Radeon ROCm 2.1 Bringing RocTracer Preview, Improved DGEMM Performance

[phoronix]

Phoronix: Radeon ROCm 2.1 Bringing RocTracer Preview, Improved DGEMM Performance

It looks like AMD is about to release ROCm 2.1 for the Radeon Open Compute stack...

http://www.phoronix.com/scan.php?pag...Cm-2.1-Prepped

[Azpegath]

Slow and steady progress, that's what we like! Great work, AMD!

[Meteorhead]

+1: RocTracer
+1: rocm-smi (PCI-E usage)
-1: No SPIR
-1: No SPIR-V
-1: No Windows
+-: CPU?

All I want is SYCL to work with OpenCL-OpenGL interop. OpenCL 1.2 would do, SYCL 1.2.1, and OpenGL 3.3. This is the stuff I use daily. SYCL via ComputeCpp (SPIR, SPIR-V) or the new Intel SYCL compiler (SPIR-V).

[Naquatis]

Did anyone tried to build this dependency monster from source? It begins with a custom python setup and after that really quick spit out a 404 Not Found!

Get https://gerrit.googlesource.com/git-repo/clone.bundle
Get https://gerrit.googlesource.com/git-repo
Get https://github.com/RadeonOpenCompute/ROCm.git
  % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                 Dload  Upload   Total   Spent    Left  Speed
  0     0    0     0    0     0      0      0 --:--:-- --:--:-- --:--:--     0
curl: (22) The requested URL returned error: 404 Not Found
Server does not provide clone.bundle; ignoring.
fatal: Couldn't find remote ref refs/heads/roc-2.1.0
fatal: Couldn't find remote ref refs/heads/roc-2.1.0
fatal: cannot obtain manifest https://github.com/RadeonOpenCompute/ROCm.git
error: in `sync`: [Errno 2] No such file or directory: '/home/erocm/ROCm/.repo/manifests/.git/HEAD'
error: manifest missing or unreadable -- please run init

I am waiting since ages to get a non closed source OpenCL >= 1.2 for my RX 480 and with the idea to get a Radeon VII I would solve this dependency puzzle if I only could get the sourcecode in a managable way!

+ already reported: https://github.com/RadeonOpenCompute/ROCm/issues/699

Ok the first problem got solved and leads to the next issue:

error: Exited sync due to fetch errors

+ already reported: https://github.com/RadeonOpenCompute/ROCm/issues/700

Ok this issue is also solved .. lets see how all this pieces stick together.

[schmidtbag]

Oooh SMI sounds nice.

[bridgman]

+ already reported: https://github.com/RadeonOpenCompute/ROCm/issues/700

700 should be solved as well now, can you please try again at your convenience ?

[Naquatis]

700 should be solved as well now, can you please try again at your convenience ?

Yep, thanks .. lets find a way to all this sources.

[tuxd3v]

Did anyone tried to build this dependency monster from source? It begins with a custom python setup and after that really quick spit out a 404 Not Found!

Yes, and I hit the wall
At the time mainline kernel was not compatible with rocm, now is different,

But anyway I hit the wall again some time ago on hardware restrictions..
You cannot have Rx480 + PCIe3.0 only
AFAIK, you need Rx480 + pcie3.0( with PCIe atomic operations support...in all connection chain, cpu/mobo/GPU )

So on Kavery, and other CPUs marked for HSA, paired with rx480/rx580, no rocm!

[Naquatis]

Yes, and I hit the wall
At the time mainline kernel was not compatible with rocm, now is different,

But anyway I hit the wall again some time ago on hardware restrictions..
You cannot have Rx480 + PCIe3.0 only
AFAIK, you need Rx480 + pcie3.0( with PCIe atomic operations support...in all connection chain, cpu/mobo/GPU )

So on Kavery, and other CPUs marked for HSA, paired with rx480/rx580, no rocm!

Sounds like problems while using ROCm on an APU + RX 480. That is not what I try to archive. The reason why I do all this ROCm stuff is that the old opencl-amd package had problems to use newer libdrm versions (well that is not the case anymore). Setting up ROCm is not as easy as I thought. Even compiling every single ROCm package by a time will need more knowledge. Would be nice to get some kind of documentation how this packages should be compiled and work together. I see it comes with a whole clang/llvm environment.