Add optional symjit generator. #530
Conversation
|
Here is the benchmark up to 24 links: The benchmark script gives the same settings to all generators such that they generate the full mass matrix and full forcing vector and use the same linear system solver (numpy.linalg.solve()). The integration is LSODA (default in PyDy). |
|
You state symjit >= 2.5.0 is required, so I guess, you added the 'high speed version'. |
I just made it the latest release. It may work with older versions. |
|
At the very bottom of this: siravan/symjit#3 |
|
Using the undocumented apply method: |
|
Doesn't look like symjit 2.5 is available for Python 3.9 from conda forge. |
Just so I am very clerar: Did you use numerical or symbolic inversion here? |
I don't think I will answer, so that you read the code here :) |
I read it for 30 min, but still not clear. :-(( |
|
Just as a record this is lambdify and symjit using |
This adds a new ODEFunctionGenerator that uses symjit as sympy->numeric code generator and compiler. symjit does not support iterable of iterables as function arguments and it can only output flat arrays, so a fair amount of input/output conversion is needed that slows down both the code generation and the numerical evaluation. Replacing symbols in the equations of motion does not seem to slow symjit's code generation which is clearly very fast but the numeric manipulation and type conversion does slow down the evaluation. If symjit can manage some things faster internally to manage similar inputs that lambdify accepts, then symjit may have equal evaluation speed to the Cython generation code and generate extremely faster.
Here is the benchmark run with 16 pendulum links, 100 integration steps, and 2 second duration.
This needs a bit more tidying and management of the optional dependency before merging.