Asking the VideoCore to decode my JPEGs, from bare metal
Back to the Pi 1 hub · measured 2026-10-26

Situation
My face detector on the Pi 1 took frames from a USB webcam (an A4 Tech, behind the LAN9512 hub). In raw
YUYV at 352x288, 15 fps. The camera can also send MJPEG, so I asked myself whether this CPU has an
MJPEG decoder. It doesn’t. The ARM1176 has nothing. But the
VideoCore IV firmware does: its video_decode component lists MJPG. On Linux you reach it through
VCHIQ (a message queue in shared memory between the ARM and the GPU) and MMAL (the media API on top).
I had no Linux.
Task
Talk VCHIQ and MMAL from FreeRTOS, hand the firmware a JPEG, get I420 back. Then feed the face detector from it.
Where this fits in the whole thing: webcam MJPEG through the LAN9512 and my DWC2 host code, about 74 KB a frame, then the VideoCore video_decode over VCHIQ/MMAL (~5 ms) before the QPUs see a pixel.
Action
I took VCHIQ and MMAL from Linux’s drivers/staging/vc04_services and rewrote what I needed in bare
metal (vchiq.c, mmal.c, vcdec.c), keeping Linux’s six MMAL headers unchanged. Buffers are copied
(bulk transfers), not zero-copy. Simpler, and I can measure it later.
The firmware said no. First vc init:
pz> vc init
vchiq_init: -1, VCHIQ_INIT refused (answer not 0) (tag code 80000004, answer FFFFFFFF, VideoCore initialised 0)
Over and over. Two things had to change: slot zero needed its debug area filled in (1288 bytes, the size the firmware expects), and the board had to be rebooted between tries, because VCHIQ only opens once per GPU boot. After that the answer was 0 but no CONNECT came back, until I rang the doorbell every time instead of only when I thought it was needed.
MMAL. Once connected:
pz> vc mmal
MMAL version 16.1
ril.video_decode: handle 00000040, 1 input(s), 1 output(s)
First JPEG: nothing back. Bulk transfers aborted until I rounded the length up to a multiple of 4. Then the 19 534-byte test JPEG arrived on the GPU side (19 536 with padding) and… nothing came back. To see why, I raised the firmware’s own log levels (by poking its logging symbol from the ARM) and read its log. The decoder answers one frame late. Send the JPEG again and the first frame pops out:
decode: 152064 bytes of I420 from 19534 of JPEG; ...
Y mean 101.878, Y at 16 points: 94 122 124 48 102 126 194 155 53 39 131 206 159 71 197 96
video_decode:0:RIL: decoding 19534 bytes, 0us, flags 130, (cb:0x1f4c968c)
video_decode:0:RIL: finished passing imageJ
To check the pixels I decoded the same test JPEG on the PC with libjpeg: Y mean 101.635 there, 101.878 on the GPU. Close enough for two different decoders.
Then live. face live 352 288 15 mjpeg keeps whole JPEGs from the UVC stream (by SOI and length),
decodes each on the VideoCore and runs the detector on the I420. With me in front of the webcam it
found one face.
Result
MJPEG gets decoded on the VideoCore from bare metal: 19 534 B of JPEG -> 152 064 B of I420 in 5.3 ms in steady state, and the Y mean is 101.878 against libjpeg’s 101.635. Live on the webcam it did 597 frames with 0 decode errors, 5 ms a decode, and found one face (3.0 s a run, the detector was still on the CPU then).
The gotchas, so I don’t hit them twice: VCHIQ opens once per GPU boot, bulk lengths go in multiples of 4, and the decoder returns frame N when you send frame N+1.
Straight off the HDMI output, 27 s after boot: my floor through the webcam, and the panel says GPU VideoCore IV: giải MJPEG 5 ms, 80 KB/khung (decode 5 ms, 80 KB per frame). The 29 USB errors on it are the next story. (Vietnamese panel.)
What I’d do next
Two input and two output buffers instead of one each, so the decode overlaps the next USB frame. And try the zero-copy path: right now every JPEG gets copied once more than it needs to be.