gr-plasma: A New GNU Radio-based Tool for Software-defined Radar

Shane Flandermeyer SHANE . FLANDERMEYER @ OU . EDU

School of Electrical and Computer Engineering, Advanced Radar Research Center, University of Oklahoma
Rylee Mattingly RMATTINGLY @ OU . EDU
School of Electrical and Computer Engineering, Advanced Radar Research Center, University of Oklahoma
Justin Metcalf JMETCALF @ OU . EDU
School of Electrical and Computer Engineering, Advanced Radar Research Center, University of Oklahoma
Abstract t=0 t = TPRI t = 2TPRI t = MTPRI
A low-cost experimental setup is an invaluable
tool for rapid prototyping of radar signal pro-
PDU 1 PDU 2 PDU M
cessing algorithms. Here, the gr-plasma
out-of-tree module is presented as a convenient
way to collect radar data using UHD-compatible
software-defined radio (SDR) systems. The
module operates entirely over the GNU Ra- t
dio message passing interface, packetizing each
pulse received by the radar into a protocol data
unit (PDU) to simplify downstream processing Figure 1. Pulsed radar data structure
and minimize latency. It includes blocks for
waveform generation, monostatic transmission
features from previous projects to facilitate experimenta-
and reception, and data storage, and supports
tion by both academics and hobbyists while making it easy
both CPU and GPU backend processing. All
for developers to extend the module to other use cases.
signal processing functionality in gr-plasma
is an implementation of plasma dsp, a sepa- There are two key challenges with software-defined radar
rate companion library that can also be used for that differ from most communications protocols: trans-
projects outside the GNU Radio ecosystem. In mit/receive synchronization and ultra-reliable high-speed
this paper, the utility and structure of each block duplex operation. Radar systems operate by transmitting
is discussed, and the results of range-Doppler electromagnetic waves into the environment and measuring
processing are shown for data collected with gr- their reflections off of objects. To achieve localization ac-
plasma and an Ettus X310 in an open-air test. curacy, the transmit and receive processes must be synchro-
nized down to the sample level. The pulsed radar developed
for this work maintains precise knowledge of the beginning
1. Introduction and end of each pulse along with the beginning of the re-
ceive interval. Since the range resolution of a radar sen-
Building experimental radar systems has traditionally re- sor is inversely proportional to the bandwidth of the emit-
quired expensive custom hardware. Such designs are of- ted waveform, the radar module must also be capable of
ten difficult to program or modify, preventing easy exper- streaming data at high sample rates in order to maximize
imentation with new radar signal processing algorithms. performance.
While SDR hardware is not inherently designed to operate
as a radar, there have been numerous documented examples This work will demonstrate the utility of GNU Radio’s
of software-defined radar implementations on commercial- message passing interface in achieving these goals. Mes-
off-the-shelf systems, such as Christiansen & Smith (2019) sages are well-suited for bursty data, and they make it easy
and Wunsch (2014). This work aims to combine the best to determine the bounds of each radar pulse compared to
traditional stream buffers that are not large enough to pro-
cess entire pulses of data unless the pulse duration is short
Proceedings of the 1 st GNU Radio Conference, Copyright 2016 or the sample rate is sufficiently small. Each pulse rep-
by the author(s).
etition interval (PRI) of received radar data can then be
 gr-plasma: A New GNU Radio-based Tool for Software-defined Radar

packaged into a protocol data unit (PDUs) as shown in Fig.

1, which simplifies processing by providing a standardized
data format for downstream signal processing blocks. Mes-
sage passing also increases the maximum sample rate that
can be achieved for USRP devices (compared to the in-
tree UHD source and sink blocks) because timing and data
throughput limitations can be managed with no reliance on
the scheduler.
The remainder of this paper is outlined as follows: Sec-
tion 2 outlines the signal processing blocks that are im-
plemented in gr-plasma, along with select features in
plasma dsp that are useful for development. Section
3 describes an open-air testing setup used for collecting
radar data, including system specifications and required
software. Section 4 shows the results of this experiment
along with range-doppler maps that were generated during
data collection. Finally, Section 5 outlines the conclusions Figure 2. LFM waveform in the time domain
and presents future plans for the gr-plasma module.
Unlike traditional waveforms (such as the LFM) which are
2. Blocks and Features a function of only a few parameters, phase-coded wave-
forms are parameterized by the individual phase values in
2.1. Waveform Generators
the code, giving them many degrees of freedom that can be
gr-plasma currently implements two types of radar used for optimization. However, due to the instantaneous
waveforms. The first is a linear frequency-modulated transitions between phase values, the resulting waveform
waveform (LFM), which is the most common waveform will be distorted when it is passed through a power ampli-
used in pulsed radar systems due to its simplicity in imple- fier operating in the saturation region. PCFM addresses this
mentation and desirable pulse-Doppler ambiguity proper- shortcoming by applying a smoothing filter to the wave-
ties. As the name implies, an LFM “chirp” sweeps linearly form’s phase trajectory, resulting in an emission that is (ap-
across its frequency range over a finite time duration. In proximately) both continuous and constant modulus and
gr-plasma, the Linear FM block generates an LFM at thus amenable to transmission through a high-power am-
complex baseband (i.e., centered around 0 Hz) that sweeps plifier (Blunt et al., 2014a).
upwards from the lower frequency to the higher frequency,
or    
B 2
x(t) = exp jπ −Bt + t + θ0 (1)
Tp
where B is the bandwidth of the waveform, Tp is the du-
ration of the frequency sweep, and θ0 is an arbitrary start-
ing phase. The time-domain output of the LFM block is
shown in Fig. 2 for a waveform with B = 40 MHz and
Tp = 20 µs.
Figure 3. Forming a phase-coded waveform from individual
gr-plasma can also generate polyphase-coded FM phase values
(PCFM) waveforms, which were first introduced in (Blunt
et al., 2014a) and (Blunt et al., 2014b). PCFM wave-
forms are based on the continuous phase modulation The PCFM Source block in gr-plasma implements first-
(CPM) scheme from the communications literature, which order PCFM (Tan et al., 2015), which linearly interpo-
maps a sequence of data symbols into a nearly continu- lates to K intermediate values between each phase code
ous constant-amplitude waveform. Rather than modulat- value on the unit circle. Mathematically, the phase func-
ing symbols in a data stream, PCFM modulates discrete se- tion at complex baseband resulting from phase code θ =
quences of phase values known as phase codes (Levanon & [θ1 , θ2 , . . . , θNc ] can be described as
Mozeson, 2004). A waveform can be formed from a length- c −1
NX
Z t
Nc code by transmitting each code value for some amount ψ(t; θ) = wc (λ) ∗ δ(t − nTc )αn dλ + θ0 (2)
of time Tc , producing a waveform of length Tp (Fig. 3). 0 n=0
 gr-plasma: A New GNU Radio-based Tool for Software-defined Radar

where wc (t) is a unit-energy shaping filter and αi is instan- The delay is computed using a streamlined version of the
taneous frequency, which is the difference between succes- Estimator Sync Pulse block from gr-radar: with the Tx
sive phase code values (wrapped into the range [−π, π]) port connected directly to the Rx port (with a 30 dB atten-
with α0 = 0. In the PCFM source block, wc is rectan- uator in the loop), a known reference pulse is transmitted
gular with length K and Barker, Frank, and P4 codes are and then cross-correlated with the received data, and the
currently supported (Fig. 4). first peak is taken to be the delay. The calibration can be
performed for any number of sample rates. For example, a
2.2. USRP Interfacing and Control B210 may be calibrated for 10 MS/s and 20 MS/s with the
following command
Unlike in communications systems, a radar’s transmitter
and receiver are often co-located and managed by the same plasma calibrate delay --rates 10e6
host system (known as a monostatic configuration). There- 20e6 --filename $HOME/delay.json
fore, gr-plasma uses a single custom block to handle
” B210 ” : [
both the transmission and reception of radar pulses for
{
UHD-compatible SDR devices. Unlike the USRP Source
” delay ” : 46 ,
and Sink in the main tree, the USRP Radar block operates
” m a s t e r c l o c k r a t e ”: 40000000.0 ,
entirely in the message domain. The block operates as fol-
” samp rate ” : 10000000.0
lows: first, a waveform must be passed to the input port
},
as a PDU. The metadata dictionary in this PDU must con-
{
tain an item with a key called prf whose value defines the
” delay ” : 165 ,
pulse repetition frequency (PRF) at which the waveform is
” m a s t e r c l o c k r a t e ”: 20000000.0 ,
to be transmitted. The PRF tag can be added to any PDU
” samp rate ” : 20000000.0
using the Waveform Controller block. Although the Wave-
},
form Controller block is currently only used to set the PRF,
]
it will eventually be extended to facilitate cognitive opera-
tion in which it uses the results of downstream processing Listing 1. Delay calibration output file format
to update the transmitted waveform.
When the PDU has been processed, the radar block will The output of the calibration script is then saved to a JSON
repeatedly transmit and receive the waveform until a new file that stores a list of delays, master clock rates, and sam-
PDU is received. This method minimizes the number of ple rates for each device configuration that has been cal-
messages that must be processed by the block and makes it ibrated. For the example above, the JSON in Listing 1
possible to operate at higher sample rates since no overhead would be produced. The USRP radar block takes this file
is incurred from the scheduler. It also reduces latency dur- as an input, and skips the first Ndelay received samples if
ing processing compared to a stream-based approach since the configuration is present in the file.
messages can be processed immediately when they are re-
ceived, making it possible to update the waveform on a sub- 2.3. Signal Processing Blocks
millisecond timescale. Moreover, unlike in tagged stream
Matched filtering, also known as pulse compression in
blocks, the size of the waveform is not limited by the size of
radar and sonar applications, is a method for detecting a
each block’s buffer. Since radar returns are often processed
signal in the presence of comparatively high-power noise.
on a pulse-by-pulse basis, the received data is processed
The goal of matched filtering is to maximize the signal-to-
into pulses and passed to the output as a new PDU. Meta-
noise ratio (SNR) at the filter output for a given transmit-
data from upstream blocks is only propagated on the first
ted waveform. It can be shown that for a waveform xtx (t)
received pulse for which it applies in order to further min-
with duration τp in additive white Gaussian noise (AWGN),
imize the amount of data to be processed by downstream
the matched filter’s impulse response is given as (Richards,
blocks. The complete USRP processing chain is summa-
2013)
rized in Fig. 5.
One hardware quirk that had to be addressed before the
h(t) = αx∗tx (Tm − t) (3)
USRP radar could be used for open-air experiments was
the delay between transmit (Tx) and receive (Rx) ports due
to the internal DSP of each device, which manifests as a where scaling constant α has no impact on SNR and is of-
range offset for all targets during processing. This delay is ten set to unity, and Tm is the time at which the SNR is
constant for a given device, sample rate, and master clock maximized. Tm can also be any value, but for causality it
rate, and can thus be accounted for in a pre-processing step. is required that Tm ≥ τp (in the matched filter GNU Ra-
dio block, it is assumed that Tm = τp ). Since h(t) is a
 gr-plasma: A New GNU Radio-based Tool for Software-defined Radar

(a) Barker Code (Nc = 13, K = 16) (b) Frank Code (Nc = 64, K = 16) (c) P4 Code (Nc = 64, K = 16)

Figure 4. PCFM waveforms generated from common phase codes

PDU
(Waveform + Metadata)
equal to the duration of the waveform. The width of the
Waveform
Controller
peak, known as the range resolution, is primarily a func-
State information tion of the waveform bandwidth and determines the ability
(Detections,
spectrum occupancy, etc.) of the radar to distinguish targets that are similar in range.
While no new
PDU received

Radar
Processing PDU Pulses
UHD Radar

Figure 5. USRP radar operational logic

time-reversed, complex conjugated version of the xtx (t), it

is said to be “matched” to xtx (t). Performing a convolu-
tion of the matched filter with some received signal xrx (t)
gives
Z ∞
(xtx ∗ xrx )(τ ) = h(τ − t)xrx (t)dt
−∞ (a) Before matched filtering
Z ∞
(4)
= x∗tx (t − τ )xrx (t)dt
−∞
= (xtx ⋆ xrx )(τ )
where a(t)⋆b(t) denotes the cross-correlation between sig-
nals a(t) and b(t). Therefore, the matched filter is mathe-
matically equivalent to the cross-correlation between the
transmitted and received signals. When a copy of the trans-
mitted pulse is present in the received signal, there will be
a peak in the matched filter response at the corresponding
delay. Noise and interference, on the other hand, will be fil-
tered as long as they are not highly correlated to the trans-
mitted pulse. An example matched filter output is shown
in Fig. 6 for an LFM transmitted by an NI USRP-2901 in (b) After Matched filtering
loopback mode so that τ = 0. The waveform is sampled
at 50 MS/s with bandwidth B = 25 MHz and pulse width Figure 6. Time-domain data from an LFM pulse in loopback
τp = 50 µs. Before applying the matched filter, the signal is
hardly visible above the noise (Fig. 6a). However, applying
the matched filter to the received pulse produces a distinc- Once the matched filter has been used to determine the
tive peak more than 30 dB above the noise floor (Fig. 6b). range of a target, the target’s Doppler shift can be ex-
Although there is no delay, the peak exhibits a time shift tracted using a technique known as pulse-Doppler process-
 gr-plasma: A New GNU Radio-based Tool for Software-defined Radar

ing. In gr-plasma, this algorithm is implemented in the

“Doppler Processing” block and works as follows. Con-
sider a radar that transmits M pulses that are reflected from
an object moving with constant radial velocity such that the
received signal is Doppler shifted by FD Hz. If the matched
filter output for a single pulse is yp (t), the matched filter
output for all M pulses is given by

M
X −1
y(t) = α exp(j2πFD t)yp (t − mTP RI − τm ) (5)
m=0 Figure 7. Range-Doppler map for a loopback configuration

where τm is the target delay at each pulse, and α is a com-

plex scale factor that encapsulates global amplitude and On a host system with an AMD Ryzen 5800x CPU and
phase factors that do not vary significantly over the process- Nvidia RTX 3070 GPU, range-Doppler processing could
ing interval. Next, the signal is sampled in both fast-time be performed in real time at sample rates up to 100 MS/s
(analog-to-digital converter (ADC) samples) and slow-time on the CUDA and OpenCL backends. If the message queue
(pulses). Defining Ts as the sampling interval of the ADC of a signal processing block has overflowed above a user-
and k as the sample index, Eq. 5 can be re-written as specified size (i.e., the host PC cannot achieve the required
throughput for a given rate), messages are dropped until
there is again space in the queue. Since data is processed
on a CPI-by-CPI basis, this ensures that the processed data
y[k, m] = α exp(j2πFD mTP RI )yp (kTs − τm ) (6) remains coherent even when samples must be dropped.

which is a matrix where each column is the matched filter 2.4. Miscellaneous Blocks
output for each pulse. For a given Doppler shift, a peak
In addition to the waveform generation and signal process-
can be produced at the target range by coherently integrat-
ing blocks described above, gr-plasma also contains
ing the signal in slow-time so that the pulse-to-pulse phase
blocks for managing PDUs. As mentioned in Section 2.2,
rotation is eliminated. In practice, the Doppler shift is un-
the USRP radar block outputs one PDU per pulse. How-
known a priori and the coherent integration must be per-
ever, many radar signal processing tasks operate on a num-
formed for a Doppler filter bank, where each filter applies
ber of pulses combined to form a coherent processing in-
a phase correction for a different Doppler shift. Thus, for
terval (CPI). For example, the DFT operation in Doppler
each possible Doppler FˆD , the Doppler processing block
processing assumes coherence over many pulses. There-
computes
fore, the “Pulses to CPI” block can be used to consolidate
NCP I PDUs containing individual pulses into one large
PDU. This block also propagates any metadata in the in-
M −1
X put directly to the output for use by downstream process-
Y [k, F̂D ] = y[k, m] exp(−j2π F̂D mTP RI ) (7)
ing. A custom PDU file sink block was also designed to
m=0
facilitate saving the collected data for future use without
first converting back to a traditional stream. The new file
which is equivalent to the discrete Fourier transform (DFT)
sink block can also (optionally) save metadata to a sepa-
of each row in y. Each element in the resulting matrix is a
rate file by converting PDU metadata dictionaries to JSON.
range-Doppler parameter pair, and a statistical hypothesis
This makes it simple to save metadata that complies to
test can be applied to each element to determine if an object
the SigMF standard (Hilburn et al., 2018), and it is flexi-
is present. Fig. 7 shows the range-Doppler response for
ble enough to easily extend to other formats. A sample of
an NI USRP-2901 in loopback, which produces a “target”
the JSON metadata output that is generated automatically
at zero delay and Doppler. gr-plasma also includes a
from the blocks is shown in Listing 2. In future work, this
Range Doppler Sink block for displaying these maps in real
non-standard format will be replaced with a formal SigMF
time.
extension for processing radar data.
The signal processing and waveform generation blocks dis-
cussed above are implemented using ArrayFire (Yalaman-
chili et al., 2015), and the user can choose to perform the
computations using CUDA, OpenCL, or CPU backends.
 gr-plasma: A New GNU Radio-based Tool for Software-defined Radar

{
Rx Antenna Tx antenna
” annotations ”: [
{
” c o r e : l a b e l ” : ” pcfm ” ,
” core : s am p l e s t ar t ”: 0 ,
” r a d a r : d o p p l e r f f t s i z e ” : 1024 ,
” r a d a r : d u r a t i o n ” : 5 . 1 2 e −05 ,
” r a d a r : num phase code chips ” : 128 ,
” r a d a r : n u m p u l s e c p i ” : 1024 , Host PC
NI USRP-2901
” r a d a r : p h a s e c o d e c l a s s ” : ” p4 ” ,
” r a d a r : p r f ” : 10000.0
}
],
” captures ”: [
{
” core : frequency ” : 5000000000.0 ,
” core : s am p l e s t ar t ”: 0
} Figure 8. Experimental test bed setup
],
” global ”: {
” core : datatype ”: ” cf64 le ” ,
” core : sample rate ” : 20000000.0 ,
” core : version ”: ”1.0.0”
}
}

Listing 2. PDU file sink metadata output

3. Experiment Design
Figure 9. Pulse-Doppler processing experiment
To verify the functionality of the module, an open-air ex-
perimental test bed was created using only commercial-off-
the-shelf (COTS) components. In the experiment, an NI with its associated metadata) for further offline process-
USRP-2901 was connected to 25 dBi parabolic dish anten- ing. This process continues until the flowgraph is manually
nas in a simultaneous transmit and recieve (STAR) con- stopped. Relevant radar processing parameters are summa-
figuration (Fig. 8) using only the internal amplification in rized in Table 1. Note that the USRP source and sink blocks
the radio. The host PC was a Dell XPS 13 9310 laptop in the main tree are unable to support this sample rate for
with an Intel i7-1165G7 CPU, 32 GB of RAM, and inte- B210 devices due to overhead from the scheduler, resulting
grated graphics. This laptop was used for the experiment in fatal underflows and late commands.
to demonstrate the utility of gr-plasma for performing Fig. 10 outlines the geography of the area used for the
radar tasks with low-cost, general-purpose systems. Since test. The target is a vehicle that accelerates in the direction
it does not have a dedicated GPU, it is unable to perform shown until it reaches the speed limit of 25 mph. When it
pulse-Doppler processing in real time. Instead, it simply reaches the end of the street, it turns around and accelerates
performs range-Doppler processing as quickly as it can and
drops any excess PDUs it receives while data is being pro-
cessed. Parameter Value
The signal processing chain and the parameter specifica- Sample rate 50 MS/s
tions used in the experiment are as shown in the flowgraph Center frequency 5 GHz
in Fig. 9. Here, an LFM waveform is transmitted and re- Bandwidth 40 MHz
ceived at the constant PRF. When a full CPI has been col- Pulse width 20 µs
lected, matched filtering and Doppler processing are per- PRF 10 kHz
formed and the resulting range-Doppler map is plotted at Range resolution 3.75 m
the fastest rate that the host PC can sustain. At the same Pulses per CPI 256
time, raw I/Q data from each pulse is saved to a file (along
Table 1. Radar Operational Parameters
 gr-plasma: A New GNU Radio-based Tool for Software-defined Radar

(a) Vehicle moving away from radar

Figure 10. Data collection geometry

again toward the radar before decelerating to a stop. This

test setup is desirable for a number of reasons. First, a short
collection range is needed since the output power is limited
(b) Vehicle moving toward radar
by the internal amplification in the SDR. Moreover, a short
collection range can be used since the radar has no blind
Figure 11. Range-Doppler maps
range in the STAR configuration. Although the vehicle is
not instrumented with GPS, having a known speed limit
also makes it easy to validate the range-Doppler output as 5. Conclusion and Future Work
the car accelerates and decelerates.
In this paper, the gr-plasma module has been presented
as a new tool for generating and processing radar data
4. Experiment Results from within GNU Radio. Every block in the module
Fig. 11 shows the range-Doppler maps that were produced uses the message passing architecture to process data as
from the Range Doppler Sink block during the experiment, PDUs, which better captures the bursty nature of radar pro-
zoomed in to the range and Doppler values that were rele- cessing while improving performance by avoiding over-
vant for the given geometry. No windowing was performed head from the scheduler. The module currently imple-
in range or Doppler to maximize SNR and resolution at the ments blocks that generate LFM and PCFM waveforms,
expense of higher sidelobes. In Fig. 11a, the vehicle is ac- collect data from USRP SDRs, and perform pulse-Doppler
celerating away from the radar. There is a pronounced clut- processing on both the CPU and GPU. In future work,
ter ridge near zero Doppler from stationary scatterers in the gr-plasma will be extended to include a more diverse
scene (e.g., buildings, trees, and the ground), and the vehi- collection of waveform generators and more signal pro-
cle can be clearly seen at a range of 50 m as it moves away cessing blocks (e.g., for constant false alarm rate (CFAR)
from the radar at roughly 10 m/s (with a negative Doppler detection). Adaptive capabilities that utilize the results of
shift). downstream processing are also planned for a future re-
lease.
In the range-Doppler map in Fig. 11b, multiple vehicles
are traveling along the road. In this figure, the car from
Fig. 11a is moving towards the radar and can be seen with References
range and velocity (R, v) = (115 m, 10 m/s). At the same Blunt, Shannon D., Cook, Matthew, Jakabosky, John,
time, a large truck and a small car can be seen driving away De Graaf, Jean, and Perrins, Erik. Polyphase-coded
from the radar (on the left). The large truck is located at fm (pcfm) radar waveforms, part i: implementation.
(135 m, −6 m/s), and is the brightest scatterer in the scene IEEE Transactions on Aerospace and Electronic Sys-
besides the clutter. The small car is following the truck tems, 50(3):2218–2229, 2014a. doi: 10.1109/TAES.
at the same speed, but at 120 m from the radar. The car’s 2014.130361.
range-Doppler response has a much lower amplitude than
the truck’s since the truck has a much larger radar cross Blunt, Shannon D., Jakabosky, John, Cook, Matthew,
section (RCS) than the car. Stiles, James, Seguin, Sarah, and Mokole, E. L.
 gr-plasma: A New GNU Radio-based Tool for Software-defined Radar

Polyphase-coded fm (pcfm) radar waveforms, part ii:

optimization. IEEE Transactions on Aerospace and
Electronic Systems, 50(3):2230–2241, 2014b. doi: 10.
1109/TAES.2014.130362.
Christiansen, Jonas Myhre and Smith, Graeme E. Develop-
ment and calibration of a low-cost radar testbed based on
the universal software radio peripheral. IEEE Aerospace
and Electronic Systems Magazine, 34(12):50–60, 2019.
doi: 10.1109/MAES.2019.2953803.
Hilburn, Ben, West, Nathan, O’Shea, Tim, and Roy,
Tamoghna. Sigmf: The signal metadata format.
Proceedings of the GNU Radio Conference, 3(1),
2018. URL https://pubs.gnuradio.org/
index.php/grcon/article/view/52.
Levanon, Nadav and Mozeson, Eli. Radar Signals.
John Wiley Sons, Ltd, 2004. ISBN 9780471663089.
doi: https://doi.org/10.1002/0471663085.ch6. URL
https://onlinelibrary.wiley.com/doi/
abs/10.1002/0471663085.ch6.
Richards, M.A. Fundamentals of Radar Signal Processing,
Second Edition. McGraw-Hill Education, 2013. ISBN
9780071798334. URL https://books.google.
com/books?id=zQlVAgAAQBAJ.
Tan, Peng Seng, Jakabosky, John, Stiles, James M., and
Blunt, Shannon D. On higher-order representations of
polyphase-coded fm radar waveforms. In 2015 IEEE
Radar Conference (RadarCon), pp. 0467–0472, 2015.
doi: 10.1109/RADAR.2015.7131044.
Wunsch, Stefan. gr-radar. https://github.com/
kit-cel/gr-radar, 2014.
Yalamanchili, Pavan, Arshad, Umar, Mohammed, Zakiud-
din, Garigipati, Pradeep, Entschev, Peter, Kloppenborg,
Brian, Malcolm, James, and Melonakos, John. ArrayFire
- A high performance software library for parallel com-
puting with an easy-to-use API, 2015. URL https:
//github.com/arrayfire/arrayfire.