[feature request] - implementing a pd external in cuda using c++

[ghost]

what if this could be the starting point to create a new pd external using cuda and c++:

• https://gist.github.com/tiagomoraismorgado2022/ea81ca8277a3e04b3e37866f98f4b471

[umlaeute]

it would be cool, but i don't see what can be done on the Pd side.

specifically, i doubt that @millerpuckette would add glue to natively support some proprietary acceleration framework (however fast), but of course that is his decision.

but that shouldn't keep you from creating your own framework (if that is required at all to create CUDA-based externals).

for now i'm going to close this a wontfix 😞

[umlaeute]

and just to add. back in the days (when the first CUDA-like devices came out), it was accepted knowledge that while the processing itself is super fast, there's a bottleneck when it comes to transferring data between main memory and GPU memory. this might not be a problem if you work with large amounts of data (as the throughput is high), but in Pd we have small blocks that should return fast (and unfortunately the latency is high as well).

i don't think that things have changed considerably, as shown in this article from 2021:

GPUs are a latency hiding device, so the naive approach of transferring 4MB files over PCIe from the CPU to the GPU for further processing would cripple workflow performance.

[ghost]

this could be interesting for instance for an intelligent counterpoint based harmonizer, or some control voltage crazy behaviour emulation

[Spacechild1]

This is AI generated code without any concrete purpose, copy pasted and posted as a "feature request" to three different repositories: (here, csound/csound#1688, supercollider/supercollider#5947).

Please don't do this.