Finally made Multik wrapper work. Performance for ...

# mathematics

[altavir]

Finally made Multik wrapper work. Performance for linear add operations is of the same order for KMath-specialized buffer, Viktor and Multik (about 10-20% difference which is not relevant). It is much faster for @Ролан's tensor in-place operations (like 300%). I guess that memory allocation is the most expensive thing for simple operations. In KMath it is possible to do memory buffer pooling inside fixed size algebra context. I wonder if it makes sense to do.

[Ilya Muradyan]

Is it possible to operate on tensors over finite fields with KMath using Multik engine?

[altavir]

I am not sure what you mean by finite. But I've just implemented a full connector for basic ND algebra and partial connector for Rolan's tensor algebra. Some operations for tensors need to be reimplemented because they use featuresmissing in multik.

From here we can construct some advanced features. For example, we can integrate a tensor field.

[Ролан]

@altavir "It is much faster for @Ролан's tensor in-place operations (like 300%)" what's that?

[altavir]

I mean that in-place tensor operations in your tensor algebra are like 3 times faster.

[altavir]

Regular operations are a bit faster, but within error margin. I did not test multik in-place operations though.

[altavir]

I've added multik in-place additions and it is consistent with tensor algebra in-place additions (a bit fater, but within error margin).

So in-place operations are significantly faster.