KPerfIR: Towards a Open and Compiler-centric Ecosystem for GPU Kernel Performance...
summary
KPerfIR is a compiler‑centric infrastructure integrated into Triton that instruments a record op across a multi‑layer IR, adds compiler passes for automatic instrumentation, and ships a runtime for trace collection and trace replay. It enables a region-based timing tool that reveals intra‑kernel producer/consumer bubbles; fixing those made a FlashAttention kernel exceed a hand‑tuned CUDA version. Reported costs are low (≈8.2% runtime overhead) with accurate timing (≈2% relative error) and portability across GPU backends.
AI-written summary · about
chapters (8)
- 0:00KPerfIR introduction
- 1:21GPU hardware and programming models
- 2:47AI compilers challenges
- 4:33Compiler-centric profiling design
- 5:58KPerfIR implementation
- 7:44Region-based timing example
- 10:08KPerfIR performance evaluation
- 12:38Takeaways and future vision
related talks
29mAuto-instrumentation for GPU performance using eBPF
08mGPU Programming with Triton Kernels
28mAuto-instrumentation for GPU Performance Using eBPF
25mLeveraging AI Ecosystem for Portable and Sustainable GPU Kernels in HPC
14mCrystalPerf: Learning to Characterize the Performance of Dataflow Computation...
32mA Practical Guide To Benchmarking AI and GPU Workloads in Kubernetes
34mBridging the Gap: Compiling and Optimizing Triton Kernels Onto RISC-V Targets Based on MLIR
16mLightning Talk: Triton Compiler