Kernel Tuner Ecosystem

Kernel Tuner is a software ecosystem for the creation of highly-optimized GPU applications through auto-tuning.

Optimize GPU kernels efficiently and painlessly from Python

Kernel Tuner simplifies the development of highly-optimized and auto-tuned CUDA, OpenCL, HIP, and C/C++ code. It works as an external tool to benchmark and optimize GPU kernels in isolation. By enabling Python-based unit testing of GPU code and seamless auto-tuning, it allows developers to directly tune user-defined parameters without requiring extensive changes to the original kernel code or introducing new dependencies in the production code.

The rational
Developing optimal GPU programs is complex due to the vast number of potential parameter configurations and hardware specificities. The goal of auto-tuning is to automate finding the best performing setup for a given device. Kernel Tuner bridges this gap by making kernels written in low-level languages accessible from Python, simplifying the auto-tuning of kernel parameters while taking care of the required host-device memory synchronization, ultimately leading to faster and more energy-efficient applications.
Key features
- Test GPU code directly from Python
- Tune parameters in both host and device code
- Output correctness verification of all benchmarked kernels
- Over 20 optimization algorithms (e.g., Simulated Annealing, Genetic Algorithms, Particle Swarm) to drastically speed up the tuning process
- Mixed-precision and accuracy tuning capabilities
- Custom metrics and tuning objectives, including optimizing for energy consumption versus time to solution
Models and platforms support
Supports mainstream GPU programming languages including CUDA, HIP, OpenCL, OpenACC, OpenMP, C, C++, and Fortran. Compatible with Python (>=3.10) and interfaces with multiple backends such as PyCUDA, CuPy, PyOpenCL, and HIP Python.