__declspec(dllexport) int __cdecl GetPhysicalProcessorCount()
{
ULONG p;
if (GetNumaHighestNodeNumber(&p))
return (int)p + 1;
else
return 1;
}
mask_type init_thread_affinity_mask(
std::size_t num_thread
, bool numa_sensitive
)
{ // {{{
std::size_t num_of_cores = hardware_concurrency();
std::size_t affinity = num_thread % num_of_cores;
ULONG numa_nodes = 1;
if (GetNumaHighestNodeNumber(&numa_nodes))
++numa_nodes;
std::size_t num_of_cores_per_numa_node = num_of_cores / numa_nodes;
ULONGLONG node_affinity_mask = 0;
ULONGLONG mask = 0x01LL;
if (numa_sensitive) {
UCHAR numa_node = UCHAR(affinity % numa_nodes);
if (!GetNumaNodeProcessorMask(numa_node, &node_affinity_mask))
{
HPX_THROW_EXCEPTION(kernel_error
, "hpx::threads::windows_topology::init_thread_affinity_mask"
, boost::str(boost::format(
"failed to initialize thread %1% affinity mask")
% num_thread));
}
mask = least_significant_bit(node_affinity_mask) <<
(affinity / numa_nodes);
}
else {
UCHAR numa_node = UCHAR(get_numa_node_number(num_thread));
if (!GetNumaNodeProcessorMask(numa_node, &node_affinity_mask))
{
HPX_THROW_EXCEPTION(kernel_error
, "hpx::threads::windows_topology::init_thread_affinity_mask"
, boost::str(boost::format(
"failed to initialize thread %1% affinity mask")
% num_thread));
}
mask = least_significant_bit(node_affinity_mask) <<
(affinity % num_of_cores_per_numa_node);
}
while (!(mask & node_affinity_mask)) {
mask <<= 1LL;
if (0 == mask)
mask = 0x01LL;
}
return static_cast<mask_type>(mask);
} // }}}
uint32_t osNumaNodes(void)
{
/* Cache the amount of NUMA values. */
static ULONG numNumaNodes = 0;
/* Cache the amount of NUMA nodes. */
if (!numNumaNodes && !GetNumaHighestNodeNumber(&numNumaNodes))
{
numNumaNodes = 1;
}
return numNumaNodes;
}
/* static */
int ThreadPool::getNumaNodeCount()
{
#if defined(_WIN32_WINNT) && _WIN32_WINNT >= _WIN32_WINNT_WIN7
ULONG num = 1;
if (GetNumaHighestNodeNumber(&num))
num++;
return (int)num;
#elif HAVE_LIBNUMA
if (numa_available() >= 0)
return numa_max_node() + 1;
else
return 1;
#else
return 1;
#endif
}
mask_type init_numa_node_affinity_mask(
std::size_t num_thread
, bool numa_sensitive
)
{ // {{{
std::size_t num_of_cores = hardware_concurrency();
UCHAR affinity = UCHAR(num_thread % num_of_cores);
ULONG numa_nodes = 1;
if (GetNumaHighestNodeNumber(&numa_nodes))
++numa_nodes;
ULONGLONG mask = 0;
if (numa_sensitive) {
UCHAR numa_node = affinity % numa_nodes;
if (!GetNumaNodeProcessorMask(numa_node, &mask))
{
HPX_THROW_EXCEPTION(kernel_error
, "hpx::threads::windows_topology::init_numa_node_affinity_mask"
, boost::str(boost::format(
"failed to initialize NUMA node affinity mask for "
"thread %1%")
% num_thread));
}
return static_cast<mask_type>(mask);
}
UCHAR numa_node = UCHAR(get_numa_node_number(num_thread));
if (!GetNumaNodeProcessorMask(numa_node, &mask))
{
HPX_THROW_EXCEPTION(kernel_error
, "hpx::threads::windows_topology::init_numa_node_affinity_mask"
, boost::str(boost::format(
"failed to initialize NUMA node affinity mask for "
"thread %1%")
% num_thread));
}
return static_cast<mask_type>(mask);
} // }}}
std::size_t init_numa_node_number(
std::size_t num_thread
)
{ // {{{
if (std::size_t(-1) == num_thread)
return std::size_t(-1);
UCHAR node_number = 0;
if (GetNumaProcessorNode(UCHAR(num_thread), &node_number))
return node_number;
std::size_t num_of_cores = hardware_concurrency();
if (0 == num_of_cores)
num_of_cores = 1; // assume one core
std::size_t num_of_numa_cores = num_of_cores;
ULONG numa_nodes = 0;
if (GetNumaHighestNodeNumber(&numa_nodes) && 0 != numa_nodes)
num_of_numa_cores = num_of_cores / (numa_nodes + 1);
return num_thread / num_of_numa_cores;
} // }}}
