Comment by saltcured

When PSR or adaptive refresh rate systems suspend or re-clock the link, this requires reengineering of the link and its controls. All of this evolved out of earlier display links, which evolved out of earlier display DACs for CRTs, which continuously scanned the system framebuffer to serialize pixel data into output signals. This scanning was synchronized to the current display mode and only changed timings when the display mode was set, often which a disruptive glitch and resynchronization period. Much of this design cruft is still there, including the whole idea of "sync to vblank".

When you have display persistence, you can imagine a very different architecture where you address screen regions and send update packets all the way to the screen. The screen in effect becomes a compositor. But then you may also want transactional boundaries, so do you end up wanting the screen's embedded buffers to also support double or triple buffering and a buffer-swap command? Or do you just want a sufficiently fast and coordinated "blank and refill" command that can send a whole screen update as a fast burst, and require the full buffer to be composited upstream of the display link?

This persistence and selective addressing is actually a special feature of the MIP screens embedded in watches etc. They have a link mode to address and update a small rectangular area of the framebuffer embedded in the screen. It sends a smaller packet of pixel data over the link, rather than sending the whole screen worth of pixels again. This requires different application and graphics driver structure to really support properly and with power efficiency benefits. I.e. you don't want to just set a smaller viewport and have the app continue to render into off-screen areas. You want it to focus on only rendering the smaller updated pixel area.

There are multiple problems here, coming from opposite needs.

A compositor could request new frames when it needs them to composite, in order to reduce its own buffering. But how does it know it is needed? Only in a case like window management where you decided to "reveal" a previously hidden application output area. This is a like older "damage" signals to tell an X application to draw its content again.

But for power-saving, display-persistence scenarios, an application would be the one that knows it needs to update screen content. It isn't because of a compositor event demanding pixels, it is because something in the domain logic of the app decided its display area (or a small portion of it) needs to change.

In the middle, naive apps that were written assuming isochronous input/process/output event loops are never going to be power efficient in this regard. They keep re-drawing into a buffer whether the compositor needs it or not, and they keep re-drawing whether their display area is actually different or not. They are not structured around diffs between screen updates.

It takes a completely different app architecture and mindset to try to exploit the extreme efficiency realms here. Ideally, the app should be completely idle until an async event wakes it, causes it to change its internal state, and it determines that a very small screen output change should be conveyed back out to the display-side compositor. Ironically, it is the oldest display pipelines that worked this way with immediate-mode text or graphics drawing primitives, with some kind of targeted addressing mode to apply mutations to a persistent screen state model.

Think of a graphics desktop that only updates the seconds digits of an embedded clock every second, and the minutes digits every minute. And an open text messaging app only adds newly typed characters to the screen, rather than constantly re-rendering an entire text display canvas. But, if it re-flows the text and has to move existing characters around, it addresses a larger screen region to do so. All those other screen areas are not just showing static imagery, but actually having a lack of application CPU, GPU, framebuffer, and display link activities burning energy to maintain that static state.