The opaque, undocumented nature of mobile GPU is outstanding (not in a good sense).
Intel CPUs are well document both by their documentation or by third party docs (like Patterson-Hennessy.) Although NVIDIA is less document than these CPUs, there are some well-written whitepapers including one for academic conferences. And these are the basis for the textbook coverage.
Compared to that, ARM or Qualcomm GPUs are totally a blackbox. I'd love to learn the specifics of mobile GPUs but there is no clue. I wonder how Collabora folks have figure out the instruction set architecture and other details.
Wasn't aware of this, but cool move on ARM side! Thanks for the tip!
Looking at the code [1], some of files are indeed copyrighted to ARM. So they are actually providing some data. Also, the linked story mentions that qcom is also providing some support to thee opensource driver. Pardon my ignorance here.
While this is true now, Collabora devs did a lot of reverse engineering with little documentation to get a functional driver, then ARM started to pay attention. For a long time, ARM neglected efforts to develop open source drivers.
The key thing about open-source drivers is they retain compatibility with newer Linux kernels, and receive bug fixes and improvements long after the original vendor has lost interest. This reduces e-waste and allows system integrators (like Google with Chromebooks) to easily maintain their hardware far longer than Broadcom, Qualcomm and ARM itself ever would.
It's great that shipping open-source drivers over the proprietary one is fast becoming the norm not the outlier. And this change is not because of hippie free software ideals, but because of cold-hearted business reasons.
I was surprised to find a while back that Google's now ancient "Nexus 7" tablet is allegedly able to run a mainline Linux kernel!
Unfortunately the one I built and "flashed" to it never booted and I lack the soldering skills to build the modified headphone connector/serial TTL cable thing it needs, so it sits in my drawer unused
I really wish we could get to a better upstream ecosystem with ARM devices but it just never seems to happen
They (or maybe the Raspberry Pi foundation, not exactly sure which) were paying someone to work on the open source Mesa driver. Not sure if that's still the case.
This really makes Samsung Exynos chips with Mali GPUs appealing for Linux use, and hopefully when Exynos with RDNA2 comes out, support will be great there as well. I'm not really seeing Mali GPUs and unlocked bootloader phone options in the US (other than the PinePhone), but I'll have to look more another time. Thanks Collabora!
It feels like we now have open drivers (at some stage of the development process) for all the relevant GPUs that are sold today, are there any still missing? Even the Apple M1 GPU has reverse engineering work being done on it.
As part of the new architecture, we are removing some functionality. Most notably:
OpenCL rendering kernels. The combination of the limited Cycles split kernel implementation, driver bugs, and stalled OpenCL standard has made maintenance too difficult. We can only make the kinds of bigger changes we are working on now by starting from a clean slate.
Vulkan Compute shaders (or GLSL compute shaders, if supported by the OpenGL version) should be quite a bit more robust than OpenCL kernels. The main issue w/ shaders compared to OpenCL kernels is some missing features in the former, that will have to be optionally supported via standard extensions.
The programming model is also quite different, so code will have to be rewritten.
I'll piggyback on your comment for a question: I only very recently started getting into the basic of GPGPU coding. I hate vendor lock in, so I started out with OpenCL. I had a good time, but I do understand that more and more of non-CUDA world (however small that may be) is moving away from OpenCL. What are they moving to? Where should one direct one's "compute learning energy" (if one already excludes CUDA or other proprietary solutions)?
Not OP, but: Unfortunately you need to pick a platform and target that. Which API you use depends on the platform. DirectX, OpenGL, Vulkan and Metal all have compute capability.
So, I think you need to abstract your code at the services level, and provide an implementations for each platform you need to support (the number may be 1 or more than 1).
I think you need to direct your "compute learning" energy at the algorithm level. Numerics and all that. The GPU API:s are just the obtuse implementation detail, but not the meat of the thing.
In addition to picking a platform, you also essentially need to pick a generation. Vulkan 1.0 for example is missing subgroups and has lots of other limitations. Vulkan 1.2 on a modern GPU with an updated driver is pretty darned powerful, with an explicit memory model, control over subgroup size, a form of pointers, and a range of scalar datatype sizes (f16 is especially useful for machine learning, commonly having 2x the throughput as f32). It is possible to have multiple versions and select at runtime, but it's an additional complexity that needs to be considered, and can constrain your architecture to accommodate the compatibility modes.
> So, I think you need to abstract your code at the services level, and provide an implementations for each platform you need to support (the number may be 1 or more than 1).
There are some cross-platform capabilities via SPIRV-Cross. The LLVM folks are working on MLIR that might provide even better features wrt. a common abstraction layer going forward.
See https://news.ycombinator.com/item?id=27396634 for a basic intro, but IIRC the main differences involved are (1) support for general pointers in OpenCL, whereas data indirection in compute shaders involves varying levels of messiness, and (2) OpenCL introduces support for some mathematical functions, that is not in Vulkan Compute. (However, note that OpenCL 3.0 has also made many features optional, so the basic featureset may be more comparable as a result.)
Note that they are not removing GPU rendering. Just the OpenCL kernels. They'll still have NVIDIA and AMD kernels. As long as the open source drivers keep working towards supporting the required features those will eventually work on them.
Intel CPUs are well document both by their documentation or by third party docs (like Patterson-Hennessy.) Although NVIDIA is less document than these CPUs, there are some well-written whitepapers including one for academic conferences. And these are the basis for the textbook coverage.
Compared to that, ARM or Qualcomm GPUs are totally a blackbox. I'd love to learn the specifics of mobile GPUs but there is no clue. I wonder how Collabora folks have figure out the instruction set architecture and other details.