void QthreadExec::resize_worker_scratch( const int reduce_size , const int shared_size )
{
const int exec_all_reduce_alloc = align_alloc( reduce_size );
const int shepherd_scan_alloc = align_alloc( 8 );
const int shepherd_shared_end = exec_all_reduce_alloc + shepherd_scan_alloc + align_alloc( shared_size );
if ( s_worker_reduce_end < exec_all_reduce_alloc ||
s_worker_shared_end < shepherd_shared_end ) {
// Clear current worker memory before allocating new worker memory
clear_workers();
// Increase the buffers to an aligned allocation
s_worker_reduce_end = exec_all_reduce_alloc ;
s_worker_shared_begin = exec_all_reduce_alloc + shepherd_scan_alloc ;
s_worker_shared_end = shepherd_shared_end ;
// Need to query which shepherd this main 'process' is running...
// Have each worker resize its memory for proper first-touch
for ( int jshep = 0 ; jshep < s_number_shepherds ; ++jshep ) {
for ( int i = jshep ? 0 : 1 ; i < s_number_workers_per_shepherd ; ++i ) {
// Unit tests hang with this call:
//
// qthread_fork_to_local_priority( driver_resize_workers , NULL , NULL , jshep );
//
qthread_fork_to( driver_resize_worker_scratch , NULL , NULL , jshep );
}}
driver_resize_worker_scratch( NULL );
// Verify all workers allocated
bool ok = true ;
for ( int iwork = 0 ; ok && iwork < s_number_workers ; ++iwork ) { ok = 0 != s_exec[iwork] ; }
if ( ! ok ) {
std::ostringstream msg ;
msg << "Kokkos::Impl::QthreadExec::resize : FAILED for workers {" ;
for ( int iwork = 0 ; iwork < s_number_workers ; ++iwork ) {
if ( 0 == s_exec[iwork] ) { msg << " " << ( s_number_workers - ( iwork + 1 ) ); }
}
msg << " }" ;
Kokkos::Impl::throw_runtime_exception( msg.str() );
}
}
}
void *mcache_alloc(mcache_t mc){
struct mem_cache *m = mc;
void *ptr;
ASSERT(m != NULL);
ptr = align_alloc(m->mc_size, m->mc_align);
if(ptr != NULL){
atomic_inc(&m->mc_allocs);
}
return ptr;
}
int main(int argc, char *argv[])
{
int nchann, nsample, niter;
int k, opt, ns;
hfilter filt = NULL;
void *buffin, *buffout;
int filein = -1, fileout = -1;
size_t samsize;
uint32_t nchan32;
uint32_t type;
uint32_t r = 4;
int retval = 1;
int keepfiles = 0;
// Process command-line options
nchann = NCHANN;
nsample = NSAMPLE;
niter = NITER;
type = TYPE_DEF;
while ((opt = getopt(argc, argv, "hc:s:i:t:r:k:")) != -1) {
switch (opt) {
case 'c':
nchann = atoi(optarg);
break;
case 's':
nsample = atoi(optarg);
break;
case 'i':
niter = atoi(optarg);
break;
case 'r':
r = atoi(optarg);
break;
case 't':
type = atoi(optarg);
break;
case 'k':
keepfiles = atoi(optarg);
break;
case 'h':
default: /* '?' */
fprintf(stderr, "Usage: %s [-c numchannel] [-s numsample] [-i numiteration] "
"[-t (0 for float/1 for double)]\n",
argv[0]);
exit(EXIT_FAILURE);
}
}
printf("\tnumber of channels: %i \t\tlength of batch: %i\n", nchann, nsample);
samsize = sizeof_data(type) * nchann;
nchan32 = nchann;
// Allocate buffers
buffin = align_alloc(16, nsample*samsize);
buffout = align_alloc(16, nsample*samsize);
if (!buffin || !buffout) {
fprintf(stderr, "buffer allocation failed\n");
goto out;
}
// Open files for writing
filein = open(infilename, O_WRONLY|O_CREAT, S_IRWXU);
fileout = open(outfilename, O_WRONLY|O_CREAT, S_IRWXU);
if ((filein < 0) || (fileout < 0)) {
fprintf(stderr, "File opening failed\n");
goto out;
}
// write filter params on fileout
if (write(filein, &type, sizeof(type)) == -1 ||
write(filein, &nchan32, sizeof(nchan32)) == -1 ||
write(fileout, &type, sizeof(type)) == -1 ||
write(fileout, &nchan32, sizeof(nchan32)) == -1 ||
write(fileout, &r, sizeof(r)) == -1)
goto out;
// create filters
filt = rtf_create_downsampler(nchann, type, r);
if (!filt) {
fprintf(stderr,"Creation of filter failed\n");
goto out;
}
// Filter chunks of data and write input and output on files
for (k=0; k<niter; k++) {
set_signals(nchann, nsample, type, buffin);
ns = rtf_filter(filt, buffin, buffout, nsample);
if ( write(filein, buffin, nsample*samsize) == -1
|| write(fileout, buffout, ns*samsize) == -1 ) {
fprintf(stderr,"Error while writing file\n");
break;
}
}
retval = 0;
out:
rtf_destroy_filter(filt);
align_free(buffin);
align_free(buffout);
if (filein != -1)
close(filein);
if (fileout != -1)
close(fileout);
#if HAVE_MATLAB
if (retval == 0)
retval = compare_results(0.02);
#endif
if (!keepfiles) {
unlink(infilename);
unlink(outfilename);
}
return retval;
}
void QthreadExec::resize_worker_scratch( const int reduce_size , const int shared_size )
{
const int exec_all_reduce_alloc = align_alloc( reduce_size );
const int shepherd_scan_alloc = align_alloc( 8 );
const int shepherd_shared_end = exec_all_reduce_alloc + shepherd_scan_alloc + align_alloc( shared_size );
if ( s_worker_reduce_end < exec_all_reduce_alloc ||
s_worker_shared_end < shepherd_shared_end ) {
/*
fprintf( stdout , "QthreadExec::resize\n");
fflush(stdout);
*/
// Clear current worker memory before allocating new worker memory
clear_workers();
// Increase the buffers to an aligned allocation
s_worker_reduce_end = exec_all_reduce_alloc ;
s_worker_shared_begin = exec_all_reduce_alloc + shepherd_scan_alloc ;
s_worker_shared_end = shepherd_shared_end ;
// Need to query which shepherd this main 'process' is running...
const int main_shep = qthread_shep();
// Have each worker resize its memory for proper first-touch
#if 1
for ( int jshep = 0 ; jshep < s_number_shepherds ; ++jshep ) {
for ( int i = jshep != main_shep ? 0 : 1 ; i < s_number_workers_per_shepherd ; ++i ) {
qthread_fork_to( driver_resize_worker_scratch , NULL , NULL , jshep );
}}
#else
// If this function is used before the 'qthread.task_policy' unit test
// the 'qthread.task_policy' unit test fails with a seg-fault within libqthread.so.
for ( int jshep = 0 ; jshep < s_number_shepherds ; ++jshep ) {
const int num_clone = jshep != main_shep ? s_number_workers_per_shepherd : s_number_workers_per_shepherd - 1 ;
if ( num_clone ) {
const int ret = qthread_fork_clones_to_local_priority
( driver_resize_worker_scratch /* function */
, NULL /* function data block */
, NULL /* pointer to return value feb */
, jshep /* shepherd number */
, num_clone - 1 /* number of instances - 1 */
);
assert(ret == QTHREAD_SUCCESS);
}
}
#endif
driver_resize_worker_scratch( NULL );
// Verify all workers allocated
bool ok = true ;
for ( int iwork = 0 ; ok && iwork < s_number_workers ; ++iwork ) { ok = 0 != s_exec[iwork] ; }
if ( ! ok ) {
std::ostringstream msg ;
msg << "Kokkos::Impl::QthreadExec::resize : FAILED for workers {" ;
for ( int iwork = 0 ; iwork < s_number_workers ; ++iwork ) {
if ( 0 == s_exec[iwork] ) { msg << " " << ( s_number_workers - ( iwork + 1 ) ); }
}
msg << " }" ;
Kokkos::Impl::throw_runtime_exception( msg.str() );
}
}
}
int main(int argc, char *argv[])
{
int nchann, nsample, niter, filtorder;
int k, opt;
hfilter filt = NULL;
void *buffin, *buffout;
int filein = -1, fileout = -1;
float fc = FC_DEF;
size_t buffinsize, buffoutsize;
uint32_t nchan32, ntotsample32;
int32_t datintype, datouttype;
int retval = 1;
int keepfiles = 0;
// Process command-line options
nchann = NCHANN;
nsample = NSAMPLE;
niter = NITER;
filtorder = FILTORDER;
datouttype = datintype = TYPE_DEF;
while ((opt = getopt(argc, argv, "hc:s:i:o:f:d:p:k:")) != -1) {
switch (opt) {
case 'c':
nchann = atoi(optarg);
break;
case 's':
nsample = atoi(optarg);
break;
case 'i':
niter = atoi(optarg);
break;
case 'o':
filtorder = atoi(optarg);
break;
case 'f':
fc = atof(optarg);
break;
case 'd':
datintype = atoi(optarg);
break;
case 'p':
ptype = atoi(optarg);
break;
case 'k':
keepfiles = atoi(optarg);
break;
case 'h':
default: /* '?' */
fprintf(stderr, "Usage: %s [-c numchannel] [-s numsample] [-i numiteration] "
"[-o filterorder] [-f cutoff_freq] [-d (0 for float/1 for double)]\n",
argv[0]);
exit(EXIT_FAILURE);
}
}
printf("\tfilter order: %i \tnumber of channels: %i \t\tlength of batch: %i\n",filtorder, nchann, nsample);
set_param(ptype);
datouttype = datintype;
if (ptype & RTF_COMPLEX_MASK)
datouttype |= RTF_COMPLEX_MASK;
buffinsize = sizeof_data(datintype) * nchann * nsample;
buffoutsize = sizeof_data(datouttype) * nchann * nsample;
nchan32 = nchann;
ntotsample32 = nsample*niter;
// Allocate buffers
buffin = align_alloc(16, buffinsize);
buffout = align_alloc(16, buffoutsize);
if (!buffin || !buffout) {
fprintf(stderr, "buffer allocation failed\n");
goto out;
}
// Open files for writing
filein = open(infilename, O_WRONLY|O_CREAT, S_IRWXU);
fileout = open(outfilename, O_WRONLY|O_CREAT, S_IRWXU);
if ((filein < 0) || (fileout < 0)) {
fprintf(stderr, "File opening failed\n");
goto out;
}
fprintf(stdout, "datin=%i datout=%i ptype=%i\n", datintype, datouttype, ptype );
// write filter params on fileout
if (write(fileout, &ptype, sizeof(ptype)) == -1 ||
write(fileout, &numlen, sizeof(numlen)) == -1 ||
write(fileout, num, numlen*sizeof_data(ptype)) == -1 ||
write(fileout, &denumlen, sizeof(denumlen)) == -1 ||
write(fileout, denum, denumlen*sizeof_data(ptype)) == -1 ||
write(filein, &datintype, sizeof(datintype)) == -1 ||
write(filein, &nchan32, sizeof(nchan32)) == -1 ||
write(filein, &ntotsample32, sizeof(ntotsample32)) == -1 ||
write(fileout, &datouttype, sizeof(datouttype)) == -1 ||
write(fileout, &nchan32, sizeof(nchan32)) == -1 ||
write(fileout, &ntotsample32, sizeof(ntotsample32)) == -1)
goto out;
// set signals (ramps)
set_signals(nchann, nsample, datintype, buffin);
// create filters
filt = rtf_create_filter(nchann, datintype, numlen, num, denumlen, denum, ptype);
if (!filt) {
fprintf(stderr,"Creation of filter failed\n");
goto out;
}
// Test the type
if (datintype != rtf_get_type(filt, 1)
|| datouttype != rtf_get_type(filt, 0)) {
fprintf(stderr, "Unexpected data type\n"
"expected: in=%i out=%i\n"
"returned: in=%i out=%i\n",
datintype, datouttype,
rtf_get_type(filt, 1),
rtf_get_type(filt, 0));
goto out;
}
// Filter chunks of data and write input and output on files
for (k=0; k<niter; k++) {
rtf_filter(filt, buffin, buffout, nsample);
if ( write(filein, buffin, buffinsize) == -1
|| write(fileout, buffout, buffoutsize) == -1 ) {
fprintf(stderr,"Error while writing file\n");
break;
}
}
retval = 0;
out:
rtf_destroy_filter(filt);
align_free(buffin);
align_free(buffout);
if (filein != -1)
close(filein);
if (fileout != -1)
close(fileout);
#if HAVE_MATLAB
if (retval == 0)
retval = compare_results( (datintype & RTF_PRECISION_MASK) ? 1e-12 : 1e-4);
#endif
if (!keepfiles) {
unlink(infilename);
unlink(outfilename);
}
return retval;
}
