Initial AMD Radeon "Vega 20" Support For RADV Vulkan Driver

[bridgman]

If you are trying to mix APU and dGPU then your best bet is probably the packaged AMDGPU/PRO driver, installed without the -PRO option but with OpenCL.

Which dGPU are you using ?

[tuxd3v]

If you want to use packaged drivers, can you try an officially supported distro? If you are using a newer distro, the in box drivers should work fine for most asics. Alternatively you can try a ppa is you want to try the latest bleeding edge stuff.

The problem is, that I already tried 16.04,18.04 debian stretch.
Graphics work ok, in igpu and dgpu

OpenCl

With 18.04 and amdgpu-pro I can compile and launch opencl Kernels, only in the IGPU(kavery 7800 - R7 graphics), not in the DGPU, even changing it from pcie slot, and even changing it from x8 to pcie3.0 x16...it fails to execute the compiled code..

clinfo works, everything works(Full opencl1.2 only, for igpu and dgpu), but no juice..
In the igpu it works...some times...some times crash(maybe the driver is not dealing well with 2 graphics..I noticed some time around 2014 that kavery cannot fork exec on graphic does it in cpu and the passes to igpu..it was fixed in carrizo I think..)
disabling the iGPU, its the same...

I can change profile to Compute("4"), and power cap, and so on, and when I launch kernels, it indeed consumes power, it seems that it is processing, but it fails.
A simple helloworld, it tries to launch it, and then says something wrong, it will try on CPU, and prints a lot of trash, on stdout..

This machine were Running on Cuda before, the pcie slots should be fine..

I will give a try to rocm1.9

[agd5f]

Thankyou for your reply.

This is driving me crazy..

If you want to use packaged drivers, can you try an officially supported distro? If you are using a newer distro, the in box drivers should work fine for most asics. Alternatively you can try a ppa is you want to try the latest bleeding edge stuff.

[fuzz]

Thankyou for your reply.

This is driving me crazy..

I need to choose a bunch of Graphics for our employees, and it needs to be or Ubuntu or Debian, but we want to use the APUs we have, attached to display, and the dGPU for Opencl.
I have ordered 2 amd cards( for testing ), but its a dead end..

I tested Rocm1.8.3,
No way, because we have a bunch of kavery machines, and also because of the 'Integer Status' in Struct on linux/fence.h( Kernels <= 4.16 ).
I tried to put a 'int Status',the end of the struct... after compile and reboot...the module segfault, as obvious..

Then I tried AMDGPU-PRO, the last version also without success, on Debian, and on Ubuntu.

On Ubuntu it compiles OK, but no Screen, and no way to edit xorg, or jump into single mode( only via serial attached to motherboard perhaps..).
What seems to me is that /var/log/Xorg.0.log, is empty, seems that nothing happen.

The problem here is Using the Kaverys R7 iGPU for graphics and the dGPU for OpenCl...I thing that when I install AMDGPU-PRO, it smashes mesa driver , and so the Kaverys R7 can´ t work( but I am not sure of it .. ).

I wanted AMD cards, but I don´ t know if this is possible.

You can install amdgpu-pro with --compute-only. it works very well for opencl.

[tuxd3v]

Unless you have somehow gotten your hands on an early 1.9 ROCm stack release you won't have 18.04 support yet:



The big push for recent releases was getting CentOS/RHEL support into the published stacks; I think 1.9 adds support for newer Ubuntu versions as well.

Even more importantly, 1.9 should align user/kernel interfaces with upstream code, allowing you to get most of the functionality via upstream kernels.

Thankyou for your reply.

This is driving me crazy..

I need to choose a bunch of Graphics for our employees, and it needs to be or Ubuntu or Debian, but we want to use the APUs we have, attached to display, and the dGPU for Opencl.
I have ordered 2 amd cards( for testing ), but its a dead end..

I tested Rocm1.8.3,
No way, because we have a bunch of kavery machines, and also because of the 'Integer Status' in Struct on linux/fence.h( Kernels <= 4.16 ).
I tried to put a 'int Status',the end of the struct... after compile and reboot...the module segfault, as obvious..

Then I tried AMDGPU-PRO, the last version also without success, on Debian, and on Ubuntu.

On Ubuntu it compiles OK, but no Screen, and no way to edit xorg, or jump into single mode( only via serial attached to motherboard perhaps..).
What seems to me is that /var/log/Xorg.0.log, is empty, seems that nothing happen.

The problem here is Using the Kaverys R7 iGPU for graphics and the dGPU for OpenCl...I thing that when I install AMDGPU-PRO, it smashes mesa driver , and so the Kaverys R7 can´ t work( but I am not sure of it .. ).

I wanted AMD cards, but I don´ t know if this is possible.

[bridgman]

Unless you have somehow gotten your hands on an early 1.9 ROCm stack release you won't have 18.04 support yet:

Supported Operating Systems - New operating systems available

The ROCm 1.8.3 platform has been tested on the following operating systems:

• Ubuntu 16.04
• CentOS 7.4 &. 7.5 (Using devetoolset-7 runtime support)
• RHEL 7.4. &. 7.5 (Using devetoolset-7 runtime support)

The big push for recent releases was getting CentOS/RHEL support into the published stacks; I think 1.9 adds support for newer Ubuntu versions as well.

Even more importantly, 1.9 should align user/kernel interfaces with upstream code, allowing you to get most of the functionality via upstream kernels.

[tuxd3v]

Another But, this time in Ubuntu, with Rocm:

root@desktop:~# lsb_release -a
No LSB modules are available.
Distributor ID: Ubuntu
Description: Ubuntu 18.04.1 LTS
Release: 18.04
Codename: bionic
root@desktop:~# uname -a
Linux desktop 4.15.0-33-generic #36-Ubuntu SMP Wed Aug 15 16:00:05 UTC 2018 x86_64 x86_64 x86_64 GNU/Linux
root@desktop:~# /opt/rocm/bin/rocminfo
hsa api call failure at line 900, file: /home/jenkins/jenkins-root/workspace/compute-rocm-rel-1.8/rocminfo/rocminfo.cc. Call returned 4104

[tuxd3v]

Looks like that function was removed between the 4.16 and 4.17 kernels... are you sure you are using one of the supported distros / kernels ?

We Use Debian 9,
On Kernels til 4.16, the function exist, on 4.17 it doesn't exist.

AFAIK the highest kernel version in any of the supported distros is 4.15. Which distro/kernel are you installing on ?

The problem is,
On Kernels bellow 4.17 there are another bug,
When Building amdgpu kernel driver, via dkms,
/var/lib/dkms/amdgpu/18.30-641594/build/amd/amdgpu/amdgpu_drv.c:163: ( fence->status == error )

Doesn't exist in kernels <= 4.16

We can build the module, going on header file and adding at the end of the 'Struct fence':
char * status, but I don't know the efect that it has, on the driver when it will be loaded and used?!
Is this safe?

Thanks in Advance,

[bridgman]

Looks like that function was removed between the 4.16 and 4.17 kernels... are you sure you are using one of the supported distros / kernels ?

AFAIK the highest kernel version in any of the supported distros is 4.15. Which distro/kernel are you installing on ?

[tuxd3v]

In the meantime we are still Strugling to install AMDGPU-PRO-18.30, due to a Bug..
When compiling the kernel module:

/var/lib/dkms/amdgpu/18.30-641594/build/amd/amdgpu/amdgpu_drv.c:768:2: error: implicit declaration of function ‘vga_switcheroo_set_dynamic_switch’ [-Werror=implicit-function-declaration]
vga_switcheroo_set_dynamic_switch(pdev, VGA_SWITCHEROO_OFF);

This function doesn't exist in the Kernel sources..

Rocm suffers also from same problem:

https://github.com/RadeonOpenCompute/ROCm/issues/473