Killer Technologies

Wireless Connectivity ❙ PCI Express ❙ ATCA ❙ Configurable Computing ❙ Real-Time Java

System-Level Design Using

FPGAs and DSPs
An Example Showing Software-Defined Radio
DSPs have traditionally tackled signal processing applications. But the
complement of FPGA-based processing can add levels of performance
only possible with custom ASICs, along with configurability and cost
savings that far surpass what ASICs can offer.

by Dick Benson, The MathWorks

and Narinder Lall, Xilinx

ignal processing has traditionally hours instead of the months or even years phase shift across the audio frequency range

S been the realm of DSP processors

(for low sample rate applications)
and ASICs (for high sample rates).
required for many complex ASICs.

Software Defined Radios

(300-3000 Hz). Deviations from 90 degrees
resulted in less suppression in the unwanted
sideband signal. The filtering method, using
Recently the FPGA has emerged as an A software defined radio is a radio crystal and mechanical filters, eventually
alternative option for DSP designers. The made from reconfigurable/programmable won out. In 1956, Donald Weaver developed
FPGA is organized as an array of logic ele- general-purpose components. It is the pro- a third method that used filters and phase
ments and programmable routing resources gramming of these components that define shifts, although the filters were low-pass, not
used to provide the connectivity between its operational characteristics. For instance, band-pass, and the phase shifts were fixed
the logic elements, I/O pins and other bandwidth and modulation (ssb, cw, am, with respect to frequency. However, the
resources such as on-chip memory, digital fsk, psk, qpsk, etc.) are completely deter- technology was not quite up to the task, and
clock managers, embedded hardware mul- mined by how the reconfigurable parts are the filtering approach worked well, so the
tipliers, embedded microprocessors and programmed, not by hardware such as fil- new “Third Method” was not adopted.
multi-gigabit transceivers. A large number ters, mixers, amplifiers or other “tradi- The main problems of DC offset and
of transistors are used to support the pro- tional” components. Figure 1 illustrates the filter gain-phase match, which prevented the
grammable routing resources. architectural simplicity of this radio. idea from being adopted in 1956, simply do
As a result, an FPGA is a highly config- Although millions of transistors are not exist in today’s DSP implementation.
urable device that can be used to construct required, modern semiconductor manufac- Applying Weaver’s technique in DSP hard-
customized, and yet reconfigurable, data turing techniques make the cost per tran- ware is quite easy.
paths to solve the problem at hand. sistor less than that of a one turn wire loop Figure 2 shows a block diagram of a
Additionally, the significant combined serving as an inductor. The radio uses mul- simple SSB generator using Weaver’s
computational power of dedicated DSP tiple sampling rates for signal processing. scheme. To provide a recognizable shape to
resources such as 500+ multipliers, accu- The high-speed work is done in the FPGA the spectrum, a weighted sum of sine waves
mulators, distributed and block memory, while the audio bandwidth work is done in is used as the audio input. The first transla-
and shift-register logic allows the imple- the DSP—a rather perfect division of labor, tion and filtering operation centers the
mentation of highly parallel structures that with each signal processing component desired sideband around DC and removes
support computational throughput rates doing what makes the most sense. the other sideband. Then it is a simple matter
approaching that of today’s ASICs. Prior to 1956 there were two primary to up-convert this complex (I & Q) signal to
The FPGA approach to signal pro- methods for generating single sideband the desired RF frequency. The output of the
cessing provides a high-performance, (SSB) signals: phasing and filtering. Each final two mixers is either summed or differ-
highly flexible, miniature silicon foundry at had pros and cons. The phasing method enced to get the upper or lower sideband.
your disposal, with a turn-around time of required a circuit that generated a 90-degree The spectra shown in Figure 2 tell the story.
Reprinted from January 2004
 Killer Technologies
Configurable Computing

The previous simplified example shifted

the upper sideband to approximately 25 kHz.
The actual radio will need to go far higher, to
almost 30 MHz. To accomplish this, multi-
stage, multirate filtering is required. The
MathWorks filter design tools can simplify
this design job.
Naturally, to cover the 30 MHz HF spec-
trum, a sampling clock of greater than 2x30
MHz is needed. The SignalMaster onboard
64 MHz clock satisfies this requirement. For
voice communications, the desired audio
sampling rate is around 8000 Hz, and a dec-
imation factor of 64e6/8192 = 7812.5 Hz is a
good fit. Focusing on the receiver for a
moment, the first stage of filtering will be a
cascaded integrator-comb (CIC) filter since
Figure 1 The entire radio architecture can be captured in 12 functional it maps very well to Virtex-II FPGA hard-
blocks. The black signal paths indicate continuous time sig- ware. The remaining receiver filters will be
nals, while the colored paths (red, green, blue) are discrete FIR decimators or interpolators for the
time with red indicating the highest sampling rate of 64M transmit chain. The filter processing is as fol-
Samples/s. The FPGA is used for frequency translation, while lows: CIC (D=64), FIR_1 (D=8), FIR_2
the DSP takes care of modulation, demodulation, compres- (D=8), FIR_3 (D=2) for an overall sample
sion, AGC and other lower data rate functions.
rate reduction of 64*8*8*2=8192. The last
filter (D=2) determines the ultimate audio
frequency response characteristics. It is
trivial to change the filter characteristics
since they are all defined by software.
Once the basics have been laid out, more
detail can be added. The first step in this
process is to create a hardware-independent
model. This is can be accomplished using the
Simulink DSP Blockset. Figure 3 shows the
general signal flow in the radio. The local
oscillators are shared by both receive and
transmit processing chains. The receiver con-
sists of a digital downconverter followed by a
final filter-demodulator stage. The transmitter
is simply the receiver blocks turned around
with interpolating rather than decimating fil-
ters. Once satisfactory simulation results have
been obtained, indicating the general signal
flow and filtering are correct, it is time to split
the model between the FPGA and the DSP.
For this design, partitioning the functions is
straightforward.
Because the light green blocks in Figure 2
contain data rates that are all at or below
Figure 2 There are several methods that may be used to generate a 15.625K Samples/s, these sub-systems will be
single sideband signal. The method first proposed in 1956 by implemented in the DSP chip. The light red
Donald Weaver works nicely with today’s digital signal pro- blocks have the higher rate data running from
cessing technology. The above model shows the spectra 64M Samples/s down to the 15.625K
involved with the evolution of a single sideband signal: audio Samples/s rate and will therefore be imple-
input (left), translation and removal of one sideband (center), mented in the Xilinx FPGA. By the way, this
and frequency translation and conversion to a real signal con- is still the Weaver scheme, but we will
taining the desired sideband (right). describe it in modern terminology as digital

Reprinted from January 2004

 Killer Technologies
Configurable Computing

down / upconverters with SSB modulators and

demodulators.

FPGA Implementation
The Xilinx System Generator for DSP is an
extension of Simulink that provides design
entry, data path definition, bit- and cycle-true
simulations, test bench generation, hardware
co-simulation and VHDL code generation.
The tool’s block library maps to Xilinx DSP
LogiCores, which are optimized implementa-
tions of typical DSP functions such as filters,
direct digital synthesizer and FFT. Engineers
pick blocks from the library, define the fixed
point parameters (word size, binary point posi-
tion, rounding, saturation), and hook them
together like standard Simulink blocks.
Gateways are used to convert between the Figure 3 Detail is added to the design to create the transmitter and
standard Simulink and the Xilinx-specific data receiver processing, which shares a common set of local oscilla-
types used by the Xilinx blocks. tors. This top-level model captures the five main subsystems
As Figure 4 indicates, there are also blocks needed to create the radio. Each block contains several levels
for Lyr SignalMaster hardware, marked here as of hierarchy, getting down to fixed-point implementation details
“LSP.” The ADC (red) represents the 64M where required. The light red blocks handle the 64 M Samples/s
data rates and will therefore be implemented in the FPGA. The
Samples/s converter, while the DAC (red) is
green blocks handle data at both 15.625 and 7.8125K
the 64M Samples/s DAC. The gate_1 gateway
Samples/s, and will be implemented in the DSP chip.
block is a 32-bit register that the DSP can write
to change the frequency of the direct digital
synthesizer in the RF Local Oscillator block.
The down-converted IQ stream from the
Rx_Mix_Filters is fed to a 32-bit gateway that
interfaces to the TI DSP. On the transmit side,
another gateway (IQ from DSP) takes data
from the TI DSP and drives the IQ input of the
TX_Filters_Mix block. This diagram shows
the top level; further detail is contained within
each block.
Once the data paths and processing are
defined, simulation reveals whether the design
meets the objectives of dynamic range, and
whether spurious responses introduced by the
fixed-point implementation have been rejected.
If not, chances are that more bits must be used
in the filter coefficients and/or data paths. Once
the performance objectives are met, you can
generate the approximately 200 files of VHDL
required to implement the design.
It is then a matter of using the normal
Xilinx tool flow of synthesis, place and rout,
and bit-stream generation to program the Figure 4 The FPGA signal processing portion of the design is simulated to
FPGA. This process requires virtually no user bit- and cycle-true accuracy by using the Xilinx System Generator
intervention and bit-stream generation can be for DSP in The MathWorks Simulink environment. Data path size,
accomplished with a single mouse click. filter coefficient quantization, and all fixed-point attributes are
One of the more interesting reports that can defined using parameterized blocks. The LSP blocks represent
be generated is a “floor plan” of the FPGA gateways to and from the DSP chip (cyan) while the red block rep-
design. The digital frequency translator con- resents the high-speed ADC and DAC hardware on the Lyr
sumed virtually all (95%) of the FPGA. The SignalMaster development hardware.

Reprinted from January 2004

 Killer Technologies
Configurable Computing

fed to the multipliers, which mix the IQ

data with the audio quadrature oscillator.
The result of this is then fed to an adder.
This implements the SSB demodulator.
To control tuning frequency,
transmit/receive and sideband selection, a
simple user interface was made from
switches and sliders. The light bulb icon
activates LEDs on the SignalMaster during
transmit. A wire from the LED driver is the
electrical transmit/receive signal for the
external (analog) hardware.
The real-time code to implement the
example in Figure 5 can be generated auto-
matically from the Simulink environment
using the Real-Time Workshop product.
The process that generates the C code can be
modified to support a variety of target hard-
ware. A pre-packaged target exists for the
Figure 5 The top-level view of the partition executing in the DSP chip is Lyr SignalMaster, making it remarkably
shown above. The C code to implement the design is automatically easy to get up and running. Not a single line
generated by The MathWorks Real-Time Workshop option to the of C or VHDL code was manually written.
Simulink environment. Both the FPGA bitstream and DSP chip The block diagram now becomes the
binary image are downloaded to the Lyr SignalMaster with a single user interface to control the radio. The
mouse click. The block diagram then turns into a graphical debug- switches and sliders on the block diagram
ging tool where scopes, meters and numerical displays can show
change parameters (T/R, frequency,
the signals while the target DSP is executing the real-time code.
Sliders, switches and other parameters can be changed on-the-fly to
USB/LSB) on-the-fly while the code is
control the hardware. This design controls upper/lower sideband executing in the hardware. Beyond this,
selection, transmit/receive, tuning frequency (fine and coarse), parameters can also be changed while the
audio levels and more. code is running to adjust settings such as
reverb level, AGC time constant and signal
design did not initially fit in the xc2v1000 processing algorithms to be implemented limiting. Scopes and digital readouts can be
part that was on the SignalMaster. Bits on the DSP. used to monitor the signal levels in the
were trimmed and both rounding and satu- The Audio Processing block contains system under operating conditions. The
ration were abandoned in the transmit data the low-frequency ADC for the micro- system block diagram becomes a graphical
paths. The simulation results indicated that phone input, a reverberation subsystem, an tuning, debugging environment as well.
this would not cause any severe problems if audio compressor and the ability to gen-
the proper audio signal conditioning were erate a two-tone signal for test purposes.
handled in the DSP chip portion of the The resulting audio data stream is fed to the
design. The ability to easily make these transmitter (Tx) SSB Modulator subsystem
tradeoffs is a key attribute of the design that contains the audio frequency quadra-
flow and the rationale for using a program- ture local oscillator, interpolating filter
mable logic device. In this case, the design (I=2), and mixers to generate the transmit
fits on the available device, but if we IQ data stream. This data stream is then fed
wanted to add more functionality or space to the SMC6xx Interface block as a 32-bit
efficiency to the system, we could use this word (16 for I, 16 for Q), which drives the
analysis to shrink the design further. input of the FPGA digital upconverter.
On the receive side, the output of the www.mathworks.com
DSP Processor SMC6xx Interface block is fed to an IQ info@mathworks.com
Implementation demux that splits the 32-bit word into two 508.647.7000
Figure 5 shows the top level of the 16-bit words for I and Q. This complex 91192V00
radio partition, which is implemented with signal then goes though the D=2 deci-
the TI C6711 DSP chip on the Lyr mating stage. The filter output drives an
SignalMaster development hardware. AGC stage that is similar in nature to the
Like the preceding models, this model is audio compression stage. The amplitude-
hierarchical, and contains details of audio stabilized result of the AGC stage is then

Reprinted from January 2004