Added DEVICE_INFO_PROGRAM_SET_SPECIALIZATION_CONSTANTS

[RossBrunton]

THis is similar to the existing
DEVICE_INFO_KERNEL_SET_SPECIALIZATION_CONSTANTS device info, only
it applies to programs rather than kernels. All of the adapters should
be updated to report it correctly, and tests have been added.

[frasercrmck]

CUDA/HIP LGTM

[RossBrunton]

@oneapi-src/unified-runtime-native-cpu-write @oneapi-src/unified-runtime-level-zero-write Can you quickly look at this? I've just set it so that Native CPU and Level Zero don't advertise support for specialization constants.

[coldav]

@oneapi-src/unified-runtime-native-cpu-write @oneapi-src/unified-runtime-level-zero-write Can you quickly look at this? I've just set it so that Native CPU and Level Zero don't advertise support for specialization constants.

It might makes sense to get @PietroGhg to review (not sure if he is still in these groups or not) - we will want specialized constants for native cpu though I think.

[kbenzie]

we will want specialized constants for native cpu though I think.

It doesn't support them today so it should say so to the user.

[coldav]

we will want specialized constants for native cpu though I think.

It doesn't support them today so it should say so to the user.

I thought we did, just that we fail tests.

[RossBrunton]

we will want specialized constants for native cpu though I think.

It doesn't support them today so it should say so to the user.

I thought we did, just that we fail tests.

Both urKernelSetSpecializationConstants and urProgramSetSpecializationConstants return UR_RESULT_ERROR_UNSUPPORTED_FEATURE in native CPU.

[coldav]

we will want specialized constants for native cpu though I think.

It doesn't support them today so it should say so to the user.

I thought we did, just that we fail tests.

Both urKernelSetSpecializationConstants and urProgramSetSpecializationConstants return UR_RESULT_ERROR_UNSUPPORTED_FEATURE in native CPU.

okay that's fine, as long as we can easily switch this back on.

[RossBrunton]

@oneapi-src/unified-runtime-level-zero-write I want to get this merged, can you look at it?

@coldav @PietroGhg It should be as simple as implementing it and returning true for this device info. Once that is done, it should "just work" and (properly written) user code will switch over to using spec constants. If you're okay with it, can I get an approval?

[PietroGhg]

we will want specialized constants for native cpu though I think.

It doesn't support them today so it should say so to the user.

I thought we did, just that we fail tests.

Both urKernelSetSpecializationConstants and urProgramSetSpecializationConstants return UR_RESULT_ERROR_UNSUPPORTED_FEATURE in native CPU.

For Native CPU we are going through the "emulated" path for Spec Constants, which means (I may be wrong, it's been a while since I looked into it) that the values for them are set as kernel arguments by the SYCL runtime, so I think that the values for urKernelSetSpecializationConstants and urProgramSetSpecializationConstants won't affect Native CPU, so this change looks fine to me