NPU page improvements

[superm1]

I wrote some kernel patches for exposing utilization and power on AMD NPUs.

1. 8 columns of utilization data.
2. power for the IP block.

I have some other questions/comments.

1. I think the tool sends a confusing message right now with utilization staying 0 when really it just doesn't have an API at the time to use. I think the utilization graph on the NPU page at least should be hidden unless API is present. Another idea might be to grey it out. Thoughts?
2. I don't think the kernel driver has a good way of tracking memory usage (as it's applications locking the memory). I think the graph should be hidden for now. Thoughts?
3. How would you want the new utilization data exposed? The current "utilization" graph is for a device, but it's actually 8 individual columns. There's a lot of ways to slice and dice (8 graphs, average, pick top, etc).
4. How would you want the new power data exposed? Historical data is kinda neat for power IMO. Maybe that's a better graph than memory is.

[nokyan]

Hi, thanks for your input and thanks for your work on the kernel to expose NPU usage! :)

1. This is a problem that's also apparent on GPUs: right now I'm using a hybrid approach of gathering GPU/NPU usage from both the sysfs files and from accumulating per-process usages exposed in procfs. If both are available, Resources will pick the higher of the two. Unfortunately, when using the per-process approach, I haven't found a way of differentiating between the GPU/NPU just having 0 usage and the GPU/NPU not being able to be properly monitored since both are possible (afaik) when there's absolutely nothing in procfs regarding that GPU/NPU, hence Resources displaying 0 % if no measurements could be found.
2. Would you say it's still a useful metric? I don't have an NPU so I couldn't play around with one so far, but if it's just apps locking the memory and if that's rather irrelevant, hiding it for now could be the right call. I also don't know if Intel NPUs work differently in that regard.
3. If per-column usage is a useful metric, maybe we could display it like on the CPU page with a toggle between the average load on all columns and 8 (or however many in the future, I've read that Phoenix and Hawk Point have 5 columns) columns displaying individual column usage. Otherwise, for one graph I think picking the average load on all columns is the best fit, that's how CPU usage works already too. Again, I don't know how Intel NPUs work in that regard, keeping two seperate UI designs could be cumbersome in the future.
4. If I add a power usage graph for NPUs, I should probably add a power usage graph for GPUs as well to keep the UI consistent and the GPU page is already quite full with graphs. But since other people also have been asking for graphs on power usage, I might as well just do it and add power graphs to GPUs and NPUs

[superm1]

This is a problem that's also apparent on GPUs: right now I'm using a hybrid approach of gathering GPU/NPU usage from both the sysfs files and from accumulating per-process usages exposed in procfs. If both are available, Resources will pick the higher of the two. Unfortunately, when using the per-process approach, I haven't found a way of differentiating between the GPU/NPU just having 0 usage and the GPU/NPU not being able to be properly monitored since both are possible (afaik) when there's absolutely nothing in procfs regarding that GPU/NPU, hence Resources displaying 0 % if no measurements could be found.
Would you say it's still a useful metric? I don't have an NPU so I couldn't play around with one so far, but if it's just apps locking the memory and if that's rather irrelevant, hiding it for now could be the right call. I also don't know if Intel NPUs work differently in that regard.

For #1 and #2 I came up with this which hopefully works well enough without breaking others.
#641

If per-column usage is a useful metric, maybe we could display it like on the CPU page with a toggle between the average load on all columns and 8 (or however many in the future, I've read that Phoenix and Hawk Point have 5 columns) columns displaying individual column usage. Otherwise, for one graph I think picking the average load on all columns is the best fit, that's how CPU usage works already too. Again, I don't know how Intel NPUs work in that regard, keeping two seperate UI designs could be cumbersome in the future.

I don't really have a sense of how useful it is to most people. How about for now just do an an average of the utilization on all the columns and then decide later to adjust if people want more data.

If I add a power usage graph for NPUs, I should probably add a power usage graph for GPUs as well to keep the UI consistent and the GPU page is already quite full with graphs. But since other people also have been asking for graphs on power usage, I might as well just do it and add power graphs to GPUs and NPUs

OK great.