neon-benchmarks

benchmarks for neon, for various backends, including OpenCL direct.

This is a separate repo, so it can be used against both repos

Pluggable test backend lets tests run against both

How to use: neoncl

• install neoncl https://github.com/hughperkins/neonCl-underconstruction
• install neon-benchmarks, eg:

pip install -e ./

• run:

neon_benchmarks.py

How to use: neon base

• Install neon base from https://github.com/nervanasystems/neon
• Install neon-benchmarks. From the neon directory:

source .venv/bin/activate

• Then, from the neon-benchmarks directory do:

pip install -e ./

• run, Maxwell Winograd SASS kernels:

neon_benchmarks.py --backend sass_winograd

• run, Kepler direct CUDA kernels:

neon_benchmarks.py --backend cuda_direct

Model selection

There are different network models in the models directory. Simply use the --model parameter to choose one, eg:

neon_benchmarks.py --model simple1  # this is mostly for debugging / dev
neon_benchmarks.py --model alexnet
neon_benchmarks.py --model vgga

Results

• VGG A on Titan X