Additional Notes For Running on a Windows Machine

The support for running on Windows is experimental since we mainly tested on Linux machines. In additional to the code setup instructions, you also need to meet some hardware requirement.

Hardware requirement

While this code runs with any modern desktop GPU in Windows, latency measures might be off and do not reflect the full capability of the hardware. By default, GTX/RTX GPUs operate in WDDM mode (intended for graphics use) instead of TCC mode (intended for general-purpose computation). Latency is not optimized for computation tasks under WDDM and you might observe runtime much larger than what's reported in the paper. To avoid such issues, you need to operate in TCC mode. Measured runtime in TCC mode in Windows should match that in Linux. However, due to NVIDIA's restriction, only certain GPUs can be put into TCC mode in Windows. Commonly known models that allow TCC mode in Windows are TITAN RTX and all Tesla series including V100. In summary, if you want to obtain reasonable results for latency-sensitive parts of this codebase in Windows, you need one of the above GPUs.

Alternative solution to the above hardware requirement

While not directly supported in this repository, you can alternatively use a deep learning framework with DirectML (instead of CUDA) backend and write some glue code in det/det_apis.py for efficient GPU computation alongside with graphics applications in Windows.

Running the scripts

Instead of running the *.sh scripts, you run *.cmd counterparts in Windows.