Repository navigation
perm_row of a uniform LU batch reports batch member 1 only #92
Description
Found while working on: T19 (PR #85), review nit.
What: getparam(solver, "perm_row") in src/solver.jl (the "perm_row" branch of getparam, structure "G", stage ≥ factorized) returns sp.perm[piv[1:n]], i.e. the reordering composed with the local row interchanges of batch member 1 only. In a uniform batch every member pivots independently and has its own row order, so members 2..nbatch get no perm_row. Every other per-member output ("info", "npivots", "inertia", "pivot_stats", "diag") is returned per member. The docstring table already says "batch member 1", so the behaviour is documented, but it differs from the other per-member outputs.
Why it is out of scope: T19 delivered LU with uniform batches; the per-member output shape is an API decision. It is not in TASKS.md and blocks no later task. Check what cuDSS returns for CUDSS_DATA_PERM_ROW on a uniform batch before deciding.
Suggested fix or plan change: either return an n × nbatch matrix for a batch (as "diag" concatenates members) and make getparam! accept it, or keep member 1 and state in the docstring that this is by design.
Activity
Tests: split test_numeric_lu per element type; cite #92 in the T19 re…
michel2323 commentedon Oct 7, 2026
What cuDSS returns (measured)
cuDSS 0.8.0 (CUDSS.jl 0.8, RTX 4080): a uniform LU batch of 3 members, dense 4×4 pattern, with members that need different row interchanges. Every vector parameter was queried with host and device buffers of
nand ofn·nbatchentries.nbuffern·nbatchbufferperm_row,perm_colCUDSS_STATUS_NOT_SUPPORTEDNOT_SUPPORTEDINVALID_VALUEperm_reorder_rowINVALID_VALUEdiagINVALID_VALUEperm_row/perm_colat all, not even for a single matrix. CUDSS.jl's own tests skipcudss_getfor both (test/test_cudss.jl). No MadNLP code that works with cuDSS today can depend on it.n·nbatchentries. Settingubatch_indexafter the factorization does not change whatdiagreturns."diag"differs from cuDSS: it returns all members one after the other (n·nbatchentries), where cuDSS returns member 1.Decision
Keep member 1 for
perm_row. It follows cuDSS's convention for batched vector outputs and needs no new output shape. Ann × nbatchmatrix would makeperm_rowthe only output whose type depends on the batch size, andgetparam!would then need matrix buffers. The docstring now says this is by design and names"diag"as the deliberate exception. If someone needs every member's row order later, amemberkeyword ongetparamcan add it without changing the default. Docstring PR follows.Docs: perm_row of a uniform LU batch is member 1 by design (#92) (#94)
PLAN.md: refresh to the implemented design after T21 [skip ci]
PLAN.md: refresh to the implemented design after T21 [skip ci] (#99)