Toolkit Optimization

[mtao]

Now that the toolkit is beginning to work we are identifying a number of performance issues. These issues were predicted to some degree, but it is finally time to start optimizing performance a bit.

Attributes

One category where there are certainly optimizations to be performed is the attribute system.

Reduce attribute creation

There are a few ways how we could potentially reduce attribute creation

• [ ] Implement an abstraction that holds multiple accessors to minimizes the number of attributes that are created
• [ ] More rigorously implement the _slow paradigm to let users be more aware of when they might accidentally create accessors. We want to keep these slow pipelines to allow users to easily declare assertions / allow for rapid prototyping but it should be easy to identify which interfaces are creating unecessary accessors.
• [ ] Implement per-thread accessors (this might be in conjunction with the aforementioned multiple-accessor abstraction.

Reduce access costs

• [ ] Implement templated attribute accessors for vector types to reduce dependency on runtime attribute sizes as most applications depend on static sizes
• [ ] Change AttributeCache to use a hashmap instead of map (like unordered_map)
• [ ] Change AttributeCache underlying data storage to something more memory coherent (like a flat vector with enough size reserved)

[mtao]

@Zhouyuan-Chen i believe you have a test that is good for showing off some performance deficiencies - could you post which test it is / where we can find it?

[Zhouyuan-Chen]

@Zhouyuan-Chen i believe you have a test that is good for showing off some performance deficiencies - could you post which test it is / where we can find it?

in this branch

under /tests/components/test_embedded_remeshing_pipeline.cpp

• embedded_remeshing_2D_pipeline (you may need a picture to run this test, I uploaded it in the channel)
• embedded_remeshing_2D_pipeline_for_test (you can use this one instead of the one above)
• embedded_remeshing_3D_pipeline