GCC C++17 Parallel STL Offloading Execution Policy AKA 'par_offload'
This repository contains an experimental proof-of-concept of an heterogeneous offload-capable Parallel STL. It introduces a new execution policy which attempts to offload the execution to a device that supports HSA 1.0 Full Profile. When the new 'par_offload' policy is used, the implementation tries to offload the algorithm launch through the HSA runtime, and if it fails for some reason, it gracefully falls back to host (CPU) execution.
NOTE: this project is in a preview stage and has received practically no proper testing outside the included test suite. It is also not yet optimized for any particular target.
In your program add
#include <experimental/par_offload> to your sources after including
a PSTL implementation you want to fallback to.
Then build your program with the '-fpar_offload' switch:
g++ --std=c++17 -fpar_offload program.cc -o program
And run it normally:
If you have a working HSA 1.0 Full Profile environment installed, it might silently offload the algorithm's execution. To verify that offloading happens, you can set the environment variable HSA_DEBUG to 1 to get verbose output.
Currently the following algorithms can be offloaded:
Things to Do
An incomplete list of improvements/fixes to do:
- Split the direct kernel invocation API to two functions; one that reflects the host functions name to a string (to get the placeholder function name), and one that takes an array of placeholder-address pairs.
par_offload was developed by Parmance engineers for General Processor Technologies who published the code to the open source community in May 2018.
It is heavily based on the excellent GCC HSA offloading work mostly done by Martin Jambor and Martin Liška of SUSE.