Repository navigation

Batched SVD support for ROCSOLVER and CUSOLVER - #275

Merged

kshyatt merged 51 commits into

mainfrom

ksh/batched_svd

Oct 2, 2026

Merged

Batched SVD support for ROCSOLVER and CUSOLVER#275
kshyatt merged 51 commits into
mainfrom
ksh/batched_svd

Commits on Oct 1, 2026
- Batched SVD support for ROCSOLVER and CUSOLVER
  Katharine Hyatt
  authored
  f15d10a
- Go back to QRIteration default
  Katharine Hyatt
  authored
  0dfb4ae
- Fix default driver for CUDA
  kshyatt
  authored
  756bd02
- Fix typo
  kshyatt
  authored
  76b6665
- And fix bad check in yacusolver
  kshyatt
  authored
  6fc6a76
- One more API fix
  kshyatt
  authored
  ead34c6
- Fix size of U and V bar
  kshyatt
  authored
  db0d4b8
- More API fixes
  kshyatt
  authored
  e09586c
- Remove redundant dummy one liners
  kshyatt
  authored
  39349fe
- Switch to batched_svd_f and get rid of Batched versions of algorithms
  kshyatt
  authored
  5aa0617
- Fixup AMD default algos
  kshyatt
  authored
  cdcf956
- Move ragged batch handling and cutoffs etc into MAK
  kshyatt
  authored
  ce3565b
- Add support for a batched svd_via_adjoint for algos requiring tall matrices
  Katharine Hyatt
  authored
  b7f9c93
- Add batched svd_full too
  kshyatt
  authored
  12c2d70
- Add special path for AMDGPU, implement gesvdx batching there, and fix some errors
  Katharine Hyatt
  authored
  7d00e25
- Some fixes and actually test Bisection on AMD
  Katharine Hyatt
  authored
  0578c38
- A few more fixes for Bisection
  Katharine Hyatt
  authored
  5b7a058
- Support svd_full for Bisection
  Katharine Hyatt
  authored
  939f9c1
- Remove batched and non-batched bisection for now
  kshyatt
  authored
  6199602
- Restore Bisection test
  kshyatt
  authored
  0527bab
- Avoid some allocs
  kshyatt
  authored
  e67c712
- one to one-liner and more checks for batch_size
  kshyatt
  authored
  9616379
- Apply batched suggestions from code review
  
  authored
  2d775bb
- Fix sizes of dummy arrays
  kshyatt
  authored
  9f41e62
- Don't return foreach
  kshyatt
  authored
  1bf1afc
- Missing batched Bisection piping
  Katharine Hyatt
  authored
  6859f58
- Refactor out gaugefixing for batches
  Katharine Hyatt
  authored
  83bf9c4
- Add a flag for whether the driver supports ragged batches
  kshyatt
  authored
  92a39d1
- Dumb typos
  kshyatt
  authored
  b9c72c0
- Have batched_svd_compact use Diagonal
  kshyatt
  authored
  4b513b4
- Fix dimension check in gesvdx_strided_batched!
  Katharine Hyatt
  authored
  9ec647d
- supports_ragged_batch takes a driver, not an algorithm
  Katharine Hyatt
  authored
  8011ad9
- Make sure to mark bisection as supporting svd full here
  Katharine Hyatt
  authored
  974a1d4
- Move some batching utilities into a common file
  kshyatt
  authored
  9368acc
- Move supports_ragged_batch to Driver definition
  kshyatt
  authored
  2a838f1
- Just put batches below algos
  kshyatt
  authored
  dbb09c9
- Put the basis completion in the implementation file
  Katharine Hyatt
  authored
  370b91e
- Move batch_size length checks into else
  kshyatt
  authored
  305b1bb
- Apply batched suggestions from code review
  
  kshyatt
  and
  Jutho
  authored
  c0996d9
- Strides for safety
  kshyatt
  authored
  d398490
- Use both alg and driver in the batch checks
  kshyatt
  authored
  6d3d77e
- batch_size checks for batched gesvdx
  Katharine Hyatt
  authored
  16972d7
- Force driver lookup for support checks
  kshyatt
  authored
  fea0ae3
- Dumb typo
  kshyatt
  authored
  b287776
- Actually make sure matrices that are too big aren't sent to the batched function
  kshyatt
  authored
  4dafcbb
- Move batch_size check for CUSOLVER too
  kshyatt
  authored
  fef45dd
- Reorder check_inputs
  kshyatt
  authored
  66d8f71
- Add TODO for CUSOLVER
  kshyatt
  authored
  ee46715
- Use supports_pointer_batch and fix oversized test
  kshyatt
  authored
  06e7f13
- Formatter
  kshyatt
  authored
  30d11c1
- Check that inputs are preserved for in-place batch methods
  Katharine Hyatt
  authored
  5832020

Commits

Commits on Oct 1, 2026