Re: [PATCH 2/2] Input: add motion-tracking ABS_* bits and docs
From: Roderick Colenbrander
Date: Thu Sep 29 2016 - 17:46:38 EST
On Thu, Sep 29, 2016 at 1:55 AM, <jic23@xxxxxxxxxxxxxxxxxxxxx> wrote:
> On 28.09.2016 18:39, Dmitry Torokhov wrote:
>> On Tue, Sep 27, 2016 at 4:38 PM, Roderick Colenbrander
>> <roderick@xxxxxxxxxx> wrote:
>>> From: Roderick Colenbrander <roderick.colenbrander@xxxxxxxx>
>>> This patch introduces new axes for acceleration and angular velocity.
>>> David Herrmann's work served as a base, but we extended the specification
>>> with various changes inspired by real devices and challenges we see
>>> when doing motion tracking.
>>> - Changed unit of acceleration to G instead of m/s^2. We felt that m/s^2
>>> is not the appropriate unit to return, because accelerometers are most
>>> often calibrated based on gravity.
> I'd so like to believe they are referenced to a standard g rather than
> whatever is true at the plant. Almost no consumer parts are able to
> adjust their calibrations (or at least it's undocumented if they are).
Main motivation for g is that it is often easy to calibrate for.
Typically there are registers with device specific constants for
calibrated 1g values. At least in our case we are able to leverage
>>> They return values in multiples of
>>> G and since we don't know the device location on earth, we should not
>>> blindly multiply by '9.8' for accuracy reasons. Such conversion is left
>>> to userspace.
> Hmm. If userspace is involved in trimming the readings, then the units don't
> Ah well, not important really. However, there is a question of generalised
> interfaces. There are plenty of other devices (distance sensors etc) which
> in m. For IIO at least we have to support those as well so SI units are
> preferred. I remember having the same discussion long long ago!
I think it really depends on the use case if trimming is needed. For
various situations just 'g' is fine (it is better than raw values).
Just the use case to go to m/s^2 is not possible at all in kernel
(without passing parameters to the kernel module, which is undesired).
Doing real calibration in user space would be a big challenge I think
and not sure if it is easy to standardize. Some devices calibration is
just about a multiplication and an offset, but other devices are
non-linear or temperature dependent and needing frequent
recalibration. It would best I think to basic calibration in kernel.
>>> - Resolution field is used for acceleration and gyro to report precision.
>>> The previous spec, specified to map 1 unit to e.g. 0.001 deg/s or 0.001
>>> This is of course simpler for applications, but unit definition is a
>>> arbitrary. Previous axes definitions used the resolution field, which
>>> felt more consistent.
>>> - Added section on timestamps, which are important for accurate motion
>>> tracking purposes. The use of MSC_TIMESTAMP was recommended in this
>>> situation to get access to the hardware timestamp if available.
> The moment you are into doing motion tracking I'm really thinking shoving it
> through input makes little sense... This is one of the main things
> IIO is set up to do...
The devices we intend to support through this API are actual HID
devices. They have the timestamp and motion sensor values within the
same HID report as the other HID values come in (buttons, sticks, ..).
That's basically the background on why we thought to add this to the
spec, we noticed significant enough variation to the Linux timestamps.
>>> - Changed motion axes to be defined as a right-handed coordinate system.
>>> Due to this change the gyro vectors are now defined as
>>> The overall changes makes the definitions consistent with computer
>>> [PATCH 4/4] Input: add motion-tracking ABS_* bits and docs
>>> David Herrmann <dh.herrmann@xxxxxxxxx>
>>> Tue Dec 17 07:48:54 PST 2013
>>> Motion sensors are getting quite common in mobile devices. To avoid
>>> returning accelerometer data via ABS_X/Y/Z and irritating the Xorg
>>> mouse-driver, this adds separate ABS_* bits for that.
>> We have IIO for motions sensors that are not strictly human input
>> devices; I believe there is also IIO->input bridge where generic IIO
>> sensors could be mapped to input device if they are supposed to be
>> used as such in given product.
> Yeah, *looks guilty* I've been failing to actually submit the input
> bridge for a quite some time due to the open question of how the heck
> we describe the connectivity (device tree etc) after all the fuss
> over the hwmon bridge.
> Need to get this moving again - in principle it is very straight forward.
> The infrastructure is used for a few other purposes so is reasonably
> well tested.
> Basically there has been very push on this previously so good
> if someone wants to pick it up and push it forward.
> (basically I did it as a proof of concept years ago but didn't care
> that much as I don't have a screen on any of the boards I care about).
>>> This is needed if gaming devices want to report their normal data plus
>>> accelerometer/gyro data. Usually, ABS_X/Y are already used by analog
>>> sticks, so need separate definitions, anyway.
>> I am not sure if this direction is sustainable. We can't keep adding
>> more and more ABS axes every time we add another control to something
>> that is basically a composite device. What if you add another stick?
>> Magnetometer? Some other sensor?
>> I think the only reasonable way it to come up with a notion of
>> "composite" input device consisting of several event nodes and have
>> userspace "assemble" it all together.
> That sounds sensible if we are going to pipe this stuff out to userspace.
> If the actual need is to do fusion on the various sensors then I'd be
> inclined to pull it from an IIO device directly then use uinput to
> push the fused result back in so that it emerges from input as one
> would expect a magic 'this is pose' data type.
> There is obviously the open question of how to describe the various
> axis if they don't make sense to be passed directly to input.
> So there are valid points raised by this email.
>>> Signed-off-by: David Herrmann <dh.herrmann@xxxxxxxxx>
>>> Signed-off-by: Roderick Colenbrander <roderick.colenbrander@xxxxxxxx>
>>> Documentation/input/gamepad.txt | 9 +-
>>> Documentation/input/motion-tracking.txt | 176
>>> include/uapi/linux/input-event-codes.h | 7 ++
>>> 3 files changed, 190 insertions(+), 2 deletions(-)
>>> create mode 100644 Documentation/input/motion-tracking.txt
>>> diff --git a/Documentation/input/gamepad.txt
>>> index 3f6d8a5..ed13782 100644
>>> --- a/Documentation/input/gamepad.txt
>>> +++ b/Documentation/input/gamepad.txt
>>> @@ -57,6 +57,9 @@ Most gamepads have the following features:
>>> - Rumble
>>> Many devices provide force-feedback features. But are mostly just
>>> simple rumble motors.
>>> + - Motion-tracking
>>> + Gamepads may include motion-tracking sensors like accelerometers and
>>> + gyroscopes.
>>> 3. Detection
>>> @@ -138,8 +141,6 @@ Triggers:
>>> Upper trigger buttons are reported as BTN_TR or ABS_HAT1X (right) and
>>> or ABS_HAT1Y (left). Lower trigger buttons are reported as BTN_TR2 or
>>> ABS_HAT2X (right/ZR) and BTN_TL2 or ABS_HAT2Y (left/ZL).
>>> - If only one trigger-button combination is present (upper+lower), they
>>> - reported as "right" triggers (BTN_TR/ABS_HAT1X).
>>> (ABS trigger values start at 0, pressure is reported as positive
>>> @@ -155,5 +156,9 @@ Menu-Pad:
>>> Rumble is advertised as FF_RUMBLE.
>>> + Motion-tracking is defined in
>>> ./Documentation/input/motion-tracking.txt and
>>> + gamepads shall comply to the rules defined there.
>>> Written 2013 by David Herrmann <dh.herrmann@xxxxxxxxx>
>>> diff --git a/Documentation/input/motion-tracking.txt
>>> new file mode 100644
>>> index 0000000..d34a290
>>> --- /dev/null
>>> +++ b/Documentation/input/motion-tracking.txt
>>> @@ -0,0 +1,176 @@
>>> + Motion Tracking API
>>> +1. Intro
>>> +Motion tracking devices produce device motion events generated from an
>>> +accelerometer, gyroscope or compass. These data can be returned to
>>> +via input events. This document defines how these data are reported.
>>> +2. Devices
>>> +In this document, a "device" is one of:
>>> + - accelerometer
>>> + - gyroscope
>>> + - compass
>>> +These devices returned their information via different APIs in the past.
>>> +unify them and define a common API, a set of input evdev codes was
>>> created. Old
>>> +drivers might continue using their API, but developers are encouraged to
>>> +the input evdev API for new drivers.
> Agreed, it was a mess which is one of the reasons IIO came about. Another
> being that there are lots of usecases for these devices which have nothing
> do with user input, where the filtering etc that occurs there is
>>> +2.1 Axes
>>> +Movement data is usually returned as absolute data for the 3 axes of a
>>> +In this context, the three axes are defined in a right-handed coordinate
>>> +system as:
>>> + - X: Axis goes from the left to the right side of the device
>>> + - Y: Axis goes from the bottom to the top of the device
>>> + - Z: Axis goes from the back to the front of the device
>>> +The front of a device is the side faced to the user. For a mobile-phone
>>> +would be the screen. For devices without a screen, the top is usually
>>> +side with the most buttons on it.
> You need the means to describe a transformation from the chip to device
> It can be spectacularly non obvious. See the mounting matrix stuff we have
> in IIO. Doing the nasty floating point maths to fix this stuff is
> a usespace job!
The consumer devices I worked with typically do this transformation on
the HID device itself and give kind of transformed values to the
driver, though values are not calibrated. I guess this is different
for some other devices.
>>> + Example: Mobile-Phone
>>> + | TOP
>>> + |
>>> + |
>>> + | +---------------------------+
>>> + | |\ ________________________ \ .__
>>> + | \ \ \ \ \ |\
>>> + | \ \ \ __ \ \ \
>>> + | \ \ \ /| \ \ \__
>>> + | \ \ \ __/ \ \ |\
>>> + | \ \ \ /| \ \ \ (Y Axis)
>>> + | \ \ \ __/ (Z axis) \ \ \__
>>> + | \ \ \ /| \ \ |\
>>> + | LEFT \ \ \ / \ \ \
>>> + | \ \ \ FRONT \ \ \
>>> + | \ \ \ \ \
>>> + | \ \ \_______________________\ \
>>> + | \ \ ___ \
>>> + | /\ \ \__\ \
>>> + | __/ \ +---------------------------+
>>> + | /| \|___________________________|
>>> + | / BACK
>>> + | (X axis)
>>> + | ------->------->------->------->
>>> + |
>>> + |
>>> + | BOTTOM
>>> +Rotation-data is reported as counter-clockwise rotation on an axis when
>>> +from the top of the axis, as given by the right hand rule. For a given
>>> +the reported rotation would be:
>>> + ____
>>> + //|
>>> + // | (axis)
>>> + //
>>> + //
>>> + . // __
>>> + / // /\
>>> + | // |
>>> + \ // / (counter-clockwise rotation)
>>> + *.___.*
>>> + //
>>> + //
>>> +2.2 Calibration
>>> +Motion sensors are often highly sensitive and need precise calibration.
>>> +are advised to perform neutral-point calibration themselves or to
>>> implement a
>>> +state-machine to normalize input data automatically.
>>> +Kernel devices may perform their own calibration and/or normalization.
>>> +this is usually sparse and, if implemented, transparent to the user.
>>> +There is currently no way to feed calibration data into the kernel in a
>>> +way. Proposals welcome!
> There is in IIO and always has been.
>>> +2.3 Units
>>> +(NOTE: This section describes an experimental API. Currently, no device
>>> +to these rules so this might change in the future.)
>>> +Reported data shall be returned as:
>>> + - Acceleration: 1/(input_absinfo.resolution) G
>>> + - Rotation: 1/(input_absinfo.resolution) degree per second
>>> +Acceleration is reported in units of G as opposed to m/s^2, because
>>> +sensors internally work based on gravitation.
> No they don't. They work based on acceleration. Gravity just happens to
> acceleration. Please don't confuse this issue. Take a sensor calibrated in
> Cambridge UK (it's a bit of local joke round here and catches undergraduate
> physicists every year apparently) and carry it to almost anywhere else in
> world and you'll find that the value goes up.
>>> Since the conversion to m/s^2 is
>>> +location dependent, applications should either approximate the
>>> +factor as 9.8 m/s^2 or if more precision is desired obtain a scaling
>>> +by other means e.g. GPS.
>>> +However, for most devices the reported units are unknown (more
>>> precisely: no
>>> +one has the time to measure them and figure them out). Therefore,
>>> +shall use abs-minimum and abs-maximum to calculate relative data and use
>>> +instead. Devices which return wrong units may be fixed in the future to
>>> +to these rules.
> Most devices they are known for, it's a characteristic of the relevant chip
> typically factory calibrated (or based on design rules).
>>> +2.4 Timestamps
>>> +For motion tracking purposes the time delta between consecutive motion
>>> +is important for mathematical operations such as differentiation and
>>> +The time delta could be derived from the 'time' field in 'struct
>>> input_event' by
>>> +subtracting the time between consecutive events. However, this timestamp
>>> may not
>>> +provide enough accuracy depending on the use case, since it is based
>>> upon time of
>>> +processing within the input layer versus time of arrival in the kernel
>>> or the
>>> +time the hardware sent the data. There is often a small variable time
>>> +between these.
>>> +Optionally, hardware may provide a hardware timestamp produced at the
>>> time it
>>> +sampled the motion sensors. This timestamp is is exposed through
>>> +'MSC_TIMESTAMP' event, which provides timing information in
>>> +If available, MSC_TIMESTAMP is the recommended approach for calculation
>>> of time
> This is actually pretty unusual, but as you might expect we have supported
> in IIO from the start.
For raw sensors I agree this may not always be around. At least in the
consumer devices I have seen there is (and our devices have them as
>>> +3.1 Accelerometer
>>> +Accelerometers measure movement acceleration of devices. Any combination
>>> of the
>>> +three available axes can be used. Usually, all three are supported.
>>> +Data is provided as absolute acceleration. A positive integer defines
>>> +acceleration in the direction of an axis. A negative integer defines
>>> +acceleration in the opposite direction.
>>> +The evdev ABS codes used are:
>>> + - ABS_ACCEL_X: X axis
>>> + - ABS_ACCEL_Y: Y axis
>>> + - ABS_ACCEL_Z: Z axis
>>> +3.2 Gyroscope
>>> +A gyroscope measures rotational speed (*not* acceleration!). Any
>>> combination of
>>> +the three available axes can be used. Usually, all three are supported.
>>> +Data is provided as absolute speed. A positive integer defines the
>>> +speed in counter-clockwise order around a given axis when viewed from
>>> the top of
>>> +the axis. A negative integer defines it in clockwise order.
>>> +The evdev ABS codes used are:
>>> + - ABS_GYRO_X: X axis (also: Pitch)
>>> + - ABS_GYRO_Y: Y axis (also: Roll)
>>> + - ABS_GYRO_Z: Z axis (also: Azimuth/Yaw)
>>> +3.3 Compass
>>> +(NOTE: No compass device currently uses the evdev input subsystem. Thus,
>>> +API is only a proposal, it hasn't been implemented, yet.)
>>> +A compass measures the ambient magnetic field of the three defined axes.
>>> +makes the data self-contained and independent of the current device
>>> +Any combination of the three axes can be used. Usually all three are
>>> +otherwise, it's not really useful as a compass.
>>> +Proposed evdev ABS codes are:
>>> + - ABS_COMPASS_X: X axis
>>> + - ABS_COMPASS_Y: Y axis
>>> + - ABS_COMPASS_Z: Z axis
>>> + (c) 2013 David Herrmann <dh.herrmann at gmail.com>
>>> + (c) 2016 Roderick Colenbrander <roderick.colenbrander@xxxxxxxx>
>>> diff --git a/include/uapi/linux/input-event-codes.h
>>> index 7bf2a2e..0cacfe7 100644
>>> --- a/include/uapi/linux/input-event-codes.h
>>> +++ b/include/uapi/linux/input-event-codes.h
>>> @@ -763,6 +763,13 @@
>>> #define ABS_MAX 0x3f
>>> #define ABS_CNT (ABS_MAX+1)
>>> +#define ABS_GYRO_X 0x40 /* Gyroscope X axis */
>>> +#define ABS_GYRO_Y 0x41 /* Gyroscope Y axis */
>>> +#define ABS_GYRO_Z 0x42 /* Gyroscope Z axis */
>>> +#define ABS_ACCEL_X 0x43 /* Accelerometer X axis */
>>> +#define ABS_ACCEL_Y 0x44 /* Accelerometer Y axis */
>>> +#define ABS_ACCEL_Z 0x45 /* Accelerometer Z axis */
>>> * Due to API restrictions the legacy evdev API only supports ABS values
>>> up to
>>> * ABS_MAX/CNT. Use the extended *ABS2 ioctls to operate on the full
>>> range of