[SYCLcompat] Add support for device attribute MAX_PITCH and ASYNC_ENGINE_COUNT

[the-slow-one]

CU_DEVICE_ATTRIBUTE_AYSNC_ENGINE_COUNT is associated with value 2 for bidirectional data transfer between host and device while executing the kernel and is associated to 1 in case of unidirectional data transfer and for no support of data transfer and parallel execution it's 0.

CU_DEVICE_ATTRIBUTE_MAX_PITCH return INT_MAX, indicating there is no limits! Intel GPU shows the same behavior too!

Signed-off-by: Deepak Raj H R

[zhiweij1]

Please don't mention CUDA-related text here.

[GeorgeWeb]

Minor comment, otherwise syclcompat looks okay.

[GeorgeWeb]

I think you may prefer to use numeric_limits<int>::max() from the C++ <limits> library instead which is equivalent to INT_MAX from <climits>.

Also, while this compiles fine due to limits probably being included indirectly from another include, it is better to include the library/headers at the top of this file explicitly.

[JackAKirk]

CU_DEVICE_ATTRIBUTE_AYSNC_ENGINE_COUNT is associated with value 2 for bidirectional data transfer between host and device while executing the kernel and is associated to 1 in case of unidirectional data transfer and for no support of data transfer and parallel execution it's 0.

I may be misremembering, but I don't think the above is correct. IIRC querying CU_DEVICE_ATTRIBUTE_ASYNC_ENGINE_COUNT can return greater than 2 depending upon the NVLINK capabilities.

In any case I don't understand why you are always returning 0 for get_async_engine_count if you want to map it to the CU_DEVICE_ATTRIBUTE_ASYNC_ENGINE_COUNT device query?
The correct way of dealing with such device queries in a future-proof manner for all existing and future devices is to add a query in urDeviceGetInfo that calls cuDeviceGetAttribute using CU_DEVICE_ATTRIBUTE_ASYNC_ENGINE_COUNT https://github.com/oneapi-src/unified-runtime/blob/main/source/adapters/cuda/device.cpp#L41
and then calling this UR API via the sycl compat entry point using a mapped CU_DEVICE_ATTRIBUTE_ASYNC_ENGINE_COUNT query.

It may also have different mappings to other backends (e.g. AMD), and the call to urDeviceGetInfo allows such backends to return an appropriate value.

[GeorgeWeb]

Apart from the minor comment I left earlier, looking at this again, @JackAKirk is right that it seems a little strange to return 0 for all backends from get_async_engine_count when this should be correctly queried via device::get_info which calls into the respective backend query. I am still okay with this if the above stated may not be needed or if there's a plan to follow up with it in case it is needed. Thanks!