Merge topic 'ci-nvhpc'

16a5c36795 gitlab-ci: add nightly job testing NVHPC with Ninja on Linux
9474d02386 Tests: Clarify CudaOnly.SeparateCompilationPTX output

Acked-by: Kitware Robot <[email protected]>
Merge-request: !7946

.gitlab-ci.yml

@@ -237,6 +237,16 @@ t:fedora37-makefiles-nospace:
         CMAKE_CI_BUILD_NAME: fedora37_makefiles_nospace
         CMAKE_CI_JOB_NIGHTLY: "true"
 
+t:nvhpc22.9-ninja:
+    extends:
+        - .nvhpc_ninja
+        - .cmake_test_linux_release
+        - .linux_x86_64_v3_builder_tags_cuda
+        - .run_dependent
+        - .needs_centos6_x86_64
+    variables:
+        CMAKE_CI_JOB_NIGHTLY: "true"
+
 t:cuda9.2-nvidia:
     extends:
         - .cuda9.2_nvidia

.gitlab/ci/configure_nvhpc_ninja.cmake

@@ -0,0 +1,5 @@
+set(CMake_TEST_CUDA "NVIDIA" CACHE STRING "")
+
+set(configure_no_sccache 1)
+
+include("${CMAKE_CURRENT_LIST_DIR}/configure_external_test.cmake")

.gitlab/ci/ctest_exclusions.cmake

@@ -27,6 +27,16 @@ if ("$ENV{CMAKE_CONFIGURATION}" MATCHES "_jom")
     )
 endif()
 
+if ("$ENV{CMAKE_CONFIGURATION}" MATCHES "nvhpc_")
+  list(APPEND test_exclusions
+    # FIXME(#24187): This test fails with NVHPC as the CUDA host compiler.
+    "^CudaOnly.SeparateCompilationPTX$"
+
+    # FIXME(#24188): FindCUDAToolkit breaks on some symlink layouts.
+    "^Cuda.Toolkit$"
+    )
+endif()
+
 string(REPLACE ";" "|" test_exclusions "${test_exclusions}")
 if (test_exclusions)
   set(test_exclusions "(${test_exclusions})")

.gitlab/ci/docker/nvhpc22.9/Dockerfile

@@ -0,0 +1,6 @@
+# https://catalog.ngc.nvidia.com/orgs/nvidia/containers/nvhpc/tags
+FROM nvcr.io/nvidia/nvhpc:22.9-devel-cuda_multi-ubuntu22.04
+MAINTAINER Brad King <[email protected]>
+
+COPY install_deps.sh /root/install_deps.sh
+RUN sh /root/install_deps.sh

.gitlab/ci/docker/nvhpc22.9/install_deps.sh

@@ -0,0 +1,11 @@
+#!/bin/sh
+
+set -e
+
+apt-get update
+
+# Install development tools.
+apt-get install -y \
+    curl
+
+apt-get clean

.gitlab/ci/env_nvhpc_ninja.sh

@@ -0,0 +1,5 @@
+export CC=nvc
+export CXX=nvc++
+export FC=nvfortran
+export CUDACXX=nvcc
+export CUDAHOSTCXX=nvc++

.gitlab/os-linux.yml

@@ -255,6 +255,18 @@
         CMAKE_CONFIGURATION: inteloneapi_makefiles
         CMAKE_GENERATOR: "Unix Makefiles"
 
+### NVHPC Compiler
+
+.nvhpc:
+    image: "kitware/cmake:ci-nvhpc22.9-x86_64-2022-11-22"
+    variables:
+        CMAKE_ARCH: x86_64
+
+.nvhpc_ninja:
+    extends: .nvhpc
+    variables:
+        CMAKE_CONFIGURATION: nvhpc_ninja
+
 ### CUDA builds
 
 .cuda9.2:
@@ -392,6 +404,13 @@
         - docker
         - linux
 
+.linux_x86_64_v3_builder_tags_cuda:
+    tags:
+        - cmake
+        - cuda-rt
+        - docker
+        - linux-x86_64-v3
+
 .linux_builder_tags_radeon:
     tags:
         - cmake

Tests/CudaOnly/SeparateCompilationPTX/main.cu

@@ -21,10 +21,11 @@ int main()
   cuCtxCreate(&context, 0, device);
 
   CUmodule module;
-  cuModuleLoadData(&module, kernels);
-  if (module == nullptr) {
-    std::cerr << "Failed to load the embedded ptx" << std::endl;
+  CUresult result = cuModuleLoadData(&module, kernels);
+  std::cout << "module pointer: " << module << '\n';
+  if (result != CUDA_SUCCESS || module == nullptr) {
+    std::cerr << "Failed to load the embedded ptx with error: "
+              << static_cast<unsigned int>(result) << '\n';
     return 1;
   }
-  std::cout << module << std::endl;
 }