ROCm 3.7 Has Open-Source OpenCL Image Module Included

[phoronix]

Phoronix: ROCm 3.7 Has Open-Source OpenCL Image Module Included

With last week's release of ROCm 3.7 for the Radeon Open Compute stack, it turns out there is a welcome change that wasn't published as part of the announcement...

http://www.phoronix.com/scan.php?pag...7-OpenCL-Image

[Jabberwocky]

Can someone explain this? https://github.com/RadeonOpenCompute/ROCm/issues/1180

Specifically the comment by a community member: "AMD has a separate product line called CDNA for computing. I'm afraid RDNA architecture is not suitable for ML by design."

I'm assuming it's related to hip stack and not opencl, but still wondering what's going on.

[santoniuk]

Can someone explain this? https://github.com/RadeonOpenCompute/ROCm/issues/1180

Specifically the comment by a community member: "AMD has a separate product line called CDNA for computing. I'm afraid RDNA architecture is not suitable for ML by design."

I'm assuming it's related to hip stack and not opencl, but still wondering what's going on.

It's a completely stupid policy of AMD. Currently I'm able to use OpenCL with RX 5700 on Fedora using rocm 3.3 so RDNA is capable to do ML tasks.

Rocm 3.5 and 3.7 do not work.

[zaphod_]

Jabberwocky I would also be interested in using amd gpus for compute/ml tasks but it seems like amd is just not interested in compute/ml...
This news is a first step but since rdna1 is still not working with rocm, I dont think rdna2 will work in the foreseeable future. My only hope for an working opensource ml/compute gpu is now intel... But for now I'll just buy a geforce 3080/90...

[darkbasic]

This SUCKS. AMD please fix this...

[nuetzel]

'OpenCL Image support is generally needed by imaging applications with the likes of Darktable and DaVinci Resolve,...'

Blender, too?

Now, do it (ROCm) for RX5xx (580) without PCIe atomics, please!
(Marek had it working @ the beginning.)

BTW Back from vacation. ;-)

[bridgman]

Can someone explain this? https://github.com/RadeonOpenCompute/ROCm/issues/1180

Specifically the comment by a community member: "AMD has a separate product line called CDNA for computing. I'm afraid RDNA architecture is not suitable for ML by design." I'm assuming it's related to hip stack and not opencl, but still wondering what's going on.

I'm not aware of anything in RDNA that makes it "not suitable for ML by design" or even by accident. One is optimized for large scale compute, the other is optimized for graphics, but the only exclusion is that CDNA parts generally won't be including a graphics pipeline. They do include video encode/decode though.

This SUCKS. AMD please fix this...

What sucks ? We open sourced one of the last bits of closed source in the stack.

If you're talking about lack of official support in ROCM for Navi parts that is being worked on and we are trying to get funding to work on it faster.

[darkbasic]

and we are trying to get funding to work on it faster.

Finally, I guess better late than never...

[oleid]

If you're talking about lack of official support in ROCM for Navi parts that is being worked on and we are trying to get funding to work on it faster.

I know it's probably very hard to say and you are probably not free to say...
But I'm wondering what time frame to expect for RNDA2, once hardware is available. More like instant support, or more like another year till it's usable ?

[MadeUpName]

This has the potential to be HUGE for me. Unfortunately I am on Fedora and FC32 is still stuck on the rocm-runtime 2.0.0-5 which I think was built while George the third was still on the thrown. I had a look in Koji and there is a version 3.5 for FC33 sitting in there so this may filter through fairly soon. Not having to reboot several times a day into RH in order to use the functionality of my graphics card will save me a crazy amount of time and vastly improve my productivity. It might also mean I don't have to switch to NVidia for my next graphics card. There are no two ways about it. AMD done good with this one.

https://koji.fedoraproject.org/koji/...ackageID=26312