Slackware and AMD GPU PRO

Pixxt · 01-15-2018, 08:58 PM

Quote:

Originally Posted by Roboto

Why does AMD's stuff suck such gassy monkey bum in Linux?

Does it? I only buy AMD gfx cards because they beat Nvidia in terms of less headaches, better drivers and support under Linux and other free operating systems.

Roboto · 01-19-2018, 10:13 AM

One reason,

Video production in Linux, baby. VIDEO!!!!

Nvidia, can be done.

AMD/ATI, not by the hair of my chinny-chin-chin!

Nille_kungen · 01-20-2018, 04:59 AM

Your got an Southern Island card and those has bad support for OpenCL and isn't really supported with Rocm and you can add vulcan to that list.
The Southern Island is the first GCN cards and the first cards often miss things that later cards supports and is a sort of hybrid between old and new.
There's a lot that SI cards does different than later gen GCN and that is why it took such time to get it ported to AMDGPU and been experimental for some time.
IIRC the differences in Southern Island and the later GCN, something in my head says different IOCTLs, missing MEC block and so on.
This means that it takes much extra work and resources to add support for SI and that is why SI support is lacking in some areas.
I got an SI dGPU but i still use the old radeon driver and has no need for OpenCL so it works for me.
When SI cards was released the support was ok with the blob driver but then AMD new driver plans happened for the newer cards that did things differently and SI got caught in the middle of it.
You cannot compare windows support with Linux support, it's different teams with different resources and developing different drivers.

This is just from the top of my head so things might not be accurate, i recently had an hard drive dying and can't check my notes on the subject on top of that i got a bad cold messing with me.
I don't know if a newer cards is much better but the odds are much better and AMD did help blender with OpenCL support.

Roboto · 01-20-2018, 08:58 AM

Hey Nille,

It is too bad that AMD lies and says that the GPU is supported but clearly it is not. I blew last weekend installing Slackware and trying to install the AMDGPUPRO and so am not sure I want to go again with a different card even though that is listed as supported as well on their driver site. There was a fellow before who posted about using the open drivers, but I tested that out with the newer card for video editing last night and it was a dog's breakfast, yet again. I will not bother trying AMDGPUPRO again, maybe until they get newer, updated drivers or something. Already wasted enough time trying to video edit in Linux.

Gumboherpy · 01-20-2018, 08:33 PM

Do not use AMD Radeon if you plan on installing proprietary drivers. I have plenty of nightmare stories with fiddling with the catalyst drivers in the past. I am using Nvidia at the moment but I will be switching to AMD soon enough cause their open source drivers are vastly superior to any of their proprietary drivers that it makes me wonder why they even bother to release them in the first place.

Roboto · 01-21-2018, 12:39 AM

@Gumbotherapy If you are video editing and composting, you need open CL. The opendrivers won't give you openCL, thus, you need to install the proprietary drivers to get it. Yes, trying to install AMD proprietary drivers GPUPRO is a miserable experience. Bass Madrigal was cool in helping, I wish him success.

bassmadrigal · 01-21-2018, 01:34 AM

Quote:

Originally Posted by Roboto

The opendrivers won't give you openCL, thus, you need to install the proprietary drivers to get it.

They actually can with ROCm and a supported CPU and GPU. Right now, Intel's Haswell and newer and AMD's Ryzen are the recommended CPUs, and Polaris and newer cards are supported.

Per ROCm's supported hardware page, the following GPUs are supported:

GFX8: Radeon RX 480,Radeon RX 470,Radeon RX 460,R9 Nano,Radeon R9 Fury,Radeon R9 Fury X Radeon Pro WX7100, FirePro S9300 x2
Radeon Vega Frointer Edition
Radeon Instinct: MI6, MI8, and MI25

But as far as I know, I don't think anyone's done work to get ROCm on Slackware yet (they provide Ubuntu and Fedora packages)...

For more details, read up on their site.

But yeah, for simple OpenCL support, you're best using proprietary drivers, which means you're best using an Nvidia product...

gaitos · 10-05-2019, 02:37 PM

Just one observation: I have used the 17.10 version of the SlackBuild (it seemed that 16.60 might not support my card) and at first the amdgpu module would not load:

Code:

# modprobe -v amdgpu
insmod /lib/modules/4.4.14/extra/amd/amdkcl/amdkcl.ko
modprobe: ERROR: could not insert 'amdgpu': Exec format error

relevant dmesg message:

Code:

[   67.073449] amdkcl: exports duplicate symbol drm_atomic_helper_duplicate_state (owned by drm_kms_helper)

The 17.10 build script no longer included the patch for the standard 4.4 kernel. However, after copying the patch from 16.60 and modifying the SlackBuild script to apply it, everything is working properly.

I have revived this thread since it was the most relevant to what I think might be a typical problem: using Slackware 14.2 with a somewhat recent AMD card. The opensource drivers present in 14.2 do not support this card (and I suspect many others produced in the last years) and this was the only solution to avoid complete software rendering (in fact the first notice I took of the problem was that gmplayer wouldn't render movies, then I noticed that xvinfo would show no extensions and so on).

gaitos · 10-06-2019, 03:47 AM

All is not well, unfortunatelly both opera and vivaldi fail to start with:

Code:

libGL error: pci id for fd 34: 1002:67df, driver (null)
libGL error: No driver found
libGL error: failed to load driver: (null)
libGL error: pci id for fd 34: 1002:67df, driver (null)
Segmentation fault

I assume the same would happen with chromium. Firefox starts and runs apparently fine but one of its threads also seems to segfault:

Code:

[  772.596382] firefox[2472]: segfault at 0 ip           (null) sp 00007ffee1246658 error 14 in firefox[400000+24000]

glxheads, glxinfo work fine.

orbea · 10-06-2019, 10:06 AM

If you have a recent amd GPU you really should not use amdgpu pro and should use the the free drivers instead which are much better maintained and tested. The only caveat is that there are likely many issues fixed with a graphics stack newer than 14.2, this means updating llvm, libdrm, mesa, the kernel and maybe xorg.

bassmadrigal · 10-06-2019, 03:00 PM

Quote:

Originally Posted by gaitos

The 17.10 build script no longer included the patch for the standard 4.4 kernel. However, after copying the patch from 16.60 and modifying the SlackBuild script to apply it, everything is working properly.

I added the 4.4 patch back to the SlackBuild. Thanks for the testing and update!

gaitos · 10-07-2019, 11:50 AM

Since I can't get vivaldi or opera to run, I'm going to try and use the free version. Apart from updating my whole system to -current (and I'd rather not do that yet), do you think I could try and compile only the relevant packages? I assume simply installing the compiled ones from -current won't be a good idea.

Side-note: before amdgpu pro, I tried compiling the newest stable kernel (5.3.4) just to see what would happen. First thing I noticed, the screen went black when drm initialized (when switching from 80x25 textmode to graphics console). Logging in blindly and typing startx works, but still with software rendering, of course. Exiting X returns to a blank text screen with a blinking cursor (no text output is visible).

Looking at dmesg (this is with 5.3.4, no amdpro), it looks like it's complaining about missing firmware (this is before starting X)

Code:

[    6.647869] amdgpu 0000:01:00.0: Direct firmware load for amdgpu/polaris10_pfp_2.bin failed with error -2
[    6.650512] intel_rapl_common: Found RAPL domain package
[    6.650517] intel_rapl_common: Found RAPL domain core
[    6.654715] amdgpu 0000:01:00.0: Direct firmware load for amdgpu/polaris10_me_2.bin failed with error -2
[    6.663940] amdgpu 0000:01:00.0: Direct firmware load for amdgpu/polaris10_ce_2.bin failed with error -2
[    6.683366] [drm] Chained IB support enabled!
[    6.705246] amdgpu 0000:01:00.0: Direct firmware load for amdgpu/polaris10_mec_2.bin failed with error -2
[    6.725885] amdgpu 0000:01:00.0: Direct firmware load for amdgpu/polaris10_mec2_2.bin failed with error -2
[    6.732888] amdgpu: [powerplay] hwmgr_sw_init smu backed is polaris10_smu
[    6.790520] [drm] Found UVD firmware Version: 1.130 Family ID: 16
[    6.817584] [drm] Found VCE firmware Version: 53.26 Binary ID: 3
[    6.818293] amdgpu 0000:01:00.0: Direct firmware load for amdgpu/polaris10_k_smc.bin failed with error -2
[    6.818409] [drm:amdgpu_cgs_get_firmware_info [amdgpu]] *ERROR* Failed to request firmware
[    6.856564] amdgpu: [powerplay] SMU Firmware start failed!
[    6.856567] amdgpu: [powerplay] Failed to load SMU ucode.
[    6.856569] amdgpu: [powerplay] fw load failed
[    6.856570] smu firmware loading failed
[    6.856572] amdgpu 0000:01:00.0: amdgpu_device_ip_init failed
[    6.856575] amdgpu 0000:01:00.0: Fatal error during GPU init
[    6.856580] [drm] amdgpu: finishing device.
[    6.996567] WARNING: CPU: 2 PID: 645 at drivers/gpu/drm/amd/amdgpu/amdgpu_object.c:929 amdgpu_bo_unpin+0xbe/0xe0 [amdgpu]

I assume that besides the kernel I should get some firmware as part of the free drivers? If so, where from?

At present I only care about it working, either with proprietary or free drivers.

Thank you all for your patience!

bassmadrigal · 10-07-2019, 03:20 PM

You can get updated kernel firmware by installing the package from -current. Most of the time, it isn't recommended to pull packages from -current, but this is just repackaging the files from kernel.org. If you're at all wary about it, you can just grab the SlackBuild for kernel-header package and run it. It will clone the kernel-firmware git and then make a Slackware package out of it.

Then run upgradepkg on it.

orbea · 10-07-2019, 04:03 PM

Using Pat's Slackbuilds from current (Not pacakges) to rebuild the entire graphics stack on 14.2 should be safe and the main potential problem are build issues.

Like bassmadrigal said the current package for linux-firmware is safe, its easy to build a newer kernel without making a package following the directions on slackdocs and you can leave the kernel-headers package alone. As for the vulkan-sdk try building it from SBo, it has some build fixes for 14.2.

gaitos · 10-08-2019, 11:55 AM

Just for a first simple experiment I've installed the 4.19.77-huge kernel from -current (I think that for the kernel it doesn't matter if it's installed as a binary or recompiled) and indeed amdgpu no longer complains about lack of firmware, however...

Code:

[    8.780604] ------------[ cut here ]------------
[    8.780746] WARNING: CPU: 3 PID: 751 at drivers/gpu/drm/amd/amdgpu/../display/dc/gpio/gpio_base.c:64 dal_gpio_open_
ex.cold+0xc/0x14 [amdgpu]

..followed by the full debug info (let me know if I need to post that, it's rather long) and after that, something that is quite rare nowadays:

Code:

[    8.925929] BUG: unable to handle kernel NULL pointer dereference at 0000000000000008

To be clear, this happens right at kernel initialization of drm and fbcon (on boot), without X being started. What is more, if I reboot after booting with 4.19.77, the card fails to initialize, I have to push the reset button in order to get an image on the monitor.

This warning and NULL pointer BUG don't happen with amdgpu pro (in fact the only trouble I have with that approach is that chromium based browsers don't work).

Before posting this I have tried the 5.3.4 kernel again and now that there is a firmware present it doesn't complain about that (and fbcon properly initializs), however it somewhat randomly (i.e. 1 out of 5 cold reboots, but on every warm reboot) gives similar warnings in dmesg (and on the console, of course), but no NULL pointer BUG messages. It also leaves the card in a better state, so that warm reboots work.

I don't know if these warnings would prevent the rest of the amdgpu drivers (xorg and so on) from working, but I suspect so!

I will try and get the 5.3.4 firmware to see if there are any changes.
The newer firmware is the same for this card (same version) so that doesn't change anything.

However, should I have known this I would have went with nVidia, it's funny that I bought AMD after reading that they have better open-source drivers. Right!

Also: should I create a new thread for this? It's getting somwhat unrelated to amdgpu pro.