gpu: nvidia: skip unsupported gtests #2202
Conversation
sgeor255
force-pushed
the
fix-nvidia-gtests
branch
from
5143029 to
2861953
Compare
tests/gtests/sycl/api/CMakeLists.txt
Outdated
| @@ -28,7 +28,7 @@ if(DNNL_WITH_SYCL) | |||
| endif() | |||
|
|
|||
| # Enable linking SYCL kernels. | |||
| if(DNNL_SYCL_CUDA OR (DNNL_SYCL_GENERIC AND NVIDIA_TARGET_SUPPORTED)) | |||
| if(DNNL_SYCL_GENERIC AND NVIDIA_TARGET_SUPPORTED) | |||
There was a problem hiding this comment.
Could you explain, please, why DNNL_SYCL_CUDA is removed?
There was a problem hiding this comment.
The tests Simple/sycl_memory_buffer_test.InteropReorderAndUserKernel/cpu and Simple/sycl_memory_buffer_test.EltwiseWithUserKernel/cpu run with a CPU engine and try to run a SYCL kernel on CPU. They currently fail because the nvidia target triple is set. This results in a UR_ERROR_INVALID_BINARY. I am not sure if this is intended behaviour or the tests should be skipped, but removing the DNNL_SYCL_CUDA from the check here allows them to run successfully on CPU as it results in not setting the nvidia target triple.
There was a problem hiding this comment.
What you are saying is that target shouldn't be set for CPU part of the test. You removed linker options for pure Nvidia target, but how this change would help for a generic case targeting Nvidia backend?
There was a problem hiding this comment.
@sgeor255, I think the problem arises due to overriding the default SYCL targets with -fsycl-targets=nvptx64-nvidia-cuda which throws off the default SYCL target, which is spir64. We could try to use both targets here as we expect the kernels to be compiled for CPU as well:
-fsycl-targets=nvptx64-nvidia-cuda,spir64
There was a problem hiding this comment.
That's a good idea @densamoilov , thanks. Added the change.
sgeor255
force-pushed
the
fix-nvidia-gtests
branch
from
2861953 to
e57e7a6
Compare
|
make test |
Description
This PR adds skips in gtests for unimplemented primitives on nvidia.