Magic Mask 2 + Intel GPU (140V) UI problems

Testing the release version V20. For Magic Mask 2, I still have these observations:
- First time starting the Mask, the pop-up window says "Tracking" but fps and time remaining indications do not show up and video preview window does not update. Unclear if tracking is on-going and the UI info is not updated or if there is an issue with tracking. Waiting some 5 minutes, no change..
- Trying to Stop the tracking, the first click on Stop seems to have no effect, the second or third kills the pop-up window but apparently tracking process (or thread) is still working as when I try to restart tracking, an error appears in the video preview window stating that "tracking is on-going. Please stop tracking"
- I wait perhaps a minute or so and then I can restart tracking, now the pop-up window works and shows "Tracking" + fps performance and time remaining estimates + the preview window steps through the video clip.
- Performance is pretty bad (1 minute clip takes about 1 hour to track in "Better" mode), however the final result is good. As for the performance, for most other functions seems Resolve is able to keep the GPU load at or very close to 100% but for magic mask, load fluctuates between 100 and 0, see attached performance graph from Task manager. Compute function is the bottleneck, all other functions are at about 1% load. Hopefully it is possible to optimize GPU utilization.

compute load.png (44.44 KiB) Viewed 433 times

The tracking was done on an vertical excerpt from a 4k, the vertical segment was reframed in the edit window and the person being tracked is centered in the vertical segment.

- With these (re)tracking times, I still think it would be very useful to get a way to freeze/lock the track so that no upstream actions will trigger a retrack. My preferred solution would be the ability to generate a "Compound clip" from the mask that could then be easily edited as a regular clip. Some way to toggle "Freeze/lock" on the tracked data would be an other option.

Regarding your last point, if the node where you're using Magic Mask gets its input before (or in parallel to) any other nodes you'll have a much better experience. This way it will be much less common to have upstream changes that invalidates the cached mask.

As you've noticed, that iGPU is a bit underpowered for that type of workload. The AI inference performance is about 12 times slower than a medium GPU from the previous gen Nvidia cards.

Regarding your last point, if the node where you're using Magic Mask gets its input before (or in parallel to) any other nodes you'll have a much better experience. This way it will be much less common to have upstream changes that invalidates the cached mask.

Yes, that's how I would try to use it. However, I really like using pre-clip stuff so editing that triggers the re-track. And then at times, something seems to trigger it and I don't even really know what..

However, being able to convert the mask to a compound clip would have a number of additional benefits as then the mask could be edited and not just graded. Sometimes being able to edit the mask would be nice.

As you've noticed, that iGPU is a bit underpowered for that type of workload. The AI inference performance is about 12 times slower than a medium GPU from the previous gen Nvidia cards.

Yes, this was not a complaint about the speed per se, this was a choice in getting a minimal, super-light laptop. It was just an observation that Resolve does not do a very good job in this case of keeping the GPU pipeline filled, something it does quite well for many other things. Possible room for optimization in the code? However, the other observations I would consider software bugs.

Haven't put V20 on my "production" (read: desktop) system yet, but seems my laptop in general does roughly as well as the 6600 GPU card I have on my desktop.