AMDGPU-PRO 16.50 Rolls Out With GCN 1.0 Support, FreeSync

[Amarildo]

whitecat AMD can be pretty stupid when it comes to Linux, but their Windows driver is light-years ahead.

They now have radeon, amdgpu, and AMDGPU-PRO on Linux.

Radeon is "meh" for some games, for others it's total crap;
amdgpu was going in the right direction but now they've screwed mouse cursor for SI users;
and AMDGPU-PRO seems just like the ancient Catalyst OGL/OCL with the amdgpu Kernel driver, except AMDGPU-PRO for SI users is broken: CSGO has black artifacts and a black dome in the sky; X-Plane crawls if you enable clouds (which doesn't happen on amdgpu+radeonSI or the Windows driver); I can't patch xorg-server to make my mouse wheel work on Counter Strike; and OpenCL is completely broken: not only it detects my card as a HD 8000 series and codename "Hainan" (which it's not, it's a Pitcairn R9 270X), but it says my OpenCL device has 32 KB of memory and a clock of 555 MHz instead of 1070.

All of this just makes me wanna say a big "fuck you" to AMD on Linux and go to Windows where the drivers are actually good.

[bridgman]

I just installed a fresh CentOS 7.3 for my system with RX470 (MSI brand).
yum update && reboot
Installed the EPEL7 repository.
Executed "amdgpu-pro-preinstall.sh --check" => OK
Then installed the amdgpu-pro 16.50 package => everything downloaded/installed perfectly.
Reboot
Starting... GNOME3 crash!

Does the crash happen without EPEL7 ? We definitely do QA prior to release but I'm not sure if we would have included EPEL7 in that testing.

[bridgman]

amdgpu was going in the right direction but now they've screwed mouse cursor for SI users;

and AMDGPU-PRO seems just like the ancient Catalyst OGL/OCL with the amdgpu Kernel driver, except AMDGPU-PRO for SI users is broken: CSGO has black artifacts and a black dome in the sky; X-Plane crawls if you enable clouds (which doesn't happen on amdgpu+radeonSI or the Windows driver); I can't patch xorg-server to make my mouse wheel work on Counter Strike; and OpenCL is completely broken: not only it detects my card as a HD 8000 series and codename "Hainan" (which it's not, it's a Pitcairn R9 270X), but it says my OpenCL device has 32 KB of memory and a clock of 555 MHz instead of 1070.

re: mouse cursor, are you talking about a regression in the upstream amdgpu tree or just in 16.50 (our first -PRO driver with limited SI support) ? Note that 16.50 does not yet advertise support for your card and I don't believe it has been QA'ed on desktop consumer boards yet. If you run it on the HW it claims to support I don't think you will find it broken (although X-Plane clouds won't run any faster).

There are still reporting problems in OpenCL over the new kernel driver - we have been picking them off but there are still some there. Initial SI support in 16.50 is for specific workstation and mobile parts; we'll be extending support (and QA) to other SI parts in subsequent releases.

From an OpenCL/OpenGL perspective AMDGPU-PRO *is* the Catalyst OGL/OCL with the amdgpu Kernel driver (plus X driver, libdrm etc...), and that's what we have been saying all along - what were you expecting ?

[Amarildo]

re: mouse cursor, are you talking about a regression in the upstream amdgpu tree or just in 16.50 (our first -PRO driver with limited SI support) ? Note that 16.50 does not yet advertise support for your card and I don't believe it has been QA'ed on desktop consumer boards yet. If you run it on the HW it claims to support I don't think you will find it broken (although X-Plane clouds won't run any faster).

There are still reporting problems in OpenCL over the new kernel driver - we have been picking them off but there are still some there. Initial SI support in 16.50 is for specific workstation and mobile parts; we'll be extending support (and QA) to other SI parts in subsequent releases.

From an OpenCL/OpenGL perspective AMDGPU-PRO *is* the Catalyst OGL/OCL with the amdgpu Kernel driver (plus X driver, libdrm etc...), and that's what we have been saying all along - what were you expecting ?

About mouse cursor: no, I'm talking about amdgpu + RadeonSI. I had to leave Arch because mouse cursor is not working there anymore. I re-installed Arch, used a 4.7-staging Kernel I built on Dec 7 (which was working fine), and then installed mesa-git and llvm-svn. Mouse cursor is gone, just like when support for SI cards was initially out. Using the same Kernel and mesa-git/llvm-svn, but with the RADEON kernel driver, mouse cursor is there.
I also built the latest staging Kernel but it didn't help my case.

About OpenCL: Please tell me you're not going to take 4 months to push a new -PRO stack that works correctly with all SI cards

[mokurai82]

OK So I installed Ubuntu MATE 16.04.1 LTS and the drivers don't spit out the error message that they did in Mint 18 MATE. Freesync seems to be working in Witcher 2, but I'm not sure as the monitor doesn't report it in the OSD - its refresh rate is constant (144hz), unlike in windows where it fluctuates depending on the FPS. Either I have FreeSync working or the perceived smoothness is a placebo effect.

[bridgman]

amdgpu was going in the right direction but now they've screwed mouse cursor for SI users;

About mouse cursor: no, I'm talking about amdgpu + RadeonSI. I had to leave Arch because mouse cursor is not working there anymore. I re-installed Arch, used a 4.7-staging Kernel I built on Dec 7 (which was working fine), and then installed mesa-git and llvm-svn. Mouse cursor is gone, just like when support for SI cards was initially out. Using the same Kernel and mesa-git/llvm-svn, but with the RADEON kernel driver, mouse cursor is there.
I also built the latest staging Kernel but it didn't help my case.

OK, I've read this a few times and I'm still not sure if there has been a regression in amdgpu re: mouse cursor or whether we just haven't gotten it working on your board yet.

When you "had to leave Arch because mouse cursor is not working there any more" was that because Arch switched to amdgpu for SI support, or were they *already* using amdgpu for SI support with working mouse cursor and then amdgpu SI mouse cursor stopped working ?

About OpenCL: Please tell me you're not going to take 4 months to push a new -PRO stack that works correctly with all SI cards

Nope... plan is every 2 months and there was an initial backlog so I would expect less than 2 months between releases while we catch up.

[dungeon]

About OpenCL: Please tell me you're not going to take 4 months to push a new -PRO stack that works correctly with all SI cards

They will need to update sooner as Ubuntu LTS on Desktop will swtich to new kernel sooner, so again they will do it sooner because they will be forced to do it

So, set your clock around next Ubi LTS point release with hope that your SI issues will be fixed in the meantime.

[Amarildo]

OK, I've read this a few times and I'm still not sure if there has been a regression in amdgpu re: mouse cursor or whether we just haven't gotten it working on your board yet.

It was working before, I've been using the amdgpu driver for a few months without problems, it was really going great. Then I went to test the -PRO stack on Ubuntu, and when I came back to Arch mouse cursor was gone. Something updated, I'm not sure what. It wasn't the Kernel, though, because I used the same Kernel that was working before.
I do remember that mouse cursor wasn't present in the very early stages of SI support on amdgpu, but then xf86-video-amdgpu updated and it started working and was working since. So I do believe the problem is with this package.

When you "had to leave Arch because mouse cursor is not working there any more" was that because Arch switched to amdgpu for SI support, or were they *already* using amdgpu for SI support with working mouse cursor and then amdgpu SI mouse cursor stopped working ?

They don't yet. Arch is known for not steering away from upstream, so unless upstream enables amdgpu support for SI cards in the config by default, Arch won't do that too.

I was using this kernel with the following:

clang-svn
clang-tools-extra-svn
lib32-libdrm-git
lib32-llvm-svn
lib32-vulkan-radeon-git
libclc-git libdrm-git
llvm-svn
opencl-mesa-git
vulkan-radeon-git
xf86-video-amdgpu-git
lib32-mesa-libgl-git

I did revert to mesa-13.0.2, to older xf86-video-amdgpu, libdrm, etc, but no login manager starts with these, for some reason I need the -git variants to use the amdgpu kernel driver with my card.

[Amarildo]

They will need to update sooner as Ubuntu LTS on Desktop will swtich to new kernel sooner, so again they will do it sooner because they will be forced to do it

So, set your clock around next Ubi LTS point release with hope that your SI issues will be fixed in the meantime.

I won't. I'll either go to UbuntuMATE/Kubuntu with amdgpu + mesa-git (if they work), or I'll go straight to Windows. I'm sick and tired of AMD's mistakes on Linux, they've proven to be good with so many things but yet very incompetent with others. I can't be productive on either Radeon, AMDGPU, or with the -PRO stack, so my only choice will have to be Windows. I'll be spied there and will hate myself for doing so, but at least the drivers are actually good (and not crap they are on Linux) and everything else works too, no need to compile a single program.

[dungeon]

I won't.

Ah OK then ... i probably missed that was not normal, but figurative question.

Windows is for people too, most use that on Desktop anyway but i can't nor won't choose for you . I don't have GCN 1.0 cards but i feel a pain about linux drivers for these, not that others are bug free but probably GCN 1.0 leading in amount of issues these couple years since it existed It was major pain in the arse developing andswitching to llvm for shader compiler back then... but what you actually get now is probably the best of it and that is still not amdgpu for sure So with that in mind and if you are not satisfied now, you will never be i guess

And likely never will be if you compare it with apps for Windows and drivers Especially last year where dGPUs get about 30 drivers on Windows, but on Linux about 10 times less attention...

Sort of as often as like APUs support on Windows, which are actually not supported under new Linux blob officialy AT ALL... that droped last year with fglrx, but no one informed us nor they like to write changelogs, etc...