Randomness

[MaardinK522]

Hello, Juan Fumero.
I am really excited to complete my custom implementation of a JNeuralNetwork from scratch on my NVidia 1050Ti GPU. But, In the implementation I needed an object of random number generator class. Is there any way I can implement randomness into my code?
Or Should I just go sequentially on my GPU?
I hope I am able to explain you my request.
And again Thank you very much for releasing TornadoVM build scripts.

[jjfumero]

Hi @MaardinK522 . Thanks for your interest. This is a topic I personally would like to explore but unfortunately, no immediate plans.

So, to answer your question, TornadoVM does not have any utility for random number generators, but we have examples that can help, for instance the Montecarlo-based for pseudo random number genenerator:

https://github.com/beehive-lab/TornadoVM/blob/master/tornado-examples/src/main/java/uk/ac/manchester/tornado/examples/compute/MonteCarlo.java#L41-L64
Hope this helps.

Also, as feedback for us (TornadoVM team), is there anything specific you would like to see?

[MaardinK522]

Yes.
Every time I am executing something it says:
[Bailout] Running the sequential implementation. Enable --debug to see the reason.
I don't how to fix it?

[jjfumero]

This means that the code gen failed. Which code are you running? Is it possible to share a test-case that breaks?

You can use the --debug and --fullDebug to see the cause.

[MaardinK522]

`public class Demo {
    public static void main(String[] args) {
        // TornadoVM version: 1.0.7
        Matrix2DDouble matrix1 = new Matrix2DDouble(5, 5);
        Matrix2DDouble matrix2 = new Matrix2DDouble(5, 5);
        Matrix2DDouble matrix3 = new Matrix2DDouble(5, 5);

        TornadoAPIProvider.loadScheduleRuntime("s0");
        var device = TornadoExecutionPlan.getDevice(0, 0);
        System.out.println(device);

        TaskGraph taskGraph = new TaskGraph("Testing")
                .transferToDevice(DataTransferMode.FIRST_EXECUTION, matrix1, matrix2)
                .task("", MatrixOperationSolver::fill, 1.0, matrix1, matrix2, matrix3)
                .task("addition", MatrixOperationSolver::solveAddition, matrix1, matrix2, matrix3)
                .transferToDevice(DataTransferMode.EVERY_EXECUTION, matrix1);
        ImmutableTaskGraph immutableTaskGraph = taskGraph.snapshot();
        try (TornadoExecutionPlan plan = new TornadoExecutionPlan(immutableTaskGraph)) {
            plan.withDevice(device).withProfiler(ProfilerMode.CONSOLE).execute();
        } catch (TornadoExecutionPlanException e) {
            throw new RuntimeException(e);
        }
        System.out.println(matrix3);
    }
}
public class MatrixOperationSolver{
   public static void solveAddition(Matrix2DDouble matrix1, Matrix2DDouble matrix2, Matrix2DDouble result) {
        if (matrix1.getNumRows() != matrix2.getNumRows() && matrix2.getNumColumns() != matrix2.getNumColumns())
            return;
        for (@Parallel int i = 0; i < result.getNumRows(); i++)
            for (@Parallel int j = 0; j < result.getNumColumns(); j++)
                result.set(i, j, matrix1.get(i, j) + matrix2.get(i, j));
    }
}
`

Here is the code I am trying to execute.

[jjfumero]

I suspect the error in the codegen is related to the if condition within the solveAddition method, since it is before a parallel region. This is clearly a bug in TornadoVM. As a work-around, I would remove the if and do the checks on the host side (on the CPU) before invoking TornadoVM.

public class MatrixOperationSolver{
   public static void solveAddition(Matrix2DDouble matrix1, Matrix2DDouble matrix2, Matrix2DDouble result) {
        for (@Parallel int i = 0; i < result.getNumRows(); i++) {
            for (@Parallel int j = 0; j < result.getNumColumns(); j++) {
                result.set(i, j, matrix1.get(i, j) + matrix2.get(i, j));
            }
        }
   }
}

[jjfumero]

We will look into this bug.

[jjfumero]

@mairooni is taking a look to this issue

[mairooni]

Hi @MaardinK522, could you also provide the code of the function MatrixOperationSolver::fill?