[SYCL][CUDA] Returns minimum mandated capabilities for atomic_fence device queries

[maarquitos14]

Currently, we were returning an error because it was unimplemented. I believe it makes more sense to return minimum mandated capabilities, as we do in other backends (e.g. HIP).

[hdelan]

Unfortunately the minimum mandated capabilities are not supported by all CUDA arches (especially acq_rel, which is not supported on GTX 1050, for instance). In the PR for HIP atomics I didn't include acq_rel for the same reason https://github.com/intel/llvm/pull/8003/files#diff-b88f648055cd14bce67433ad94075c14ff434a79d971857da7cf1a5daf596dd4R1855 Would you consider removing acq_rel?

[hdelan]

This is beyond the scope of this PR, but if we want this query to return useful and correct information relating to a particular device we may have to implement some logic by hand to return the correct capabilities for each correct arch.

I am also unsure of subgroup level atomics. Do all CUDA arches support this (natively)?

[maarquitos14]

This is beyond the scope of this PR, but if we want this query to return useful and correct information relating to a particular device we may have to implement some logic by hand to return the correct capabilities for each correct arch.

I am also unsure of subgroup level atomics. Do all CUDA arches support this (natively)?

Totally agree, as soon as I get this merged I will open an issue for that.

[maarquitos14]

Unfortunately the minimum mandated capabilities are not supported by all CUDA arches (especially acq_rel, which is not supported on GTX 1050, for instance). In the PR for HIP atomics I didn't include acq_rel for the same reason https://github.com/intel/llvm/pull/8003/files#diff-b88f648055cd14bce67433ad94075c14ff434a79d971857da7cf1a5daf596dd4R1855 Would you consider removing acq_rel?

I'm not sure we can return something including less than minimum mandated capabilities, to be honest. @gmlueck what do you think?

[gmlueck]

I'm not sure we can return something including less than minimum mandated capabilities, to be honest. @gmlueck what do you think?

No, you cannot. All SYCL devices must support at least the required minimum capabilities when the spec says this.

I think you are asking about info::device::atomic_fence_order_capabilities here, correct? If the hardware does not natively support an acq_rel fence operation, can't you implement it with separate acquire and release fences?

[maarquitos14]

I'm not sure we can return something including less than minimum mandated capabilities, to be honest. @gmlueck what do you think?

No, you cannot. All SYCL devices must support at least the required minimum capabilities when the spec says this.

I think you are asking about info::device::atomic_fence_order_capabilities here, correct? If the hardware does not natively support an acq_rel fence operation, can't you implement it with separate acquire and release fences?

Honestly, I'm not sure if @hdelan is talking about info::device::atomic_fence_order_capabilities. I just checked HIP code and the query for info::device::atomic_fence_order_capabilities is returning minimum mandated capabilities (including acq_rel), but info::device::atomic_memory_order_capabilities is the one not including acq_rel. @hdelan can you confirm?

[gmlueck]

but info::device::atomic_memory_order_capabilities is the one not including acq_rel.

The SYCL spec has extremely minimal requirements for this query. The only requirement is that the device must support the relaxed memory order (which is basically a no-op).

[hdelan]

Aha sorry I mistakenly conflated this with atomic_memory_order_capabilities, apologies. LGTM

[npmiller]

LGTM