Google blocked OpenCL on Nexus with Android 4.3

renderscript-eats-openclImportant: this is only for Google-branded Nexus phones – other brands are free to do what they want, and they most-probably will.

Also important: this doesn’t mean that OpenCL on Android devices will be over, but that there is a bump in the road now Google tries to lock-in customers to their own APIs.

The big idea behind OpenCL is that higher level languages and libraries can be built on top of it. This is exactly what was done under Android: RenderScript Compute (a higher-level language) was implemented using OpenCL for ARM Mali GPUs.

Having OpenCL drivers on Android has several advantages, such that OpenCL can directly be used on Android and that there is room for other high-level languages that have OpenCL as back-end. Especially the latter is what probably made Google decide to cripple the OpenCL-drivers.

Google seems to be afraid of competition, and that’s a shame, as competition is the key factor that drives innovation. The OpenCL community is not the only one complaining about Google’s intentions concerning Android. Read page 3 of that article to understand how Google is controlling handset-vendors and chip-makers.

Google’s statement

In February OpenCL drivers were discovered on two Nexus tablets using a MALI T604 GPU. Around the same time there was one public answer from Google employee Tim Murray (twitter) why Google did not want to choose OpenCL:

1) The answer is that Android’s needs are very different than what OpenCL tries to provide.

2) OpenCL uses the execution model first introduced in CUDA. In this model, a kernel is made up of one or many groups of workers, and each group has fast shared memory and synchronization primitives within that group. What this does is cause the description of an algorithm to be intermingled with how that algorithm should be scheduled on a particular architecture (because you’re deciding the size of a group and when to synchronize within that group).

3) That’s great when you’re writing for one architecture and you want absolute peak performance, but it gets peak performance at the expense of performance portability. Maybe on your architecture, you have enough registers and shared memory to run 256 workers per group for best performance, but on another architecture, you’d end up with massive register spills with anything above 128 workers per group, causing an 80% performance regression. Meanwhile, because your code is written explicitly for 256 workers per group, the runtime can’t do anything to try to improve the situation on another architecture–it has to obey what you’ve written. This sort of situation is common when moving from architecture to architecture on the desktop/HPC side of GPU compute.

4) On mobile, Android needs performance portability between many different GPU and CPU vendors with very different architectures. If Android were to rely on a CUDA-style execution model, it would be almost impossible to write a single kernel and have it run acceptably on a range of devices.

5) RenderScript abstracts control over scheduling away from the developer at the cost of some peak performance; however, we’re constantly closing the gap in terms of what’s possible with RenderScript. For example, ScriptGroup, an API introduced in Android 4.2, is a big part of our plans to further improve GPU code generation. There are plenty of new improvements coming that make writing fast code even easier, too.

This is no good explanation, but just tries to convince people who have little or no experience with OpenCL or CUDA.

1) Google targets on a subset of algorithms that OpenCL can tackle, then says it only needs the subset. This is where the big disagreement is, and now it is hard, or even impossible, to implement certain algorithms in an optimal way. Google did not foresee a need for FFT, but you can find an implementation anyway.

2) This is exactly the big difference between OpenCL and RenderScript. The RenderScript host-driver tries to take care of the number of GPU/CPU cores being used by the software, but in return you cannot specify the worker-configuration.

3) In OpenCL it is also possible to leave the workgroup-configuration to the driver, just like in RenderScript. The big advantage is that it scales very well and gets better with improved drivers. The bad thing is that not all algorithms can be implemented, or at least not efficiently. The examples Tim gives here, would not get through our code-review!

4) True, this is the continuous discussion on advancing OpenCL. As the processors advance each year and we cannot look into the future, it is quite something to say that there is a solution already.

5) But it still is not possible to get the ID of the worker in Renderscript. OpenCL architects are working on improving the solution, but if algorithms could be implemented without knowing the worker-ID then of course this would never been introduced to the OpenCL model. Tim is claiming that OpenCL/CUDA’s model does not work, but does not provide a solution.

See this Linkedin-thread for further discussion.

Also read the reply from Scott Le Grand in the comments below.

Google blocked OpenCL in Android 4.3

It was announced on Twitter that on Android 4.3 OpenCL stopped working. An ARM intern told that Google blocked access to the OpenCL-driver, such that it can now only be used by the RenderScript-compiler. The block was quite harsh, as it blocked at the LLVM-frontend level in a way that an easy hack was not possible. ARM could not do much to reverse it, as it is the choice of Google and the block is implemented in Android’s core-libraries.

So Google blocked OpenCL drivers on purpose, overruling the decision of their hardware-partners to provide OpenCL-support for their customers.

The only way out seems to go to Linux. An OpenCL-driver for Ubuntu-on-ARM on the Chromebook will be released via the ARM Mali developer webpage in the coming month, probably at the SDK-page.

The exact state is in the below table.

Why is Google against OpenCL?

We can speculate a little around the reasons behind not opting for OpenCL:

  1. OpenCL is an open standard. Google cannot have the influence it wants. Its future depends on the opinion of many, not on theirs alone.
  2. On top of OpenCL various high-level languages can be built, also the high-level languages from competitors.
  3. OpenCL is initiated by Apple. Apple is a competitor of Google in the smartphone OS market. Even though I think it shows strength if a company supports open standards, they probably don’t think that themselves.
  4. Android is fragmented. Google tries to fix this image of the fragmented android space, and must show teeth&balls. OpenCL is the victim.
  5. Avoiding (accelerated) software to be ported to other operating systems, makes Android unique by its apps.
  6. Pushing a proprietary standard is hard, even though getting support for open standards is harder to accomplish. Once a company gets their own proprietary software accepted, the profits raise. In other words: vendor-lockin sells.

So, it actually could be that it is not specifically OpenCL, but any open standard that can be blocked without to much trouble. They tried replacing OpenGL with RenderScript (the RenderScript-API was much larger before), but they did not succeed.

Why not simply choose RenderScript?

RenderScript has various design-decisions which various experienced developers do not agree upon. If there is competition, the bad decisions will be reversed sooner. One unpopular design decision is that RenderScript does not inform the software if it is running on a CPU or GPU, and therefore cannot be optimized for speed. Compare it to driving a car and not knowing you are at an urban road or at a 16-lane motorway – do you drive the same in both situations or optimize your driving-style?

Here is a list of blogs describing RenderScript from an OpenCL perspective:

[list1]

[/list1]

If you want to learn more about RenderScript Compute The official documents page is the place to go since it’s not possible to find much on the net. In any case, please try it out and let me know in the comments in what cases you would prefer RenderScript over OpenCL, and when would you prefer OpenCL over RenderScript.

You can also let me know in the comments what you think my readers should also read. I’ll add the info here.

Share your opinion

Once a new bug-report gets submitted for the OpenCL-driver being blocked on Android 4.3, I’ll add it here.

If you want to add to the discussion on OpenCL-on-Android, please add your comments here:

[list1]

[/list1]

To get OpenCL back on Android, we need to complain, cry and whine. This is how OpenGL got back, and unfortunately the only way there is. If we succeed then innovative projects like AccelerEyes Mobile, ViennaCL and VexCL may end op on Android. If not, then the project developers need to rewrite all the code, but only if the developers want to create two different source-trees and if the limitations of RenderScript do not obstruct porting the already efficient code.

Related Posts

8 thoughts on “Google blocked OpenCL on Nexus with Android 4.3

  1. Pingback: Google revela porque eliminó el soporte OpenCL en Android 4.3 “Jelly Bean” | SINETEC

  2. Scott Le Grand

    Great article…

    I’m not so sure that Google is against competition here, but rather that the RenderScript team is composed mainly of ex-Nvidia and ex-Intel driver developers, for whom I have nothing but respect, but I think they’re completely out to lunch on this one. And that’s because none of them actually wrote much CUDA or OpenCL code to my knowledge so I believe they lack the instinct and insight to do the right thing here.

    5 years ago, I was the guy who ported Folding@Home to CUDA while I was working at Nvidia (ironically alongside much of the current RenderScript team when they worked there as well). I bring this up because this effort points out first why a Brook-like language was already obsolete in 2008 (and it’s even more so today) and second that it’s easy to target diverse hardware in OpenCL and CUDA just by using finer-grained task decomposition (but far less so than thread-level ala Renderscript).

    The Folding@Home port exploited 3 major features of CUDA: shared memory, workgroup synchronization, and warp-level programming. This was enough for G80 to best AMD’s otherwise faster GPU running a Brook-like implementation of Folding@Home by a factor of 6, or nearly an order of magnitude. But it didn’t stop there, because the lowest end laptop variant of G80 was still nearly twice as fast as the Brook edition on AMD’s highest end GPU. And it absolutely demolished the Cell in PS3.

    And that’s because workgroup synchronization is trivial in CUDA and
    OpenCL but it requires an entire kernel call to do so in Brook/Renderscript. This is a horrendous waste of resources and battery life. My Galaxy Nexus has a battery life of approximately 4 hours so I want to make every electron count. Forcing developers to go through those cycles of kernel calls just for the purpose of synchronization is just daft IMO.

    Further I cannot disagree more with the assertion that: “That’s great when you’re writing for one architecture and you want absolute peak performance, but it gets peak performance at the expense of performance portability. Maybe on your architecture, you have enough registers and shared memory to run 256 workers per group for best performance, but on another architecture, you’d end up with massive register spills with anything above 128 workers per group, causing an 80% performance regression.”

    And that’s because if one is targeting multiple architectures, all one needs to do is break those 256 workers per group into smaller groups of, say, 64 workers, and run 4x as many. And bingo, the task will now automagically scale up and down architectural variants and even across multiple architectures. Need evidence? Go take a look at what I did after Folding@Home: the GPU port of AMBER runs transparently on every single NVIDIA GPU from GTX280 to GTX Titan. I again think the viewpoint above is something I would expect to hear from a driver writer and not an actual developer with products in the field.

    http://ambermd.org/gpus/

    And while we’re at it: “If Android were to rely on a CUDA-style execution model, it would be almost impossible to write a single kernel and have it run acceptably on a range of devices.”

    To which I can only respond by repeating: AMBER runs transparently on every single NVIDIA GPU from GTX 280 to GTX Titan, spanning three (and arguably a half) markedly different SM architectures. Further, an OpenCL port could be done in a matter of months to make it run on every single OpenCL device in existence from Android phones to Radeon 7970. The reason this hasn’t happened is that I’ve prioritized adding features for AMBER’s userbase and improving its overall performance, but it’s not a hard task for someone with sufficient manycore experience. The hard part is finding people like that when companies like Google continually kick sand in the face of those trying to get that experience.

    Really disappointed in Google here, but not the least bit surprised. I’m not saying you guys should become 100% OpenCL fanboys, but when you’re already using OpenCL as a base layer for your own middleware, the least you could do is not get in the way of other people trying to do the same.

    • mpeniak

      Great reply! I am also ex NVIDIA employee (Santa Clara) and could not agree more with your views.

  3. Gourmet

    Google is trying (like the others in this case) to replace Javascript with Dart.
    But, curiously, Javascript is always in Chrome 🙂
    If Google wants to replace such a common thing that is OpenCL it should to it more smooothly like it does with Javascript.
    OpenCL is now a well known standard. If Google wants to replace it it should choose to make it evolve not to fight it.

    There are always good reasons for a leader on a market to cancel a standard that exists on this market (µsoft trying to replace the HTML, and now the javascript, Google doing the same, Apple that replaced the TRRS connector on iPod and iPhone by moving the ground to a ring [has been on the sleeve for decades], IBM, etc). The first reason, non said of course, is to catch the customer base in the nets.

    That’s always a very bad idea!

    db

    • StreamHPC

      Corporations will always do this, but it is not completely a bad thing (and a very good discussion in a political café). In the end the user decides what should be done. We at StreamHPC support OpenCL, as this allows great languages/DSLs like “RenderScript Compute” to be built. (sorry for the late reply)

  4. mdriftmeyer

    “Also important: this doesn’t mean the end of OpenCL on mobile
    devices, but that there is a bump in the road now Google tries to
    lock-in customers to their own APIs.”

    Thanks for the laughter with such hyperbolic rhetoric. I’m not referencing Google, but the notion “this doesn’t mean the end of OpenCL on mobile devices…”

    Which devices? OpenCL is alive and thriving on iOS. You’ve got Android and then what’s left over.

    If you meant to say, “this doesn’t mean the end of installing non-sanctioned/independent OpenCL on Android…” then you have something.

    • StreamHPC

      Changed “mobile devices” to “Android devices” to make it more clear I refer to Android only. I disagree with your suggestion.

      Could you show me the clinfo-output of any iOS device, because I didn’t manage to run OpenCL software on iOS. You seem to have a better experience.

Comments are closed.