fix a special case in parallelization autotuning

src/main.cpp

@@ -922,27 +922,32 @@ void autotune_threads(RaxmlInstance& instance)
   max_workers = std::min(max_workers, (size_t) opts.num_workers_max);
   if (opts.num_workers == 0)
   {
-    auto rank_threads = opts.num_threads > 0 ? opts.num_threads : opts.num_threads_max;
-    auto opt_workers = rank_threads / res.num_threads_throughput;
-    opt_workers *= num_ranks;
-    opts.num_workers = std::min(opt_workers, max_workers);
+    if (max_workers > 1)
+    {
+      auto rank_threads = opts.num_threads > 0 ? opts.num_threads : opts.num_threads_max;
+      auto opt_workers = std::max(rank_threads / res.num_threads_throughput, 1ul);
+      opt_workers *= num_ranks;
+      opts.num_workers = std::min(opt_workers, max_workers);
 
-    while (num_ranks*opts.num_threads % opts.num_workers != 0)
-      opts.num_workers--;
+      while (num_ranks*opts.num_threads % opts.num_workers != 0)
+        opts.num_workers--;
 
-    /* make sure we have integer number of workers per rank */
-    opts.num_workers -= opts.num_workers % num_ranks;
+      /* make sure we have integer number of workers per rank */
+      opts.num_workers -= opts.num_workers % num_ranks;
 
-    opts.num_workers = std::max(opts.num_workers, 1u);
+      opts.num_workers = std::max(opts.num_workers, 1u);
 
-    /* workers spanning multiple MPI ranks are not supported atm -> check for this */
-    if (opts.num_workers > 1 && opts.num_workers < num_ranks)
-    {
-      if (num_ranks <= max_workers)
-        opts.num_workers = num_ranks;
-      else
-        opts.num_workers = 1;
+      /* workers spanning multiple MPI ranks are not supported atm -> check for this */
+      if (opts.num_workers > 1 && opts.num_workers < num_ranks)
+      {
+        if (num_ranks <= max_workers)
+          opts.num_workers = num_ranks;
+        else
+          opts.num_workers = 1;
+      }
     }
+    else
+      opts.num_workers = 1;
   }
 
   assert(opts.num_workers > 0);