Hi Mikhail, On Fri, Sep 16, 2022 at 12:27:42AM +0300, Mikhail Rudenko wrote: > > Hi Dave, > > On 2022-09-14 at 10:58 +01, Dave Stevenson <dave.stevenson@xxxxxxxxxxxxxxx> wrote: > > Hi Mikhail > > > > On Sun, 11 Sept 2022 at 21:02, Mikhail Rudenko <mike.rudenko@xxxxxxxxx> wrote: > >> > >> Hello, > >> > >> this series implements support for Omnivision OV4689 image > >> sensor. The Omnivision OV4689 is a high performance, 1/3-inch, 4 > >> megapixel image sensor. Ihis chip supports high frame rate speeds up > >> to 90 fps at 2688x1520 resolution. It is programmable through an I2C > >> interface, and sensor output is sent via 1/2/4 lane MIPI CSI-2 > >> connection. > >> > >> The driver is based on Rockchip BSP kernel [1]. It implements 4-lane CSI-2 > >> and single 2688x1520 @ 30 fps mode. The driver was tested on Rockchip > >> 3399-based FriendlyElec NanoPi M4 board with MCAM400 camera module. > >> > >> While porting the driver, I stumbled upon two issues: > >> > >> (1) In the original driver, horizontal total size (HTS) was set to a > >> value (2584) lower then the frame width (2688), resulting in negative > >> hblank. In this driver, I increased HTS to 2688, but fps dropped from > >> 29.88 to 28.73. What is the preferred way to handle this? > > > > This is one of the joys of sensors - they don't all work in the same way. > > > > I don't have an official datasheet for OV4689 from Omnivision, but > > found one on the internet [1]. That should allow you to reverse the > > PLL configuration to confirm that the pixel rate is the value you've > > computed based on link frequency (they aren't necessarily related). Do > > the frame rate calculations work using width + HBLANK, height + > > VBLANK, and pixel rate? > > The datasheet claims the sensor supports 2688x1520 @ 90 fps, so > > something doesn't hold true between 4 data lanes at 500MHz/1Gbit/s per > > lane when your default hts/vts is 2688x1554 and it only gives > > 28.73fps. > > Seems like those 90 fps is about CSI throughput, not actual sensor > performance. I've checked the datasheet and the register values, and it > seems like the pixel clock is 126 Mhz in this configuration (the maximum > is 150 MHz according to the datasheet). This corresponds to a > theoretical fps of 30.16 at hts=2688 and vts=1554. At the same time the > observed fps is 28.73. I'm not sure where those 1.43 frames are lost, > hope to do more experimentation with VTS and HTS over the weekend. > > > I have seen modes in sensors where the HTS register is in units of 2 > > pixels, so what range of HTS (and VTS) values actually works on this > > sensor? (I don't see it documented, but I'm not surprised). > > > > [1] https://cdn.hackaday.io/files/19354828041536/OV4689-OmniVision.pdf > > > >> (2) The original driver exposes analog gain range 0x0 - 0x7ff, but the > >> gain is not linear across that range. Instead, it is piecewise linear > >> (and discontinuous). 0x0-0xff register values result in 0x-2x gain, > >> 0x100-0x1ff to 0x-4x, 0x300-0x3ff to 0x-8x, and 0x700-0x7ff to 0x-16x, > >> with more linear segments in between. Rockchip's camera engine code > >> chooses one of the above segments depenging on the desired gain > >> value. The question is, how should we proceed keeping in mind > >> libcamera use case? Should the whole 0x0-0x7ff be exposed as-is and > >> libcamera will do the mapping, or the driver will do the mapping > >> itself and expose some logical gain units not tied to the actual gain > >> register value? Meanwhile, this driver conservatively exposes only > >> 0x0-0xf8 gain register range. > > > > The datasheet linked above says "for the gain formula, please contact > > your local OmniVision FAE" :-( > > I would assume that the range is from 1x rather than 0x - people > > rarely want a totally black image that 0x would give. Or is it ranges > > of 1x - 2x, 2x - 4x, 4x - 8x, and 8x - 16x? > > A picture is worth a thousand words, so I've attached the results of my > experimentation with the gain register. They were obtained with Rockchip > 3399, with AEC, AGC and black level subtraction disabled. The image was > converted from 10-bit RGGB to 8-bit YUV 4:2:0 by the Rockchip ISP. Based on that it looks like their medication may have been a little too strong. Could this be implemented so that the control value would be linear linear but its range would correspond 1x--16x values? libcamera will be able to cope with that. > > > Other sensors expose the full range of the register via > > V4L2_CID_ANALOGUE_GAIN, and require userspace (mainly libcamera now) > > to know how to convert a gain into the register value. If the gain > > range goes up to x16, then exposing that would be useful. I'd advocate > > just exposing the full range of 0x000 - 0x7ff, as then you can have > > the accuracy of 256 values between x1 to x2, but also the full range. > > I also like this approach, although libcamera's CameraSensorHelper > doesn't support piecewise-linear gain code mapping yet. Nevertheless, > I believe exposing the full range is a good idea and will do so in v3. > > > I might see if I can pick up one of these sensors and see if I can get > > it running on a Raspberry Pi. Thanks for trying to upstream this - > > it's nice to have such a range of sensor drivers to choose from. > > > > Dave > > > > Thanks for your elucidating tips! > > -- > Best regards, > Mikhail Rudenko > -- Sakari Ailus
