Call for Discussion: New Project: Babylon

Wed Sep 13 22:31:17 UTC 2023

Hi Juan,

On Sep 13, 2023, at 10:03 AM, Juan Fumero <juan.fumero at paravox.ai> wrote:

Hi Paul,
   I think this is a great initiative and very well-needed in the Java world. I have a few questions.

1)
> Babylon will ensure that code reflection is fit for purpose by creating a  GPU programming model for Java that leverages code reflection and is  implemented as a Java library.

Does this mean that one of the goals of the project is to define how GPUs should be programmed using the Code Reflection API, or for Java in General?

The intent is a general approach that depends on the support of code reflection (and Panama FFM).

I think it is up to us, as members of the OpenJDK community, to determine where we head with regards to the GPU programming model, any concrete artifacts that could be produced, and where the dividing lines may be between APIs, implementations, and vendors. Gary can speak more to this than I.

Is Babylon limited to GPUs? Are you also considering other types of accelerators (e.g., AI accelerators, RISC-V accelerators, etc).

In principle it's not limited. As you have shown with TornadoVM the same programming model for GPUs can apply to other forms of hardware that are highly parallel processors, like FPGAs where a program is “printed out” (?) or uniquely arranged in some malleable hardware. In this case, assuming the programming model is applicable, it seems predominantly an area of implementation focus someone could choose to take on in their own implementations.

I think the more specialized the hardware the more limited the programming. So in some cases a parallel programming model may not apply, like with hardware that specializes only in multiplying tensors, which in effect reduces to some form of library calls.

We have other programming models such as TornadoVM [1], which can be programmed using different styles (e.g., loop parallel programs and kernel APIs). How the new model/s will accommodate existing solutions? Is this to be defined?

Again Gary can speak more to this, but I suspect the design will focus predominantly on a range-based kernel model (similar to Tornado’s kernel API). But, in principle I imagine it may be possible to plugin different kernel models (or copy parts of the design) where code reflection could be applied with different and more sophisticated approaches to program analysis and compilation, such as for a loop-based kernel model.

Two key ares of focus I see are:

1) the extraction of kernel call graphs using code reflection, as discussed in Gary’s JVMLS talk. Thus a developer does not have to explicitly build a task graph (as currently required by TornadoVM) and instead a specialized compiler does that work. (Note, it does not render any existing task graph API redundant, it just moves it more into the background as an important lower-level building block where the developer is not required to use it).

2) the ability to call pre-defined “native” kernels that exist in some where else e.g., GPU-enabled library, which may also be a solution for leveraging more exotic but constrained limited hardware.

2)
> We do not currently plan to deliver the GPU programming model into the JDK. However, work on that model could identify JDK features and enhancements of general utility which could be addressed in future work.

Does this mean that the GPU programming model will be only used as a motivation to develop the Code Reflection APIs for different use cases?

3) Is there any intent to support JVM languages with these models (e.g., R, Scala, etc), or will it be specific for the Java language?

It’s specific to the Java language and reflection of Java code.

4) I believe we also need new types. As we discussed in JVMLS this year, we will also need NDArray and Tensor types, Vector types and Panama-based types for AI and Heterogeneous Computing. This is aligned to the Gary's talk at JVMLS [2] in which he proposed the HAT initiative (Heterogeneous Accelerator Toolkit) and Panama-based types. Will be this also part of the Babylon project?

I think we will inevitably explore some of that, and they may be of such “general utility” we could decide to address in future work. However, I am wary of overly focusing on imperfections in this effort, esp. as in many of these cases there is a tendency to focus on syntax rather than the underlying model e.g., arrays (which requires much deeper and careful thinking, but result will be much better for that). It won’t be perfect and we can feed those imperfections into possible future work.

Paul.

[1] https://tornadovm.readthedocs.io/en/latest/programming.html#core-programming

[2] https://www.youtube.com/watch?v=lbKBu3lTftc

Thanks
Juan

On 13/09/2023 01:37, Paul Sandoz wrote:
Hi Ethan,

Current/prior work includes Mojo, MLIR, C# LINQ, Julia [1], Swift for TensorFlow [2], Haskell [3].

In the context of lunch and Python what I had in mind is machine learning and all those frameworks, and I was also thinking about introspection of Python code which IIUC is what TorchDynamo [4] does.

Paul.

[1] https://arxiv.org/abs/1712.03112

[2] https://llvm.org/devmtg/2018-10/slides/Hong-Lattner-SwiftForTensorFlowGraphProgramExtraction.pdf

[3] http://conal.net/papers/essence-of-ad/essence-of-ad-icfp.pdf

[4] https://pytorch.org/docs/stable/dynamo/index.html

On Sep 12, 2023, at 12:31 PM, Ethan McCue <ethan at mccue.dev><mailto:ethan at mccue.dev> wrote:

Can you elaborate more on prior work / the state of affairs in other language ecosystems? In the talk you reference Python "eating Java's lunch" - do they have a comparable set of features or some mechanism that serves the same goal (write code in Python, derive GPU kernel/autodiffed/etc. code)?

On Wed, Sep 6, 2023 at 12:44 PM Paul Sandoz <paul.sandoz at oracle.com<mailto:paul.sandoz at oracle.com>> wrote:
I hereby invite discussion of a new Project, Babylon, whose primary goal
will be to extend the reach of Java to foreign programming models such as
SQL, differentiable programming, machine learning models, and GPUs.

Focusing on the last example, suppose a Java developer wants to write a GPU
kernel in Java and execute it on a GPU. The developer’s Java code must,
somehow, be analyzed and transformed into an executable GPU kernel. A Java
library could do that, but it requires access to the Java code in symbolic
form. Such access is, however, currently limited to the use of non-standard
APIs or to conventions at different points in the program’s life cycle
(compile time or run time), and the symbolic forms available (abstract
syntax trees or bytecodes) are often ill-suited to analysis and transformation.

Babylon will extend Java's reach to foreign programming models with an
enhancement to reflective programming in Java, called code reflection. This
will enable standard access, analysis, and transformation of Java code in a
suitable form. Support for a foreign programming model can then be more
easily implemented as a Java library.

Babylon will ensure that code reflection is fit for purpose by creating a
GPU programming model for Java that leverages code reflection and is
implemented as a Java library. To reduce the risk of bias we will also
explore, or encourage the exploration of, other programming models such as
SQL and differentiable programming, though we may do so less thoroughly.

Code reflection consists of three parts:

1) The modeling of Java programs as code models, suitable for access,
   analysis, and transformation.
2) Enhancements to Java reflection, enabling access to code models at compile
   time and run time.
3) APIs to build, analyze, and transform code models.

For further details please see the JVM Language Summit 2023 presentations
entitled "Code Reflection" [1] and "Java and GPU … are we nearly there yet?"
[2].

I propose to lead this Project with an initial set of Reviewers that
includes, but is not limited to, Maurizio Cimadamore, Gary Frost, and
Sandhya Viswanathan.

For code reflection this Project will start with a clone of the current JDK
main-line release, JDK 22, and track main-line releases going forward.
For the GPU programming model this Project will create a separate repository,
that is dependent on code reflection features as they are developed.

We expect to deliver Babylon over time, in a series of JEPs that will likely
span multiple feature releases.
We do not currently plan to deliver the GPU programming model into the JDK.
However, work on that model could identify JDK features and enhancements of
general utility which could be addressed in future work.

Comments?

Paul.

[1] https://cr.openjdk.org/~psandoz/conferences/2023-JVMLS/Code-Reflection-JVMLS-23-08-07.pdf
    https://youtu.be/xbk9_6XA_IY<https://urldefense.com/v3/__https://youtu.be/xbk9_6XA_IY__;!!ACWV5N9M2RV99hQ!Pi_JEFeTachQ7GPUzCbX43Gh_znVj4rdfF5nwlwB6Ge37ghWGq6BLIbq-KlIM2mmm18hSL0CdCRECtQy0Q$>

[2] https://youtu.be/lbKBu3lTftc<https://urldefense.com/v3/__https://youtu.be/lbKBu3lTftc__;!!ACWV5N9M2RV99hQ!Pi_JEFeTachQ7GPUzCbX43Gh_znVj4rdfF5nwlwB6Ge37ghWGq6BLIbq-KlIM2mmm18hSL0CdCRhl4eQWQ$>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/discuss/attachments/20230913/b5d25749/attachment-0001.htm>