SVD on GPU: complex values, interface cleanup (Discussion)

[shamanDevel]

Hi,
pull request #11878 brought an implementation of the SVD on the GPU. But at the moment, only real values are supported.

The current status when applying the SVD (M=USV') on a complex matrix M:

• The python interface declares U,V as complex, S as real
• The C++ kernel definition declares both U,V and S as complex
  (This simplified the CPU implementation using Eigen)
• The python code then immediately casts S to the reals
• The GPU solver (cuSolver) would, however, output the singular values directly as reals

This leads to the following questions / ideas / suggestions
(credit goes also to @rmlarsen for discussion this the first time with me)

• Change the kernel definition: Add a new kernel (V2 suffix) that returns S as a real type
• Implement the complex support on GPUs
• Adopt the CPU code to also use the new kernel definition
  OR
  Keep both definitions and let the python wrapper to choose between the two versions based on the target device (CPU vs GPU)
• ...?

What do you think?

[tensorflowbutler]

Nagging Assignee @rmlarsen: It has been 379 days with no activity and this issue has an assignee. Please update the label and/or status accordingly.

[Harshini-Gadige]

Automatically closing due to lack of recent activity. Please update the issue when new information becomes available, and we will reopen the issue. Thanks!

[amilsted]

It would be great to have this available - it would be very handy for quantum physics applications! Any news?

[amilsted]

@shamanDevel Any chance of reopening this?