Raspberry Pi V3D Graphics Driver Preps For Super Pages To Boost Performance

[phoronix]

Phoronix: Raspberry Pi V3D Graphics Driver Preps For Super Pages To Boost Performance

Igalia continues maintaining the Broadcom V3D open-source graphics driver code that is used by the Raspberry Pi single board computers. With a new patch series posted today for the V3D DRM driver, support for Super Pages is enabled to help with enhancing the graphics performance. In many benchmarks having Super Pages can enhance the performance by a few percent but in some extreme cases can be 19~42% faster...

https://www.phoronix.com/news/Raspberry-Pi-V3D-Super-Pages

[schmidtbag]

Not sure if Intel or AMD enable super pages, but if they don't, makes me wonder if that might help their iGPU performance.

[brent]

AMD GPUs supports variable sized pages and amdgpu/radeon drivers have been using them for a long time.

Not sure about Intel, but it would surprise me if they don't have any optimization in that area.

[intelfx]

Not sure if Intel or AMD enable super pages, but if they don't, makes me wonder if that might help their iGPU performance.

These "super pages" are specific to the Raspberry Pi 4+ GPU (V3D). If Intel or AMD have an equivalent, I bet they use it in their drivers.

[schmidtbag]

These "super pages" are specific to the Raspberry Pi 4+ GPU (V3D). If Intel or AMD have an equivalent, I bet they use it in their drivers.

I understand it's specific to V3D, but what got me thinking that is the other recent article about Nvidia's driver's use of transparent huge pages. I know that's not the same thing, but it got me thinking how perhaps AMD and/or Intel could take a different approach in memory management.

I am not saying they need to; I don't know if they need to. But, perhaps it's something that could help in bandwidth-limited environments.

[gnarlin]

Any chance that we can get some benchmarks for this to compare it off and on?

[NathanG]

Kloster Andechs is good beer.

[monkeynut]

Any chance that we can get some benchmarks for this to compare it off and on?

I'd also be interested in benchmarks to compare.

[mangeek]

perhaps AMD and/or Intel could take a different approach in memory management..

The Intel drivers can use huge pages. Huge pages got a really bad reputation a decade ago, but people are coming around to the idea that they're really a Good Thing and that memory should probably be allocated in 1-2MB blocks instead of 4K blocks on all but very constrained systems. It used to be the thinking that a specific app might benefit from it, a specialized setting for weird use-cases, but I think it's logical that having as few 4K pages as possible optimizes TLB hit rates.