search

bulletphysics / bullet3

Bullet Physics SDK: real-time collision detection and multi-physics simulation for VR, games, visual effects, robotics, machine learning etc.
http://bulletphysics.org
Other
12.61k stars 2.87k forks source link

resetDebugVisualizerCamera very slow now #3045

Closed 222464 closed 4 years ago

222464 commented 4 years ago

Hello,

I have used PyBullet on and off since its release. I noticed that the latest version(s) of PyBullet made my projects run really slow. I narrowed it down to the resetDebugVisualizerCamera function (I use it to implement a follow camera). With that line removed, it is as fast as it used to be. However, with it I get like 20fps.

I tried narrowing it down further by commenting out a part of resetDebugVisualizerCamera's source, it doesn't seem to be the connection check. So, it must be in one of the b3 calls. That's as far as I got, as I don't know the API well enough.

Any ideas?

erwincoumans commented 4 years ago

Thanks for the report. I have never seen the performance impact of this resetDebugVisualizerCamera's.

0) Are you using the latest PyBullet version 2.9.5?

1) What GPU and operating system do you use?

2) Did you try profiling it? In the GUI window, hold 'p' key for 1 second, then use Google Chrome to open the timings.json file. It usually is stored in the /tmp folder on Linux and Mac OSX, or in the current working dir for Windows.

3) A small reproduction case is important too, it could be the assets etc.

4) Recently, the reset camera goes to a 'sync', so it happens in the rendering thread. Can you make sure to disable wait-for-vsync in your driver? I can have a look to avoid this sync, it will likely revert performance to previous.

5) How many times per simulation step do you call resetDebugVisualizerCamera? What is the simulation frequency?

222464 commented 4 years ago

Did you just change something? Since I upgraded and now it works fine again!

So here is what I did today:

In case you still need it:

  1. Manjaro Linux, Vega 64.
  2. Once per simulation step
erwincoumans commented 4 years ago

Yes, I just fixed it in 2.9.6. You may still want to disable wait on v-sync. For example, using nvidia-settings (if you have a NVIDIA GPU), AMD and Intel and other GPUs have similar setting.

222464 commented 4 years ago

Excellent, thanks!

erwincoumans commented 4 years ago

By the way, did you check out this MPC example? https://www.youtube.com/watch?v=KPFJdC8Tcm0 You can try it out, the instructions are in the video description. Latest PyBullet has mipmapping enabled, looks much smoother.

Do you have any simulation work to share?

222464 commented 4 years ago

Just tried that example, seems like an interesting task. I will play around with it!

I have contributed to some other simulations, but I haven't created any of my own with PyBullet. I fleshed out a simulation for the Stanford Pupper robot to make experimenting with that easier. I didn't use it for sim2real, though - I work on stuff that works IRL directly. It was still handy to debug the RL algorithms, though.

Here is a video if what came of that, in case it interests you: https://www.youtube.com/watch?v=Zl6Rfb3OQoY

Also, I have used regular C++ bullet for a reinforcement learning environment a while ago: https://github.com/222464/PGE. I haven't touched that engine in a while, but perhaps I should revive it. It has modern 3D graphics (screen space reflections / global illumination, deferred shading, parallax occlusion mapping, CSM, surface net voxel terrain, etc.), and bullet does all the physics.

erwincoumans commented 4 years ago

Thanks for sharing, the PGE project looks cool!

For PyBullet, I kept the OpenGL 3 features simpler than PGE (just shadow mapping, single light, a few GLSL shaders), but it takes effort to maintain compatibility with Windows, Linux and Mac OSX and EGL.