Add VerilatedContext::useNumaAssign and set on threads() call (#6954)

include/verilated.cpp

@@ -2816,6 +2816,7 @@ void VerilatedContext::threads(unsigned n) {
             "%Error: Cannot set simulation threads after the thread pool has been created.");
     }
 
+    m_useNumaAssign = true;
     if (m_threads == n) return;  // To avoid unnecessary warnings
     m_threads = n;
     const unsigned threadsAvailableToProcess = VlOs::getProcessDefaultParallelism();
@@ -2826,6 +2827,8 @@ void VerilatedContext::threads(unsigned n) {
     }
 }
 
+void VerilatedContext::useNumaAssign(bool flag) { m_useNumaAssign = flag; }
+
 void VerilatedContext::commandArgs(int argc, const char** argv) VL_MT_SAFE_EXCLUDES(m_argMutex) {
     // Not locking m_argMutex here, it is done in impp()->commandArgsAddGuts
     // m_argMutex here is the same as in impp()->commandArgsAddGuts;

include/verilated.h

@@ -433,6 +433,8 @@ protected:
     const std::unique_ptr<VerilatedContextImpData> m_impdatap;
     // Number of threads to use for simulation (size of m_threadPool + 1 for main thread)
     unsigned m_threads = VlOs::getProcessDefaultParallelism();
+    // Use numa automatic CPU-to-thread assignment
+    bool m_useNumaAssign = false;
     // Number of threads in added models
     unsigned m_threadsInModels = 0;
     // The thread pool shared by all models added to this context
@@ -599,6 +601,13 @@ public:
     /// Can only be called before the thread pool is created (before first model is added).
     void threads(unsigned n);
 
+    /// Use numa automatic CPU-to-thread assignment.
+    bool useNumaAssign() const VL_MT_SAFE { return m_useNumaAssign; }
+    /// Set numa assignment of threads to cores
+    /// Defaults false; set true automatically when threads() called;
+    /// call this to override back to false if numa assignment not wanted.
+    void useNumaAssign(bool flag);
+
     /// Trace signals in models within the context; called by application code
     void trace(VerilatedTraceBaseC* tfp, int levels, int options = 0);
     /// Allow traces to at some point be enabled (disables some optimizations)

include/verilated_threads.cpp

@@ -137,7 +137,7 @@ VlThreadPool::VlThreadPool(VerilatedContext* contextp, unsigned nThreads) {
         m_workers.push_back(new VlWorkerThread{contextp});
         m_unassignedWorkers.push(i);
     }
-    m_numaStatus = numaAssign();
+    m_numaStatus = numaAssign(contextp);
 }
 
 VlThreadPool::~VlThreadPool() {
@@ -145,8 +145,9 @@ VlThreadPool::~VlThreadPool() {
     for (auto& i : m_workers) delete i;
 }
 
-std::string VlThreadPool::numaAssign() {
+std::string VlThreadPool::numaAssign(VerilatedContext* contextp) {
 #if defined(__linux) || defined(CPU_ZERO) || defined(VL_CPPCHECK)  // Linux-like pthreads
+    if (contextp && !contextp->useNumaAssign()) { return "NUMA assignment not requested"; }
     std::string numa_strategy = VlOs::getenvStr("VERILATOR_NUMA_STRATEGY", "default");
     if (numa_strategy == "none") {
         return "no NUMA assignment requested";

include/verilated_threads.h

@@ -254,7 +254,7 @@ public:
 private:
     VL_UNCOPYABLE(VlThreadPool);
 
-    std::string numaAssign();
+    std::string numaAssign(VerilatedContext* contextp);
 };
 
 #endif