minhhn2910 / cuda-half2

Convert CUDA programs from float data type to half or half2 with SIMDization

Half datatype conversion tool for CUDA programs

Overview

This tool will help you to convert your program from the version using float to half and half2. It is written in Clang libtooling (version 4.0) because that is the only option I can find to parse CUDA code easily for now. All contribution and pull requests are welcome.

Requirements

• Clang 4.0, I have not tested it on other versions of Clang but I will make it compatible to newer versions if necessary.
• G++ (g++ build is optional, you can build the tool with clang using the bash script files provided
• The tutorial on how to install Clang and setup your system can be found here [https://releases.llvm.org/4.0.1/tools/clang/LibASTMatchersTutorial.html]
• Linux & Unix OS for the bash scripts to run, otherwise you have to write your own Windows .bat scripts to invoke clang plugins with appropriate flags.

Building the binaries

• Changing the environment variables to reflect clang path in your system in the file env_path.conf
• Update the environment variables: source env_path.conf
• There are three cpp files to be built [detect_half2_vars.cpp, half2_rewrite.cpp, half_rewrite.cpp] , you can build them using clang or g++.Using clang to build the tool is more stable but slower, while g++ can build faster but some versions may give linking/compile error.
• Example using clang to build detect_half2_vars.cpp, (the argument for the script is a cpp file without extension): ./build_clang.sh detect_half2_vars
• or you can do the same process using g++ (optional): ./build_g++.sh detect_half2_vars

Running the example

• First step is to identify the floating point variables that can be converted to half precision: ./detect_vars.sh test/vectorAdd.cu
• After completing this step, you should have half2VarsList.txt file generated with the content is the list of function and floating point variables in them.
• Choose your configuration from rewrite.conf:
  • use half2 data type simd_mode=true
  • use half data type simd_mode=false
  • use operator overload header to make your code readible operator_overload=true
• Next step is to run the conversion tool, to convert the sample in test/vectorAdd.cu use: ./run_rewrite.sh test/vectorAdd.cu

Publication

Ho, N. M., & Wong, W. F. (2017, September). Exploiting half precision arithmetic in Nvidia GPUs. In 2017 IEEE High Performance Extreme Computing Conference (HPEC) (pp. 1-7). IEEE.

Bibtex:

@inproceedings{ho2017exploiting,
  title={Exploiting half precision arithmetic in Nvidia GPUs},
  author={Ho, Nhut-Minh and Wong, Weng-Fai},
  booktitle={2017 IEEE High Performance Extreme Computing Conference (HPEC)},
  pages={1--7},
  year={2017},
  organization={IEEE}
}

Todo

Clang 7.0

Upcoming release

Block localized threads fusing on latest Clang.