krnopt-cuda-generic-optimization

To pick the next CUDA optimization experiment, map the diagnosed bottleneck to a specific source region and workload context. This process proposes a single primary experiment with a clear validation signal, enforcing an evidence-driven decision loop for your kernel.

Choosing a CUDA kernel optimization direction requires attributing the profiling bottleneck to a source region. The selection process then maps this evidence to a focused fix family and proposes one testable action to validate the performance improvement.

After bottleneck attribution, you proceed by requesting one primary optimization experiment. The workflow assesses the evidence, selects an appropriate primitive fix family, and defines a specific validation signal to verify the resulting kernel performance change.

Yes, you need a diagnosed hotspot with an attributed source region and workload context before starting. The optimization selection relies on this existing evidence to scope the fix family and propose a single, testable action for your CUDA kernel.

CUDA kernel optimization requires a defined validation signal to enforce an evidence-driven decision loop. By proposing a single testable action with a clear validation metric, you ensure the selected fix family measurably addresses the diagnosed hotspot.