Open
Description
We've got a few CTS tests that only pass on one or two adapters. The behaviour these tests are enforcing isn't being missed upstream in the backends that are failing, which I think is a sign that either: these tests should have some skipping mechanism, we need a spec adjustment, or in the cases where only one adapter passes we could consider making them adapter specific tests.
These are the ones I noticed while working on the device parameterization PR:
- urEventSetCallbackTest.ValidateParameters has a match entry for every adapter (although the opencl one may be config specific) (Wrap urEventSetCallback when ran through loader #2527 and maybe look into other adapters)
- urEventGetProfilingInfo.Success - UR_PROFILING_INFO_COMMAND_COMPLETE fails on all adapters apart from CL, maybe just needs to be optional (Make profiling info optional and update tests #2533 )
- urProgramSetSpecializationConstants InvalidValueSize, InvalidValueId and InvalidValuePtr have match entries for all adapters apart form CL - likely there's a missing skip for these one because cuda and hip don't support the entry point at all (Added
DEVICE_INFO_PROGRAM_SET_SPECIALIZATION_CONSTANTS#2539 ) - urUSM*Alloc InvalidUSMSize tests all fail for cuda, hip and native cpu - might simply not be possible to check this on these backends ( Add size validation for USM device/shared allocs #2557 )
- urDeviceGetGlobalTimestampTest SuccessSynchronizedTime fails on every adapter, only passes on CL CPU ( Be less strict with timer test #2547 )
- urEnqueueKernelLaunchKernelSubGroupTest fails on every adapter but CL
- urEnqueueKernelLaunchTest.InvalidKernelArgs fails on every adapter but CL ( Update spec to make kernel validation optional #2564 )
- urMemBufferCreateWithNativeHandleTest has a lot of match entries, all native handle entry points should be optional so almost certainly just missing a skip
( Update MemBufferCreate with unsupported checks #2562 )