sofglk — Windows sound driver for the Dell Chromebook 3100 / 3100 2-in-1 (Gemini Lake)

Free, open source speaker audio for Dell Chromebook 3100 and Chromebook 3100 2-in-1 (board fleex, Celeron N4000/N4020) running Windows 10 or 11 on MrChromebox firmware. Also relevant to other Gemini Lake chromebooks with the same codecs.

If Windows says "No audio output device is installed" on a chromebook, or the speakers are silent with no audio device in Device Manager, this is the missing piece: the Intel AudioDSP (SOF / Sound Open Firmware) with a DA7219 headset codec and a MAX98357A speaker amplifier has no in-box Windows driver.

Written from open sources only: the Linux SOF driver for the hardware protocol, and CoolStar's BSD-licensed sklhdaudbus and codec drivers for everything around it. No part of it is derived from disassembling anyone's binary.

This is experimental kernel code that drives an amplifier connected to your speakers. It has already destroyed one speaker on the development machine — read "Damage warning" below before installing it.

Does this fix my machine?

your situation	this driver
Dell Chromebook 3100 or 3100 2-in-1, Windows installed, no sound	yes, speakers
"No audio output device is installed" in Sound settings	yes - there is no endpoint until a driver publishes one
Another Gemini Lake chromebook with DA7219 + MAX98357A	very likely, same topology
Gemini Lake chromebook with cs42l42 or rt5682 instead	needs its own .tplg and widget names - see docs/PROGRESS.md
Headphone jack or microphone	not yet
Apollo Lake, Braswell, Tiger Lake, AMD	no - different DSP or different firmware

You also need CoolStar's sklhdaudbus and codec drivers installed first; they are free and open source, and the install script checks for them.

What works

• Boots SOF firmware on the AudioDSP and talks IPC3 both ways
• Parses the real .tplg topology and takes the SSP clocking from it
• Publishes a WaveRT render endpoint, so ordinary applications play to it
• The Windows volume slider and mute drive the DSP's volume stage in hardware, not the software mixer. QueryHardwareSupport on the endpoint returns 0x3.
• A ceiling the slider cannot exceed, enforced three times over: in the topology handler, in the driver's setter, and by the firmware's own max_value

What does not

• Speakers only. No headphones, no microphone. SSP2 and DMIC are unwired.
• 48 kHz, 16 bit, stereo only. Windows resamples everything else, which works, but the driver refuses other formats outright.
• Mono in practice — and this is the hardware, not the driver. The MAX98357A is one mono amplifier; both speakers get the same signal, so there is no left / right separation to be had and volume is a single control.
• Hardcoded for GLK with DA7219 + MAX98357A. A chromebook with cs42l42 or rt5682 needs its own topology file and different widget names.
• No power management worth the name: D3 means a full firmware re-download.

If the sound stops after a while

Fixed in the current build — the full hunt is in docs/PROGRESS.md, phases 14 and 15. Two distinct bugs stacked on top of each other:

1. The mailbox was polled from a timer that could not keep up with the firmware's notification rate, so under CPU load the doorbell went unanswered and the IPC channel deadlocked. The driver now takes the interrupt through sklhdaudbus's RegisterInterrupt, acknowledged at DIRQL where nothing can starve it.
2. PCM_FREE was only sent when a stream closed, not when it stopped. Linux frees the PCM on every stop (pcm.c, reset_hw_params); this driver did not, so a pause/resume inside one long-lived stream — which is what changing videos in a browser is — sent PCM_PARAMS on top of a still-configured host component. Some time later the firmware's IPC task simply stopped: cores powered, no panic, no notifications, every later request a pure timeout. The stop path now runs TRIG_STOP → DMA stop → PCM_FREE, in Linux's order.

Health is readable without a debugger, under the device's hardware key:

value	healthy reading
IrqRegistered	1
MailboxIrq	climbing steadily while audio plays
DspStalls	0 — increments when the DSP goes quiet mid-stream
IpcFailures	0 — increments when a request times out
FwPanicHipct	0 — non-zero is a firmware oops, with its code

Belt and braces, both in this repo:

• fix-audio.bat (run as administrator) restarts the device and brings the sound back in seconds, without rebooting.
• sofglk-watchdog.ps1 — register it as a SYSTEM scheduled task and forget it: it watches IpcFailures + DspStalls, restarts the device when they move, frees the browser's audio process if that blocks the restart, and re-enables a device found disabled. 60 s cooldown so it cannot loop.

Requirements

Hardware	Gemini Lake chromebook, DA7219 + MAX98357A
Firmware	MrChromebox (or anything that leaves Secure Boot off)
sklhdaudbus	installed first — it publishes the CSAUDIO\ADSP device
max98357a driver	installed — it owns the amplifier's SD_MODE pin
Test signing	on (bcdedit /set testsigning on, then reboot)

sklhdaudbus, max98357a and da7219 are at https://github.com/chrultrabook.

Test signing is needed because this driver is signed with a self-signed WDK certificate. Attestation signing, which is what CoolStar's drivers carry, needs an EV code signing certificate, which needs a registered company; Azure Trusted Signing does not cover kernel drivers. There is no free path to a signature that Windows accepts without test mode.

Install

powershell -ExecutionPolicy Bypass -File install.ps1

It checks the prerequisites, trusts the bundled test certificate, installs the driver, and prints what the driver reported: firmware state, topology counts, whether the graph came up, and the volume ceiling.

To remove it again:

powershell -ExecutionPolicy Bypass -File uninstall.ps1

Volume

Use the normal Windows volume slider — it controls the DSP directly.

The driver also holds a ceiling that the slider cannot pass, because this amplifier can drive more power than the speakers are rated for. Slider at maximum means "as loud as the ceiling allows". To move the ceiling:

powershell -File set-volume.ps1 -AttenDb 3

Lower numbers are louder; the default is 6, meaning 6 dB below full scale. Dell rates these speakers at 2 W average / 2.5 W peak and the MAX98357A can deliver about 3.2 W into 4 ohms, so 0 dB with continuous content is outside what the hardware is rated for. Going below 6 is your call.

Damage warning

The MAX98357A has no registers. It has an enable pin, and this driver decides when that pin is high. Get that wrong and the amplifier drives a constant level into the voice coil: the cone parks off centre, no sound comes out, and every watt becomes heat in one spot. That destroyed a speaker on the development machine before the cause was understood — the smell was the first symptom, and by then it was too late.

The driver now:

• enables the amplifier only while it can see audio actually moving through the DMA, and drops it within ~100 ms of silence — flowing zeros do not count
• enables it after the DMA is running and disables it before the DMA stops
• zero fills the buffer and lets it drain before stopping anything
• keeps enable/disable notifications at least 25 ms apart, because the codec driver raises its pin from a work item that sleeps 5 ms first and a Stop landing inside that window is silently undone (see max98357a-sdmode-race.md)
• answers Stop when a codec announces itself, so nothing is left enabled from a previous session
• kills the DSP clocks outright if the DMA position freezes while the amplifier is on — the amp mutes itself without a bit clock — and does the same from a bugcheck callback, so a crash cannot leave the speaker energised

None of that is a guarantee. If the speakers ever get warm, unplug them and say so.

Licensing

Driver source: BSD-3-Clause, same as the surrounding ecosystem.

Redistributed binaries, both from https://github.com/thesofproject/sof-bin:

file	what it is	licence
sofglk-fw.ri	sof-apl.ri from sof-bin v2.2 — Gemini Lake runs the Apollo Lake image, and sof-bin's sof-glk.ri is a symlink to it	BSD-3-Clause plus Intel's limited patent licence (LICENCE.Intel)
sof-glk-da7219.tplg	topology from sof-bin v2.2	same

The firmware is installed under our own name so that uninstalling this driver cannot disturb a sof-glk.ri that belongs to somebody else's package.

If it does not work

Everything the driver does is reported under the device's hardware key, readable without a debugger:

powershell -c "$k=(Get-ChildItem -Recurse 'HKLM:\SYSTEM\CurrentControlSet\Enum\CSAUDIO' | ? PSChildName -eq 'Device Parameters')[0].PSPath; Get-ItemProperty $k | Format-List"

value	healthy reading
FsrAfterDlState	5, FW_ENTERED; anything else means the firmware did not boot
TplgParseStatus	0, with TplgWidgets 60 / TplgRoutes 28 / TplgLinks 6 on this board
GraphReady	1 once the pipeline is built
SubdeviceStatus / TopologyStatus / ConnectionStatus	all 0 — the wave filter, the topology filter, and the link between them
*ReplyError	per-IPC firmware errors, all 0 on a healthy bring-up
AmpEnabled	the amplifier right now — should be 0 whenever nothing is playing
VolumeGain	current DSP gain, Qx.16; 65536 would be 0 dB, the ceiling is lower
ClocksKilled	1 means the stall watchdog fired; restart the device to play again
WaveRtMaxPos	how far the DMA got; should reach the buffer size
CsAudioListening	1 once the driver is receiving codec notifications too
FmtValidBits / FmtForce32Container	what the amplifier asked for: 16 valid bits in a 32-bit container