[llvm] [Offload] libomptarget force dlopen vendor libraries by default. (PR #92788)

[Ye Luo via llvm-commits]

https://github.com/ye-luo created https://github.com/llvm/llvm-project/pull/92788

Since #87009, libomptarget directly links all the plugins statically. All the dependencies of plugins got exposed to libomptarget. The CUDA plugin depends on libcuda and the amdgpu plugin depends on libhsa if not forced using dlopen. On a cluster with different compute node architectures, libomptarget can be built and run on different nodes. In the build stage, if cmake founds libcuda and `LIBOMPTARGET_FORCE_DLOPEN_LIBCUDA=OFF`, libomptarget links libcuda.so directly and the result libomptarget may not run a node without a NVIDIA driver for example a CPU or AMD GPU only machine with a complaint that libcuda.so not found.

The solution is setting  `LIBOMPTARGET_FORCE_DLOPEN_LIBCUDA` and `LIBOMPTARGET_FORCE_DLOPEN_LIBHSA` `ON`. Preferably this should be default to maximize the usability of libomptarget. If cmake detects NVIDIA or AMD software on an OS imaging building node, the resulted libomptarget may not be able to function on the user side due to the requirement the existence of vendor runtime libraries.

>From 498fd037283f0e8937a6fd04f43efcdf35837566 Mon Sep 17 00:00:00 2001
From: Ye Luo <yeluo at anl.gov>
Date: Mon, 20 May 2024 11:43:30 -0500
Subject: [PATCH] [Offload] libomptarget force dlopen vendor libraries by
 default.

Since #87009, libomptarget directly links all the plugins statically. All the dependencies of plugins got exposed to libomptarget.
The CUDA plugin depends on libcuda and the amdgpu plugin depends on libhsa if not forced using dlopen.
On a cluster with different compute node architectures, libomptarget can be built and run on different nodes.
In the build stage, if cmake founds libcuda and `LIBOMPTARGET_FORCE_DLOPEN_LIBCUDA=OFF`, libomptarget links libcuda.so directly and the result libomptarget may not run a node without a NVIDIA driver for example a CPU or AMD GPU only machine with a complaint that libcuda.so not found.

The solution is setting  `LIBOMPTARGET_FORCE_DLOPEN_LIBCUDA` and `LIBOMPTARGET_FORCE_DLOPEN_LIBHSA` `ON`.
Preferably this should be default. If cmake detects NVIDIA or AMD software on an OS imaging building node, the resulted libomptarget may not be able to function on the user side due to the requirement the existence of vendor runtime libraries.
---
 offload/plugins-nextgen/amdgpu/CMakeLists.txt | 2 +-
 offload/plugins-nextgen/cuda/CMakeLists.txt   | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/offload/plugins-nextgen/amdgpu/CMakeLists.txt b/offload/plugins-nextgen/amdgpu/CMakeLists.txt
index 2f4057c0ae7ef..7630e3788dae0 100644
--- a/offload/plugins-nextgen/amdgpu/CMakeLists.txt
+++ b/offload/plugins-nextgen/amdgpu/CMakeLists.txt
@@ -13,7 +13,7 @@ target_sources(omptarget.rtl.amdgpu PRIVATE src/rtl.cpp)
 target_include_directories(omptarget.rtl.amdgpu PRIVATE
                            ${CMAKE_CURRENT_SOURCE_DIR}/utils)
 
-option(LIBOMPTARGET_FORCE_DLOPEN_LIBHSA "Build with dlopened libhsa" OFF)
+option(LIBOMPTARGET_FORCE_DLOPEN_LIBHSA "Build with dlopened libhsa" ON)
 if(hsa-runtime64_FOUND AND NOT LIBOMPTARGET_FORCE_DLOPEN_LIBHSA)
   message(STATUS "Building AMDGPU plugin linked against libhsa")
   target_link_libraries(omptarget.rtl.amdgpu PRIVATE hsa-runtime64::hsa-runtime64)
diff --git a/offload/plugins-nextgen/cuda/CMakeLists.txt b/offload/plugins-nextgen/cuda/CMakeLists.txt
index 10ff612848ad2..fa5559c5e7dcb 100644
--- a/offload/plugins-nextgen/cuda/CMakeLists.txt
+++ b/offload/plugins-nextgen/cuda/CMakeLists.txt
@@ -10,7 +10,7 @@ add_target_library(omptarget.rtl.cuda CUDA)
 
 target_sources(omptarget.rtl.cuda PRIVATE src/rtl.cpp)
 
-option(LIBOMPTARGET_FORCE_DLOPEN_LIBCUDA "Build with dlopened libcuda" OFF)
+option(LIBOMPTARGET_FORCE_DLOPEN_LIBCUDA "Build with dlopened libcuda" ON)
 if(LIBOMPTARGET_DEP_CUDA_FOUND AND NOT LIBOMPTARGET_FORCE_DLOPEN_LIBCUDA)
   message(STATUS "Building CUDA plugin linked against libcuda")
   target_link_libraries(omptarget.rtl.cuda PRIVATE CUDA::cuda_driver)