You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
When using a dense VLM mesh, compute_partials in some components (e.g., eval_mtx in aerodynamics) becomes a bottleneck for derivative computations. These partials can be accelerated by replacing the current analytical derivatives with AD.
Aditya Deshpande and Sriram Bommakanti tried it out for the AE588 project, and they showed that AD actually accelerated the partials. Their prototype implementation can be found in their fork.
AD support should be optional because we don't want to add JAX as a hard dependency (for now), and AD likely doesn't offer performance benefits for moderate mesh size.
Potential solution
Run profiling and identify the components that can be accelerated by AD. eval_mtx is one, but there could be others.
Replace (part of) the compute_partials method with AD. We'll need to try out multiple AD options as Aditya and Sriram did.
Add an optional dependency on JAX in setup.py
Add a documentation page on AD - ideally, suggest a mesh size threshold at which the AD becomes faster than the default analytical partials.
The text was updated successfully, but these errors were encountered:
Description of feature
When using a dense VLM mesh,
compute_partials
in some components (e.g., eval_mtx in aerodynamics) becomes a bottleneck for derivative computations. These partials can be accelerated by replacing the current analytical derivatives with AD.Aditya Deshpande and Sriram Bommakanti tried it out for the AE588 project, and they showed that AD actually accelerated the partials. Their prototype implementation can be found in their fork.
AD support should be optional because we don't want to add JAX as a hard dependency (for now), and AD likely doesn't offer performance benefits for moderate mesh size.
Potential solution
eval_mtx
is one, but there could be others.compute_partials
method with AD. We'll need to try out multiple AD options as Aditya and Sriram did.The text was updated successfully, but these errors were encountered: