Repository navigation
Conversation
The Tensor G4 (zumapro) camera front end has the same Samsung DC-PHY receiver with nine slaves for its twelve CSIS links, the clock lane first on each and one reset bit per slave. Add its compatible; the PHY cell is the CSIS link. Signed-off-by: W4HO <87742609+W4HO-ham@users.noreply.github.com>
zumapro's DC-PHY slaves sit at 0x1c100000 with the bias block at +0x1000 as on gs201, but every slave has its clock lane first and up to four data lanes (three C-PHY trios) after it, and three slaves are a clock lane only, which links 4 and 6 enable next to their own in four-lane D-PHY mode. The layout comes from the Pixel 9a camera HAL's register table and Google's GPL zuma DT; the reset bits are one per slave in the ISP-FE system registers at +0x500. Make the link layout per SoC with the reset bits and the partner clock lane in it, and add zumapro's four-lane links (0, 1, 2, 4 and 6). The data file format is unchanged. On the Pixel 9a, link 1 (the ultrawide IMX712, four lanes at 1105.92 Mbps) powers on with settle 55, and the CSIS link receives about 30 frames a second with no SOT, CRC or ECC errors. Signed-off-by: W4HO <87742609+W4HO-ham@users.noreply.github.com>
zumapro has the same CSIS link (version 0x08010000 against gs201's 0x05040000, the same register layout) with twelve link contexts, but a different DMA path. Add its compatible with the link registers and the link interrupt only. Signed-off-by: W4HO <87742609+W4HO-ham@users.noreply.github.com>
zumapro's CSIS link has gs201's layout, but its frames reach memory through the ISP front end's CSIS core, which this driver doesn't drive yet. Add a zumapro match that maps the link only and always runs as link_only does: PHY, link and sensor, counting frames, with no WDMA, no routing and none of gs201's CSIS sysreg accesses (an unknown sysreg word aborted gs201). Signed-off-by: W4HO <87742609+W4HO-ham@users.noreply.github.com>
On zumapro a link's frames reach memory through the ISP front end's CSIS core, one page shared by all links, with one interrupt for all of them. Describe it as an optional second register range and interrupt of a zumapro link, with the image logical channel the link writes through. Signed-off-by: W4HO <87742609+W4HO-ham@users.noreply.github.com>
zumapro's CSIS core takes a link's data type on one virtual channel into an image logical channel (LoCH) and writes it to memory through its Bayer WDMA. The register layout and fields come from the Pixel 9a camera HAL's register table and value packers and, for the interrupt bits, Google's GPL zuma ISP event description. Route the link's channel 0 into a LoCH, write packed RAW10 (or RAW8) at the stride the video node set, and program the batch slot at each start of frame only: a configuration written during a frame is a late config and stops the LoCH, so frames with no queued buffer go to a scratch buffer. A status queue page takes the LoCH's per-frame record. The link differs from gs201 in three bits the zumapro HAL sets or clears: pixel mode 3, CMN_CTRL bit 12, and DBG_OPTION_SUITE cleared. gs201 sets DBG_OPTION_SUITE bit 23, and with it the link hands out unpacked 16-bit words, which gs201's WDMA expects; zumapro's core reads them as packed RAW10 and scrambles every pixel. On the Pixel 9a the ultrawide's solid colour test pattern (R 1023, Gr 0, B 341, Gb 682) reads back exactly in RAW10 and RAW16, and libcamera's simple pipeline streams it at 30 fps. Signed-off-by: W4HO <87742609+W4HO-ham@users.noreply.github.com>
The Pixel 9a main camera is a Samsung part with model ID 0x08e8 that takes the GN1's register layout and data file, with a second auxiliary supply and its digital supply switched by the board. Signed-off-by: W4HO <87742609+W4HO-ham@users.noreply.github.com>
The Pixel 9a main camera (model 0x08e8) powers up as stock's module power sequence has it: the interface supply, two 2.9 V auxiliary supplies, the analog and the digital supply, reset released, then the clock and 10 ms. Its soft reset goes through the direct page, its standard registers too, and its horizontal addresses count pixels. Its initialisation and modes come from the data file, as for the other parts. On the Pixel 9a it identifies at 0x56 and streams 2000x1128 over three C-PHY trios. Signed-off-by: W4HO <87742609+W4HO-ham@users.noreply.github.com>
All zumapro links write through image LoCH 0, so a stream that stopped with the LoCH in error (the main camera's first C-PHY attempts) left it taking a few frames and then nothing for the next link. Disable, abort and clear the LoCH before routing it. Count the core interrupts each link handles in debugfs, with the last aggregate and processing sources. Bring-up parameters for the link and the core (pixel and parallel mode, CMN_CTRL bits, LoCH data type and memory format) stay for the C-PHY work. Signed-off-by: W4HO <87742609+W4HO-ham@users.noreply.github.com>
zumapro's camera software never writes the link's PHY_CMN_CTRL: its PHY lanes are enabled in the DC-PHY transaction. Write it on gs201 only. The ultrawide and front cameras stream as before without it. Signed-off-by: W4HO <87742609+W4HO-ham@users.noreply.github.com>
…ocks A zumapro link takes the INT DVFS domain, the rail of the ISP front end's NOC clock, and that clock, as "int" and "fe". gs201 keeps its "cam" and "int" DVFS domains. Signed-off-by: W4HO <87742609+W4HO-ham@users.noreply.github.com>
ABL leaves the ISP front end's NOC clock at a sixth of SHARED3 (111 MHz). The Pixel 9a main camera's three C-PHY trios at 2740 Msps (18.8 Gbps) overrun the front end at that rate: the link reports ERR_DESKEW_OVER and the LoCH a size mismatch and stream error, with a few hundred bytes of each frame in memory. At 332 MHz every frame lands whole. Take the INT DVFS domain (the clock's rail, through ACPM) and the front end clock as "int" and "fe", and raise both while a link streams, the rail first, restoring them in reverse afterwards. Signed-off-by: W4HO <87742609+W4HO-ham@users.noreply.github.com>
The Pixel 9a's ultrawide and front cameras are Sony IMX712s (model 0x0712), with the IMX386's register map and their own initialisation and modes in the data file. Signed-off-by: W4HO <87742609+W4HO-ham@users.noreply.github.com>
The Pixel 9a's ultrawide and front cameras answer with model 0x0712 at the IMX386's chip ID register and take the same CCS-style register map. Take the expected model ID from the compatible, for the probe's check and the data file's header. Both cameras stream 4208x3072 RAW10 on four D-PHY lanes on the palmtop with their data file from the 9a's camera HAL. Signed-off-by: W4HO <87742609+W4HO-ham@users.noreply.github.com>
…used define The LoCH interrupt registers' offsets are also in Google's GPL zuma-isp-event-info.dtsi, not only their bits: say so. ZP_STRIDE_ALIGN was never used (CSIS_STRIDE_ALIGN, 64, already covers it). Signed-off-by: W4HO <87742609+W4HO-ham@users.noreply.github.com>
…roles The GN8's power-up order is rear_main_camera_power_seq in Google's GPL zuma-tegu-camera.dtsi, which also names ldo5 and ldo4 as the actuator's and OIS's AVDD and AF_VM (the sensor's own being ldo1 and vdig). Replace the open question with that; both stay required, as the module's sequence powers them before the sensor. Link: https://android.googlesource.com/kernel/devices/google/tegu/+/6276884d9940694b25a6934671918ec8ac412ee8/dts/zuma-tegu-camera.dtsi Signed-off-by: W4HO <87742609+W4HO-ham@users.noreply.github.com>
None of gs201-csis, the MIPI DC-PHY receiver or the S5KGN1 and IMX386 sensor drivers had an entry. Add one block for the four and their bindings. Signed-off-by: W4HO <87742609+W4HO-ham@users.noreply.github.com>
On the Pixel 9a the first capture after switching cameras came back empty whenever the previous stream ran on a D-PHY link (the ultrawide or the front camera) and worked: the new link received its frames, but the LoCH raised no interrupt and wrote nothing. The next capture on the same link worked. The LoCH's EN_STATUS register (0x0a8, named in the camera HAL's register table) shows why. After a stream that stopped cleanly it still reads active with the old link in its low bits (0x106 after link 6, 0x101 after link 1), and every failing start began that way, every good start with it clear. The abort at route time didn't clear it, because the LoCH ignores an abort while it is disabled. So when the LoCH is still active, enable it (its DMA is already off) for the abort, and disable it afterwards. Nine captures in a row, switching between all three cameras in every order, now all write frames; before, about half of the switches failed. Signed-off-by: W4HO <87742609+W4HO-ham@users.noreply.github.com>
…errupt ERR_WRONG_CONFIG flags a packet that no channel is configured to take. The Pixel 9a's sensors send such packets by design on VC 0: one embedded-data line a frame on the front and ultrawide cameras, and about 527 PDAF packets a frame on the main camera. At 30 fps that is some 16,000 link interrupts a second on the main camera, each a read and two writes of the link's interrupt registers, for an error that never affects the image. Leave it out of the default INT0 mask; the other link errors still interrupt, and the int0_mask parameter can still take them all for bring-up. Signed-off-by: W4HO <87742609+W4HO-ham@users.noreply.github.com>
|
Thank you. The numbers explain the change well (about 527 interrupts a frame on the main camera, for packets the image never uses). One thing for when it leaves draft: |
|
Run on the palmtop ( |
|
Thank you for the run. Nine captures across all three cameras in every order, with the main camera down from 2,611 and 3,078 link interrupts to 10 for five frames, is a clear result. The You've taken it out of draft with the default still shared, so I'll treat it as a deliberate change on both phones. The reason holds on the Pixel 7 Pro too: my camera notes record ERR_WRONG_CONFIG as the only link error there, about 180-200 interrupts a frame from the main camera's VC 1 lines and once a frame on the others. Two things before it merges:
The checks pass on today's |
|
A second zumapro data point, from the Pixel 9 Pro Fold (comet,
So the default mask holds on comet's five sensors too. Comet's board DTS will set the LRTE property on the ultrawide's link when it goes up. |
|
@kdegeek, thank you for the comet numbers. The main IMX787 going from about 2,248 of these a frame to a handful per capture, with the other sensors quiet and no WDMA errors, is the same picture as on the 9a, now on five more sensors. I'll note the ultrawide's @W4HO-ham, where this stands. The head hasn't changed since 28 October, and I re-ran our gate on today's Two things are still open from my comment of the 28th:
|
Draft, stacked on #103 (and #63 under it); only the last commit is new. It waits for one camera run on the palmtop: the nine-switch capture test, plus the link interrupt count per main-camera frame, which should drop from about 530 to about 2.
Every camera stream on the 9a shows link errors, and all of them are INT0 bit 3, ERR_WRONG_CONFIG (the name in Google's
isp_fe.h): a packet on VC 0 with a data type no channel takes. The front and ultrawide cameras send one a frame, an embedded-data line. The main camera sends about 527 a frame, its PDAF data; the link's per-VC counters show frames only on VC 0. At 30 fps that's some 16,000 interrupts a second, each a read and two writes of the link's interrupt registers, for an error that never touches the image.The default INT0 mask now leaves ERR_WRONG_CONFIG out, and the other link errors still interrupt.
int0_maskcan still take them all for bring-up.Checks: checkpatch --strict clean. The module with this change (together with #103's fix) is built for the palmtop's kernel.