RADV Vulkan Driver Manages Launch-Day Support For AMD Navi 10/12/14 GPUs

[phoronix]

Phoronix: RADV Vulkan Driver Manages Launch-Day Support For AMD Navi 10/12/14 GPUs

Leading up to today's Radeon RX 5700 "Navi" series launch it was looking like there wouldn't be any support within Mesa's Radeon "RADV" Vulkan driver for this community-maintained open-source implementation. But the open-source developers at Valve managed to not only deliver Navi 10 support but also Navi 12 and Navi 14 are also supported with this new Mesa 19.2 code...

http://www.phoronix.com/scan.php?page=news_item&px=AMD-Navi-RADV-Mesa-19.2

[tildearrow]

Beating AMDVLK in release delay. Nice.

[R41N3R]

Very impressive! Thanks, radv is the best graphics driver I've ever used.

[entropy]

The code is just about one thousand lines on top of the existing RADV GFX9 support. Interestingly, some 600 lines of that code is just for bringing up the NGG (Next Generation Geometry) support for vertex shaders. So overall at least from the Vulkan driver side it wasn't much new code for driving Navi, granted, it also leverages the Navi support within the AMDGPU kernel driver and LLVM shader compiler.

Isn't that a result from what John Bridgman said about the confusion when it comes to the "GCN/RDNA" architecture?
Namely, that contrary to some news sites Navi is not a hybrid of GCN and RDNA
(they deduced that from Phoronix' initial findings in the first code drops).
While the architecture (RDNA) is significantly different its ISA has a large overlap
with the existing GCN ISA and, hence, makes it possible to leverage existing driver code.

[ResponseWriter]

The MR says this:

There are still significant areas disabled because we haven't had the time to get them fully working:

Tessellation
Geometry shaders
Transform feedback
Binning

This will limit the amount of games that you can play for now, but we'll expect to enable these over the coming days.

So looks like there is still work needed before benchmarks will be meaningful but hopefully won't be too long. Still, great work and thanks to Sam and Bas for getting this working!

[shmerl]

While the architecture (RDNA) is significantly different its ISA has a large overlap
with the existing GCN ISA and, hence, makes it possible to leverage existing driver code.

While it supports backwards compatibility, it still needs different approach to get optimal performance. Check the "waves" topic.

[Brisse]

Isn't that a result from what John Bridgman said about the confusion when it comes to the "GCN/RDNA" architecture?
Namely, that contrary to some news sites Navi is not a hybrid of GCN and RDNA
(they deduced that from Phoronix' initial findings in the first code drops).
While the architecture (RDNA) is significantly different its ISA has a large overlap
with the existing GCN ISA and, hence, makes it possible to leverage existing driver code.

Yea that makes sense. Check out Gamers Nexus interview with David Kanter where they talk about the architecture and the importance of backwards compatibility with GCN.

https://youtu.be/V7wwDnp8p6Y

[ernstp]

Geometry shaders and tessellation are pretty big deals yes... But nice anyway! Give them another day or two 😁

[bridgman]

While the architecture (RDNA) is significantly different its ISA has a large overlap with the existing GCN ISA and, hence, makes it possible to leverage existing driver code.

Yea that makes sense. Check out Gamers Nexus interview with David Kanter where they talk about the architecture and the importance of backwards compatibility with GCN.

Yep... the ISA and programming model have changed, but we only changed what we had to.

You will probably see more divergence between GCN and RDNA driver code over time as more optimization happens, but not enough to justify starting a new driver. Most of GFX10 is about HW implementation changes although some of that does bleed through into the programming model.

[Adarion]

People, THIS is the power of FOSS, documentation and whatnot. Some hours ago I read Michael's test and his complaint that there was no Vulkan functionality yet (maybe in the blob) and bam! a few hours later we have code. (Of course it was prepared and held back / worked at til now, but wow, that's a nice surprise!)