An lv2 IR-loader to go with jalv and neural-amp-modeler-lv2

[torgeir]

Just getting my hands dirty with this project 🥳

I was trying to get a barebones version of jalv and neural-amp-modeler-lv2 up and running on the Raspberry PI 4 and was wondering if you had suggestions for some good IR-loaders also lv2 that would go with it. I was planning to use jalv for that too, and connect it up with jack_connect on the command line.

Had a go with convoLV2 but I cant seem to get any sound out of it. Ran it with jalv.gtk3 http://gareus.org/oss/lv2/convoLV2#Mono on my linux system as well (used jalv, the plain version, on the pi), but I could not get any sound through it from there, nor on the raspberry pi 4. Tried a few ML sound lab IRs I had lying around.

Suggestions or ideas welcome!

Also saw your stompbox repo, great stuff! Might give that a go some day as well.

Best regards,
Torgeir

[mikeoliphant]

You could try @brummer10's IR loader: https://github.com/brummer10/ImpulseLoader.lv2

[torgeir]

Unless you have the very latest version

I have been building it from source like described in the readme

git clone --recurse-submodules -j4 https://github.com/mikeoliphant/neural-amp-modeler-lv2
cd neural-amp-modeler-lv2/build
cmake .. -DCMAKE_BUILD_TYPE="Release"
make -j4

Doing cmake with "-DWAVENET_PREFER_NAM=ON" will make sure it is using NAM.

Alright, I'll have a go with that some time later. I just nuked the sdcard to start fresh and try pipedal.

I can do 64

Wow that would be a world of difference from what I was experiencing 😬 Any system setup tricks needed to make jack perform like that or are we talking a bog standard raspberry pi 4 install using the rpi imager?

[mikeoliphant]

Wow that would be a world of difference from what I was experiencing 😬 Any system setup tricks needed to make jack perform like that or are we talking a bog standard raspberry pi 4 install using the rpi imager?

Just a standard OS. I've been using "Raspberry Pi OS Lite" 64bit. Make sure you use the 64bit OS - it makes a pretty big difference.

[torgeir]

That's what I have been using too.

[mikeoliphant]

I'm pretty sure using "-DWAVENET_PREFER_NAM=ON" will sort things out. Adding a 50% extra CPU hit takes standard NAM from running comfortably to being right on the edge of usability.

If you get the latest lv2 source, it should now be using NAM Core by default - the cmake output tells you which engine it is using.

[torgeir]

Solved!

[brummer10]

@mikeoliphant do you use some special cxx flags?

[mikeoliphant]

@mikeoliphant do you use some special cxx flags?

For building what?

[brummer10]

for building on arm

[mikeoliphant]

for building on arm

Nope. When I was running a 32bit OS I did, but it seems to do proper NEON optimization for 64bit Arm by default.

[brummer10]

I just ask because some guitarix users reported issues with NAM when running guitarix on a pi. As I didn't own one I can't check it out. So I wonder which flags you use to build NAM for a pi. Would you mind to share the flags you've in use when running your default build?

[mikeoliphant]

cmake config is here:

https://github.com/mikeoliphant/neural-amp-modeler-lv2/blob/main/src/CMakeLists.txt

So, just "-Wall -Ofast" on Arm.

[torgeir]

So that worked out perfectly! No need to muck around with rtprio tweaks and irqs, which was the path I was about to head down..

Compiling like this did the trick

git clone --recurse-submodules -j4 https://github.com/mikeoliphant/neural-amp-modeler-lv2
pushd neural-amp-modeler-lv2/build
cmake .. -DCMAKE_BUILD_TYPE="Release" -DWAVENET_PREFER_NAM=ON
make -j4
mv neural_amp_modeler.lv2 ~/.lv2/
popd

I documented the steps I took to get this running. Its still a bit rough, I might improve that some day

https://gist.github.com/torgeir/e1fc51da45d3491dd4b08ebf35710ebc

Thanks both of you for the work on your plugins! This is going to be a lot of fun. @mikeoliphant @brummer10

[mikeoliphant]

👍 glad you got it sorted!

[mikeoliphant]

@torgeir You may want to grab the latest version of this repo. I've got a new WaveNet implementation and I'm seeing ~30% better performance for standard NAM models on Pi4.

[brummer10]

You may try with
make -j1
to keep your CPU responsible.

[rusk911]

You may try with make -j1 to keep your CPU responsible.

Thank you. Now it just dies

https://pastebin.com/f2kVnvdr

[brummer10]

So you running out of resources.
what did you get on
ulimit

[rusk911]

So you running out of resources. what did you get on ulimit

$ ulimit -f
unlimited

So I will try docker mod plugin builder on desktop.

[rusk911]

So you running out of resources. what did you get on ulimit

plugin builder also failed. I'm not familiar with plugins programming, seems something is missing.

/home/dima/mod-workdir/generic-aarch64/build/neural-amp-modeler-lv2-3ca0fdc51fc7a747e921606530be7eb4509eb0ae/deps/NeuralAudio/NeuralAudio/../deps/RTNeural-NAM/wavenet/arena.hpp:5:10: fatal error: span: No such file or directory
    5 | #include <span>
      |          ^~~~~~
compilation terminated.

[mikeoliphant]

👍

Btw, by default it is optimized for 64-sample buffers, but you can change it on the cmake commandline. For example, adding:

-DWAVENET_FRAMES=128

will optimize for 128 samples instead. You get slightly better performance if it is optimized for the exact buffer size you are using.

[torgeir]

Nice, I'll have a go with 96 - seems like a sweet spot for the pi4

[rusk911]

I'll try mod docker builder tomorrow. вс, 9 февр. 2025 г., 21:33 Torgeir Thoresen ***@***.***>:

…

Mine is an 8GB one. — Reply to this email directly, view it on GitHub <#83 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABWVGSE5FZIATGKJ3NC6DUD2O63PZAVCNFSM6AAAAABV4YCQCWVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTEMJRGEYDANY> . You are receiving this because you commented.Message ID: <mikeoliphant/neural-amp-modeler-lv2/repo-discussions/83/comments/12111007 @github.com>