Testing The Open-Source "RADV" Radeon Vulkan Driver vs. AMDGPU-PRO

[phoronix]

Phoronix: Testing The Open-Source "RADV" Radeon Vulkan Driver vs. AMDGPU-PRO

With word coming out last week that the RADV open-source Vulkan driver can now render Dota 2 correctly, I've been running some tests the past few days of this RADV Vulkan driver compared to AMD's official (but currently closed-source) Vulkan driver bundled with the AMDGPU-PRO Vulkan driver.

http://www.phoronix.com/vr.php?view=23490

[jrch2k8]

Holy shit i expected it to be way less than this, Holy Moly AMDGPU-PRO took a beating this week.

So basically they work a bit on hardware and Texture Upload bandwith(my guess why it tank on 4k), pass the remaining conformance tests and Voila!!!

now AMD would just need to release their OpenCL stack or focus hard on OpenCL 2.0 on clover.

AMD FOSS team and RADV team, thank you very frakking much, Amazing Amazing work.

All hats off people we got badasses over here

[shmerl]

What's the reason for using NIR, instead of translating SPIR-V directly into lower representations?

[darkbasic]

SPOILER: AMD's implementation will be a DAL-like compatibility-layer nightmare and everybody will simply stick with RADVD.

P.S. The funny thing is that the legal review is taking more than 6 months while writing a full driver from scratch took just a week.

[darkbasic]

Forgot to thank David Airlie for the great work.

[BNieuwenhuizen]

What's the reason for using NIR, instead of translating SPIR-V directly into lower representations?

Currently we don't really use the advantages that much, but we expect that inter-stage optimization will be easier to do in NIR than in either SPIR-V or LLVM IR. Basically SPIR-V is annoying to use for the result of transformations, and LLVM IR makes it complicated to represent some graphics shader concepts.

[Evil Penguin]

What's the reason for using NIR, instead of translating SPIR-V directly into lower representations?

I'm curious about this as well, can someone pitch in?
The performance is very surprising! I hope someone from AMD will give us an update regarding their Vulkan open-source driver. They've been a little too quiet on this matter.

[chimpy]

BNieuwenhuizen What are the eventual plans for radv? I mean will you keep developing this driver if AMD decides not use it as the base for their open source Vulkan driver?

[jrch2k8]

Currently we don't really use the advantages that much, but we expect that inter-stage optimization will be easier to do in NIR than in either SPIR-V or LLVM IR. Basically SPIR-V is annoying to use for the result of transformations, and LLVM IR makes it complicated to represent some graphics shader concepts.

Well i think using NIR is actually good idea B, after all intel uses it too for their anv driver.

My guess is the extra eyes may help a lot on the SPIR side instead of going to LLVM IR alone and this benchies demonstrate so far the performance hit is negible enough, at least until more complex vulkan code reach Linux that could justify going that low level but if the shader cache hits vulkan then is even less important.

just one question, NIR goes straight to LLVM or you using NIR->TGSI->LLVM or simply NIR->reuse mesa shader compiler that intel uses??

[xeekei]

Promising for the future. Now if RX 480 could stop being the same price as a GTX 1060 in Sweden, maybe I could justify buying one.