Incorrect matrix representation
Bug #807400 reported by
Niclas Jansson
This bug affects 1 person
| Affects | Status | Importance | Assigned to | Milestone | ||
|---|---|---|---|---|---|---|
| Unicorn | Status tracked in Hpc | |||||
| Hpc |
Fix Released
|
High
|
Niclas Jansson | |||
Bug Description
The macro used for accessing matrices in ElasticSmoother is invalid, it is supposed to be row major but it's currently column order.
Revision history for this message
| Niclas Jansson (njansson) wrote | #1 |
| Changed in unicorn: | |
| importance: | Undecided → Critical |
| importance: | Critical → High |
| status: | New → Confirmed |
| Changed in unicorn: | |
| assignee: | nobody → Niclas Jansson (njansson) |
| status: | Confirmed → In Progress |
| Changed in unicorn: | |
| milestone: | none → 0.2.0-hpc |
| Changed in unicorn: | |
| status: | In Progress → Fix Committed |
To post a comment you must log in.
This explains the difference between the stable and hpc branch implementations (why we compute Finv * Finv' instead of Finv' * Finv)
If we keep the column major order, we can directly interface the solver with fortran blas (Cray libsci for example). We have to pay a small penalty for cblas interfaces, but since the matrices are small (3x3) it should be negligible.