void Qthread::initialize( int thread_count )
{
// Environment variable: QTHREAD_NUM_SHEPHERDS
// Environment variable: QTHREAD_NUM_WORKERS_PER_SHEP
// Environment variable: QTHREAD_HWPAR
{
char buffer[256];
snprintf(buffer,sizeof(buffer),"QTHREAD_HWPAR=%d",thread_count);
putenv(buffer);
}
const bool ok_init = ( QTHREAD_SUCCESS == qthread_initialize() ) &&
( thread_count == qthread_num_shepherds() * qthread_num_workers_local(NO_SHEPHERD) ) &&
( thread_count == qthread_num_workers() );
bool ok_symmetry = true ;
if ( ok_init ) {
Impl::s_number_shepherds = qthread_num_shepherds();
Impl::s_number_workers_per_shepherd = qthread_num_workers_local(NO_SHEPHERD);
Impl::s_number_workers = Impl::s_number_shepherds * Impl::s_number_workers_per_shepherd ;
for ( int i = 0 ; ok_symmetry && i < Impl::s_number_shepherds ; ++i ) {
ok_symmetry = ( Impl::s_number_workers_per_shepherd == qthread_num_workers_local(i) );
}
}
if ( ! ok_init || ! ok_symmetry ) {
std::ostringstream msg ;
msg << "Kokkos::Qthread::initialize(" << thread_count << ") FAILED" ;
msg << " : qthread_num_shepherds = " << qthread_num_shepherds();
msg << " : qthread_num_workers_per_shepherd = " << qthread_num_workers_local(NO_SHEPHERD);
msg << " : qthread_num_workers = " << qthread_num_workers();
if ( ! ok_symmetry ) {
msg << " : qthread_num_workers_local = {" ;
for ( int i = 0 ; i < Impl::s_number_shepherds ; ++i ) {
msg << " " << qthread_num_workers_local(i) ;
}
msg << " }" ;
}
Impl::s_number_workers = 0 ;
Impl::s_number_shepherds = 0 ;
Impl::s_number_workers_per_shepherd = 0 ;
if ( ok_init ) { qthread_finalize(); }
Kokkos::Impl::throw_runtime_exception( msg.str() );
}
Impl::QthreadExec::resize_worker_scratch( 256 , 256 );
// Init the array for used for arbitrarily sized atomics
Impl::init_lock_array_host_space();
}
// Test that writeFF waits for empty var to be filled, writes, and leaves full.
// Requires that only one worker is running. Basically does:
// 1: empty var
// 1: fork(writeFF)
// 1: yields
// 2: starts runnning
// 2: hits writeFF, and yields since var is empty
// 1: writeEF
// 1: hits readFF on forked task and yield
// 2: running again, finishes writeFF, task returns
// 1: readFF competes, finishes
static void testWriteFFWaits(void)
{
aligned_t ret;
concurrent_t=45;
qthread_empty(&concurrent_t);
assert(qthread_num_workers() == 1);
iprintf("1: Forking writeFF wrapper\n");
qthread_fork_to(writeFF_wrapper, NULL, &ret, qthread_shep());
iprintf("1: Forked, now yielding to 2\n");
qthread_yield();
iprintf("1: Back from yield\n");
// verify that writeFF has not completed
assert(qthread_feb_status(&concurrent_t) == 0);
assert(concurrent_t != 55);
iprintf("1: Writing EF\n");
qthread_writeEF_const(&concurrent_t, 35);
// wait for writeFF wrapper to complete
qthread_readFF(NULL, &ret);
// veify that writeFF completed and that FEB is full
iprintf("1: concurrent_t=%d\n", concurrent_t);
assert(qthread_feb_status(&concurrent_t) == 1);
assert(concurrent_t == 55);
}
int main(int argc,
char *argv[])
{
CHECK_VERBOSE();
assert(qthread_init(1) == 0);
iprintf("%i shepherds...\n", qthread_num_shepherds());
iprintf(" %i threads total\n", qthread_num_workers());
testWriteFFWaits();
}
/******************************************************
* Unbalanced Tree Search v2.1 *
* Based on the implementation available at *
* http://sourceforge.net/projects/uts-benchmark *
******************************************************/
#ifdef HAVE_CONFIG_H
# include "config.h" /* for _GNU_SOURCE */
#endif
#include <assert.h>
#include <stdio.h>
#include <stdlib.h>
#include <limits.h> /* for INT_MAX */
#include <math.h> /* for floor, log, sin */
#include <qthread/qthread.h>
#include <qthread/qtimer.h>
#include "argparsing.h"
#define BRG_RNG // Select RNG
#include "../../utils/rng/rng.h"
#define PRINT_STATS 1
#define MAXNUMCHILDREN 100
typedef enum {
BIN = 0,
GEO,
HYBRID,
BALANCED
} tree_t;
static char *type_names[] = {
"Binomial",
"Geometric",
"Hybrid",
"Balanced"
};
typedef enum {
LINEAR = 0,
EXPDEC,
CYCLIC,
FIXED
} shape_t;
static char *shape_names[] = {
"Linear decrease",
"Exponential decrease",
"Cyclic",
"Fixed branching factor"
};
typedef struct {
int height; // Depth of node in the tree
struct state_t state; // Local RNG state
int num_children;
aligned_t *acc;
aligned_t *dc;
aligned_t expect;
} node_t;
// Default values
static tree_t tree_type = GEO;
static double bf_0 = 4.0;
static int root_seed = 0;
static int num_samples = 1;
static int tree_depth = 6;
static shape_t shape_fn = LINEAR;
static int non_leaf_bf = 4;
static double non_leaf_prob = 15.0 / 64.0;
static double shift_depth = 0.5;
// Tree metrics
static uint64_t tree_height = 0;
static uint64_t num_leaves = 0;
static double normalize(int n)
{/*{{{*/
if (n < 0) {
printf("*** toProb: rand n = %d out of range\n", n);
}
return ((n < 0) ? 0.0 : ((double)n) / (double)INT_MAX);
}/*}}}*/
static int calc_num_children_bin(node_t *parent)
{/*{{{*/
int v = rng_rand(parent->state.state);
double d = normalize(v);
return (d < non_leaf_prob) ? non_leaf_bf : 0;
}/*}}}*/
static int calc_num_children(node_t *parent)
{/*{{{*/
int num_children = 0;
if (parent->height == 0) { num_children = (int)floor(bf_0); } else { num_children = calc_num_children_bin(parent); }
if (parent->height == 0) {
int root_bf = (int)ceil(bf_0);
if (num_children > root_bf) {
printf("*** Number of children truncated from %d to %d\n",
num_children, root_bf);
num_children = root_bf;
}
} else {
if (num_children > MAXNUMCHILDREN) {
printf("*** Number of children truncated from %d to %d\n",
num_children, MAXNUMCHILDREN);
num_children = MAXNUMCHILDREN;
}
}
return num_children;
}/*}}}*/
// Notes:
// - Each task receives distinct copy of parent
// - Copy of child is shallow, be careful with `state` member
static aligned_t visit(void *args_)
{
node_t *parent = (node_t *)args_;
int parent_height = parent->height;
int num_children = parent->num_children;
aligned_t expect = parent->expect;
aligned_t num_descendants[num_children];
aligned_t sum_descendants = 1;
if (num_children != 0) {
node_t child __attribute__((aligned(8)));
aligned_t donec = 0;
// Spawn children, if any
child.height = parent_height + 1;
child.dc = &donec;
child.expect = num_children;
qthread_empty(&donec);
for (int i = 0; i < num_children; i++) {
child.acc = &num_descendants[i];
for (int j = 0; j < num_samples; j++) {
rng_spawn(parent->state.state, child.state.state, i);
}
child.num_children = calc_num_children(&child);
qthread_fork_syncvar_copyargs(visit, &child, sizeof(node_t), NULL);
}
// Wait for children to finish up, accumulate descendants counts
if (donec != expect) qthread_readFF(NULL, &donec);
for (int i = 0; i < num_children; i++) {
sum_descendants += num_descendants[i];
}
}
*parent->acc = sum_descendants;
if (qthread_incr(parent->dc, 1) + 1 == expect) {
qthread_fill(parent->dc);
}
return 0;
}
#ifdef PRINT_STATS
static void print_stats(void)
{/*{{{*/
printf("tree-type %d\ntree-type-name %s\n",
tree_type, type_names[tree_type]);
printf("root-bf %.1f\nroot-seed %d\n",
bf_0, root_seed);
if ((tree_type == GEO) || (tree_type == HYBRID)) {
printf("gen_mx %d\nshape-fn %d\nshape-fn-name %s\n",
tree_depth, shape_fn, shape_names[shape_fn]);
}
if ((tree_type == BIN) || (tree_type == HYBRID)) {
double q = non_leaf_prob;
int m = non_leaf_bf;
double es = (1.0 / (1.0 - q * m));
printf("q %f\nm %d\nE(n) %f\nE(s) %.2f\n",
q, m, q * m, es);
}
if (tree_type == HYBRID) {
printf("root-to-depth %d\n",
(int)ceil(shift_depth * tree_depth));
}
if (tree_type == BALANCED) {
printf("gen_mx %d\n", tree_depth);
printf("expected-num-nodes %llu\nexpected-num-leaves %llu\n",
(unsigned long long)((pow(bf_0, tree_depth + 1) - 1.0) / (bf_0 - 1.0)),
(unsigned long long)pow(bf_0, tree_depth));
}
printf("compute-granularity %d\n", num_samples);
printf("num-sheps %d\n", qthread_num_shepherds());
printf("num-workers %d\n", qthread_num_workers());
printf("\n");
fflush(stdout);
}/*}}}*/
#else /* ifdef PRINT_STATS */
static void print_banner(void)
{/*{{{*/
printf("UTS - Unbalanced Tree Search 2.1 (C/Qthreads)\n");
printf("Tree type:%3d (%s)\n", tree_type, type_names[tree_type]);
printf("Tree shape parameters:\n");
printf(" root branching factor b_0 = %.1f, root seed = %d\n",
bf_0, root_seed);
if ((tree_type == GEO) || (tree_type == HYBRID)) {
printf(" GEO parameters: gen_mx = %d, shape function = %d (%s)\n",
tree_depth, shape_fn, shape_names[shape_fn]);
}
if ((tree_type == BIN) || (tree_type == HYBRID)) {
double q = non_leaf_prob;
int m = non_leaf_bf;
double es = (1.0 / (1.0 - q * m));
printf(" BIN parameters: q = %f, m = %d, E(n) = %f, E(s) = %.2f\n",
q, m, q * m, es);
}
if (tree_type == HYBRID) {
printf(" HYBRID: GEO from root to depth %d, then BIN\n",
(int)ceil(shift_depth * tree_depth));
}
if (tree_type == BALANCED) {
printf(" BALANCED parameters: gen_mx = %d\n", tree_depth);
printf(" Expected size: %llu nodes, %llu leaves\n",
(unsigned long long)((pow(bf_0, tree_depth + 1) - 1.0) / (bf_0 - 1.0)),
(unsigned long long)pow(bf_0, tree_depth));
}
printf("Random number generator: ");
printf("SHA-1 (state size = %ldB)\n", sizeof(struct state_t));
printf("Compute granularity: %d\n", num_samples);
printf("Execution strategy:\n");
printf(" Shepherds: %d\n", qthread_num_shepherds());
printf(" Workers: %d\n", qthread_num_workers());
printf("\n");
fflush(stdout);
}/*}}}*/
int main(int argc,
char *argv[])
{
assert(qthread_initialize() == QTHREAD_SUCCESS);
CHECK_VERBOSE();
NUMARG(numincrs, "NUM_INCRS");
// future_init(128);
iprintf("%i shepherds\n", qthread_num_shepherds());
iprintf("%i threads\n", qthread_num_workers());
qt_loop_balance_sinc(0, numincrs, sum, NULL);
if (threads != numincrs) {
iprintf("threads == %lu, not %lu\n", (unsigned long)threads, (unsigned long)numincrs);
}
assert(threads == numincrs);
return 0;
}
static void print_stats(void)
{/*{{{*/
printf("tree-type %d\ntree-type-name %s\n",
tree_type, type_names[tree_type]);
printf("root-bf %.1f\nroot-seed %d\n",
bf_0, root_seed);
if ((tree_type == GEO) || (tree_type == HYBRID)) {
printf("gen_mx %d\nshape-fn %d\nshape-fn-name %s\n",
tree_depth, shape_fn, shape_names[shape_fn]);
}
if ((tree_type == BIN) || (tree_type == HYBRID)) {
double q = non_leaf_prob;
int m = non_leaf_bf;
double es = (1.0 / (1.0 - q * m));
printf("q %f\nm %d\nE(n) %f\nE(s) %.2f\n",
q, m, q * m, es);
}
if (tree_type == HYBRID) {
printf("root-to-depth %d\n",
(int)ceil(shift_depth * tree_depth));
}
if (tree_type == BALANCED) {
printf("gen_mx %d\n", tree_depth);
printf("expected-num-nodes %llu\nexpected-num-leaves %llu\n",
(unsigned long long)((pow(bf_0, tree_depth + 1) - 1.0) / (bf_0 - 1.0)),
(unsigned long long)pow(bf_0, tree_depth));
}
printf("compute-granularity %d\n", num_samples);
printf("num-sheps %d\n", qthread_num_shepherds());
printf("num-workers %d\n", qthread_num_workers());
printf("\n");
fflush(stdout);
}/*}}}*/
int main(int argc,
char *argv[])
{
aligned_t return_value = 0;
int status, ret;
CHECK_VERBOSE(); // part of the testing harness; toggles iprintf() output
NUMARG(THREADS_ENQUEUED, "THREADS_ENQUEUED");
status = qthread_initialize();
assert(status == QTHREAD_SUCCESS);
iprintf("%i shepherds...\n", qthread_num_shepherds());
iprintf(" %i threads total\n", qthread_num_workers());
iprintf("Creating the queue...\n");
the_queue = qthread_queue_create(QTHREAD_QUEUE_MULTI_JOIN_LENGTH, 0);
assert(the_queue);
iprintf("---------------------------------------------------------\n");
iprintf("\tSINGLE THREAD TEST\n\n");
iprintf("1/4 Spawning thread to be queued...\n");
status = qthread_fork(tobequeued, NULL, &return_value);
assert(status == QTHREAD_SUCCESS);
iprintf("2/4 Waiting for thread to queue itself...\n");
while(qthread_queue_length(the_queue) != 1) qthread_yield();
assert(qthread_readstate(NODE_BUSYNESS) == 1);
iprintf("3/4 Releasing the queue...\n");
qthread_queue_release_all(the_queue);
ret = qthread_readFF(NULL, &return_value);
assert(ret == QTHREAD_SUCCESS);
assert(threads_in == 1);
assert(awoke == 1);
assert(qthread_queue_length(the_queue) == 0);
assert(qthread_readstate(NODE_BUSYNESS) == 1);
iprintf("4/4 Test passed!\n");
iprintf("---------------------------------------------------------\n");
iprintf("\tMULTI THREAD TEST\n\n");
threads_in = 0;
awoke = 0;
aligned_t *retvals = malloc(sizeof(aligned_t) * THREADS_ENQUEUED);
iprintf("1/6 Spawning %u threads to be queued...\n", THREADS_ENQUEUED);
for (int i=0; i<THREADS_ENQUEUED; i++) {
status = qthread_fork(tobequeued, NULL, retvals + i);
assert(status == QTHREAD_SUCCESS);
}
iprintf("2/6 Waiting for %u threads to queue themselves...\n", THREADS_ENQUEUED);
while(qthread_queue_length(the_queue) != THREADS_ENQUEUED) qthread_yield();
assert(threads_in == THREADS_ENQUEUED);
assert(qthread_readstate(NODE_BUSYNESS) == 1);
iprintf("3/6 Releasing a single thread...\n");
qthread_queue_release_one(the_queue);
iprintf("4/6 Waiting for that thread to exit\n");
while (awoke == 0) qthread_yield();
assert(qthread_queue_length(the_queue) == (THREADS_ENQUEUED - 1));
assert(qthread_readstate(NODE_BUSYNESS) == 1);
iprintf("5/6 Releasing the rest of the threads...\n");
qthread_queue_release_all(the_queue);
for (int i=0; i<THREADS_ENQUEUED; i++) {
ret = qthread_readFF(NULL, retvals + i);
assert(ret == QTHREAD_SUCCESS);
}
assert(qthread_queue_length(the_queue) == 0);
assert(qthread_readstate(NODE_BUSYNESS) == 1);
iprintf("6/6 Test passed!\n");
return EXIT_SUCCESS;
}
uint32_t chpl_task_getNumThreads(void)
{
return (uint32_t)qthread_num_workers();
}
int main(int argc, char *argv[])
{
aligned_t *ui_array, *ui_array2;
double *d_array, *d_array2;
size_t len = 1000000;
qtimer_t timer = qtimer_create();
double cumulative_time_qutil = 0.0;
double cumulative_time_libc = 0.0;
int using_doubles = 0;
unsigned long iterations = 10;
qthread_initialize();
CHECK_VERBOSE();
printf("%i threads\n", (int)qthread_num_workers());
NUMARG(len, "TEST_LEN");
NUMARG(iterations, "TEST_ITERATIONS");
NUMARG(using_doubles, "TEST_USING_DOUBLES");
printf("using %s\n", using_doubles ? "doubles" : "aligned_ts");
if (using_doubles) {
d_array = calloc(len, sizeof(double));
printf("array is %s\n", human_readable(len * sizeof(double)));
assert(d_array);
// madvise(d_array,len*sizeof(double), MADV_SEQUENTIAL);
for (unsigned int i = 0; i < len; i++) {
d_array[i] = ((double)random()) / ((double)RAND_MAX) + random();
}
d_array2 = calloc(len, sizeof(double));
assert(d_array2);
// madvise(d_array2,len*sizeof(double), MADV_RANDOM);
iprintf("double array generated...\n");
for (unsigned int i = 0; i < iterations; i++) {
memcpy(d_array2, d_array, len * sizeof(double));
qtimer_start(timer);
qutil_qsort(d_array2, len);
qtimer_stop(timer);
cumulative_time_qutil += qtimer_secs(timer);
iprintf("\t%u: sorting %lu doubles with qutil took: %f seconds\n",
i, (unsigned long)len, qtimer_secs(timer));
}
cumulative_time_qutil /= (double)iterations;
printf("sorting %lu doubles with qutil took: %f seconds (avg)\n",
(unsigned long)len, cumulative_time_qutil);
for (unsigned int i = 0; i < iterations; i++) {
memcpy(d_array2, d_array, len * sizeof(double));
qtimer_start(timer);
qsort(d_array2, len, sizeof(double), dcmp);
qtimer_stop(timer);
cumulative_time_libc += qtimer_secs(timer);
iprintf("\t%u: sorting %lu doubles with libc took: %f seconds\n",
i, (unsigned long)len, qtimer_secs(timer));
}
cumulative_time_libc /= (double)iterations;
printf("sorting %lu doubles with libc took: %f seconds\n",
(unsigned long)len, cumulative_time_libc);
free(d_array);
free(d_array2);
} else {
ui_array = calloc(len, sizeof(aligned_t));
printf("array is %s\n", human_readable(len * sizeof(aligned_t)));
for (unsigned int i = 0; i < len; i++) {
ui_array[i] = random();
}
ui_array2 = calloc(len, sizeof(aligned_t));
iprintf("ui_array generated...\n");
for (int i = 0; i < iterations; i++) {
memcpy(ui_array2, ui_array, len * sizeof(aligned_t));
qtimer_start(timer);
qutil_aligned_qsort(ui_array2, len);
qtimer_stop(timer);
cumulative_time_qutil += qtimer_secs(timer);
}
cumulative_time_qutil /= (double)iterations;
printf("sorting %lu aligned_ts with qutil took: %f seconds\n",
(unsigned long)len, cumulative_time_qutil);
for (int i = 0; i < iterations; i++) {
memcpy(ui_array2, ui_array, len * sizeof(aligned_t));
qtimer_start(timer);
qsort(ui_array2, len, sizeof(double), acmp);
qtimer_stop(timer);
cumulative_time_libc += qtimer_secs(timer);
}
cumulative_time_libc /= (double)iterations;
printf("sorting %lu aligned_ts with libc took: %f seconds (avg)\n",
(unsigned long)len, cumulative_time_libc);
free(ui_array);
free(ui_array2);
}
if (cumulative_time_qutil < cumulative_time_libc) {
printf("qutil with %lu threads provides a %0.2fx speedup.\n", (unsigned long)qthread_num_shepherds(), cumulative_time_libc/cumulative_time_qutil);
} else {
printf("qutil with %lu threads provides a %0.2fx slowdown.\n", (unsigned long)qthread_num_shepherds(), cumulative_time_libc/cumulative_time_qutil);
}
qtimer_destroy(timer);
return 0;
}
static void hazardous_scan(hazard_freelist_t *hfl)
{/*{{{*/
const size_t num_hps = qthread_num_workers() * HAZARD_PTRS_PER_SHEP;
void **plist = MALLOC(sizeof(void *) * (num_hps + hzptr_list_len));
hazard_freelist_t tmpfreelist;
assert(plist);
tmpfreelist.freelist = calloc(freelist_max, sizeof(hazard_freelist_entry_t));
assert(tmpfreelist.freelist);
do {
/* Stage 1: Collect hazardpointers */
{
qthread_shepherd_id_t i;
for (i = 0; i < qthread_num_shepherds(); ++i) {
for (qthread_worker_id_t j = 0; j < qlib->nworkerspershep; ++j) {
if (&(qlib->shepherds[i].workers[j].hazard_free_list) != hfl) {
memcpy(plist + (i * qlib->nworkerspershep * HAZARD_PTRS_PER_SHEP) + (j * HAZARD_PTRS_PER_SHEP),
qlib->shepherds[i].workers[j].hazard_ptrs,
sizeof(void *) * HAZARD_PTRS_PER_SHEP);
} else {
memset(plist + (i * qlib->nworkerspershep * HAZARD_PTRS_PER_SHEP) + (j * HAZARD_PTRS_PER_SHEP),
0,
sizeof(void *) * HAZARD_PTRS_PER_SHEP);
}
}
}
uintptr_t *hzptr_tmp = QTHREAD_CASLOCK_READ(hzptr_list);
while (hzptr_tmp != NULL) {
memcpy(plist + (i * qlib->nworkerspershep * HAZARD_PTRS_PER_SHEP),
hzptr_tmp,
sizeof(uintptr_t) * HAZARD_PTRS_PER_SHEP);
hzptr_tmp = (uintptr_t *)hzptr_tmp[HAZARD_PTRS_PER_SHEP];
}
}
/* Stage 2: free pointers that are not in the set of hazardous pointers */
tmpfreelist.count = 0;
qsort(plist, num_hps, sizeof(void *), void_cmp);
assert(hfl->count == freelist_max);
for (size_t i = 0; i < freelist_max; ++i) {
const uintptr_t ptr = (uintptr_t)hfl->freelist[i].ptr;
if (ptr == 0) { break; }
/* look for this ptr in the plist */
if (binary_search((uintptr_t *)plist, ptr, num_hps)) {
/* if found, cannot free it */
tmpfreelist.freelist[tmpfreelist.count] = hfl->freelist[i];
tmpfreelist.count++;
} else {
/* not found, therefore, we can free it */
hfl->freelist[i].freefunc((void *)ptr);
}
}
if (tmpfreelist.count == freelist_max) {
/* This will ONLY happen under *extremely* heavy contention. */
MACHINE_FENCE;
}
} while (tmpfreelist.count == freelist_max);
assert(tmpfreelist.count < freelist_max);
memcpy(hfl->freelist, tmpfreelist.freelist, tmpfreelist.count * sizeof(hazard_freelist_entry_t));
hfl->count = tmpfreelist.count;
FREE(tmpfreelist.freelist, sizeof(hazard_freelist_entry_t));
FREE(plist, sizeof(void *) * (num_hps + hzptr_list_len));
}/*}}}*/
int main(int argc,
char *argv[])
{
size_t threads, i;
aligned_t *rets;
qtimer_t t;
unsigned int iter, iterations = 10;
double tot = 0.0;
assert(qthread_initialize() == 0);
t = qtimer_create();
CHECK_VERBOSE();
NUMARG(iterations, "ITERATIONS");
threads = qthread_num_workers();
iprintf("%i shepherds...\n", qthread_num_shepherds());
iprintf("%i threads...\n", (int)threads);
initme = calloc(threads, sizeof(aligned_t));
assert(initme);
rets = malloc(threads * sizeof(aligned_t));
assert(rets);
iprintf("Creating a barrier to block %i threads\n", threads);
wait_on_me = qt_barrier_create(threads, REGION_BARRIER, 0); // all my spawnees plus me
assert(wait_on_me);
for (iter = 0; iter < iterations; iter++) {
iprintf("%i: forking the threads\n", iter);
for (i = 1; i < threads; i++) {
void *arg[2] = {wait_on_me, (void*)(intptr_t)i};
qthread_spawn(barrier_thread, arg, sizeof(void*)*2, rets + i, 0, NULL, i, 0);
}
iprintf("%i: done forking the threads, entering the barrier\n", iter);
qtimer_start(t);
qt_barrier_enter(wait_on_me, 0);
qtimer_stop(t);
iprintf("%i: main thread exited barrier in %f seconds\n", iter, qtimer_secs(t));
tot += qtimer_secs(t);
// reset
initme_idx = 1;
// check retvals
for (i = 1; i < threads; i++) {
qthread_readFF(NULL, rets + i);
if (initme[i] != iter + 1) {
iprintf("initme[%i] = %i (should be %i)\n", (int)i,
(int)initme[i], iter + 1);
}
assert(initme[i] == iter + 1);
}
}
iprintf("Average barrier time = %f\n", tot / iterations);
iprintf("Destroying the barrier...\n");
qt_barrier_destroy(wait_on_me);
iprintf("Success!\n");
return 0;
}
void chpl_task_init(void)
{
chpl_bool we_set_worker_unit = false;
int32_t numThreadsPerLocale;
int32_t commMaxThreads;
int32_t hwpar;
size_t callStackSize;
pthread_t initer;
char newenv_stack[100] = { 0 };
char *noWorkSteal;
// Set up available hardware parallelism.
// Experience has shown that we hardly ever win by using more than
// one PU per core, so default to that. If this was explicitly
// set by the user we won't override it, however.
if (getenv("QTHREAD_WORKER_UNIT") == NULL) {
we_set_worker_unit = (getenv("QT_WORKER_UNIT") == NULL);
(void) setenv("QT_WORKER_UNIT", "core", 0);
}
// Determine the thread count. CHPL_RT_NUM_THREADS_PER_LOCALE has
// the highest precedence but we limit it to the number of PUs.
// QTHREAD_HWPAR has the next precedence. We don't impose the
// same limit on it, so it can be used to overload the hardware.
// In either case the number of threads can be no greater than any
// maximum imposed by the comm layer. This limit is imposed
// silently.
numThreadsPerLocale = chpl_task_getenvNumThreadsPerLocale();
commMaxThreads = chpl_comm_getMaxThreads();
hwpar = 0;
if (numThreadsPerLocale != 0) {
int32_t numPUsPerLocale;
hwpar = numThreadsPerLocale;
numPUsPerLocale = chpl_numCoresOnThisLocale();
if (0 < numPUsPerLocale && numPUsPerLocale < hwpar) {
if (2 == verbosity) {
printf("QTHREADS: Reduced numThreadsPerLocale=%d to %d "
"to prevent oversubscription of the system.\n",
hwpar, numPUsPerLocale);
}
// Do not oversubscribe the system, use all available resources.
hwpar = numPUsPerLocale;
}
if (0 < commMaxThreads && commMaxThreads < hwpar) {
hwpar = commMaxThreads;
}
} else {
if (0 < commMaxThreads) {
hwpar = qt_internal_get_env_num("HWPAR", 0, 0);
if (commMaxThreads < hwpar) {
hwpar = commMaxThreads;
}
}
}
if (hwpar > 0) {
char newenv[100];
char *sched;
// Unset relevant Qthreads environment variables. Currently
// QTHREAD_HWPAR has precedence over the QTHREAD_NUM_* ones,
// but that isn't documented and may not be true forever, so
// we unset them all.
qt_internal_unset_envstr("HWPAR");
qt_internal_unset_envstr("NUM_SHEPHERDS");
qt_internal_unset_envstr("NUM_WORKERS_PER_SHEPHERD");
// The current check for scheduler and setting HWPAR or
// NUM_SHEPHERDS/WORKERS_PER_SHEPHERD is just to experiment with
// the performance of different schedulers. This is not production code
// and if it's around after July 2014, yell at Elliot.
sched = getenv("CHPL_QTHREAD_SCHEDULER");
if (sched != NULL && strncmp(sched, "nemesis", 7) == 0) {
// Set environment variable for Qthreads
snprintf(newenv, sizeof(newenv), "%i", (int)hwpar);
setenv("QT_NUM_SHEPHERDS", newenv, 1);
setenv("QT_NUM_WORKERS_PER_SHEPHERD", "1", 1);
// Unset QT_WORKER_UNIT iff we set it.
if (we_set_worker_unit) {
(void) unsetenv("QT_WORKER_UNIT");
}
} else {
// Set environment variable for Qthreads
snprintf(newenv, sizeof(newenv), "%i", (int)hwpar);
setenv("QT_HWPAR", newenv, 1);
}
}
// Precedence (high-to-low):
// 1) Chapel minimum
// 2) QTHREAD_STACK_SIZE
// In practice we never get to #2, because the Chapel minimum is
// always > 0, but we cover that case as a backstop.
callStackSize = chpl_task_getMinCallStackSize();
if (callStackSize <= 0)
callStackSize = 1024 * 1024 * sizeof(size_t);
snprintf(newenv_stack, 99, "%zu", callStackSize);
setenv("QT_STACK_SIZE", newenv_stack, 1);
// Turn on informative Qthreads setting messages with Chapel's verbose flag
if (verbosity == 2) {
setenv("QT_INFO", "1", 1);
}
pthread_create(&initer, NULL, initializer, NULL);
while (chpl_qthread_done_initializing == 0) SPINLOCK_BODY();
// Now that Qthreads is up and running, make sure that the number
// of workers is less than any comm layer limit. This is mainly
// checking that the default thread count without QTHREAD_HWPAR
// being set is within any comm layer limit, because we can't
// determine that default ahead of time. Secondarily, it's a
// sanity check on the thread count versus comm limit logic
// above.
assert(0 == commMaxThreads || qthread_num_workers() < commMaxThreads);
if (blockreport || taskreport) {
if (signal(SIGINT, SIGINT_handler) == SIG_ERR) {
perror("Could not register SIGINT handler");
}
}
// Turn off work stealing if it was configured to be off
noWorkSteal = getenv("CHPL_QTHREAD_NO_WORK_STEALING");
if (noWorkSteal != NULL && strncmp(noWorkSteal, "yes", 3) == 0) {
qthread_steal_disable();
}
}
int main(int argc,
char *argv[])
{
uint64_t total_num_nodes = 0;
qtimer_t timer;
double total_time = 0.0;
CHECK_VERBOSE();
{
unsigned int tmp = (unsigned int)tree_type;
NUMARG(tmp, "UTS_TREE_TYPE");
if (tmp <= BALANCED) {
tree_type = (tree_t)tmp;
} else {
fprintf(stderr, "invalid tree type\n");
return EXIT_FAILURE;
}
tmp = (unsigned int)shape_fn;
NUMARG(tmp, "UTS_SHAPE_FN");
if (tmp <= FIXED) {
shape_fn = (shape_t)tmp;
} else {
fprintf(stderr, "invalid shape function\n");
return EXIT_FAILURE;
}
}
DBLARG(bf_0, "UTS_BF_0");
NUMARG(root_seed, "UTS_ROOT_SEED");
NUMARG(tree_depth, "UTS_TREE_DEPTH");
DBLARG(non_leaf_prob, "UTS_NON_LEAF_PROB");
NUMARG(non_leaf_bf, "UTS_NON_LEAF_NUM");
NUMARG(shift_depth, "UTS_SHIFT_DEPTH");
NUMARG(num_samples, "UTS_NUM_SAMPLES");
// If the operator did not attempt to set a stack size, force
// a reasonable lower bound
if (!getenv("QT_STACK_SIZE") && !getenv("QTHREAD_STACK_SIZE"))
setenv("QT_STACK_SIZE", "32768", 0);
assert(qthread_initialize() == 0);
#ifdef PRINT_STATS
print_stats();
#else
print_banner();
#endif
timer = qtimer_create();
qtimer_start(timer);
node_t root;
root.height = 0;
rng_init(root.state.state, root_seed);
root.num_children = calc_num_children(&root);
aligned_t donecount = 0;
root.dc = &donecount;
qthread_empty(&donecount);
aligned_t tot = 0;
root.acc = &tot;
root.expect = 1;
qthread_fork_syncvar(visit, &root, NULL);
qthread_readFF(NULL, root.dc);
total_num_nodes = tot;
qtimer_stop(timer);
total_time = qtimer_secs(timer);
qtimer_destroy(timer);
#ifdef PRINT_STATS
printf("tree-size %lu\ntree-depth %d\nnum-leaves %llu\nperc-leaves %.2f\n",
(unsigned long)total_num_nodes,
(int)tree_height,
(unsigned long long)num_leaves,
num_leaves / (float)total_num_nodes * 100.0);
printf("exec-time %.3f\ntotal-perf %.0f\npu-perf %.0f\n\n",
total_time,
total_num_nodes / total_time,
total_num_nodes / total_time / qthread_num_workers());
#else
printf("Tree size = %lu, tree depth = %d, num leaves = %llu (%.2f%%)\n",
(unsigned long)total_num_nodes,
(int)tree_height,
(unsigned long long)num_leaves,
num_leaves / (float)total_num_nodes * 100.0);
printf("Wallclock time = %.3f sec, performance = %.0f "
"nodes/sec (%.0f nodes/sec per PE)\n\n",
total_time,
total_num_nodes / total_time,
total_num_nodes / total_time / qthread_num_workers());
#endif /* ifdef PRINT_STATS */
return 0;
}
GLT_func_prefix void glt_subthread_get_num(GLT_subthread *num) {
*num = qthread_num_workers();
}