Dependencies

Mandatory:

OpenCL SDK
Python3 for building
cmake

Recommended:

OpenBLAS for CPU only inference/training
Sqlite3 for ceching of compiled kernels for non-NVidia GPUs to improve startup times

Optional:

boost python and boost numpy - for python bindings
HDF5 C++ bindings for model loading/saving in HDF5 format instead of internal

Building

Use following:

mkdir build
cd build
cmake .. -DCMAKE_BUILD_TYPE=RelWithDebInfo

Using CL/cl.hpp

Some OpenCL programs use older header CL/cl.hpp, in order to make dlprimitives compatible and use the old header pass a parameter to cmake:

cmake -DUSE_CL_HPP=ON ..

Running Unit Test

make test

or ctest

By default ctest used 0:0 GPU - 0 platform, 0 device if you want to change it to another one, lets say platofm 1 and device 0, pass parameter to cmake: -DTEST_DEV="1:0"

Installation

Provide -DCMAKE_INSTALL_PREFIX=/path/to/installation/location to cmake, for example

cmake -DCMAKE_INSTALL_PREFIX=/opt/dlprim ..

In order to use python bindings export PYTHONPATH=/opt/dlprim/python - according to your installed location

Benchmarking

Running benchmakrs on opencl platform:device 0:0

# Inference
./dlprim_benchmark 0:0 ../docs/nets_for_benchmark/resnet18-b16.json
# Train
./dlprim_benchmark -b 0:0 ../docs/nets_for_benchmark/resnet18-b16.json

Windows Notes

MinGW compiler has troubles with cl2.hpp, so it is recommended to set an option -DUSE_CL_HPP=ON
Under windows the library is build as static library