roystgnr

Updates: I do understand C++11 moves correctly after all. I'm just seeing 6 allocations instead of 3 because I was forgetting that each of the three remaining DNSA copies does...

Testing `DualNumber` was easier done than said. I currently get runs in 7 seconds with the simple kernel, 11.5 seconds with an empty-derivative DualNumber, and ~55 seconds with a full...

empty-DSNA is what you get if you use DynamicSparseNumberArray but you comment out the call to insert() so the derivatives vectors remain empty. No data, no indices, no allocation calls....

The only remaining allocation in empty-DSNA is on the stack as part of the DualNumber, and so is nearly free (there's a chance that the extra bytes will increase cache...

Switching to Howard Hinnant's stack allocator and futzing with the size option a bit gives me no improvement, maybe a couple percent slowdown, no matter how large I make the...

Okay, putting this down for a while. I think I've exhausted all the low-hanging fruit.

"object pool" would be a subset (and probably the most tempting subset) of "custom allocator", wouldn't it? Unless I'm misunderstanding, I do think that's worth trying, but I was pretty...

So the motivation for HybridClass is that we'd avoid the dynamic memory allocation, but we'd still keep the operations sparse? That sounds like an optimization worth trying, yeah. The idea...

`atan2()` is defined for `DualNumber`, `DualExpression`, and all the other MetaPhysicL types. What is the type of `_grad_u_qp`? `git grep _grad_u_qp` isn't finding anything in MOOSE for me ... there's...

> wouldn't make sense with vector (as opposed to array) arguments. which I guess should qualify > all the other MetaPhysicL types. e.g. in `metaphysicl/numberarray.h` we define `atan2` term-by-term, but...