Patches Prep The Merging Of AMDKFD + AMDGPU Linux Drivers

[phoronix]

Phoronix: Patches Prep The Merging Of AMDKFD + AMDGPU Linux Drivers

The plans talked about in early July for merging the AMDKFD driver into the AMDGPU DRM driver are moving ahead and out today are the initial patches working towards this merger...

http://www.phoronix.com/scan.php?page=news_item&px=Patches-Prep-AMDKFD-AMDGPU

[tuxd3v]

I don't know what to think..
The plan is good, but should be done some 1-2 years ago already..

I own a fm2+ Kavery with r7, and mobo with A88X (Bolton-D4) chipset, 32GB RAM 2133Mhz
I have amdgpu for graphics,( no control panel in Debian 9 )
But I still don't have opencl 2.0 in ROCm

So basically today, AMD is telling its Userbase...to buy Nvidia Graphics cards...
Because they work everywere with CUDA, even in a 2008 processor with a chipset from 2006 with pcie1.1 x16, you can get cuda 9.2 working.

I live, and work in the Opensource world, and I like more AMD, than intel or Nvidia, for Several Reasons.
But this reality, is a lot dificult to ignore, in Debian 9, we don't have a option now.

I need to buy some cards in the near future, and I don't know what to do sincerelly.
I like this idea, its a dream, and a must for AMD and its userbase,it's an idea from 2014 I think, we are in 2018..

I hope it really materialize.

[ZiisusChrist]

Could this bring anything new to 7870 stability? With all kinds of driver setups and distros it just crashes the whole pc in almost every single game that there is.
Prime95 and Furmark stable machine on both Windows and Linux. Somehow even Unigine's Valley benchmark crashes my pc.

If they don't get things sorted out soon i think that Nvidia is the only gpu to buy if you are using Linux.

[s_j_newbury]

So basically today, AMD is telling its Userbase...to buy Nvidia Graphics cards...
Because they work everywere with CUDA, even in a 2008 processor with a chipset from 2006 with pcie1.1 x16, you can get cuda 9.2 working.

I think this unlikely. I have a 2012 GPU in my laptop (NVS5200M) which is only supported up to CUDA7 and gcc-4.9. The situation with support for non-current GPUs with NVIDIA compute is certainly not better than AMD. At least with AMD it is possible to work to improve support for compute with old pre-GCN chips as the information is out there, and you're not tied to old insecure toolchains. To be honest, I find the the way NVIDIA ties their SDK/CUDA compiler to specific releases of GCC quite a horrible hack.

[fuzz]

Could this bring anything new to 7870 stability? With all kinds of driver setups and distros it just crashes the whole pc in almost every single game that there is.
Prime95 and Furmark stable machine on both Windows and Linux. Somehow even Unigine's Valley benchmark crashes my pc.

If they don't get things sorted out soon i think that Nvidia is the only gpu to buy if you are using Linux.

You must be kidding. AMD's more recent GPUs have much better support. Also that sounds like an individual hardware issue, unless Unigine Valley doesn't crash on Windows.

[tuxd3v]

I think this unlikely. I have a 2012 GPU in my laptop (NVS5200M) which is only supported up to CUDA7 and gcc-4.9. The situation with support for non-current GPUs with NVIDIA compute is certainly not better than AMD. At least with AMD it is possible to work to improve support for compute with old pre-GCN chips as the information is out there, and you're not tied to old insecure toolchains. To be honest, I find the the way NVIDIA ties their SDK/CUDA compiler to specific releases of GCC quite a horrible hack.

I understand the Cuda limitation for Hardware, but CUDA its a lot more extensive..

i.e: a 8400GS cuda1.1 is last supported by Cuda Toolkit 6.5(for Cuda <=1.1)
A GTX 1060 6GB is Cuda 6.1 capable ( CUDA toolkit 9.2 and later ones to be released, with support >=Cuda 3 )

Both are suported in a core2duo 7200 and a mobo p5gcmx1333 ddr2 533(pcie1.1), and in what cuda is related, with minor performance hit, a machine from last decade.

I understand , and its also my opinion, AMD is doing a Great work, in bringing a common infraestrusture that supports both, proprietary and opensource.
I also tend to go AMD, not only because of the color...but because inovation on x86 platform basically have been done by AMD, and the advances in Graphics merged with cpus, and so on is also bring by AMD, a lot of things to resume, of course including their opensource effort, which I prise.

But then, you look into Rocm, and basically it supports intel cpus, only rizen cpus and needs pcie3.0.

It doesn't even support Opencl 2.0 on carrizo and Kavery, so what this people will do when they buy a new card?
They are locked on fm2+, and CUDA works great there..that was my point.

I know that the majority of problems will be solved by Rocm 1.3, at least was the idea I get, and they will align their Structures with Linux kernel, which will help to compile the necesary kernel modules via dkms, easier, that will come in kernel 5.0.

I which that will hapen.

[pal666]

But I still don't have opencl 2.0 in ROCm

So basically today, AMD is telling its Userbase...to buy Nvidia Graphics cards...
Because they work everywere with CUDA

but you asked for opencl 2.0. they don't work with opencl 2.0 at all

[theriddick]

Maybe after this is merged AMD will work on OpenCL2.0 support. Thought from what I understand OpenCL is from Apple and so like Apple, they have dropped support for it?

[bridgman]

We are already working on OpenCL 2.0 support and have been for a while. I have mentioned this at least once previously. Expect it to come out on PAL paths first, since PAL gives the widest HW coverage.

[bridgman]

I don't know what to think..
The plan is good, but should be done some 1-2 years ago already..

I own a fm2+ Kavery with r7, and mobo with A88X (Bolton-D4) chipset, 32GB RAM 2133Mhz
I have amdgpu for graphics,( no control panel in Debian 9 )
But I still don't have opencl 2.0 in ROCm

So basically today, AMD is telling its Userbase...to buy Nvidia Graphics cards...
Because they work everywere with CUDA, even in a 2008 processor with a chipset from 2006 with pcie1.1 x16, you can get cuda 9.2 working.

I live, and work in the Opensource world, and I like more AMD, than intel or Nvidia, for Several Reasons.
But this reality, is a lot dificult to ignore, in Debian 9, we don't have a option now.

I need to buy some cards in the near future, and I don't know what to do sincerelly.
I like this idea, its a dream, and a must for AMD and its userbase,it's an idea from 2014 I think, we are in 2018..

I hope it really materialize.

It's not at all clear from your post what you think these changes will do, but it does seem very likely that whatever you have in mind is not what we are actually doing.

This is about building the amdgpu and amdkfd code into a single module rather than two separate modules, so that data sharing between the components will be easier and we can simplify the current interface between them (or at least stop it from getting more complex as we add features).

Nothing to do with forcing people onto ROCM paths or anything like that.

Now, if you are saying that people planning to buy cards from 2006 in 2018 for compute should go with NVidia, we're not telling people that but I wouldn't disagree, since CUDA has been around longer than ROCm. For people buying newer cards, however, different story.