daylen / leela_ane

leela_ane

An attempt to get the lc0 network to run on the Neural Engine of the M1 chip. Why? For faster chess programs!

Using this repo

Github file size limit is 100MB. Download the .mlmodel file and place next to AppDelegate.swift.

Creating the .mlmodel file

To run on the ANE, one must have a CoreML .mlmodel file, unless one is doing this.

1. Start with network "42850" from https://lczero.org/play/networks/bestnets/. This is in the Leela network format.
2. Also download this t40.yml file: https://gist.github.com/daylen/7ac1e9d9c9d38a9eaadff133f3546df2
3. Use net_to_model.py in my fork of lczero-training to create a .mlmodel file from the network weights and yml. Note that unfortunately this only supports POLICY_CLASSICAL, not POLICY_CONVOLUTION

Benchmarking

For reasons I don't understand yet, inference on ANE and GPU is slower than inference on CPU:

.all (Activates ANE) Time to evaluate: 28.928249542 seconds
.cpuAndGpu Time to evaluate: 29.006600709 seconds
.cpuOnly Time to evaluate: 16.062404083 seconds

Verifying ANE usage

I found these instructions useful. (tl;dr there is no programmatic way to identify whether the ANE is being used, so the trick is to set a breakpoint on -[_ANEModel program]. When using .all for computeUnits it does break on that line.