add an extension to query device IP version

[bashbaug]

We need to add a query for the "device IP version". This would be a simple addition to zeDeviceGetProperties, and could be added in a similar way as the device LUID extension.

Is there anything else we should add to the same extension while we're at it?

See: https://registry.khronos.org/OpenCL/extensions/intel/cl_intel_device_attribute_query.html

Notes:

• We already have the device ID via ze_device_properties_t.deviceId.
• We already have the number of slices via ze_device_properties_t.numSlices.
• We already have the number of sub-slices per slice via ze_device_properties_t.numSubslicesPerSlice.
• We already have the number of EUs per sub-slice via ze_device_properties_t.numEUsPerSubslice.
• We already have the number of threads per EU via ze_device_properties_t.numThreadsPerEU.
• We already have the dp4a feature flag via ze_device_module_flags_t.ZE_DEVICE_MODULE_FLAG_DP4A.
• We don't appear to have a feature flag for dpas.

[petercad]

@bashbaug Who's in charge of specifying the result of the device IP version query for Intel GPUs?
(I'm assuming that wouldn't be part of the L0 spec.)

[bashbaug]

Yeah, spec-wise I think the returned value would be "implementation-defined".

It looks like our current Intel GPU driver encodes the device major, minor, and revision in the returned value:
https://github.com/intel/compute-runtime/blob/1a28c2a15472c37998f1c586bda0de54f3464bfb/opencl/source/helpers/cl_hw_helper.cpp#L24

e.g.:
https://github.com/intel/compute-runtime/blob/8424b2775470a740c3c2ef1e4ed8b34d9c78a7ec/opencl/source/xe_hpg_core/cl_hw_helper_xe_hpg_core.cpp#L55

We could change this, though it would be nice to keep the property that "newer devices should have a higher version than older devices."

[jandres742]

@bashbaug
from the opencl:

CL_DEVICE_IP_VERSION_INTEL	cl_version	The IP version for the device. The meaning of this version is implementation-defined, but newer devices should have a higher version than older devices.

so that's kind of vague. What is the actual meaning of this? or is just that we want to have L0 returning a similar number as OCL?