Recently it has become apparent that Batched BLAS (BBLAS) routines, for computing multiple BLAS operations simultaneously, are required to obtain maximal performance in a number of HPC applications.
Although there is some initial support for BBLAS in CuBLAS, MKL, and MAGMA, there is no standard set of functions and calling sequence between these libraries. Setting such a standard would allow these libraries to interoperate easily and give users a standard set of functionality for use in their applications.
There is a draft standard available at http://eprints.ma.man.ac.uk/2464/ which is open to comments and suggestions from the community.
At the recent "Workshop on Batched, Reproducible, and Reduced Precision BLAS" (http://www.netlib.org/utk/people/JackDongarra/WEB-PAGES/Batched-BLAS-2016/) a number of alternative APIs were suggested by the participants.
The aim of this repository is to create sample code for the different APIs analyzed in the technical report at http://eprints.ma.man.ac.uk/2493/. Please use the repository in combination with the tech. report in order to fully understand the different approaches.
This repository contains implementations of three possible APIs for BBLAS routines. The APIs we implement are:
- Separate fixed and variable batch API
- Flag-based API
- Group API
We also include a version of DGEMM using a strided memory layout, as opposed to the pointer-to-pointer approach used in the other functions.
To clone the repository please use the following command.
git clone git@github.com:sdrelton/bblas_api_test.git
After this you may need to edit the make.inc file.
If you are intending to use MKL BLAS and LAPACK functions, which are installed in the default directories, then you should not need to change anything.
Otherwise, please set the include paths and libraries of BLAS, CBLAS, LAPACK, and LAPACKE appropriately.
Once this is done please type make and wait for the compilation to complete.
In the examples directory there is a simple test of each API for the BBLAS versions of ZGEMM, ZGEMV, and ZDOT. For example, running testzgemm has the following output.
./testzgemm Testing gemm using some very simple tests. Creating inputs for testing grouped zgemm... Finished creating inputs, running each API... Computing with first API... Computing with second API... Computing with third API... Completed computations with all APIs...
If you inspect the code in test_zgemm.c then you can see how each of the three APIs is called. The actual implementations can be found in the src directory.
Furthermore in the utils directory there is a xerbla_batch.c file to handle any errors in the calling sequence.