What is the way to lead the OpenCL SDK compiler to reduce kernel logic utilization ?

There is another version of the optimization guide coming that will discuss this more but here are some attributes you can try:

max_share_resources (x)

num_share_resources (y)

max_share_resources tells the compiler to attempt to reuse common portions of the compute unit if it *does not* affect the throughput of the kernel

num_share_resources tells the compiler to attempt to reuse common portions of the compute unit *regardless* if it affects the throughput of the kernel

By the sounds of it you will probably want to use num_share_resources. This attribute gets expored when you run -O3 but you can apply it manually to avoid having to recompile if you already have a specific performance target in mind. There is another attibute called "max_unroll_loops(x)" were you can put a limit on how much unrolling occurs with the kernel (this is also explored with the -O3 flag)