[SYCL][CUDA] Eliminate incorrect assertions and enable cuda usm tests #2557
Conversation
Signed-off-by: James Brodman <james.brodman@intel.com>
Signed-off-by: James Brodman <james.brodman@intel.com>
Signed-off-by: James Brodman <james.brodman@intel.com>
|
SYCL :: basic_tests/queue/queue_parallel_for_generic.cpp now works as well. |
Signed-off-by: James Brodman <james.brodman@intel.com>
|
Anyone know what these new XPASSes are about? |
Looks like you are right and the issue is gone with the new GPU driver. I've created a PR with removing XFAILs to check this #2564 |
| @@ -4133,7 +4133,7 @@ pi_result cuda_piextUSMHostAlloc(void **result_ptr, pi_context context, | |||
| } catch (pi_result error) { | |||
| result = error; | |||
| } | |||
| assert(reinterpret_cast<std::uintptr_t>(*result_ptr) % alignment == 0); | |||
There was a problem hiding this comment.
I see the issue when alignment == 0 - but removing it there are no checks that the allocation respects the alignment ?
There was a problem hiding this comment.
@fwyzard, do you suggest adding something like:
assert(alignment && (reinterpret_cast<std::uintptr_t>(*result_ptr) % alignment == 0));There was a problem hiding this comment.
I was actually thinking
assert(alignment == 0 || reinterpret_cast<std::uintptr_t>(*result_ptr) % alignment == 0);since alignment == 0 means no alignment requirements IIRC ?
There was a problem hiding this comment.
Right. This seems to be an easy change: the tests are already enabled, so it should be just three new lines of code in pi_cuda.cpp.
Cuda plugin basically had divide by 0 assertions that weren't necessary.
Signed-off-by: James Brodman james.brodman@intel.com