void collect_reduction(const T& val) {
static T total;
static Core cores_in = 0;
DCHECK(mycore() == HOME_CORE);
if (cores_in == 0) {
total = val;
} else {
total = ReduceOp(total, val);
}
cores_in++;
DVLOG(4) << "cores_in: " << cores_in;
if (cores_in == cores()) {
cores_in = 0;
T tmp_total = total;
for (Core c = 0; c < cores(); c++) {
send_heap_message(c, [tmp_total] {
Reduction<T>::result.writeXF(tmp_total);
});
}
}
}
static GlobalAddress<GlobalBag> create(size_t total_capacity) {
auto self = symmetric_global_alloc<GlobalBag>();
auto n = total_capacity / cores()
+ total_capacity % cores();
call_on_all_cores([=]{
new (self.localize()) GlobalBag(self, n);
});
return self;
}
// Export to R
RcppExport SEXP rflann_RadiusSearch(SEXP query_SEXP,
SEXP ref_SEXP,
SEXP radiusSEXP,
SEXP max_neighbourSEXP,
SEXP buildSEXP,
SEXP coresSEXP,
SEXP checksSEXP) {
BEGIN_RCPP
Rcpp::RObject __result;
Rcpp::RNGScope __rngScope;
Rcpp::traits::input_parameter< Rcpp::NumericMatrix >::type
query_(query_SEXP);
Rcpp::traits::input_parameter< Rcpp::NumericMatrix >::type
ref_(ref_SEXP);
Rcpp::traits::input_parameter< double >::type
radius(radiusSEXP);
Rcpp::traits::input_parameter< int >::type
max_neighbour(max_neighbourSEXP);
Rcpp::traits::input_parameter< std::string >::type
build(buildSEXP);
Rcpp::traits::input_parameter< int >::type
cores(coresSEXP);
Rcpp::traits::input_parameter< int >::type
checks(checksSEXP);
__result = Rcpp::wrap(RadiusSearch(query_, ref_, radius,
max_neighbour, build, cores, checks));
return __result;
END_RCPP
}
/// Mark a certain number of things completed. When the global count on all cores goes to 0, all
/// tasks waiting on the GCE will be woken.
///
/// Note: this can be called in a message handler (e.g. remote completes from stolen tasks).
void complete(int64_t dec = 1) {
count -= dec;
DVLOG(4) << "complete (" << count << ") -- gce(" << this << ")";
// out of work here
if (count == 0) { // count[dec -> 0]
// enter cancellable barrier
send_heap_message(master_core, [this] {
cores_out--;
DVLOG(4) << "core entered barrier (cores_out:"<< cores_out <<")";
// if all are in
if (cores_out == 0) { // cores_out[1 -> 0]
CHECK_EQ(count, 0);
// notify everyone to wake
for (Core c = 0; c < cores(); c++) {
send_heap_message(c, [this] {
CHECK_EQ(count, 0);
DVLOG(3) << "broadcast";
broadcast(&cv); // wake anyone who was waiting here
reset(); // reset, now anyone else calling `wait` should fall through
});
}
}
});
}
}
void call_on_all_cores(F work) {
Core origin = mycore();
CompletionEvent ce(cores()-1);
auto lsz = [&ce,origin,work]{};
MessagePool pool(cores()*(sizeof(Message<decltype(lsz)>)));
for (Core c = 0; c < cores(); c++) if (c != mycore()) {
pool.send_message(c, [&ce, origin, work] {
work();
send_heap_message(origin, [&ce]{ ce.complete(); });
});
}
work(); // do my core's work
ce.wait();
}
void on_all_cores(F work) {
CompletionEvent ce(cores());
auto ce_addr = make_global(&ce);
auto lsz = [ce_addr,work]{};
MessagePool pool(cores()*(sizeof(Message<decltype(lsz)>)));
for (Core c = 0; c < cores(); c++) {
pool.send_message(c, [ce_addr, work] {
spawn([ce_addr, work] {
work();
complete(ce_addr);
});
});
}
ce.wait();
}
T reduce( GlobalAddress<P> localizable ) {
CompletionEvent ce(cores() - 1);
T total = Accessor(localizable);
Core origin = mycore();
for (Core c=0; c<cores(); c++) {
if (c != origin) {
send_heap_message(c, [localizable, &ce, &total, origin]{
T val = Accessor(localizable);
send_heap_message(origin, [val,&ce,&total] {
total = ReduceOp(total, val);
ce.complete();
});
});
}
}
ce.wait();
return total;
}
GlobalAddress<T> symmetric_global_alloc() {
static_assert(sizeof(T) % block_size == 0,
"must pad global proxy to multiple of block_size, or use GRAPPA_BLOCK_ALIGNED");
// allocate enough space that we are guaranteed to get one on each core at same location
auto qac = global_alloc<char>(cores()*(sizeof(T)+block_size));
while (qac.core() != MASTER_CORE) qac++;
auto qa = static_cast<GlobalAddress<T>>(qac);
CHECK_EQ(qa, qa.block_min());
CHECK_EQ(qa.core(), MASTER_CORE);
return qa;
}
T reduce(T * global_ptr) {
//NOTE: this is written in a continuation passing
//style to avoid the use of a GCE which async delegates only support
CompletionEvent ce(cores()-1);
// TODO: look into optionally stack-allocating pool storage like in IncoherentAcquirer.
MessagePool pool(cores() * sizeof(Message<std::function<void(T*)>>));
T total = *global_ptr;
Core origin = mycore();
for (Core c=0; c<cores(); c++) {
if (c != origin) {
pool.send_message(c, [global_ptr, &ce, &total, origin]{
T val = *global_ptr;
send_heap_message(origin, [val,&ce,&total] {
total = ReduceOp(total, val);
ce.complete();
});
});
}
}
ce.wait();
return total;
}
//////////////////////////////////
// Main entry function of the GPU cache model
//////////////////////////////////
int main(int argc, char** argv) {
srand(time(0));
std::cout << SPLIT_STRING << std::endl;
message("");
// Flush messages as soon as possible
std::cout.setf(std::ios_base::unitbuf);
// Read the hardware settings from file
Settings hardware = get_settings();
// Print cache statistics
message("Cache configuration:");
std::cout << "### \t Cache size: ~" << hardware.cache_bytes/1024 << "KB" << std::endl;
std::cout << "### \t Line size: " << hardware.line_size << " bytes" << std::endl;
std::cout << "### \t Layout: " << hardware.cache_ways << " ways, " << hardware.cache_sets << " sets" << std::endl;
message("");
// Parse the input argument and make sure that there is only one
if (argc != 3) {
message("Error: provide one argument only (a folder containing input trace files)");
message("");
std::cout << SPLIT_STRING << std::endl;
exit(1);
}
std::string benchname = argv[1];
std::string suitename = argv[2];
// Loop over all found traces in the folder (one trace per kernel)
for (unsigned kernel_id = 0; kernel_id < 20; kernel_id++) {
std::vector<Thread> threads(MAX_THREADS);
for (unsigned t=0; t<MAX_THREADS; t++) {
threads[t] = Thread();
}
// Set the kernelname and include a counter
std::string kernelname;
if (kernel_id < 10) {
kernelname = benchname+"_0"+std::to_string(kernel_id);
}
else {
kernelname = benchname+"_" +std::to_string(kernel_id);
}
// Load a memory access trace from a file
Dim3 blockdim = read_file(threads, kernelname, benchname, suitename);
unsigned blocksize = blockdim.x*blockdim.y*blockdim.z;
// There was not a single trace that could be found - exit with an error
if (blocksize == 0 && kernel_id == 0) {
std::cout << "### Error: could not read file 'output/" << benchname << "/" << kernelname << ".trc'" << std::endl;
message("");
std::cout << SPLIT_STRING << std::endl;
exit(1);
}
// The final tracefile is already processed, exit the loop
if (blocksize == 0) {
break;
}
// Assign threads to warps, threadblocks and GPU cores
message("");
std::cout << "### Assigning threads to warps/blocks/cores...";
unsigned num_blocks = ceil(threads.size()/(float)(blocksize));
unsigned num_warps_per_block = ceil(blocksize/(float)(hardware.warp_size));
std::vector<std::vector<unsigned>> warps(num_warps_per_block*num_blocks);
std::vector<std::vector<unsigned>> blocks(num_blocks);
std::vector<std::vector<unsigned>> cores(hardware.num_cores);
schedule_threads(threads, warps, blocks, cores, hardware, blocksize);
std::cout << "done" << std::endl;
// Model only a single core, modelling multiple cores requires a loop over 'cid'
unsigned cid = 0;
// Compute the number of active blocks on this core
unsigned hardware_max_active_blocks = std::min(hardware.max_active_threads/blocksize, hardware.max_active_blocks);
unsigned active_blocks = std::min((unsigned)cores[cid].size(), hardware_max_active_blocks);
// Start the computation of the reuse distance profile
message("");
std::cout << "### [core " << cid << "]:" << std::endl;
std::cout << "### Running " << active_blocks << " block(s) at a time" << std::endl;
std::cout << "### Calculating the reuse distances";
// Create a Gaussian distribution to model memory latencies
std::random_device random;
std::mt19937 gen(random());
// Compute the reuse distance for 4 different cases
std::vector<map_type<unsigned,unsigned>> distances(NUM_CASES);
for (unsigned runs = 0; runs < NUM_CASES; runs++) {
std::cout << "...";
unsigned sets, ways;
unsigned ml, ms, nml;
unsigned mshr;
// CASE 0 | Normal - full model
sets = hardware.cache_sets;
ways = hardware.cache_ways;
ml = hardware.mem_latency;
ms = hardware.mem_latency_stddev;
nml = NON_MEM_LATENCY;
mshr = hardware.num_mshr;
// CASE 1 | Only 1 set: don't model associativity
if (runs == 1) {
sets = 1;
ways = hardware.cache_ways*hardware.cache_sets;
}
// CASE 2 | Memory latency to 0: don't model latencies
if (runs == 2) {
ml = 0;
ms = 0;
nml = 0;
}
// CASE 3 | MSHR count to infinite: don't model MSHRs
if (runs == 3) {
mshr = INF;
}
// Calculate the reuse distance profile
std::normal_distribution<> distribution(0,ms);
reuse_distance(cores[cid], blocks, warps, threads, distances[runs], active_blocks, hardware,
sets, ways, ml, nml, mshr, gen, distribution);
}
std::cout << "done" << std::endl;
// Process the reuse distance profile to obtain the cache hit/miss rate
message("");
output_miss_rate(distances, kernelname, benchname,suitename, hardware);
// Display the cache hit/miss rate from the output of the verifier (if available)
message("");
verify_miss_rate(kernelname, benchname);
message("");
}
// End of the program
std::cout << SPLIT_STRING << std::endl;
return 0;
}
int Graph::colore(int precision, bool cor)
{
int nCrom = this->n;
std::vector<Vertex> verts(this->n);
std::vector<int> cores(this->n);
std::priority_queue<Vertex, std::vector<Vertex>, std::less<Vertex> > heaps;
int k;
for(k=0; k < precision; k++)
{
for(int j = 0; j < this->n; j++)
{
verts[j].value = rand() % vDeg[j];
verts[j].index = j+1;
heaps.push(verts[j]);
}
Vertex v;
int i;
int l;
bool ok = true;
int crom = 0;
//Vetor de filas, em que a i-ésima fila corresponde aos vértices que possuem a cor i
std::vector< std::deque<int> > C(this->n);
while (!heaps.empty())
{
v = heaps.top();
heaps.pop();
i = 0;
while (true)
{
for(std::deque<int>::iterator it = C[i].begin(); it < C[i].end(); it++)
{
for (std::vector<int>::iterator j = this->vec[v.index-1].begin(); j < this->vec[v.index-1].end(); j++)
{
if(*j == *it)
{
ok = false;
break; //o loop interno
}
}
if (!ok)
{
break;
}
}
if(ok)
{
C[i].insert(C[i].end(),v.index);
if (crom < i+1)
{
crom = i+1;
}
break;
}
i++;
ok = true;
}
}
if (nCrom > crom)
{
nCrom = crom;
if (cor)
{
l = 0;
while (true)
{
if (C[l].empty()) break;
for (std::deque<int>::iterator pos = C[l].begin(); pos < C[l].end(); pos++)
{
cores[*pos-1] = l;
}
l++;
}
}
}
}
if (cor) this->colors = cores;
return nCrom;
}
/// SPMD, must be called on static/file-global object on all cores
/// blocks until reduction is complete
void call_allreduce(T * in_array, size_t nelem) {
// setup everything (block to make sure HOME_CORE is done)
this->array = in_array;
this->nelem = nelem;
size_t n_per_msg = MAX_MESSAGE_SIZE / sizeof(T);
size_t nmsg = nelem / n_per_msg + (nelem % n_per_msg ? 1 : 0);
auto nmsg_total = nmsg*(cores()-1);
CompletionEvent local_ce;
this->ce = &local_ce;
this->ce->enroll( (mycore() == HOME_CORE) ? nmsg_total : nmsg );
barrier();
if (mycore() != HOME_CORE) {
for (size_t k=0; k<nelem; k+=n_per_msg) {
size_t this_nelem = MIN(n_per_msg, nelem-k);
// everyone sends their contribution to HOME_CORE, last one wakes HOME_CORE
send_heap_message(HOME_CORE, [this,k](void * payload, size_t payload_size) {
DCHECK(mycore() == HOME_CORE);
auto in_array = static_cast<T*>(payload);
auto in_n = payload_size/sizeof(T);
auto total = this->array+k;
for (size_t i=0; i<in_n; i++) {
total[i] = ReduceOp(total[i], in_array[i]);
}
DVLOG(3) << "incrementing HOME sem, now at " << ce->get_count();
this->ce->complete();
}, (void*)(in_array+k), sizeof(T)*this_nelem);
}
DVLOG(3) << "about to block for " << nelem << " with sem == " << ce->get_count(); this->ce->wait();
} else {
auto nmsg_total = nmsg*(cores()-1);
// home core waits until woken by last received message from other cores
this->ce->wait();
DVLOG(3) << "woke with sem == " << ce->get_count();
// send total to everyone else and wake them
char msg_buf[(cores()-1)*sizeof(PayloadMessage<std::function<void(decltype(this),size_t)>>)];
MessagePool pool(msg_buf, sizeof(msg_buf));
for (Core c = 0; c < cores(); c++) {
if (c != HOME_CORE) {
// send totals back to all the other cores
size_t n_per_msg = MAX_MESSAGE_SIZE / sizeof(T);
for (size_t k=0; k<nelem; k+=n_per_msg) {
size_t this_nelem = MIN(n_per_msg, nelem-k);
pool.send_message(c, [this,k](void * payload, size_t psz){
auto total_k = static_cast<T*>(payload);
auto in_n = psz / sizeof(T);
for (size_t i=0; i<in_n; i++) {
this->array[k+i] = total_k[i];
}
this->ce->complete();
DVLOG(3) << "incrementing sem, now at " << ce->get_count();
}, this->array+k, sizeof(T)*this_nelem);
}
}
}
// once all messages are sent, HOME_CORE's task continues
}
}
