UCBerkeleySETI / dedopplerperf

Performance testing of different dedoppler kernels.

dedopplerperf

Performance testing of different dedoppler kernels. There are currently three different dedoppler kernels tested here, the CudaTaylor5demo code in C+cuda, the turboSETI Python code, and raw_kernel.py which is Python + cuda using the cupy RawKernel interface.

They are set up to run on the same input data. Check the comments in the code to see how to alter it.

I used Ubuntu for testing. Performance on a 256 x 2^19 matrix with a GTX 1080:

• cudataylor5demo: 0.04s
• raw_kernel: 0.09s
• turboseti: 0.45s

How to run CudaTaylor5demo

You need to install the Cuda toolkit, make sure nvcc is on your path, then:

nvcc CudaTaylor5demo.cu
./a.out

How to run the Python dedoppler code

You need a Python environment with cupy and turboseti installed. I provided an environment.yml if you are using conda. Then:

./turboseti_wrapper.py
./raw_kernel.py