Question

anyone tried BEPU Physic Engine ?
http://bepuphysics.codeplex.com/

It’s a fully managed physic engine written in C# … I know it mostly used for XNA ( XBOX and WP7 Projects ) Because Unmanaged Codes are not Allowed.

But what I want to know is how a fully Managed Physic Engine is compared to a P/Invoked One ( For example tao.ODE ) in Windows Environment ( in term of performance ) ?

In other words which method makes more overhead, fully managed code or P/Invoke Wrapper around unmanaged Engines like ODE or PhysX in a Real Project?

Answer 1

I cannot comment on the specific physics engines, however I can offer some experience in writing high performance code (both unmanaged and managed).

A few years ago I worked on a piece of simulation software written in Delphi and ported to .NET (before my arrival I might say). It was pure managed code and computed ion-trajectories for mass spectrometer simulations. The code involved numerical integration, differentiation, N-body electrostatic charge calculations so was certainly testing the CPU.

I found through various experiments trying to find the highest performance that some C++ versions of the simulation routines could be beaten by well optimized C# code.

By well optimized I mean reduction of new operators (re-use of objects), caching, object pooling, use of structs where possible, minimizing method calls where possible, moving method calls to static / sealed where possible, minimize number of parameters passed to methods, compile in release, x64, detached from debugger. Once I had done this, to actually beat the CLR using C++ I had to resort to low level techniques such as SSE/SSE2 and inline assembler.

Ok, I’ll admit, C# and managed languages are no match for hand-optimized C++ code in experienced hands, but I’ve seen poorly optimized code on both platforms. It’s easy to blame the CLR when C# code is slow but when developers are using the new operator at will I find it odd they get surprised when the GC is run so frequently. new and delete in C++ also imposes a performance hit so don’t expect C++ compilation to just make things faster.

Back to your specific engine – you will have to of course do some testing and performance analysis yourself. Regarding platform invoke, it does incur a performance hit known as thunking when pointers and structs are marshalled across the managed/unmanaged boundary. Pure managed code will not have this but also it will miss out on optimisations such as low level memory copy, SSE/SSE2 extensions etc… that can be coded in C++.

Finally I will say that for an example of managed->Platform invoke library that is extremely powerful and fast, take a look at SlimDX. Ok you’re going to get a performance hit over native code and DirectX (some sources say ~5%) but for the productivity benefits of developing in C# I’d say its worth it!