offerrall/windowsscreencapture

windowsscreencapture's Issues

Just an idea and some thoughts.

I have been on a constant journey to create the fastest windows capture->(ml, onnx whatever) pipeline. I am a general technologist and sometime lack the true core fundamental principals of static languages so if this is ignorant, I apologize.

When I see a project like this, I assume it's ml based every time or you at least have an intent to do something with said frame. All that said if speed is what you want, for python at least there's this:
https://github.com/NiiightmareXD/windows-capture/tree/main/windows-capture-python

This will be your best shot as it uses the native windows capture in like windows 10 1804+. bitblt is slow and always will be.

The biggest problem is always going to be the frame pull though. To go from gpu->cpu for me is almost impossible to get under 30ms. That why I had to even abandon rust. It does vector conversions and I don't know enough to get under the hood for that performance. Python will be the same I think. Idk, they're always getting lower and aligning the python buffers to direct C so I'm still planning on playing more, but at this time I've decided if I want the ms I need c++ no choice.

So far I have the dx11 frame going right into a cudarray within 5ms, stripping alpha channel with cuda kernels etc. I'm still working out some pain points. But I believe with nano models etc I'll be able to take it to 10-20ms frame to detection. We'll see sometimes with these nuanced objects it looks like you're getting there and dead ending on hard to debug issues. I'm having threading issues, but I hope to figure that by today and understand if I've actually created a viable onnx tensor or I just made some broken binary blobs.

Anywho if you could make a DLL that uses https://learn.microsoft.com/en-us/uwp/api/windows.graphics.capture?view=winrt-22621 I would be overjoyed. But there be dragons. The whole RT space is very opinionated.

OBS uses it though they do this whole c export thing. Again, Im just not good enough to architect entire new build model for these opinionated frameworks.