Request for least squares solver using SVD

[adayton1]

Your Environment (If applicable)

• Multiple different platforms: Windows 10-11, macOS Monterey and newer, and Linux (RHEL 8 based)
• Compilers: amdclang (ROCm 6), nvcc (CUDA 11), intel (2022), clang (12-14), gcc (10-12)

Describe the problem

Description

I would like to request a least squares solver routine that uses SVD for robustness. I'm especially looking for the __device__ versions (these solves take place in the middle of large GPU kernels) of these routines. My application can be compiled for both NVIDIA and AMD GPUs. LAPACK has gelsd and gelss, but I'd be open to other algorithms using SVD if they are more amenable to GPU execution.

At the moment my application is using a handrolled solver, but I can't just replace it outright with LAPACK because I need GPU support. I am eager to see the performance difference from using an optimized linear algebra library!

The sizes of the systems I am solving are small (2x2, 3x3 and 6x6).

Additional context

https://www.netlib.org/lapack/explore-html/d9/d67/group__gelsd.html
https://www.netlib.org/lapack/explore-html/da/d55/group__gelss.html
https://www.netlib.org/lapack/lug/node71.html (Fig 3.3)

Desired timelines

Within 3-12 months would be appreciated.

[adayton1]

My apologies, this was the wrong magma repo.