tiny-differentiable-simulator VS Numba

I need differentiable rigid body dynamics because I want to do nonlinear MPC. One library that can do this is C++ is Tiny Differentiable Simulator https://github.com/erwincoumans/tiny-differentiable-simulator. As I understand it, this software uses a C++ auto-diff library and code generation to create CUDA kernels to compute fast derivatives in parallel. This seems pretty fast because it's C++. Another option is Brax https://github.com/google/brax. Brax uses JAX which I've never used, but from what I've seen online, JAX is popular for researchers and probably very good.

https://github.com/google-research/tiny-differentiable-simul...

Yes, Chrome about://tracing is great to visualize your custom timing data. Happy used for the last 5 years in Bullet and recent physics engines, including events across tracing multiple threads:

https://github.com/google-research/tiny-differentiable-simul...

Or our C++ and CUDA Tiny Differentiable Simulator: https://github.com/google-research/tiny-differentiable-simulator

Drake (and also tiny-differentiable-simulator that I know of) are using meshcat and it seems neat to me

Around the same time, I discovered Numba and was fascinated by how easily it could bring huge performance improvements to Python code.

Simulations are, at least in my experience, numba’s [0] wheelhouse.

[0]: https://numba.pydata.org/

That's very cool. Numba introduces just-in-time compilation to Python via decorators and its sole reason for being is to turn everything it can into abstract syntax trees.

A super-fast backtesting engine built in NumPy and accelerated with Numba.

Regarding speed, I don't agree this is a good argument against Python. For example, it seems no one here has yet mentioned numba, a Python JIT compiler. With a simple decorator you can compile a function to machine code with speeds on par with C. Numba also allows you to easily write cuda kernels for GPU computation. I've never had to drop down to writing C or C++ to write fast and performant Python code that does computationally demanding tasks thanks to numba.

Just for reference,

* Nuitka[0] "is a Python compiler written in Python. It's fully compatible with Python 2.6, 2.7, 3.4, 3.5, 3.6, 3.7, 3.8, 3.9, 3.10, and 3.11."

* Pypy[1] "is a replacement for CPython" with builtin optimizations such as on the fly JIT compiles.

* Cython[2] "is an optimising static compiler for both the Python programming language and the extended Cython programming language... makes writing C extensions for Python as easy as Python itself."

* Numba[3] "is an open source JIT compiler that translates a subset of Python and NumPy code into fast machine code."

* Pyston[4] "is a performance-optimizing JIT for Python, and is drop-in compatible with ... CPython 3.8.12"

[0] https://github.com/Nuitka/Nuitka

[1] https://www.pypy.org/

[2] https://cython.org/

[3] https://numba.pydata.org/

[4] https://github.com/pyston/pyston

For the benefit of future readers: https://numba.pydata.org/

Taichi (similar to numba) is a python library that allows you to write high speed code within python. So your program consists of slow python that gets interpreted regularly, and fast python (fully type annotated and restricted to a subset of the language) that gets parallellized and jitted for CPU or GPU. And you can mix the two within the same source file.