Guard CUDA runtime APIs with error checking #266
Conversation
PointKernel
added
type: feature request
| @@ -66,7 +66,7 @@ TEMPLATE_TEST_CASE_SIG("Parallel insert-or-update", | |||
| static constexpr int Blocks = 1024; | |||
| static constexpr int Threads = 128; | |||
| parallel_sum<<<Blocks, Threads>>>(m.get_device_mutable_view()); | |||
| cudaDeviceSynchronize(); | |||
| CUCO_CUDA_TRY(cudaDeviceSynchronize()); | |||
There was a problem hiding this comment.
How does catch2 handle exceptions or failing assertions? Will the test case just fail or will this nuke the whole test suite? As a workaround, we could wrap CUDA calls with a catch2 clause, e.g., REQUIRE(cudaSuccess == cudaDeviceSynchronize())
There was a problem hiding this comment.
REQUIRE would return if a test case fails while CHECK won't stop at the first failure and will complete the whole test suite.
There was a problem hiding this comment.
We can also wrap CUCO_CUDA_TRY(...) with REQUIRE_NOTHROW(CUCO_CUDA_TRY(...))
There was a problem hiding this comment.
How does catch2 handle exceptions or failing assertions? Will the test case just fail or will this nuke the whole test suite?
Just did a quick test, throwing an error will fail the current test case but other tests in the test suite will continue. So the behavior is similar to CHECK_NOTHROW. In that sense, wrapping those calls with catch2 macros is not necessary.
Similar to rapidsai/cudf#12531
This PR adds error checking for CUDA runtime APIs.