void operator()(Block* b, const ReduceProxy& rp) const
{
if (rp.round() == 0)
{
// enqueue values to the correct locations
for (size_t i = 0; i < (b->*values).size(); ++i)
{
int to = std::lower_bound((b->*samples).begin(), (b->*samples).end(), (b->*values)[i], cmp) - (b->*samples).begin();
rp.enqueue(rp.out_link().target(to), (b->*values)[i]);
}
(b->*values).clear();
} else
{
dequeue_values((b->*values), rp, false);
std::sort((b->*values).begin(), (b->*values).end(), cmp);
}
}
void operator()(Block* b, const ReduceProxy& srp, const RegularSwapPartners& partners) const
{
int k_in = srp.in_link().size();
int k_out = srp.out_link().size();
std::vector<T> samples;
if (k_in == 0)
{
// draw random samples
for (size_t i = 0; i < num_samples; ++i)
samples.push_back((b->*values)[std::rand() % (b->*values).size()]);
} else
dequeue_values(samples, srp, false);
if (k_out == 0)
{
// pick subsamples that separate quantiles
std::sort(samples.begin(), samples.end(), cmp);
std::vector<T> subsamples(srp.nblocks() - 1);
int step = samples.size() / srp.nblocks(); // NB: subsamples.size() + 1
for (size_t i = 0; i < subsamples.size(); ++i)
subsamples[i] = samples[(i+1)*step];
(b->*dividers).swap(subsamples);
}
else
{
for (int i = 0; i < k_out; ++i)
{
MemoryBuffer& out = srp.outgoing(srp.out_link().target(i));
save(out, &samples[0], samples.size());
}
}
}
static void dequeue_values(std::vector<T>& v, const ReduceProxy& rp, bool skip_self = true)
{
auto log = get_logger();
int k_in = rp.in_link().size();
log->trace("dequeue_values(): gid={}, round={}; v.size()={}", rp.gid(), rp.round(), v.size());
if (detail::is_default< Serialization<T> >::value)
{
// add up sizes
size_t sz = 0;
size_t end = v.size();
for (int i = 0; i < k_in; ++i)
{
log->trace(" incoming size from {}: {}", rp.in_link().target(i).gid, sz);
if (skip_self && rp.in_link().target(i).gid == rp.gid()) continue;
MemoryBuffer& in = rp.incoming(rp.in_link().target(i).gid);
sz += in.size() / sizeof(T);
}
log->trace(" incoming size: {}", sz);
v.resize(end + sz);
for (int i = 0; i < k_in; ++i)
{
if (skip_self && rp.in_link().target(i).gid == rp.gid()) continue;
MemoryBuffer& in = rp.incoming(rp.in_link().target(i).gid);
size_t sz = in.size() / sizeof(T);
T* bg = (T*) &in.buffer[0];
std::copy(bg, bg + sz, &v[end]);
end += sz;
}
} else
{
for (int i = 0; i < k_in; ++i)
{
if (skip_self && rp.in_link().target(i).gid == rp.gid()) continue;
MemoryBuffer& in = rp.incoming(rp.in_link().target(i).gid);
while(in)
{
T x;
diy::load(in, x);
v.emplace_back(std::move(x));
}
}
}
log->trace(" v.size()={}", v.size());
}
void operator()(void* b, const ReduceProxy& srp, const RegularSwapPartners& partners) const
{
int k_in = srp.in_link().size();
int k_out = srp.out_link().size();
if (k_in == 0 && k_out == 0) // special case of a single block
{
ReduceProxy all_srp_out(srp, srp.block(), 0, srp.assigner(), empty_link, all_neighbors_link);
ReduceProxy all_srp_in (srp, srp.block(), 1, srp.assigner(), all_neighbors_link, empty_link);
op(b, all_srp_out);
MemoryBuffer& in_queue = all_srp_in.incoming(all_srp_in.in_link().target(0).gid);
in_queue.swap(all_srp_out.outgoing(all_srp_out.out_link().target(0)));
in_queue.reset();
op(b, all_srp_in);
return;
}
if (k_in == 0) // initial round
{
ReduceProxy all_srp(srp, srp.block(), 0, srp.assigner(), empty_link, all_neighbors_link);
op(b, all_srp);
Master::OutgoingQueues all_queues;
all_queues.swap(*all_srp.outgoing()); // clears out the queues and stores them locally
// enqueue outgoing
int group = all_srp.out_link().size() / k_out;
for (int i = 0; i < k_out; ++i)
{
std::pair<int,int> range(i*group, (i+1)*group);
srp.enqueue(srp.out_link().target(i), range);
for (int j = i*group; j < (i+1)*group; ++j)
{
int from = srp.gid();
int to = all_srp.out_link().target(j).gid;
srp.enqueue(srp.out_link().target(i), std::make_pair(from, to));
srp.enqueue(srp.out_link().target(i), all_queues[all_srp.out_link().target(j)]);
}
}
} else if (k_out == 0) // final round
{
// dequeue incoming + reorder into the correct order
ReduceProxy all_srp(srp, srp.block(), 1, srp.assigner(), all_neighbors_link, empty_link);
Master::IncomingQueues all_incoming;
all_incoming.swap(*srp.incoming());
std::pair<int, int> range; // all the ranges should be the same
for (int i = 0; i < k_in; ++i)
{
int gid_in = srp.in_link().target(i).gid;
MemoryBuffer& in = all_incoming[gid_in];
load(in, range);
while(in)
{
std::pair<int, int> from_to;
load(in, from_to);
load(in, all_srp.incoming(from_to.first));
all_srp.incoming(from_to.first).reset();
}
}
op(b, all_srp);
} else // intermediate round: reshuffle queues
{
// add up buffer sizes
std::vector<size_t> sizes_out(k_out, sizeof(std::pair<int,int>));
std::pair<int, int> range; // all the ranges should be the same
for (int i = 0; i < k_in; ++i)
{
MemoryBuffer& in = srp.incoming(srp.in_link().target(i).gid);
load(in, range);
int group = (range.second - range.first)/k_out;
std::pair<int, int> from_to;
size_t s;
while(in)
{
diy::load(in, from_to);
diy::load(in, s);
int j = (from_to.second - range.first) / group;
sizes_out[j] += s + sizeof(size_t) + sizeof(std::pair<int,int>);
in.skip(s);
}
in.reset();
}
// reserve outgoing buffers of correct size
int group = (range.second - range.first)/k_out;
for (int i = 0; i < k_out; ++i)
{
MemoryBuffer& out = srp.outgoing(srp.out_link().target(i));
out.reserve(sizes_out[i]);
std::pair<int, int> out_range;
out_range.first = range.first + group*i;
out_range.second = range.first + group*(i+1);
save(out, out_range);
}
// re-direct the queues
for (int i = 0; i < k_in; ++i)
{
MemoryBuffer& in = srp.incoming(srp.in_link().target(i).gid);
std::pair<int, int> range;
load(in, range);
std::pair<int, int> from_to;
while(in)
{
load(in, from_to);
int j = (from_to.second - range.first) / group;
MemoryBuffer& out = srp.outgoing(srp.out_link().target(j));
save(out, from_to);
MemoryBuffer::copy(in, out);
}
}
}
}
static void dequeue_values(std::vector<T>& v, const ReduceProxy& rp, bool skip_self = true)
{
int k_in = rp.in_link().size();
//printf("dequeue_values(): gid=%d, round=%d; v.size()=%lu\n", rp.gid(), rp.round(), v.size());
if (detail::is_default< Serialization<T> >::value)
{
// add up sizes
size_t sz = 0;
size_t end = v.size();
for (int i = 0; i < k_in; ++i)
{
//printf(" incoming size from %d: %lu\n", rp.in_link().target(i).gid, sz);
if (skip_self && rp.in_link().target(i).gid == rp.gid()) continue;
MemoryBuffer& in = rp.incoming(rp.in_link().target(i).gid);
sz += in.size() / sizeof(T);
}
//printf(" incoming size: %lu\n", sz);
v.resize(end + sz);
for (int i = 0; i < k_in; ++i)
{
if (skip_self && rp.in_link().target(i).gid == rp.gid()) continue;
MemoryBuffer& in = rp.incoming(rp.in_link().target(i).gid);
size_t sz = in.size() / sizeof(T);
T* bg = (T*) &in.buffer[0];
std::copy(bg, bg + sz, &v[end]);
end += sz;
}
} else
{
for (int i = 0; i < k_in; ++i)
{
if (skip_self && rp.in_link().target(i).gid == rp.gid()) continue;
MemoryBuffer& in = rp.incoming(rp.in_link().target(i).gid);
while(in)
{
T x;
diy::load(in, x);
#if __cplusplus > 199711L // C++11
v.emplace_back(std::move(x));
#else
v.push_back(x);
#endif
}
}
}
//printf(" v.size()=%lu\n", v.size());
}
