comp-physics
Code gen to C++
Hi,
I'd be interested in creating a direct to C++ code generator instead of using the wrapper. We've had some issues with dependencies and getting to run on device functions in C++ and I'd much rather just directly gen the code. Is this something that is of interest to you all? I imagine we can build off the existing generated fortran code and get it to generate C++ instead.
For demonstration purposes I'd like to start with a single layer NN my colleague has provided for me Sutherland's law. It's a silly test case because it's just a polyfit to begin with, but it can test the work flow and easily end up in CFD. What do you think?
(this is Ryan Johnson from NRL btw)
@rfj2c -- Yes just doing code-gen is possible. This is basically what is currently done in the code. So, in principle, one can use the existing code here to generate the Fortran files for a given NN architecture with weights/biases/activation functions, then cross-compile them in an existing C/C++ code.
Alternatively, one could write a different code-generator that writes native C++ code. This is presumably not terribly challenging because the more difficult work of translating ONNX files into loops and if statements is already done. One could potentially just interpret this as having two native backends: Fortran and C++.
