GNURadio on Zynq: radiomodem and fast ADC sources, processing on the PL

I wish to inform the GNURadio community on some of the latest results on
the use of the Xilinx Zynq platform
for real time radiofrequency signal processing aimed at making the best
use of the PS/PL architecture.
Let me emphasize that although I am writing this email, all the
technical work has been achieved by Guillaume Bres-Saix,
in Cc of this email, who has been working full time on this topic since
the beginning of the year, with the help
of Gwenhael Goavec-Merou (also in Cc) who attended FOSDEM this year with
me.

As discussed during the FOSDEM meeting in Brussels last february, we are
interested in interfacing custom
hardware for radiofrequency signal processing, in which the Zynq
platform seems most appropriate. The two
targeted hardware interfaces are an RF frontend (Semtech SX1255,
operating in the 400-500 MHz range, providing
separate frequency sources for the output and input signals), and fast
A/D converters (mostly for RADAR
applications). In this context, I believe that sending the datastream to
the CPU before using the FPGA as
a coprocessor is inefficient, and I consider and architecture made of
two separate processing blocks, one
including the data source and fast pre-processing steps using basic
functions such as decimation, FIR or
auto/cross-correlation to be implemented in the PL, followed by more
complex processing steps best handled by the
CPU (PS), in our case WFM demodulation and graphical display (through
the exported X client). In such an
architeture, only one data transfer occurs between PL and PS, namely the
pre-processed data leave the
FPGA memory when fetched by the linux kernel before being transfered to
userspace when the GNURadio source
scheduler requests data. At the moment this data transfer is performed
in polling mode, although DMA
transfer has been demonstrated in another context.

Guillaume has documented these applications: I believe the most
impressive result is shown in the video at
http://guillaumebres.wix.com/eportfolio#!zedboard1/c171q and
- YouTube
In this example, the I/Q stream from the RF frontend is decimated,
transfered to the GNURadio source and
processed by displaying the Fourier transform of the recorded signal,
demonstrating that the data
transfer is sound and the data processing is consistent with
expectations. The modulated signal source
is a sine wave modulating a FM-modulated radiomodem, and the Fourier
transform indeed exhibits the
sidebands at a frequency offset from the carrier equal to the emitted
sine wave frequency. The continuity
of the datastream and no loss of data is demonstrated in the second half
of the video in which an audio
signal modulates the radiomodem, and the WFM demodulation block of
GNURadio feeding the sound card output
generates a proper continuous signal, hence demonstrating that no
sample has been lost in the transfer process.

The second demonstration of a radiofrequency source made of fast A/D
converters is presented at
http://guillaumebres.wix.com/eportfolio#!zynqadc/c3pa
and its application to RADAR measurements is illustrated at
http://guillaumebres.wix.com/eportfolio#!zynqradar/coj3

I believe these are fascinating demonstrations of the flexibility of the
Zynq based SDR platform with
minimal money investment but huge technical and scientific opportunities
with the close ties of the PS/PL
architecture. Beyond these technical demonstrations, one issue we are
now facing is the integration of
the PL in the GNURadio(-companion) framework. I believe the solution is
easier since I have decided to
keep all the FPGA part in the initial (fast) processing steps and high
level software processing once
the datastream has been decimated. At the moment, one /dev/ entry is
needed for each functionality
in the FPGA so a GNURadio-companion block allows for the reconfiguration
of the associated functionality.
How the gnuradio processing stream is kept consistent with the bitstream
configuring the FPGA remains
an open question and at the moment the developer must provide this
consistency.

All this work must still be properly documented and shared with the
community with a properly accessible
repository: if interested in getting preliminary access to these source
codes while all the files
are being organized, let me know by email and we will try to make a
usable archive.

JM


JM Friedt, FEMTO-ST Time & Frequency/SENSeOR, 32 av. observatoire, 25044
Besancon, France