3D and 4D support for Linear Algebra functions

[pavanky]

Do this inside a for loop with offsets. This will be needed for GFOR

CUDA

• matmul
• [ ] dot
• lu
• qr
• cholesky
• solve
• inverse

---

OpenCL

• matmul
• [ ] dot
• lu
• qr
• cholesky
• solve
• inverse

---

CPU

• matmul
• [ ] dot
• lu
• qr
• cholesky
• solve
• inverse

---

The first draft can be done inside a for loop using multiple queues / streams.

Note: dot product function has been removed from the above list since a batch dot product can be easily done using matmul function.

[pavanky]

These functions error out in 3.0. Pushing the feature request for a later release.

[ckeitz]

Pavan, can you clarify explicitly whether this means that matrix multiply is not available within gfor? I am currently getting incorrect results when I use it inside gfor for batch sizes larger than ~5 (small matricies on the order of [40,20] x [20, 40]) - interestingly it seems to work correctly in small batches and I'm nowhere near out of memory.

An aside - is there a registry of functions that are and are not supported in gfor that's available somewhere? Maybe we could put something together for easy reference.

Thanks!

[pavanky]

It is most definitely not supported in the open source version. We were not throwing errors before. I changed this behavior couple of days ago. We will add support for it before next release.

[ckeitz]

Thanks for the quick reply! Glad to know the issue 👍

[mchandra]

Will release 3.2 allow for solve() inside a gfor() loop?

Thanks,
Mani

[botev]

Maybe adding and batched version of gaussian kernel would be good as well. Again by batch I mean having 3D tensor inputs, where one dimension is the same.

[InBetweenNames]

What would it take to get this implemented? I could really use this feature in my code. Right now I'm using a workaround that involves tiling out matrices (and transposing one) and doing:

af::sum(A * B)

followed by a moddims to get the proper shape. It does work, although memory usage is a concern and I know this probably isn't the most efficient method.

[Thormidable]

Could I also request this feature

[WilliamTambellini]

CPU matmul seems now to be done from 3.6.2, right ?

[umar456]

Yes, Updated the list