Hi there,
We are searching for someone to code a Multichannel Convolution Engine. Alternatively, if someone can point me to a good library it could work as well. The IRs are up to 4 s long and recorded in 5.0.
Thanks!
Edit: The engine should run in a plugin in real-time
To get a delay free convolution you’ll need to run at least 2 FFTs in different threads plus a standard time domain convolution to prevent the delay. For multiple channels you’ll be using up a lot of CPU. I don’t think there’s an off the shelf solution, it’s been a while since I looked TBH.
You don’t necessarily need several threads, also you don’t need standard time domain convolution. See JUCE’s implementation of the uniformly-partitioned convolution.
However, for such long IRs, a non-uniformly partitioned convolution is necessary, which makes use of different partition-sizes (gardener scheme) and several threads to become efficient.
See also: Convolution Reverb - #2 by danielrudrich
I‘m aware of all that. Interesting that no one has done this yet and that there is no library around. That Gardener paper has been around for ages.
I think the patent makes people a little nervous(!). Even though it’s quite specific.
I’ve heard the patent was dropped, but I can’t be sure without looking further into it, sorry.
For long IRs I found that I needed a small 128 or 256 size time domain section just to fill the gap before the first FFT came back. It was the conclusion of a long struggle with generic CPU a few years ago.
I could live with a latency of FFTSize samp actually.
I’m aware that Overloud does something similar with their REmatrix, check if that may be what you are looking for:
I did started to work on partioned convolution back in 2015 during a univerity project. But never really finished the library.
So far I rememeber it correctly it would be possible to support uniform partioned convolution with minor adaptions. 
HIRT
and WDL convoengine might be interesting.
although not sure either offers more than JUCE’s own convolution engine
