Passing in a class, running functions and updating the class data

[ggiavelli]

I notice that you streamout a variable, but I am working on a connectionist architecture where I want to pass in an array of class that has variables internal to the class, have it compute, and then remain updated. Does this require streamingOUT each class then reading it back in as serialized or some such strategy?

I was also wondering how to have a global memory array which the passed in class would update

thanks!

[jjfumero]

Hi @ggiavelli , thank you for your interest in TornadoVM.

I am not sure I understood your question.

The calls streamIn and streamOut will force a copy (from the host to the device and vice versa) every time the execute method is invoked.

For instance:

 ts.streamIn(a, b)
     .task("task1", ... )
     .task("task2", ...)
     .task("taskN", ...)
     .streamOut(result1, result2, resultN);

ts.execute();

ts.execute();
... 

However, there are strategies in TornadoVM to keep the updated data on the device without sending back to the host. This strategy assumes data are not manipulated by the host before executing a task-schedule again. To do so, you will do something like this:

 ts..task("task1", ... )
     .task("task2", ...)
     .task("taskN", ...)
     
ts.execute();

ts.execute();

## Force copy out of the variables of interest 
ts.syncObjects(result1, result2, ..., resultN);

Note that syncObjects is a blocking call, and it will update the result in the variables passed as arguments. Those variables must be already present when invoking some of the tasks.
Here you can find an example:

TornadoVM/benchmarks/src/main/java/uk/ac/manchester/tornado/benchmarks/nbody/NBodyTornado.java

Line 119 in d0f4506

ts.syncObjects(posSeq, velSeq);

For your question:

but I am working on a connectionist architecture where I want to pass in an array of class that has variables internal to the class, have it compute, and then remain updated

TornadoVM supports a subset of Java. For objects support, TornadoVM can run the supported types (primitive arrays, vector types and matrix types). Having said this, however, the TornadoVM JIT Compiler will also inline all objects passed in order to analyze their containing primitive types or object supported types. Nevertheless, we do not recommend this way since it is still very experimental. For this reason, we recommend using arrays of primitives or the supported types:

All supported types here:
https://github.com/beehive-lab/TornadoVM/tree/master/tornado-api/src/main/java/uk/ac/manchester/tornado/api/collections/types

I was also wondering how to have a global memory array which the passed in class would update

What do you mean here? Could you expand on it?

[ggiavelli]

*Question: Does it run under windows 10 as well?* Thanks. Really appreciate the quick reply. This was useful to get started. I will have to experiment with fully serializable objects and see how that goes. Great job with the library it does seem to be a step up from working with opencl directly. In a connectionist architecture there is a neuron which has a state value, an activation function, a propagation function, and about 16 connections to other neuronal units. Each connection also has a weight and may have an activation function. So if you stick to a 2D representation (but most likely things would evaluate in 3D) the challenge is from a parallel standpoint that you have one neuronal unit and the goal is to determine its central value. To do this then you must sum the connections to 16 other neuronal units. However, the entirety of all the neurons can be quite large so you dont want to pass in that large of a structure. I believe the solution would be to maintain the array of neurons outside of the parallelism with synchronized access and then before passing in the neuron for parallel compute you would simply fetch an array as well which would correspond to the inputs. The last complexity is that you want the order of processing to be plane by plane. So if the neural structure is 8 across and 12 down, then you want all 8 to evaluate before evaluating the next row of 8 below. I believe that should be simply a matter of how you construct the loops as they can be out of order within one. But the outer loop which then increments the Y position, that might not be parallelable as you may get execution tasks from a prior row still executing. As they would not then have their core values set yet, this could produce errors. Essentially what has happened is the dim low IQ savages at google pushed a model of neural networks designed for image processing. These luddites have utterly corrupted EVERY LAST API and LIBRARY so that it's all matrix tensor functions. Worse, its high jibberish to extend functions in C. Final horror these libraries are so restrictive even with their so called ability to customize that they cannot break out of this dim savage low IQ luddite image processing model. It's derivation comes from that idiot Hinton and Sutskever at google. So after huge frustrations with THE ENTIRE FRIGGIN WORLD DOING IT WRONG I am hopeful to do something that can create a proper neural modeled universe of different neural types, different connection strategies and functions. Our first product is morphological NLP enterprise search. We built the fastest most precise search engine in the world. No one cared. sigh. The second product is our cognitive engine, which is a bit of a regular conversational language interface to information but through neural activations and propagations not information retrieval techniques. We built a large conversational engine based on our advanced morphological engine and were not happy that it was advanced enough to really be like how humans solve problems. Our next product, where we ran into challenges with the frameworks is NeuralAlert, a warning/insurance system for crypto and commodity stock trading but with application to corporate data as well. We wanted to do time series functions such as acceleration calculations and to do advanced connection architectures. NOPE nothing does it. At that point we began to look at a way to do it ourselves but all the OpenCL / Cuda / Ibm 8 tech edition jvm were all just not really written to make that so easy. The reason Keras and Tensorflow have such limited customizations is because of the low level coding to do so and so they build in many functions they think should cover most situations - if you are a dumb google idiot. So we arrived at tornado and think although its a fair bit of kit to make all this happen, would be massively cool and a technical advantage. I'll start in on it. thanks Gia

…

On Tue, Jan 4, 2022 at 1:49 AM Juan Fumero ***@***.***> wrote: Hi @ggiavelli <https://github.com/ggiavelli> , thank you for your interest in TornadoVM. I am not sure I understood your question. The calls streamIn and streamOut will force a copy (from the host to the device and vice versa) every time the execute method is invoked. For instance: ts.streamIn(a, b) .task("task1", ... ) .task("task2", ...) .task("taskN", ...) .streamOut(result1, result2, resultN); ts.execute(); ts.execute();... However, there are strategies in TornadoVM to keep the updated data on the device without sending back to the host. This strategy assumes data are not manipulated by the host before executing a task-schedule again. To do so, you will do something like this: ts..task("task1", ... ) .task("task2", ...) .task("taskN", ...) ts.execute(); ts.execute(); ## Force copy out of the variables of interest ts.syncObjects(result1, result2, ..., resultN); Note that syncObjects is a blocking call, and it will update the result in the variables passed as arguments. Those variables must be already present when invoking some of the tasks. Here you can find an example: https://github.com/beehive-lab/TornadoVM/blob/d0f450642603bed9721452acd824f580fd582dac/benchmarks/src/main/java/uk/ac/manchester/tornado/benchmarks/nbody/NBodyTornado.java#L119 For your question: but I am working on a connectionist architecture where I want to pass in an array of class that has variables internal to the class, have it compute, and then remain updated TornadoVM supports a subset of Java. For objects support, TornadoVM can run the supported types (primitive arrays, vector types and matrix types). Having said this, however, the TornadoVM JIT Compiler will also inline all objects passed in order to analyze their containing primitive types or object supported types. Nevertheless, we do not recommend this way since it is still very experimental. For this reason, we recommend using arrays of primitives or the supported types: All supported types here: https://github.com/beehive-lab/TornadoVM/tree/master/tornado-api/src/main/java/uk/ac/manchester/tornado/api/collections/types I was also wondering how to have a global memory array which the passed in class would update What do you mean here? Could you expand on it? — Reply to this email directly, view it on GitHub <#172 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABMIZW2FYQBISKI3OOMVEVTUUKRCNANCNFSM5LF7OYTA> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>. You are receiving this because you were mentioned.Message ID: ***@***.***>

-- --------------------------- Gianna Giavelli, president www.Noonean.com <http://www.noonean.com/> tutto ciò che pensi in un sogno diventerà realtà

[stratika]

Question: Does it run under windows 10 as well?

Hi @ggiavelli, the TornadoVM unit-tests and examples work in Windows 10. We have tested both the OpenCL and PTX backends. The SPIR-V backend has not been tested yet.

[jjfumero]

@ggiavelli ,
thank you for the detailed description of your applications. Along with this, we finished last year an EU project (e2data). One of the companies in the consortium, NEUROCOM (https://www.neurocom.eu/) provided NPL applications written in TornadoVM. Although they did not use neural networks, I think it is a good starting point.

In case of interest, the public deliverable is here:
Deliverable 6.3, Section 3.2.3
https://e2data.eu/download/deliverables