There's Finally An Easy Way To Track Mesa's OpenCL Support

[Creak]

But that's just chicken-and-egg. If OpenCL were ubiquitous, then more stuff would use it and it's be a subject of broader interest.

I think that's what would be interesting if Mesa's support started to improve. We could conceivably find ourselves in a situation where you could reasonable assume that a modern Linux machine has it, and then suddenly the floodgates could open for apps to start using it.

Okay, it's a little late for that, by now. At this point, I guess it's more likely that happens with Vulkan compute.

It's not a simple as a chicken-egg problem. Companies in need of GPGPU investigate different solutions. If OpenCL was a viable, reliable and high-performance solution, it would have been chosen when compared to NVIDIA's Cuda. The difference between OpenCL and Cuda is not just marketing. There are all the ecosystem around it that simply.. works well. If OpenCL were to take over Cuda, it has to bring a better alternative. Just being the same but open source is often not enough for a company to make the switch when another solution is already well established.

But you're right, at least having mesa's OpenCL at the same level of features as Cuda would already be a good start. But even on Windows OpenCL isn't the chosen one, so the problem is deeper than just mesa's OpenCL completion, OpenCL is missing something compared to Cuda (I don't know what since I'm no expert TBH).

[qarium]

Agreed, we need to make that easier.

AMD should not make this easier instead it should be the only way...

and move the OpenGL/workstation users to mesa instead of closed source openGL...

[qarium]

Agreed, we need to make that easier.

And please make a chiplet design ot dual-gpu with one RDNA chip and one CDNA chip.

by this way people can use the RDNA plart to play games and use Desktop and use the CDNA part to run Compute/OpenCL

[coder]

Companies in need of GPGPU investigate different solutions. If OpenCL was a viable, reliable and high-performance solution, it would have been chosen when compared to NVIDIA's Cuda. The difference between OpenCL and Cuda is not just marketing. There are all the ecosystem around it that simply.. works well. If OpenCL were to take over Cuda, it has to bring a better alternative.

It's quite simple, really. Because CUDA is proprietary, Nvidia can just do whatever they want with it, and lets them move a lot faster than an open standard. So, OpenCL is never going to win on features or performance. OpenCL's only major* selling point is vendor-independence. That's why OpenGL and Direct3D dominated, and pushed out hardware-specific, proprietary APIs.

However, in order to have meaningful vendor-independence, you need a variety of vendors with high-quality and up-to-date implementations, like we have with OpenGL and DirectX. But that's where it got derailed. Everything just kind of fizzled, before the market could reach critical mass. It wasn't helped by Apple's abandonment of it, and Google's decision to stick with Renderscript, depriving it of even the kind of momentum that Vulkan derived from Android's adoption of it.

But you're right, at least having mesa's OpenCL at the same level of features as Cuda would already be a good start.

That's not what I said. I was talking about being able to depend on it being everywhere. Even with inferior capabilities vs. CUDA, simply having one compute API that you could be sure was available on virtually all users' systems would definitely cause more developers to embrace it and maybe even use it in some places or ways you wouldn't expect. I feel like that's something we eventually saw with OpenGL.

But even on Windows OpenCL isn't the chosen one,

Because Microsoft did its usual thing and created its own DirectCompute as an answer to OpenCL, rather than jumping on the bandwagon and being a supporter.

* From what I've seen of CUDA, I think OpenCL also has a few minor advantages:

• Simpler
• Cleaner
• More self-consistent

Generally, OpenCL seems to have benefited by learning from some of the things Nvidia tried in CUDA. There are also advantages to being second.

[qarium]

I was talking about being able to depend on it being everywhere. Even with inferior capabilities vs. CUDA, simply having one compute API that you could be sure was available on virtually all users' systems would definitely cause more developers to embrace it and maybe even use it in some places or ways you wouldn't expect. I feel like that's something we eventually saw with OpenGL.

we get this with Vulkan Compute.... thats why apple abandonet it because OpenCL is a waste of time.

people have to unterstand that OpenCL was created by Nvidia and Apple... the point that apple abandonet it proof it is Nvidia only.

if we do Vulkan compute all will be fine.

[LinAGKar]

It would be nice if distros actually shipped OpenCL drivers out of the box, like with OpenGL and Vulkan.

[coder]

we get this with Vulkan Compute.... thats why apple abandonet it because OpenCL is a waste of time.

No, Apple abandoned it to create Metal. Vulkan wasn't even a thing, when Apple walked away from OpenCL.

people have to unterstand that OpenCL was created by Nvidia and Apple... the point that apple abandonet it proof it is Nvidia only.

It's proof of nothing. Here's what really happened:

OpenCL was initially developed by Apple Inc., which holds trademark rights, and refined into an initial proposal in collaboration with technical teams at AMD, IBM, Qualcomm, Intel, and Nvidia. Apple submitted this initial proposal to the Khronos Group. On June 16, 2008, the Khronos Compute Working Group was formed^[23] with representatives from CPU, GPU, embedded-processor, and software companies.

if we do Vulkan compute all will be fine.

I don't believe you actually know how the Vulkan and OpenCL host APIs compare. A lot of people seem to latch onto the idea that we should just switch everything to Vulkan, because it's nearly ubiquitous, but they miss that Vulkan is a lower-level API and that makes it a lot more difficult to use. It also raises questions about how well it can scale, given the way that OpenCL reflects the device memory-hierarchy in its task structure.

[dragon321]

Agreed, we need to make that easier.

Nice to hear that. I'm fine with AMDGPU-Pro for OpenCL so easier setup on more distributions would be nice.