ithemal / Ithemal

Instruction THroughput Estimator using MAchine Learning (ITHEMAL)
MIT License
142 stars 31 forks source link

Public API and Demo #16

Closed CharithYMendis closed 5 years ago

CharithYMendis commented 6 years ago

Shall we give a public API for Ithemal Predictions? Also, we can have a demo website for Ithemal for people to try out.

@mcarbin @alexrenda what are your thoughts?

thisiscam commented 6 years ago

Maybe you can look into this https://github.com/warmspringwinds/pytorch-cpp

alexrenda commented 6 years ago

https://onnx.ai/getting-started This may also be worth looking in to, if going directly between PyTorch and c/c++ is hard

mcarbin commented 6 years ago

I think hosting a web form would be valuable.

That is independent of generating C++, no? For that, the actual challenge here is that this is a dynamic model, right? ONNX or pytorch should be reasonable to use for the LSMT cell, embedding layer, and final regression. However, we'll still need to code the basic block to DAG-RNN traversal yourself, right?

CharithYMendis commented 6 years ago

@mcarbin The tokenization is done at by the DynamoRIO clients, but we can do that using x64asm as well. This is independent from generating C++. Essentially this is like a demo, for example consider http://binsim.com/ or http://tensor-compiler.org/codegen.html.