Experimental Ray-Tracing For Open-Source Radeon Vulkan Driver Nears Upstream Mesa

[phoronix]

Phoronix: Experimental Ray-Tracing For Open-Source Radeon Vulkan Driver Nears Upstream Mesa

It looks like within the coming days that the Vulkan ray-tracing support for Mesa's "RADV" Radeon Vulkan driver will be upstreamed for Mesa 21.3...

https://www.phoronix.com/scan.php?page=news_item&px=RADV-Ray-Tracing-Near-Mainline

[schmidtbag]

Pretty exciting news. That's no minor achievement either, especially to be done in such a relatively short amount of time. Even more impressive for someone who isn't an AMD employee.

[perpetually high]

As an avid q2 player to this day (I go by Frank White on tastyspleen.net, come get some), this warms my heart. Sincere thanks for all the work put into this.

[tildearrow]

Pretty exciting news. That's no minor achievement either, especially to be done in such a relatively short amount of time. Even more impressive for someone who isn't an AMD employee.

Agree! What is this secrecy about ray-tracing hardware?
I am glad we finally have open-source hardware-accelerated ray-tracing support.

[tildearrow]

Typo:

the weekend Bas opened a MR for adding this experimental tay-tracing support to mainline.

[bridgman]

Agree! What is this secrecy about ray-tracing hardware?
I am glad we finally have open-source hardware-accelerated ray-tracing support.

The ray tracing instructions are documented (~5 pages) in the ISA guide. What we haven't published AFAIK is our code for accelerating the building of BVH structures but as far as I know that part is fairly generic other than making it go fast.

Agree that "generic" is not the same as "easy", particularly when there is no existing code from other vendors to leverage - this is great work.

https://developer.amd.com/wp-content...vember2020.pdf pages 80-82 and 222-223.

[arun54321]

do people really care if a app works 2% faster or slower ? the benefits from opensource drivers are much higher than the loss of 2% of the performance.

so please stop this madness... stop making drivers no one use stop making drivers that makes no sense from a technical standoint...

on linux i wast to plug in my gpu card start my fedora and run the games ... i do not want to install amd driver from amd.com i do not want to use some strange driver no one use who lose all benchmarks anyway i do not want to use closed source opengl driver for 2% performance benefit because aside from this everything workstation customers are are better on the opensource side.
.

Speaking for yourself. There are people who are fine with closed source drivers for extra performance. world doesn't revolve around you.

[zboszor]

what extra performance ? the opensource solutions wins 98% of all benchmarks...

The extra performance the niche market of CAD/CAM users enjoy with the Pro driver. You know, the large grey zone between masons and the proverbial joke of programmers vs woodpeckers. "If builders built buildings the way programmers wrote programs, then the first woodpecker that came along would destroy civilization."

Also, Marek stated in a comment that the Pro driver still has 2% advantage over the Mesa OpenGL driver in performance and that's after a huge amount of optimizations. The last ones are definitely with diminishing returns. I am sure that this last 2% gap can be eliminated but with quite some future work. The Pro driver's low level solutions for its advantage may not be applicable because (1) the codebase is sufficiently different from Mesa and (2) the Pro driver may contain licensed 3rd party code that can't be opensourced.

[phoronix_is_awesome]

AMD should rething they opensource stradegie on linux... do we really need 3 different drivers for the same purpose ?
95% of the people dont use this 3 different drivers... most of them use RADV+ACO
people are very irritated by this and every forum thread turns into bullshit bingo about what driver to use...

i bought cyberpunk 2077 i play it with proton on wine on my vega64 everything at max and then FidelityFX Cas at 75% looks very good and works perfekt. of course it is RADV+ACO ...

from a PR standpoint amd should end this 3 drivers for one task... also why not end the closed source OpenGL driver?

do people really care if a app works 2% faster or slower ? the benefits from opensource drivers are much higher than the loss of 2% of the performance.

so please stop this madness... stop making drivers no one use stop making drivers that makes no sense from a technical standoint...

on linux i wast to plug in my gpu card start my fedora and run the games ... i do not want to install amd driver from amd.com i do not want to use some strange driver no one use who lose all benchmarks anyway i do not want to use closed source opengl driver for 2% performance benefit because aside from this everything workstation customers are are better on the opensource side.

and do even more start to port the linux opengl driver to windows, start to port the linux vulkan driver to windows.

i can not even unterstand this LLVM compiler bullshit all tests i know ACO is better even for compute...

so whats the point?... does -AMD do it for the LOLs for yet another benchmark who RADV+ACO and RadeonSI-opengl driver win all benchmarks ?...

maybe a joke i dont unterstand.

I agree that the 3 drivers for one task is bad for AMD. But there is a reason most people don't talk about. Even though the closed source driver is slower in most cases, GPU makers need closed drivers for "hiding" proprietary algorithms or optimizations that they don't want the open source driver to have It also has some Professional "guarantees" that corporations need for professional applications. I mean, AMDVLK sucks and garbles my laptop screens all the time. The open source driver, mesa+aco acquired by the Ubuntu ernstp ppa is "Mesa Almost Stable", and you have to accept that some updates bork some of your random Proton games. And the "Stable Mesa" Kisak for whatever reason is always 10-15% slower than the ernstp PPA mesa even based on the same code( AVX2 optimizations enabled on ernstp and not on Kisak?)

The problem is actually that AMD releases the official "PRO" drivers roughly 2-3 times a year, and it doesn't always support the latest linux kernels( It only supports Ubuntu LTS line of kernels, which is 6-12 months behind mainline itself.) So you would have to wait for AMD's ass for a quarter minimum even on top of Ubuntu LTS kernels for support, so the PRO driver is always 16months behind mainline in terms of kernel support. Unlike Nvidia's binary drivers, which is usually only one point release behind linux mainline kernel for support. If AMD supported PRO drivers are released monthly,and supports the latest mainline kernels better, then I would actually suggest the "PRO" driver instead of the open source mesa compiled by a "Third Party" awesome German. Shit, a standalone German engineer is releasing driver faster and more predictably than the whole "PAID by AMD people team".

The whole situation is in a toilet, that's why I don't recommend AMD GPUs for anything other than a very small primary display adapter running Xorg and Firefox just so you can save the VRAM on other Nvidia's GPUs to do real computational work. Having a ton of firefox and Xorg running on Nvidia GPUs waste 1-2GB of precious Nvidia VRAM, and you can offload it to a cheap $10 AMD GPU to actually speed up Tensorflow on Nvidia GPUs).

[obri]

I doubt that this Raytracing [EDIT: deleted RTX] driver will ever have great performance.
Perhaps I am to pessimistic here - but without the knowledge how hardware works, how can you really optimize it to get the best performance out of the hardware?
Seems to me that this affors just can not lead to an optimal usage of the hardware aka great performance.