Winter 2025 linops #115
Winter 2025 linops #115
Conversation
…ration of orthonormality loss in the code itself
…now); modified ABRIK testing to expose an inefficiency
rileyjmurray
left a comment
rileyjmurray
left a comment
There was a problem hiding this comment.
You should just make a CompositeOperator class that accepts two abstract linear operators as template parameters. That would be much more flexible.
Yeah, you're right. |
…or. This allowed me to significantly simplkify the composite operator (which no longer is hardcoded to have sparse right side).
|
Riley says "I strongly dislike this densification. I'd ask that we log a warning to std::cerr when this codepath is hit. Later on we can figure out when's okay to be quiet about densifying." I guess I dont need the sparse * sparse case rn, so I'l comment the densification out and add a cerr for now. |
…ding how regular CQRRT should properly function for low-rank cases
…now is said to only accept full-rank inputs
…omposit operator because Eigen's solvers are column-major
…n cholsolver kills us
|
For the past couple of days, I was working on figuring out an issue where CQRRT_linop with nested compisut operator input (cholsolver op, sparse op, dense op) was exhibiting poor orthogonality in the Q factor upon any sparse inputs into the cholsolver_linop (dense cases worked fine). The solution was to disable Eigen's use of fill-reduce pertmutations in sparse cholsolver. Root Cause: Eigen's Solution: Force natural ordering (no permutations) by specifying |
|
At the moment, there is lots of diagnostic code in the repo. |
|
This PR is completed; I will brgin breaking it up into multiple smaller PRs now. |
This PR is a part of of #115, which will be broken up into more digestible parts and closed. This PR intoduces refined abstract linear operator support (with testing) and adds a new "composite" linear operator type. Core changes: 1. Linear Operators (rl_linops.hh): - Extended DenseLinOp and SparseLinOp with full multiplication support (both Side::Left and Side::Right) - Added row-major layout support for dense operators - Added CompositeLinOp class for operator composition (e.g., A = B * C) - Added tests for each linop type 2. Matrix Generation (rl_gen.hh) - Added gen_random_dense() with layout support - Added gen_sparse_coo() and gen_sparse_csc() for sparse matrix generation - New matrix generation functions do not rely on "mat_info" struct so that the matrices can be generated straightforwardly 3. Utilities (rl_util.hh, rl_util_test.hh) - Reorganized utility functions used specifically in tests into separate header (rl_util_test.hh) - Added sparse_to_dense_summing_duplicates() for sparse-to-dense conversion 4. Drivers - Minor updates to rl_abrik and its benchmarks for linop compatability (unsure why the whole driver file appears reworked) - rl_cqrrt is refactored to make sure it only processes full-rank matrices
WIP