Support buffer_location for shared/host USM allocations

[bsyrowik]

This change adds a new memory property to the aocl_mmd.h file to facilitate passing of buffer location information from the OpenCL runtime to the MMD USM allocation APIs.

The IP authoring project requires this change: USM allocations for memory-mapped host interfaces will have the buffer_location property set, and this should be propagated to the simulation MMD to ensure appropriate allocations are made (e.g. with the correct address bits set).

This complements the following SYCL runtime changes:

• [SYCL][L0][CUDA][HIP] Allow empty properties in backends llvm#6225
• [SYCL] Support USM buffer location property in malloc_shared llvm#6218
• [SYCL] Support USM buffer location property in malloc_host llvm#6220
• [SYCL] Align USM buffer location implementation for malloc_device with malloc_shared llvm#6269
• [SYCL] Query USM buffer location extension only if property is passed to malloc_shared llvm#6268

[pcolberg]

Thanks Bain! Could you provide a brief motivation of the change in the commit message? Why is the change needed, what flows does it apply to (only IP authoring in simulation?), etc.

[pcolberg]

I suggest we add the runtime implementation to this PR that demonstrates how the new API would work. The idea is to pass the newly received buffer_location property in clSharedMemAllocINTEL() and clHostMemAllocINTEL() through to the MMD using this new attribute, correct? Besides that, is any additional logic needed in these runtime functions?

[bsyrowik]

I suggest we add the runtime implementation to this PR that demonstrate how the new API would work. The idea is to pass the newly received buffer_location property in clSharedMemAllocINTEL() and clHostMemAllocINTEL() through to the MMD using this new attribute, correct? Besides that, is any additional logic needed in these runtime functions?

Sounds like a good plan! Yes, you are correct. And no, I don't think there is any additional logic required. :)

[bsyrowik]

Alright, I think we're good to go now. @tiwaria1 @pcolberg - please let me know if there are any other changes you'd like :)

[pcolberg]

Thanks @bsyrowik, looks good!

For clDeviceMemAllocINTEL(), commit f79a9c2 includes a unit test that queries and verifies the buffer location using clGetMemAllocInfoINTEL(). However, the buffer location is stored within a cl_mem struct, which is populated in clDeviceMemAllocINTEL() but not in clHostMemAllocINTEL() and clSharedMemAllocINTEL(). I wonder if we should add a mem_id field to acl_usm_allocation_t to persist the buffer location for later query?

Wait for integration testing before approval

[pcolberg]

@bsyrowik I have made a slight modification to store the properties in an array on the stack similar to intel/llvm#6220 and intel/llvm#6218. Once this PR passes integration testing, this is ready to be merged.