[QUESTION]: is there reasonably easy way to call CUDA nvJPEG or NPP functions?

[ilehtoranta]

Question

We have been using CUDA for years and we would like to migrate to ILGPU, but we depend on CUDA libraries, which are difficult to replace. The most notably, we have a demand for nvJPEG library.

CUDA stream pointer is private in AcceleratorStream but I could dig it out with reflection. Are there any obstacles going this way? Is there an easier way to call nvJPEG or NPP functions, keeping data in video memory between calls? Ask an AI to write a JPEG encoder for ILGPU?

Environment

• ILGPU version: 1.5.3
• .NET version: 10
• Operating system: Ubuntu 18/20/22

Additional context

No response

[MoFtZ]

@ilehtoranta the Accelerator class has a DefaultStream property. For Cuda, the Accelerator is an instance of CudaAccelerator, and the Stream is an instance of CudaStream. The CudaStream class has a public property StreamPtr, which you can use, if required.

Some Cuda libraries are easier to integrate that others. ILGPU has some C# to Native C bindings for some Cuda libraries, including a small subset of nvJPEG.

Sample code here:
https://github.com/m4rs-mt/ILGPU/blob/v1.5.3/Samples/AlgorithmsNvJpeg/Program.cs

The work is in creating the C# bindings to the Cuda libraries, including data structures, alignments, etc.

[ilehtoranta]

Nice! Unfortunately there is only a decoder when we need an encoder, but this assures us that ILGPU is a way to go!