[2/5] libcamera: Add ClockRecovery class to generate wallclock timestamps
diff mbox series

Message ID 20241206142742.7931-3-david.plowman@raspberrypi.com
State New
Headers show
Series
  • Implement wallclock timestamps for frames
Related show

Commit Message

David Plowman Dec. 6, 2024, 2:27 p.m. UTC
The ClockRecovery class takes pairs of timestamps from two different
clocks, and models the second ("output") clock from the first ("input")
clock.

We can use it, in particular, to get a good wallclock estimate for a
frame's SensorTimestamp.

Signed-off-by: David Plowman <david.plowman@raspberrypi.com>
---
 include/libcamera/internal/clock_recovery.h |  72 +++++++
 include/libcamera/internal/meson.build      |   1 +
 src/libcamera/clock_recovery.cpp            | 207 ++++++++++++++++++++
 src/libcamera/meson.build                   |   1 +
 4 files changed, 281 insertions(+)
 create mode 100644 include/libcamera/internal/clock_recovery.h
 create mode 100644 src/libcamera/clock_recovery.cpp

Comments

Kieran Bingham Dec. 6, 2024, 3:50 p.m. UTC | #1
Quoting David Plowman (2024-12-06 14:27:39)
> The ClockRecovery class takes pairs of timestamps from two different
> clocks, and models the second ("output") clock from the first ("input")
> clock.
> 
> We can use it, in particular, to get a good wallclock estimate for a
> frame's SensorTimestamp.
> 
> Signed-off-by: David Plowman <david.plowman@raspberrypi.com>
> ---
>  include/libcamera/internal/clock_recovery.h |  72 +++++++
>  include/libcamera/internal/meson.build      |   1 +
>  src/libcamera/clock_recovery.cpp            | 207 ++++++++++++++++++++
>  src/libcamera/meson.build                   |   1 +
>  4 files changed, 281 insertions(+)
>  create mode 100644 include/libcamera/internal/clock_recovery.h
>  create mode 100644 src/libcamera/clock_recovery.cpp
> 
> diff --git a/include/libcamera/internal/clock_recovery.h b/include/libcamera/internal/clock_recovery.h
> new file mode 100644
> index 00000000..c874574e
> --- /dev/null
> +++ b/include/libcamera/internal/clock_recovery.h
> @@ -0,0 +1,72 @@
> +/* SPDX-License-Identifier: BSD-2-Clause */
> +/*
> + * Copyright (C) 2024, Raspberry Pi Ltd
> + *
> + * Camera recovery algorithm

Recovering cameras ? ;-)

> + */
> +#pragma once
> +
> +#include <stdint.h>
> +
> +namespace libcamera {
> +
> +class ClockRecovery
> +{
> +public:
> +       ClockRecovery();
> +
> +       /* Set configuration parameters. */
> +       void configure(unsigned int numPts = 100, unsigned int maxJitter = 2000, unsigned int minPts = 10,
> +                      unsigned int errorThreshold = 50000);
> +       /* Erase all history and restart the fitting process. */
> +       void reset();
> +
> +       /*
> +        * Add a new input clock / output clock sample, taking the input from the Linux
> +        * CLOCK_BOOTTIME and the output from the CLOCK_REALTIME.
> +        */
> +       void addSample();
> +       /*
> +        * Add a new input clock / output clock sample, specifying the clock times exactly. Use this
> +        * when you want to use clocks other than the ones described above.
> +        */
> +       void addSample(uint64_t input, uint64_t output);
> +       /* Calculate the output clock value for this input. */
> +       uint64_t getOutput(uint64_t input);
> +
> +private:
> +       unsigned int numPts_; /* how many samples contribute to the history */
> +       unsigned int maxJitter_; /* smooth out any jitter larger than this immediately */
> +       unsigned int minPts_; /* number of samples below which we treat clocks as 1:1 */
> +       unsigned int errorThreshold_; /* reset everything when the error exceeds this */
> +
> +       unsigned int count_; /* how many samples seen (up to numPts_) */
> +       uint64_t inputBase_; /* subtract this from all input values, just to make the numbers easier */
> +       uint64_t outputBase_; /* as above, for the output */
> +
> +       uint64_t lastInput_; /* the previous input sample */
> +       uint64_t lastOutput_; /* the previous output sample */
> +
> +       /*
> +        * We do a linear regression of y against x, where:
> +        * x is the value input - inputBase_, and
> +        * y is the value output - outputBase_ - x.
> +        * We additionally subtract x from y so that y "should" be zero, again making the numnbers easier.
> +        */
> +       double xAve_; /* average x value seen so far */
> +       double yAve_; /* average y value seen so far */
> +       double x2Ave_; /* average x^2 value seen so far */
> +       double xyAve_; /* average x*y value seen so far */
> +
> +       /*
> +        * Once we've seen more than minPts_ samples, we recalculate the slope and offset according
> +        * to the linear regression normal equations.
> +        */
> +       double slope_; /* latest slope value */
> +       double offset_; /* latest offset value */
> +
> +       /* We use this cumulative error to monitor spontaneous system clock updates. */
> +       double error_;
> +};
> +
> +} /* namespace libcamera */
> diff --git a/include/libcamera/internal/meson.build b/include/libcamera/internal/meson.build
> index 7d6aa8b7..41500636 100644
> --- a/include/libcamera/internal/meson.build
> +++ b/include/libcamera/internal/meson.build
> @@ -11,6 +11,7 @@ libcamera_internal_headers = files([
>      'camera_manager.h',
>      'camera_sensor.h',
>      'camera_sensor_properties.h',
> +    'clock_recovery.h',
>      'control_serializer.h',
>      'control_validator.h',
>      'converter.h',
> diff --git a/src/libcamera/clock_recovery.cpp b/src/libcamera/clock_recovery.cpp
> new file mode 100644
> index 00000000..966599ee
> --- /dev/null
> +++ b/src/libcamera/clock_recovery.cpp
> @@ -0,0 +1,207 @@
> +/* SPDX-License-Identifier: BSD-2-Clause */
> +/*
> + * Copyright (C) 2024, Raspberry Pi Ltd
> + *
> + * Clock recovery algorithm
> + */
> +
> +#include "libcamera/internal/clock_recovery.h"
> +
> +#include <time.h>
> +
> +#include <libcamera/base/log.h>
> +
> +/**
> + * \file clock_recovery.h
> + * \brief Clock recovery - deriving one clock from another independent clock
> + */
> +
> +namespace libcamera {
> +
> +LOG_DEFINE_CATEGORY(ClockRec)
> +
> +/**
> + * \class ClockRecovery
> + * \brief Recover an output clock from an input clock
> + *
> + * The ClockRecovery class derives an output clock from an input clock,
> + * modelling the output clock as being linearly related to the input clock.
> + * For example, we may use it to derive wall clock timestamps from timestamps
> + * measured by the internal system clock which counts local time since boot.
> + *
> + * When pairs of corresponding input and output timestamps are available,
> + * they should be submitted to the model with addSample(). The model will
> + * update, and output clock values for known input clock values can be
> + * obtained using getOutput().
> + *
> + * As a convenience, if the input clock is indeed the time since boot, and the
> + * output clock represents a real wallclock time, then addSample() can be
> + * called with no arguments, and a pair of timestamps will be captured at
> + * that moment.
> + *
> + * The configure() function accepts some configuration parameters to control
> + * the linear fitting process.
> + */
> +
> +/**
> + * \brief Construct a ClockRecovery
> + */
> +ClockRecovery::ClockRecovery()
> +{
> +       configure();
> +       reset();
> +}
> +
> +/**
> + * \brief Set configuration parameters
> + * \param[in] numPts The approximate duration for which the state of the model
> + * is persistent, measured in samples
> + * \param[in] maxJitter New output samples are clamped to no more than this
> + * amount of jitter, to prevent sudden swings from having a large effect
> + * \param[in] minPts The fitted clock model is not used to generate outputs
> + * until this many samples have been received
> + * \param[in] errorThreshold If the accumulated differences between input and
> + * output clocks reaches this amount over a few frames, the model is reset
> + */
> +void ClockRecovery::configure(unsigned int numPts, unsigned int maxJitter, unsigned int minPts,
> +                             unsigned int errorThreshold)
> +{
> +       LOG(ClockRec, Debug)
> +               << "configure " << numPts << " " << maxJitter << " " << minPts << " " << errorThreshold;
> +
> +       numPts_ = numPts;
> +       maxJitter_ = maxJitter;
> +       minPts_ = minPts;
> +       errorThreshold_ = errorThreshold;
> +}
> +
> +/**
> + * \brief Reset the clock recovery model and start again from scratch
> + */
> +void ClockRecovery::reset()
> +{
> +       LOG(ClockRec, Debug) << "reset";
> +
> +       lastInput_ = 0;
> +       lastOutput_ = 0;
> +       xAve_ = 0;
> +       yAve_ = 0;
> +       x2Ave_ = 0;
> +       xyAve_ = 0;
> +       count_ = 0;
> +       slope_ = 0.0;

Should slope be initialised to 1.0 or anything 'normalised' for any
startup conditions?

> +       offset_ = 0.0;
> +       error_ = 0.0;
> +}
> +
> +/**
> + * \brief Add a sample point to the clock recovery model, for recovering a wall
> + * clock value from the internal system time since boot
> + *
> + * This is a convenience function to make it easy to derive a wall clock value
> + * (using the Linux CLOCK_REALTIME) from the time since the system started
> + * (measured by CLOCK_BOOTTIME).
> + */
> +void ClockRecovery::addSample()
> +{
> +       LOG(ClockRec, Debug) << "addSample";
> +
> +       struct timespec bootTime;
> +       struct timespec wallTime;
> +
> +       /* Get boot and wall clocks in microseconds. */
> +       clock_gettime(CLOCK_BOOTTIME, &bootTime);
> +       clock_gettime(CLOCK_REALTIME, &wallTime);
> +       uint64_t boot = bootTime.tv_sec * 1000000ULL + bootTime.tv_nsec / 1000;
> +       uint64_t wall = wallTime.tv_sec * 1000000ULL + wallTime.tv_nsec / 1000;
> +
> +       addSample(boot, wall);
> +}
> +
> +/**
> + * \brief Add a sample point to the clock recovery model, specifying the exact
> + * input and output clock values
> + *
> + * This function should be used for corresponding clocks other than the Linux
> + * BOOTTIME and REALTIME clocks.

I like that this object has both an 'easy addSample()' and the paired
version.

I think having the addSample() can define that it pairs BOOTTIME and
REALTIME clocks, but this addSample(uint64_t input, uint64_t output)
itself isn't actually tied to those clocks.

If this ClockRecovery object is used elsewhere, the 'only' reason I
could imagine this function being used is if it were to use a clock
specifically different to the defaults - so I'd probably drop the text
here, and keep the specification of BOOTTIME/REALTIME to the addSample()
- or specific to the usage in V4L2VideoDevice.

> +void ClockRecovery::addSample(uint64_t input, uint64_t output)

As an example that I'm not sure if it's crazy or not yet - I could
imagine setting up a ClockRecovery object that maps sequence number and
timestamp so that you could determine the frame sequence number (output)
from timestamp (input) which would give you quantised sequence numbers
determined from the timestamp - so you could spot frame drops if a CSI
receiver always supplies monotonic sequence numbers. (which I think the
i.MX8MP does... ?)




> +{
> +       LOG(ClockRec, Debug) << "addSample " << input << " " << output;
> +
> +       if (count_ == 0) {
> +               inputBase_ = input;
> +               outputBase_ = output;
> +       }
> +
> +       /*
> +        * We keep an eye on cumulative drift over the last several frames. If this exceeds a
> +        * threshold, then probably the system clock has been updated and we're going to have to
> +        * reset everything and start over.
> +        */

Throughout the series, there's lots of text that could easily be wrapped
to 80 chars. It shows up a lot in my email because I have an 92 char
terminal for my email which shows lots of lines wrapping, but it's not a
hard requirement.


> +       if (lastOutput_) {
> +               int64_t inputDiff = getOutput(input) - getOutput(lastInput_);
> +               int64_t outputDiff = output - lastOutput_;
> +               error_ = error_ * 0.95 + (outputDiff - inputDiff);
> +               if (std::abs(error_) > errorThreshold_) {
> +                       reset();
> +                       inputBase_ = input;
> +                       outputBase_ = output;
> +               }
> +       }
> +       lastInput_ = input;
> +       lastOutput_ = output;
> +
> +       /*
> +        * Never let the new output value be more than maxJitter_ away from what we would have expected.
> +        * This is just to reduce the effect of sudden large delays in the measured output.
> +        */
> +       uint64_t expectedOutput = getOutput(input);
> +       output = std::clamp(output, expectedOutput - maxJitter_, expectedOutput + maxJitter_);
> +
> +       /*
> +        * We use x, y, x^2 and x*y sums to calculate the best fit line. Here we update them by
> +        * pretending we have count_ samples at the previous fit, and now one new one. Gradually
> +        * the effect of the older values gets lost. This is a very simple way of updating the
> +        * fit (there are much more complicated ones!), but it works well enough. Using averages
> +        * instead of sums makes the relative effect of old values and the new sample clearer.
> +        */
> +       double x = static_cast<int64_t>(input - inputBase_);
> +       double y = static_cast<int64_t>(output - outputBase_) - x;
> +       unsigned int count1 = count_ + 1;
> +       xAve_ = (count_ * xAve_ + x) / count1;
> +       yAve_ = (count_ * yAve_ + y) / count1;
> +       x2Ave_ = (count_ * x2Ave_ + x * x) / count1;
> +       xyAve_ = (count_ * xyAve_ + x * y) / count1;
> +
> +       /* Don't update slope and offset until we've seen "enough" sample points. */
> +       if (count_ > minPts_) {

What's the output from getOuput before this happens ? Anything
problematic? Or just things that can be ignored for the first few
frames?

> +               /* These are the standard equations for least squares linear regression. */
> +               slope_ = (count1 * count1 * xyAve_ - count1 * xAve_ * count1 * yAve_) /
> +                        (count1 * count1 * x2Ave_ - count1 * xAve_ * count1 * xAve_);
> +               offset_ = yAve_ - slope_ * xAve_;
> +       }
> +
> +       /* Don't increase count_ above numPts_, as this controls the long-term amount of the residual fit. */
> +       if (count1 < numPts_)
> +               count_++;
> +}
> +
> +/**
> + * \brief Calculate the output clock value according to the model from an input
> + * clock value
> + *
> + * \return Output clock value
> + */
> +uint64_t ClockRecovery::getOutput(uint64_t input)
> +{
> +       double x = static_cast<int64_t>(input - inputBase_);
> +       double y = slope_ * x + offset_;
> +       uint64_t output = y + x + outputBase_;
> +
> +       LOG(ClockRec, Debug) << "getOutput " << input << " " << output;
> +
> +       return output;
> +}
> +
> +} /* namespace libcamera */
> diff --git a/src/libcamera/meson.build b/src/libcamera/meson.build
> index 57fde8a8..4eaa1c8e 100644
> --- a/src/libcamera/meson.build
> +++ b/src/libcamera/meson.build
> @@ -21,6 +21,7 @@ libcamera_internal_sources = files([
>      'byte_stream_buffer.cpp',
>      'camera_controls.cpp',
>      'camera_lens.cpp',
> +    'clock_recovery.cpp',
>      'control_serializer.cpp',
>      'control_validator.cpp',
>      'converter.cpp',
> -- 
> 2.39.5
>
David Plowman Dec. 6, 2024, 4:29 p.m. UTC | #2
Hi Kieran

Thanks for the review!

On Fri, 6 Dec 2024 at 15:50, Kieran Bingham
<kieran.bingham@ideasonboard.com> wrote:
>
> Quoting David Plowman (2024-12-06 14:27:39)
> > The ClockRecovery class takes pairs of timestamps from two different
> > clocks, and models the second ("output") clock from the first ("input")
> > clock.
> >
> > We can use it, in particular, to get a good wallclock estimate for a
> > frame's SensorTimestamp.
> >
> > Signed-off-by: David Plowman <david.plowman@raspberrypi.com>
> > ---
> >  include/libcamera/internal/clock_recovery.h |  72 +++++++
> >  include/libcamera/internal/meson.build      |   1 +
> >  src/libcamera/clock_recovery.cpp            | 207 ++++++++++++++++++++
> >  src/libcamera/meson.build                   |   1 +
> >  4 files changed, 281 insertions(+)
> >  create mode 100644 include/libcamera/internal/clock_recovery.h
> >  create mode 100644 src/libcamera/clock_recovery.cpp
> >
> > diff --git a/include/libcamera/internal/clock_recovery.h b/include/libcamera/internal/clock_recovery.h
> > new file mode 100644
> > index 00000000..c874574e
> > --- /dev/null
> > +++ b/include/libcamera/internal/clock_recovery.h
> > @@ -0,0 +1,72 @@
> > +/* SPDX-License-Identifier: BSD-2-Clause */
> > +/*
> > + * Copyright (C) 2024, Raspberry Pi Ltd
> > + *
> > + * Camera recovery algorithm
>
> Recovering cameras ? ;-)

Hehe. I'll fix that!!

>
> > + */
> > +#pragma once
> > +
> > +#include <stdint.h>
> > +
> > +namespace libcamera {
> > +
> > +class ClockRecovery
> > +{
> > +public:
> > +       ClockRecovery();
> > +
> > +       /* Set configuration parameters. */
> > +       void configure(unsigned int numPts = 100, unsigned int maxJitter = 2000, unsigned int minPts = 10,
> > +                      unsigned int errorThreshold = 50000);
> > +       /* Erase all history and restart the fitting process. */
> > +       void reset();
> > +
> > +       /*
> > +        * Add a new input clock / output clock sample, taking the input from the Linux
> > +        * CLOCK_BOOTTIME and the output from the CLOCK_REALTIME.
> > +        */
> > +       void addSample();
> > +       /*
> > +        * Add a new input clock / output clock sample, specifying the clock times exactly. Use this
> > +        * when you want to use clocks other than the ones described above.
> > +        */
> > +       void addSample(uint64_t input, uint64_t output);
> > +       /* Calculate the output clock value for this input. */
> > +       uint64_t getOutput(uint64_t input);
> > +
> > +private:
> > +       unsigned int numPts_; /* how many samples contribute to the history */
> > +       unsigned int maxJitter_; /* smooth out any jitter larger than this immediately */
> > +       unsigned int minPts_; /* number of samples below which we treat clocks as 1:1 */
> > +       unsigned int errorThreshold_; /* reset everything when the error exceeds this */
> > +
> > +       unsigned int count_; /* how many samples seen (up to numPts_) */
> > +       uint64_t inputBase_; /* subtract this from all input values, just to make the numbers easier */
> > +       uint64_t outputBase_; /* as above, for the output */
> > +
> > +       uint64_t lastInput_; /* the previous input sample */
> > +       uint64_t lastOutput_; /* the previous output sample */
> > +
> > +       /*
> > +        * We do a linear regression of y against x, where:
> > +        * x is the value input - inputBase_, and
> > +        * y is the value output - outputBase_ - x.
> > +        * We additionally subtract x from y so that y "should" be zero, again making the numnbers easier.
> > +        */
> > +       double xAve_; /* average x value seen so far */
> > +       double yAve_; /* average y value seen so far */
> > +       double x2Ave_; /* average x^2 value seen so far */
> > +       double xyAve_; /* average x*y value seen so far */
> > +
> > +       /*
> > +        * Once we've seen more than minPts_ samples, we recalculate the slope and offset according
> > +        * to the linear regression normal equations.
> > +        */
> > +       double slope_; /* latest slope value */
> > +       double offset_; /* latest offset value */
> > +
> > +       /* We use this cumulative error to monitor spontaneous system clock updates. */
> > +       double error_;
> > +};
> > +
> > +} /* namespace libcamera */
> > diff --git a/include/libcamera/internal/meson.build b/include/libcamera/internal/meson.build
> > index 7d6aa8b7..41500636 100644
> > --- a/include/libcamera/internal/meson.build
> > +++ b/include/libcamera/internal/meson.build
> > @@ -11,6 +11,7 @@ libcamera_internal_headers = files([
> >      'camera_manager.h',
> >      'camera_sensor.h',
> >      'camera_sensor_properties.h',
> > +    'clock_recovery.h',
> >      'control_serializer.h',
> >      'control_validator.h',
> >      'converter.h',
> > diff --git a/src/libcamera/clock_recovery.cpp b/src/libcamera/clock_recovery.cpp
> > new file mode 100644
> > index 00000000..966599ee
> > --- /dev/null
> > +++ b/src/libcamera/clock_recovery.cpp
> > @@ -0,0 +1,207 @@
> > +/* SPDX-License-Identifier: BSD-2-Clause */
> > +/*
> > + * Copyright (C) 2024, Raspberry Pi Ltd
> > + *
> > + * Clock recovery algorithm
> > + */
> > +
> > +#include "libcamera/internal/clock_recovery.h"
> > +
> > +#include <time.h>
> > +
> > +#include <libcamera/base/log.h>
> > +
> > +/**
> > + * \file clock_recovery.h
> > + * \brief Clock recovery - deriving one clock from another independent clock
> > + */
> > +
> > +namespace libcamera {
> > +
> > +LOG_DEFINE_CATEGORY(ClockRec)
> > +
> > +/**
> > + * \class ClockRecovery
> > + * \brief Recover an output clock from an input clock
> > + *
> > + * The ClockRecovery class derives an output clock from an input clock,
> > + * modelling the output clock as being linearly related to the input clock.
> > + * For example, we may use it to derive wall clock timestamps from timestamps
> > + * measured by the internal system clock which counts local time since boot.
> > + *
> > + * When pairs of corresponding input and output timestamps are available,
> > + * they should be submitted to the model with addSample(). The model will
> > + * update, and output clock values for known input clock values can be
> > + * obtained using getOutput().
> > + *
> > + * As a convenience, if the input clock is indeed the time since boot, and the
> > + * output clock represents a real wallclock time, then addSample() can be
> > + * called with no arguments, and a pair of timestamps will be captured at
> > + * that moment.
> > + *
> > + * The configure() function accepts some configuration parameters to control
> > + * the linear fitting process.
> > + */
> > +
> > +/**
> > + * \brief Construct a ClockRecovery
> > + */
> > +ClockRecovery::ClockRecovery()
> > +{
> > +       configure();
> > +       reset();
> > +}
> > +
> > +/**
> > + * \brief Set configuration parameters
> > + * \param[in] numPts The approximate duration for which the state of the model
> > + * is persistent, measured in samples
> > + * \param[in] maxJitter New output samples are clamped to no more than this
> > + * amount of jitter, to prevent sudden swings from having a large effect
> > + * \param[in] minPts The fitted clock model is not used to generate outputs
> > + * until this many samples have been received
> > + * \param[in] errorThreshold If the accumulated differences between input and
> > + * output clocks reaches this amount over a few frames, the model is reset
> > + */
> > +void ClockRecovery::configure(unsigned int numPts, unsigned int maxJitter, unsigned int minPts,
> > +                             unsigned int errorThreshold)
> > +{
> > +       LOG(ClockRec, Debug)
> > +               << "configure " << numPts << " " << maxJitter << " " << minPts << " " << errorThreshold;
> > +
> > +       numPts_ = numPts;
> > +       maxJitter_ = maxJitter;
> > +       minPts_ = minPts;
> > +       errorThreshold_ = errorThreshold;
> > +}
> > +
> > +/**
> > + * \brief Reset the clock recovery model and start again from scratch
> > + */
> > +void ClockRecovery::reset()
> > +{
> > +       LOG(ClockRec, Debug) << "reset";
> > +
> > +       lastInput_ = 0;
> > +       lastOutput_ = 0;
> > +       xAve_ = 0;
> > +       yAve_ = 0;
> > +       x2Ave_ = 0;
> > +       xyAve_ = 0;
> > +       count_ = 0;
> > +       slope_ = 0.0;
>
> Should slope be initialised to 1.0 or anything 'normalised' for any
> startup conditions?

Actually, 0 is the value that ensures "output = input" (allowing for
the different starting offset).
Maybe worth adding a comment to that end?

>
> > +       offset_ = 0.0;
> > +       error_ = 0.0;
> > +}
> > +
> > +/**
> > + * \brief Add a sample point to the clock recovery model, for recovering a wall
> > + * clock value from the internal system time since boot
> > + *
> > + * This is a convenience function to make it easy to derive a wall clock value
> > + * (using the Linux CLOCK_REALTIME) from the time since the system started
> > + * (measured by CLOCK_BOOTTIME).
> > + */
> > +void ClockRecovery::addSample()
> > +{
> > +       LOG(ClockRec, Debug) << "addSample";
> > +
> > +       struct timespec bootTime;
> > +       struct timespec wallTime;
> > +
> > +       /* Get boot and wall clocks in microseconds. */
> > +       clock_gettime(CLOCK_BOOTTIME, &bootTime);
> > +       clock_gettime(CLOCK_REALTIME, &wallTime);
> > +       uint64_t boot = bootTime.tv_sec * 1000000ULL + bootTime.tv_nsec / 1000;
> > +       uint64_t wall = wallTime.tv_sec * 1000000ULL + wallTime.tv_nsec / 1000;
> > +
> > +       addSample(boot, wall);
> > +}
> > +
> > +/**
> > + * \brief Add a sample point to the clock recovery model, specifying the exact
> > + * input and output clock values
> > + *
> > + * This function should be used for corresponding clocks other than the Linux
> > + * BOOTTIME and REALTIME clocks.
>
> I like that this object has both an 'easy addSample()' and the paired
> version.
>
> I think having the addSample() can define that it pairs BOOTTIME and
> REALTIME clocks, but this addSample(uint64_t input, uint64_t output)
> itself isn't actually tied to those clocks.
>
> If this ClockRecovery object is used elsewhere, the 'only' reason I
> could imagine this function being used is if it were to use a clock
> specifically different to the defaults - so I'd probably drop the text
> here, and keep the specification of BOOTTIME/REALTIME to the addSample()
> - or specific to the usage in V4L2VideoDevice.
>
> > +void ClockRecovery::addSample(uint64_t input, uint64_t output)
>
> As an example that I'm not sure if it's crazy or not yet - I could
> imagine setting up a ClockRecovery object that maps sequence number and
> timestamp so that you could determine the frame sequence number (output)
> from timestamp (input) which would give you quantised sequence numbers
> determined from the timestamp - so you could spot frame drops if a CSI
> receiver always supplies monotonic sequence numbers. (which I think the
> i.MX8MP does... ?)
>

Interesting thought, though I think I'd prefer to keep the basic
ClockRecovery restricted to, well, clocks. If someone wants to spot
frame drops... obviously they're welcome to do that, but maybe that
gets implemented on top? Remember of course that frame times may be
variable.

>
>
>
> > +{
> > +       LOG(ClockRec, Debug) << "addSample " << input << " " << output;
> > +
> > +       if (count_ == 0) {
> > +               inputBase_ = input;
> > +               outputBase_ = output;
> > +       }
> > +
> > +       /*
> > +        * We keep an eye on cumulative drift over the last several frames. If this exceeds a
> > +        * threshold, then probably the system clock has been updated and we're going to have to
> > +        * reset everything and start over.
> > +        */
>
> Throughout the series, there's lots of text that could easily be wrapped
> to 80 chars. It shows up a lot in my email because I have an 92 char
> terminal for my email which shows lots of lines wrapping, but it's not a
> hard requirement.

I guess I'm not clear how strict we are about line lengths... mostly I
don't worry if the style checker doesn't complain. But I'm happy to do
a bit of reflowing if that helps, though some slightly more "official"
guidance on when to do that would be helpful.

>
>
> > +       if (lastOutput_) {
> > +               int64_t inputDiff = getOutput(input) - getOutput(lastInput_);
> > +               int64_t outputDiff = output - lastOutput_;
> > +               error_ = error_ * 0.95 + (outputDiff - inputDiff);
> > +               if (std::abs(error_) > errorThreshold_) {
> > +                       reset();
> > +                       inputBase_ = input;
> > +                       outputBase_ = output;
> > +               }
> > +       }
> > +       lastInput_ = input;
> > +       lastOutput_ = output;
> > +
> > +       /*
> > +        * Never let the new output value be more than maxJitter_ away from what we would have expected.
> > +        * This is just to reduce the effect of sudden large delays in the measured output.
> > +        */
> > +       uint64_t expectedOutput = getOutput(input);
> > +       output = std::clamp(output, expectedOutput - maxJitter_, expectedOutput + maxJitter_);
> > +
> > +       /*
> > +        * We use x, y, x^2 and x*y sums to calculate the best fit line. Here we update them by
> > +        * pretending we have count_ samples at the previous fit, and now one new one. Gradually
> > +        * the effect of the older values gets lost. This is a very simple way of updating the
> > +        * fit (there are much more complicated ones!), but it works well enough. Using averages
> > +        * instead of sums makes the relative effect of old values and the new sample clearer.
> > +        */
> > +       double x = static_cast<int64_t>(input - inputBase_);
> > +       double y = static_cast<int64_t>(output - outputBase_) - x;
> > +       unsigned int count1 = count_ + 1;
> > +       xAve_ = (count_ * xAve_ + x) / count1;
> > +       yAve_ = (count_ * yAve_ + y) / count1;
> > +       x2Ave_ = (count_ * x2Ave_ + x * x) / count1;
> > +       xyAve_ = (count_ * xyAve_ + x * y) / count1;
> > +
> > +       /* Don't update slope and offset until we've seen "enough" sample points. */
> > +       if (count_ > minPts_) {
>
> What's the output from getOuput before this happens ? Anything
> problematic? Or just things that can be ignored for the first few
> frames?

No, you just get "output = input" allowing for the different starting
offsets. Again, maybe worth a comment?

Thanks!
David

>
> > +               /* These are the standard equations for least squares linear regression. */
> > +               slope_ = (count1 * count1 * xyAve_ - count1 * xAve_ * count1 * yAve_) /
> > +                        (count1 * count1 * x2Ave_ - count1 * xAve_ * count1 * xAve_);
> > +               offset_ = yAve_ - slope_ * xAve_;
> > +       }
> > +
> > +       /* Don't increase count_ above numPts_, as this controls the long-term amount of the residual fit. */
> > +       if (count1 < numPts_)
> > +               count_++;
> > +}
> > +
> > +/**
> > + * \brief Calculate the output clock value according to the model from an input
> > + * clock value
> > + *
> > + * \return Output clock value
> > + */
> > +uint64_t ClockRecovery::getOutput(uint64_t input)
> > +{
> > +       double x = static_cast<int64_t>(input - inputBase_);
> > +       double y = slope_ * x + offset_;
> > +       uint64_t output = y + x + outputBase_;
> > +
> > +       LOG(ClockRec, Debug) << "getOutput " << input << " " << output;
> > +
> > +       return output;
> > +}
> > +
> > +} /* namespace libcamera */
> > diff --git a/src/libcamera/meson.build b/src/libcamera/meson.build
> > index 57fde8a8..4eaa1c8e 100644
> > --- a/src/libcamera/meson.build
> > +++ b/src/libcamera/meson.build
> > @@ -21,6 +21,7 @@ libcamera_internal_sources = files([
> >      'byte_stream_buffer.cpp',
> >      'camera_controls.cpp',
> >      'camera_lens.cpp',
> > +    'clock_recovery.cpp',
> >      'control_serializer.cpp',
> >      'control_validator.cpp',
> >      'converter.cpp',
> > --
> > 2.39.5
> >
Kieran Bingham Dec. 6, 2024, 4:43 p.m. UTC | #3
Quoting David Plowman (2024-12-06 16:29:02)
> Hi Kieran
> 
> Thanks for the review!
> 
> On Fri, 6 Dec 2024 at 15:50, Kieran Bingham
> <kieran.bingham@ideasonboard.com> wrote:
> >
> > Quoting David Plowman (2024-12-06 14:27:39)
> > > The ClockRecovery class takes pairs of timestamps from two different
> > > clocks, and models the second ("output") clock from the first ("input")
> > > clock.
> > >
> > > We can use it, in particular, to get a good wallclock estimate for a
> > > frame's SensorTimestamp.
> > >
> > > Signed-off-by: David Plowman <david.plowman@raspberrypi.com>
> > > ---
> > >  include/libcamera/internal/clock_recovery.h |  72 +++++++
> > >  include/libcamera/internal/meson.build      |   1 +
> > >  src/libcamera/clock_recovery.cpp            | 207 ++++++++++++++++++++
> > >  src/libcamera/meson.build                   |   1 +
> > >  4 files changed, 281 insertions(+)
> > >  create mode 100644 include/libcamera/internal/clock_recovery.h
> > >  create mode 100644 src/libcamera/clock_recovery.cpp
> > >
> > > diff --git a/include/libcamera/internal/clock_recovery.h b/include/libcamera/internal/clock_recovery.h
> > > new file mode 100644
> > > index 00000000..c874574e
> > > --- /dev/null
> > > +++ b/include/libcamera/internal/clock_recovery.h
> > > @@ -0,0 +1,72 @@
> > > +/* SPDX-License-Identifier: BSD-2-Clause */
> > > +/*
> > > + * Copyright (C) 2024, Raspberry Pi Ltd
> > > + *
> > > + * Camera recovery algorithm
> >
> > Recovering cameras ? ;-)
> 
> Hehe. I'll fix that!!
> 
> >
> > > + */
> > > +#pragma once
> > > +
> > > +#include <stdint.h>
> > > +
> > > +namespace libcamera {
> > > +
> > > +class ClockRecovery
> > > +{
> > > +public:
> > > +       ClockRecovery();
> > > +
> > > +       /* Set configuration parameters. */
> > > +       void configure(unsigned int numPts = 100, unsigned int maxJitter = 2000, unsigned int minPts = 10,
> > > +                      unsigned int errorThreshold = 50000);
> > > +       /* Erase all history and restart the fitting process. */
> > > +       void reset();
> > > +
> > > +       /*
> > > +        * Add a new input clock / output clock sample, taking the input from the Linux
> > > +        * CLOCK_BOOTTIME and the output from the CLOCK_REALTIME.
> > > +        */
> > > +       void addSample();
> > > +       /*
> > > +        * Add a new input clock / output clock sample, specifying the clock times exactly. Use this
> > > +        * when you want to use clocks other than the ones described above.
> > > +        */
> > > +       void addSample(uint64_t input, uint64_t output);
> > > +       /* Calculate the output clock value for this input. */
> > > +       uint64_t getOutput(uint64_t input);
> > > +
> > > +private:
> > > +       unsigned int numPts_; /* how many samples contribute to the history */
> > > +       unsigned int maxJitter_; /* smooth out any jitter larger than this immediately */
> > > +       unsigned int minPts_; /* number of samples below which we treat clocks as 1:1 */
> > > +       unsigned int errorThreshold_; /* reset everything when the error exceeds this */
> > > +
> > > +       unsigned int count_; /* how many samples seen (up to numPts_) */
> > > +       uint64_t inputBase_; /* subtract this from all input values, just to make the numbers easier */
> > > +       uint64_t outputBase_; /* as above, for the output */
> > > +
> > > +       uint64_t lastInput_; /* the previous input sample */
> > > +       uint64_t lastOutput_; /* the previous output sample */
> > > +
> > > +       /*
> > > +        * We do a linear regression of y against x, where:
> > > +        * x is the value input - inputBase_, and
> > > +        * y is the value output - outputBase_ - x.
> > > +        * We additionally subtract x from y so that y "should" be zero, again making the numnbers easier.
> > > +        */
> > > +       double xAve_; /* average x value seen so far */
> > > +       double yAve_; /* average y value seen so far */
> > > +       double x2Ave_; /* average x^2 value seen so far */
> > > +       double xyAve_; /* average x*y value seen so far */
> > > +
> > > +       /*
> > > +        * Once we've seen more than minPts_ samples, we recalculate the slope and offset according
> > > +        * to the linear regression normal equations.
> > > +        */
> > > +       double slope_; /* latest slope value */
> > > +       double offset_; /* latest offset value */
> > > +
> > > +       /* We use this cumulative error to monitor spontaneous system clock updates. */
> > > +       double error_;
> > > +};
> > > +
> > > +} /* namespace libcamera */
> > > diff --git a/include/libcamera/internal/meson.build b/include/libcamera/internal/meson.build
> > > index 7d6aa8b7..41500636 100644
> > > --- a/include/libcamera/internal/meson.build
> > > +++ b/include/libcamera/internal/meson.build
> > > @@ -11,6 +11,7 @@ libcamera_internal_headers = files([
> > >      'camera_manager.h',
> > >      'camera_sensor.h',
> > >      'camera_sensor_properties.h',
> > > +    'clock_recovery.h',
> > >      'control_serializer.h',
> > >      'control_validator.h',
> > >      'converter.h',
> > > diff --git a/src/libcamera/clock_recovery.cpp b/src/libcamera/clock_recovery.cpp
> > > new file mode 100644
> > > index 00000000..966599ee
> > > --- /dev/null
> > > +++ b/src/libcamera/clock_recovery.cpp
> > > @@ -0,0 +1,207 @@
> > > +/* SPDX-License-Identifier: BSD-2-Clause */
> > > +/*
> > > + * Copyright (C) 2024, Raspberry Pi Ltd
> > > + *
> > > + * Clock recovery algorithm
> > > + */
> > > +
> > > +#include "libcamera/internal/clock_recovery.h"
> > > +
> > > +#include <time.h>
> > > +
> > > +#include <libcamera/base/log.h>
> > > +
> > > +/**
> > > + * \file clock_recovery.h
> > > + * \brief Clock recovery - deriving one clock from another independent clock
> > > + */
> > > +
> > > +namespace libcamera {
> > > +
> > > +LOG_DEFINE_CATEGORY(ClockRec)
> > > +
> > > +/**
> > > + * \class ClockRecovery
> > > + * \brief Recover an output clock from an input clock
> > > + *
> > > + * The ClockRecovery class derives an output clock from an input clock,
> > > + * modelling the output clock as being linearly related to the input clock.
> > > + * For example, we may use it to derive wall clock timestamps from timestamps
> > > + * measured by the internal system clock which counts local time since boot.
> > > + *
> > > + * When pairs of corresponding input and output timestamps are available,
> > > + * they should be submitted to the model with addSample(). The model will
> > > + * update, and output clock values for known input clock values can be
> > > + * obtained using getOutput().
> > > + *
> > > + * As a convenience, if the input clock is indeed the time since boot, and the
> > > + * output clock represents a real wallclock time, then addSample() can be
> > > + * called with no arguments, and a pair of timestamps will be captured at
> > > + * that moment.
> > > + *
> > > + * The configure() function accepts some configuration parameters to control
> > > + * the linear fitting process.
> > > + */
> > > +
> > > +/**
> > > + * \brief Construct a ClockRecovery
> > > + */
> > > +ClockRecovery::ClockRecovery()
> > > +{
> > > +       configure();
> > > +       reset();
> > > +}
> > > +
> > > +/**
> > > + * \brief Set configuration parameters
> > > + * \param[in] numPts The approximate duration for which the state of the model
> > > + * is persistent, measured in samples
> > > + * \param[in] maxJitter New output samples are clamped to no more than this
> > > + * amount of jitter, to prevent sudden swings from having a large effect
> > > + * \param[in] minPts The fitted clock model is not used to generate outputs
> > > + * until this many samples have been received
> > > + * \param[in] errorThreshold If the accumulated differences between input and
> > > + * output clocks reaches this amount over a few frames, the model is reset
> > > + */
> > > +void ClockRecovery::configure(unsigned int numPts, unsigned int maxJitter, unsigned int minPts,
> > > +                             unsigned int errorThreshold)
> > > +{
> > > +       LOG(ClockRec, Debug)
> > > +               << "configure " << numPts << " " << maxJitter << " " << minPts << " " << errorThreshold;
> > > +
> > > +       numPts_ = numPts;
> > > +       maxJitter_ = maxJitter;
> > > +       minPts_ = minPts;
> > > +       errorThreshold_ = errorThreshold;
> > > +}
> > > +
> > > +/**
> > > + * \brief Reset the clock recovery model and start again from scratch
> > > + */
> > > +void ClockRecovery::reset()
> > > +{
> > > +       LOG(ClockRec, Debug) << "reset";
> > > +
> > > +       lastInput_ = 0;
> > > +       lastOutput_ = 0;
> > > +       xAve_ = 0;
> > > +       yAve_ = 0;
> > > +       x2Ave_ = 0;
> > > +       xyAve_ = 0;
> > > +       count_ = 0;
> > > +       slope_ = 0.0;
> >
> > Should slope be initialised to 1.0 or anything 'normalised' for any
> > startup conditions?
> 
> Actually, 0 is the value that ensures "output = input" (allowing for
> the different starting offset).
> Maybe worth adding a comment to that end?

aha - no - that's probably fine - it's just me misunderstanding. For
some reason I thought '1' would be the more expected slope for
undefined/uninitialised, but I  see '0' just removes the slope which is
(until samples are ready...) undefined ;-)



> 
> >
> > > +       offset_ = 0.0;
> > > +       error_ = 0.0;
> > > +}
> > > +
> > > +/**
> > > + * \brief Add a sample point to the clock recovery model, for recovering a wall
> > > + * clock value from the internal system time since boot
> > > + *
> > > + * This is a convenience function to make it easy to derive a wall clock value
> > > + * (using the Linux CLOCK_REALTIME) from the time since the system started
> > > + * (measured by CLOCK_BOOTTIME).
> > > + */
> > > +void ClockRecovery::addSample()
> > > +{
> > > +       LOG(ClockRec, Debug) << "addSample";
> > > +
> > > +       struct timespec bootTime;
> > > +       struct timespec wallTime;
> > > +
> > > +       /* Get boot and wall clocks in microseconds. */
> > > +       clock_gettime(CLOCK_BOOTTIME, &bootTime);
> > > +       clock_gettime(CLOCK_REALTIME, &wallTime);
> > > +       uint64_t boot = bootTime.tv_sec * 1000000ULL + bootTime.tv_nsec / 1000;
> > > +       uint64_t wall = wallTime.tv_sec * 1000000ULL + wallTime.tv_nsec / 1000;
> > > +
> > > +       addSample(boot, wall);
> > > +}
> > > +
> > > +/**
> > > + * \brief Add a sample point to the clock recovery model, specifying the exact
> > > + * input and output clock values
> > > + *
> > > + * This function should be used for corresponding clocks other than the Linux
> > > + * BOOTTIME and REALTIME clocks.
> >
> > I like that this object has both an 'easy addSample()' and the paired
> > version.
> >
> > I think having the addSample() can define that it pairs BOOTTIME and
> > REALTIME clocks, but this addSample(uint64_t input, uint64_t output)
> > itself isn't actually tied to those clocks.
> >
> > If this ClockRecovery object is used elsewhere, the 'only' reason I
> > could imagine this function being used is if it were to use a clock
> > specifically different to the defaults - so I'd probably drop the text
> > here, and keep the specification of BOOTTIME/REALTIME to the addSample()
> > - or specific to the usage in V4L2VideoDevice.
> >
> > > +void ClockRecovery::addSample(uint64_t input, uint64_t output)
> >
> > As an example that I'm not sure if it's crazy or not yet - I could
> > imagine setting up a ClockRecovery object that maps sequence number and
> > timestamp so that you could determine the frame sequence number (output)
> > from timestamp (input) which would give you quantised sequence numbers
> > determined from the timestamp - so you could spot frame drops if a CSI
> > receiver always supplies monotonic sequence numbers. (which I think the
> > i.MX8MP does... ?)
> >
> 
> Interesting thought, though I think I'd prefer to keep the basic
> ClockRecovery restricted to, well, clocks. If someone wants to spot
> frame drops... obviously they're welcome to do that, but maybe that
> gets implemented on top? Remember of course that frame times may be
> variable.

Oh absolutly - I'm not expecting anything here - I'm just speculating on
what crazy things we could match between input/output to get the linear
regression on ;-)

And that's why I think "ClockRecovery::addSample(uint64_t input, uint64_t
output)" as a function with two parameters isn't itself tied to a
specific clock. (but the usage in V4L2VideoDevice, and addSample() is).


> 
> >
> >
> >
> > > +{
> > > +       LOG(ClockRec, Debug) << "addSample " << input << " " << output;
> > > +
> > > +       if (count_ == 0) {
> > > +               inputBase_ = input;
> > > +               outputBase_ = output;
> > > +       }
> > > +
> > > +       /*
> > > +        * We keep an eye on cumulative drift over the last several frames. If this exceeds a
> > > +        * threshold, then probably the system clock has been updated and we're going to have to
> > > +        * reset everything and start over.
> > > +        */
> >
> > Throughout the series, there's lots of text that could easily be wrapped
> > to 80 chars. It shows up a lot in my email because I have an 92 char
> > terminal for my email which shows lots of lines wrapping, but it's not a
> > hard requirement.
> 
> I guess I'm not clear how strict we are about line lengths... mostly I
> don't worry if the style checker doesn't complain. But I'm happy to do
> a bit of reflowing if that helps, though some slightly more "official"
> guidance on when to do that would be helpful.

I think we normally aim for blocks of 80 - but the exceptions are up to
120 if it helps/makes sense in that instance.

It only shows up in this series, as I think everything is targetting
~100/120ish instead.



> 
> >
> >
> > > +       if (lastOutput_) {
> > > +               int64_t inputDiff = getOutput(input) - getOutput(lastInput_);
> > > +               int64_t outputDiff = output - lastOutput_;
> > > +               error_ = error_ * 0.95 + (outputDiff - inputDiff);
> > > +               if (std::abs(error_) > errorThreshold_) {
> > > +                       reset();
> > > +                       inputBase_ = input;
> > > +                       outputBase_ = output;
> > > +               }
> > > +       }
> > > +       lastInput_ = input;
> > > +       lastOutput_ = output;
> > > +
> > > +       /*
> > > +        * Never let the new output value be more than maxJitter_ away from what we would have expected.
> > > +        * This is just to reduce the effect of sudden large delays in the measured output.
> > > +        */
> > > +       uint64_t expectedOutput = getOutput(input);
> > > +       output = std::clamp(output, expectedOutput - maxJitter_, expectedOutput + maxJitter_);
> > > +
> > > +       /*
> > > +        * We use x, y, x^2 and x*y sums to calculate the best fit line. Here we update them by
> > > +        * pretending we have count_ samples at the previous fit, and now one new one. Gradually
> > > +        * the effect of the older values gets lost. This is a very simple way of updating the
> > > +        * fit (there are much more complicated ones!), but it works well enough. Using averages
> > > +        * instead of sums makes the relative effect of old values and the new sample clearer.
> > > +        */
> > > +       double x = static_cast<int64_t>(input - inputBase_);
> > > +       double y = static_cast<int64_t>(output - outputBase_) - x;
> > > +       unsigned int count1 = count_ + 1;
> > > +       xAve_ = (count_ * xAve_ + x) / count1;
> > > +       yAve_ = (count_ * yAve_ + y) / count1;
> > > +       x2Ave_ = (count_ * x2Ave_ + x * x) / count1;
> > > +       xyAve_ = (count_ * xyAve_ + x * y) / count1;
> > > +
> > > +       /* Don't update slope and offset until we've seen "enough" sample points. */
> > > +       if (count_ > minPts_) {
> >
> > What's the output from getOuput before this happens ? Anything
> > problematic? Or just things that can be ignored for the first few
> > frames?
> 
> No, you just get "output = input" allowing for the different starting
> offsets. Again, maybe worth a comment?

Hrm, ok so maybe it might be useful to state 'somewhere' that the system
will produce output values matching the input value until sufficient
sampling points have been collected.

And following that back, I see minPts_ is configurable (which is a good
thing :D), so it's up to the use case to determine what the minimum
requirements are.

Seems pretty good to me.

> 
> Thanks!
> David
> 
> >
> > > +               /* These are the standard equations for least squares linear regression. */
> > > +               slope_ = (count1 * count1 * xyAve_ - count1 * xAve_ * count1 * yAve_) /
> > > +                        (count1 * count1 * x2Ave_ - count1 * xAve_ * count1 * xAve_);
> > > +               offset_ = yAve_ - slope_ * xAve_;
> > > +       }
> > > +
> > > +       /* Don't increase count_ above numPts_, as this controls the long-term amount of the residual fit. */
> > > +       if (count1 < numPts_)
> > > +               count_++;
> > > +}
> > > +
> > > +/**
> > > + * \brief Calculate the output clock value according to the model from an input
> > > + * clock value
> > > + *
> > > + * \return Output clock value
> > > + */
> > > +uint64_t ClockRecovery::getOutput(uint64_t input)
> > > +{
> > > +       double x = static_cast<int64_t>(input - inputBase_);
> > > +       double y = slope_ * x + offset_;
> > > +       uint64_t output = y + x + outputBase_;
> > > +
> > > +       LOG(ClockRec, Debug) << "getOutput " << input << " " << output;
> > > +
> > > +       return output;
> > > +}
> > > +
> > > +} /* namespace libcamera */
> > > diff --git a/src/libcamera/meson.build b/src/libcamera/meson.build
> > > index 57fde8a8..4eaa1c8e 100644
> > > --- a/src/libcamera/meson.build
> > > +++ b/src/libcamera/meson.build
> > > @@ -21,6 +21,7 @@ libcamera_internal_sources = files([
> > >      'byte_stream_buffer.cpp',
> > >      'camera_controls.cpp',
> > >      'camera_lens.cpp',
> > > +    'clock_recovery.cpp',
> > >      'control_serializer.cpp',
> > >      'control_validator.cpp',
> > >      'converter.cpp',
> > > --
> > > 2.39.5
> > >
Barnabás Pőcze Dec. 12, 2024, 5:56 p.m. UTC | #4
Hi


2024. december 6., péntek 15:27 keltezéssel, David Plowman <david.plowman@raspberrypi.com> írta:

> The ClockRecovery class takes pairs of timestamps from two different
> clocks, and models the second ("output") clock from the first ("input")
> clock.
> 
> We can use it, in particular, to get a good wallclock estimate for a
> frame's SensorTimestamp.
> 
> Signed-off-by: David Plowman <david.plowman@raspberrypi.com>
> ---
>  include/libcamera/internal/clock_recovery.h |  72 +++++++
>  include/libcamera/internal/meson.build      |   1 +
>  src/libcamera/clock_recovery.cpp            | 207 ++++++++++++++++++++
>  src/libcamera/meson.build                   |   1 +
>  4 files changed, 281 insertions(+)
>  create mode 100644 include/libcamera/internal/clock_recovery.h
>  create mode 100644 src/libcamera/clock_recovery.cpp
> 
> diff --git a/include/libcamera/internal/clock_recovery.h b/include/libcamera/internal/clock_recovery.h
> new file mode 100644
> index 00000000..c874574e
> --- /dev/null
> +++ b/include/libcamera/internal/clock_recovery.h
> @@ -0,0 +1,72 @@
> +/* SPDX-License-Identifier: BSD-2-Clause */
> +/*
> + * Copyright (C) 2024, Raspberry Pi Ltd
> + *
> + * Camera recovery algorithm
> + */
> +#pragma once
> +
> +#include <stdint.h>
> +
> +namespace libcamera {
> +
> +class ClockRecovery
> +{
> +public:
> +	ClockRecovery();
> +
> +	/* Set configuration parameters. */
> +	void configure(unsigned int numPts = 100, unsigned int maxJitter = 2000, unsigned int minPts = 10,
> +		       unsigned int errorThreshold = 50000);
> +	/* Erase all history and restart the fitting process. */
> +	void reset();
> +
> +	/*
> +	 * Add a new input clock / output clock sample, taking the input from the Linux
> +	 * CLOCK_BOOTTIME and the output from the CLOCK_REALTIME.
> +	 */
> +	void addSample();
> +	/*
> +	 * Add a new input clock / output clock sample, specifying the clock times exactly. Use this
> +	 * when you want to use clocks other than the ones described above.
> +	 */
> +	void addSample(uint64_t input, uint64_t output);
> +	/* Calculate the output clock value for this input. */
> +	uint64_t getOutput(uint64_t input);
> +
> +private:
> +	unsigned int numPts_; /* how many samples contribute to the history */
> +	unsigned int maxJitter_; /* smooth out any jitter larger than this immediately */
> +	unsigned int minPts_; /* number of samples below which we treat clocks as 1:1 */
> +	unsigned int errorThreshold_; /* reset everything when the error exceeds this */
> +
> +	unsigned int count_; /* how many samples seen (up to numPts_) */
> +	uint64_t inputBase_; /* subtract this from all input values, just to make the numbers easier */
> +	uint64_t outputBase_; /* as above, for the output */
> +
> +	uint64_t lastInput_; /* the previous input sample */
> +	uint64_t lastOutput_; /* the previous output sample */
> +
> +	/*
> +	 * We do a linear regression of y against x, where:
> +	 * x is the value input - inputBase_, and
> +	 * y is the value output - outputBase_ - x.
> +	 * We additionally subtract x from y so that y "should" be zero, again making the numnbers easier.
> +	 */
> +	double xAve_; /* average x value seen so far */
> +	double yAve_; /* average y value seen so far */
> +	double x2Ave_; /* average x^2 value seen so far */
> +	double xyAve_; /* average x*y value seen so far */
> +
> +	/*
> +	 * Once we've seen more than minPts_ samples, we recalculate the slope and offset according
> +	 * to the linear regression normal equations.
> +	 */
> +	double slope_; /* latest slope value */
> +	double offset_; /* latest offset value */
> +
> +	/* We use this cumulative error to monitor spontaneous system clock updates. */
> +	double error_;
> +};
> +
> +} /* namespace libcamera */
> diff --git a/include/libcamera/internal/meson.build b/include/libcamera/internal/meson.build
> index 7d6aa8b7..41500636 100644
> --- a/include/libcamera/internal/meson.build
> +++ b/include/libcamera/internal/meson.build
> @@ -11,6 +11,7 @@ libcamera_internal_headers = files([
>      'camera_manager.h',
>      'camera_sensor.h',
>      'camera_sensor_properties.h',
> +    'clock_recovery.h',
>      'control_serializer.h',
>      'control_validator.h',
>      'converter.h',
> diff --git a/src/libcamera/clock_recovery.cpp b/src/libcamera/clock_recovery.cpp
> new file mode 100644
> index 00000000..966599ee
> --- /dev/null
> +++ b/src/libcamera/clock_recovery.cpp
> @@ -0,0 +1,207 @@
> +/* SPDX-License-Identifier: BSD-2-Clause */
> +/*
> + * Copyright (C) 2024, Raspberry Pi Ltd
> + *
> + * Clock recovery algorithm
> + */
> +
> +#include "libcamera/internal/clock_recovery.h"
> +
> +#include <time.h>
> +
> +#include <libcamera/base/log.h>
> +
> +/**
> + * \file clock_recovery.h
> + * \brief Clock recovery - deriving one clock from another independent clock
> + */
> +
> +namespace libcamera {
> +
> +LOG_DEFINE_CATEGORY(ClockRec)
> +
> +/**
> + * \class ClockRecovery
> + * \brief Recover an output clock from an input clock
> + *
> + * The ClockRecovery class derives an output clock from an input clock,
> + * modelling the output clock as being linearly related to the input clock.
> + * For example, we may use it to derive wall clock timestamps from timestamps
> + * measured by the internal system clock which counts local time since boot.
> + *
> + * When pairs of corresponding input and output timestamps are available,
> + * they should be submitted to the model with addSample(). The model will
> + * update, and output clock values for known input clock values can be
> + * obtained using getOutput().
> + *
> + * As a convenience, if the input clock is indeed the time since boot, and the
> + * output clock represents a real wallclock time, then addSample() can be
> + * called with no arguments, and a pair of timestamps will be captured at
> + * that moment.
> + *
> + * The configure() function accepts some configuration parameters to control
> + * the linear fitting process.
> + */
> +
> +/**
> + * \brief Construct a ClockRecovery
> + */
> +ClockRecovery::ClockRecovery()
> +{
> +	configure();
> +	reset();
> +}
> +
> +/**
> + * \brief Set configuration parameters
> + * \param[in] numPts The approximate duration for which the state of the model
> + * is persistent, measured in samples
> + * \param[in] maxJitter New output samples are clamped to no more than this
> + * amount of jitter, to prevent sudden swings from having a large effect
> + * \param[in] minPts The fitted clock model is not used to generate outputs
> + * until this many samples have been received
> + * \param[in] errorThreshold If the accumulated differences between input and
> + * output clocks reaches this amount over a few frames, the model is reset
> + */
> +void ClockRecovery::configure(unsigned int numPts, unsigned int maxJitter, unsigned int minPts,
> +			      unsigned int errorThreshold)
> +{
> +	LOG(ClockRec, Debug)
> +		<< "configure " << numPts << " " << maxJitter << " " << minPts << " " << errorThreshold;
> +
> +	numPts_ = numPts;
> +	maxJitter_ = maxJitter;
> +	minPts_ = minPts;
> +	errorThreshold_ = errorThreshold;
> +}
> +
> +/**
> + * \brief Reset the clock recovery model and start again from scratch
> + */
> +void ClockRecovery::reset()
> +{
> +	LOG(ClockRec, Debug) << "reset";
> +
> +	lastInput_ = 0;
> +	lastOutput_ = 0;
> +	xAve_ = 0;
> +	yAve_ = 0;
> +	x2Ave_ = 0;
> +	xyAve_ = 0;
> +	count_ = 0;
> +	slope_ = 0.0;
> +	offset_ = 0.0;
> +	error_ = 0.0;
> +}
> +
> +/**
> + * \brief Add a sample point to the clock recovery model, for recovering a wall
> + * clock value from the internal system time since boot
> + *
> + * This is a convenience function to make it easy to derive a wall clock value
> + * (using the Linux CLOCK_REALTIME) from the time since the system started
> + * (measured by CLOCK_BOOTTIME).
> + */
> +void ClockRecovery::addSample()
> +{
> +	LOG(ClockRec, Debug) << "addSample";
> +
> +	struct timespec bootTime;
> +	struct timespec wallTime;
> +
> +	/* Get boot and wall clocks in microseconds. */
> +	clock_gettime(CLOCK_BOOTTIME, &bootTime);
> +	clock_gettime(CLOCK_REALTIME, &wallTime);
> +	uint64_t boot = bootTime.tv_sec * 1000000ULL + bootTime.tv_nsec / 1000;
> +	uint64_t wall = wallTime.tv_sec * 1000000ULL + wallTime.tv_nsec / 1000;

It could be that I am missing something that accounts for this, but I am wondering
if it would make sense to sample one of the clocks twice, and average the two samples.
i.e.

  x1 = clock_boottime()
  y = clock_realtime()
  x2 = clock_boottime()

  addSample(midpoint(x1, x2), y)

Otherwise I'd expect a constant offset to be present in the mapping, although
I am not sure if that makes an appreciable difference.


Regards,
Barnabás Pőcze


> +
> +	addSample(boot, wall);
> +}
> +
> +/**
> + * \brief Add a sample point to the clock recovery model, specifying the exact
> + * input and output clock values
> + *
> + * This function should be used for corresponding clocks other than the Linux
> + * BOOTTIME and REALTIME clocks.
> + */
> +void ClockRecovery::addSample(uint64_t input, uint64_t output)
> +{
> +	LOG(ClockRec, Debug) << "addSample " << input << " " << output;
> +
> +	if (count_ == 0) {
> +		inputBase_ = input;
> +		outputBase_ = output;
> +	}
> +
> +	/*
> +	 * We keep an eye on cumulative drift over the last several frames. If this exceeds a
> +	 * threshold, then probably the system clock has been updated and we're going to have to
> +	 * reset everything and start over.
> +	 */
> +	if (lastOutput_) {
> +		int64_t inputDiff = getOutput(input) - getOutput(lastInput_);
> +		int64_t outputDiff = output - lastOutput_;
> +		error_ = error_ * 0.95 + (outputDiff - inputDiff);
> +		if (std::abs(error_) > errorThreshold_) {
> +			reset();
> +			inputBase_ = input;
> +			outputBase_ = output;
> +		}
> +	}
> +	lastInput_ = input;
> +	lastOutput_ = output;
> +
> +	/*
> +	 * Never let the new output value be more than maxJitter_ away from what we would have expected.
> +	 * This is just to reduce the effect of sudden large delays in the measured output.
> +	 */
> +	uint64_t expectedOutput = getOutput(input);
> +	output = std::clamp(output, expectedOutput - maxJitter_, expectedOutput + maxJitter_);
> +
> +	/*
> +	 * We use x, y, x^2 and x*y sums to calculate the best fit line. Here we update them by
> +	 * pretending we have count_ samples at the previous fit, and now one new one. Gradually
> +	 * the effect of the older values gets lost. This is a very simple way of updating the
> +	 * fit (there are much more complicated ones!), but it works well enough. Using averages
> +	 * instead of sums makes the relative effect of old values and the new sample clearer.
> +	 */
> +	double x = static_cast<int64_t>(input - inputBase_);
> +	double y = static_cast<int64_t>(output - outputBase_) - x;
> +	unsigned int count1 = count_ + 1;
> +	xAve_ = (count_ * xAve_ + x) / count1;
> +	yAve_ = (count_ * yAve_ + y) / count1;
> +	x2Ave_ = (count_ * x2Ave_ + x * x) / count1;
> +	xyAve_ = (count_ * xyAve_ + x * y) / count1;
> +
> +	/* Don't update slope and offset until we've seen "enough" sample points. */
> +	if (count_ > minPts_) {
> +		/* These are the standard equations for least squares linear regression. */
> +		slope_ = (count1 * count1 * xyAve_ - count1 * xAve_ * count1 * yAve_) /
> +			 (count1 * count1 * x2Ave_ - count1 * xAve_ * count1 * xAve_);
> +		offset_ = yAve_ - slope_ * xAve_;
> +	}
> +
> +	/* Don't increase count_ above numPts_, as this controls the long-term amount of the residual fit. */
> +	if (count1 < numPts_)
> +		count_++;
> +}
> +
> +/**
> + * \brief Calculate the output clock value according to the model from an input
> + * clock value
> + *
> + * \return Output clock value
> + */
> +uint64_t ClockRecovery::getOutput(uint64_t input)
> +{
> +	double x = static_cast<int64_t>(input - inputBase_);
> +	double y = slope_ * x + offset_;
> +	uint64_t output = y + x + outputBase_;
> +
> +	LOG(ClockRec, Debug) << "getOutput " << input << " " << output;
> +
> +	return output;
> +}
> +
> +} /* namespace libcamera */
> diff --git a/src/libcamera/meson.build b/src/libcamera/meson.build
> index 57fde8a8..4eaa1c8e 100644
> --- a/src/libcamera/meson.build
> +++ b/src/libcamera/meson.build
> @@ -21,6 +21,7 @@ libcamera_internal_sources = files([
>      'byte_stream_buffer.cpp',
>      'camera_controls.cpp',
>      'camera_lens.cpp',
> +    'clock_recovery.cpp',
>      'control_serializer.cpp',
>      'control_validator.cpp',
>      'converter.cpp',
> --
> 2.39.5
David Plowman Dec. 13, 2024, 9:21 a.m. UTC | #5
Hi Barnabas

Thanks for looking at this!

On Thu, 12 Dec 2024 at 17:56, Barnabás Pőcze <pobrn@protonmail.com> wrote:
>
> Hi
>
>
> 2024. december 6., péntek 15:27 keltezéssel, David Plowman <david.plowman@raspberrypi.com> írta:
>
> > The ClockRecovery class takes pairs of timestamps from two different
> > clocks, and models the second ("output") clock from the first ("input")
> > clock.
> >
> > We can use it, in particular, to get a good wallclock estimate for a
> > frame's SensorTimestamp.
> >
> > Signed-off-by: David Plowman <david.plowman@raspberrypi.com>
> > ---
> >  include/libcamera/internal/clock_recovery.h |  72 +++++++
> >  include/libcamera/internal/meson.build      |   1 +
> >  src/libcamera/clock_recovery.cpp            | 207 ++++++++++++++++++++
> >  src/libcamera/meson.build                   |   1 +
> >  4 files changed, 281 insertions(+)
> >  create mode 100644 include/libcamera/internal/clock_recovery.h
> >  create mode 100644 src/libcamera/clock_recovery.cpp
> >
> > diff --git a/include/libcamera/internal/clock_recovery.h b/include/libcamera/internal/clock_recovery.h
> > new file mode 100644
> > index 00000000..c874574e
> > --- /dev/null
> > +++ b/include/libcamera/internal/clock_recovery.h
> > @@ -0,0 +1,72 @@
> > +/* SPDX-License-Identifier: BSD-2-Clause */
> > +/*
> > + * Copyright (C) 2024, Raspberry Pi Ltd
> > + *
> > + * Camera recovery algorithm
> > + */
> > +#pragma once
> > +
> > +#include <stdint.h>
> > +
> > +namespace libcamera {
> > +
> > +class ClockRecovery
> > +{
> > +public:
> > +     ClockRecovery();
> > +
> > +     /* Set configuration parameters. */
> > +     void configure(unsigned int numPts = 100, unsigned int maxJitter = 2000, unsigned int minPts = 10,
> > +                    unsigned int errorThreshold = 50000);
> > +     /* Erase all history and restart the fitting process. */
> > +     void reset();
> > +
> > +     /*
> > +      * Add a new input clock / output clock sample, taking the input from the Linux
> > +      * CLOCK_BOOTTIME and the output from the CLOCK_REALTIME.
> > +      */
> > +     void addSample();
> > +     /*
> > +      * Add a new input clock / output clock sample, specifying the clock times exactly. Use this
> > +      * when you want to use clocks other than the ones described above.
> > +      */
> > +     void addSample(uint64_t input, uint64_t output);
> > +     /* Calculate the output clock value for this input. */
> > +     uint64_t getOutput(uint64_t input);
> > +
> > +private:
> > +     unsigned int numPts_; /* how many samples contribute to the history */
> > +     unsigned int maxJitter_; /* smooth out any jitter larger than this immediately */
> > +     unsigned int minPts_; /* number of samples below which we treat clocks as 1:1 */
> > +     unsigned int errorThreshold_; /* reset everything when the error exceeds this */
> > +
> > +     unsigned int count_; /* how many samples seen (up to numPts_) */
> > +     uint64_t inputBase_; /* subtract this from all input values, just to make the numbers easier */
> > +     uint64_t outputBase_; /* as above, for the output */
> > +
> > +     uint64_t lastInput_; /* the previous input sample */
> > +     uint64_t lastOutput_; /* the previous output sample */
> > +
> > +     /*
> > +      * We do a linear regression of y against x, where:
> > +      * x is the value input - inputBase_, and
> > +      * y is the value output - outputBase_ - x.
> > +      * We additionally subtract x from y so that y "should" be zero, again making the numnbers easier.
> > +      */
> > +     double xAve_; /* average x value seen so far */
> > +     double yAve_; /* average y value seen so far */
> > +     double x2Ave_; /* average x^2 value seen so far */
> > +     double xyAve_; /* average x*y value seen so far */
> > +
> > +     /*
> > +      * Once we've seen more than minPts_ samples, we recalculate the slope and offset according
> > +      * to the linear regression normal equations.
> > +      */
> > +     double slope_; /* latest slope value */
> > +     double offset_; /* latest offset value */
> > +
> > +     /* We use this cumulative error to monitor spontaneous system clock updates. */
> > +     double error_;
> > +};
> > +
> > +} /* namespace libcamera */
> > diff --git a/include/libcamera/internal/meson.build b/include/libcamera/internal/meson.build
> > index 7d6aa8b7..41500636 100644
> > --- a/include/libcamera/internal/meson.build
> > +++ b/include/libcamera/internal/meson.build
> > @@ -11,6 +11,7 @@ libcamera_internal_headers = files([
> >      'camera_manager.h',
> >      'camera_sensor.h',
> >      'camera_sensor_properties.h',
> > +    'clock_recovery.h',
> >      'control_serializer.h',
> >      'control_validator.h',
> >      'converter.h',
> > diff --git a/src/libcamera/clock_recovery.cpp b/src/libcamera/clock_recovery.cpp
> > new file mode 100644
> > index 00000000..966599ee
> > --- /dev/null
> > +++ b/src/libcamera/clock_recovery.cpp
> > @@ -0,0 +1,207 @@
> > +/* SPDX-License-Identifier: BSD-2-Clause */
> > +/*
> > + * Copyright (C) 2024, Raspberry Pi Ltd
> > + *
> > + * Clock recovery algorithm
> > + */
> > +
> > +#include "libcamera/internal/clock_recovery.h"
> > +
> > +#include <time.h>
> > +
> > +#include <libcamera/base/log.h>
> > +
> > +/**
> > + * \file clock_recovery.h
> > + * \brief Clock recovery - deriving one clock from another independent clock
> > + */
> > +
> > +namespace libcamera {
> > +
> > +LOG_DEFINE_CATEGORY(ClockRec)
> > +
> > +/**
> > + * \class ClockRecovery
> > + * \brief Recover an output clock from an input clock
> > + *
> > + * The ClockRecovery class derives an output clock from an input clock,
> > + * modelling the output clock as being linearly related to the input clock.
> > + * For example, we may use it to derive wall clock timestamps from timestamps
> > + * measured by the internal system clock which counts local time since boot.
> > + *
> > + * When pairs of corresponding input and output timestamps are available,
> > + * they should be submitted to the model with addSample(). The model will
> > + * update, and output clock values for known input clock values can be
> > + * obtained using getOutput().
> > + *
> > + * As a convenience, if the input clock is indeed the time since boot, and the
> > + * output clock represents a real wallclock time, then addSample() can be
> > + * called with no arguments, and a pair of timestamps will be captured at
> > + * that moment.
> > + *
> > + * The configure() function accepts some configuration parameters to control
> > + * the linear fitting process.
> > + */
> > +
> > +/**
> > + * \brief Construct a ClockRecovery
> > + */
> > +ClockRecovery::ClockRecovery()
> > +{
> > +     configure();
> > +     reset();
> > +}
> > +
> > +/**
> > + * \brief Set configuration parameters
> > + * \param[in] numPts The approximate duration for which the state of the model
> > + * is persistent, measured in samples
> > + * \param[in] maxJitter New output samples are clamped to no more than this
> > + * amount of jitter, to prevent sudden swings from having a large effect
> > + * \param[in] minPts The fitted clock model is not used to generate outputs
> > + * until this many samples have been received
> > + * \param[in] errorThreshold If the accumulated differences between input and
> > + * output clocks reaches this amount over a few frames, the model is reset
> > + */
> > +void ClockRecovery::configure(unsigned int numPts, unsigned int maxJitter, unsigned int minPts,
> > +                           unsigned int errorThreshold)
> > +{
> > +     LOG(ClockRec, Debug)
> > +             << "configure " << numPts << " " << maxJitter << " " << minPts << " " << errorThreshold;
> > +
> > +     numPts_ = numPts;
> > +     maxJitter_ = maxJitter;
> > +     minPts_ = minPts;
> > +     errorThreshold_ = errorThreshold;
> > +}
> > +
> > +/**
> > + * \brief Reset the clock recovery model and start again from scratch
> > + */
> > +void ClockRecovery::reset()
> > +{
> > +     LOG(ClockRec, Debug) << "reset";
> > +
> > +     lastInput_ = 0;
> > +     lastOutput_ = 0;
> > +     xAve_ = 0;
> > +     yAve_ = 0;
> > +     x2Ave_ = 0;
> > +     xyAve_ = 0;
> > +     count_ = 0;
> > +     slope_ = 0.0;
> > +     offset_ = 0.0;
> > +     error_ = 0.0;
> > +}
> > +
> > +/**
> > + * \brief Add a sample point to the clock recovery model, for recovering a wall
> > + * clock value from the internal system time since boot
> > + *
> > + * This is a convenience function to make it easy to derive a wall clock value
> > + * (using the Linux CLOCK_REALTIME) from the time since the system started
> > + * (measured by CLOCK_BOOTTIME).
> > + */
> > +void ClockRecovery::addSample()
> > +{
> > +     LOG(ClockRec, Debug) << "addSample";
> > +
> > +     struct timespec bootTime;
> > +     struct timespec wallTime;
> > +
> > +     /* Get boot and wall clocks in microseconds. */
> > +     clock_gettime(CLOCK_BOOTTIME, &bootTime);
> > +     clock_gettime(CLOCK_REALTIME, &wallTime);
> > +     uint64_t boot = bootTime.tv_sec * 1000000ULL + bootTime.tv_nsec / 1000;
> > +     uint64_t wall = wallTime.tv_sec * 1000000ULL + wallTime.tv_nsec / 1000;
>
> It could be that I am missing something that accounts for this, but I am wondering
> if it would make sense to sample one of the clocks twice, and average the two samples.
> i.e.
>
>   x1 = clock_boottime()
>   y = clock_realtime()
>   x2 = clock_boottime()
>
>   addSample(midpoint(x1, x2), y)
>
> Otherwise I'd expect a constant offset to be present in the mapping, although
> I am not sure if that makes an appreciable difference.
>
>
> Regards,
> Barnabás Pőcze

That's a good suggestion, so thanks for that, I'll give it a try.

I think you're also right that, in reality, there will be no
meaningful difference - we're only counting microseconds - and for the
purposes of camera synchronisation, fixed offsets don't matter. But
it's still a good thought, so I'll try it!

Thanks
David

>
>
> > +
> > +     addSample(boot, wall);
> > +}
> > +
> > +/**
> > + * \brief Add a sample point to the clock recovery model, specifying the exact
> > + * input and output clock values
> > + *
> > + * This function should be used for corresponding clocks other than the Linux
> > + * BOOTTIME and REALTIME clocks.
> > + */
> > +void ClockRecovery::addSample(uint64_t input, uint64_t output)
> > +{
> > +     LOG(ClockRec, Debug) << "addSample " << input << " " << output;
> > +
> > +     if (count_ == 0) {
> > +             inputBase_ = input;
> > +             outputBase_ = output;
> > +     }
> > +
> > +     /*
> > +      * We keep an eye on cumulative drift over the last several frames. If this exceeds a
> > +      * threshold, then probably the system clock has been updated and we're going to have to
> > +      * reset everything and start over.
> > +      */
> > +     if (lastOutput_) {
> > +             int64_t inputDiff = getOutput(input) - getOutput(lastInput_);
> > +             int64_t outputDiff = output - lastOutput_;
> > +             error_ = error_ * 0.95 + (outputDiff - inputDiff);
> > +             if (std::abs(error_) > errorThreshold_) {
> > +                     reset();
> > +                     inputBase_ = input;
> > +                     outputBase_ = output;
> > +             }
> > +     }
> > +     lastInput_ = input;
> > +     lastOutput_ = output;
> > +
> > +     /*
> > +      * Never let the new output value be more than maxJitter_ away from what we would have expected.
> > +      * This is just to reduce the effect of sudden large delays in the measured output.
> > +      */
> > +     uint64_t expectedOutput = getOutput(input);
> > +     output = std::clamp(output, expectedOutput - maxJitter_, expectedOutput + maxJitter_);
> > +
> > +     /*
> > +      * We use x, y, x^2 and x*y sums to calculate the best fit line. Here we update them by
> > +      * pretending we have count_ samples at the previous fit, and now one new one. Gradually
> > +      * the effect of the older values gets lost. This is a very simple way of updating the
> > +      * fit (there are much more complicated ones!), but it works well enough. Using averages
> > +      * instead of sums makes the relative effect of old values and the new sample clearer.
> > +      */
> > +     double x = static_cast<int64_t>(input - inputBase_);
> > +     double y = static_cast<int64_t>(output - outputBase_) - x;
> > +     unsigned int count1 = count_ + 1;
> > +     xAve_ = (count_ * xAve_ + x) / count1;
> > +     yAve_ = (count_ * yAve_ + y) / count1;
> > +     x2Ave_ = (count_ * x2Ave_ + x * x) / count1;
> > +     xyAve_ = (count_ * xyAve_ + x * y) / count1;
> > +
> > +     /* Don't update slope and offset until we've seen "enough" sample points. */
> > +     if (count_ > minPts_) {
> > +             /* These are the standard equations for least squares linear regression. */
> > +             slope_ = (count1 * count1 * xyAve_ - count1 * xAve_ * count1 * yAve_) /
> > +                      (count1 * count1 * x2Ave_ - count1 * xAve_ * count1 * xAve_);
> > +             offset_ = yAve_ - slope_ * xAve_;
> > +     }
> > +
> > +     /* Don't increase count_ above numPts_, as this controls the long-term amount of the residual fit. */
> > +     if (count1 < numPts_)
> > +             count_++;
> > +}
> > +
> > +/**
> > + * \brief Calculate the output clock value according to the model from an input
> > + * clock value
> > + *
> > + * \return Output clock value
> > + */
> > +uint64_t ClockRecovery::getOutput(uint64_t input)
> > +{
> > +     double x = static_cast<int64_t>(input - inputBase_);
> > +     double y = slope_ * x + offset_;
> > +     uint64_t output = y + x + outputBase_;
> > +
> > +     LOG(ClockRec, Debug) << "getOutput " << input << " " << output;
> > +
> > +     return output;
> > +}
> > +
> > +} /* namespace libcamera */
> > diff --git a/src/libcamera/meson.build b/src/libcamera/meson.build
> > index 57fde8a8..4eaa1c8e 100644
> > --- a/src/libcamera/meson.build
> > +++ b/src/libcamera/meson.build
> > @@ -21,6 +21,7 @@ libcamera_internal_sources = files([
> >      'byte_stream_buffer.cpp',
> >      'camera_controls.cpp',
> >      'camera_lens.cpp',
> > +    'clock_recovery.cpp',
> >      'control_serializer.cpp',
> >      'control_validator.cpp',
> >      'converter.cpp',
> > --
> > 2.39.5
Laurent Pinchart Dec. 18, 2024, 2:21 a.m. UTC | #6
On Fri, Dec 06, 2024 at 04:43:43PM +0000, Kieran Bingham wrote:
> Quoting David Plowman (2024-12-06 16:29:02)
> > On Fri, 6 Dec 2024 at 15:50, Kieran Bingham wrote:
> > > Quoting David Plowman (2024-12-06 14:27:39)
> > > > The ClockRecovery class takes pairs of timestamps from two different
> > > > clocks, and models the second ("output") clock from the first ("input")
> > > > clock.
> > > >
> > > > We can use it, in particular, to get a good wallclock estimate for a
> > > > frame's SensorTimestamp.
> > > >
> > > > Signed-off-by: David Plowman <david.plowman@raspberrypi.com>
> > > > ---
> > > >  include/libcamera/internal/clock_recovery.h |  72 +++++++
> > > >  include/libcamera/internal/meson.build      |   1 +
> > > >  src/libcamera/clock_recovery.cpp            | 207 ++++++++++++++++++++
> > > >  src/libcamera/meson.build                   |   1 +
> > > >  4 files changed, 281 insertions(+)
> > > >  create mode 100644 include/libcamera/internal/clock_recovery.h
> > > >  create mode 100644 src/libcamera/clock_recovery.cpp
> > > >
> > > > diff --git a/include/libcamera/internal/clock_recovery.h b/include/libcamera/internal/clock_recovery.h
> > > > new file mode 100644
> > > > index 00000000..c874574e
> > > > --- /dev/null
> > > > +++ b/include/libcamera/internal/clock_recovery.h
> > > > @@ -0,0 +1,72 @@
> > > > +/* SPDX-License-Identifier: BSD-2-Clause */

All headers are LGPL-licensed, can we do the same here ?

> > > > +/*
> > > > + * Copyright (C) 2024, Raspberry Pi Ltd
> > > > + *
> > > > + * Camera recovery algorithm
> > >
> > > Recovering cameras ? ;-)
> > 
> > Hehe. I'll fix that!!
> > 
> > > > + */
> > > > +#pragma once
> > > > +
> > > > +#include <stdint.h>
> > > > +
> > > > +namespace libcamera {
> > > > +
> > > > +class ClockRecovery
> > > > +{
> > > > +public:
> > > > +       ClockRecovery();
> > > > +
> > > > +       /* Set configuration parameters. */

All documentation goes to the .cpp file.

> > > > +       void configure(unsigned int numPts = 100, unsigned int maxJitter = 2000, unsigned int minPts = 10,
> > > > +                      unsigned int errorThreshold = 50000);

	void configure(unsigned int numPts = 100, unsigned int maxJitter = 2000,
		       unsigned int minPts = 10, unsigned int errorThreshold = 50000);

> > > > +       /* Erase all history and restart the fitting process. */
> > > > +       void reset();
> > > > +
> > > > +       /*
> > > > +        * Add a new input clock / output clock sample, taking the input from the Linux
> > > > +        * CLOCK_BOOTTIME and the output from the CLOCK_REALTIME.
> > > > +        */
> > > > +       void addSample();
> > > > +       /*
> > > > +        * Add a new input clock / output clock sample, specifying the clock times exactly. Use this
> > > > +        * when you want to use clocks other than the ones described above.
> > > > +        */
> > > > +       void addSample(uint64_t input, uint64_t output);
> > > > +       /* Calculate the output clock value for this input. */
> > > > +       uint64_t getOutput(uint64_t input);
> > > > +
> > > > +private:
> > > > +       unsigned int numPts_; /* how many samples contribute to the history */
> > > > +       unsigned int maxJitter_; /* smooth out any jitter larger than this immediately */
> > > > +       unsigned int minPts_; /* number of samples below which we treat clocks as 1:1 */
> > > > +       unsigned int errorThreshold_; /* reset everything when the error exceeds this */
> > > > +
> > > > +       unsigned int count_; /* how many samples seen (up to numPts_) */
> > > > +       uint64_t inputBase_; /* subtract this from all input values, just to make the numbers easier */
> > > > +       uint64_t outputBase_; /* as above, for the output */
> > > > +
> > > > +       uint64_t lastInput_; /* the previous input sample */
> > > > +       uint64_t lastOutput_; /* the previous output sample */
> > > > +
> > > > +       /*
> > > > +        * We do a linear regression of y against x, where:
> > > > +        * x is the value input - inputBase_, and
> > > > +        * y is the value output - outputBase_ - x.
> > > > +        * We additionally subtract x from y so that y "should" be zero, again making the numnbers easier.
> > > > +        */
> > > > +       double xAve_; /* average x value seen so far */
> > > > +       double yAve_; /* average y value seen so far */
> > > > +       double x2Ave_; /* average x^2 value seen so far */
> > > > +       double xyAve_; /* average x*y value seen so far */
> > > > +
> > > > +       /*
> > > > +        * Once we've seen more than minPts_ samples, we recalculate the slope and offset according
> > > > +        * to the linear regression normal equations.
> > > > +        */
> > > > +       double slope_; /* latest slope value */
> > > > +       double offset_; /* latest offset value */
> > > > +
> > > > +       /* We use this cumulative error to monitor spontaneous system clock updates. */
> > > > +       double error_;
> > > > +};
> > > > +
> > > > +} /* namespace libcamera */
> > > > diff --git a/include/libcamera/internal/meson.build b/include/libcamera/internal/meson.build
> > > > index 7d6aa8b7..41500636 100644
> > > > --- a/include/libcamera/internal/meson.build
> > > > +++ b/include/libcamera/internal/meson.build
> > > > @@ -11,6 +11,7 @@ libcamera_internal_headers = files([
> > > >      'camera_manager.h',
> > > >      'camera_sensor.h',
> > > >      'camera_sensor_properties.h',
> > > > +    'clock_recovery.h',
> > > >      'control_serializer.h',
> > > >      'control_validator.h',
> > > >      'converter.h',
> > > > diff --git a/src/libcamera/clock_recovery.cpp b/src/libcamera/clock_recovery.cpp
> > > > new file mode 100644
> > > > index 00000000..966599ee
> > > > --- /dev/null
> > > > +++ b/src/libcamera/clock_recovery.cpp
> > > > @@ -0,0 +1,207 @@
> > > > +/* SPDX-License-Identifier: BSD-2-Clause */
> > > > +/*
> > > > + * Copyright (C) 2024, Raspberry Pi Ltd
> > > > + *
> > > > + * Clock recovery algorithm
> > > > + */
> > > > +
> > > > +#include "libcamera/internal/clock_recovery.h"
> > > > +
> > > > +#include <time.h>
> > > > +
> > > > +#include <libcamera/base/log.h>
> > > > +
> > > > +/**
> > > > + * \file clock_recovery.h
> > > > + * \brief Clock recovery - deriving one clock from another independent clock
> > > > + */
> > > > +
> > > > +namespace libcamera {
> > > > +
> > > > +LOG_DEFINE_CATEGORY(ClockRec)
> > > > +
> > > > +/**
> > > > + * \class ClockRecovery
> > > > + * \brief Recover an output clock from an input clock
> > > > + *
> > > > + * The ClockRecovery class derives an output clock from an input clock,
> > > > + * modelling the output clock as being linearly related to the input clock.
> > > > + * For example, we may use it to derive wall clock timestamps from timestamps
> > > > + * measured by the internal system clock which counts local time since boot.
> > > > + *
> > > > + * When pairs of corresponding input and output timestamps are available,
> > > > + * they should be submitted to the model with addSample(). The model will
> > > > + * update, and output clock values for known input clock values can be
> > > > + * obtained using getOutput().
> > > > + *
> > > > + * As a convenience, if the input clock is indeed the time since boot, and the
> > > > + * output clock represents a real wallclock time, then addSample() can be
> > > > + * called with no arguments, and a pair of timestamps will be captured at
> > > > + * that moment.
> > > > + *
> > > > + * The configure() function accepts some configuration parameters to control
> > > > + * the linear fitting process.
> > > > + */
> > > > +
> > > > +/**
> > > > + * \brief Construct a ClockRecovery
> > > > + */
> > > > +ClockRecovery::ClockRecovery()
> > > > +{
> > > > +       configure();
> > > > +       reset();
> > > > +}
> > > > +
> > > > +/**
> > > > + * \brief Set configuration parameters
> > > > + * \param[in] numPts The approximate duration for which the state of the model
> > > > + * is persistent, measured in samples

Instead of saying "measured in samples" you could rename the variable to
numSamples. Same elsewhere where applicable.

> > > > + * \param[in] maxJitter New output samples are clamped to no more than this
> > > > + * amount of jitter, to prevent sudden swings from having a large effect
> > > > + * \param[in] minPts The fitted clock model is not used to generate outputs
> > > > + * until this many samples have been received
> > > > + * \param[in] errorThreshold If the accumulated differences between input and
> > > > + * output clocks reaches this amount over a few frames, the model is reset
> > > > + */
> > > > +void ClockRecovery::configure(unsigned int numPts, unsigned int maxJitter, unsigned int minPts,
> > > > +                             unsigned int errorThreshold)
> > > > +{
> > > > +       LOG(ClockRec, Debug)
> > > > +               << "configure " << numPts << " " << maxJitter << " " << minPts << " " << errorThreshold;
> > > > +
> > > > +       numPts_ = numPts;
> > > > +       maxJitter_ = maxJitter;
> > > > +       minPts_ = minPts;
> > > > +       errorThreshold_ = errorThreshold;
> > > > +}
> > > > +
> > > > +/**
> > > > + * \brief Reset the clock recovery model and start again from scratch
> > > > + */
> > > > +void ClockRecovery::reset()
> > > > +{
> > > > +       LOG(ClockRec, Debug) << "reset";
> > > > +
> > > > +       lastInput_ = 0;
> > > > +       lastOutput_ = 0;
> > > > +       xAve_ = 0;
> > > > +       yAve_ = 0;
> > > > +       x2Ave_ = 0;
> > > > +       xyAve_ = 0;
> > > > +       count_ = 0;
> > > > +       slope_ = 0.0;
> > >
> > > Should slope be initialised to 1.0 or anything 'normalised' for any
> > > startup conditions?
> > 
> > Actually, 0 is the value that ensures "output = input" (allowing for
> > the different starting offset).
> > Maybe worth adding a comment to that end?
> 
> aha - no - that's probably fine - it's just me misunderstanding. For
> some reason I thought '1' would be the more expected slope for
> undefined/uninitialised, but I  see '0' just removes the slope which is
> (until samples are ready...) undefined ;-)
> 
> > > > +       offset_ = 0.0;
> > > > +       error_ = 0.0;
> > > > +}
> > > > +
> > > > +/**
> > > > + * \brief Add a sample point to the clock recovery model, for recovering a wall
> > > > + * clock value from the internal system time since boot
> > > > + *
> > > > + * This is a convenience function to make it easy to derive a wall clock value
> > > > + * (using the Linux CLOCK_REALTIME) from the time since the system started
> > > > + * (measured by CLOCK_BOOTTIME).
> > > > + */
> > > > +void ClockRecovery::addSample()
> > > > +{
> > > > +       LOG(ClockRec, Debug) << "addSample";
> > > > +
> > > > +       struct timespec bootTime;
> > > > +       struct timespec wallTime;
> > > > +
> > > > +       /* Get boot and wall clocks in microseconds. */
> > > > +       clock_gettime(CLOCK_BOOTTIME, &bootTime);
> > > > +       clock_gettime(CLOCK_REALTIME, &wallTime);
> > > > +       uint64_t boot = bootTime.tv_sec * 1000000ULL + bootTime.tv_nsec / 1000;
> > > > +       uint64_t wall = wallTime.tv_sec * 1000000ULL + wallTime.tv_nsec / 1000;

Apart from this, there's nothing in this class that makes it specific to
clocks. I wonder if we could pass a sampler function to the constructor
(or maybe to the configure function, or use a template parameter ...),
and make the class handle linear regressions in a more generic way.

> > > > +
> > > > +       addSample(boot, wall);
> > > > +}
> > > > +
> > > > +/**
> > > > + * \brief Add a sample point to the clock recovery model, specifying the exact
> > > > + * input and output clock values

Missing \param

> > > > + *
> > > > + * This function should be used for corresponding clocks other than the Linux
> > > > + * BOOTTIME and REALTIME clocks.
> > >
> > > I like that this object has both an 'easy addSample()' and the paired
> > > version.
> > >
> > > I think having the addSample() can define that it pairs BOOTTIME and
> > > REALTIME clocks, but this addSample(uint64_t input, uint64_t output)
> > > itself isn't actually tied to those clocks.
> > >
> > > If this ClockRecovery object is used elsewhere, the 'only' reason I
> > > could imagine this function being used is if it were to use a clock
> > > specifically different to the defaults - so I'd probably drop the text
> > > here, and keep the specification of BOOTTIME/REALTIME to the addSample()
> > > - or specific to the usage in V4L2VideoDevice.

The class could be useful in the uvcvideo pipeline handler, to convert
from the hardware clock to the system clock (we'll need two instances
though, as we need to go through the USB clock in the middle).

> > > > +void ClockRecovery::addSample(uint64_t input, uint64_t output)
> > >
> > > As an example that I'm not sure if it's crazy or not yet - I could
> > > imagine setting up a ClockRecovery object that maps sequence number and
> > > timestamp so that you could determine the frame sequence number (output)
> > > from timestamp (input) which would give you quantised sequence numbers
> > > determined from the timestamp - so you could spot frame drops if a CSI
> > > receiver always supplies monotonic sequence numbers. (which I think the
> > > i.MX8MP does... ?)
> > 
> > Interesting thought, though I think I'd prefer to keep the basic
> > ClockRecovery restricted to, well, clocks. If someone wants to spot
> > frame drops... obviously they're welcome to do that, but maybe that
> > gets implemented on top? Remember of course that frame times may be
> > variable.
> 
> Oh absolutly - I'm not expecting anything here - I'm just speculating on
> what crazy things we could match between input/output to get the linear
> regression on ;-)
> 
> And that's why I think "ClockRecovery::addSample(uint64_t input, uint64_t
> output)" as a function with two parameters isn't itself tied to a
> specific clock. (but the usage in V4L2VideoDevice, and addSample() is).

A point that bothers me a bit with the two functions is that it opens
the door to incorrect usage. I wonder if we could improve that. Not
something to look at right now, let's focus on the rest of the series
first. I'll see if I can find a good idea to improve the next version in
this regards.

> > > > +{
> > > > +       LOG(ClockRec, Debug) << "addSample " << input << " " << output;
> > > > +
> > > > +       if (count_ == 0) {
> > > > +               inputBase_ = input;
> > > > +               outputBase_ = output;
> > > > +       }
> > > > +
> > > > +       /*
> > > > +        * We keep an eye on cumulative drift over the last several frames. If this exceeds a
> > > > +        * threshold, then probably the system clock has been updated and we're going to have to
> > > > +        * reset everything and start over.
> > > > +        */
> > >
> > > Throughout the series, there's lots of text that could easily be wrapped
> > > to 80 chars. It shows up a lot in my email because I have an 92 char
> > > terminal for my email which shows lots of lines wrapping, but it's not a
> > > hard requirement.
> > 
> > I guess I'm not clear how strict we are about line lengths... mostly I
> > don't worry if the style checker doesn't complain. But I'm happy to do
> > a bit of reflowing if that helps, though some slightly more "official"
> > guidance on when to do that would be helpful.
> 
> I think we normally aim for blocks of 80 - but the exceptions are up to
> 120 if it helps/makes sense in that instance.

That's right, we have a 80 columns soft limit and a 120 columns harder
limit. For comment blocks there are very few reasons to go over 80
columns.

> It only shows up in this series, as I think everything is targetting
> ~100/120ish instead.
> 
> > > > +       if (lastOutput_) {
> > > > +               int64_t inputDiff = getOutput(input) - getOutput(lastInput_);
> > > > +               int64_t outputDiff = output - lastOutput_;
> > > > +               error_ = error_ * 0.95 + (outputDiff - inputDiff);
> > > > +               if (std::abs(error_) > errorThreshold_) {
> > > > +                       reset();
> > > > +                       inputBase_ = input;
> > > > +                       outputBase_ = output;
> > > > +               }
> > > > +       }
> > > > +       lastInput_ = input;
> > > > +       lastOutput_ = output;
> > > > +
> > > > +       /*
> > > > +        * Never let the new output value be more than maxJitter_ away from what we would have expected.
> > > > +        * This is just to reduce the effect of sudden large delays in the measured output.
> > > > +        */
> > > > +       uint64_t expectedOutput = getOutput(input);
> > > > +       output = std::clamp(output, expectedOutput - maxJitter_, expectedOutput + maxJitter_);
> > > > +
> > > > +       /*
> > > > +        * We use x, y, x^2 and x*y sums to calculate the best fit line. Here we update them by
> > > > +        * pretending we have count_ samples at the previous fit, and now one new one. Gradually
> > > > +        * the effect of the older values gets lost. This is a very simple way of updating the
> > > > +        * fit (there are much more complicated ones!), but it works well enough. Using averages
> > > > +        * instead of sums makes the relative effect of old values and the new sample clearer.
> > > > +        */
> > > > +       double x = static_cast<int64_t>(input - inputBase_);
> > > > +       double y = static_cast<int64_t>(output - outputBase_) - x;
> > > > +       unsigned int count1 = count_ + 1;
> > > > +       xAve_ = (count_ * xAve_ + x) / count1;
> > > > +       yAve_ = (count_ * yAve_ + y) / count1;
> > > > +       x2Ave_ = (count_ * x2Ave_ + x * x) / count1;
> > > > +       xyAve_ = (count_ * xyAve_ + x * y) / count1;
> > > > +
> > > > +       /* Don't update slope and offset until we've seen "enough" sample points. */
> > > > +       if (count_ > minPts_) {
> > >
> > > What's the output from getOuput before this happens ? Anything
> > > problematic? Or just things that can be ignored for the first few
> > > frames?
> > 
> > No, you just get "output = input" allowing for the different starting
> > offsets. Again, maybe worth a comment?
> 
> Hrm, ok so maybe it might be useful to state 'somewhere' that the system
> will produce output values matching the input value until sufficient
> sampling points have been collected.

What should be the behaviour of the FrameWallClock control during that
time ? That should be documented in the control definition. I'm tempted
to say it should not be generated until it becomes stable.

> And following that back, I see minPts_ is configurable (which is a good
> thing :D), so it's up to the use case to determine what the minimum
> requirements are.
> 
> Seems pretty good to me.
> 
> > > > +               /* These are the standard equations for least squares linear regression. */
> > > > +               slope_ = (count1 * count1 * xyAve_ - count1 * xAve_ * count1 * yAve_) /
> > > > +                        (count1 * count1 * x2Ave_ - count1 * xAve_ * count1 * xAve_);
> > > > +               offset_ = yAve_ - slope_ * xAve_;
> > > > +       }
> > > > +
> > > > +       /* Don't increase count_ above numPts_, as this controls the long-term amount of the residual fit. */
> > > > +       if (count1 < numPts_)
> > > > +               count_++;
> > > > +}
> > > > +
> > > > +/**
> > > > + * \brief Calculate the output clock value according to the model from an input
> > > > + * clock value

Missing \param

> > > > + *
> > > > + * \return Output clock value
> > > > + */
> > > > +uint64_t ClockRecovery::getOutput(uint64_t input)
> > > > +{
> > > > +       double x = static_cast<int64_t>(input - inputBase_);
> > > > +       double y = slope_ * x + offset_;
> > > > +       uint64_t output = y + x + outputBase_;
> > > > +
> > > > +       LOG(ClockRec, Debug) << "getOutput " << input << " " << output;
> > > > +
> > > > +       return output;
> > > > +}
> > > > +
> > > > +} /* namespace libcamera */
> > > > diff --git a/src/libcamera/meson.build b/src/libcamera/meson.build
> > > > index 57fde8a8..4eaa1c8e 100644
> > > > --- a/src/libcamera/meson.build
> > > > +++ b/src/libcamera/meson.build
> > > > @@ -21,6 +21,7 @@ libcamera_internal_sources = files([
> > > >      'byte_stream_buffer.cpp',
> > > >      'camera_controls.cpp',
> > > >      'camera_lens.cpp',
> > > > +    'clock_recovery.cpp',
> > > >      'control_serializer.cpp',
> > > >      'control_validator.cpp',
> > > >      'converter.cpp',
David Plowman Dec. 18, 2024, 2:10 p.m. UTC | #7
Hi Laurent

Thanks for the review and comments.

On Wed, 18 Dec 2024 at 02:21, Laurent Pinchart
<laurent.pinchart@ideasonboard.com> wrote:
>
> On Fri, Dec 06, 2024 at 04:43:43PM +0000, Kieran Bingham wrote:
> > Quoting David Plowman (2024-12-06 16:29:02)
> > > On Fri, 6 Dec 2024 at 15:50, Kieran Bingham wrote:
> > > > Quoting David Plowman (2024-12-06 14:27:39)
> > > > > The ClockRecovery class takes pairs of timestamps from two different
> > > > > clocks, and models the second ("output") clock from the first ("input")
> > > > > clock.
> > > > >
> > > > > We can use it, in particular, to get a good wallclock estimate for a
> > > > > frame's SensorTimestamp.
> > > > >
> > > > > Signed-off-by: David Plowman <david.plowman@raspberrypi.com>
> > > > > ---
> > > > >  include/libcamera/internal/clock_recovery.h |  72 +++++++
> > > > >  include/libcamera/internal/meson.build      |   1 +
> > > > >  src/libcamera/clock_recovery.cpp            | 207 ++++++++++++++++++++
> > > > >  src/libcamera/meson.build                   |   1 +
> > > > >  4 files changed, 281 insertions(+)
> > > > >  create mode 100644 include/libcamera/internal/clock_recovery.h
> > > > >  create mode 100644 src/libcamera/clock_recovery.cpp
> > > > >
> > > > > diff --git a/include/libcamera/internal/clock_recovery.h b/include/libcamera/internal/clock_recovery.h
> > > > > new file mode 100644
> > > > > index 00000000..c874574e
> > > > > --- /dev/null
> > > > > +++ b/include/libcamera/internal/clock_recovery.h
> > > > > @@ -0,0 +1,72 @@
> > > > > +/* SPDX-License-Identifier: BSD-2-Clause */
>
> All headers are LGPL-licensed, can we do the same here ?

Will do.

>
> > > > > +/*
> > > > > + * Copyright (C) 2024, Raspberry Pi Ltd
> > > > > + *
> > > > > + * Camera recovery algorithm
> > > >
> > > > Recovering cameras ? ;-)
> > >
> > > Hehe. I'll fix that!!
> > >
> > > > > + */
> > > > > +#pragma once
> > > > > +
> > > > > +#include <stdint.h>
> > > > > +
> > > > > +namespace libcamera {
> > > > > +
> > > > > +class ClockRecovery
> > > > > +{
> > > > > +public:
> > > > > +       ClockRecovery();
> > > > > +
> > > > > +       /* Set configuration parameters. */
>
> All documentation goes to the .cpp file.

I'll delete the comments here, presumably they're still OK when
documenting private members (?) so I'll leave those.

>
> > > > > +       void configure(unsigned int numPts = 100, unsigned int maxJitter = 2000, unsigned int minPts = 10,
> > > > > +                      unsigned int errorThreshold = 50000);
>
>         void configure(unsigned int numPts = 100, unsigned int maxJitter = 2000,
>                        unsigned int minPts = 10, unsigned int errorThreshold = 50000);
>
> > > > > +       /* Erase all history and restart the fitting process. */
> > > > > +       void reset();
> > > > > +
> > > > > +       /*
> > > > > +        * Add a new input clock / output clock sample, taking the input from the Linux
> > > > > +        * CLOCK_BOOTTIME and the output from the CLOCK_REALTIME.
> > > > > +        */
> > > > > +       void addSample();
> > > > > +       /*
> > > > > +        * Add a new input clock / output clock sample, specifying the clock times exactly. Use this
> > > > > +        * when you want to use clocks other than the ones described above.
> > > > > +        */
> > > > > +       void addSample(uint64_t input, uint64_t output);
> > > > > +       /* Calculate the output clock value for this input. */
> > > > > +       uint64_t getOutput(uint64_t input);
> > > > > +
> > > > > +private:
> > > > > +       unsigned int numPts_; /* how many samples contribute to the history */
> > > > > +       unsigned int maxJitter_; /* smooth out any jitter larger than this immediately */
> > > > > +       unsigned int minPts_; /* number of samples below which we treat clocks as 1:1 */
> > > > > +       unsigned int errorThreshold_; /* reset everything when the error exceeds this */
> > > > > +
> > > > > +       unsigned int count_; /* how many samples seen (up to numPts_) */
> > > > > +       uint64_t inputBase_; /* subtract this from all input values, just to make the numbers easier */
> > > > > +       uint64_t outputBase_; /* as above, for the output */
> > > > > +
> > > > > +       uint64_t lastInput_; /* the previous input sample */
> > > > > +       uint64_t lastOutput_; /* the previous output sample */
> > > > > +
> > > > > +       /*
> > > > > +        * We do a linear regression of y against x, where:
> > > > > +        * x is the value input - inputBase_, and
> > > > > +        * y is the value output - outputBase_ - x.
> > > > > +        * We additionally subtract x from y so that y "should" be zero, again making the numnbers easier.
> > > > > +        */
> > > > > +       double xAve_; /* average x value seen so far */
> > > > > +       double yAve_; /* average y value seen so far */
> > > > > +       double x2Ave_; /* average x^2 value seen so far */
> > > > > +       double xyAve_; /* average x*y value seen so far */
> > > > > +
> > > > > +       /*
> > > > > +        * Once we've seen more than minPts_ samples, we recalculate the slope and offset according
> > > > > +        * to the linear regression normal equations.
> > > > > +        */
> > > > > +       double slope_; /* latest slope value */
> > > > > +       double offset_; /* latest offset value */
> > > > > +
> > > > > +       /* We use this cumulative error to monitor spontaneous system clock updates. */
> > > > > +       double error_;
> > > > > +};
> > > > > +
> > > > > +} /* namespace libcamera */
> > > > > diff --git a/include/libcamera/internal/meson.build b/include/libcamera/internal/meson.build
> > > > > index 7d6aa8b7..41500636 100644
> > > > > --- a/include/libcamera/internal/meson.build
> > > > > +++ b/include/libcamera/internal/meson.build
> > > > > @@ -11,6 +11,7 @@ libcamera_internal_headers = files([
> > > > >      'camera_manager.h',
> > > > >      'camera_sensor.h',
> > > > >      'camera_sensor_properties.h',
> > > > > +    'clock_recovery.h',
> > > > >      'control_serializer.h',
> > > > >      'control_validator.h',
> > > > >      'converter.h',
> > > > > diff --git a/src/libcamera/clock_recovery.cpp b/src/libcamera/clock_recovery.cpp
> > > > > new file mode 100644
> > > > > index 00000000..966599ee
> > > > > --- /dev/null
> > > > > +++ b/src/libcamera/clock_recovery.cpp
> > > > > @@ -0,0 +1,207 @@
> > > > > +/* SPDX-License-Identifier: BSD-2-Clause */
> > > > > +/*
> > > > > + * Copyright (C) 2024, Raspberry Pi Ltd
> > > > > + *
> > > > > + * Clock recovery algorithm
> > > > > + */
> > > > > +
> > > > > +#include "libcamera/internal/clock_recovery.h"
> > > > > +
> > > > > +#include <time.h>
> > > > > +
> > > > > +#include <libcamera/base/log.h>
> > > > > +
> > > > > +/**
> > > > > + * \file clock_recovery.h
> > > > > + * \brief Clock recovery - deriving one clock from another independent clock
> > > > > + */
> > > > > +
> > > > > +namespace libcamera {
> > > > > +
> > > > > +LOG_DEFINE_CATEGORY(ClockRec)
> > > > > +
> > > > > +/**
> > > > > + * \class ClockRecovery
> > > > > + * \brief Recover an output clock from an input clock
> > > > > + *
> > > > > + * The ClockRecovery class derives an output clock from an input clock,
> > > > > + * modelling the output clock as being linearly related to the input clock.
> > > > > + * For example, we may use it to derive wall clock timestamps from timestamps
> > > > > + * measured by the internal system clock which counts local time since boot.
> > > > > + *
> > > > > + * When pairs of corresponding input and output timestamps are available,
> > > > > + * they should be submitted to the model with addSample(). The model will
> > > > > + * update, and output clock values for known input clock values can be
> > > > > + * obtained using getOutput().
> > > > > + *
> > > > > + * As a convenience, if the input clock is indeed the time since boot, and the
> > > > > + * output clock represents a real wallclock time, then addSample() can be
> > > > > + * called with no arguments, and a pair of timestamps will be captured at
> > > > > + * that moment.
> > > > > + *
> > > > > + * The configure() function accepts some configuration parameters to control
> > > > > + * the linear fitting process.
> > > > > + */
> > > > > +
> > > > > +/**
> > > > > + * \brief Construct a ClockRecovery
> > > > > + */
> > > > > +ClockRecovery::ClockRecovery()
> > > > > +{
> > > > > +       configure();
> > > > > +       reset();
> > > > > +}
> > > > > +
> > > > > +/**
> > > > > + * \brief Set configuration parameters
> > > > > + * \param[in] numPts The approximate duration for which the state of the model
> > > > > + * is persistent, measured in samples
>
> Instead of saying "measured in samples" you could rename the variable to
> numSamples. Same elsewhere where applicable.

Yes, a name change is probably helpful.

>
> > > > > + * \param[in] maxJitter New output samples are clamped to no more than this
> > > > > + * amount of jitter, to prevent sudden swings from having a large effect
> > > > > + * \param[in] minPts The fitted clock model is not used to generate outputs
> > > > > + * until this many samples have been received
> > > > > + * \param[in] errorThreshold If the accumulated differences between input and
> > > > > + * output clocks reaches this amount over a few frames, the model is reset
> > > > > + */
> > > > > +void ClockRecovery::configure(unsigned int numPts, unsigned int maxJitter, unsigned int minPts,
> > > > > +                             unsigned int errorThreshold)
> > > > > +{
> > > > > +       LOG(ClockRec, Debug)
> > > > > +               << "configure " << numPts << " " << maxJitter << " " << minPts << " " << errorThreshold;
> > > > > +
> > > > > +       numPts_ = numPts;
> > > > > +       maxJitter_ = maxJitter;
> > > > > +       minPts_ = minPts;
> > > > > +       errorThreshold_ = errorThreshold;
> > > > > +}
> > > > > +
> > > > > +/**
> > > > > + * \brief Reset the clock recovery model and start again from scratch
> > > > > + */
> > > > > +void ClockRecovery::reset()
> > > > > +{
> > > > > +       LOG(ClockRec, Debug) << "reset";
> > > > > +
> > > > > +       lastInput_ = 0;
> > > > > +       lastOutput_ = 0;
> > > > > +       xAve_ = 0;
> > > > > +       yAve_ = 0;
> > > > > +       x2Ave_ = 0;
> > > > > +       xyAve_ = 0;
> > > > > +       count_ = 0;
> > > > > +       slope_ = 0.0;
> > > >
> > > > Should slope be initialised to 1.0 or anything 'normalised' for any
> > > > startup conditions?
> > >
> > > Actually, 0 is the value that ensures "output = input" (allowing for
> > > the different starting offset).
> > > Maybe worth adding a comment to that end?
> >
> > aha - no - that's probably fine - it's just me misunderstanding. For
> > some reason I thought '1' would be the more expected slope for
> > undefined/uninitialised, but I  see '0' just removes the slope which is
> > (until samples are ready...) undefined ;-)
> >
> > > > > +       offset_ = 0.0;
> > > > > +       error_ = 0.0;
> > > > > +}
> > > > > +
> > > > > +/**
> > > > > + * \brief Add a sample point to the clock recovery model, for recovering a wall
> > > > > + * clock value from the internal system time since boot
> > > > > + *
> > > > > + * This is a convenience function to make it easy to derive a wall clock value
> > > > > + * (using the Linux CLOCK_REALTIME) from the time since the system started
> > > > > + * (measured by CLOCK_BOOTTIME).
> > > > > + */
> > > > > +void ClockRecovery::addSample()
> > > > > +{
> > > > > +       LOG(ClockRec, Debug) << "addSample";
> > > > > +
> > > > > +       struct timespec bootTime;
> > > > > +       struct timespec wallTime;
> > > > > +
> > > > > +       /* Get boot and wall clocks in microseconds. */
> > > > > +       clock_gettime(CLOCK_BOOTTIME, &bootTime);
> > > > > +       clock_gettime(CLOCK_REALTIME, &wallTime);
> > > > > +       uint64_t boot = bootTime.tv_sec * 1000000ULL + bootTime.tv_nsec / 1000;
> > > > > +       uint64_t wall = wallTime.tv_sec * 1000000ULL + wallTime.tv_nsec / 1000;
>
> Apart from this, there's nothing in this class that makes it specific to
> clocks. I wonder if we could pass a sampler function to the constructor
> (or maybe to the configure function, or use a template parameter ...),
> and make the class handle linear regressions in a more generic way.

That's an interesting thought. Mostly I'm dead nervous about
generalising stuff when I have only a single use case, but I certainly
see the idea here. Folks could supply a "sampler", or maybe derive
from the base class, or something. Maybe the version of addSample with
two parameters becomes "protected"?

On the other hand, as soon as you're doing that, then maybe the
configuration parameters become something you might want to tweak as
well. Maybe making intialise() and addSample() virtual is enough?
There's also a bit of an assumption about the input-to-output rate
defaulting to 1-to-1, and whether that's transferable. The fact that
you know this is pretty helpful - it means you get good output
estimates right off the bat. So... don't know. Maybe? Maybe wait for
another use case to turn up?

>
> > > > > +
> > > > > +       addSample(boot, wall);
> > > > > +}
> > > > > +
> > > > > +/**
> > > > > + * \brief Add a sample point to the clock recovery model, specifying the exact
> > > > > + * input and output clock values
>
> Missing \param

Will add!

>
> > > > > + *
> > > > > + * This function should be used for corresponding clocks other than the Linux
> > > > > + * BOOTTIME and REALTIME clocks.
> > > >
> > > > I like that this object has both an 'easy addSample()' and the paired
> > > > version.
> > > >
> > > > I think having the addSample() can define that it pairs BOOTTIME and
> > > > REALTIME clocks, but this addSample(uint64_t input, uint64_t output)
> > > > itself isn't actually tied to those clocks.
> > > >
> > > > If this ClockRecovery object is used elsewhere, the 'only' reason I
> > > > could imagine this function being used is if it were to use a clock
> > > > specifically different to the defaults - so I'd probably drop the text
> > > > here, and keep the specification of BOOTTIME/REALTIME to the addSample()
> > > > - or specific to the usage in V4L2VideoDevice.
>
> The class could be useful in the uvcvideo pipeline handler, to convert
> from the hardware clock to the system clock (we'll need two instances
> though, as we need to go through the USB clock in the middle).
>
> > > > > +void ClockRecovery::addSample(uint64_t input, uint64_t output)
> > > >
> > > > As an example that I'm not sure if it's crazy or not yet - I could
> > > > imagine setting up a ClockRecovery object that maps sequence number and
> > > > timestamp so that you could determine the frame sequence number (output)
> > > > from timestamp (input) which would give you quantised sequence numbers
> > > > determined from the timestamp - so you could spot frame drops if a CSI
> > > > receiver always supplies monotonic sequence numbers. (which I think the
> > > > i.MX8MP does... ?)
> > >
> > > Interesting thought, though I think I'd prefer to keep the basic
> > > ClockRecovery restricted to, well, clocks. If someone wants to spot
> > > frame drops... obviously they're welcome to do that, but maybe that
> > > gets implemented on top? Remember of course that frame times may be
> > > variable.
> >
> > Oh absolutly - I'm not expecting anything here - I'm just speculating on
> > what crazy things we could match between input/output to get the linear
> > regression on ;-)
> >
> > And that's why I think "ClockRecovery::addSample(uint64_t input, uint64_t
> > output)" as a function with two parameters isn't itself tied to a
> > specific clock. (but the usage in V4L2VideoDevice, and addSample() is).
>
> A point that bothers me a bit with the two functions is that it opens
> the door to incorrect usage. I wonder if we could improve that. Not
> something to look at right now, let's focus on the rest of the series
> first. I'll see if I can find a good idea to improve the next version in
> this regards.

OK, we'll wait and see. I'll probably post an update in the meantime
just to try and settle at least some other things where we can!

>
> > > > > +{
> > > > > +       LOG(ClockRec, Debug) << "addSample " << input << " " << output;
> > > > > +
> > > > > +       if (count_ == 0) {
> > > > > +               inputBase_ = input;
> > > > > +               outputBase_ = output;
> > > > > +       }
> > > > > +
> > > > > +       /*
> > > > > +        * We keep an eye on cumulative drift over the last several frames. If this exceeds a
> > > > > +        * threshold, then probably the system clock has been updated and we're going to have to
> > > > > +        * reset everything and start over.
> > > > > +        */
> > > >
> > > > Throughout the series, there's lots of text that could easily be wrapped
> > > > to 80 chars. It shows up a lot in my email because I have an 92 char
> > > > terminal for my email which shows lots of lines wrapping, but it's not a
> > > > hard requirement.
> > >
> > > I guess I'm not clear how strict we are about line lengths... mostly I
> > > don't worry if the style checker doesn't complain. But I'm happy to do
> > > a bit of reflowing if that helps, though some slightly more "official"
> > > guidance on when to do that would be helpful.
> >
> > I think we normally aim for blocks of 80 - but the exceptions are up to
> > 120 if it helps/makes sense in that instance.
>
> That's right, we have a 80 columns soft limit and a 120 columns harder
> limit. For comment blocks there are very few reasons to go over 80
> columns.
>
> > It only shows up in this series, as I think everything is targetting
> > ~100/120ish instead.
> >
> > > > > +       if (lastOutput_) {
> > > > > +               int64_t inputDiff = getOutput(input) - getOutput(lastInput_);
> > > > > +               int64_t outputDiff = output - lastOutput_;
> > > > > +               error_ = error_ * 0.95 + (outputDiff - inputDiff);
> > > > > +               if (std::abs(error_) > errorThreshold_) {
> > > > > +                       reset();
> > > > > +                       inputBase_ = input;
> > > > > +                       outputBase_ = output;
> > > > > +               }
> > > > > +       }
> > > > > +       lastInput_ = input;
> > > > > +       lastOutput_ = output;
> > > > > +
> > > > > +       /*
> > > > > +        * Never let the new output value be more than maxJitter_ away from what we would have expected.
> > > > > +        * This is just to reduce the effect of sudden large delays in the measured output.
> > > > > +        */
> > > > > +       uint64_t expectedOutput = getOutput(input);
> > > > > +       output = std::clamp(output, expectedOutput - maxJitter_, expectedOutput + maxJitter_);
> > > > > +
> > > > > +       /*
> > > > > +        * We use x, y, x^2 and x*y sums to calculate the best fit line. Here we update them by
> > > > > +        * pretending we have count_ samples at the previous fit, and now one new one. Gradually
> > > > > +        * the effect of the older values gets lost. This is a very simple way of updating the
> > > > > +        * fit (there are much more complicated ones!), but it works well enough. Using averages
> > > > > +        * instead of sums makes the relative effect of old values and the new sample clearer.
> > > > > +        */
> > > > > +       double x = static_cast<int64_t>(input - inputBase_);
> > > > > +       double y = static_cast<int64_t>(output - outputBase_) - x;
> > > > > +       unsigned int count1 = count_ + 1;
> > > > > +       xAve_ = (count_ * xAve_ + x) / count1;
> > > > > +       yAve_ = (count_ * yAve_ + y) / count1;
> > > > > +       x2Ave_ = (count_ * x2Ave_ + x * x) / count1;
> > > > > +       xyAve_ = (count_ * xyAve_ + x * y) / count1;
> > > > > +
> > > > > +       /* Don't update slope and offset until we've seen "enough" sample points. */
> > > > > +       if (count_ > minPts_) {
> > > >
> > > > What's the output from getOuput before this happens ? Anything
> > > > problematic? Or just things that can be ignored for the first few
> > > > frames?
> > >
> > > No, you just get "output = input" allowing for the different starting
> > > offsets. Again, maybe worth a comment?
> >
> > Hrm, ok so maybe it might be useful to state 'somewhere' that the system
> > will produce output values matching the input value until sufficient
> > sampling points have been collected.
>
> What should be the behaviour of the FrameWallClock control during that
> time ? That should be documented in the control definition. I'm tempted
> to say it should not be generated until it becomes stable.

Well, actually I find that I get pretty good estimates right from the
start, in large part because of the 1-to-1 initial assumption. Of
course, there is a risk that, if you're unlucky, you'll end up with an
offset that's slightly skewed, and this would last for several
samples.

If that initial sample is a worry, we could take multiple readings and
average them. Another slightly interesting variation is that, with
Barnabas's modification, you could actually generate a passable bound
for this error (namely bootTime2 - bootTime1). So you could take the
sample where this is smallest.

In fact, there's a lot we could do with something like an error bound,
the question in my mind is really... should we? Or are we
(prematurely) over-complicating everything?

>
> > And following that back, I see minPts_ is configurable (which is a good
> > thing :D), so it's up to the use case to determine what the minimum
> > requirements are.
> >
> > Seems pretty good to me.
> >
> > > > > +               /* These are the standard equations for least squares linear regression. */
> > > > > +               slope_ = (count1 * count1 * xyAve_ - count1 * xAve_ * count1 * yAve_) /
> > > > > +                        (count1 * count1 * x2Ave_ - count1 * xAve_ * count1 * xAve_);
> > > > > +               offset_ = yAve_ - slope_ * xAve_;
> > > > > +       }
> > > > > +
> > > > > +       /* Don't increase count_ above numPts_, as this controls the long-term amount of the residual fit. */
> > > > > +       if (count1 < numPts_)
> > > > > +               count_++;
> > > > > +}
> > > > > +
> > > > > +/**
> > > > > + * \brief Calculate the output clock value according to the model from an input
> > > > > + * clock value
>
> Missing \param

Yep.

Thanks!
David

>
> > > > > + *
> > > > > + * \return Output clock value
> > > > > + */
> > > > > +uint64_t ClockRecovery::getOutput(uint64_t input)
> > > > > +{
> > > > > +       double x = static_cast<int64_t>(input - inputBase_);
> > > > > +       double y = slope_ * x + offset_;
> > > > > +       uint64_t output = y + x + outputBase_;
> > > > > +
> > > > > +       LOG(ClockRec, Debug) << "getOutput " << input << " " << output;
> > > > > +
> > > > > +       return output;
> > > > > +}
> > > > > +
> > > > > +} /* namespace libcamera */
> > > > > diff --git a/src/libcamera/meson.build b/src/libcamera/meson.build
> > > > > index 57fde8a8..4eaa1c8e 100644
> > > > > --- a/src/libcamera/meson.build
> > > > > +++ b/src/libcamera/meson.build
> > > > > @@ -21,6 +21,7 @@ libcamera_internal_sources = files([
> > > > >      'byte_stream_buffer.cpp',
> > > > >      'camera_controls.cpp',
> > > > >      'camera_lens.cpp',
> > > > > +    'clock_recovery.cpp',
> > > > >      'control_serializer.cpp',
> > > > >      'control_validator.cpp',
> > > > >      'converter.cpp',
>
> --
> Regards,
>
> Laurent Pinchart

Patch
diff mbox series

diff --git a/include/libcamera/internal/clock_recovery.h b/include/libcamera/internal/clock_recovery.h
new file mode 100644
index 00000000..c874574e
--- /dev/null
+++ b/include/libcamera/internal/clock_recovery.h
@@ -0,0 +1,72 @@ 
+/* SPDX-License-Identifier: BSD-2-Clause */
+/*
+ * Copyright (C) 2024, Raspberry Pi Ltd
+ *
+ * Camera recovery algorithm
+ */
+#pragma once
+
+#include <stdint.h>
+
+namespace libcamera {
+
+class ClockRecovery
+{
+public:
+	ClockRecovery();
+
+	/* Set configuration parameters. */
+	void configure(unsigned int numPts = 100, unsigned int maxJitter = 2000, unsigned int minPts = 10,
+		       unsigned int errorThreshold = 50000);
+	/* Erase all history and restart the fitting process. */
+	void reset();
+
+	/*
+	 * Add a new input clock / output clock sample, taking the input from the Linux
+	 * CLOCK_BOOTTIME and the output from the CLOCK_REALTIME.
+	 */
+	void addSample();
+	/*
+	 * Add a new input clock / output clock sample, specifying the clock times exactly. Use this
+	 * when you want to use clocks other than the ones described above.
+	 */
+	void addSample(uint64_t input, uint64_t output);
+	/* Calculate the output clock value for this input. */
+	uint64_t getOutput(uint64_t input);
+
+private:
+	unsigned int numPts_; /* how many samples contribute to the history */
+	unsigned int maxJitter_; /* smooth out any jitter larger than this immediately */
+	unsigned int minPts_; /* number of samples below which we treat clocks as 1:1 */
+	unsigned int errorThreshold_; /* reset everything when the error exceeds this */
+
+	unsigned int count_; /* how many samples seen (up to numPts_) */
+	uint64_t inputBase_; /* subtract this from all input values, just to make the numbers easier */
+	uint64_t outputBase_; /* as above, for the output */
+
+	uint64_t lastInput_; /* the previous input sample */
+	uint64_t lastOutput_; /* the previous output sample */
+
+	/*
+	 * We do a linear regression of y against x, where:
+	 * x is the value input - inputBase_, and
+	 * y is the value output - outputBase_ - x.
+	 * We additionally subtract x from y so that y "should" be zero, again making the numnbers easier.
+	 */
+	double xAve_; /* average x value seen so far */
+	double yAve_; /* average y value seen so far */
+	double x2Ave_; /* average x^2 value seen so far */
+	double xyAve_; /* average x*y value seen so far */
+
+	/*
+	 * Once we've seen more than minPts_ samples, we recalculate the slope and offset according
+	 * to the linear regression normal equations.
+	 */
+	double slope_; /* latest slope value */
+	double offset_; /* latest offset value */
+
+	/* We use this cumulative error to monitor spontaneous system clock updates. */
+	double error_;
+};
+
+} /* namespace libcamera */
diff --git a/include/libcamera/internal/meson.build b/include/libcamera/internal/meson.build
index 7d6aa8b7..41500636 100644
--- a/include/libcamera/internal/meson.build
+++ b/include/libcamera/internal/meson.build
@@ -11,6 +11,7 @@  libcamera_internal_headers = files([
     'camera_manager.h',
     'camera_sensor.h',
     'camera_sensor_properties.h',
+    'clock_recovery.h',
     'control_serializer.h',
     'control_validator.h',
     'converter.h',
diff --git a/src/libcamera/clock_recovery.cpp b/src/libcamera/clock_recovery.cpp
new file mode 100644
index 00000000..966599ee
--- /dev/null
+++ b/src/libcamera/clock_recovery.cpp
@@ -0,0 +1,207 @@ 
+/* SPDX-License-Identifier: BSD-2-Clause */
+/*
+ * Copyright (C) 2024, Raspberry Pi Ltd
+ *
+ * Clock recovery algorithm
+ */
+
+#include "libcamera/internal/clock_recovery.h"
+
+#include <time.h>
+
+#include <libcamera/base/log.h>
+
+/**
+ * \file clock_recovery.h
+ * \brief Clock recovery - deriving one clock from another independent clock
+ */
+
+namespace libcamera {
+
+LOG_DEFINE_CATEGORY(ClockRec)
+
+/**
+ * \class ClockRecovery
+ * \brief Recover an output clock from an input clock
+ *
+ * The ClockRecovery class derives an output clock from an input clock,
+ * modelling the output clock as being linearly related to the input clock.
+ * For example, we may use it to derive wall clock timestamps from timestamps
+ * measured by the internal system clock which counts local time since boot.
+ *
+ * When pairs of corresponding input and output timestamps are available,
+ * they should be submitted to the model with addSample(). The model will
+ * update, and output clock values for known input clock values can be
+ * obtained using getOutput().
+ *
+ * As a convenience, if the input clock is indeed the time since boot, and the
+ * output clock represents a real wallclock time, then addSample() can be
+ * called with no arguments, and a pair of timestamps will be captured at
+ * that moment.
+ *
+ * The configure() function accepts some configuration parameters to control
+ * the linear fitting process.
+ */
+
+/**
+ * \brief Construct a ClockRecovery
+ */
+ClockRecovery::ClockRecovery()
+{
+	configure();
+	reset();
+}
+
+/**
+ * \brief Set configuration parameters
+ * \param[in] numPts The approximate duration for which the state of the model
+ * is persistent, measured in samples
+ * \param[in] maxJitter New output samples are clamped to no more than this
+ * amount of jitter, to prevent sudden swings from having a large effect
+ * \param[in] minPts The fitted clock model is not used to generate outputs
+ * until this many samples have been received
+ * \param[in] errorThreshold If the accumulated differences between input and
+ * output clocks reaches this amount over a few frames, the model is reset
+ */
+void ClockRecovery::configure(unsigned int numPts, unsigned int maxJitter, unsigned int minPts,
+			      unsigned int errorThreshold)
+{
+	LOG(ClockRec, Debug)
+		<< "configure " << numPts << " " << maxJitter << " " << minPts << " " << errorThreshold;
+
+	numPts_ = numPts;
+	maxJitter_ = maxJitter;
+	minPts_ = minPts;
+	errorThreshold_ = errorThreshold;
+}
+
+/**
+ * \brief Reset the clock recovery model and start again from scratch
+ */
+void ClockRecovery::reset()
+{
+	LOG(ClockRec, Debug) << "reset";
+
+	lastInput_ = 0;
+	lastOutput_ = 0;
+	xAve_ = 0;
+	yAve_ = 0;
+	x2Ave_ = 0;
+	xyAve_ = 0;
+	count_ = 0;
+	slope_ = 0.0;
+	offset_ = 0.0;
+	error_ = 0.0;
+}
+
+/**
+ * \brief Add a sample point to the clock recovery model, for recovering a wall
+ * clock value from the internal system time since boot
+ *
+ * This is a convenience function to make it easy to derive a wall clock value
+ * (using the Linux CLOCK_REALTIME) from the time since the system started
+ * (measured by CLOCK_BOOTTIME).
+ */
+void ClockRecovery::addSample()
+{
+	LOG(ClockRec, Debug) << "addSample";
+
+	struct timespec bootTime;
+	struct timespec wallTime;
+
+	/* Get boot and wall clocks in microseconds. */
+	clock_gettime(CLOCK_BOOTTIME, &bootTime);
+	clock_gettime(CLOCK_REALTIME, &wallTime);
+	uint64_t boot = bootTime.tv_sec * 1000000ULL + bootTime.tv_nsec / 1000;
+	uint64_t wall = wallTime.tv_sec * 1000000ULL + wallTime.tv_nsec / 1000;
+
+	addSample(boot, wall);
+}
+
+/**
+ * \brief Add a sample point to the clock recovery model, specifying the exact
+ * input and output clock values
+ *
+ * This function should be used for corresponding clocks other than the Linux
+ * BOOTTIME and REALTIME clocks.
+ */
+void ClockRecovery::addSample(uint64_t input, uint64_t output)
+{
+	LOG(ClockRec, Debug) << "addSample " << input << " " << output;
+
+	if (count_ == 0) {
+		inputBase_ = input;
+		outputBase_ = output;
+	}
+
+	/*
+	 * We keep an eye on cumulative drift over the last several frames. If this exceeds a
+	 * threshold, then probably the system clock has been updated and we're going to have to
+	 * reset everything and start over.
+	 */
+	if (lastOutput_) {
+		int64_t inputDiff = getOutput(input) - getOutput(lastInput_);
+		int64_t outputDiff = output - lastOutput_;
+		error_ = error_ * 0.95 + (outputDiff - inputDiff);
+		if (std::abs(error_) > errorThreshold_) {
+			reset();
+			inputBase_ = input;
+			outputBase_ = output;
+		}
+	}
+	lastInput_ = input;
+	lastOutput_ = output;
+
+	/*
+	 * Never let the new output value be more than maxJitter_ away from what we would have expected.
+	 * This is just to reduce the effect of sudden large delays in the measured output.
+	 */
+	uint64_t expectedOutput = getOutput(input);
+	output = std::clamp(output, expectedOutput - maxJitter_, expectedOutput + maxJitter_);
+
+	/*
+	 * We use x, y, x^2 and x*y sums to calculate the best fit line. Here we update them by
+	 * pretending we have count_ samples at the previous fit, and now one new one. Gradually
+	 * the effect of the older values gets lost. This is a very simple way of updating the
+	 * fit (there are much more complicated ones!), but it works well enough. Using averages
+	 * instead of sums makes the relative effect of old values and the new sample clearer.
+	 */
+	double x = static_cast<int64_t>(input - inputBase_);
+	double y = static_cast<int64_t>(output - outputBase_) - x;
+	unsigned int count1 = count_ + 1;
+	xAve_ = (count_ * xAve_ + x) / count1;
+	yAve_ = (count_ * yAve_ + y) / count1;
+	x2Ave_ = (count_ * x2Ave_ + x * x) / count1;
+	xyAve_ = (count_ * xyAve_ + x * y) / count1;
+
+	/* Don't update slope and offset until we've seen "enough" sample points. */
+	if (count_ > minPts_) {
+		/* These are the standard equations for least squares linear regression. */
+		slope_ = (count1 * count1 * xyAve_ - count1 * xAve_ * count1 * yAve_) /
+			 (count1 * count1 * x2Ave_ - count1 * xAve_ * count1 * xAve_);
+		offset_ = yAve_ - slope_ * xAve_;
+	}
+
+	/* Don't increase count_ above numPts_, as this controls the long-term amount of the residual fit. */
+	if (count1 < numPts_)
+		count_++;
+}
+
+/**
+ * \brief Calculate the output clock value according to the model from an input
+ * clock value
+ *
+ * \return Output clock value
+ */
+uint64_t ClockRecovery::getOutput(uint64_t input)
+{
+	double x = static_cast<int64_t>(input - inputBase_);
+	double y = slope_ * x + offset_;
+	uint64_t output = y + x + outputBase_;
+
+	LOG(ClockRec, Debug) << "getOutput " << input << " " << output;
+
+	return output;
+}
+
+} /* namespace libcamera */
diff --git a/src/libcamera/meson.build b/src/libcamera/meson.build
index 57fde8a8..4eaa1c8e 100644
--- a/src/libcamera/meson.build
+++ b/src/libcamera/meson.build
@@ -21,6 +21,7 @@  libcamera_internal_sources = files([
     'byte_stream_buffer.cpp',
     'camera_controls.cpp',
     'camera_lens.cpp',
+    'clock_recovery.cpp',
     'control_serializer.cpp',
     'control_validator.cpp',
     'converter.cpp',