Cr8-xyz/blender
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

Fix T60585: Cycles not using all cores on threadripper, after recent changes.

Browse Source
This commit is contained in:
Brecht Van Lommel
2019-01-18 19:15:38 +01:00
parent 1e3203986c
commit c5eb10b110
3 changed files with 14 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
intern/cycles/util/util_task.cpp
View File

@@ -225,9 +225,9 @@ int get_num_total_processors(const vector<int>& num_per_node_processors)
/* Assign every thread a node on which is should be running, for the best /* Assign every thread a node on which is should be running, for the best
* performance. */ * performance. */
void distribute_threads_on_nodes(const vector<thread*>& threads) vector<int> distribute_threads_on_nodes(const int num_threads)
{ {
const int num_threads = threads.size(); vector<int> thread_nodes(num_threads, -1);
const int num_active_group_processors = const int num_active_group_processors =
system_cpu_num_active_group_processors(); system_cpu_num_active_group_processors();
VLOG(1) << "Detected " << num_active_group_processors << " processors " VLOG(1) << "Detected " << num_active_group_processors << " processors "
@@ -241,14 +241,14 @@ void distribute_threads_on_nodes(const vector<thread*>& threads)
* have two Cycles/Blender instances running manually set to a different * have two Cycles/Blender instances running manually set to a different
* dies on a CPU. */ * dies on a CPU. */
VLOG(1) << "Not setting thread group affinity."; VLOG(1) << "Not setting thread group affinity.";
return; return thread_nodes;
} }
vector<int> num_per_node_processors; vector<int> num_per_node_processors;
get_per_node_num_processors(&num_per_node_processors); get_per_node_num_processors(&num_per_node_processors);
if(num_per_node_processors.size() == 0) { if(num_per_node_processors.size() == 0) {
/* Error was already repported, here we can't do anything, so we simply /* Error was already repported, here we can't do anything, so we simply
* leave default affinity to all the worker threads. */ * leave default affinity to all the worker threads. */
return; return thread_nodes;
} }
const int num_nodes = num_per_node_processors.size(); const int num_nodes = num_per_node_processors.size();
int thread_index = 0; int thread_index = 0;
@@ -273,11 +273,11 @@ void distribute_threads_on_nodes(const vector<thread*>& threads)
{ {
VLOG(1) << "Scheduling thread " << thread_index << " to node " VLOG(1) << "Scheduling thread " << thread_index << " to node "
<< current_node_index << "."; << current_node_index << ".";
threads[thread_index]->schedule_to_node(current_node_index); thread_nodes[thread_index] = current_node_index;
++thread_index; ++thread_index;
if(thread_index == num_threads) { if(thread_index == num_threads) {
/* All threads are scheduled on their nodes. */ /* All threads are scheduled on their nodes. */
return; return thread_nodes;
} }
} }
++current_node_index; ++current_node_index;
@@ -305,6 +305,8 @@ void distribute_threads_on_nodes(const vector<thread*>& threads)
++thread_index; ++thread_index;
current_node_index = (current_node_index + 1) % num_nodes; current_node_index = (current_node_index + 1) % num_nodes;
} }
return thread_nodes;
} }
} // namespace } // namespace
@@ -325,13 +327,17 @@ void TaskScheduler::init(int num_threads)
num_threads = system_cpu_thread_count(); num_threads = system_cpu_thread_count();
} }
VLOG(1) << "Creating pool of " << num_threads << " threads."; VLOG(1) << "Creating pool of " << num_threads << " threads.";
/* Compute distribution on NUMA nodes. */
vector<int> thread_nodes = distribute_threads_on_nodes(num_threads);
/* Launch threads that will be waiting for work. */ /* Launch threads that will be waiting for work. */
threads.resize(num_threads); threads.resize(num_threads);
for(int thread_index = 0; thread_index < num_threads; ++thread_index) { for(int thread_index = 0; thread_index < num_threads; ++thread_index) {
threads[thread_index] = new thread( threads[thread_index] = new thread(
function_bind(&TaskScheduler::thread_run, thread_index + 1)); function_bind(&TaskScheduler::thread_run, thread_index + 1),
thread_nodes[thread_index]);
} }
distribute_threads_on_nodes(threads);
} }
void TaskScheduler::exit() void TaskScheduler::exit()

5
intern/cycles/util/util_thread.cpp
View File

@@ -58,9 +58,4 @@ bool thread::join()
} }
} }
void thread::schedule_to_node(int node)
{
node_ = node;
}
CCL_NAMESPACE_END CCL_NAMESPACE_END

4
intern/cycles/util/util_thread.h
View File

@@ -54,10 +54,6 @@ public:
static void *run(void *arg); static void *run(void *arg);
bool join(); bool join();
/* For an existing thread descriptor which is NOT running yet, assign node
* on which it should be running. */
void schedule_to_node(int node);
protected: protected:
function<void()> run_cb_; function<void()> run_cb_;
std::thread thread_; std::thread thread_;