Offscreen buffer slow because of to_ram = True?

Scripting Issues
1

Hi everyone,

I was trying to use an offscreen buffer to be able to export and externally alter some camera image of my scene, read it in again and display it somewhere on the screen in the next frame.

This is what I have so far:

framebuffer = self.world.win.makeTextureBuffer("", 800, 600, None, True)
camera = self.world.makeCamera(framebuffer, scene=self.world.render)
camera.reparentTo(self.world.render)
tex = framebuffer.getTexture()
image = tex.getRamImageAs("RGB")

I had to set to_ram = True to get it running. So now it is a bit slow (just like the documentation told me beforehand). Though, when set to False, my buffer is empty.

Then I tried to understand GraphicsStateGuardians, since having multiple graphics contexts might probably be a reason for the lack of performance as I read in the documentation.

So I tried to use GraphicOutput’s makeBuffer()-function. Though, I don’t really get it to work as i wish:

framebuffer = base.graphicsEngine.makeBuffer(base.win, "", 1, 800, 600)
camera = self.world.makeCamera(framebuffer, scene=self.world.render)
camera.reparentTo(self.world.render)
tex = Texture()
base.win.addRenderTexture(tex, GraphicsOutput.RTMCopyTexture, GraphicsOutput.RTPColor)
# texture is empy

When I use GraphicsOutput.RTMCopyRam it kind of works but I guess it does not help in terms of performance. Also, now this renders the whole window and not just the desired camera view. Is this the consequence of having only one GSG, which basically means that you only have one view of the scene? I think I do not understand it that well yet.

Does anyone know how to proceed here or if there is any performance improvement possible at all?

Thanks in advance to everyone who tries to give me some smart hints!

2

Yes, to_ram is slow. There’s no way around it. The driver has to flush all pending commands in the pipeline, and wait for the GPU to finish all queued work up to the present frame, then transfer the data back over a bus that is really designed to transfer data the other way.

But if you just want to display it during the next frame, you don’t need to use to_ram. You just won’t have the data available in getRamImage and friends.

Please be mindful when using the lower-level graphicsEngine.makeOutput, if you don’t pass in a gsg and host window, you’ll get a very inefficient situation with multiple GL contexts. The makeTextureBuffer method is a higher-level abstraction around this function.

3

But then there is no way to somehow get the image’s bytes then, right? The only way is with getRamImage and to_ram=True?

With makeBuffer you can’t add both host window and gsg though. Anyways, I feel like for my use case it doesn’t really help to go with those low level stuff if I need to get the image from RAM in the end.

Thanks for your quick reply btw!

4

Yes, basically. The alternative, if you don’t need the bytes every frame, is to use the RTMTriggeredCopyRam mode, and you can call buf.triggerCopy() in order to get the data in the next frame, but it will still be slow.

1 Like
5

Alright, thank you for your help!