As my research always involves a huge amount of computation, parallel technique is a super fancy way to save my time. But the resource of cluster at my university is limited. I recently received a warning that CIT has detected multiple-account usage under my IP. To avoid the risk of being fired, I have to figure out another way to parallelize my job instead of parallelizing accounts. Right on time, GPU programming came out as an ideal option. I hope the GPU resource of my university is not full loaded yet.
Due to some incompatibilities of CUDA and Visual Studio 2017, I spent half a day to figure out the solutions and finally succeeded building up GPU programming environment on Windows 10 with Pycuda installed in Python. Here, I post the procedure of the build and some solutions to the incompatibilities.
Prerequisite
For now, I am working on Windows 10 with Python 3.7.0. The CUDA version that is compatibly built up for me is CUDA 10.0. I used Pycharm 2018.3 as the Python IDE. Probably I will try to build on OS 10 soon in the future.
Simple procedure
Install Python 3.7 and PyCharm 2018.3.
Install Visual Studio 2017 as CUDA needs C++ compile.
Open the Visual Studio installer under the folder Visual Studio 2017.
Select: Modify under Visual Studio 2017 -> Installation details.
Install options: select only the Windows 10 SDK.
Tips
- To make sure that CUDA is successfully installed, check
nvcc -vat your terminal. - The step to build samples in Nvidia’s guide is not necessary. I couldn’t build the samples due to some incompatibilities of vs c++ but still have it work on Python.
Problem shooting
- CUDA install failed
I only came across this issue on my desktop at office that has an old monitor. Then, I resolved it by customizing the installation with uncheck of Visual Studio integration. No idea why but it works. Also check answer here
- **nvcc fatal : Cannot find compiler ‘cl.exe’ in PATH **
Check here. In principle, you should add cl.exe to the environment variables.