Merge pull request #4257 from jsquyres/pr/moar-hwloc-cuda-cleanup

hwloc2a/configure.m4: be more careful in with_cuda->enable_cuda

opal/mca/hwloc/hwloc2a/configure.m4

@@ -124,14 +124,36 @@ AC_DEFUN([MCA_opal_hwloc_hwloc2a_CONFIG],[
 
     # Per https://github.com/open-mpi/ompi/issues/4219, if
     # --without-cuda was specified, be sure to disable it in hwloc,
-    # too.  Note that hwloc uses --disable-cuda, so just set
-    # enable_cuda=$with_cuda to get the same value that was passed in
-    # via the top-level --with-cuda CLI option.
-    enable_cuda=$with_cuda
+    # too.  Note that hwloc uses --disable-cuda (i.e., a yes or no
+    # value), whereas we use --with-cuda here in Open MPI (i.e., a
+    # yes, no, or path value).  Need to translate appropriately.
+    #
+    # Set enable_cuda to yes if:
+    #
+    # 1. --with-cuda was specified (i.e., a human specifically asked
+    # for it)
+    # 2. --with-cuda=blah was specified (i.e., a human specifically
+    # asked for it)
+    # 3. --with-cuda was not specified, but Open MPI is building CUDA
+    # support
+    #
+    # Set enable_cuda to no in all other cases.  This logic could be
+    # compressed into a smaller set of if tests, but for readability /
+    # clarity, I left it expanded.
+    AC_MSG_CHECKING([for hwloc --enable-cuda value])
+    enable_cuda=no
+    AS_IF([test "$with_cuda" = "yes"],
+          [enable_cuda=yes],
+          [AS_IF([test -n "$with_cuda" && test "$with_cuda" != "no"],
+                 [enable_cuda=yes],
+                 [AS_IF([test "$opal_check_cuda_happy" = "yes"],
+                        [enable_cuda=yes])
+                 ])
+          ])
+    AC_MSG_RESULT(["$enable_cuda"])
 
-    # Disable NVML support if CUDA support is not built
-    AS_IF([test "$opal_check_cuda_happy" != "yes"],
-          [enable_nvml=no])
+    # Open MPI currently does not use hwloc's NVML support
+    enable_nvml=no
 
     # hwloc checks for compiler visibility, and its needs to do
     # this without "picky" flags.