[UR][L0] Add L0 teardown notification support and simply teardown with proxy loader support

[nrspruit]

• Add support for L0 teardown notification in the static L0 loader.
• Add loading of the L0 dynamic loader in the windows proxy loader to ensure that the L0 dynamic loader is usable during teardown.

[nrspruit]

related to oneapi-src/level-zero#333

[cperkinsintel]

@nrspruit - the test failures reported here seem legit. But they are strange.

The buffer_create.cpp test works fine for me if I build and run it manually. Everything working exactly as it should. But when run from llvm-lit, I can get this output:

# | 
# | ZE ---> zeDeviceGet(ZeDrivers[I], &ZeDeviceCount, nullptr)
# | ZE ---> zeDeviceGet(ZeDrivers[I], &ZeDeviceCount, ZeDevices.data())
# | ZE ---> zeDeviceGetProperties(ZeDevices[D], &device_properties)
# | ZE ---> zeDriverGetApiVersion(ZeDriver, &ZeApiVersion)
# | ZE ---> zeDriverGetExtensionProperties(ZeDriver, &Count, nullptr)
# | ZE ---> zeDriverGetExtensionProperties(ZeDriver, &Count, ZeExtensions.data())
# | ZE ---> zelLoaderTranslateHandle(ZEL_HANDLE_DRIVE
# | ...
# `---data was truncated--------

The data was truncated comes from llvm-lit. Essentially, there seemed to have been an error when zelLoaderTranslateHandle was called. I think the next argument it is supposed to log is the "ZeDriver", so there might be something going on. It didn't even output the last "R"

The test failed on Arc, I think. But I can repro the failure on pvc-01.

[nrspruit]

@nrspruit - the test failures reported here seem legit. But they are strange.

The buffer_create.cpp test works fine for me if I build and run it manually. Everything working exactly as it should. But when run from llvm-lit, I can get this output:

# | 
# | ZE ---> zeDeviceGet(ZeDrivers[I], &ZeDeviceCount, nullptr)
# | ZE ---> zeDeviceGet(ZeDrivers[I], &ZeDeviceCount, ZeDevices.data())
# | ZE ---> zeDeviceGetProperties(ZeDevices[D], &device_properties)
# | ZE ---> zeDriverGetApiVersion(ZeDriver, &ZeApiVersion)
# | ZE ---> zeDriverGetExtensionProperties(ZeDriver, &Count, nullptr)
# | ZE ---> zeDriverGetExtensionProperties(ZeDriver, &Count, ZeExtensions.data())
# | ZE ---> zelLoaderTranslateHandle(ZEL_HANDLE_DRIVE
# | ...
# `---data was truncated--------

The data was truncated comes from llvm-lit. Essentially, there seemed to have been an error when zelLoaderTranslateHandle was called. I think the next argument it is supposed to log is the "ZeDriver", so there might be something going on. It didn't even output the last "R"

The test failed on Arc, I think. But I can repro the failure on pvc-01.

that error has existed since before this change, so it may be a real issue, but it is not part of this PR.

See here:
https://github.com/intel/llvm/actions/runs/15042177600/job/42277602962

[cperkinsintel]

Test failures are known: #18463

[gmlueck]

I assume this PR has some sort of merge problem? Seems like a lot of unrelated things changed.

[nrspruit]

I assume this PR has some sort of merge problem? Seems like a lot of unrelated things changed.

the rebase failed, let me fix

[nrspruit]

I assume this PR has some sort of merge problem? Seems like a lot of unrelated things changed.

the rebase failed, let me fix

fixed, for some reason the pulldown was included with I rebased.

[jchlanda]

Suggested change

	// unloaded prematurely. the Only trusted loader is the one that is loaded
	// unloaded prematurely. The only trusted loader is the one that is loaded

[jchlanda]

I guess cuda reviewers were pulled in because of the rebase problem, but still, LGTM :)

[maarquitos14]

LGTM.

[nrspruit]

@intel/llvm-gatekeepers , please merge when available, this enables usage of an improved short term and long term solution to the teardown issues in SYCL when using L0.