Convolution Kernel profiling

Jan 26, 2026

I optimized a convolution kernel by removing shared-memory bottlenecks, moving suitable constants to __constant__, and applying load-widening techniques. These changes reduced warp-stalling and increased ALU utilization (approx. 5% → 50%), yielding the measured latency improvement.

Highlights

• Memory strategy: Moved kernel constants to __constant__ where beneficial (intra-warp broadcast).
• Load widening: #pragma unroll and autotuning further alleviated MIO throttle stalls.
• Profiling: Nsight Compute (NCU) for hotspots and SASS/PTX inspection for instruction mix.
• Outcome: Reduced stalls, higher ALU throughput, and small codegen shifts (more integer ops).

See the kernels: ConvProfNCU/kernels and the full analysis: profiling_summary.md.