Optimize Kernels

• Choose grid/block sizes based on occupancy and memory patterns
• Use shared memory for data reuse; avoid bank conflicts
• Ensure coalesced global memory accesses
• Minimize divergence; restructure conditionals
• Use profiler-guided iteration (Nsight Systems/Compute)