MAGMA batched for non-symmetric eigenvalues?

[Beichuan Yan]

Hi there,

 

I am new in this group. Thank you for looking at my question.

 

Typically, I need to resolve millions of 6th-order eigenvalue problems at each time step in my simulations, which often involves millions of steps. So this would amount to a large number of small and non-symmetric 6x6 matrix eigenvalue problems to be resolved efficiently. Only eigenvalues are needed and eigenvectors are not needed in my work.

 

Would this fit into the Magma batched processing on CPU/GPU? As I read from https://icl.cs.utk.edu/projectsfiles/magma/doxygen/group__batched.html, eigenvalues are not covered, but the pages might be outdated.

 

Thank you very much.

 

Beichuan

 

 

 

 

 

[Ahmad Abdelfattah]

Hi Beichuan,

Thank you for reaching out to us. Unfortunately, MAGMA does not currently provide a batch non-symmetric eigensolver. This is something that we are definitely planning to add to the library, but we don’t have a specific timeline for it. 

The link you sent is a little outdated, the latest release is 2.6.2. Here is a summary of what is available for batch computation:

• Batch L3 BLAS — fixed and variable size
• Batch mat-vec (gemv/symv) — fixed and variable size
• Batch LU and Cholesky factorizations — fixed and variable size 
• Batch QR factorization — only fixed size at the moment 

The current developments are 

• Performance improvement for variable-size batch LU (pre-release)
• Performance improvement for fixed-size batch QR (pre-release)
• Batch SVD (fixed size) based on one-sided Jacobi algorithm 

Thanks,

Ahmad 

On Apr 22, 2022, at 5:49 PM, Beichuan Yan <Beichu...@colorado.edu> wrote:

Hi there,
 
I am new in this group. Thank you for looking at my question.
 
Typically, I need to resolve millions of 6th-order eigenvalue problems at each time step in my simulations, which often involves millions of steps. So this would amount to a large number of small and non-symmetric 6x6 matrix eigenvalue problems to be resolved efficiently. Only eigenvalues are needed and eigenvectors are not needed in my work.
 

Would this fit into the Magma batched processing on CPU/GPU? As I read fromhttps://icl.cs.utk.edu/projectsfiles/magma/doxygen/group__batched.html, eigenvalues are not covered, but the pages might be outdated.

 
Thank you very much.
 
Beichuan
 
 
 
 
 

-- 
You received this message because you are subscribed to the Google Groups "MAGMA User" group.
To unsubscribe from this group and stop receiving emails from it, send an email to magma-user+...@icl.utk.edu.
To view this discussion on the web visit https://groups.google.com/a/icl.utk.edu/d/msgid/magma-user/CY4PR03MB3383D8031164EDADAF9C8DEDF0F79%40CY4PR03MB3383.namprd03.prod.outlook.com.

[Beichuan Yan]

Ahmad,

 

Thank you very much for the updated information.

 

Beichuan

[Ahmad Abdelfattah]

Dear Alexander, 

We are actively working on the batch SVD. We don’t have plans in place for the non-symmetric eigensolvers. This could change in the future, though. 

Thanks,

Ahmad Abdelfattah
Research Assistant Professor
Innovative Computing Laboratory
University of Tennessee, USA
ah...@icl.utk.edu

On Jul 13, 2024, at 9:17 AM, Alexander Maeder <alexander.st...@gmail.com> wrote:

Dear Ahmad,

I am new magma user and want to use it to accelerate our quantum transport simulations.

Batched GEMM and GETRF bring a good speedup, but SVD/non symmetric eigenvalue decomposition are becoming a bottleneck

where batching would be ideal. Matrices sizes are between 256 to 1024.

What is the current developement in the direction of batched SVD/EIG?

Best,

Alexander Maeder

[Alexander Maeder]

Dear Ahmad,

I am new magma user and want to use it to accelerate our quantum transport simulations.

Batched GEMM and GETRF bring a good speedup, but SVD/non symmetric eigenvalue decomposition are becoming a bottleneck

where batching would be ideal. Matrices sizes are between 256 to 1024.

What is the current developement in the direction of batched SVD/EIG?

Best,

Alexander Maeder

On Sunday, April 24, 2022 at 7:40:54 PM UTC+2 Ahmad Abdelfattah wrote: