slow ONNX runtime

[enthropy7]

Now it’s CPU bound, gpu backend is implemented but not connected to ONNX runtime. so, gpu flag is useless. Also, gpu calculations should be more fast. That’s what i’m fixing this whole day.

Gpu backend should be connected and ONNX runtime must be optimised.

[enthropy7]

currently faster than MS onnxruntime (cpp) with cpu only mode, but slower than CoreML. trying to blaze code as it possible. tried direct Metal (apple silicon) API.

[enthropy7]

so, yeee, added Metal backend for apple silicons, custom optimized kernels, etc. Will push it soon after enough tests. new flag here