| Issue |
71747
|
| Summary |
[libomptarget][tests] Maximum assumed device heap size.
|
| Labels |
openmp:libomp
|
| Assignees |
jdoerfert
|
| Reporter |
Meinersbur
|
Despite the GPU of the [openmp-offload-cuda-project](https://lab.llvm.org/staging/#/builders/68) and [openmp-offload-cuda-runtime](https://lab.llvm.org/staging/#/builders/7) buildbots having 4 GiB of memory, the default device heap size as returned by `cuCtxGetLimit(CU_LIMIT_MALLOC_HEAP_SIZE, ...)` is 8388608 (8 MiB), while [offloading/malloc.c](https://github.com/llvm/llvm-project/blob/main/openmp/libomptarget/test/offloading/malloc.c) requires ~55MB and [offloading/malloc_parallel.c](https://github.com/llvm/llvm-project/blob/main/openmp/libomptarget/test/offloading/malloc_parallel.c) even more.
I don't know how CUDA determines the default heap size limit, but I assume it is constant and inherited from the earliest days of CUDA.
To fix this, we either
1. limit the amount of heap allocated by any test to 8 MiB (e.g. reducing the number of teams in parallel.c to 48), or
2. set `LIBOMPTARGET_HEAP_SIZE` to the maximum heap size allocated by any test. This patch fixes the two malloc tests, reducing the number of failed tests to 30:
```
diff --git a/openmp/libomptarget/test/lit.cfg b/openmp/libomptarget/test/lit.cfg
index 6dab31bd35a9..e288827c50f6 100644
--- a/openmp/libomptarget/test/lit.cfg
+++ b/openmp/libomptarget/test/lit.cfg
@@ -31,6 +31,8 @@ if 'LIBOMPTARGET_LOCK_MAPPED_HOST_BUFFERS' in os.environ:
if 'OMP_TARGET_OFFLOAD' in os.environ:
config.environment['OMP_TARGET_OFFLOAD'] = os.environ['OMP_TARGET_OFFLOAD']
+config.environment['LIBOMPTARGET_HEAP_SIZE'] = '134217728' # 128 MiB
+
# set default environment variables for test
if 'CHECK_OPENMP_ENV' in os.environ:
test_env = os.environ['CHECK_OPENMP_ENV'].split()
```
A 64 MiB heap is sufficient for the malloc.c test, but not for malloc_parallel.c.
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
