ppc64le: Workaround "error: identifier "__ieee128" is undefined"

This appears to be a bug with GCC 8+ and CUDA 10. We need to
disable quadruple precision there, which shouldn't be a huge
issue.
LLNL/blt#341 (comment)

libstdc++8 with gcc 8.2 has a bug that it doesn't compile with
-mno-float128. https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84654
We would need to upgrade to 8.3 at least. However, I didn't
manage to do that via conda in our Jenkinsfile. Went with 7.2
instead.

Signed-off-by: Max H. Gerlach <git@maxgerlach.de>

CMakeLists.txt

@@ -156,8 +156,13 @@ endif()
 include(CheckLanguage)
 check_language(CUDA)
 if (CMAKE_CUDA_COMPILER)
+    if ((CMAKE_CXX_COMPILER_ID MATCHES GNU) AND (CMAKE_SYSTEM_PROCESSOR MATCHES ppc64le))
+        if (CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 8.0)
+            set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Xcompiler -mno-float128")
+        endif ()
+    endif ()
     enable_language(CUDA)
-endif()
+endif ()
 
 macro(ADD_CUDA)
     find_package(CUDAToolkit REQUIRED)

Jenkinsfile.ppc64le

@@ -24,9 +24,9 @@ pipeline {
                 sh '''#!/usr/bin/env bash
                       git submodule update --init --recursive
                       . ${CONDA_INIT}
-                      conda activate ${CONDA_ENV}
-                      conda install -y "cmake>=3.18"
                       set -xe
+                      conda activate ${CONDA_ENV}
+                      conda install -y "gcc=7.2"
                       HOROVOD_WITHOUT_MXNET=1 HOROVOD_WITHOUT_GLOO=1 HOROVOD_WITH_PYTORCH=1 HOROVOD_WITH_TENSORFLOW=1 \
                           HOROVOD_CUDA_HOME="/usr/local/cuda" HOROVOD_GPU_OPERATIONS=NCCL \
                           pip install -v . --no-cache-dir --no-deps