Tutorial

Introduction

• Useful commands on Linux

Operators

• Using C implementation of operator Conv

Build

The packages relies on cmake to build the C++ extensions. whether it wrapped with pybind11 or cython. Both options are available and can be linked with openmp, eigen, onnxruntime, CUDA. cmake is called from setup.py with two instructions:

• python setup.py build_ext --inplace, the legacy way

• pip install -e ., the new way

By default, cmake builds with CUDA if it is available. It can be disabled:

• python setup.py build_ext --inplace --with-cuda=0, the legacy way

• pip install -e . --config-settings="--with-cuda=0", the new way

• Build with cython
• Build with pybind11
• Build with CUDA
• Build with onnxruntime

Validation, Experiments

• Measuring CPU performance
• Measuring CPU performance with a vector sum
• Measuring CPU performance with a parallelized vector sum
• Measuring CPU performance with a parallelized vector sum and AVX
• Measuring CPU/GPU performance with a vector sum
• Measuring onnxruntime performance
• Measuring performance about Gemm

Technical details in practice

• How float format has an impact on speed computation