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;
} // }}}
ThreadPool* ThreadPool::allocThreadPools(x265_param* p, int& numPools)
{
enum { MAX_NODE_NUM = 127 };
int cpusPerNode[MAX_NODE_NUM + 1];
memset(cpusPerNode, 0, sizeof(cpusPerNode));
int numNumaNodes = X265_MIN(getNumaNodeCount(), MAX_NODE_NUM);
int cpuCount = getCpuCount();
bool bNumaSupport = false;
#if defined(_WIN32_WINNT) && _WIN32_WINNT >= _WIN32_WINNT_WIN7
bNumaSupport = true;
#elif HAVE_LIBNUMA
bNumaSupport = numa_available() >= 0;
#endif
for (int i = 0; i < cpuCount; i++)
{
#if defined(_WIN32_WINNT) && _WIN32_WINNT >= _WIN32_WINNT_WIN7
UCHAR node;
if (GetNumaProcessorNode((UCHAR)i, &node))
cpusPerNode[X265_MIN(node, (UCHAR)MAX_NODE_NUM)]++;
else
#elif HAVE_LIBNUMA
if (bNumaSupport >= 0)
cpusPerNode[X265_MIN(numa_node_of_cpu(i), MAX_NODE_NUM)]++;
else
#endif
cpusPerNode[0]++;
}
if (bNumaSupport && p->logLevel >= X265_LOG_DEBUG)
for (int i = 0; i < numNumaNodes; i++)
x265_log(p, X265_LOG_DEBUG, "detected NUMA node %d with %d logical cores\n", i, cpusPerNode[i]);
/* limit nodes based on param->numaPools */
if (p->numaPools && *p->numaPools)
{
const char *nodeStr = p->numaPools;
for (int i = 0; i < numNumaNodes; i++)
{
if (!*nodeStr)
{
cpusPerNode[i] = 0;
continue;
}
else if (*nodeStr == '-')
cpusPerNode[i] = 0;
else if (*nodeStr == '*')
break;
else if (*nodeStr == '+')
;
else
{
int count = atoi(nodeStr);
cpusPerNode[i] = X265_MIN(count, cpusPerNode[i]);
}
/* consume current node string, comma, and white-space */
while (*nodeStr && *nodeStr != ',')
++nodeStr;
if (*nodeStr == ',' || *nodeStr == ' ')
++nodeStr;
}
}
// In the case that numa is disabled and we have more CPUs than 64,
// spawn the last pool only if the # threads in that pool is > 1/2 max (heuristic)
if ((numNumaNodes == 1) && (cpusPerNode[0] % MAX_POOL_THREADS < (MAX_POOL_THREADS / 2)))
{
cpusPerNode[0] -= (cpusPerNode[0] % MAX_POOL_THREADS);
x265_log(p, X265_LOG_DEBUG, "Creating only %d worker threads to prevent asymmetry in pools; may not use all HW contexts\n", cpusPerNode[0]);
}
numPools = 0;
for (int i = 0; i < numNumaNodes; i++)
{
if (bNumaSupport)
x265_log(p, X265_LOG_DEBUG, "NUMA node %d may use %d logical cores\n", i, cpusPerNode[i]);
if (cpusPerNode[i])
numPools += (cpusPerNode[i] + MAX_POOL_THREADS - 1) / MAX_POOL_THREADS;
}
if (!numPools)
return NULL;
if (numPools > p->frameNumThreads)
{
x265_log(p, X265_LOG_DEBUG, "Reducing number of thread pools for frame thread count\n");
numPools = X265_MAX(p->frameNumThreads / 2, 1);
}
ThreadPool *pools = new ThreadPool[numPools];
if (pools)
{
int maxProviders = (p->frameNumThreads + numPools - 1) / numPools + 1; /* +1 is Lookahead, always assigned to threadpool 0 */
int node = 0;
for (int i = 0; i < numPools; i++)
{
while (!cpusPerNode[node])
node++;
int cores = X265_MIN(MAX_POOL_THREADS, cpusPerNode[node]);
if (!pools[i].create(cores, maxProviders, node))
{
X265_FREE(pools);
numPools = 0;
return NULL;
}
if (numNumaNodes > 1)
x265_log(p, X265_LOG_INFO, "Thread pool %d using %d threads on NUMA node %d\n", i, cores, node);
else
x265_log(p, X265_LOG_INFO, "Thread pool created using %d threads\n", cores);
cpusPerNode[node] -= cores;
}
}
else
numPools = 0;
return pools;
}
