[Feature request] Batched SYRK in cuBLAS #375

Activity

1. JanuszL

   added

   cuBLAS

   on Oct 29, 2026
2. vamsi-sripathi commented on Oct 29, 2026

   

   vamsi-sripathi

   on Oct 29, 2026

   More actions

   Hi @FenglongSong,

   Thanks for sharing your input! As you mentioned, computing only upper|lower triangle is exactly the saving a batched SYRK would provide for the Schur complement use case (form A^T A, then potrfBatched). To help us scope this, could you share below details?

   1. Precision: FP32, FP64, or complex, to match the potrfBatched variant you use?
   2. Typical sizes: n (output dimension), k, and batch count.
   3. Uniform or variable sizes: are all problems in a batch the same size, or do sizes vary across the batch?
   4. Target GPUs: which architectures are you running on?
3. FenglongSong commented on Oct 29, 2026

   

   FenglongSong

   on Oct 29, 2026

   Author

   More actions

   Hi @vamsi-sripathi,

   Thanks for the response! My use case is to develop batched quadratic programming (QP) solver on GPU: https://github.com/PREDICT-EPFL/cupiqp.

   When solving a batch of QPs, I need to construct and factorize the Schur complement of the KKT system: $L=\text{chol}(P+A^TA)$

   1. Precision: I only use FP32 and FP64. I personally do not need to deal with complex numbers.
   2. Typical sizes: I mainly consider small problems where the matrix sizes are usually below 100*100, batch size can reach hundreds or thousands, depending on the applications. Although under these sizes I can write custom cuda kernels and achieve good performance, I believe extending the SYRK in cuBLAS to batched is valuable itself.
   3. Uniform or variable sizes: in my case I only deal with uniform batch size.
   4. Target GPUs: I'm currently using an NVIDIA GeForce RTX 5090.

   👍React with 👍1Reacted by Vamsi Sripathi
4. rsdubtso

   assigned

   vamsi-sripathi

   on Nov 10, 2026