How do I benchmark deep learning code (C/C++/CUDA) to compare performance of the generated code running on CPUs and GPUs?

I'm looking at different deep learning networks (e.g. squeezenet, mobilenet, etc) and I want to be able to profile on different types of hardware (CPU and GPU).

I know I can generate code for these using MATLAB Coder and GPU Coder. I'd like an automated way of benchmarking the code running on different hardware so I can quickly compare performance.

Answer 1

[Bill Chou]

You can use the dlCodegenBench function to benchmark the runtime performance of deep learning models running as C/C++/CUDA code generated from MATLAB Coder and GPU Coder.

dlCodegenBench automates the execution running with different code generation configurations so you can compare performance on different hardware (CPU/GPU), different deep learning optimization libraries (target-independent code, MKL-DNN, cuDNN, TensorRT, etc).

More details here:

• dlCodegenBench: https://www.mathworks.com/matlabcentral/fileexchange/180806-dlcodegenbench