Function optimal_block_size 

pub fn optimal_block_size<F>(
    registers_per_thread: u32,
    shared_mem_per_block: F,
    limits: &SmResourceLimits,
) -> Result<(u32, OccupancyResult), GpuOptimError>
where
    F: Fn(u32) -> usize,

Search for the block size that maximises occupancy (à la cudaOccupancyMaxPotentialBlockSize).

Candidate block sizes are every multiple of limits.warp_size from warp_size up to and including limits.max_threads_per_block. The usage model is intentionally simple and documented:

• registers_per_thread is a constant independent of block size (the usual assumption — register usage is a property of the compiled kernel).
• shared_mem_per_block is a closure block_size -> bytes, which covers both a constant footprint (|_| BYTES) and block-size-dependent allocations such as a reduction kernel’s |threads| threads as usize * size_of::<f32>().

Returns the (block_size, occupancy_result) with the highest occupancy. Ties are broken toward the larger block size (i.e. fewer resident blocks), matching the CUDA runtime’s preference.