GPU acceleration with SPI displays (STM32MP157 with Vivante GPU)

[Lucas Stach]

Am Dienstag, dem 05.08.2025 um 08:46 +0200 schrieb Josef Luštický:
> On Mon, Aug 4, 2025 at 11:59 AM Lucas Stach <l.stach at pengutronix.de> wrote:
> > 
> > Hi Josef,
> Hello Lucas, thank you for answering.
> > 
> > Am Freitag, dem 01.08.2025 um 13:06 +0200 schrieb Josef Luštický:
> > > Hello,
> > > I am trying to get GPU acceleration to work with SPI display using
> > > ili9225_dri.so,
> > > and I use kmscube to test.
> > > The SoC is STM32MP157c with Vivante GC400 GPU and I use Mesa3D 24.0.9
> > > (rootfs generated with Buildroot 2025.02.3) and kernel 5.10.176 (from
> > > STM).
> > > 
> > > I also have the same system with MIPI Display attached.
> > > While kmscube renders at 60 fps with the MIPI Display system, I get
> > > only 11 fps with the SPI display. Apart from that both systems are the
> > > same.
> > 
> > The question is what can you expect from a SPI display? SPI is not a
> > great interface for isochronous traffic with quite high bandwidth
> > demands. Even at really low resolutions you need a lot of bandwidth to
> > update a display. What is the resolution and color depth of you
> > display? How fast is your SPI connection?
> The display is 320x480 with 16bpp, so the screen is 307200 bytes.
> The SPI clock is set to 50MHz in device-tree and I also tried 100MHz
> with no change.
> From the clock tree configuration, it seems the SPI peripheral gets 64MHz,
> but I'll check what's the real frequency of SPI CLK output.

If this is a standard SPI connection with a single bit per clock
transfer rate, 50 MHz SPI clock rate works out to ~20 frames per
second, limited by pure transfer time. This means no other overhead
like SPI message setup in the host controller and userspace waiting for
flip completion or anything like that.

Given that, your measured 11 fps doesn't sound that much off from what
you could expect from such a setup.

Regards,
Lucas

> > 
> > You might run modetest -v to get a reference point on how fast the
> > display can be updated without any GPU rendering going on.
> > 
> > Regards,
> > Lucas
> > 
> > > 
> > > I checked the output of kmscube and it prints the exact same extensions on both,
> > > the only difference is the framerate.
> > > I also checked output of kmscube's /proc/PID/maps and the only
> > > difference is that with MIPI Display it loads /usr/lib/dri/stm_dri.so
> > > while with the SPI Display it loads /usr/lib/dri/ili9225_dri.so,
> > > but I see those files are the same (hardlinks).
> > > Both open /dev/dri/renderD128 (GPU).
> > > 
> > > Why am I getting just 11 fps on the system with SPI Display?
> > > Is there some architectural limitation, either in Mesa, kernel, or HW-related?
> > > I suppose both setups use kernel's KMS as renderer only (kmsro), thus
> > > it should render at the same framerate.
> > > The difference in HW is the peripheral the displays use - LTDC
> > > (stm_dri.so) vs. SPI (eventually ili9225_dri.so).
> > > Why do I see no /usr/lib/etnaviv_dri.so loaded?
> > > 
> > > The MIPI Display is attached to the LTDC (LCD Timing Display
> > > Controller) peripheral of the STM SoC and the display driver is
> > > drivers/gpu/drm/panel/panel-sitronix-st7703.c (DRM panel).
> > > 
> > > The SPI display driver uses the kernel's TinyDRM API and its driver is
> > > drivers/gpu/drm/tiny/ili9225.c.
> > > 
> > > I also tried softpipe (disabled GPU in device-tree) and it renders 7
> > > fps on both systems.
> > > 
> > > I attach the output from kmscube on the SPI Display system (the MIPI
> > > Display is the same except framerate is 60 fps.)
> > > 
> > > Kind regards
> > > Josef Lusticky
> >