Call Lua from C++ in threads

Question

I have a classifier model in lua to predict the gender of a person in an image, and I need to call it from C++. I'm using the Lua C API.

I have it working for a single prediction, I mean, in C++ I call dofile("script.lua") and the model is loaded from the disk, then I call predict("image1.jpg") and I got the result. So far so good.

Problem

The problem is that I need to be able to call the predict function using threads (several at the same time), because the predict takes a while to return. The model is quite big, so I need to somehow load it once and use it like a singleton.

How can I implement such thing? Or it's not possible / feasible? So far I've tried but none of it worked:

• Having a model.lua file with only the model as a singleton and the script.lua would get an instance of the model and loading it in the first time. Then in C++, I would have a new lua_State and dofile("script.lua") every time I need to make a prediction. I tried this but the model was being loaded every time.
• Having a static object lua_State in C++ with the model.lua always in memory, which would return the model to C++ ( as a table). Then send it to lua to the script.lua, like: predict(model, "image1.jpg").
• Similar as the previous one, but instead of returning the model to C++ and send it back to lua, return a callback function to C++, send it back to lua and use it to get the model. Example: model.lua has a getModel() function which would return the model. I would return the getModel() to C++ and send it to the script.lua, like: predict(getModel, "image1.jpg"). Then the predict function would do call getModel() to have the model.

Update

Thanks for the feedback guys. Once it's not possible to implement what I need, I dropped this idea of using the C API to communicate between Lua and C++.

Instead, I'm running a server with Lua - using luasched , and from C++ I'm creating a new connection to the Lua socket each time I have to make a prediction.

Regards, Sérgio

Answer 1

What you want is not generally possible with raw Lua. It is an inherently single-threaded system. While you can create multiple independent lua_State objects that can run on independent threads, you cannot call into the same lua_State instance from two different threads.

You could try Lua Lanes, which is ostensibly a threaded framework for Lua. However, it still requires separate lua_States; the main thing it does is (at least in theory) provide a light-weight way to share data between lua_States.

Alternatively, you can create a number of processes equal to the number of cores you have. Each process will have to independently load the model, but you will never load the model more times than the number of cores you want to use. But this way, each process can execute any number of predict calls. You can even take the filename from stdin, thus allowing some master process to farm out work to an individual process.