[Milkymist-devel] HDMI/DVI mixer extension for M1 - draft schematics

[Sébastien Bourdeauducq]

Hi,

I've been working on a M1 daughterboard that can turn it into a simple 
video mixer. It has two HDMI ports, one "crossfade" linear 
potentiometer, and one "fade to black" knob.

Draft schematics are here:
http://milkymist.org/vmixext-20121010.pdf

Connection with the M1
======================
Power and video streams are via the expansion port J21. The design is 
compatible with all M1 revisions, including the very first six R1 
prototypes. The two 5V pins that have been added in R2/R3 are not used, 
instead, this voltage is supplied by the HDMI source.
I/Os are severely lacking on this port, thus:
1) TMDS needs to be converted to a single-ended I/O standard.
2) The HDMI HPD signal cannot be controlled.
3) The connection of the potentiometers is to be done using the M1's 
memory card pins.

TMDS to LVCMOS conversion
=========================
Since there are no existing chips that can do this, we use a LVDS to 
LVCMOS converter instead and tweak the voltage levels to make the 
signals palatable to all devices involved.
TMDS requires a 50 ohm termination to 3.3V. It is not allowed to 
terminate it to another DC voltage as it might cause some "monitor 
detection" features to fail [1].
Since the encoding of the TMDS data makes the signal DC-balanced, we can 
use a high-impedance AC coupling after the 50 ohm termination. The DC 
bias to the 1.2V LVDS offset voltage is provided by a 8.2K/4.7K divider 
connected to the 3.3V supply.
TMDS has up to 600mV single-ended swing [2], higher than the 400mV 
maximum of LVDS. However, the MAX9121 chip can tolerate up to 1V. Using 
a series resistor to reduce the swing as suggested in [3] would not have 
been easily possible as the equivalent resistance of the divider network 
is too high: the input capacitance of the chip combined with the high 
series resistor value would quickly ruin the signal edges.
Now let's see if this works :)

Potentiometers
==============
Simple circuit here, the FPGA generates a rising edge on pin 4 and the 
charge times of the capacitors through R59+P1/R60+P2 is measured after 
thresholding by the Schmitt triggers U3/U4.

Limitations
===========
HDMI is characterized by massive and unpredictable clock skew, and 
recovering the data requires a phase detector [4]. Since the "hard" 
phase detector can only be used with differential signals, we'll need to 
reimplement it with 2X oversampling of the signal using ISERDES. This 
oversampling can occur at a maximum of 945MHz due to the slowness of the 
S6 I/O clocking, which will limit the pixel clock to 47.25MHz. A 
resolution of 800x600 60Hz 24bpp should remain possible, however.
Another limitation is that, still due to the lack of I/O, the HPD signal 
cannot be controlled. This can cause the video source to attempt DDC 
transfers while the gateware is not ready, and finally give up and 
disable the port. Connecting or enabling the video source only after the 
M1 is ready should work reliably, though.

Next steps
==========
This add-on board should be good for prototyping HDMI and a simple video 
mixer. Later, we may want to integrate those ports on the same board as 
the FPGA, and get rid of the I/O hacks (which along with the improved 
memory bandwidth in milkymist-ng should enable HD resolutions like 720p60).

Comments welcome.

Thanks,
Sébastien

[1] http://en.wikipedia.org/wiki/TMDS
[2] http://www.ddwg.org/lib/dvi_10.pdf p.36
[3] http://lekernel.net/diff_convert.pdf fig. 8
[4] 
http://www.xilinx.com/support/documentation/application_notes/xapp495_S6TMDS_Video_Interface.pdf