In this two weeks connection_pattern and infer_shape were merged. I was supposted to implemented the GPU optimization feature. As I don’t have a machine with Nvidia GPU, I truned to the c_code method after several days.

Reusing the code of CLinker would be our original idea. But thing wuold not be that simple. CLinker generate code at a function scale which mean it treat node OpFromGraph as a function. Yet we want to generate code at a OP scale. Therefore we need to remove some code in a larger scale to make OpFromGraph.c_code() return a ‘node like’ c_code.

There’re to solution. One is to add a new feature to CLinker so that it can detect OpFromGraph node and do some special behavior. The other one is to avoid high level method like code_gen() in CLinker – only use auxilary method to assemble a ‘node like’ c_code. The later solution seems easier and I am working on it. CLinker is a very complicated class, I try to understand it and work out a workable code in the next two weeks.

Wish myself good luck~~!