Under what conditions does kernel fusion not help, or e...

If an operation is compute-bound (above the ridge point), the system is waiting on compute cycles rather than memory bandwidth; therefore, fusing memory-bound operations around it does not alleviate the primary bottleneck. Furthermore, fusion can negatively impact performance if it increases register pressure within a warp to the point where registers spill into local memory, which introduces additional latency. Always validate these optimizations with a profiler rather than relying solely on rules of thumb.