Skip to content

README.md

Latest commit

 

History

History
86 lines (69 loc) · 5.05 KB

README.md

File metadata and controls

86 lines (69 loc) · 5.05 KB
Raw

Surface Pro 7 cameras (Intel IPU4P)

The SP7's Ice Lake ISP (8086:8a19) has no upstream driver. This builds the out-of-tree IPU4P port as modules for the running Ubuntu linux-surface kernel.

Sources

What From Where
Kernel 6.19.8 cdn.kernel.org (sha256 verified) vendor/dl/linux-6.19.8.tar.xz
IPU4P driver + ov5693 fixes thisiscamk/sp7-ipu4-camera overlaid on vendor/linux-6.19.8/drivers/media/
4 new IPU headers ruslanbay/ipu4-drivers patches/kernel/v6.19-ConsultingFuture4200/0001,0012,0015 camera/compat/include/
Firmware ipu4p_cpd.bin ruslanbay/ipu4-drivers firmware/ipu4-20191030.bin (LFS oid verified, Microsoft-signed) vendor/dl/

Our changes

  • compat/ipu-abi-compat.h: additive uapi #defines the port expects.
  • patches/0001-ipu4-build-against-stock-v4l2-headers.patch: the port depends on Intel's edits to core v4l2 structs. As a module for a stock kernel that's an ABI break: .get_routing inserted into v4l2_subdev_pad_ops shifts .set_routing/.enable_streams, so the core calls the wrong function (e.g. VIDIOC_SUBDEV_S_ROUTING lands in get_routing with set_routing's arguments). The patch builds against stock headers instead. Upstream candidate.

Build

camera/build.sh        # -> camera/out/$(uname -r)/*.ko

Needs build-essential and linux-headers-$(uname -r). The kernel was built with gcc 11, we build with gcc 15 (kbuild warns; expected).

Status (2026-10-14)

Camera State
Rear ov8865 (8 MP, autofocus) Works. 3264x1836/1920x1080 at 20 fps in GNOME Camera; autofocus, white balance, colour, denoise.
Front ov5693 (5 MP) Locks reliably with the shifted data-lane timing (patches/0016, csi2_shifted_ticks_port, on by default in install/surface-kit.conf), from ruslanbay/ipu4-drivers RFC v2 11/12, reverse-engineered there from the Windows driver. Measured 2026-11-02 through libcamera, alternating in the same boot: 24/24 clean 30-frame starts with no relock bounces, against 4/24 with the usual delay counters. Before this it looked like a failing cable, because the lock rate fell over days and every setting we swept failed; that was wrong. Colour: patches/ov5693/0005 reports GBRG with mirror=1 (the flip shifts the Bayer order; check with scripts/bayer-sites.py), tuning in libcamera/patches/0019. Open: thin magenta line along the top edge, magenta at exactly 1296x972 (SoftISP 2:1 scaling), mild colour shift with a bright lamp in view.
IR ov7251 Never probes on this driver (skip_fw_ports=6). Works on the 7.4 kernel: see kernel/ and face/.

What makes the rear camera good, and where it lives:

  • Kernel: IPU4P stack with our fixes (patches/0001-0014), ov8865 (patches/ov8865/: mode registers on stream start, mirror fix, vblank reset, fine gain steps, 20 fps 1836 mode, SCLK pixel rate), dw9719 lens (patches/dw9719/).
  • libcamera 0.7.0 rebuilt as .debs (libcamera/patches/0001-0016): SoftISP fixes, AGC (anti-flicker, proportional, highlight damping), sensor delays, contrast autofocus (fine sweep, approach from below, wobble tolerance, parked start), grey-locus AWB and colour matrices from the Windows calibration, lens shading, denoise pass, tuning files.
  • Windows driver data: scripts/win-regtables.py (sensor register tables), scripts/aiqb-to-lsc.py (lens shading tables), scripts/aiqb-dump.py / aiqb-to-ccm.py (Intel AIQB tuning records; this SP7 is "Haven": front MSHW0190, rear MSHW0191).

The same build also carries two non-camera kernel fixes:

  • patches/intel-ish-ipc/: reverts the 6.19 sensor-hub start-up changes (upstream 507561b00ac2, 9e097dc9df80). Without it the ISH never finishes its handshake ("HBM haven't started") and there is no accelerometer, so no auto-rotation (linux-surface#2140). Built here, installed by sensors/install.sh (see sensors/README.md): with it bound during suspend the machine eventually hung hard on every sleep.
  • patches/gpio_keys/0002: ignores power-button presses for resume_ignore_ms (1500, set in install/surface-kit.conf) after resume, so the press that woke it or an impatient second press can't put it straight back to sleep. power/ adds the matching user service that turns the screen off the moment suspend starts.

Install / update:

camera/build.sh && sudo camera/install/install.sh      # modules; reboot after
libcamera/build-deb.sh                                  # prints the install
# commands; stop PipeWire while installing the .debs (an old PipeWire that
# loads the new IPA module hangs), then start it again

Test tools: scripts/focus-raw-sweep.py (raw sharpness per lens position, ground truth for autofocus), scripts/awb-eval.py (white balance methods against a white-paper reference), scripts/settle-sweep.sh (front link sweeps), scripts/denoise-proto.py. Sweeps and raw tests need PipeWire stopped and GNOME Camera closed.