AMDGPU vs. Radeon Kernel Driver Performance On Linux 5.0 For AMD GCN 1.0/1.1 GPUs

[Dukenukemx]

I use the AMDGPU driver for my Radeon HD 7850 just for the Vulkan driver cause the Radeon driver doesn't have Vulkan.

[Berniyh]

The problem with the AMDGPU driver is that AMD does not release the updated firmware necessary to make it work with the community-written VAAPI patch, so no hardware accelerated decoding at all.
<3 proprietary firmware

Works fine over here with a GCN1.1 chip on amdgpu using va-api.

I've been using amdgpu on my Kaveri for almost 3 years now exclusively, since a few versions also in combination with the DC (display code) stack.
Apart from a few problems (especially with DC at the beginning), it's working like charm.
Actually the last kernel version I used with the radeon driver was 4.4.

Of course I can't say anything about other chips, but personally, I don't see much reason to stay on radeon, if you have the option. Should at least test it.

Edit: vdpau works as well, but with the restriction that you can only decode one video at the time, while va-api can do multiple videos in parallel. Used to be the same for vdpau, but for some reason doesn't work properly anymore.

[pegasus]

All this gaming and fps and whatnot is totally missing the point. Fact is that 7970 is still the highest performing single consumer gpu card at double precision compute performance. Not even consumer Vega offers as much. If you need more than 1TF it offers, you need to invest into enterprise Vega at minimum 5x the price.

And this is why I still need an ubuntu 14.04 installation around with amd proprietary driver where OpenCL actually works to do some research and development around it. It's becoming more and more of a pain with outdated compiler stack.

I sincerely hope that when gcc 10 comes out with openmp loop offloading to gpus, GCN 1.0 support in opensource compute driver stack will work. Otherwise it's all moot.

[Strunkenbold]

I think Powerplay missing is the major showstopper for enable it by default.
I guess AMD has other priorities and focus on newer gens but people are right: it's not like they released the last card with this tech 5 years ago.

I gave up waiting for the point they fix all bugs with gen 1 cards and implement all missing features. RX 580 it's dirt cheap. So no reason to spend more life time on those old cards.

[Namenlos]

The problem with the AMDGPU driver is that AMD does not release the updated firmware necessary to make it work with the community-written VAAPI patch, so no hardware accelerated decoding at all.
<3 proprietary firmware

This is really annoying you have to choose Vulkan or hardware decoding.

[Masush5]

This is really annoying you have to choose Vulkan or hardware decoding.

While i use hwdec pretty much every day on my desktop, I don't think i'd even notice if it would just silently stop working. The benefits are so small if your cpu is more than powerful enough to do the decoding, I'm not sure they're even there. I would need to messure power consumption to see if it's actually any lower with hwdec.

[cb88]

All this gaming and fps and whatnot is totally missing the point. Fact is that 7970 is still the highest performing single consumer gpu card at double precision compute performance. Not even consumer Vega offers as much. If you need more than 1TF it offers, you need to invest into enterprise Vega at minimum 5x the price.

And this is why I still need an ubuntu 14.04 installation around with amd proprietary driver where OpenCL actually works to do some research and development around it. It's becoming more and more of a pain with outdated compiler stack.

I sincerely hope that when gcc 10 comes out with openmp loop offloading to gpus, GCN 1.0 support in opensource compute driver stack will work. Otherwise it's all moot.

Where have you been under a rock? the Radeon VII (consumerized Vega20) has 3.46 TFLOPS DPFP... for ~$700 far outclassing the 7970 in DPFP by over 50%, and having vastly more memory (16GBvs 3GB) and bandwidth (1TB/s vs 200-288GB/s), and improved compute architecture so real world applications that use DPFP are probably well over 50% faster.

Also RoCm provides support for Radeon VII officially https://rocm.github.io/hardware.html

[pegasus]

Nice ... so Radeon VII has useable DP performance. I somehow missed this, thanks for pointing it out.

[agd5f]

The problem with the AMDGPU driver is that AMD does not release the updated firmware necessary to make it work with the community-written VAAPI patch, so no hardware accelerated decoding at all.
<3 proprietary firmware

CIK (GCN 1.1) has had VCE and UVD support from the beginning. For SI, the current radeon SI code for VCE and UVD could be ported over directly to amdgpu. Updated firmware is only required to fix some addressing limitations due to differences in how how we set up the GPU memory layout in amdgpu compared to radeon. You could switch SI support in amdgpu to the old radeon memory layout and port the support directly from radeon.

For CIK, there are some limitations on the display side (no support for analog DACs, audio, etc.) compared to radeon.

If you are concerned about Vulkan, someone could also make radv run over radeon as well.

The bigger thing however, is that these parts has been tested extensively for years on radeon. Lots of fixes for things like suspend and resume and platform quirks. There's a good chance for a lot of regressions for a lot of people if you switch the driver. Are distros going to want to take that on? It's trivial to switch between the drivers if you to use amdgpu.

All of this is open source. I'm happy to review patches to fill in any remaining gaps.

[YamashitaRen]

I have to say I'm a bit disappointed by how AMD neglected CIK and, especially, SI support in amdgpu.

It is confirmed. SI and CIK cards are having copy to host issues. This is most likely a problem with AMD as only OpenCL is suffering from problems when using experimental SI and CIK support.
Update: AMD's OpenCL implementation is crippled on GCN 1 cards as of v18.40. Guess GCN 1 owners will have to wait for AMD to fix the issue or revert back to Catalyst.

For the time being SI support is pretty much unusable for any purpose, it fails OpenCL Conformance Tests pretty spectacularly, failing 67 out of the 77 tests before segfaulting the test program during async copy test. It even fails 9 out of the 11 basic fp/int math tests, and all of the type convertion tests.
And that's on CentOS 7, listed as supported by AMD.

https://aur.archlinux.org/packages/opencl-amd/

SI doesn't even have powerplay, which means no overclocking capabilities : https://bugzilla.freedesktop.org/show_bug.cgi?id=106363