Another NVIDIA Engineer Just Made His First Contribution To Mesa

Another NVIDIA engineer has made his first contribution to Mesa in the rather interesting focus of fixing up Volta so atomic operations will work with OpenCL SVM.

NVIDIA contributions to Mesa are rare... During 2020 were just two commits from a NVIDIA.com address from one developer, no commits from NVIDIA.com addresses in 2019, and a number in 2018 all in regards to Tegra embedded graphics support. In the prior five years there have been just six NVIDIA engineers with commits to Mesa.

Just days into 2021, NVIDIA's Alistair Popple saw his first ever commits to Mesa. Alistair Popple is a NVIDIA engineer since last year out of Australia mostly focused on the company's Linux kernel support. Unlike most of the NVIDIA commits we see to Mesa that are around Tegra support, EGL code, or other common bits, this was about fixing up Volta in the Nouveau driver.

The merge request was to fix up atomic operations on Volta/GV100 so that they will work with SVM. Mesa's Nouveau driver supports Shared Virtual Memory with its ongoing OpenCL compute support. It's a bit ironic as the NVIDIA proprietary driver's lack of OpenCL SVM support has been one of the reasons it hadn't been supporting OpenCL 2.0 for these years.

So it's rather peculiar seeing NVIDIA actually make a (albeit minor) contribution to Mesa now on the desktop side as an enhancement to its OpenCL open-source compute support... This is the area Red Hat has been working on heavily the past few years with Nouveau OpenCL compute support, the NIR support and related infrastructure, etc.

It hasn't been clear why Red Hat has been investing so much on getting this open-source NVIDIA compute support improved, especially as to date all GPUs since the GTX 900 Maxwell series suffer from incredibly poor performance stemming from the re-clocking situation around signed firmware images and the Nouveau driver not having the ability to provide proper PMU handling. So beginning to see Mesa commits from NVIDIA in this area is rather interesting.

Granted, it was back in 2018 NVIDIA and Red Hat formed a partnership to "Align on Open Source Solutions to Fuel Emerging Workloads".

In 2021 it will be interesting to see how this plays out especially with the cancelled talk from last year's GTC conference over potential more NVIDIA open-source news. Given Red Hat's heavy investment into the open-source NVIDIA compute space and even seeing these few bits from NVIDIA leaves us to wonder still the grand plan and ultimately there has to be some solution to the massive performance/re-clocking hurdle in the works and to the extent of any open-source NVIDIA action.