zeloe / juce_cuda_convolution

Linear realtime convolution using CUDA

juce_cuda_convolution

Linear Convolution using CUDA

 git clone https://github.com/zeloe/juce_cuda_convolution.git
 cmake . -B build -G "Visual Studio 17 2022"

How it works

It performs time domain convolution on two different files on a realtime thread.

Time Domain Buffer

This holds all values for convolution in size of the padded impulseresponse.
Padded impulse response is a multiple of buffersize and number of paralell convolutions. Insert and shift kernel copies new buffer at beggining.
All other content gets shifted by buffersize.
Content at end of Time Domain Buffer gets discarded.

Cuda

It uses four kernels with different streams.
Inside header file you can specify number of paralell streams. (you will need to edit and add functions for that)
Each of these kernels computes convolution results and sums them inside same buffer.
I used constant memory for offsets in time and impulse response buffer.
Shared memory for each sub convolution.

Hardware

GeForce GTX 1660 Ti

Note:

It still needs optimisation.
I tried it also as VST and got it running in mono. I have really old hardware :).
As a template repo i used this template.
Cuda == 12.3.52
MSVC == 19.36.32537.0