Improve simulation speed

[CybDing]

How can I improve my parallel simulation speed when running on gpu when rolling out? Are there any tips for that?
I find that program always reminds me of turning on the performance_mode option in the setting which might help, but I can't find where to open it.
And since I'm should render the img from the obs so that videos could be recorded, is that the reason why the simulation speed is not fast as expected. I'm running on 4090 card.
By the way, are there anyway I could run the genesis on multiple gpu and manage the gs.device, I find the gs.device could only choose from the general backend not the exact one, and I don't know how to utilise multiple gpu for boosting the simulation and rendering speed.
Really appreciate your help!

[jgillick]

There was some discussions on ways to improve performance in this thread

Setting "performance_mode" is done as part of gs.init. For example:

gs.init(logging_level="warning", backend=gs.gpu, performance_mode=True)

This will improve performance by a bit, but don't expect a major performance boost from it.

Rendering cameras to video can affect simulation speed. Two things that I've found that help:

• Be sure to set the env_idx parameter when you add the camera so you're only rendering for one environment.
• Calling camera.render() for each step will be wasteful, since you only need a video at 30-60 FPS. Instead, only render on some steps.

Here's a rough idea on limiting the camera renderings to get 30 frames per second:

# Calculate the steps between rendering 
fps = 30
self.steps_per_frame = round(1.0 / fps / dt)

# Create a step counter
self.current_step = 0

def step(self, actions: torch.Tensor):
    self.current_step += 1
    if self.current_step % self.steps_per_frame == 0:
        self.camera.render()

    # ... do other step related things...

Genesis Force, a framework built on top of Genesis, has this optimization built-in. There are also several examples in that repo that might be helpful. At least, you can check the performance of those examples compared to your project.

Outside of that, it's hard to say why your simulations are running slowly without seeing the code.

[CybDing]

Thanks! Probably because I have to use the camera rendered images as the policy input, which requires me to render every active env's camera continually(not every step, but I really find changing the step per render influence quite a lot). I will check this repo to see if the speed is normal in similar configuration.

[duburcqa]

Are you using BatchRenderer for this? Yes using camera images as policy is slowing down simulation significantly. One of the examples available in this repository is training a similar policy using BatchRenderer. You may have a look as a baseline for you.