/**
* Distribute the tunecache from node 0 to all other nodes.
*/
static void broadcastTuneCache()
{
#ifdef MULTI_GPU
std::stringstream serialized;
size_t size;
if (comm_rank() == 0) {
serializeTuneCache(serialized);
size = serialized.str().length();
}
comm_broadcast(&size, sizeof(size_t));
if (size > 0) {
if (comm_rank() == 0) {
comm_broadcast(const_cast<char *>(serialized.str().c_str()), size);
} else {
char *serstr = new char[size+1];
comm_broadcast(serstr, size);
serstr[size] ='\0'; // null-terminate
serialized.str(serstr);
deserializeTuneCache(serialized);
delete[] serstr;
}
}
#endif
}
mgasptr_t mgas_all_dmalloc(size_t size, size_t n_dims,
const size_t block_size[],
const size_t n_blocks[])
{
mgas_barrier();
size_t i;
size_t whole_size = 1;
for (i = 0; i < n_dims; i++) {
whole_size *= block_size[i];
whole_size *= n_blocks[i];
}
MGAS_CHECK(size == whole_size);
mgas_proc_t me = globals_get_pid();
gmt_t *gmt = globals_get_gmt();
// allocate addresses
mgasptr_t mp = MGAS_NULL;
if (me == 0) {
mp = gmt_alloc_dist(gmt, size);
MGAS_CHECK(mp != MGAS_NULL);
MGAS_ASSERT(mp >= MGASPTR_MIN);
}
comm_broadcast(&mp, sizeof(mgasptr_t), 0);
dist_t dist;
dist_init(&dist, n_dims, block_size, n_blocks);
gmt_validate_dist(gmt, mp, size, &dist);
mgas_barrier();
return mp;
}
