Version 0.4.2

Version 0.4.2 includes a lot of work on the search space representation, application of restrictions, and optimization strategies. In addition to the addition of several new optimization strategies, most optimization strategies should see improved performance both in terms of the number of evaluated kernel configurations as well as execution time.

Added

• new optimization strategies: dual annealing, greedly ILS, ordered greedy MLS, greedy MLS
• support for constant memory in cupy backend
• constraint solver to cut down time spent in creating search spaces
• support for custom tuning objectives
• support for max_fevals and time_limit in strategy_options of all strategies

Removed

• alternative Bayesian Optimization strategies that could not be used directly
• C++ wrapper module that was too specific and hardly used

Changed

• string-based restrictions are compiled into functions for improved performance
• genetic algorithm, MLS, ILS, random, and simulated annealing use new search space object
• diff evo, firefly, PSO are initialized using population of all valid configurations
• all strategies except brute_force strictly adhere to max_fevals and time_limit
• simulated annealing adapts annealing schedule to max_fevals if supplied
• minimize, basinhopping, and dual annealing start from a random valid config

Version 0.4.1

This version adds a brand new Bayesian Optimization strategy, as well as some smaller features and fixes.

[0.4.1] - 2021-09-10

Added

• support for PyTorch Tensors as input data type for kernels
• support for smem_args in run_kernel
• support for (lambda) function and string for dynamic shared memory size
• a new Bayesian Optimization strategy

Changed

• optionally store the kernel_string with store_results
• improved reporting of skipped configurations

Version 0.4.0

This version adds a great deal of new functionality and extra flexibility and additional control to the user over what is being benchmarked and when. From the CHANGELOG:

Added

• support for (lambda) function instead of list of strings for restrictions
• support for (lambda) function instead of list for specifying grid divisors
• support for (lambda) function instead of tuple for specifying problem_size
• function to store the top tuning results
• function to create header file with device targets from stored results
• support for using tuning results in PythonKernel
• option to control measurements using observers
• support for NVML tunable parameters
• option to simulate auto-tuning searches from existing cache files
• Cupy backend to support C++ templated CUDA kernels
• support for templated CUDA kernels using PyCUDA backend
• documentation on tunable parameter vocabulary

Version 0.3.2

Version 0.3.2

This version adds several new and recent features. Most importantly is the new feature to specify user-defined metrics for Kernel Tuner to compute along with the benchmarking results. User-defined metrics are composable, so you can define metrics that build upon other metrics. The documentation pages have also been updated to include this new feature and other recent changes.

An important change that might influence benchmark results reported by Kernel Tuner is the fact that the runner will now do a warm up of the device using the first kernel in the parameter space. This is to remove any startup or cold start delays that were significantly slowing down the first benchmarked kernel on many devices.

From the changelog:

[0.3.2] - 2020-11-04

Added

• support loop unrolling using params that start with loop_unroll_factor
• always insert "define kernel_tuner 1" to allow preprocessor ifdef kernel_tuner
• support for user-defined metrics
• support for choosing the optimization starting point x0 for most strategies

Changed

• more compact output is printed to the terminal
• sequential runner runs first kernel in the parameter space to warm up device
• updated tutorials to demonstrate use of user-defined metrics

Version 0.3.1

A small release for 2 small new features and a bugfix for older GPUs.

[0.3.1] - 2020-06-11

Added

• kernelbuilder functionality for including kernels in Python applications
• smem_args option for dynamically allocated shared memory in CUDA kernels

Changed

• bugfix for NVML Error on Nvidia devices without internal current sensor

Version 0.3.0

Version 0.3.0

This is the release of version 0.3.0 of Kernel Tuner. We have done a lot of work on the internals of Kernel Tuner. This release fixes several issues, adds and extends new features, and simplifies the user interface.

[0.3.0] - 2019-12-20

Changed

• fix for output checking, custom verify functions are called just once
• benchmarking now returns multiple results not only time
• more sophisticated implementation of genetic algorithm strategy
• how the "method" option is passed, now use strategy_options

Added

• Bayesian Optimizaton strategy, use strategy="bayes_opt"
• support for kernels that use texture memory in CUDA
• support for measuring energy consumption of CUDA kernels
• option to set strategy_options to pass strategy specific options
• option to cache and restart from tuned kernel configurations cachefile

Removed

• Python 2 support, it may still work but we no longer test for Python 2
• Noodles parallel runner

Version 0.2.0

Version 0.2.0

Version 0.2.0 adds a large number of search optimization algorithms and basic support for testing and tuning Fortran kernels.

Changed

• no longer replacing kernel names with instance strings during tuning
• bugfix in tempfile creation that lead to too many open files error

Added

• A minimal Fortran example and basic Fortran support
• Particle Swarm Optimization strategy, use strategy="pso"
• Simulated Annealing strategy, use strategy="simulated_annealing"
• Firefly Algorithm strategy, use strategy="firefly_algorithm"
• Genetic Algorithm strategy, use strategy="genetic_algorithm"

Version 0.1.9

[0.1.9] - 2018-04-18

Changed

• bugfix for C backend for byte array arguments
• argument type mismatches throw warning instead of exception

Added

• wrapper functionality to wrap C++ functions
• citation file and zenodo doi generation for releases

Version 0.1.8

Version 0.1.8 brings many improvements, mostly focused on user friendliness. The installation process of optional dependencies is simplified as you can now use extras with pip. For example, pip install kernel_tuner[cuda] can be used to install both Kernel Tuner and the optional dependency PyCuda. In addition, Version 0.1.8 introduces many more checks on the user input that you pass to tune_kernel and run_kernel. For example, the kernel source code is parsed to see if the signature matches the argument list. The additional checks on input should make it easier to use and debug programs using Kernel Tuner. For a more detailed overview of the changes, see below:

[0.1.8] - 2017-11-23

Changed

• bugfix for when using iterations smaller than 3
• the install procedure now uses extras, e.g. [cuda,opencl]
• option quiet makes tune_kernel completely quiet
• extensive updates to documentation

Added

• type checking for kernel arguments and answers lists
• checks for reserved keywords in tunable paramters
• checks for whether thread block dimensions are specified
• printing units for measured time with CUDA and OpenCL
• option to print all measured execution times

Version 0.1.7

[0.1.7] - 2017-10-11

Changed

• bugfix install when scipy not present
• bugfix for GPU cleanup when using Noodles runner
• reworked the way strings are handled internally

Added

• option to set compiler name, when using C backend