Some OpenCL platforms have a SIMDwarpwavefront-based execution model that execute groups of threads in lockstep, permitting us to perform cross-thread synchronisation within each such group without the use of barriers. Unfortunately, there seems to be no reliable way to query these sizes at runtime. Instead, we use builtin tables to figure out which size we should use for a specific platform and device. If nothing matches here, the wave size should be set to one.

We also use this to select reasonable default group sizes and group counts.