Batched SVD support for ROCSOLVER and CUSOLVER - #275

Open

kshyatt wants to merge 46 commits into

mainfrom

ksh/batched_svd

Commits on Oct 1, 2026

Batched SVD support for ROCSOLVER and CUSOLVER
Katharine Hyatt
authored
7475023
Go back to QRIteration default
Katharine Hyatt
authored
606c42b
Fix default driver for CUDA
kshyatt
authored
57b5bc9
Fix typo
kshyatt
authored
7ff48b3
And fix bad check in yacusolver
kshyatt
authored
e626f2b
One more API fix
kshyatt
authored
b681779
Fix size of U and V bar
kshyatt
authored
4e59aaf
More API fixes
kshyatt
authored
f8f46f2
Remove redundant dummy one liners
kshyatt
authored
e0f4f8b
Switch to batched_svd_f and get rid of Batched versions of algorithms
kshyatt
authored
a02ce63
Fixup AMD default algos
kshyatt
authored
e8834e7
Move ragged batch handling and cutoffs etc into MAK
kshyatt
authored
721c09e
Add support for a batched svd_via_adjoint for algos requiring tall matrices
Katharine Hyatt
authored
dbedd55
Add batched svd_full too
kshyatt
authored
8e4bcc4
Add special path for AMDGPU, implement gesvdx batching there, and fix some errors
Katharine Hyatt
authored
dc48639
Some fixes and actually test Bisection on AMD
Katharine Hyatt
authored
da3adb8
A few more fixes for Bisection
Katharine Hyatt
authored
d3d4fd4
Support svd_full for Bisection
Katharine Hyatt
authored
91b3dc8
Remove batched and non-batched bisection for now
kshyatt
authored
a2a45f0
Restore Bisection test
kshyatt
authored
97c4b7b
Avoid some allocs
kshyatt
authored
1db311e
one to one-liner and more checks for batch_size
kshyatt
authored
379538b
Apply batched suggestions from code review

authored
4f4974e
Fix sizes of dummy arrays
kshyatt
authored
7a77866
Don't return foreach
kshyatt
authored
fcb5c9c
Missing batched Bisection piping
Katharine Hyatt
authored
ead9a69
Refactor out gaugefixing for batches
Katharine Hyatt
authored
0995349
Add a flag for whether the driver supports ragged batches
kshyatt
authored
bfe9245
Dumb typos
kshyatt
authored
4d02a3c
Have batched_svd_compact use Diagonal
kshyatt
authored
1e9646e
Fix dimension check in gesvdx_strided_batched!
Katharine Hyatt
authored
25529da
supports_ragged_batch takes a driver, not an algorithm
Katharine Hyatt
authored
23f4870
Make sure to mark bisection as supporting svd full here
Katharine Hyatt
authored
04df157
Move some batching utilities into a common file
kshyatt
authored
b9ddc59
Move supports_ragged_batch to Driver definition
kshyatt
authored
d404f67
Just put batches below algos
kshyatt
authored
d9d8c5e
Put the basis completion in the implementation file
Katharine Hyatt
authored
dfada5b
Move batch_size length checks into else
kshyatt
authored
bb619b5
Apply batched suggestions from code review

kshyatt
and
Jutho
authored
3aef615
Strides for safety
kshyatt
authored
0b7a6cc
Use both alg and driver in the batch checks
kshyatt
authored
7cbfec5
batch_size checks for batched gesvdx
Katharine Hyatt
authored
382f50e
Force driver lookup for support checks
kshyatt
committed
818792a
Dumb typo
kshyatt
committed
27df039
Actually make sure matrices that are too big aren't sent to the batched function
kshyatt
committed
e37c010
Move batch_size check for CUSOLVER too
kshyatt
committed
bb0d96d