CUDA Kernel Optimization and Counter-Free Performance Analysis for Depthwise Convolution in Cloud Environments

Researchers Huriyeh Babak and Melanie Schaller have published a paper on arXiv detailing CUDA kernel optimization for depthwise convolution operators used in Structured State Space Model Convolutional Diagonal (S4ConvD). The study, submitted to IEEE TPDS, evaluates four kernel variants: naive, global-memory-coalesced, shared-memory cache-blocked, and warp-tiled. The warp-tiled kernel achieved a 3.26x runtime reduction compared to the naive baseline, with end-to-end training speedup reaching 1.29x. The work introduces a counter-free performance analysis methodology that combines CUDA-event timing, execution-path decomposition, memory-traffic modeling, effective-bandwidth estimation, and roofline analysis, enabling profiling insights without hardware performance counters.

Forward and input-gradient paths benefited substantially from improved locality and on-chip data reuse, while the reduction-dominated weight-gradient path remained the primary bottleneck. The results demonstrate that meaningful architecture-level GPU kernel analysis can be performed reproducibly in restricted cloud environments without privileged profiling access. This approach is particularly relevant for professionals tuning AI workloads in cloud settings where hardware counters are often unavailable. The paper provides a controlled operator-level study with fixed operator, model, dataset, and training configuration, varying only the CUDA kernel implementation across forward, input-gradient, and weight-gradient execution paths under steady-state training conditions.