KernelTuner/kernel_tuner

Kernel Tuner

What it solves

Kernel Tuner is an external tool designed to benchmark and optimize GPU kernels in isolation. It removes the manual trial-and-error process of tuning GPU code to achieve maximum performance, allowing developers to find the best parameters for their kernels across different hardware and programming languages.

How it works

The tool works by treating the GPU kernel as a search space of possible parameters (such as block sizes). It then uses one of over 20 optimization algorithms to search through these parameters to find the configuration that yields the best performance. It can be integrated into existing code without extensive changes and supports multiple GPU programming languages including CUDA, HIP, OpenCL, OpenACC, and OpenMP.

Who it’s for

Developers and researchers who write high-performance GPU applications and need to optimize their kernels for speed, energy efficiency, or accuracy (mixed-precision).

Highlights

  • Multi-language support: Works with CUDA, HIP, OpenCL, OpenACC, and OpenMP.
  • Broad optimization goals: Supports tuning for execution time, energy consumption (via power capping and clock frequency tuning), and accuracy.
  • Extensive algorithm library: Includes over 20 different optimization algorithms to speed up the tuning process.
  • Verification: Includes built-in output verification to ensure that optimized kernels remain correct.
  • Flexible integration: Can be used with applications written in any host programming language.

Related

  • Project
    cfregly/ai-performance-engineering

    A comprehensive collection of code and resources for optimizing AI systems, focusing on GPU tuning, distributed training, and high-throughput inference scaling.

  • Project
    ekondis/mixbench

    A benchmark tool for evaluating the performance bounds of GPUs and CPUs by testing mixed operational intensity kernels across various precision levels.

  • Project
    gpu-mode/reference-kernels

    A collection of reference implementations and specifications for GPU kernel competitions, used to verify the correctness of custom kernels.

  • Project
    TongmingLAIC/AKO4ALL

    AKO4ALL is an automated GPU kernel optimization tool packaged as a Claude Code skill that iteratively profiles, rewrites, and benchmarks kernels to achieve expert-level performance.

  • Project
    ROCm/composable_kernel

    A library providing a programming model for writing performance-critical machine learning kernels that are portable across different GPU and CPU architectures.