tf-perf-kernels

This repository contains scripts calling kernels for TensorFlow along with profiling scripts for Cori-GPU.

Prerequisites (Cori)

module load python/3.7-anaconda-2019.07 cuda/10.2.89

Install TensorFlow (Conda)

Create a new virtual environment py3.7-tf2 by choosing Python 3.7:

conda create -n py3.7-tf2 pip python=3.7

Activate the virtual environment and install TensorFlow (Python 3.7 GPU support):

source activate py3.7-tf2
(env) pip install --ignore-installed --upgrade https://storage.googleapis.com/tensorflow/linux/gpu/tensorflow_gpu-2.0.0-cp37-cp37m-manylinux2010_x86_64.whl
conda deactivate

Install PyCUDA

Download PyCUDA and unpack it:

wget https://files.pythonhosted.org/packages/5e/3f/5658c38579b41866ba21ee1b5020b8225cec86fe717e4b1c5c972de0a33c/pycuda-2019.1.2.tar.gz
tar xvf pycuda-*.tar.gz

Configure and build within the Conda env:

cd pycuda-*
source activate py3.7-tf2
(env) python configure.py --cuda-root=${CUDA_HOME}
(env) make install -j8
conda deactivate