static int _delete_handler(const hpx_addr_t * const lcos, size_t n) {
hpx_lco_wait(lcos[2]);
hpx_lco_delete(lcos[2], HPX_NULL);
hpx_lco_delete(lcos[0], HPX_NULL);
hpx_lco_set(lcos[1], 0, NULL, HPX_NULL, HPX_NULL);
return HPX_SUCCESS;
}
/// A utility that tests a certain leaf function through I iterations.
static int _benchmark(char *name, hpx_action_t op, int iters, size_t size) {
int ranks = HPX_LOCALITIES * HPX_THREADS;
hpx_addr_t allreduce = hpx_lco_allreduce_new(ranks, ranks, size, _init, _min);
hpx_addr_t done = hpx_lco_and_new(ranks);
hpx_time_t start = hpx_time_now();
hpx_bcast(_fill_node, HPX_NULL, HPX_NULL, &op, &done, &allreduce, &iters, &size);
hpx_lco_wait(done);
double elapsed = hpx_time_elapsed_ms(start);
hpx_lco_delete(allreduce, HPX_NULL);
hpx_lco_delete(done, HPX_NULL);
printf("%s: %.7f\n", name, elapsed/iters);
return HPX_SUCCESS;
}
/// A utility that tests a certain leaf function through I iterations.
static int _test(hpx_action_t leaf) {
int L = HPX_LOCALITIES;
int n = N * L;
hpx_addr_t allreduce = hpx_lco_allreduce_new(n, n, sizeof(int), _init, _sum);
hpx_addr_t sum = hpx_lco_reduce_new(n, sizeof(int), _init, _sum);
for (int i = 0; i < I; ++i) {
int r;
CHECK( hpx_bcast(_test_bcast, HPX_NULL, HPX_NULL, &allreduce, &sum, &leaf) );
CHECK( hpx_lco_get_reset(sum, sizeof(r), &r) );
test_assert(r == n * HPX_LOCALITIES * (N + 1) * N / 2);
}
hpx_lco_delete(sum, HPX_NULL);
hpx_lco_delete(allreduce, HPX_NULL);
return HPX_SUCCESS;
}
static int _jacobi_main_handler(int n, int nsteps) {
double h = 1.0/n;
// allocate and initialize arrays
hpx_addr_t u = hpx_gas_calloc_local_attr((n+1), BSIZE, 0, HPX_GAS_ATTR_LB);
hpx_addr_t f = hpx_gas_alloc_local((n+1), BSIZE, 0);
hpx_addr_t and = hpx_lco_and_new(n+1);
for (int i = 0; i <= n; ++i) {
double val = i*h;
hpx_gas_memput_lsync(IDX(f,i), &val, sizeof(val), and);
}
hpx_lco_wait(and);
hpx_lco_delete(and, HPX_NULL);
printf("starting jacobi iterations...\n");
hpx_time_t start = hpx_time_now();
jacobi(n, nsteps, u, f);
double elapsed = hpx_time_elapsed_ms(start)/1e3;
// run the solver
printf("n: %d\n", n);
printf("nsteps: %d\n", nsteps);
printf("seconds: %.7f\n", elapsed);
// write the results
if (fname) {
write_solution(n, u, fname);
}
hpx_gas_free(f, HPX_NULL);
hpx_gas_free(u, HPX_NULL);
hpx_exit(HPX_SUCCESS);
}
int _hpx_process_call(hpx_addr_t process, hpx_addr_t addr, hpx_action_t id,
hpx_addr_t result, int n, ...) {
va_list args;
va_start(args, n);
hpx_action_t set = hpx_lco_set_action;
hpx_parcel_t *p = action_new_parcel_va(id, addr, result, set, n, &args);
va_end(args);
if (hpx_thread_current_pid() == hpx_process_getpid(process)) {
hpx_parcel_send_sync(p);
return HPX_SUCCESS;
}
hpx_addr_t sync = hpx_lco_future_new(0);
hpx_parcel_t *q = hpx_parcel_acquire(NULL, parcel_size(p));
q->target = process;
q->action = _proc_call;
q->c_target = sync;
q->c_action = hpx_lco_set_action;
hpx_parcel_set_data(q, p, parcel_size(p));
q->pid = 0;
q->credit = 0;
EVENT_PROCESS_CALL(process, q->pid);
hpx_parcel_send_sync(q);
parcel_delete(p);
hpx_lco_wait(sync);
hpx_lco_delete(sync, HPX_NULL);
return HPX_SUCCESS;
}
// Do sweeps of Jacobi iteration on a 1D Poisson problem
// discretized by n+1 equally spaced mesh points on [0,1].
// u is subject to Dirichlet boundary conditions specified in
// the u[0] and u[n] entries of the initial vector.
void jacobi(int nsteps, int n, hpx_addr_t u, hpx_addr_t f) {
double h = 1.0/n;
double h2 = h*h;
hpx_addr_t utmp = hpx_gas_alloc_local((n+1), BSIZE, 0);
// fill boundary conditions into utmp
hpx_gas_memcpy_sync(IDX(utmp,0), IDX(u,0), sizeof(double));
hpx_gas_memcpy_sync(IDX(utmp,n), IDX(u,n), sizeof(double));
hpx_addr_t and = hpx_lco_and_new(n-1);
for (int sweep = 0; sweep < nsteps; sweep += 2) {
for (int i = 1; i < n; ++i) {
hpx_xcall(IDX(utmp, i), _op, and, i, h2, u, f);
}
hpx_lco_wait_reset(and);
for (int i = 1; i < n; ++i) {
hpx_xcall(IDX(u, i), _op, and, i, h2, utmp, f);
}
hpx_lco_wait_reset(and);
}
hpx_lco_delete(and, HPX_NULL);
hpx_gas_free(utmp, HPX_NULL);
}
static int thread_cont_action_handler(void) {
printf("Starting the Thread continue target and action test\n");
// Start the timer
hpx_time_t t1 = hpx_time_now();
hpx_addr_t *cont_and = calloc(hpx_get_num_ranks(), sizeof(hpx_addr_t));
for (int i = 0; i < hpx_get_num_ranks(); i++) {
cont_and[i] = hpx_lco_and_new(2);
hpx_parcel_t *p = hpx_parcel_acquire(NULL, DATA_SIZE);
hpx_parcel_set_target(p, HPX_THERE(i));
hpx_parcel_set_action(p, _thread_current_cont_target);
hpx_parcel_set_cont_target(p, cont_and[i]);
hpx_parcel_set_cont_action(p, hpx_lco_set_action);
hpx_parcel_send_sync(p);
printf("Started index %d.\n", i);
}
for (int i = 0; i < hpx_get_num_ranks(); i++) {
hpx_lco_wait(cont_and[i]);
printf("Received continuation from %d\n",i);
hpx_lco_delete(cont_and[i], HPX_NULL);
}
free(cont_and);
printf(" Elapsed: %g\n", hpx_time_elapsed_ms(t1));
return HPX_SUCCESS;
}
static int lco_error_handler(void) {
printf("Starting the HPX LCO get all test\n");
hpx_time_t t1 = hpx_time_now();
hpx_addr_t lco = hpx_lco_future_new(0);
hpx_addr_t done = hpx_lco_future_new(0);
hpx_call(HPX_HERE, _errorset, done, &lco, sizeof(lco));
hpx_status_t status = hpx_lco_wait(lco);
printf("status == %d\n", status);
assert(status == HPX_ERROR);
hpx_lco_wait(done);
hpx_lco_delete(lco, HPX_NULL);
hpx_lco_delete(done, HPX_NULL);
printf(" Elapsed: %.7f\n", hpx_time_elapsed_ms(t1)/1e3);
return HPX_SUCCESS;
}
static int _test_recursion_top_handler(void) {
static int DEPTH = 500;
hpx_addr_t and = hpx_lco_and_new(DEPTH);
int e = hpx_xcall(HPX_HERE, _test_recursion, and, DEPTH, and);
if (HPX_SUCCESS == e) {
e = hpx_lco_wait(and);
}
hpx_lco_delete(and, HPX_NULL);
return e;
}
/// Use the join_async operation in the allreduce leaf.
static int
_join_async_leaf_handler(hpx_addr_t allreduce, int i, int j, hpx_addr_t sum) {
int r;
hpx_addr_t f = hpx_lco_future_new(0);
CHECK( hpx_lco_allreduce_join_async(allreduce, i, sizeof(j), &j, &r, f) );
CHECK( hpx_lco_wait(f) );
hpx_lco_delete(f, HPX_NULL);
test_assert(r == HPX_LOCALITIES * N * (N + 1) / 2);
return hpx_call_cc(sum, hpx_lco_set_action, &r, sizeof(r));
}
// Testcase to test hpx_lco_get_all function
static int _getAll_handler(uint32_t *args, size_t size) {
uint32_t n = *args;
if (n < 2) {
return HPX_THREAD_CONTINUE(n);
}
hpx_addr_t peers[] = {
HPX_HERE,
HPX_HERE
};
uint32_t ns[] = {
n - 1,
n - 2
};
hpx_addr_t futures[] = {
hpx_lco_future_new(sizeof(uint32_t)),
hpx_lco_future_new(sizeof(uint32_t))
};
uint32_t ssn[] = {
0,
0
};
void *addrs[] = {
&ssn[0],
&ssn[1]
};
size_t sizes[] = {
sizeof(uint32_t),
sizeof(uint32_t)
};
hpx_call(peers[0], _getAll, futures[0], &ns[0], sizeof(uint32_t));
hpx_call(peers[1], _getAll, futures[1], &ns[1], sizeof(uint32_t));
hpx_lco_get_all(2, futures, sizes, addrs, NULL);
hpx_lco_wait(futures[0]);
hpx_lco_wait(futures[1]);
hpx_addr_t wait = hpx_lco_future_new(0);
hpx_lco_delete_all(2, futures, wait);
hpx_lco_wait(wait);
hpx_lco_delete(wait, HPX_NULL);
uint32_t sn = ssn[0] * ssn[0] + ssn[1] * ssn[1];
return HPX_THREAD_CONTINUE(sn);
}
static int lco_wait_handler(void) {
printf("Starting the LCO wait test.\n");
// allocate and start a timer
const hpx_time_t t1 = hpx_time_now();
const hpx_addr_t termination_lco = hpx_lco_and_new(2 * LCOS_PER_LOCALITY * HPX_LOCALITIES);
hpx_bcast(_spawn, HPX_NULL, HPX_NULL, &termination_lco);
hpx_lco_wait(termination_lco);
hpx_lco_delete(termination_lco, HPX_NULL);
printf(" Elapsed: %g\n", hpx_time_elapsed_ms(t1));
return HPX_SUCCESS;
}
static int _test_try_task_handler(void) {
barrier = sr_barrier_new(HPX_THREADS);
assert(barrier);
hpx_addr_t and = hpx_lco_and_new(HPX_THREADS + 1);
assert(and);
for (int i = 0; i < HPX_THREADS; ++i) {
int e = hpx_call(HPX_HERE, _test_action, and);
assert(e == HPX_SUCCESS);
}
hpx_lco_wait(and);
hpx_lco_delete(and, HPX_NULL);
sync_barrier_delete(barrier);
return HPX_SUCCESS;
}
static int _main_action(int *args, size_t size) {
int n = *args;
printf("seqspawn(%d)\n", n); fflush(stdout);
hpx_addr_t and = hpx_lco_and_new(n);
hpx_time_t now = hpx_time_now();
for (int i = 0; i < n; i++)
hpx_call(HPX_HERE, _nop, and, 0, 0);
hpx_lco_wait(and);
double elapsed = hpx_time_elapsed_ms(now)/1e3;
hpx_lco_delete(and, HPX_NULL);
printf("seconds: %.7f\n", elapsed);
printf("localities: %d\n", HPX_LOCALITIES);
printf("threads: %d\n", HPX_THREADS);
hpx_exit(HPX_SUCCESS);
}
/// Use a synchronous join for the allreduce operation.
static int
_allreduce_join_handler(hpx_addr_t allreduce, int iters, size_t size) {
unsigned char sbuf[size];
unsigned char rbuf[size];
for (int i = 0, e = size; i < e; ++i) {
sbuf[i] = rand();
}
int id = (HPX_LOCALITY_ID * HPX_THREADS) + HPX_THREAD_ID;
hpx_addr_t f = hpx_lco_future_new(0);
for (int i = 0; i < iters; ++i) {
hpx_lco_allreduce_join_async(allreduce, id, size, sbuf, rbuf, f);
hpx_lco_wait_reset(f);
}
hpx_lco_delete(f, HPX_NULL);
return HPX_SUCCESS;
}
int parallel_nqueens(int n, int col, int *hist)
{
hpx_addr_t theThread = HPX_HERE;
struct thread_data td;
//td.lyst = hist;
td.n = n;
td.col = col;
memcpy(td.lyst, hist, MAX_SIZE*sizeof(int));
//printf("thread_data size:%d\n", sizeof(struct thread_data));
mutex = hpx_lco_sema_new(1);
//solve(td.n, td.col, td.lyst);
hpx_addr_t done = hpx_lco_future_new(sizeof(uint64_t));
hpx_call(theThread, _nqueens, done, &td, sizeof(td));
hpx_lco_wait(done);
hpx_lco_delete(done, HPX_NULL);
return HPX_SUCCESS;
}
// Test code -- ThreadCreate
static int thread_create_handler(int *args, size_t size) {
printf("Starting the Threads test\n");
// Start the timer
hpx_time_t t1 = hpx_time_now();
hpx_addr_t addr = hpx_gas_alloc_cyclic(NUM_THREADS, sizeof(initBuffer_t), 0);
// HPX Threads are spawned as a result of hpx_parcel_send() / hpx_parcel_
// sync().
for (int t = 0; t < NUM_THREADS; t++) {
hpx_addr_t done = hpx_lco_and_new(1);
hpx_parcel_t *p = hpx_parcel_acquire(NULL, sizeof(initBuffer_t));
// Fill the buffer
initBuffer_t *init = hpx_parcel_get_data(p);
init->index = t;
strcpy(init->message, "Thread creation test");
// Set the target address and action for the parcel
hpx_parcel_set_target(p, hpx_addr_add(addr, sizeof(initBuffer_t) * t, sizeof(initBuffer_t)));
hpx_parcel_set_action(p, _initData);
// Set the continuation target and action for parcel
hpx_parcel_set_cont_target(p, done);
hpx_parcel_set_cont_action(p, hpx_lco_set_action);
// and send the parcel, this spawns the HPX thread
hpx_parcel_send(p, HPX_NULL);
hpx_lco_wait(done);
hpx_lco_delete(done, HPX_NULL);
}
hpx_gas_free(addr, HPX_NULL);
printf(" Elapsed: %g\n", hpx_time_elapsed_ms(t1));
hpx_exit(HPX_SUCCESS);
}
static int parcel_get_continuation_handler(void) {
printf("Testing parcel contination target and action\n");
hpx_time_t t1 = hpx_time_now();
hpx_addr_t addr = hpx_gas_alloc_cyclic(1, sizeof(uint64_t), sizeof(uint64_t));
hpx_addr_t done = hpx_lco_and_new(1);
hpx_parcel_t *p = hpx_parcel_acquire(NULL, sizeof(uint64_t));
// Get access to the data, and fill it with the necessary data.
uint64_t *result = hpx_parcel_get_data(p);
*result = 1234;
// Set the target address and action for the parcel
hpx_parcel_set_target(p, addr);
hpx_parcel_set_action(p, _get_cont_value);
// Set the continuation target and action for the parcel
hpx_parcel_set_cont_target(p, done);
hpx_parcel_set_cont_action(p, hpx_lco_set_action);
hpx_action_t get_act = hpx_parcel_get_cont_action(p);
assert_msg(get_act == hpx_lco_set_action,
"Error in getting cont action");
assert(hpx_parcel_get_cont_target(p) == done);
// Send the parcel
hpx_parcel_send(p, HPX_NULL);
hpx_lco_wait(done);
hpx_lco_delete(done, HPX_NULL);
hpx_gas_free(addr, HPX_NULL);
printf("Elapsed: %g\n", hpx_time_elapsed_ms(t1));
return HPX_SUCCESS;
}
static int _nqueens_action(void *args, size_t size)
{
int i, j;
struct thread_data *my_data;
my_data = (struct thread_data *) args;
/*
printf("n = %d, col = %d, count = %d\n", my_data->n
, my_data->col
, count);
*/
if (my_data->col == my_data->n) {
hpx_lco_sema_p(mutex);
++count;
/*
printf("\nNo. %d\n-----\n", count);
for (i = 0; i < my_data->n; i++, putchar('\n'))
for(j = 0; j < my_data->n; j++)
putchar(j == my_data->lyst[i] ? 'Q' : ((i + j) & 1) ? ' ' : '.');
*/
hpx_lco_sema_v_sync(mutex);
hpx_thread_exit(HPX_SUCCESS);
//hpx_thread_continue(NULL, 0);
//return HPX_SUCCESS;
}
#define p_attack(i, j) (my_data->lyst[j] == i || abs(my_data->lyst[j] - i) == my_data->col - j)
int dummy=0;
int num_spawns=0;
for(i = 0, j = 0; i < my_data->n; i++) {
for (j = 0; j < my_data->col && !p_attack(i, j); j++);
if (j < my_data->col) {
dummy++;
}
}
//printf("dummy/spawns: %d/%d\n", dummy, my_data->n);
num_spawns = my_data->n - dummy;
bool D_CALL = false;
//printf("num_spawns = %d\n", num_spawns);
if( num_spawns == 0 ) {
num_spawns = 1;
D_CALL = true;
}
//num_spawns = my_data->n;
struct thread_data temp[num_spawns];
hpx_addr_t futures[num_spawns];
hpx_addr_t threads[num_spawns];
int pqs[num_spawns];
size_t p_size[num_spawns];
void *addrs[num_spawns];
for(i = 0; i < num_spawns; i++) {
futures[i] = hpx_lco_future_new(sizeof(int));
threads[i] = HPX_HERE;
pqs[i] = 0;
addrs[i] = &pqs[i];
p_size[i] = sizeof(size_t);
}
int k=0; // counter for hpx data
for(i = 0, j = 0; i < my_data->n; i++) {
for (j = 0; j < my_data->col && !p_attack(i, j); j++);
if (j < my_data->col) {
//printf("[%d] call continue.\n", i);
continue;
}
//printf("[%d] call nqueens %d\n", i, k);
my_data->lyst[my_data->col] = i;
memcpy(temp[k].lyst, my_data->lyst, MAX_SIZE*sizeof(int));
temp[k].n = my_data->n;
temp[k].col = my_data->col+1;
//solve(n, col + 1, hist);
hpx_call(threads[k], _nqueens, futures[k], (void *)&temp[k], sizeof(temp[k]));
k++;
}
if( !D_CALL ) {
hpx_lco_get_all(num_spawns, futures, p_size, addrs, NULL);
for(i = 0; i < num_spawns; i++)
hpx_lco_delete(futures[i], HPX_NULL);
}
return HPX_SUCCESS;
}
static int _main_action(int *args, size_t size) {
hpx_time_t t;
int count;
fprintf(stdout, HEADER);
fprintf(stdout, "# Latency in (ms)\n");
t = hpx_time_now();
hpx_addr_t done = hpx_lco_future_new(0);
fprintf(stdout, "Creation time: %g\n", hpx_time_elapsed_ms(t));
value = 1234;
t = hpx_time_now();
hpx_call(HPX_HERE, _set_value, done, &value, sizeof(value));
fprintf(stdout, "Value set time: %g\n", hpx_time_elapsed_ms(t));
t = hpx_time_now();
hpx_lco_wait(done);
fprintf(stdout, "Wait time: %g\n", hpx_time_elapsed_ms(t));
t = hpx_time_now();
hpx_lco_delete(done, HPX_NULL);
fprintf(stdout, "Deletion time: %g\n", hpx_time_elapsed_ms(t));
fprintf(stdout, "%s\t%*s%*s%*s\n", "# NumReaders " , FIELD_WIDTH,
"Get_Value ", FIELD_WIDTH, " LCO_Getall ", FIELD_WIDTH, "Delete");
for (int i = 0; i < sizeof(num_readers)/sizeof(num_readers[0]); i++) {
fprintf(stdout, "%d\t\t", num_readers[i]);
count = num_readers[i];
int values[count];
void *addrs[count];
size_t sizes[count];
hpx_addr_t futures[count];
for (int j = 0; j < count; j++) {
addrs[j] = &values[j];
sizes[j] = sizeof(int);
futures[j] = hpx_lco_future_new(sizeof(int));
}
t = hpx_time_now();
for (int j = 0; j < count; j++) {
t = hpx_time_now();
hpx_call(HPX_HERE, _get_value, futures[j], NULL, 0);
hpx_lco_wait(futures[j]);
}
fprintf(stdout, "%*g", FIELD_WIDTH, hpx_time_elapsed_ms(t));
t = hpx_time_now();
hpx_lco_get_all(count, futures, sizes, addrs, NULL);
fprintf(stdout, "%*g", FIELD_WIDTH, hpx_time_elapsed_ms(t));
t = hpx_time_now();
for (int j = 0; j < count; j++)
hpx_lco_delete(futures[j], HPX_NULL);
fprintf(stdout, "%*g\n", FIELD_WIDTH, hpx_time_elapsed_ms(t));
}
hpx_exit(HPX_SUCCESS);
}
