Hello Sliverfish,

I spotted your proposal for GPU-accelerated signal processing (gr-cuda +
MatX) and wanted to say great work on outlining your approach! I’ve also
explored gr-cuda before but couldn’t pursue it deeply due to hardware
constraints. However, I did have some helpful discussions with Josh Morman
– one of the main contributors to gr-cuda – and he pointed me to a resource
that might interest you:
https://github.com/luigifcruz/CyberEther?tab=readme-ov-file#compatibility

Here are a few suggestions you might find useful:

   1.

   *Scoping and Timeline:*
   -

      Adding a MatX-specific buffer type plus writing specialized blocks
      (e.g., kernel wrappers, file I/O) can get tricky quickly. Make sure you
      have a strong plan for incremental testing. Wrapping entire sets of MatX
      capabilities might be too big for a single GSoC. You could consider
      focusing on a smaller, high-impact set of operations (e.g., FFTs and FIR
      filters) and mark the rest as stretch goals.
      2.

   *Integration with gr-cuda Memory Model:*
   -

      gr-cuda provides a zero-copy approach for device memory. Ensure that
      layering higher-level MatX buffers on top of that doesn’t cause redundant
      data transfers or overhead. A quick proof-of-concept verifying you can
      maintain zero-copy usage with MatX is a good first step.
      3.

   *Testing Strategy:*
   -

      GPU debugging is often more involved than CPU-based code. Set aside
      extra time to build robust test harnesses, especially for performance
      comparisons. Mentors typically appreciate seeing a clear plan to measure
      throughput gains vs. CPU-based blocks under practical scenarios (large
      chunk sizes, typical sample rates).
      4.

   *Documentation & Examples:*
   -

      In addition to developer docs, consider example flowgraphs that
      compare CPU vs. GPU blocks. Demonstrating real performance
improvements for
      something like FFT or FIR filtering will highlight the immediate
community
      value.

Your plan looks thorough, and it’s exciting to see further GPU integration
in GNURadio. I’ll be following your progress—best of luck, and feel free to
reach out if you want to compare notes on GPU aspects!

All the best,
*Krish Gupta*

On Sat, Apr 5, 2025 at 6:42 AM Shen Yu <jiayu....@gmail.com> wrote:

> Hello!
>   My name's Silverfish, I'm a master's student studying wireless
> communications, and I use GNURadio for my work. I would like to contribute
> to the gnuradio community through GSoC.
>    I've written an early version of my proposal and would love some
> feedback to improve it. I look forward to hearing your thoughts. Thank you
> for your time!
> Best Wishes,
> Sliverfish
>

Reply via email to