On 07.06.22 10:28, Heinrich Schuchardt wrote:
On 6/7/22 01:43, Alexander Graf wrote:
This patch set speeds up graphics output on ARM by a factor of 60x.
On most ARM SBCs, we keep the frame buffer in DRAM and map it as cached,
but need it accessible by the display controller which reads directly
from a later point of consistency. Hence, we flush the frame buffer to
DRAM on every change. The full frame buffer.
Isn't a similar problem already solved by CONFIG_VIDEO_COPY?
Leaving the frame buffer uncached would convert the ARM problem into the
X86 case?
It solves a similar problem, yes. However, it requires us to allocate
the frame buffer size twice, and we would need to dynamically toggle the
MMU mappings of the frame buffer to WC instead of cached. That's code we
don't have today.
VIDEO_COPY is also terribly inefficient in the most common case: Drawing
one or multiple characters. It basically copies every line that contains
the character, for every character printed. The damage code in this
patch set only flushes the relevant rectangles after a string is fully
printed.
I think overall, damage tracking with cached memory is simple enough
that it gives us the best of all worlds.
Unfortunately, with the advent of 4k displays, we are seeing frame
buffers
that can take a while to flush out. This was reported by Da Xue with
grub,
which happily print 1000s of spaces on the screen to draw a menu. Every
printed space triggers a cache flush.
This patch set implements the easiest mitigation against this problem:
Damage tracking. We remember the lowest common denominator region
that was
touched since the last video_sync() call and only flush that.
If by "lowest common denominator region" you should mean a rectangle,
drawing a point in the upper left corner and another in the lower right
corner would require a full flush. So nothing gained in this case.
Glad you asked! :)
While theoretically possible, this is a case that just never happens in
U-Boot's code flow. All code that draws to the screen is either blt
based (like gop, character drawing or logo display) or moves large
portions of the screen (scrolling). The largest granularity we have
between syncs is when printing strings. So the worst case you'll have
today is a wrap around where you'd end up flushing full lines.
With this patch set applied, we reduce drawing a large grub menu (with
serial console attached for size information) on an RK3399-ROC system
at 1440p from 55 seconds to less than 1 second.
Alternatives considered:
1) Lazy sync - Sandbox does this. It only calls video_sync(true) ever
so often. We are missing timers to do this generically.
2) Double buffering - We could try to identify whether anything
changed
at all and only draw to the FB if it did. That would require
maintaining a second buffer that we need to scan.
3) Text buffer - Maintain a buffer of all text printed on the
screen with
respective location. Don't write if the old and new character are
identical. This would limit applicability to text only and is an
optimization on top of this patch set.
4) Hash screen lines - Create a hash (sha256?) over every line
when it
changes. Only flush when it does. I'm not sure if this would waste
more time, memory and cache than the current approach. It would
make
full screen updates much more expensive.
Alexander Graf (6):
dm: video: Add damage tracking API
dm: video: Add damage notification on display clear
vidconsole: Add damage notifications to all vidconsole drivers
video: Add damage notification on bmp display
efi_loader: GOP: Add damage notification on BLT
video: Only dcache flush damaged lines
We need documentation describing the difference between
CONFIG_VIDEO_COPY and CONFIG_VIDEO_DAMAGE.
Hm, maybe we should implement CONFIG_VIDEO_COPY as a flush mechanism
behind CONFIG_VIDEO_DAMAGE? That way we only have a single code path for
producers left and in addition also optimize drawing individual
characters. It would also make the feature useful beyond ARM dcache
flushing.
Alex