mask_type decode_mapping(hwloc_topology const& t,
mapping_type& m, std::vector<mask_type>& affinities,
std::size_t thread_index, error_code& ec)
{
std::size_t size = affinities.size();
mask_type mask;
switch (m[0].type_) {
case spec_type::socket:
// requested top level is a socket
mask = decode_mapping_socket(t, m, size, thread_index, ec);
break;
case spec_type::numanode:
// requested top level is a NUMA node
mask = decode_mapping_numanode(t, m, size, thread_index, ec);
break;
case spec_type::unknown:
// no top level is requested
mask = decode_mapping0_unknown(t, m, size,
t.get_machine_affinity_mask(), 0, thread_index, ec);
break;
default:
HPX_THROWS_IF(ec, bad_parameter, "decode_mapping",
boost::str(boost::format("unexpected specification type at "
"index zero: %x (%s)") %
static_cast<unsigned>(m[0].type_) %
spec_type::type_name(m[0].type_)));
return mask_type();
}
return mask;
}
mask_type decode_mapping_core(hwloc_topology const& t,
mapping_type& m, std::size_t size, mask_type mask,
std::size_t core_base_index, std::size_t thread_index, error_code& ec)
{
bounds_type b = extract_bounds(m[1], size, ec);
if (ec) return mask_type();
// We have to account for the thread index at this level if there are
// no specifications related to processing units.
std::size_t index = std::size_t(-1);
if (m[2].type_ == spec_type::unknown && b.size() > 1)
index = thread_index;
mask_type core_mask = mask_type();
resize(core_mask, size);
std::size_t core_index = 0;
for (bounds_type::const_iterator it = b.begin(); it != b.end();
++it, ++core_index)
{
if (index == std::size_t(-1) || core_index == index)
{
core_mask |= t.init_core_affinity_mask_from_core(
std::size_t(*it+core_base_index), mask_type());
}
}
core_base_index += std::size_t(*b.begin());
if (thread_index != std::size_t(-1) && b.size() > 1)
core_base_index += thread_index;
std::size_t base_index = 0;
for (std::size_t i = 0; i != core_base_index; ++i)
base_index += t.get_number_of_core_pus(i);
return decode_mapping1_unknown(t, m, size, mask & core_mask,
base_index, thread_index, ec);
}
mask_type decode_mapping_pu(hwloc_topology const& t,
mapping_type& m, std::size_t size, mask_type mask,
std::size_t pu_base_index, std::size_t thread_index, error_code& ec)
{
bounds_type b = extract_bounds(m[2], size, ec);
if (ec) return mask_type();
std::size_t index = std::size_t(-1);
if (b.size() > 1)
index = thread_index;
mask_type pu_mask = mask_type();
resize(pu_mask, size);
std::size_t pu_index = 0;
for (bounds_type::const_iterator it = b.begin(); it != b.end(); ++it, ++pu_index)
{
if (index == std::size_t(-1) || pu_index == index)
pu_mask |= t.init_thread_affinity_mask(std::size_t(*it+pu_base_index));
}
return mask & pu_mask;
}
mask_type decode_mapping_numanode(hwloc_topology const& t,
mapping_type& m, std::size_t size, std::size_t thread_index,
error_code& ec)
{
bounds_type b = extract_bounds(m[0], size, ec);
if (ec) return mask_type();
std::size_t index = std::size_t(-1);
if (m[1].type_ == spec_type::unknown &&
m[2].type_ == spec_type::unknown &&
b.size() > 1)
{
index = thread_index;
}
mask_type mask = mask_type();
resize(mask, size);
std::size_t node_index = 0;
for (bounds_type::const_iterator it = b.begin(); it != b.end();
++it, ++node_index)
{
if (index == std::size_t(-1) || node_index == index)
mask |= t.init_numa_node_affinity_mask_from_numa_node(std::size_t(*it));
}
std::size_t node_base_index = std::size_t(*b.begin());
if (thread_index != std::size_t(-1) && b.size() > 1)
node_base_index += thread_index;
std::size_t base_index = 0;
for (std::size_t i = 0; i != node_base_index; ++i)
base_index += t.get_number_of_numa_node_cores(i);
return decode_mapping0_unknown(t, m, size, mask, base_index,
thread_index, ec);
}
namespace hpx { namespace threads
{
mask_type noop_topology::empty_mask =
mask_type(noop_topology::hardware_concurrency());
}}
void on_start_thread(std::size_t num_thread)
{
if (nullptr == queues_[num_thread])
{
queues_[num_thread] =
new thread_queue_type(max_queue_thread_count_);
if (num_thread < high_priority_queues_.size())
{
high_priority_queues_[num_thread] =
new thread_queue_type(max_queue_thread_count_);
}
}
// forward this call to all queues etc.
if (num_thread < high_priority_queues_.size())
high_priority_queues_[num_thread]->on_start_thread(num_thread);
if (num_thread == queues_.size()-1)
low_priority_queue_.on_start_thread(num_thread);
queues_[num_thread]->on_start_thread(num_thread);
std::size_t num_threads = queues_.size();
// get numa domain masks of all queues...
std::vector<mask_type> numa_masks(num_threads);
std::vector<mask_type> core_masks(num_threads);
for (std::size_t i = 0; i != num_threads; ++i)
{
std::size_t num_pu = get_pu_num(i);
numa_masks[i] =
topology_.get_numa_node_affinity_mask(num_pu, numa_sensitive_ != 0);
core_masks[i] =
topology_.get_core_affinity_mask(num_pu, numa_sensitive_ != 0);
}
// iterate over the number of threads again to determine where to
// steal from
std::ptrdiff_t radius =
static_cast<std::ptrdiff_t>((num_threads / 2.0) + 0.5);
victim_threads_[num_thread].reserve(num_threads);
std::size_t num_pu = get_pu_num(num_thread);
mask_cref_type pu_mask =
topology_.get_thread_affinity_mask(num_pu, numa_sensitive_ != 0);
mask_cref_type numa_mask = numa_masks[num_thread];
mask_cref_type core_mask = core_masks[num_thread];
// we allow the thread on the boundary of the NUMA domain to steal
mask_type first_mask = mask_type();
resize(first_mask, mask_size(pu_mask));
std::size_t first = find_first(numa_mask);
if (first != std::size_t(-1))
set(first_mask, first);
else
first_mask = pu_mask;
auto iterate = [&](hpx::util::function_nonser<bool(std::size_t)> f)
{
// check our neighbors in a radial fashion (left and right
// alternating, increasing distance each iteration)
int i = 1;
for (/**/; i < radius; ++i)
{
std::ptrdiff_t left =
(static_cast<std::ptrdiff_t>(num_thread) - i) %
static_cast<std::ptrdiff_t>(num_threads);
if (left < 0)
left = num_threads + left;
if (f(std::size_t(left)))
{
victim_threads_[num_thread].push_back(
static_cast<std::size_t>(left));
}
std::size_t right = (num_thread + i) % num_threads;
if (f(right))
{
victim_threads_[num_thread].push_back(right);
}
}
if ((num_threads % 2) == 0)
{
std::size_t right = (num_thread + i) % num_threads;
if (f(right))
{
victim_threads_[num_thread].push_back(right);
}
}
};
// check for threads which share the same core...
iterate(
[&](std::size_t other_num_thread)
{
return any(core_mask & core_masks[other_num_thread]);
}
);
// check for threads which share the same numa domain...
iterate(
[&](std::size_t other_num_thread)
{
return
!any(core_mask & core_masks[other_num_thread])
&& any(numa_mask & numa_masks[other_num_thread]);
}
);
// check for the rest and if we are numa aware
if (numa_sensitive_ != 2 && any(first_mask & pu_mask))
{
iterate(
[&](std::size_t other_num_thread)
{
return !any(numa_mask & numa_masks[other_num_thread]);
}
);
}
}