libxsmm / libxsmm-dnn

Reference implementation of Deep Neural Network primitives using LIBXSMM's Tensor Processing Primitives (TPP)

LIBXSMM-DNN

LIBXSMM-DNN is a reference implementation of Deep Neural Network primitives using LIBXSMM's Tensor Processing Primitives (TPP).

To checkout the source code:

git pull --recurse-submodules
git submodule update --init --recursive

To build from source:

make -j $(nproc)

NOTE: To rely on a specific version of LIBXSMM (Git submodule not required), set LIBXSMMROOT environment/make variable and rely on a separately built LIBXSMM.

To mimic the performance of ResNet-50 inference:

tests/conv/run_convs.sh 1 1000 -1 f32 F 0 0 64 64 1 | tee output.log

NOTE: To mimic performance, FLOPS are shown and the time spent for convolutional layers is measured. This is different from running inference end-to-end using a real framework such as PyTorch).

To summarize timings:

tests/performance.sh output.log