Arch Linux User Repository

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.

Here's my patch to get yourself version 18.40

https://gist.githubusercontent.com/rigred/d5df1647d4c718dd28c1a9035cadbdd7/raw/25c443d1c1e60fe61efe808c2e5ff83e2cbf7f35/opencl-amd-18.40.patch

Let me know if anyone finds any issues upgrading

To solve the blender/libdrm related issues you can use:

LD_PRELOAD=/usr/lib/libdrm_amdgpo.so.1.0.0

I found a better solution here: https://bugs.archlinux.org/task/60061

Upgrade your system, compile libdrm and just save libdrm_amdgpu.so.1.0.0 library and preload it to Blender. It works fine here.

Does anyone know if the blender issue

amdgpu_device_initialize: amdgpu_query_info(ACCEL_WORKING) failed (-9)

when using openCL and cycles, currently I'm running a downgraded libdrm (version 2.4.93-1) to solve this; or workaround.

If this needs more work or investigation I'll happily help debug this, I think the problem is in AMD's upstream driver that opencl_amd uses. Please correct me if I'm wrong...

@linnaea only AMD can fix the issue but most likely they will never fully develop SI and CIK support.

@Olympus593 I just gave up and ended up using pci passthrough with VM running an older distro. I'm toying with cltorch here.

That crash is probably caused by the same problem causing OpenCL CTS to segfault. CTS segfaults randomly during async copy to host, and if it doesn't segfault, only the first several hundred bytes of the data can pass CTS verification so it's still worthless.

Feels like a timing issue to me, but there's not much anyone can do about that.

@linnaea I am able to run DaVinci Resolve with only one problem. When using power windows and GPU accelerated fusion, it crashes. Mostly because it does not even meet minimum standards when it comes to usability.

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.

Updated to 18.40 and latest linux and linux-firmware package and still very buggy SI support. Or maybe I have everything misconfigured.

@urbenlegend Does this opencl works if you hold back libdrm to ver 2.4.93 with the updated package?

EDIT: Just tested the new PKGBUILD. Updated package works as long libdrm is at ver 2.4.93

As long AMD supplies libdrm at ver 2.4.92, the last backward compatible version is upstream's libdrm 2.4.93 as hinted in the following line

ar x "${srcdir}/${prefix}${major}-${minor}-${targetos}/libdrm-amdgpu-amdgpu1_2.4.92-${minor}_amd64.deb"