/*
* Generate all children of the parent
*
* details depend on tree type, node type and shape function
*
*/
void genChildren(Node * parent, void * child_buf, Node * child, StealStack * ss) {
int parentHeight = parent->height;
int numChildren, childType;
ss->maxTreeDepth = max(ss->maxTreeDepth, parent->height);
numChildren = uts_numChildren(parent);
childType = uts_childType(parent);
// record number of children in parent
parent->numChildren = numChildren;
// construct children and push onto stack
if (numChildren > 0) {
int i, j;
child->type = childType;
child->height = parentHeight + 1;
for (i = 0; i < numChildren; i++) {
for (j = 0; j < computeGranularity; j++) {
// TBD: add parent height to spawn
// computeGranularity controls number of rng_spawn calls per node
rng_spawn(parent->state.state, child->state.state, i);
}
ss_put_work(ss, child_buf);
}
} else {
ss->nLeaves++;
}
}
unsigned long long parTreeSearch(int depth, Node *parent, int numChildren)
{
Node *n = (Node *)malloc(numChildren * sizeof(Node));
Node *nodePtr;
int i, j;
unsigned long long subtreesize = 1;
unsigned long long *partialCount = (unsigned long long *)malloc(numChildren * sizeof(unsigned long long));
// Recurse on the children
for (i = 0; i < numChildren; i++) {
nodePtr = &n[i];
nodePtr->height = parent->height + 1;
// The following line is the work (one or more SHA-1 ops)
for (j = 0; j < computeGranularity; j++) {
rng_spawn(parent->state.state, nodePtr->state.state, i);
}
nodePtr->numChildren = uts_numChildren(nodePtr);
hclib_pragma_marker("omp", "task untied firstprivate(i, nodePtr) shared(partialCount)", "pragma221_omp_task");
partialCount[i] = parTreeSearch(depth+1, nodePtr, nodePtr->numChildren);
}
hclib_pragma_marker("omp", "taskwait", "pragma225_omp_taskwait");
for (i = 0; i < numChildren; i++) {
subtreesize += partialCount[i];
}
free(n);
free(partialCount);
return subtreesize;
}
// Notes:
// - Each task receives distinct copy of parent
// - Copy of child is shallow, be careful with `state` member
static long visit(node_t parent)
{
node_t child;
uint64_t *child_descendants = calloc(sizeof(long), parent.num_children);
CILK_C_REDUCER_OPADD(num_descendants, ulong, 0);
uint64_t tmp;
// Spawn children, if any
for (int i = 0; i < parent.num_children; i++) {
child.height = parent.height + 1;
for (int j = 0; j < num_samples; j++) {
rng_spawn(parent.state.state, child.state.state, i);
}
child.num_children = calc_num_children(&child);
child_descendants[i] = _Cilk_spawn visit(child);
}
_Cilk_sync;
CILK_C_REGISTER_REDUCER(num_descendants);
_Cilk_for(int i = 0; i < parent.num_children; i++) {
REDUCER_VIEW(num_descendants) += child_descendants[i];
}
tmp = 1 + REDUCER_VIEW(num_descendants);
CILK_C_UNREGISTER_REDUCER(num_descendants);
return tmp;
}
unsigned long long parTreeSearch(int depth, Node *parent, int numChildren)
{
Node n[numChildren], *nodePtr;
int i, j;
unsigned long long subtreesize = 1, partialCount[numChildren];
// Recurse on the children
for (i = 0; i < numChildren; i++) {
nodePtr = &n[i];
nodePtr->height = parent->height + 1;
// The following line is the work (one or more SHA-1 ops)
for (j = 0; j < computeGranularity; j++) {
rng_spawn(parent->state.state, nodePtr->state.state, i);
}
nodePtr->numChildren = uts_numChildren(nodePtr);
#pragma omp task untied firstprivate(i, nodePtr) shared(partialCount)
partialCount[i] = parTreeSearch(depth+1, nodePtr, nodePtr->numChildren);
}
#pragma omp taskwait
for (i = 0; i < numChildren; i++) {
subtreesize += partialCount[i];
}
return subtreesize;
}
static counter_t _uts_action(void *args, size_t size)
{
int i, j;
struct thread_data *my_data;
struct thread_data temp, input;
my_data = (struct thread_data *)args;
Node n[my_data->numChildren], *nodePtr;
counter_t subtreesize = 1, partialCount[my_data->numChildren];
temp.depth = my_data->depth;
memcpy(&temp.parent, &my_data->parent, sizeof(Node));
temp.numChildren = my_data->numChildren;
//hpx_lco_sema_p (mutex);
//printf("D: %d; child: %d; spawns:%.0f\n", temp.depth, temp.numChildren, spawns_counter++);
//hpx_lco_sema_v_sync (mutex);
/*
printf("\n[Node] height = %d; numChildren = %d\n"
, temp.parent.height
, temp.parent.numChildren);
*/
hpx_addr_t theThread = HPX_HERE;
hpx_addr_t done = hpx_lco_future_new(sizeof(uint64_t));
// Recurse on the children
for (i = 0; i < temp.numChildren; i++) {
nodePtr = &n[i];
nodePtr->height = temp.parent.height + 1;
// The following line is the work (one or more SHA-1 ops)
for (j = 0; j < computeGranularity; j++) {
rng_spawn(temp.parent.state.state, nodePtr->state.state, i);
}
nodePtr->numChildren = uts_numChildren(nodePtr);
input.depth = temp.depth+1;
memcpy(&input.parent, nodePtr, sizeof(Node));
input.numChildren = nodePtr->numChildren;
//partialCount[i] = parTreeSearch(depth+1, nodePtr, nodePtr->numChildren);
hpx_call_sync(theThread, _uts, &partialCount[i], sizeof(partialCount[i]), &input, sizeof(input));
}
for (i = 0; i < temp.numChildren; i++) {
subtreesize += partialCount[i];
}
HPX_THREAD_CONTINUE(subtreesize);
return HPX_SUCCESS;
}
// Notes:
// - Each task receives distinct copy of parent
// - Copy of child is shallow, be careful with `state` member
static long visit(node_t *parent,
int num_children)
{
uint64_t num_descendants = 1;
#ifdef BIG_STACKS
uint64_t child_descendants[num_children];
node_t child_nodes[num_children];
#else
uint64_t *child_descendants;
node_t *child_nodes;
if (num_children > 0) {
child_descendants = calloc(sizeof(uint64_t), num_children);
child_nodes = malloc(sizeof(node_t) * num_children);
}
#endif
// Spawn children, if any
for (int i = 0; i < num_children; i++) {
node_t *child = &child_nodes[i];
child->height = parent->height + 1;
for (int j = 0; j < num_samples; j++) {
rng_spawn(parent->state.state, child->state.state, i);
}
child->num_children = calc_num_children(child);
#pragma omp task untied firstprivate(i, child) shared(child_descendants)
child_descendants[i] = visit(child, child->num_children);
}
#pragma omp taskwait
// #pragma omp parallel for reduction(+:num_descendants)
for (int i = 0; i < num_children; i++) {
num_descendants += child_descendants[i];
}
#ifndef BIG_STACKS
if (num_children > 0) {
free(child_descendants);
free(child_nodes);
}
#endif
return num_descendants;
}
// 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;
}
unsigned long long serTreeSearch(int depth, Node *parent, int numChildren)
{
unsigned long long subtreesize = 1, partialCount[numChildren];
Node n[numChildren];
int i, j;
// Recurse on the children
for (i = 0; i < numChildren; i++) {
n[i].height = parent->height + 1;
// The following line is the work (one or more SHA-1 ops)
for (j = 0; j < computeGranularity; j++) {
rng_spawn(parent->state.state, n[i].state.state, i);
}
partialCount[i] = serTreeSearch(depth+1, &n[i], uts_numChildren(&n[i]));
}
// computing total size
for (i = 0; i < numChildren; i++) subtreesize += partialCount[i];
return subtreesize;
}
counter_t parTreeSearch(int depth, Node *parent, int numChildren)
{
//JK
//Node n[numChildren], *nodePtr;
Node *n, *nodePtr;
int i, j;
counter_t subtreesize = 1;
counter_t *partialCount;
//counter_t partialCount[numChildren];
n = (Node*)malloc(numChildren * sizeof(Node));
partialCount = (counter_t*)malloc(numChildren * sizeof(counter_t));
// Recurse on the children
for (i = 0; i < numChildren; i++) {
nodePtr = &n[i];
nodePtr->height = parent->height + 1;
// The following line is the work (one or more SHA-1 ops)
for (j = 0; j < computeGranularity; j++) {
rng_spawn(parent->state.state, nodePtr->state.state, i);
}
nodePtr->numChildren = uts_numChildren(nodePtr);
#pragma omp task firstprivate(i, nodePtr) shared(partialCount) untied
partialCount[i] = parTreeSearch(depth+1, nodePtr, nodePtr->numChildren);
}
#pragma omp taskwait
for (i = 0; i < numChildren; i++) {
subtreesize += partialCount[i];
}
free(n);
free(partialCount);
return subtreesize;
}
TASK_2(Result, parTreeSearch, int, depth, Node *, parent) {
int numChildren, childType;
counter_t parentHeight = parent->height;
Result r = { depth, 1, 0 };
numChildren = uts_numChildren(parent);
childType = uts_childType(parent);
// record number of children in parent
parent->numChildren = numChildren;
// Recurse on the children
if (numChildren > 0) {
int i, j;
for (i = 0; i < numChildren; i++) {
Node *child = (Node*)alloca(sizeof(Node));
child->type = childType;
child->height = parentHeight + 1;
child->numChildren = -1; // not yet determined
for (j = 0; j < computeGranularity; j++) {
rng_spawn(parent->state.state, child->state.state, i);
}
SPAWN(parTreeSearch, depth+1, child);
}
/* Wait a bit */
struct timespec tim = (struct timespec){0, 100L*numChildren};
nanosleep(&tim, NULL);
for (i = 0; i < numChildren; i++) {
Result c = SYNC(parTreeSearch);
if (c.maxdepth>r.maxdepth) r.maxdepth = c.maxdepth;
r.size += c.size;
r.leaves += c.leaves;
}
} else {
counter_t parTreeSearch(int depth, Node *parent, int numChildren)
{
Node n[numChildren], *nodePtr;
int i, j;
counter_t subtreesize = 1, partialCount[numChildren];
//printf("[p] *** depth = %d ***\n", depth);
//printf("[p] *** height = %d ***\n", parent->height);
//printf("[p] *** numChildren = %d ***\n", parent->numChildren);
// Recurse on the children
for (i = 0; i < numChildren; i++) {
nodePtr = &n[i];
nodePtr->height = parent->height + 1;
// The following line is the work (one or more SHA-1 ops)
for (j = 0; j < computeGranularity; j++) {
rng_spawn(parent->state.state, nodePtr->state.state, i);
}
nodePtr->numChildren = uts_numChildren(nodePtr);
//#pragma omp task firstprivate(i, nodePtr) shared(partialCount) untied
partialCount[i] = parTreeSearch(depth+1, nodePtr, nodePtr->numChildren);
}
//#pragma omp taskwait
for (i = 0; i < numChildren; i++) {
subtreesize += partialCount[i];
}
return subtreesize;
}
/*
* Generate all children of the parent
*
* details depend on tree type, node type and shape function
*
*/
void genChildren(Node * parent, Node * child) {
int parentHeight = parent->height;
int numChildren, childType;
#ifdef THREAD_METADATA
t_metadata[omp_get_thread_num()].ntasks += 1;
#endif
thread_info[hclib::get_current_worker()].n_nodes++;
numChildren = uts_numChildren(parent);
childType = uts_childType(parent);
// record number of children in parent
parent->numChildren = numChildren;
// construct children and push onto stack
if (numChildren > 0) {
int i, j;
child->type = childType;
child->height = parentHeight + 1;
#ifdef UTS_STAT
if (stats) {
child->pp = parent; // pointer to parent
}
#endif
const unsigned char * parent_state = parent->state.state;
unsigned char * child_state = child->state.state;
for (i = 0; i < numChildren; i++) {
for (j = 0; j < computeGranularity; j++) {
// TBD: add parent height to spawn
// computeGranularity controls number of rng_spawn calls per node
rng_spawn(parent_state, child_state, i);
}
Node parent = *child;
int made_available_for_stealing = 0;
if (hclib::get_current_worker() == 0 && n_buffered_steals < N_BUFFERED_STEALS) {
hclib::shmem_set_lock(&steal_buffer_locks[pe]);
if (n_buffered_steals < N_BUFFERED_STEALS) {
steal_buffer[n_buffered_steals++] = parent;
made_available_for_stealing = 1;
}
hclib::shmem_clear_lock(&steal_buffer_locks[pe]);
}
if (!made_available_for_stealing) {
if (parent.height < 9) {
hclib::async([parent] {
Node child;
initNode(&child);
Node tmp = parent;
genChildren(&tmp, &child);
});
} else {
Node child;
initNode(&child);
genChildren(&parent, &child);
}
}
}
} else {
thread_info[hclib::get_current_worker()].n_leaves++;
}
}
