Support dgRMatrix, dgCMatrix and dgTMatrix

[kafku]

This PR includes the following features.

• new typed_sparray<af::dtype, af::storage> class
• Exporter and wrap fucntion that support dgRMatrix, dgCMatrix and dgTMatrix in Matrix package

[kafku]

I'll add Expoter for lgCMatrix and lgRMtarix. So, please wait till it's done.

[kafku]

I found that af_create_sparse_arry only works with f64/f32/c64/c32. So this PR does not support lgCMatrix and lgRMatrix.

[rstub]

Thanks a lot! For reference, the data type limitation for sparse arrays is documented at http://arrayfire.org/docs/group__sparse__func.htm

Do you have an idea for a good unit test?

[kafku]

Do you have an idea for a good unit test?

How about defining a function that returns given typed_sparray without doing anything to check if it returns the same matrix? If it is okay, I'll add some tests to this PR in this way.

By the way, I have a question. In this PR, I added typed_sparray<af::dtype, af::storage>. However, it is also possible to define typed_array<af::dtype, af::storage> (with AF_STORAGE_DENSE as its defulat value), and then I replace old typed_array<af::dtype> with typed_array<af::dtype, AF_STORAGE_DENSE>. Which do you think is better?

[rstub]

How about defining a function that returns given typed_sparray without doing anything to check if it returns the same matrix? If it is okay, I'll add some tests to this PR in this way.

Sounds excellent (we don't need to test AF capability after all). Just go ahead.

By the way, I have a question. In this PR, I added typed_sparray<af::dtype, af::storage>. However, it is also possible to define typed_array<af::dtype, af::storage> (with AF_STORAGE_DENSE as its defulat value), and then I replace old typed_array<af::dtype> with typed_array<af::dtype, AF_STORAGE_DENSE>. Which do you think is better?

Good question. My gut feeling would be the latter approach, since there is only one af::array and typed_array<> is meant as a thin rapper around it. How is it w.r.t. implementation? Would it be more complicated this way?

[kafku]

Would it be more complicated this way?

I don't think it would be that complicated. I'll fix this PR with the latter approach.

[kafku]

update:

• use typed_array<af::dtype, af::storage> instead of typed_sparray
• added simple tests to check the Exporter and the wrap function for sparse array.

[dselivanov]

Thank you gents for this! super curios to try to port some of matrix factorizations from my rsparse pkg.