Benchmarking on Android

[gcervantes8]

trafficstars

I've seen that for Icefall, the 2 ways to export models are using either ONNX (this package) or NCNN.

Has there been any benchmarking done for the 2 methods? I'm wondering which one would be faster.

I did find that there's this page https://github.com/k2-fsa/sherpa-ncnn/issues/44 which includes some NCNN run times.

[csukuangfj]

We have not benchmarked sherpa-onnx on Android. However, we have compared the RTF of sherpa-ncnn and sherpa-onnx on macOS and Raspberry pi 4 Model B with streaming zipformer.

The following table compares the RTF for greedy search with 1 thread

	sherpa-ncnn	sherpa-onnx
macOS	0.159	0.125
Raspberry Pi 4 Model B	0.871	0.697

If speed is the only thing you care about, then I suggest that you choose sherpa-onnx.

---

It is a pain to compile onnxruntime from source if you don't use pre-compiled onnxruntime libs. We have not managed to compile onnxruntime for 32-bit arm.

I don't know how easy it is to add a custom operator to onnxruntime.

---

The source code of ncnn is very well readable and it is easy to extend it. It also provides a tool PNNX to convert models from PyTorch. If there is an op that cannot be converted, it is straightforward to change PNNX and ncnn to support it.

One thing I want to mention is that the file size of libncnn.so for Android is less than 1.2 MB. If you customize it, you can get an even smaller lib. I don't know if there is any open-source inference framework that can produce such a small lib.

Also, ncnn supports non-NVidia GPUs, e.g., GPUs on your mobile phones and ARM GPUs on your embedded boards. ncnn also supports RISC-V.