Bump gloo

Summary:
Latest version of Gloo takes care of MPI_Init/MPI_Finalize for us, so
this commit removes handling that from caffe2/contrib/gloo. It also
imports CMake NCCL module changes from Gloo to stay consistent and
allow setting NCCL_INCLUDE_DIR and NCCL_LIB_DIR separately.
Closes https://github.com/caffe2/caffe2/pull/1295

Reviewed By: dzhulgakov

Differential Revision: D5979364

Pulled By: pietern

fbshipit-source-id: 794b00b0a445317c30a13cc8f0f4dc38e590cc77

caffe2/contrib/gloo/common.cc

@@ -19,14 +19,6 @@
 #include "caffe2/core/logging.h"
 #include "caffe2/core/tensor.h"
 
-#if defined(GLOO_USE_MPI) && GLOO_USE_MPI
-#include <mutex>
-#endif
-
-#if defined(GLOO_USE_MPI) && GLOO_USE_MPI
-#include <mpi.h>
-#endif
-
 #include <gloo/transport/tcp/device.h>
 #if defined(GLOO_USE_IBVERBS) && GLOO_USE_IBVERBS
 #include <gloo/transport/ibverbs/device.h>
@@ -68,26 +60,5 @@ std::shared_ptr<::gloo::transport::Device> createDevice(
   CAFFE_THROW("Invalid transport: ", attr.transport);
 }
 
-#if defined(GLOO_USE_MPI) && GLOO_USE_MPI
-static std::mutex mpiMutex;
-static int mpiRefs = 0;
-
-void mpiInitialize() {
-  std::lock_guard<std::mutex> lock(mpiMutex);
-  if (mpiRefs++ == 0) {
-    auto rv = MPI_Init(nullptr, nullptr);
-    CAFFE_ENFORCE_EQ(rv, 0, "MPI_Init() failed");
-  }
-}
-
-void mpiFinalize() {
-  std::lock_guard<std::mutex> lock(mpiMutex);
-  if (--mpiRefs == 0) {
-    auto rv = MPI_Finalize();
-    CAFFE_ENFORCE_EQ(rv, 0, "MPI_Finalize() failed");
-  }
-}
-#endif
-
 } // namespace gloo
 } // namespace caffe2

caffe2/contrib/gloo/common.h

@@ -40,10 +40,5 @@ struct createDeviceAttr {
 std::shared_ptr<::gloo::transport::Device> createDevice(
     const createDeviceAttr attr);
 
-#if defined(GLOO_USE_MPI) && GLOO_USE_MPI
-void mpiInitialize();
-void mpiFinalize();
-#endif
-
 } // namespace gloo
 } // namespace caffe2

caffe2/contrib/gloo/common_world_ops.h

@@ -63,25 +63,14 @@ class CreateCommonWorld final : public Operator<Context> {
       ws_->CreateBlob(status_blob_);
     }
     initialize();
-
-#if defined(GLOO_USE_MPI) && GLOO_USE_MPI
-    if (mpi_rendezvous_) {
-      mpiInitialize();
-    }
-#endif
   }
 
   virtual ~CreateCommonWorld() {
-#if defined(GLOO_USE_MPI) && GLOO_USE_MPI
-    if (mpi_rendezvous_) {
-      mpiFinalize();
-    }
-#endif
   }
 
   CommonWorld rendezvousWithMPI() {
 #if defined(GLOO_USE_MPI) && GLOO_USE_MPI
-    auto context = std::make_shared<::gloo::mpi::Context>(MPI_COMM_WORLD);
+    auto context = ::gloo::mpi::Context::createManaged();
     if (timeout_ms_ != -1) {
       context->setTimeout(std::chrono::milliseconds(timeout_ms_));
     }

cmake/Modules/FindNCCL.cmake

@@ -1,7 +1,9 @@
-# - Try to find NCCL
+# Find the nccl libraries
 #
 # The following variables are optionally searched for defaults
-#  NCCL_ROOT_DIR:            Base directory where all NCCL components are found
+#  NCCL_ROOT_DIR: Base directory where all NCCL components are found
+#  NCCL_INCLUDE_DIR: Directory where NCCL header is found
+#  NCCL_LIB_DIR: Directory where NCCL library is found
 #
 # The following are set after configuration is done:
 #  NCCL_FOUND
@@ -10,26 +12,26 @@
 
 set(NCCL_ROOT_DIR "" CACHE PATH "Folder contains NVIDIA NCCL")
 
-find_path(NCCL_INCLUDE_DIR
+find_path(NCCL_INCLUDE_DIRS
   NAMES nccl.h
   HINTS
+  ${NCCL_INCLUDE_DIR}
   ${NCCL_ROOT_DIR}
   ${NCCL_ROOT_DIR}/include)
 
-find_library(NCCL_LIBRARY
+find_library(NCCL_LIBRARIES
   NAMES nccl
   HINTS
+  ${NCCL_LIB_DIR}
   ${NCCL_ROOT_DIR}
   ${NCCL_ROOT_DIR}/lib
   ${NCCL_ROOT_DIR}/lib/x86_64-linux-gnu
   ${NCCL_ROOT_DIR}/lib64)
 
 include(FindPackageHandleStandardArgs)
-find_package_handle_standard_args(NCCL DEFAULT_MSG NCCL_INCLUDE_DIR NCCL_LIBRARY)
+find_package_handle_standard_args(NCCL DEFAULT_MSG NCCL_INCLUDE_DIRS NCCL_LIBRARIES)
 
 if(NCCL_FOUND)
-  set(NCCL_INCLUDE_DIRS ${NCCL_INCLUDE_DIR})
-  set(NCCL_LIBRARIES ${NCCL_LIBRARY})
   message(STATUS "Found NCCL (include: ${NCCL_INCLUDE_DIRS}, library: ${NCCL_LIBRARIES})")
   mark_as_advanced(NCCL_ROOT_DIR NCCL_INCLUDE_DIRS NCCL_LIBRARIES)
 endif()