mpi: use mpi attribute consistently as the default MPI implementations

Use the attribute mpi to provide a system wide default MPI
implementation. The default is openmpi (as before).
This now allows for overriding the MPI implentation by using
the overlay mechanism. Build all packages with mpich instead
of the default openmpi can now be achived like this:
self: super:
 {
   mpi = super.mpich;
 }

All derivations that have been using "mpi ? null" to provide optional
building with MPI have been change in the following way to allow for
optional builds with MPI:
{ ...
, mpi
, useMpi ? false
}

pkgs/development/python-modules/cntk/default.nix

@@ -3,7 +3,7 @@
 , pkgs
 , numpy
 , scipy
-, openmpi
+, mpi
 , enum34
 , protobuf
 , pip
@@ -17,8 +17,8 @@ in
 buildPythonPackage {
   inherit (cntk) name version src;
 
-  nativeBuildInputs = [ swig openmpi ];
-  buildInputs = [ cntk openmpi ];
+  nativeBuildInputs = [ swig mpi ];
+  buildInputs = [ cntk mpi ];
   propagatedBuildInputs = [ numpy scipy enum34 protobuf pip ];
 
   CNTK_LIB_PATH = "${cntk}/lib";
@@ -28,7 +28,7 @@ buildPythonPackage {
 
   postPatch = ''
     cd bindings/python
-    sed -i 's,"libmpi.so.12","${openmpi}/lib/libmpi.so",g' cntk/train/distributed.py
+    sed -i 's,"libmpi.so.12","${mpi}/lib/libmpi.so",g' cntk/train/distributed.py
 
     # Remove distro and libs checks; they aren't compatible with NixOS and besides we guarantee
     # compatibility by providing a package.

pkgs/development/python-modules/pytorch/default.nix

@@ -1,7 +1,7 @@
 { stdenv, lib, fetchFromGitHub, fetchpatch, buildPythonPackage, python,
   cudaSupport ? false, cudatoolkit ? null, cudnn ? null, nccl ? null, magma ? null,
   mklDnnSupport ? true, useSystemNccl ? true,
-  openMPISupport ? false, openmpi ? null,
+  MPISupport ? false, mpi,
   buildDocs ? false,
   cudaArchList ? null,
 
@@ -29,8 +29,6 @@
 
   isPy3k, pythonOlder }:
 
-assert !openMPISupport || openmpi != null;
-
 # assert that everything needed for cuda is present and that the correct cuda versions are used
 assert !cudaSupport || cudatoolkit != null;
 assert cudnn == null || cudatoolkit != null;
@@ -38,7 +36,7 @@ assert !cudaSupport || (let majorIs = lib.versions.major cudatoolkit.version;
                         in majorIs == "9" || majorIs == "10" || majorIs == "11");
 
 # confirm that cudatoolkits are sync'd across dependencies
-assert !(openMPISupport && cudaSupport) || openmpi.cudatoolkit == cudatoolkit;
+assert !(MPISupport && cudaSupport) || mpi.cudatoolkit == cudatoolkit;
 assert !cudaSupport || magma.cudatoolkit == cudatoolkit;
 
 let
@@ -224,7 +222,7 @@ in buildPythonPackage rec {
     typing-extensions
     # the following are required for tensorboard support
     pillow six future tensorflow-tensorboard protobuf
-  ] ++ lib.optionals openMPISupport [ openmpi ]
+  ] ++ lib.optionals MPISupport [ mpi ]
     ++ lib.optionals (pythonOlder "3.7") [ dataclasses ];
 
   checkInputs = [ hypothesis ninja psutil ];

pkgs/development/python-modules/tensorflow/default.nix

@@ -11,7 +11,7 @@
 # Common deps
 , git, pybind11, which, binutils, glibcLocales, cython, perl
 # Common libraries
-, jemalloc, openmpi, gast, grpc, sqlite, boringssl, jsoncpp
+, jemalloc, mpi, gast, grpc, sqlite, boringssl, jsoncpp
 , curl, snappy, flatbuffers-core, lmdb-core, icu, double-conversion, libpng, libjpeg_turbo, giflib
 # Upsteam by default includes cuda support since tensorflow 1.15. We could do
 # that in nix as well. It would make some things easier and less confusing, but
@@ -129,7 +129,7 @@ let
 
     buildInputs = [
       jemalloc
-      openmpi
+      mpi
       glibcLocales
       git