Fix T60145: Cycles resets manually set affinity

This change brings back old original logic which was checking
whether worker threads do fit into an active CPU group. But
it does it a bit smarter now and is also checking affinity
within that group. This way Cycles will use all threads on a
Threadripper2 CPU if it's set to automatic number of threads,
but on another hand will not change affinity if user requested
16 threads and changed Blender affinity.

3 files changed, 34 insertions, 10 deletions

diff --git a/intern/cycles/util/util_system.cpp b/intern/cycles/util/util_system.cpp
index cc2d7017fd8..a22bd25ce77 100644
--- a/intern/cycles/util/util_system.cpp
+++ b/intern/cycles/util/util_system.cpp
@@ -40,7 +40,7 @@ bool system_cpu_ensure_initialized()
 {
 	static bool is_initialized = false;
 	static bool result = false;
-	if (is_initialized) {
+	if(is_initialized) {
 		return result;
 	}
 	is_initialized = true;
@@ -71,8 +71,8 @@ int system_cpu_thread_count()
 {
 	const int num_nodes = system_cpu_num_numa_nodes();
 	int num_threads = 0;
-	for (int node = 0; node < num_nodes; ++node) {
-		if (!system_cpu_is_numa_node_available(node)) {
+	for(int node = 0; node < num_nodes; ++node) {
+		if(!system_cpu_is_numa_node_available(node)) {
 			continue;
 		}
 		num_threads += system_cpu_num_numa_node_processors(node);
@@ -82,7 +82,7 @@ int system_cpu_thread_count()
 
 int system_cpu_num_numa_nodes()
 {
-	if (!system_cpu_ensure_initialized()) {
+	if(!system_cpu_ensure_initialized()) {
 		/* Fallback to a single node with all the threads. */
 		return 1;
 	}
@@ -91,7 +91,7 @@ int system_cpu_num_numa_nodes()
 
 bool system_cpu_is_numa_node_available(int node)
 {
-	if (!system_cpu_ensure_initialized()) {
+	if(!system_cpu_ensure_initialized()) {
 		return true;
 	}
 	return numaAPI_IsNodeAvailable(node);
@@ -99,7 +99,7 @@ bool system_cpu_is_numa_node_available(int node)
 
 int system_cpu_num_numa_node_processors(int node)
 {
-	if (!system_cpu_ensure_initialized()) {
+	if(!system_cpu_ensure_initialized()) {
 		return system_cpu_thread_count_fallback();
 	}
 	return numaAPI_GetNumNodeProcessors(node);
@@ -107,12 +107,20 @@ int system_cpu_num_numa_node_processors(int node)
 
 bool system_cpu_run_thread_on_node(int node)
 {
-	if (!system_cpu_ensure_initialized()) {
+	if(!system_cpu_ensure_initialized()) {
 		return true;
 	}
 	return numaAPI_RunThreadOnNode(node);
 }
 
+int system_cpu_num_active_group_processors()
+{
+	if(!system_cpu_ensure_initialized()) {
+		return system_cpu_thread_count_fallback();
+	}
+	return numaAPI_GetNumCurrentNodesProcessors();
+}
+
 #if !defined(_WIN32) || defined(FREE_WINDOWS)
 static void __cpuid(int data[4], int selector)
 {
diff --git a/intern/cycles/util/util_system.h b/intern/cycles/util/util_system.h
index 15f69bcf153..0c001f11f0e 100644
--- a/intern/cycles/util/util_system.h
+++ b/intern/cycles/util/util_system.h
@@ -44,6 +44,10 @@ int system_cpu_num_numa_node_processors(int node);
  * Returns truth if affinity has successfully changed. */
 bool system_cpu_run_thread_on_node(int node);
 
+/* Number of processors within the current CPU group (or within active thread
+ * thread affinity). */
+int system_cpu_num_active_group_processors();
+
 string system_cpu_brand_string();
 int system_cpu_bits();
 bool system_cpu_support_sse2();
diff --git a/intern/cycles/util/util_task.cpp b/intern/cycles/util/util_task.cpp
index 7e9f7313fba..4241c4aa8cc 100644
--- a/intern/cycles/util/util_task.cpp
+++ b/intern/cycles/util/util_task.cpp
@@ -228,9 +228,21 @@ int get_num_total_processors(const vector<int>& num_per_node_processors)
 void distribute_threads_on_nodes(const vector<thread*>& threads)
 {
 	const int num_threads = threads.size();
-	/* TODO(sergey): Skip overriding affinity if threads fits into the current
-	 * nodes/CPU group. This will allow user to tweak affinity for weird and
-	 * wonderful reasons. */
+	const int num_active_group_processors =
+	        system_cpu_num_active_group_processors();
+	VLOG(1) << "Detected " << num_active_group_processors << " processors "
+	        << "in active group.";
+	if(num_active_group_processors >= num_threads) {
+		/* If the current thread is set up in a way that its affinity allows to
+		 * use at least requested number of threads we do not explicitly set
+		 * affinity to the worker therads.
+		 * This way we allow users to manually edit affinity of the parent
+		 * thread, and here we follow that affinity. This way it's possible to
+		 * have two Cycles/Blender instances running manually set to a different
+		 * dies on a CPU. */
+		VLOG(1) << "Not setting thread group affinity.";
+		return;
+	}
 	vector<int> num_per_node_processors;
 	get_per_node_num_processors(&num_per_node_processors);
 	if(num_per_node_processors.size() == 0) {