Expand description
Batched cuSOLVER ops.
cuSOLVER ships native batched variants for Cholesky
(Spotrf/DpotrfBatched) and Jacobi SVD
(Sgesvdj/DgesvdjBatched). The batched LU live on the cuBLAS side
(cublasSgetrf/DgetrfBatched); we adopt them into SolverActor for
API symmetry — each request takes a contiguous strided block of
batch_size × n × n matrices and stages an array-of-pointers on
demand.
All three requests share the same input shape:
a: GpuRef<T>—batch_size × m × n × stridepacked column-major, with stride equal tom * n(no extra padding).batch_size: i32— number of independent problems.
Per-batch info codes are read back into a host vec; any non-zero
entry surfaces as a LibraryError identifying the failing batch
index.