chipbuster / genetic-discrepancy

Experiemental code to accelerate a genetic algorithm for star discrepancy

Genetic Star Discrepancy

Testing repository for the Genetic Star Discrepancy algorithm.

Note: Since the test data is relatively small and self-contained (~250MB), I've chosen to include it in the repository proper.

In order to clone this repo, you'll need to get the BLIS submodule as well. You can do this in one of two ways: the simplest is to do a recursive clone, i.e.

git clone --recursive https://github.com/chipbuster/genetic-discrepancy

If you've already cloned the project, you can get the submodules with two commands:

git submodule init
git submodule update

Hey you dopes: if we ever plan to release this to public, we'll need to add GPL headers to all the source code files (see: https://www.gnu.org/licenses/gpl-howto.html)

Develop path plan:

1. Use BLIS src to benchmark raw matrix-matrix multiplication. Results: about 2 TFLOPS peak for MKL, 1.5 TFLOPS peak for BLIS

2. Use BLISLab to instatiate a derp kernel. Ongoing in tests/fitness_kern

3. Transplant BLISLab kernel into actual KNL BLIS kernel.

Other notes:

Use environmental variable RANDOM_SEED to set a seed. Else it will use a system gen.

genetic-discrepancy's Issues

Types are slightly weird

Not sure if this was intentional or a side effect of prototyping, but the types in this code can be a bit inconsistent.

E.g. in calcFitness, n and M are unsigned, but d is signed. Is there a specific reason for this? If not, I can change it around.

Also, is there any reason we wouldn't want to fix the bitwidths of these fields? e.g. if we fix on 32-bits, using uint32_t instead of unsigned.

OpenMP Threading with Fixed seed sometimes creates different values

When running with a fixed RANDOM_SEED (usually zero), it is sometimes still possible to create different outputs.

This issue disappears when OMP_NUM_THREADS=1 is added to the environment, suggesting a race condition.

Initial bug tested under g++, unconfirmed with other compilers.