Kernel Tuner Examples

Most of the examples show how to use Kernel Tuner to tune a CUDA, OpenCL, or C kernel, while demonstrating a particular usecase of Kernel Tuner.

Except for test_vector_add.py and test_vector_add_parameterized.py, which show how to write tests for GPU kernels with Kernel Tuner.

Below we list the example applications and the features they illustrate.

Vector Add

[CUDA] [CUDA-C++] [OpenCL] [C] [Fortran]

• use Kernel Tuner to tune a simple kernel

Stencil

[CUDA] [OpenCL]

• use a 2-dimensional problem domain with 2-dimensional thread blocks in a simple and clean example

Matrix Multiplication

[CUDA] [OpenCL]

• pass a filename instead of a string with code
• use 2-dimensional thread blocks and tiling in both dimensions
• tell Kernel Tuner to compute the grid dimensions for 2D thread blocks with tiling
• use the restrictions option to limit the search to only valid configurations
• use a user-defined performance metric like GFLOP/s

Convolution

There are several different examples centered around the convolution kernel [CUDA] [OpenCL]

convolution.py

[CUDA] [OpenCL]

• use tunable parameters for tuning for multiple input sizes
• pass constant memory arguments to the kernel
• write output to a json file

sepconv.py

[CUDA] [OpenCL]

• use the convolution kernel for separable filters
• write output to a csv file using Pandas

convolution_correct.py

[CUDA] [OpenCL]

• use run_kernel to compute a reference answer
• verify the output of every benchmarked kernel

convolution_streams.py

[CUDA]

• allocate page-locked host memory from Python
• overlap transfers to and from the GPU with computation
• tune parameters in the host code in combination with those in the kernel
• use the lang="C" option and set compiler options
• pass a list of filenames instead of strings with kernel code

Reduction

[CUDA] [OpenCL]

• use vector types and shuffle instructions (shuffle is only available in CUDA)
• tune the number of thread blocks the kernel is executed with
• tune the partial loop unrolling factor of a for-loop
• tune pipeline that consists of two kernels
• tune with custom output verification function

Sparse Matrix Vector Multiplication

[CUDA]

• use scipy to compute a reference answer and verify all benchmarked kernels
• express that the number of thread blocks depends on the values of tunable parameters

Point-in-Polygon

[CUDA]

• overlap transfers with device mapped host memory
• tune on different implementations of an algorithm

ExpDist

[CUDA]

• in-thread block 2D reduction using CUB library
• C++ in CUDA kernel code
• tune multiple kernels in pipeline

Code Generator

[CUDA] [OpenCL]

• use a Python function as a code generator