Precise parallel floating point applications

How to modify

This project is meant to act as a rough template to define other reduction that will still guarantee reproducibility.

How to add your own reduction

To add you own example using floating point expansions and superaccumulators, you can follow these steps.

Let's say we want to add an experiment for a reduction called 'mps'. First, you can add the source files in capps/exblas/src/cpu/blas1. We will take the 'mts' implementation as an example.

Floating point expansion

You will need one class defining the floating point expansion, such as ExMPS_FPE.hpp. For that, just copy one of the files already present (like ExMTS.hpp) and adapt the FPExpansionVectM1 class to your needs.

There are two main methods you should look into:

      void XAccumulate(double x);
      void Flush();

The first one defines the accumulation operation of your reduction operator. The second one defines how the values of the expansion are flushed into the superaccumulator. It should be called at least once per thread, since only the superaccumulators will be passed to the join.

Superaccumulator only version

The file ExMTS.hpp / ExMTS.cpp contains a TBB reduction implementation using only the superaccumulator. You should also copy and adapt this file. Look at the last two methods it contains: one of them defines the superaccumulator only function, the other one is the floating point expansion function:

template<typename CACHE> __mts ExMTSFPE(int N, double *a);

You do not need to worry about the CACHE type. The first parameter defines the size of the floating point expansion to be used, the second one the input array. The implementation if in ExMTS.cpp

Adding your reduction to ExBLAS

Once you have implemented your components, go in capps/exblas/blas1.hpp and add your main function there. For example. this main function for MTS has been defined in ExMTS.cpp.

Use your reduction

You can look at capps/test_mts.cpp for an example on how to use the reduction you just defined. You should just need to import blas1.hpp in order to use the main function of your reduction.

To be continued...