Awesome

Create optimized GPU applications in any mainstream GPU programming language (CUDA, HIP, OpenCL, OpenACC).

What Kernel Tuner does:

• Works as an external tool to benchmark and optimize GPU kernels in isolation
• Can be used directly on existing kernel code without extensive changes
• Can be used with applications in any host programming language
• Blazing fast search space construction
• More than 20 optimization algorithms to speedup tuning
• Energy measurements and optimizations (power capping, clock frequency tuning)
• ... and much more! For example, caching, output verification, tuning host and device code, user defined metrics, see the full documentation.

Installation

• First, make sure you have your CUDA, OpenCL, or HIP compiler installed
• Then type: pip install kernel_tuner[cuda], pip install kernel_tuner[opencl], or pip install kernel_tuner[hip]
• or why not all of them: pip install kernel_tuner[cuda,opencl,hip]

More information on installation, also for other languages, in the installation guide.

Example

import numpy as np
from kernel_tuner import tune_kernel

kernel_string = """
__global__ void vector_add(float *c, float *a, float *b, int n) {
    int i = blockIdx.x * block_size_x + threadIdx.x;
    if (i<n) {
        c[i] = a[i] + b[i];
    }
}
"""

n = np.int32(10000000)

a = np.random.randn(n).astype(np.float32)
b = np.random.randn(n).astype(np.float32)
c = np.zeros_like(a)

args = [c, a, b, n]

tune_params = {"block_size_x": [32, 64, 128, 256, 512]}

tune_kernel("vector_add", kernel_string, n, args, tune_params)

More examples here.

Resources

• Full documentation
• Guides:
  • Getting Started
  • Convolution
  • Diffusion
  • Matrix Multiplication
• Features & Use cases:
  • Full list of examples
  • Output verification
  • Test GPU code from Python
  • Tune code in both host and device code
  • Optimization algorithms
  • Mixed-precision & Accuracy tuning
  • Custom metrics & tuning objectives
• Kernel Tuner Tutorial slides [PDF], hands-on:
  • Vector add example [.ipynb]
  • Tuning thread block dimensions [.ipynb]
  • Search space restrictions & output verification [.ipynb]
  • Visualization & search space optimization [.ipynb]
• Energy Efficient GPU Computing tutorial slides [PDF], hands-on:
  • Kernel Tuner for GPU energy measurements [.ipynb]
  • Code optimizations for energy [.ipynb]
  • Mixed precision and accuracy tuning [.ipynb]
  • Optimzing for time vs for energy [.ipynb]

Kernel Tuner ecosystem

<img width="250px" src="https://raw.githubusercontent.com/KernelTuner/kernel_tuner/master/doc/images/kernel_launcher.png"/><br />C++ magic to integrate auto-tuned kernels into C++ applications

<img width="250px" src="https://raw.githubusercontent.com/KernelTuner/kernel_tuner/master/doc/images/kernel_float.png"/><br />C++ data types for mixed-precision CUDA kernel programming

<img width="275px" src="https://raw.githubusercontent.com/KernelTuner/kernel_tuner/master/doc/images/kernel_dashboard.png"/><br />Monitor, analyze, and visualize auto-tuning runs

Communication & Contribution

• GitHub Issues: Bug reports, install issues, feature requests, work in progress
• GitHub Discussion group: General questions, Q&A, thoughts

Contributions are welcome! For feature requests, bug reports, or usage problems, please feel free to create an issue. For more extensive contributions, check the contribution guide.

Citation

If you use Kernel Tuner in research or research software, please cite the most relevant among the publications on Kernel Tuner. To refer to the project as a whole, please cite:

@article{kerneltuner,
  author  = {Ben van Werkhoven},
  title   = {Kernel Tuner: A search-optimizing GPU code auto-tuner},
  journal = {Future Generation Computer Systems},
  year = {2019},
  volume  = {90},
  pages = {347-358},
  url = {https://www.sciencedirect.com/science/article/pii/S0167739X18313359},
  doi = {https://doi.org/10.1016/j.future.2018.08.004}
}