You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
What we now know (checked against Sony's signed sixaxis.sys, eight physical pads, and the CircumSpector PS3 captures):
Accel is a single gain of 113 over (zero - oneG) from EEPROM page 0xA0. X is mirrored after calibration (0x3FF - cal), not on the raw value. Formula matches every live pose we took.
Gyro sign: raw falls for clockwise-from-above; every sixaxis.sys path inverts it, so clockwise is positive in Sony's reported value. Scale is about 1.4 counts per deg/s (~+/-360 deg/s full scale). The old "113 vs 226" question is closed (no 2x gain).
Hardware cal byte is 26.4 counts/step (Sony's constant, confirmed on an A2). Placement is [5]/[6] if Feature 0x01 lists field 0x07, else [3]/[4] on the SIXAXIS path.
Three gyro paths come from the Feature 0x01field list, not the type bytes. One pad reports DS3-looking 18 18 18 18 but takes the SIXAXIS path. Sending the factory cal byte once is not enough on hardware-trim pads (A2 idles ~160 counts off until the tracker steps).
Counterfeits return a plausible EEPROM template, so a zero == oneG check never fires. Detection has to be behavioural (frozen sensors / 920 Hz spam).
What DsHidMini does wrong today (SIXAXIS.SYS mode): gyro sign is backwards and unzeroed; X is mirrored on the raw value; no EEPROM read and no cal byte. DS4Windows mode still has no motion.
Still open: absolute gyro scale (needs a turntable, not more hand turns), whether pads without field 0x07 ignore the cal byte, a live SIXAXIS stream, 0xEF over Bluetooth, DS4 axis permutation.
Next step is the driver work in DsHid.c / USB start, not more reverse engineering. Happy to take more pad dumps if anyone has an odd revision.
Follow-up on live USB HW_CAL pads (CECHZC2E-A1 / DS3-A1b class) after #531. Both are in master.
Viewer pose (#533). Gravity pitch/roll were swapped against the USB/trigger edge, and integrated yaw was bound to the Helix roll axis, so residual gyro looked like a constant leftward roll. Pitch is USB/trigger, roll is the grips, yaw spins about the face normal. Recenter still only zeros yaw. The viewer also ignores rest rates below a few deg/s so EEPROM leftover does not creep the 3D pad.
HW_CAL publish path (#535).sixaxis.sys reports clamp(0x3FF - raw) and leaves residuals smaller than one cal-byte step (~26.4 counts) in the HID report. On the A1 that is a rock-steady ~16 counts / ~10.7 deg/s clockwise yaw — the tracker learns zeroRef but will not step 0x77. DsHidMini now publishes the tracker output clamp(512 + zeroRef - raw) on IPC and the 49-byte SIXAXIS GetFeature when the tracker is initialized, so that bias does not reach consumers. Failed/invalid Feature 0xEF still falls back to 0x3FF - raw. The cal-byte tracker still runs for large trim errors (A2-class).
Follow-up after live USB checks on an Obigben clone and a genuine 2E A1 (PLAIN_ZERO). Both are in master.
Clone rest drift (#538). Feature 0x01 clone-heuristic PLAIN_ZERO pads (Obigben, Fake DS3) seed Sony's gyro tracker in software-only mode: zeroRef follows rest, the cal byte is never stepped or sent. Genuine PLAIN_ZERO / HW_CAL / SIXAXIS paths are unchanged. The Obigben gyro is frozen (~500 vs template EEPROM zero 512), so this only removes the false ~8.6 deg/s clockwise rest; yaw will not follow turns. Needs a rebuilt driver.
Viewer on-grip pose (#539). Rolling a genuine pad onto a grip made the 3D model flip about yaw and left heading offset after returning flat. Display used Euler pitch from atan2(ay, -az) (undefined when gravity is along X) and integrated the face-normal gyro as heading while that axis was horizontal. The viewer now uses a shortest-arc tilt quaternion and weights yaw by how level the pad is (-u_z). Driver tracker unchanged.
A 2E A1 CSV showed no ZeroRef step on either grip. Rest GyroMilliDps is about -1 deg/s on this PLAIN_ZERO path: the deadzone hid it at rest, but it rode along during handling. The viewer now learns that rate on a clean flat rest, freezes it, and subtracts it only from samples that already pass the deadzone.
The Motion viewer Record button writes %ProgramData%\DsHidMini\Log\Motion\motion-slot<N>-<timestamp>.csv (one row per SampleIndex). Pause no longer mixes new sensor columns with a frozen pose.
ControlApp-only is enough for #539; #538 needs the new driver. Still deferred: Bluetooth 0xEF, DS4Windows motion fields, counterfeit freeze detection, turntable gyro scale.
What we got wrong in #547. The first Bluetooth pass issued HIDP Feature GET for 0x01 / 0xEF over BthPS3, reasoning from a file named ..._Bluetooth_Pairing.txt. That capture — like every other one cited in PS3_USB_STARTUP.md — is a USB link-layer pairing trace. There is no Bluetooth capture of those reads in this repo. Linux hid-sony, the USB Host Shield PS3BT library, and the OpenPuck emulator also never GET_REPORT Feature 0x01 / 0xEF over the air; the console only reads them on the cable while pairing. On a live genuine 2E A1 that path never loaded EEPROM (Fallback stayed set) and the raw gyro sat at a near-zero constant (~5). The tracker formula clamp(512 + zero - raw) turns that into a full-scale yaw spin regardless of calibration.
USB SET / GET Feature 0xEF page 0xA0 still runs as before, and now persists the 64-byte reply (plus identification) on the USB devnode. A failed or invalid read still keeps nominal zero=512, oneG=399 and clears any stale cache so Bluetooth cannot pick up a corrupt blob.
Bluetooth never queries the pad for 0x01 / 0xEF. It matches the pad MAC against USB instances (including phantom / unplugged) and loads the cached blobs. A pad never seen over USB stays nominal, with no extra startup delay.
ControlApp Motion viewer reports the source: live USB, cached from USB, or “connect this controller over USB once…”. Missing source properties no longer abort the viewer.
Gated experimental SET Feature 0xF4 on Bluetooth connect for pads with a cached, non-clone identification. Current payload is the USB value 42 0C 00 00; Linux / USB-Host-Shield's Bluetooth 42 03 00 00 is in the tree to compare. This is the working hypothesis for the raw-gyro-5 symptom.
Needs a rebuilt driver (ControlApp-only is not enough for the cache or 0xF4). Pair the pad over USB once after installing, then reconnect wirelessly.
Still deferred / pending hardware
Confirming a cache hit on a live pad (identification + EEPROM match the USB session, including with the USB instance unplugged) and which 0xF4 payload — 42 0C, 42 03, or neither — brings Bluetooth gyro off the near-zero constant
DS4Windows motion-field mapping (axis permutation still needs a reference DS4 capture)
Counterfeit freeze detection
Turntable gyro scale (still ~1.4 counts/(deg/s) from hand turns)
Extra Bluetooth dumps welcome, especially raw gyro before/after 0xF4 and a pad that has never been plugged in over USB on this machine.
Live check of the #549 cache path on a genuine 2E A1 (PLAIN_ZERO, firmware 04 00 08), driver 3.9.0.2278.
Result: works. USB once, then Bluetooth: identification and EEPROM load from the cache, MotionCalibrationSource reads CachedFromUsb, viewer shows factory calibration, and the wireless yaw spin from the #547 build is gone.
Bluetooth: key 4 identical to the USB copy, key 14 = 2 (cached from USB)
Decoded: accel X 498/386, Y 494/383, Z 496/387, gyro zero 481, cal byte 0. Matches the earlier USB session for this pad.
One trap worth knowing. First attempt looked like a cache miss: the connect over USB once hint stayed up after a USB round-trip and the pad kept rotating wirelessly. Both devnodes were missing keys 13/14 entirely - not even the None marker a failed read leaves - and the System log showed BthPS3 tearing down and re-creating the pad's PDO at the exact second the driver package was installed, while the pad was connected. The on-disk dshidmini.dll was 3.9.0, but WUDFHost was still running the previous image. A reboot fixed it with no code change. If you update the driver with a pad connected and Bluetooth still reports nominal calibration afterwards, reboot before filing a bug.
Still open
0xF4: the shipped 42 0C 00 00 went out on this connect and motion is correct, but I have not isolated whether it was required or whether 42 03 00 00 behaves the same. Needs an A/B on the same pad.
HW_CAL / SIXAXIS pads over Bluetooth (cal byte via the unified output payload) - untested wirelessly so far.
Motion is now exposed to applications through two public channels. All of it is in master via #608, #609, #610 and #611.
Write-up:docs/MOTION.md, "Exposing motion to applications". Both channels are fed from the same calibrated sample the SIXAXIS GetFeature report, the IPC snapshot and the ControlApp Motion viewer already use, so everything above (EEPROM read, gain-113 accel, gyro tracker, Bluetooth cache) applies unchanged.
Shipped
DS4Windows mode now carries motion (#608). Bytes 13-18 (gyro X/Y/Z), 19-24 (accel X/Y/Z) and the 16-bit timestamp at 10-11 (5.33 µs ticks) are filled from the calibrated sample with a real DS4's scale: 16 LSB per deg/s, 8192 LSB per g. Previously these were zero.
New HID mode CGS (#608, value 0x07). "Common Gamepad + Sensors": one HID interface with three top-level collections, so Windows creates a separate device path for each:
Collection
Usage page / usage
Report ID
Input
Feature
Gamepad (identical to CGP)
0x01 / 0x05
0x01
10 bytes
PID force feedback
Accelerometer 3D
0x20 / 0x73
0x30
state, event, X/Y/Z (0.001 g)
reporting state, power state, sensor state, per-axis sensitivity, interval
Gyrometer 3D
0x20 / 0x76
0x31
state, event, X/Y/Z (0.1 deg/s)
same layout
The sensor collections are standard HID Sensor (HUTRR39) collections, which means they are picked up by the in-box SensorsHid / SensorsCx stack and show up in Windows.Devices.Sensors (Accelerometer.GetDefault(), Gyrometer.GetDefault()) with no extra driver or SDK. HIDAPI can also just hid_read the sensor paths directly. DirectInput sees only the gamepad collection. Selectable in ControlApp like any other HID mode; needs a rebuilt driver.
Windows sensor contract (#609). Reporting state, power state and sensor state are proper HID Sensor selector arrays with the official enum values, the sensor event field is present, and the driver stops emitting a collection's reports when a client sets it to No Events or power state D4.
Crash fix: Windows.Devices.Sensors clients died on first open (#610). SensorExplorer and the WinRT API crashed on CGS pads. Root cause, from an ETW capture of the sensor stack plus disassembly of SensorsHid.dll / SensorsCx.dll: on the first IOCTL after open, SensorsCx calls _BaseSensor::GetDefaultIntervalAndThresholdFloat3D, which looks up PKEY_SensorData_AccelerationX/Y/Z_Gs (gyro: AngularVelocityX/Y/Z_DegreesPerSecond) in the driver's data-thresholds collection via PropKeyFindKeySetPropVariant. That API only updates existing keys and returns STATUS_NOT_FOUND otherwise. SensorsHid only creates those threshold entries from feature usages carrying the Change Sensitivity Absolute data-field modifier (0x1000 | data field, e.g. 0x1453 for Acceleration X); our descriptor only declared the sensor-wide property 0x030F, so the collection was empty, the open returned ERROR_NOT_FOUND and the WinRT layer null-dereferenced. The descriptors now declare per-axis sensitivity features (0x1453-55 accel, 0x1457-59 gyro) and the feature report grew to 14 bytes.
Same investigation turned up a unit bug: SensorsHid!GetDoubleFromUsageValue scales a raw value by 10^UnitExponent only and never reads the HID Unit item; the unit is whatever the Sensor usage table defines for that usage (g for acceleration, deg/s for angular velocity). Our 0.001 m/s² encoding would have read 9.8x too large. Acceleration is now encoded in milli-g (exponent -3).
Configurable coordinate frame (#611). New per-mode option CGS.MotionSensorFrame, also in ControlApp as "Motion sensor coordinate frame" (only visible in CGS mode):
"Gamepad" (default): the DS4 / SDL / GameInput frame, same as the DS4Windows report. X towards the right grip, Y out of the face, Z towards the player; an axis reads +1 g when it points up, so flat face-up is (0, +1, 0). Gyro follows the right-hand rule; the DS3 only has a yaw gyro, so X and Z rotation are always 0.
"Windows": the tablet screen frame Windows' own sensor tooling assumes. X right, Y towards the trigger edge, Z out of the face, accelerometer reports the gravity vector, so flat face-up is (0, 0, -1). Use this if SensorExplorer's bubble level or orientation demos should behave as they do on a Surface.
SensorExplorer on a genuine pad over USB in CGS mode: both sensors enumerate, no crash, streaming works. At rest on a table (gamepad frame) it reads about (-0.017, 0.902, 0.185) g. Y is the up axis as intended; the 0.185 on Z is the ~11° tilt a DS3 has when it rests on its grips, and the 0.92 g magnitude is this pad's own accelerometer response under Sony's gain-113 formula (the measured pads span 103-119 counts per g), not a conversion error. In the gamepad frame SensorExplorer's level dot sits at the top edge of its circle by design; the Windows frame centres it.
Still open
DS4Windows axis permutation and signs have not been checked against a real DS4 yet (the frame is derived from the measured DS3 orientation table and documented conventions).
GameInput reporting accelerometer/gyrometer capabilities for CGS, and whether a registry mapping is needed.
CGS over Bluetooth, and on HW_CAL / SIXAXIS pads.
Sensitivity and report interval are stored and returned but do not throttle output yet; a sensor report goes out per pad report (~100 Hz).
Unchanged from before: counterfeit freeze detection, turntable gyro scale, the Bluetooth 0xF4 A/B.
Needs a rebuilt driver for all of the above; ControlApp only for the frame selector UI. Reports from other pad revisions in CGS mode are welcome, especially anything that still fails to enumerate in SensorExplorer.
Render the Motion viewer with DirectX 11 #642 moves the preview to Helix DirectX 11, with WARP for software rendering. It also corrects the preview's pitch/roll mapping and keeps the camera Z-up.
These changes need an updated ControlApp. They do not change the driver calibration formulas or public motion report formats.
Current motion support
The previously described calibrated USB path, USB-to-Bluetooth calibration cache, SIXAXIS GetFeature motion, IPC snapshot, DS4Windows motion fields, and CGS HID sensor collections remain implemented. CGS still supports the configurable Gamepad/Windows coordinate frame. The earlier live results in this thread establish USB CGS streaming and Bluetooth cached calibration on the tested genuine pad; this inspection adds no new hardware verification.
Remaining work / verification
Check DS4Windows motion axes/signs against a real DS4 and measure absolute gyro scale with a known-rate reference (currently approximately 1.4 counts per deg/s).
Verify GameInput sensor capabilities for CGS, CGS over Bluetooth, and the wider HW_CAL / SIXAXIS hardware matrix.
Enforce CGS report interval and sensitivity: they are stored/returned, but reports still follow incoming pad samples. No Events / D4 gating is implemented.
Counterfeit frozen-sensor detection.
Bluetooth 0xF4 A/B testing (42 0C versus 42 03, and whether the write is required).
docs/MOTION.md remains the detailed reference, but parts of its older roadmap are stale: DS4Windows motion mapping is implemented (reference-hardware verification remains open), and the successful Bluetooth cache check is recorded in this thread.
This was a source/history inspection; no fresh build, automated test run, or live controller test was performed. “Merged in master” here does not establish availability in a published release.
Activity
nefarius commentedon May 29, 2025
With v3 a lot has changed, both driver and config app side. Attached a diff of the relevant changes as backup to be addressed in the v3 code base.
62a024d95e77ac37d3f360cb90422effc43a8035.diff.txt
nefarius commentedon May 29, 2025
Also thanks to the efforts of @lewy20041 we might be able to rekindle the lights on this issue with material from this great project: https://github.com/lewy20041/DS3_Input_And_Report_Inspector
mrwallace888 commentedon May 19, 2026
I'm going to assume this G-axis issue is the reason why Sackboy has scoliosis in LittleBigPlanet?
At least the rest of his motion controls work fine.
docs(motion): eight-pad matrix, measured yaw scale and cal-byte sensi…
docs(motion): persist probe, dumps and implementation roadmap (#217)
nefarius commentedon Sep 6, 2026
Research pass is far enough along to implement against. No driver changes yet; notes and dumps are on
feat/217-motion-research.Write-up: docs/MOTION.md (status + checklist at the bottom). Tools/dumps: research/ds3-motion/.
What we now know (checked against Sony's signed
sixaxis.sys, eight physical pads, and the CircumSpector PS3 captures):(zero - oneG)from EEPROM page0xA0. X is mirrored after calibration (0x3FF - cal), not on the raw value. Formula matches every live pose we took.sixaxis.syspath inverts it, so clockwise is positive in Sony's reported value. Scale is about 1.4 counts per deg/s (~+/-360 deg/s full scale). The old "113 vs 226" question is closed (no 2x gain).[5]/[6]if Feature0x01lists field0x07, else[3]/[4]on the SIXAXIS path.0x01field list, not the type bytes. One pad reports DS3-looking18 18 18 18but takes the SIXAXIS path. Sending the factory cal byte once is not enough on hardware-trim pads (A2 idles ~160 counts off until the tracker steps).zero == oneGcheck never fires. Detection has to be behavioural (frozen sensors / 920 Hz spam).What DsHidMini does wrong today (SIXAXIS.SYS mode): gyro sign is backwards and unzeroed; X is mirrored on the raw value; no EEPROM read and no cal byte. DS4Windows mode still has no motion.
Still open: absolute gyro scale (needs a turntable, not more hand turns), whether pads without field
0x07ignore the cal byte, a live SIXAXIS stream,0xEFover Bluetooth, DS4 axis permutation.Next step is the driver work in
DsHid.c/ USB start, not more reverse engineering. Happy to take more pad dumps if anyone has an odd revision.docs(motion): correct cal-byte emit, offsets and dump verdicts (#217)
docs(motion): place SIXAXIS cal bytes from the field list (#217)
docs: DS3/SIXAXIS motion sensor research notes (nefarius#217)
docs(motion): genuine DS3 hardware dumps and orientation table (nefar…
docs(motion): verify DS3 motion against Sony's sixaxis.sys (nefarius#217
2 remaining items
nefarius commentedon Sep 10, 2026
Not bad!
GzT8extrEB.mp4
nefarius commentedon Sep 11, 2026
Follow-up on live USB
HW_CALpads (CECHZC2E-A1 / DS3-A1b class) after #531. Both are inmaster.Viewer pose (#533). Gravity pitch/roll were swapped against the USB/trigger edge, and integrated yaw was bound to the Helix roll axis, so residual gyro looked like a constant leftward roll. Pitch is USB/trigger, roll is the grips, yaw spins about the face normal. Recenter still only zeros yaw. The viewer also ignores rest rates below a few deg/s so EEPROM leftover does not creep the 3D pad.
HW_CALpublish path (#535).sixaxis.sysreportsclamp(0x3FF - raw)and leaves residuals smaller than one cal-byte step (~26.4 counts) in the HID report. On the A1 that is a rock-steady ~16 counts / ~10.7 deg/s clockwise yaw — the tracker learnszeroRefbut will not step0x77. DsHidMini now publishes the tracker outputclamp(512 + zeroRef - raw)on IPC and the 49-byte SIXAXIS GetFeature when the tracker is initialized, so that bias does not reach consumers. Failed/invalid Feature0xEFstill falls back to0x3FF - raw. The cal-byte tracker still runs for large trim errors (A2-class).Write-up: docs/MOTION.md.
Needs a rebuilt driver for #535 (ControlApp-only is not enough). Extra
HW_CAL/PLAIN_ZEROUSB checks still welcome.nefarius commentedon Sep 12, 2026
Follow-up after live USB checks on an Obigben clone and a genuine 2E A1 (
PLAIN_ZERO). Both are inmaster.Clone rest drift (#538). Feature
0x01clone-heuristicPLAIN_ZEROpads (Obigben, Fake DS3) seed Sony's gyro tracker in software-only mode:zeroReffollows rest, the cal byte is never stepped or sent. GenuinePLAIN_ZERO/HW_CAL/SIXAXISpaths are unchanged. The Obigben gyro is frozen (~500 vs template EEPROM zero 512), so this only removes the false ~8.6 deg/s clockwise rest; yaw will not follow turns. Needs a rebuilt driver.Viewer on-grip pose (#539). Rolling a genuine pad onto a grip made the 3D model flip about yaw and left heading offset after returning flat. Display used Euler pitch from
atan2(ay, -az)(undefined when gravity is along X) and integrated the face-normal gyro as heading while that axis was horizontal. The viewer now uses a shortest-arc tilt quaternion and weights yaw by how level the pad is (-u_z). Driver tracker unchanged.A 2E A1 CSV showed no
ZeroRefstep on either grip. RestGyroMilliDpsis about -1 deg/s on thisPLAIN_ZEROpath: the deadzone hid it at rest, but it rode along during handling. The viewer now learns that rate on a clean flat rest, freezes it, and subtracts it only from samples that already pass the deadzone.The Motion viewer Record button writes
%ProgramData%\DsHidMini\Log\Motion\motion-slot<N>-<timestamp>.csv(one row perSampleIndex). Pause no longer mixes new sensor columns with a frozen pose.Write-up: docs/MOTION.md.
ControlApp-only is enough for #539; #538 needs the new driver. Still deferred: Bluetooth
0xEF, DS4Windows motion fields, counterfeit freeze detection, turntable gyro scale.nefarius commentedon Sep 17, 2026
Bluetooth motion follow-up is in
mastervia #549 (which replaces the wireless GET path from #547).Write-up: docs/MOTION.md (Bluetooth cache + remaining checklist). Why there is no wireless EEPROM read: docs/PS3_USB_STARTUP.md.
What we got wrong in #547. The first Bluetooth pass issued HIDP Feature GET for
0x01/0xEFover BthPS3, reasoning from a file named..._Bluetooth_Pairing.txt. That capture — like every other one cited inPS3_USB_STARTUP.md— is a USB link-layer pairing trace. There is no Bluetooth capture of those reads in this repo. Linuxhid-sony, the USB Host ShieldPS3BTlibrary, and theOpenPuckemulator also neverGET_REPORTFeature0x01/0xEFover the air; the console only reads them on the cable while pairing. On a live genuine 2E A1 that path never loaded EEPROM (Fallbackstayed set) and the raw gyro sat at a near-zero constant (~5). The tracker formulaclamp(512 + zero - raw)turns that into a full-scale yaw spin regardless of calibration.Shipped (#549)
SET/GET Feature 0xEFpage0xA0still runs as before, and now persists the 64-byte reply (plus identification) on the USB devnode. A failed or invalid read still keeps nominalzero=512,oneG=399and clears any stale cache so Bluetooth cannot pick up a corrupt blob.0x01/0xEF. It matches the pad MAC against USB instances (including phantom / unplugged) and loads the cached blobs. A pad never seen over USB stays nominal, with no extra startup delay.SET Feature 0xF4on Bluetooth connect for pads with a cached, non-clone identification. Current payload is the USB value42 0C 00 00; Linux / USB-Host-Shield's Bluetooth42 03 00 00is in the tree to compare. This is the working hypothesis for the raw-gyro-5 symptom.Needs a rebuilt driver (ControlApp-only is not enough for the cache or
0xF4). Pair the pad over USB once after installing, then reconnect wirelessly.Still deferred / pending hardware
0xF4payload —42 0C,42 03, or neither — brings Bluetooth gyro off the near-zero constantExtra Bluetooth dumps welcome, especially raw gyro before/after
0xF4and a pad that has never been plugged in over USB on this machine.nefarius commentedon Sep 17, 2026
Live check of the #549 cache path on a genuine 2E A1 (
PLAIN_ZERO, firmware04 00 08), driver 3.9.0.2278.Result: works. USB once, then Bluetooth: identification and EEPROM load from the cache,
MotionCalibrationSourcereadsCachedFromUsb, viewer shows factory calibration, and the wireless yaw spin from the #547 build is gone.Devnode properties after the run:
0xEFpage0xA0reply00 EF 04 00 08 03 01 A0 … 01 F2 01 82 | 01 EE 01 7F | 01 F0 01 83 | 01 E1 00 00, key 14 = 1 (live USB)Decoded: accel X 498/386, Y 494/383, Z 496/387, gyro zero 481, cal byte 0. Matches the earlier USB session for this pad.
One trap worth knowing. First attempt looked like a cache miss: the
connect over USB oncehint stayed up after a USB round-trip and the pad kept rotating wirelessly. Both devnodes were missing keys 13/14 entirely - not even theNonemarker a failed read leaves - and the System log showed BthPS3 tearing down and re-creating the pad's PDO at the exact second the driver package was installed, while the pad was connected. The on-diskdshidmini.dllwas 3.9.0, but WUDFHost was still running the previous image. A reboot fixed it with no code change. If you update the driver with a pad connected and Bluetooth still reports nominal calibration afterwards, reboot before filing a bug.Still open
0xF4: the shipped42 0C 00 00went out on this connect and motion is correct, but I have not isolated whether it was required or whether42 03 00 00behaves the same. Needs an A/B on the same pad.HW_CAL/SIXAXISpads over Bluetooth (cal byte via the unified output payload) - untested wirelessly so far.nefarius commentedon Oct 8, 2026
Motion is now exposed to applications through two public channels. All of it is in
mastervia #608, #609, #610 and #611.Write-up: docs/MOTION.md, "Exposing motion to applications". Both channels are fed from the same calibrated sample the SIXAXIS GetFeature report, the IPC snapshot and the ControlApp Motion viewer already use, so everything above (EEPROM read, gain-113 accel, gyro tracker, Bluetooth cache) applies unchanged.
Shipped
DS4Windows mode now carries motion (#608). Bytes 13-18 (gyro X/Y/Z), 19-24 (accel X/Y/Z) and the 16-bit timestamp at 10-11 (5.33 µs ticks) are filled from the calibrated sample with a real DS4's scale: 16 LSB per deg/s, 8192 LSB per g. Previously these were zero.
New HID mode
CGS(#608, value0x07). "Common Gamepad + Sensors": one HID interface with three top-level collections, so Windows creates a separate device path for each:0x01/0x050x010x20/0x730x300x20/0x760x31The sensor collections are standard HID Sensor (HUTRR39) collections, which means they are picked up by the in-box
SensorsHid/SensorsCxstack and show up inWindows.Devices.Sensors(Accelerometer.GetDefault(),Gyrometer.GetDefault()) with no extra driver or SDK. HIDAPI can also justhid_readthe sensor paths directly. DirectInput sees only the gamepad collection. Selectable in ControlApp like any other HID mode; needs a rebuilt driver.Windows sensor contract (#609). Reporting state, power state and sensor state are proper HID Sensor selector arrays with the official enum values, the sensor event field is present, and the driver stops emitting a collection's reports when a client sets it to No Events or power state D4.
Crash fix:
Windows.Devices.Sensorsclients died on first open (#610). SensorExplorer and the WinRT API crashed on CGS pads. Root cause, from an ETW capture of the sensor stack plus disassembly ofSensorsHid.dll/SensorsCx.dll: on the first IOCTL after open,SensorsCxcalls_BaseSensor::GetDefaultIntervalAndThresholdFloat3D, which looks upPKEY_SensorData_AccelerationX/Y/Z_Gs(gyro:AngularVelocityX/Y/Z_DegreesPerSecond) in the driver's data-thresholds collection viaPropKeyFindKeySetPropVariant. That API only updates existing keys and returnsSTATUS_NOT_FOUNDotherwise.SensorsHidonly creates those threshold entries from feature usages carrying the Change Sensitivity Absolute data-field modifier (0x1000 | data field, e.g.0x1453for Acceleration X); our descriptor only declared the sensor-wide property0x030F, so the collection was empty, the open returnedERROR_NOT_FOUNDand the WinRT layer null-dereferenced. The descriptors now declare per-axis sensitivity features (0x1453-55accel,0x1457-59gyro) and the feature report grew to 14 bytes.Same investigation turned up a unit bug:
SensorsHid!GetDoubleFromUsageValuescales a raw value by10^UnitExponentonly and never reads the HIDUnititem; the unit is whatever the Sensor usage table defines for that usage (g for acceleration, deg/s for angular velocity). Our 0.001 m/s² encoding would have read 9.8x too large. Acceleration is now encoded in milli-g (exponent -3).Configurable coordinate frame (#611). New per-mode option
CGS.MotionSensorFrame, also in ControlApp as "Motion sensor coordinate frame" (only visible in CGS mode):"Gamepad"(default): the DS4 / SDL / GameInput frame, same as the DS4Windows report. X towards the right grip, Y out of the face, Z towards the player; an axis reads +1 g when it points up, so flat face-up is(0, +1, 0). Gyro follows the right-hand rule; the DS3 only has a yaw gyro, so X and Z rotation are always 0."Windows": the tablet screen frame Windows' own sensor tooling assumes. X right, Y towards the trigger edge, Z out of the face, accelerometer reports the gravity vector, so flat face-up is(0, 0, -1). Use this if SensorExplorer's bubble level or orientation demos should behave as they do on a Surface.Live result
SensorExplorer on a genuine pad over USB in CGS mode: both sensors enumerate, no crash, streaming works. At rest on a table (gamepad frame) it reads about
(-0.017, 0.902, 0.185)g. Y is the up axis as intended; the 0.185 on Z is the ~11° tilt a DS3 has when it rests on its grips, and the 0.92 g magnitude is this pad's own accelerometer response under Sony's gain-113 formula (the measured pads span 103-119 counts per g), not a conversion error. In the gamepad frame SensorExplorer's level dot sits at the top edge of its circle by design; theWindowsframe centres it.Still open
HW_CAL/SIXAXISpads.0xF4A/B.Needs a rebuilt driver for all of the above; ControlApp only for the frame selector UI. Reports from other pad revisions in CGS mode are welcome, especially anything that still fails to enumerate in SensorExplorer.
nefarius commentedon Nov 12, 2026
Status update (2026-11-12), based on inspection of
masterat1bc05e1. The local checkout matches GitHub's branch head and has a clean working tree.Since the last update
The Motion viewer has received another round of improvements, all merged:
These changes need an updated ControlApp. They do not change the driver calibration formulas or public motion report formats.
Current motion support
The previously described calibrated USB path, USB-to-Bluetooth calibration cache, SIXAXIS GetFeature motion, IPC snapshot, DS4Windows motion fields, and CGS HID sensor collections remain implemented. CGS still supports the configurable Gamepad/Windows coordinate frame. The earlier live results in this thread establish USB CGS streaming and Bluetooth cached calibration on the tested genuine pad; this inspection adds no new hardware verification.
Remaining work / verification
HW_CAL/SIXAXIShardware matrix.0xF4A/B testing (42 0Cversus42 03, and whether the write is required).docs/MOTION.md remains the detailed reference, but parts of its older roadmap are stale: DS4Windows motion mapping is implemented (reference-hardware verification remains open), and the successful Bluetooth cache check is recorded in this thread.
This was a source/history inspection; no fresh build, automated test run, or live controller test was performed. “Merged in master” here does not establish availability in a published release.