WIP -- OpenMP automate settings to get good performance

core/CMakeLists.txt

@@ -8,6 +8,7 @@ set(teca_core_srcs
     teca_bad_cast.cxx
     teca_binary_stream.cxx
     teca_common.cxx
+    teca_core_init.cxx
     teca_dataset.cxx
     teca_dataset_capture.cxx
     teca_index_executive.cxx

core/teca_core_init.cxx

@@ -0,0 +1,33 @@
+#include <iostream>
+#include <string>
+#include <cstdlib>
+#include <thread>
+
+__attribute__((constructor)) void init(void)
+{
+    std::cerr << "teca initializing ... " << std::endl;
+
+    // do something reasonable if the user doesn't. TODO -- see if we can use
+    // MPI here so we can coordinate thread placement across all ranks on the
+    // same node.
+    if (!getenv("OMP_NUM_THREADS"))
+    {
+        int n_threads = std::max(1u, std::thread::hardware_concurrency() / 2);
+        setenv("OMP_NUM_THREADS", std::to_string(n_threads).c_str(), 1);
+    }
+
+    if (!getenv("OMP_PROC_BIND"))
+        setenv("OMP_PROC_BIND", "true", 1);
+
+    if (!getenv("OMP_PLACES"))
+       setenv("OMP_PLACES", "cores", 1);
+
+    // TODO -- remove these, they just prove that the ab ove works.
+    setenv("OMP_DISPLAY_ENV", "true", 1);
+    setenv("OMP_DISPLAY_AFFINITY", "true", 1);
+}
+
+__attribute__((destructor))  void fini(void)
+{
+    std::cerr << "teca finalizing ... " << std::endl;
+}