perf-workload-profiling

What problem does it solve?

Code instrumentation for timing GPU workloads, enabling precise measurements of per-iteration latency, throughput, and data loading for training loops, as well as standalone kernel timing with warmup and per-iteration statistics; NVTX labeling for profiler timelines is also supported.

Core Features & Use Cases

• Manual timing for training loops to report per-iteration latency and throughput, with data-load timing.
• CUDA event-based timing for standalone kernels or ops, with warmup and per-iteration statistics.
• NVTX annotations to label profiler timelines for clearer visualization.

Quick Start

Provide a timing harness for your training loop or kernel to collect per-iteration latency, throughput, and data-load metrics with NVTX labeling.