[GSoC 2026] RISC-V Optimizations for DNN Module (Starting with Winograd/Winograd63)

, ,
1

Hi everyone,

I am an Electrical Engineering student from IIT Delhi interested in the RISC-V Optimization project for GSoC 2026.

My Analysis so far: I have been auditing the modules/dnn kernels and noticed that while AVX/NEON backends are well-supported, the RISC-V (RVV) backend is missing for several critical kernels. Specifically, conv_winograd_f63.simd.hpp currently falls back to generic C++ execution, which is a significant performance bottleneck for 3x3 convolutions on RISC-V hardware.

Proposed Pilot Task: I plan to implement the missing CV_RVV backend for the Winograd F(6x3) kernel as my initial contribution.

  • Current Status: I have set up a testing environment using QEMU + GCC 15.1 (RVV 1.0 support) and verified the build pipeline.

  • Next Steps: I am writing a standalone benchmark to compare the current scalar fallback vs. a draft vector intrinsic implementation.

Before I open a PR, are there any work-in-progress patches for Winograd on RVV that I should be aware of?

Thanks, Sandeep

2

gsoc 2026 has not even started yet. the last activity was for the 2025 gsoc. OpenCV_GSoC · opencv/opencv Wiki · GitHub

if you want to do something on the github, start an issue, not a PR. the core developers may then become aware of your desire to contribute to the RISCV backend of the dnn module, in the context of gsoc 2026.

3

“Understood. I will move the technical discussion to a GitHub Issue to track the missing RISC-V backend for Winograd. Thanks for the direction.”

4

Hello. I am in the same boat as the guy above and I’ve built opencv from source. Also tried raising a beginner PR for docs to get started but it was promptly rejected

They said it was not an actual community issue and there are not any good first issus labelled. Where can I find community issues that needs work so I can get started with contributing!

Thanks

5

This is your PR: docs: add pythons bindings build notes to configuration reference by AdityaMishra3000 · Pull Request #28365 · opencv/opencv · GitHub

And they rejected it, because it’s a PR “out of the blue”, which is your first PR to the repo, with no issue supporting the PR.

This is simply not how it’s done. Many PRs on the repo get closed, usually due to quality issues, no actual need (frivolous PR), or lack of coordination.

The project is popular. Everyone wants to get some commit in for bragging rights on job applications. The project cannot entertain that type of contribution.

If you’re wondering “how to contribute”, there is a wiki article on how to contribute.

IMHO, your best course of action is to follow the lifecycle of the annual Google Summer of Code event.

6

Thanks for the clarification and for taking the time to explain the process.

Understood about the need for prior coordination and community demand, especially for documentation changes.

I’m interested in contributing long-term and would like to follow the recommended path. Could you please point me to where the GSoC lifecycle and related preparation for OpenCV contributors is documented?

Thanks again