| Issue |
56453
|
| Summary |
OpenMP 5.0 target offloading: omp_alloc() undefined in nvlink step
|
| Labels |
|
| Assignees |
|
| Reporter |
vincentadamthefirst
|
I am currently trying to allocate memory on a target device in OpenMP offloading. For this I use a recently built Clang-15 and CUDA 11.6 (bundled with nvhpc 22.3).
A simple test program:
```c++
#include <omp.h>
int main() {
#pragma omp target teams distribute
for (int index = 0; index < 100; index++) {
float *shm = (float *) omp_alloc(20 * 20 * sizeof(float), omp_pteam_mem_alloc);
#pragma omp parallel num_threads(20 * 20) shared(shm) default(none)
{
int threadNum = omp_get_thread_num();
shm[threadNum] = threadNum;
#pragma omp barrier
// some work on shared memory & write-back
}
}
}
```
Compiled using
```bash
clang++ -O3 -std=c++17 -fopenmp -fopenmp-targets=nvptx64-nvidia-cuda -fopenmp-version=51 --cuda-path=/opt/nvidia/hpc_sdk/Linux_x86_64/22.3/cuda/11.6/ main.cpp
```
I expect a successful compilation but I get
```console
nvlink error : Undefined reference to 'omp_alloc' in '/tmp/test-008605-nvptx64-nvidia-cuda-sm_80-400196.cubin'
/usr/local/bin/clang-linker-wrapper: error: 'nvlink' failed
clang-15: error: linker command failed with exit code 1 (use -v to see invocation)
```
when it comes to the linking step. Am I missing something crucial? Is `omp_alloc()` not fully supported yet? (According to [the docs](https://clang.llvm.org/docs/OpenMPSupport.html#openmp-5-0-implementation-details) it should be). I have no problems when using the directives that more or less work the same (`#pragma omp allocate(shm) allocator(omp_pteam_mem_alloc)`)...
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
