( dT1 );
/* TODO - not all memory has been freed inside loop */
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE_MGPU();
return EXIT_SUCCESS;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing strmm
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, cublas_perf, cublas_time, cpu_perf, cpu_time;
float cublas_error, Cnorm, work[1];
magma_int_t M, N;
magma_int_t Ak;
magma_int_t sizeA, sizeB;
magma_int_t lda, ldb, ldda, lddb;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
float *h_A, *h_B, *h_Bcublas;
float *d_A, *d_B;
float c_neg_one = MAGMA_S_NEG_ONE;
float alpha = MAGMA_S_MAKE( 0.29, -0.86 );
magma_opts opts;
parse_opts( argc, argv, &opts );
printf("If running lapack (option --lapack), MAGMA and CUBLAS error are both computed\n"
"relative to CPU BLAS result. Else, MAGMA error is computed relative to CUBLAS result.\n\n"
"side = %c, uplo = %c, transA = %c, diag = %c \n", opts.side, opts.uplo, opts.transA, opts.diag );
printf(" M N CUBLAS Gflop/s (ms) CPU Gflop/s (ms) CUBLAS error\n");
printf("==================================================================\n");
for( int i = 0; i < opts.ntest; ++i ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[i];
N = opts.nsize[i];
gflops = FLOPS_STRMM(opts.side, M, N) / 1e9;
if ( opts.side == MagmaLeft ) {
lda = M;
Ak = M;
} else {
lda = N;
Ak = N;
}
ldb = M;
ldda = ((lda+31)/32)*32;
lddb = ((ldb+31)/32)*32;
sizeA = lda*Ak;
sizeB = ldb*N;
TESTING_MALLOC( h_A, float, lda*Ak );
TESTING_MALLOC( h_B, float, ldb*N );
TESTING_MALLOC( h_Bcublas, float, ldb*N );
TESTING_DEVALLOC( d_A, float, ldda*Ak );
TESTING_DEVALLOC( d_B, float, lddb*N );
/* Initialize the matrices */
lapackf77_slarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_slarnv( &ione, ISEED, &sizeB, h_B );
/* =====================================================================
Performs operation using CUDA-BLAS
=================================================================== */
magma_ssetmatrix( Ak, Ak, h_A, lda, d_A, ldda );
magma_ssetmatrix( M, N, h_B, ldb, d_B, lddb );
cublas_time = magma_sync_wtime( NULL );
cublasStrmm( opts.side, opts.uplo, opts.transA, opts.diag,
M, N,
alpha, d_A, ldda,
d_B, lddb );
cublas_time = magma_sync_wtime( NULL ) - cublas_time;
cublas_perf = gflops / cublas_time;
magma_sgetmatrix( M, N, d_B, lddb, h_Bcublas, ldb );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
blasf77_strmm( &opts.side, &opts.uplo, &opts.transA, &opts.diag,
&M, &N,
&alpha, h_A, &lda,
h_B, &ldb );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
}
/* =====================================================================
Check the result
=================================================================== */
if ( opts.lapack ) {
// compute relative error for both magma & cublas, relative to lapack,
// |C_magma - C_lapack| / |C_lapack|
Cnorm = lapackf77_slange( "M", &M, &N, h_B, &ldb, work );
blasf77_saxpy( &sizeB, &c_neg_one, h_B, &ione, h_Bcublas, &ione );
cublas_error = lapackf77_slange( "M", &M, &N, h_Bcublas, &ldb, work ) / Cnorm;
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e\n",
(int) M, (int) N,
cublas_perf, 1000.*cublas_time,
cpu_perf, 1000.*cpu_time,
cublas_error );
}
else {
printf("%5d %5d %7.2f (%7.2f) --- ( --- ) --- ---\n",
(int) M, (int) N,
cublas_perf, 1000.*cublas_time);
}
TESTING_FREE( h_A );
TESTING_FREE( h_B );
TESTING_FREE( h_Bcublas );
TESTING_DEVFREE( d_A );
TESTING_DEVFREE( d_B );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return 0;
}
int main(int ac, char **av)
{
int lc;
char *ptr; /* message returned from parse_opts */
gid_t group;
int i;
int entries; /* number of group entries */
initgroups("root", 0);
if ((ptr = parse_opts(ac, av, NULL, NULL)) != NULL)
tst_brkm(TBROK, NULL, "OPTION PARSING ERROR - %s", ptr);
setup();
for (lc = 0; TEST_LOOPING(lc); lc++) {
Tst_count = 0;
TEST(getgroups(-1, gidset));
if (TEST_RETURN == 0)
tst_resm(TFAIL, "getgroups succeeded unexpectedly");
else if (STD_FUNCTIONAL_TEST) {
if (errno == EINVAL)
tst_resm(TPASS,
"getgroups failed as expected with EINVAL");
else
tst_resm(TFAIL | TTERRNO,
"getgroups didn't fail as expected with EINVAL");
}
/*
* Check that if ngrps is zero that the number of groups is
* return and the the gidset array is not modified.
* This is a POSIX special case.
*/
memset(gidset, 052, NGROUPS);
memset(cmpset, 052, NGROUPS);
TEST(getgroups(0, gidset));
if (TEST_RETURN == -1)
tst_resm(TFAIL | TTERRNO, "getgroups failed");
else if (STD_FUNCTIONAL_TEST) {
if (memcmp(cmpset, gidset, NGROUPS) != 0)
tst_resm(TFAIL,
"getgroups modified the gidset array");
else
tst_resm(TPASS,
"getgroups did not modify the gidset "
"array");
}
/*
* Check to see that is -1 is returned and errno is set to
* EINVAL when ngroups is not big enough to hold all groups.
*/
if (TEST_RETURN <= 1)
tst_resm(TCONF,
"getgroups returned %ld; unable to test that using ngrps >=1 but less than number of grps",
TEST_RETURN);
else {
TEST(getgroups(TEST_RETURN - 1, gidset));
if (TEST_RETURN == -1) {
if (STD_FUNCTIONAL_TEST) {
if (errno == EINVAL)
tst_resm(TPASS,
"getgroups failed as "
"expected with EINVAL");
else
tst_resm(TFAIL | TERRNO,
"getgroups didn't fail "
"with EINVAL");
}
} else
tst_resm(TFAIL,
"getgroups succeeded unexpectedly with %ld",
TEST_RETURN);
}
TEST(getgroups(NGROUPS, gidset));
if ((entries = TEST_RETURN) == -1)
tst_resm(TFAIL | TTERRNO,
"getgroups failed unexpectedly");
else if (STD_FUNCTIONAL_TEST) {
group = getgid();
for (i = 0; i < entries; i++)
if (gidset[i] == group) {
tst_resm(TPASS,
"getgroups(NGROUPS,gidset) "
"returned %d contains gid %d "
"(from getgid)",
entries, group);
break;
}
if (i == entries)
tst_resm(TFAIL,
"getgroups(NGROUPS,gidset) ret %d, does "
"not contain gid %d (from getgid)",
entries, group);
}
}
cleanup();
tst_exit();
}
int
main(int ac, char **av)
{
int lc; /* loop counter */
char *msg; /* message returned from parse_opts */
pid_t cpid; /* process id of the child process */
int exit_status; /* exit status of child process */
sigset_t PendSig; /* variable to hold pending signal */
/* Parse standard options given to run the test. */
msg = parse_opts(ac, av, (option_t *) NULL, NULL);
if (msg != (char *) NULL) {
tst_brkm(TBROK, NULL, "OPTION PARSING ERROR - %s", msg);
tst_exit();
}
/* Perform global setup for test */
setup();
/* set the expected errnos... */
TEST_EXP_ENOS(exp_enos);
/* Check looping state if -i option given */
for (lc = 0; TEST_LOOPING(lc); lc++) {
/* Reset Tst_count in case we are looping. */
Tst_count=0;
/*
* Call vfork(2) to create a child process without
* fully copying the address space of parent.
*/
TEST(vfork());
/* check return code of vfork() */
if ((cpid = TEST_RETURN) == -1) {
TEST_ERROR_LOG(TEST_ERRNO);
tst_resm(TFAIL, "vfork() Failed, errno=%d : %s",
TEST_ERRNO, strerror(TEST_ERRNO));
} else if (cpid == 0) { /* Child process */
/*
* Perform functional verification if test
* executed without (-f) option.
*/
if (STD_FUNCTIONAL_TEST) {
/*
* Check whether the pending signal SIGUSR1
* in the parent is also pending in the child
* process by storing it in a variable.
*/
if (sigpending(&PendSig) == -1) {
tst_resm(TFAIL, "sigpending function "
"failed in child");
_exit(1);
}
/* Check if SIGUSR1 is pending in child */
if (sigismember(&PendSig, SIGUSR1) != 0) {
tst_resm(TFAIL, "SIGUSR1 also pending "
"in child process");
_exit(1);
}
/*
* Exit with normal exit code if everything
* fine
*/
_exit(0);
}
} else { /* parent process */
/*
* Let the parent process wait till child completes
* its execution.
*/
wait(&exit_status);
/* Check for the exit status of child process */
if (WEXITSTATUS(exit_status) == 0) {
tst_resm(TPASS, "Call to vfork() "
"successful");
} else if (WEXITSTATUS(exit_status) == 1) {
tst_resm(TFAIL, \
"Child process exited abnormally");
}
}
Tst_count++; /* incr. TEST_LOOP counter */
} /* End for TEST_LOOPING */
/* Call cleanup() to undo setup done for the test. */
cleanup();
return 0;
} /* End main */
/*--------------------------------------------------------------------*/
int main(int argc, char **argv)
{
/***** BEGINNING OF MAIN. *****/
int pid, npid;
int nsig, exno, nexno, status;
int ret_val = 0;
int core;
void chsig();
#ifdef UCLINUX
const char *msg;
if ((msg = parse_opts(argc, argv, NULL, NULL)) != NULL) {
tst_brkm(TBROK, NULL, "OPTION PARSING ERROR - %s", msg);
}
maybe_run_child(&do_child, "dd", &temp, &sig);
#endif
setup();
//tempdir(); /* move to new directory */ 12/20/2003
blenter();
exno = 1;
if (sigset(SIGCLD, chsig) == SIG_ERR) {
fprintf(temp, "\tsigset failed, errno = %d\n", errno);
fail_exit();
}
for (sig = 1; sig < 14; sig++) {
fflush(temp);
chflag = 0;
pid = FORK_OR_VFORK();
if (pid < 0) {
forkfail();
}
if (pid == 0) {
#ifdef UCLINUX
if (self_exec(argv[0], "dd", temp, sig) < 0) {
tst_resm(TBROK, "self_exec FAILED - "
"terminating test.");
tst_exit();
}
#else
do_child();
#endif
} else {
//fprintf(temp, "Testing signal %d\n", sig);
while (!chflag) /* wait for child */
sleep(1);
kill(pid, sig); /* child should ignroe this sig */
kill(pid, SIGCLD); /* child should exit */
#ifdef BCS
while ((npid = wait(&status)) != pid
|| (npid == -1 && errno == EINTR)) ;
if (npid != pid) {
fprintf(temp,
"wait error: wait returned wrong pid\n");
ret_val = 1;
}
#else
while ((npid = waitpid(pid, &status, 0)) != -1
|| errno == EINTR) ;
#endif
/*
nsig = status & 0177;
core = status & 0200;
nexno = (status & 0xff00) >> 8;
*/
/***** LTP Port *****/
nsig = WTERMSIG(status);
#ifdef WCOREDUMP
core = WCOREDUMP(status);
#endif
nexno = WIFEXITED(status);
/***** ** ** *****/
/* nsig is the signal number returned by wait
it should be 0, except when sig = 9 */
if ((sig == 9) && (nsig != sig)) {
fprintf(temp, "wait error: unexpected signal"
" returned when the signal sent was 9"
" The status of the process is %d \n",
status);
ret_val = 1;
}
if ((sig != 9) && (nsig != 0)) {
fprintf(temp, "wait error: unexpected signal "
"returned, the status of the process is "
"%d \n", status);
ret_val = 1;
}
/* nexno is the exit number returned by wait
it should be 1, except when sig = 9 */
if (sig == 9)
if (nexno != 0) {
fprintf(temp, "signal error: unexpected"
" exit number returned when"
" signal sent was 9, the status"
" of the process is %d \n",
status);
ret_val = 1;
} else;
else if (nexno != 1) {
fprintf(temp, "signal error: unexpected exit "
"number returned,the status of the"
" process is %d\n", status);
ret_val = 1;
}
}
}
if (ret_val)
local_flag = FAILED;
/*--------------------------------------------------------------------*/
anyfail();
tst_exit();
} /******** END OF MAIN. ********/
/**
* trans_create_rdma - Transport method for creating atransport instance
* @client: client instance
* @addr: IP address string
* @args: Mount options string
*/
static int
rdma_create_trans(struct p9_client *client, const char *addr, char *args)
{
int err;
struct p9_rdma_opts opts;
struct p9_trans_rdma *rdma;
struct rdma_conn_param conn_param;
struct ib_qp_init_attr qp_attr;
struct ib_device_attr devattr;
/* Parse the transport specific mount options */
err = parse_opts(args, &opts);
if (err < 0)
return err;
/* Create and initialize the RDMA transport structure */
rdma = alloc_rdma(&opts);
if (!rdma)
return -ENOMEM;
/* Create the RDMA CM ID */
rdma->cm_id = rdma_create_id(p9_cm_event_handler, client, RDMA_PS_TCP);
if (IS_ERR(rdma->cm_id))
goto error;
/* Associate the client with the transport */
client->trans = rdma;
/* Resolve the server's address */
rdma->addr.sin_family = AF_INET;
rdma->addr.sin_addr.s_addr = in_aton(addr);
rdma->addr.sin_port = htons(opts.port);
err = rdma_resolve_addr(rdma->cm_id, NULL,
(struct sockaddr *)&rdma->addr,
rdma->timeout);
if (err)
goto error;
err = wait_for_completion_interruptible(&rdma->cm_done);
if (err || (rdma->state != P9_RDMA_ADDR_RESOLVED))
goto error;
/* Resolve the route to the server */
err = rdma_resolve_route(rdma->cm_id, rdma->timeout);
if (err)
goto error;
err = wait_for_completion_interruptible(&rdma->cm_done);
if (err || (rdma->state != P9_RDMA_ROUTE_RESOLVED))
goto error;
/* Query the device attributes */
err = ib_query_device(rdma->cm_id->device, &devattr);
if (err)
goto error;
/* Create the Completion Queue */
rdma->cq = ib_create_cq(rdma->cm_id->device, cq_comp_handler,
cq_event_handler, client,
opts.sq_depth + opts.rq_depth + 1, 0);
if (IS_ERR(rdma->cq))
goto error;
ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
/* Create the Protection Domain */
rdma->pd = ib_alloc_pd(rdma->cm_id->device);
if (IS_ERR(rdma->pd))
goto error;
/* Cache the DMA lkey in the transport */
rdma->dma_mr = NULL;
if (devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)
rdma->lkey = rdma->cm_id->device->local_dma_lkey;
else {
rdma->dma_mr = ib_get_dma_mr(rdma->pd, IB_ACCESS_LOCAL_WRITE);
if (IS_ERR(rdma->dma_mr))
goto error;
rdma->lkey = rdma->dma_mr->lkey;
}
/* Create the Queue Pair */
memset(&qp_attr, 0, sizeof qp_attr);
qp_attr.event_handler = qp_event_handler;
qp_attr.qp_context = client;
qp_attr.cap.max_send_wr = opts.sq_depth;
qp_attr.cap.max_recv_wr = opts.rq_depth;
qp_attr.cap.max_send_sge = P9_RDMA_SEND_SGE;
qp_attr.cap.max_recv_sge = P9_RDMA_RECV_SGE;
qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
qp_attr.qp_type = IB_QPT_RC;
qp_attr.send_cq = rdma->cq;
qp_attr.recv_cq = rdma->cq;
err = rdma_create_qp(rdma->cm_id, rdma->pd, &qp_attr);
if (err)
goto error;
rdma->qp = rdma->cm_id->qp;
/* Request a connection */
memset(&conn_param, 0, sizeof(conn_param));
conn_param.private_data = NULL;
conn_param.private_data_len = 0;
conn_param.responder_resources = P9_RDMA_IRD;
conn_param.initiator_depth = P9_RDMA_ORD;
err = rdma_connect(rdma->cm_id, &conn_param);
if (err)
goto error;
err = wait_for_completion_interruptible(&rdma->cm_done);
if (err || (rdma->state != P9_RDMA_CONNECTED))
goto error;
client->status = Connected;
return 0;
error:
rdma_destroy_trans(rdma);
return -ENOTCONN;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zhegvdx
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gpu_time /*cpu_time*/;
magmaDoubleComplex *h_A, *h_R, *h_B, *h_S, *h_work;
double *w1, *w2, vl=0, vu=0;
double result[2] = {0};
magma_int_t *iwork;
magma_int_t N, n2, info, il, iu, m1, m2, nb, lwork, liwork;
magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
#if defined(PRECISION_z) || defined(PRECISION_c)
double *rwork;
magma_int_t lrwork;
#endif
//double d_one = 1.;
//double d_ten = 10.;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
double tol = opts.tolerance * lapackf77_dlamch("E");
double tolulp = opts.tolerance * lapackf77_dlamch("P");
if ( opts.check && opts.jobz == MagmaNoVec ) {
fprintf( stderr, "checking results requires vectors; setting jobz=V (option -JV)\n" );
opts.jobz = MagmaVec;
}
printf("using: itype = %d, jobz = %s, uplo = %s, check = %d, fraction = %6.4f\n",
(int) opts.itype, lapack_vec_const(opts.jobz), lapack_uplo_const(opts.uplo),
(int) opts.check, opts.fraction);
printf(" N M GPU Time (sec)\n");
printf("============================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
n2 = N*N;
nb = magma_get_zhetrd_nb(N);
#if defined(PRECISION_z) || defined(PRECISION_c)
lwork = 2*N*nb + N*N;
lrwork = 1 + 5*N +2*N*N;
#else
lwork = 1 + 6*N*nb + 2* N*N;
#endif
liwork = 3 + 5*N;
if ( opts.fraction == 0 ) {
il = N / 10;
iu = N / 5+il;
}
else {
il = 1;
iu = (int) (opts.fraction*N);
if (iu < 1) iu = 1;
}
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, n2 );
TESTING_MALLOC_CPU( h_B, magmaDoubleComplex, n2 );
TESTING_MALLOC_CPU( w1, double, N );
TESTING_MALLOC_CPU( w2, double, N );
TESTING_MALLOC_CPU( iwork, magma_int_t, liwork );
TESTING_MALLOC_PIN( h_R, magmaDoubleComplex, n2 );
TESTING_MALLOC_PIN( h_S, magmaDoubleComplex, n2 );
TESTING_MALLOC_PIN( h_work, magmaDoubleComplex, lwork );
#if defined(PRECISION_z) || defined(PRECISION_c)
TESTING_MALLOC_PIN( rwork, double, lrwork);
#endif
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &n2, h_A );
lapackf77_zlarnv( &ione, ISEED, &n2, h_B );
magma_zmake_hpd( N, h_B, N );
magma_zmake_hermitian( N, h_A, N );
// ==================================================================
// Warmup using MAGMA
// ==================================================================
if(opts.warmup){
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N );
magma_zhegvdx( opts.itype, opts.jobz, MagmaRangeI, opts.uplo,
N, h_R, N, h_S, N, vl, vu, il, iu, &m1, w1,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info );
if (info != 0)
printf("magma_zhegvdx returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N );
gpu_time = magma_wtime();
magma_zhegvdx( opts.itype, opts.jobz, MagmaRangeI, opts.uplo,
N, h_R, N, h_S, N, vl, vu, il, iu, &m1, w1,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info );
gpu_time = magma_wtime() - gpu_time;
if (info != 0)
printf("magma_zhegvdx returned error %d: %s.\n",
(int) info, magma_strerror( info ));
if ( opts.check ) {
/* =====================================================================
Check the results following the LAPACK's [zc]hegvdx routine.
A x = lambda B x is solved
and the following 3 tests computed:
(1) | A Z - B Z D | / ( |A||Z| N ) (itype = 1)
| A B Z - Z D | / ( |A||Z| N ) (itype = 2)
| B A Z - Z D | / ( |A||Z| N ) (itype = 3)
(2) | S(with V) - S(w/o V) | / | S |
=================================================================== */
#if defined(PRECISION_d) || defined(PRECISION_s)
double *rwork = h_work + N*N;
#endif
double temp1, temp2;
result[0] = 1.;
result[0] /= lapackf77_zlanhe("1", lapack_uplo_const(opts.uplo), &N, h_A, &N, rwork);
result[0] /= lapackf77_zlange("1", &N, &m1, h_R, &N, rwork);
if (opts.itype == 1) {
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &N, h_R, &N, &c_zero, h_work, &N);
for(int i=0; i < m1; ++i)
blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_neg_one, h_B, &N, h_R, &N, &c_one, h_work, &N);
result[0] *= lapackf77_zlange("1", &N, &m1, h_work, &N, rwork)/N;
}
else if (opts.itype == 2) {
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_B, &N, h_R, &N, &c_zero, h_work, &N);
for(int i=0; i < m1; ++i)
blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &N, h_work, &N, &c_neg_one, h_R, &N);
result[0] *= lapackf77_zlange("1", &N, &m1, h_R, &N, rwork)/N;
}
else if (opts.itype == 3) {
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &N, h_R, &N, &c_zero, h_work, &N);
for(int i=0; i < m1; ++i)
blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_B, &N, h_work, &N, &c_neg_one, h_R, &N);
result[0] *= lapackf77_zlange("1", &N, &m1, h_R, &N, rwork)/N;
}
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N );
magma_zhegvdx( opts.itype, MagmaNoVec, MagmaRangeI, opts.uplo,
N, h_R, N, h_S, N, vl, vu, il, iu, &m2, w2,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info );
if (info != 0)
printf("magma_zhegvdx returned error %d: %s.\n",
(int) info, magma_strerror( info ));
temp1 = temp2 = 0;
for(int j=0; j < m2; j++) {
temp1 = max(temp1, absv(w1[j]));
temp1 = max(temp1, absv(w2[j]));
temp2 = max(temp2, absv(w1[j]-w2[j]));
}
result[1] = temp2 / (((double)m2)*temp1);
}
/* =====================================================================
Print execution time
=================================================================== */
printf("%5d %5d %7.2f\n",
(int) N, (int) m1, gpu_time);
if ( opts.check ) {
printf("Testing the eigenvalues and eigenvectors for correctness:\n");
if (opts.itype == 1) {
printf("(1) | A Z - B Z D | / (|A| |Z| N) = %8.2e %s\n", result[0], (result[0] < tol ? "ok" : "failed"));
}
else if (opts.itype == 2) {
printf("(1) | A B Z - Z D | / (|A| |Z| N) = %8.2e %s\n", result[0], (result[0] < tol ? "ok" : "failed"));
}
else if (opts.itype == 3) {
printf("(1) | B A Z - Z D | / (|A| |Z| N) = %8.2e %s\n", result[0], (result[0] < tol ? "ok" : "failed"));
}
printf( "(2) | D(w/ Z) - D(w/o Z) | / |D| = %8.2e %s\n\n", result[1], (result[1] < tolulp ? "ok" : "failed"));
status += ! (result[0] < tol && result[1] < tolulp);
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( w1 );
TESTING_FREE_CPU( w2 );
TESTING_FREE_CPU( iwork );
TESTING_FREE_PIN( h_R );
TESTING_FREE_PIN( h_S );
TESTING_FREE_PIN( h_work );
#if defined(PRECISION_z) || defined(PRECISION_c)
TESTING_FREE_PIN( rwork );
#endif
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
