C++ implementation of a neural network using OpenMP and CUDA for parallelization.

Author: Mattia Orlandi

This version, in the openmp_nn/ folder, uses OpenMP to achieve parallelism.

Steps:

cd openmp_nn
mkdir build && cd build
cmake ..
make

After executing these steps, an executable file openmp_nn in openmp_nn/build/ folder will be produced.

OMP_NUM_THREADS=p ./openmp_nn N K verbosity mode

where:

• p: the number of threads to use (if not specified, it uses as many threads as all the available cores);
• N: the number of input neurons;
• K: the number of layers, with N > (K - 1) * (R - 1) and R fixed to 3;
• verbosity (optional): if 0 (default) only the execution time is printed, otherwise it will print input data, output data, execution time and validity check;
• mode: if 0 (default) it parallelizes the outer for loop (better performance), if 1 it parallelizes the inner for loop and applies a reduction (worse performance, useful for testing), else it executes the sequential version (useful for testing).

The script openmp_nn/evaluate.sh automates the execution of the program varying the number of threads and the problem size, recording each execution time (which can then be used to compute speedup and strong/weak scaling efficiency).

This version, in the cuda_nn/ folder, uses CUDA to achieve parallelism.

cd cuda_nn
mkdir build && cd build
cmake ..
make

After executing these steps, an executable file cuda_nn in cuda_nn/build/ folder will be produced.

Please notice that this build was designed for Turing GPUs (SM75); if you wish to build the program for other architectures, you'll need to edit cuda_nn/CMakeLists.txt accordingly.

./cuda_nn N K verbosity

where:

• N: the number of input neurons;
• K: the number of layers, with N > (K - 1) * (R - 1) and R fixed to 3;
• verbosity (optional): if 0 (default) only benchmarks are printed, otherwise it will print input data, output data, benchmarks and validity check.

The script cuda_nn/evaluate.sh automates the execution of the program varying the problem size, recording each effective bandwidth, computational throughput and speed-up w.r.t. CPU (in order to compute the latter, you will need to pass the path to the OpenMP version of the program as the first argument, for example ../openmp_nn/build/openmp_nn).

The file Report.pdf contains an in-depth analysis of the parallel algorithms.