[SYCL][Docs] Mark host as deprecated in SYCL_DEVICE_FILTER #7117
Conversation
The host device has been removed and related APIs have been deprecated. This includes the `host` device-type and backend in `SYCL_DEVICE_FILTER`. This commit changes the documentation for `SYCL_DEVICE_FILTER` to mention the deprecation of `host` in the filter and remove any mention of the behavior of using the deprecated values and interfaces. Signed-off-by: Larsen, Steffen <[email protected]>
There was a problem hiding this comment.
The host device has been removed and related APIs have been deprecated.
Hm... I see host device been used in CI quite extensively - https://github.com/intel/llvm/blob/sycl/devops/test_configs.json. Shouldn't we update configuration for llvm-test-suite and CTS here too?
#7160 removes the use of host you pointed out. To my knowledge the llvm-test-suite should have been updated, though intel/llvm-test-suite#1345 removes some left-over mentions of it in the READMEs. CTS I am unsure about. |
|
AFAIK, some of the tests in CTS or llvm-test-suite require at least two devices. |
|
Testing of #7160 is clean though... I guess we can merge and see where it breaks next, unless anybody has any better suggestions where to look. |
|
My wild guess is that the test requiring two devices is skipped if only one device is available. |
I believe you're right. Without host device we don't have many other choices, except maybe trying to use sub-devices, but I am not sure that would be very representative for such tests. |
|
If we don't have any scenario where these tests are going to be executed, does it make sense to have these tests in the testbase? |
|
It really depends on whether or not we ever test with systems that have multiple devices. If we don't then we may indeed run into cases where the tests were broken without us knowing. I don't know if there is much benefit in moving the few of these tests we have into a separate folder. In general I would expect the tests to have their multi-device nature stated as part of their names. |
|
@steffenlarsen, can you check post-commit fail? https://github.com/intel/llvm/actions/runs/3318840937/jobs/5483251069 |
I believe all our GPU systems should be capable of offloading to CPU. Does it make sense to modify these few tests to use GPU and CPU device? |
Actual offender is #7113. It seems to be flaky so if we need to we can revert it, but I am hoping #7154 will fix the issue. |
@steffenlarsen, looks like it keeps failing occasionally. |
Sadly it seems that way. I am investigating in will revert if necessary. |
|
@steffenlarsen, any update on flaky issue? |
I have not yet had any luck reproducing it locally. It may help to add a small patch giving more context to the failure if reproducing it continues to be a problem. |
The host device has been removed and related APIs have been deprecated. This includes the
hostdevice-type and backend inSYCL_DEVICE_FILTER. This commit changes the documentation forSYCL_DEVICE_FILTERto mention the deprecation ofhostin the filter and remove any mention of the behavior of using the deprecated values and interfaces.