Radeon Open Compute 1.3 Platform Brings Polaris & Other Features

[Amarildo]

Main point though is that there are at least three different things that get discussed when we talk about "OpenCL on SI":

- Catalyst/PRO OpenCL on SI via AMDGPU-PRO
- Catalyst/PRO OpenCL (and Vulkan) on SI with Mesa 3D
- fully open source OpenCL on SI

The first one is happening, the second is very likely but with caveats (interop APIs) and I don't think I have *ever* talked about the third (although it's possible I did without realizing context had shifted from one of the other topics).

About the second scenario: Do you mean the opened up version of your closed source OpenCL as you mentioned here? And if so, has Tom decided not to nuke it?

It's not "like we did" it's "like we are doing" and in both cases we have to make some big changes in order to open up the code. In both cases we are replacing the proprietary shader compiler with the LLVM-based open source shader compiler, and in both cases we are rewriting some of the lower level code. In the case of Vulkan we don't have an existing open platform to use so we have to do all that work from scratch, while in the case of OpenCL we already have HSA/ROC and so are using it.

And limiting support for CIK-and-up only :/ But I can see where AMD is coming from.

If it does everything you want but is only partially open source how much does that matter ? Will you actually be modifying the code yourself or just running Blender over it ?

I'll just be running Blender over it. I can't modify code as a programmer does, but I can apply/edit patches and re-compile programs just fine.

How much opensource are we talking, actually? I don't know what "partially opensource" means.

BTW, if I could somehow trick Blender into using proprietary OpenCL files from Catalyst or AMDGPU-PRO, I would. I've tried that, I extracted the catalyst-opencl package and copied the files to their respective places, and renamed most (if not all) to have the same name as the files provided by mesa-opencl. That didn't work :P I guess the way they're loaded and the way they work is different.

I'm having trouble remembering this, and it seems unlikely that I would talk about open sourcing OpenCL without mentioning HSA/ROC in the same paragraph. Can you point me to an example ?

I can't. My apologies.

[bridgman]

About the second scenario: Do you mean the opened up version of your closed source OpenCL as you mentioned here? And if so, has Tom decided not to nuke it?

Just the closed source OpenCL, but working with upstream kernel/X/Mesa etc.. and not requiring AMDGPU-PRO OpenGL (other than maybe for interop).

How much opensource are we talking, actually? I don't know what "partially opensource" means.

Maybe 3/4 open source, just not enough to modify & rebuild yourself.

BTW, if I could somehow trick Blender into using proprietary OpenCL files from Catalyst or AMDGPU-PRO, I would. I've tried that, I extracted the catalyst-opencl package and copied the files to their respective places, and renamed most (if not all) to have the same name as the files provided by mesa-opencl. That didn't work :P I guess the way they're loaded and the way they work is different.

I think you'll find that path gets a lot easier once we get further on amdgpu SI support.

[Amarildo]

Then I'm in, John!

[Marc Driftmeyer]

Which is crap. I can't use it with Blender to render my projects with my GPU.



So? It's not a bad card for 3D rendering neither gaming.


I'm a 3D artist. I've been using OpenCL to render with my R9 270X on Windows and Linux with proprietary drivers, but I *REALLY* don't like them; not only because they're proprietary and not compatible with GRSec, but they're slow as hell (amdgpu and mesa are so much far ahead in performance). The only stack I can use in this case (if I want OpenCL now) is Catalyst, not AMDGPU-PRO because it still doesn't support SI cards. This means almost a one-year-old driver/stack.
AMDGPU with mesa-git works so much better, that the only advantage of using Catalyst would be OpenCL, really. Instead of taking 8 hours to render an image, it takes only 4 if I render with OpenCL and GPU.


I would buy a newer card if I could. Mine was launched in 2013, it's not that old, and since AMD is supporting AMDGPU/Vulkan on all GCN cards I thought AMD would also support my card with OpenCL. I mean, John (bridgman) knows what I card I have, and it's been more than a year that I've been asking him about opensource OpenCL and not once did he mention SI cards wouldn't be supported.

If support for OpenCL 2.0 is already present on GCN 1.0 cards with Catalyst, I don't see why not opensource it and make it available for ALL GCN cards, like they did with AMDGPU/Vulkan.


I'm looking for an opensource OpenCL implementation that runs with current OSS drivers (amdgpu,radeon), that can be used with Blender to render with the GPU, and that works with my card (R9 270X).


I decided nothing, really. Me, you, and one other with the same card as mine, talked about opensource OpenCL, and not once you mentioned SI cards wouldn't be supported. I, for instance, had no idea "ROC" existed.


That's great! But when, though? And will it be as good as the OpenCL present in Catalyst? Meaning, can I use this OpenCL to render 3D projects on Blender using my GPU?
If only a few parts of it aren't opensourced (meaning userspace stuff, only a few pieces), than I don't mind waiting a few months more. I just hope my processor can take the load without frying.




It's not that we "like" to use older hardware, but we can't buy newer one. Or do you think I love rendering via CPU, which is at least twice as slow as my GPU, while frying it (because I can't even afford an after-market cooler)? :T



Mesa is wonderful for OpenGL. However, it's OpenCL implementation is not even close to AMD's proprietary one. Blender's Kernel can't even compile with it.

I'm with you 100% on the OpenCL and Mesa for Blender.

It's borked on my RX 480 8GB for Debian Sid. Even with Mesa 13 and Debian Experimental I get a crashing Blender. So, there is a lot of room for improvements in libdrm-amdgpu, libclc-amdgcn and Mesa OpenCL ICD.

However, the rate of improvements makes it clear that sooner rather than later it will work for us doing work in Blender. That's the good news. The bad news, is it takes more time than I'd expect, but mainly due to resources and business priorities.

My expectation is that OpenCL 1.2/2.x gets sorted out for Mesa 14, libdrm, libclc-amdgcn and either Linux 4.9 or 4.10.

Then again, once its fixed there is still Blender taking its time fixing all the Cycles issues with the OpenCL stack. There again we get help from AMD and 2 dev resources to boot.

[bridgman]

Is it fair to say that other than HPC scenarios Blender is the most important app for Linux OpenCL at the moment ?

I'm trying to make sure our internal testing aligns with what you all are actually doing with the cards.

[Amarildo]

Marc Driftmeyer I honestly doubt we'll get OpenCL working by mesa-14. Do you have any info on it's progress besides this page?

bridgman I'd say Blender is one of the most important programs to benefit from OpenCL in Linux.

[Evil Penguin]

Is it fair to say that other than HPC scenarios Blender is the most important app for Linux OpenCL at the moment ?

I'm trying to make sure our internal testing aligns with what you all are actually doing with the cards.

Folding@home and other distributed computing GPGPU applications are high on my list (I guess that falls under HPC?).
Particularly where OpenMM under CUDA just destroys the OpenCL implementation.

[mibo]

For photo editing/raw conversion darktable.org supports OpenCL (and that worked well with Catalyst).
Video transcoding would be another useful software (AFAIK handbrake only supports OpenCL in Windows).

[juno]

Who does video transcoding in opencl? I mean if you want it fast, you use fixed function hardware, that is sufficient for most private users, IMHO.

@bridgman: iirc, blender cycles did not run at all on AMD for a long time and AMD people made it happen. CUDA is still clearly faster, might also be a nice demo to site off hipification. I think this software is really widely used and one where acceleration is very welcome too.

[bridgman]

No it is not ok for me to use the AMDGPU-pro driver with my hd7850 go get closed source openCL...

Hi Q, that's not what I asked Amarildo - my question was all-open upstream-based stack plus closed source OpenCL on SI.