Adding a clEnqueueSetKernelArg() method?

It would be highly beneficial to implement a clEnqueueSetKernelArg() method, particularly for setting scalars or vectors allocated in host memory.

This need arises, for example, during a reduction operation where the resulting scalar must be used in a subsequent kernel. Currently, it's not possible to call clEnqueueNDRangeKernel() until the final value is computed. However, since the address of the host memory is already known, introducing a clEnqueueSetKernelArg() function would eliminate this syncing point, thereby significantly reducing overhead.

What do you think guys? Is there any technical limitation that I do not understand?

Start by reviewing the proposed clEnqueueSetKernelArg() entry point and the described reduction workflow involving clEnqueueNDRangeKernel(). Determine the technical limitations and specification changes required for host-memory scalars or vectors; the issue does not name files or tests, and it does not define an acceptance criterion beyond eliminating the synchronization point.

Written by the indexing model from the issue text.