See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/304346578

Open-Source SDR on Embedded Platforms

Presentation · June 2016 DOI: 10.13140/RG.2.1.1108.4409

CITATIONS READS 0 1,478

1 author:

Raj Bhattacharjea Georgia Tech Research Institute 7 PUBLICATIONS 63 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Open-Source SDR on Embedded Platforms View project

Accelerated Signal-Processing on Embedded Platforms: Paths Forward View project

All content following this page was uploaded by Raj Bhattacharjea on 23 June 2016.

The user has requested enhancement of the downloaded file. Open-Source SDR on Embedded Platforms

June 1, 2016

Dr. Rajib Bhattacharjea Georgia Tech Research Institute Information and Communications Lab Atlanta, GA 30318, USA

1 GTRI Part 1 Overview

• Prereqs • Topics to be covered • SDR background • SDR uses and users • SDR components • Demo 1

2 GTRI Prerequisites

• Radio/Comms Fundamentals - Analog and Digital Modulation - Mixers, filters, LNA, PA • Circuit analysis concepts - Transistors - Filters • concepts - High level language source code - Assembly language • Digital and analog signal processing concepts - Convolution - FFT - Dot products

3 GTRI What We’re Going to Learn About (and play with)

4 GTRI Single Board Computers!

http://hackerboards.com/ringing-in-2016-with-64-open-spec-hacker-friendly-sbcs/

5 GTRI Embedded Computers from the Living Room!

6 GTRI Signal Processing with Open Source Tools!

http://wiki.opendigitalradio.org/FM_RDS_Stereo_transmitter_using_gnuradio

7 GTRI Software Defined Radio Hardware!

8 GTRI What We’re Going to Learn About

• Overview of available open-source software and embedded hardware tools for SDR development • Signal processing • RF signals • How to get started with SDR on popular embedded platforms - ARM or embedded CPU - • Commercial SDR hardware • Optimized signal processing capabilities available on embedded platforms

9 GTRI What We’re Not Going to Learn About

• OpenEmbedded, yocto, poky, bitbake…although those would help you with devices that don’t have packages in the distros, or for developing reproducible builds for distribution. • Embedded CPU Hardware without available Linux kernel • IBM Power ISA, µC(Linux), MIPS, SPARC, etc. - Although the open source SDR ecosystem could be made to run on these kinds of platforms with various degrees of effort - MIPS is especially low hanging fruit and much of this talk applies to those devices too - Dedicated DSP chips • “IoT” but only because we’re not talking directly about what data goes over the radio links we’re talking about. If it was some kind of sensor data, this would be an IoT talk. • OSSIE (is it dead?) 10 GTRI Background (who, what, when, where, why, how?)

11 GTRI What is a Software-Defined Radio?

• A communication system where the capabilities are largely determined by the software running on the system instead of by the radio hardware. • Consists of radio hardware and something to do the signal processing. • The term “SDR” can mean the whole system, or just the radio hardware. • Let’s consider a series of examples

12 GTRI Conventional Radio Example: Super Regenerative FM Receiver

Design from: http://electronics-diy.com/simple-fm-radio.php

13 GTRI Conventional Radio Example: Super Regenerative FM Receiver • It’s purpose built for a specific task • Designed to tune over 88 MHz – 108 MHz • It isn’t easily reconfigurable and can perform no other task - The hardware defines the functionality - Many functional stages are collapsed into one; the one JFET is the: • Oscillator • Mixer • Amplifier • Filter • Uses VERY few components, only enough to achieve the function it is designed for

14 GTRI Conventional Radio Example: Superheterodyne FM Receiver

15 GTRI Conventional Radio Example: Superheterodyne FM Receiver

Superheterodyne receiver schematic by Appaloosa released under CC-BY-SA 3.0 license. https://commons.wikimedia.org/wiki/File:Tuner1.svg

16 GTRI Conventional Radio Example: Superheterodyne FM Receiver

17 GTRI Conventional Radio Example: Superheterodyne FM Receiver

Design from: http://www.daenotes.com/electronics/communication-system/superheterodyne-fm-receiver

18 GTRI Conventional Radio Example: Superheterodyne FM Receiver

Design from: http://www.radio-electronics.com/info/rf-technology-design/fm-reception/foster-seeley-fm-detector-discriminator.php

19 GTRI Conventional Radio Example: Superheterodyne FM Receiver • It’s purpose-built for a specific task • Designed to tune over 88 MHz – 108 MHz • Design is more modular than the super regenerative design - Amplification, frequency synthesis, frequency mixing, and a multistage demodulator • Hardware defines the functionality - Could change out the limiter and discriminator for a simple envelope detector and get an AM radio receiver - Or a PLL and detect FSK • Design is more complicated and uses about 7 times the number of overall components and 5 times the number of transistors

20 GTRI Software Defined Radio: Superheterodyne w/ an ADC

21 GTRI Software Defined Radio: Superheterodyne w/ an ADC • Not purpose-built for a specific task! • Tuning range can be large • Signal processing is not done by a circuit designed for demodulation or discrimination of a specific type of waveform • Software signal processing defines the functionality - Hardware determines characteristics (tuning range, max instantaneous bandwidth, noise performance)

22 GTRI Who Uses SDR and Why?

• Engineers in R&D • Enables cutting edge research • Academics • Speeds prototyping of new • US Military radio designs and signal processing algorithms - JTRS - SCA • Allows one radio to support different waveforms • US Intelligence Community • Permits radio designs to be • Amateur radio enthusiasts upgraded without buying new (hams) hardware • Others • Facilitates detailed offline signal analysis

23 GTRI Who uses SDR and Why? Video

• https://youtu.be/Lv-vkBNzZwE

24 GTRI SDR Users / Developers in the DoD and IC

• The Army, Navy, Air Force, and Marine Corps • “No Such Agency” in Fort Meade, MD - https://www.google.com/#q=nsa+redhawk • Defense Contractors - Harris

http://rf.harris.com/media/RF-7800V-HH_Enlarged_26-12108.jpg

25 GTRI Components of an SDR

• Analog front end - Quadrature demodulator - Wide or narrow tuning range - Wide or narrow bandwidths - RX Only, TX/RX half duplex, TX/RX full duplex, TX/RX MIMO configurations typical • ADC - Direct sampling ADC is a possibility, limits “tuning” range to (0,fs/2) • Baseband w/ high speed interface - FPGA - CPU - µC • Signal processing software

26 GTRI Components of an SDR: Front-end

27 GTRI Components of an SDR:

• Embedded device with a CPU or µC • FPGA

28 GTRI Demo 1 - CPU Frequency Demodulation

29 GTRI Questions?

30 GTRI Part 2 Overview

• Embedded platforms • Consumer electronics embedded • Capabilities • Types • Will it SDR? Demos 2 – 4 • Embedded ARM landscape • SDR hardware options

31 GTRI Embedded Processors (low power brains)

32 GTRI Ubiquity of Embedded Platforms

• Embedded processors or µCs are in - Car - Television - Speakers - Sound system - Wifi router - GPS receiver - Cable modem - Camera - Vacuum cleaner - Refrigerator - Watch

33 - Calculator, etc. GTRI Consumer Electronics Embedded Platforms

• Largest consumer electronics companies all sell embedded devices - Connected TV devices - Tablets - more than 250 million projected to sell globally in 2016[1] - – 1.4 billion units sold in 2015[2] • Android and iOS devices drive this market - - Apple - Samsung - Amazon

[1] http://www.gartner.com/newsroom/id/2954317 [2] http://www.statista.com/topics/840/smartphones/

34 GTRI Consumer Electronics Embedded Platforms

• These embedded CPUS go by the name “applications processors” • Hardware manufacturers amortize investment into a technology; technology trickles laterally into the maker/hacker, educational, and industrial, and commercial communities - Broadcom markets the BCM2820 as “optimized for high volume markets including mobile phones, mobile TVs, and portable audio/video/game devices”[1] - Sells the almost identical BCM2835 as a learning platform to the Pi Foundation - R&D companies ship ’s to their customers in other products • This is the class of embedded hardware we are talking about today

[1] https://www.broadcom.com/press/release.php?id=929312

35 GTRI What Drives the Capabilities?

• Often, the killer app: streaming, decoding, and playback of media - Special CPU instructions - Media processor / GPU hardware - Some of this capability might help in radio signal processing, more on this later • This talk focuses exclusively on ARM and - Linux runs effortlessly, entire open-source ecosystem is available - Can be extended to anything that runs a full, modern Linux kernel (MIPS24K in particular)

36 GTRI Types of Consumer Electronics Embedded Platforms

37 GTRI Set Top Boxes

• Class of devices used for consumption of media on TV screens • Options: - devices - Apple TV - Amazon Fire TV - WDTV - Shield - Many, many lesser known brands (check online retailers for “Android TV” or “set top ”)

38 GTRI Set Top Boxes: Will it SDR?

Yes it will!

http://www.cnx-software.com/2015/12/06/how-to-run-headless-linux-on--s905-devices-such-as-mini-mx-or-k1-plus/

39 GTRI Stick Computers

• New class of small form-factor computers for vending signs, kiosks, TV streaming, etc. • Options: - - Amazon Fire Stick - Lenovo IdeaCentre - - Azulle Quantum Access - MeeGoPad devices - And others…

40 GTRI Stick Computers: Will it SDR?

• Demo 2: Dongles on dongles • Yes it will!

41 GTRI Mini PCs

• Like a stick computer but slightly larger and the HDMI port goes the other way • Generally more ports and/or storage • Options: - Kangaroo MD2B - Vensmile iPC002 - Compulab MintBox Mini

42 GTRI Mini PCs: Will it SDR?

• Demo 3: Vensmile Mini PC SDR • Yes it will!

43 GTRI Single Board Computers (SBCs)

• Like a mini PC without the • Options: case - Raspberry Pi - BeagelBone Black - - Banana Pro M2 - Lemaker Guitar - Parallella 16 - CubieBoard4 CC-A80 - DragonBoard 410c - Odroid XU4 - Odroid C1+ - MinnowBoard Tubot - - TK1 - Nvidia Jetson TX1 - And others…

44 GTRI Single Board Computers

http://hackerboards.com/ringing-in-2016-with-64-open-spec-hacker-friendly-sbcs/

45 GTRI Single Board Computers: Will it SDR?

• Demo 4 – single board computer SDRs - Raspberry Pi 2 - BeagleBone Black - LeMaker Guitar - NVIDIA Jetson TK1 - ODROID XU4+ - ODROID C1+ • Yes it will!

46 GTRI Plug Computers

• A computer with AC mains prongs hanging off the case - Like a wall-wart, but a computer • The guts are similar ARM processors • A few vendors on the market sell this form-factor

47 GTRI Plug Computers

48 GTRI Plug Computers: Will it SDR?

49 GTRI Embedded ARM Landscape

Manufacturer Brand Name End Devices Texas Instruments OMAP Kindle Fire HD, Droid X, Droid 2 Broadcom Raspberry Pi Allwinner Tablets ASUS , TV Boxes/Sticks MediaTek Tablets, Phones Google , HTC Incredible Marvell ARMADA Toughpad A1, Google Amlogic TV Boxes Samsung HP Chromebook, S6 NVIDIA NVIDIA Shield Apple AX iPhone, iPad, Apple TV, Apple Watch Tablets, TV Boxes NXP i.MX, QorIQ Amazon Kindle HiSilicon Huawei Devices Atmel SMART SAMA5 Dev Boards Renesas Samsung Galaxy Core LTE

50 GTRI SDR Hardware

• Ettus USRP • Great Scott Gadgets HackRF • Nuand bladeRF • LimeSDR • Realtek RTL2832U-based devices • Afedri SDR • Airspy R820t • RFSpace • RedPitaya • Nutaq PicoSDR, ZeptoSDR • PicoZed SDR • FunCube Dongle Pro+ • SDRPlay RSP • And more...

51 GTRI SDR Hardware: References

• https://en.wikipedia.org/wiki/List_of_software-defined_radios • http://sdr-radio.com/Radios • http://www.rtl-sdr.com/roundup-software-defined-radios/ • http://gnuradio.org/redmine/projects/gnuradio/wiki/Hardware

52 GTRI Questions?

53 GTRI Part 3 Overview

• Open-source SDR signal processing software • Install guides – Demos 5 and 6 • Non-free tools • Core SDR signal processing operations • Speed and optimization • Intro to CPU ISA extensions

54 GTRI SDR Software Options

• GNU Radio • Mathworks MATLAB* • NI LabVIEW* • REDHAWK • Scilab • Custom signal processing code (C++, Python, etc.)

* Not publicly available on ARM architecture, not free of cost

55 GTRI Open Source SDR Software: Getting Started

• Get a Linux distro - Download and “dd” an image file on to an SD card - Or get it onto the on-board eMMC • Keyboard, monitor, mouse, etc. - Or setup networking for headless operation • Install your SDR software - More on this later

56 GTRI GNU Radio: Getting Started

• Install - package manager - Debian derivative - install using apt - Red Hat derivative – install using dnf (yum is deprecated) - Others: zypper, pacman, urpmi, RTM for your distro • Install – source - Install dependencies from package manager if possible - Then git clone, cmake, make, make install - Cross compilation or other tricks might be necessary • Embedded systems are limited in computational power, and GNURadio compilation is CPU intensive • RAM can be an issue • eMMC flash storage can be an issue • Colleague’s GNU Radio compile on a Raspberry Pi took a week

57 GTRI GNU Radio Install

Demo 5

58 GTRI REDHAWK

• Open source signal processing package largely developed for/by the US government • Supported on RHEL or CentOS Linux distributions; developers implicitly assume x86 • Port of REDHAWK for exists - Old version of RedHawk (1.8.4) - Old version of Ubuntu (12.04) • Developer community size is a bit of a mystery - Fewer than 200 questions on the official Q&A site for redhawk software - Zero questions on the official Q&A site for redhawk-related hardware problems. • ARM ports will take some work

59 GTRI REDHAWK

• Demo 6: REDHAWK on embedded ARM

60 GTRI SciLab

• Some efforts exist to use a combination of SciLab and GNU Radio • Not developed for SDR work, but for interfacing with data acquisition hardware

https://www.researchgate.net/publication/266009629_GNURadio_Scilab_Xcos_and_COMEDI_for_Data_Acquisition_and_Control_An_Open_Source_Alternative_to_LabVIEW

61 GTRI Non-Free Tools (MATLAB and LabVIEW)

• Title of the tutorial contains “Open-Source SDR…” • You can open source the code you write but development tools are not free (neither “freedom” nor “beer”) • Mentioned here for completeness

62 GTRI Custom Signal Processing / Vendor APIs

• Most SDR manufacturers have written C/C++ libraries interfacing to the hardware • These can be variously used directly or through wrapper/binding layers to other languages

63 GTRI Signal Processing on Embedded Platforms (make it go faster)

64 GTRI Embedded Processing: Operations and Speed

SDR Operations • The modern CPU is optimized • Modulation for managing the flow of instructions, not quickly • Entropy and error coding executing a small group of • Encryption instructions • Computationally heavy DSP operations • CPU manufacturers see this - Convolutions gap, have created special CPU - FFTs instructions for common DSP or multimedia computations • Potentially high data rates

65 GTRI Embedded Processing: Operations and Speed

Dot product

푀−1 푥 푛 ∙ 푦 푛 = 푥 푚 푦[푚] 푚=0

#define M (1L << 16) double dotProd_1(double *x, double *y) { double out = 0; int i; for (i = 0; i < M; i++) { out += x[i] * y[i]; } return out; }

66 GTRI Embedded Processing: Operations and Speed

Convolution (FIR filter)

푀−1 (푥 ∗ 푦) 푛 = 푥 푛 − 푚 푦[푚] 푚=0 #define M (1L << 4) #define L (1L << 4) void convolution_1(double *x, double *y, double *out) { int n; for (n = 0; n < L + M - 1; n++) { out[n] = 0; int m; for (m = 0; m < M; m++) { if(n-m < 0 || n-m > L-1) continue; out[n] += x[n-m] * y[m]; } } }

67 GTRI Embedded Processing: Operations and Speed

• Matrix multiplication - Could be for adaptive zero-forcing equalizer

푀−1

푋푌 푖푗 = 푋 푖, 푚 푌[푚, 푗] 푚=0 • Code omitted, but now three loops - One over 푖, one over 푗, and one over 푚 - But the multiply and add up bit is the same!

68 GTRI Embedded Processing: Operations and Speed

• Operation of multiplying elementwise between lists and summing is a very common operation. - Previous examples (signal convolution, matrix multiplication)

푥1푦1+푥2푦2+푥3푦3 - Vector projection 2 2 2 푦1 +푦2 +푦3 2 3 - Polynomial evaluation 푎0 + 푎1푥 + 푎2푥 + 푎3푥 • What if someone identified many of these common kernels of code and spent time optimizing them using CPU extensions for specific hardware? - The bit that boils down to pure math and runs in the innermost loops

69 GTRI Ways to Achieve Speed on Embedded CPU

Use the CPU extensions; how? • Use compiler features directly • ORC • VOLK • OpenCL • Performance-tuned libraries

70 GTRI CPU Extensions

• Additions to the instruction set for some specific application • Drivers: media, graphics, security • Speed-up computations or increase power efficiency Applications • Encoding or decoding modern video/audio standards - DCT in audio/video compression • Rasterization, shading, ray tracing for gaming • Hashing / encryption operations

71 GTRI CPU Extensions

• Vectorized or vector or SIMD • Maximal DSP performance out of an embedded processor • Intel – MMX, SSE, SSE2, SSE3, SSSE3, SSE4, AVX, AVX2 • ARM – DSP extensions, SIMD, NEON • Can be leveraged for embedded signal processing! • Writing assembly can be problematic, not for everyone • Not portable across architectures

72 GTRI CPU Extensions: Intel PMADDWD

73 GTRI CPU Extensions Example

outerloop: pxor xmm0, xmm0 // initialize 4 accumulators (32Bit) Questions: sub ebp, ebp // init Index in filter taps w/ zero innerloop: movups xmm1, qword ptr [esi+ebp] // ld 8 vec in 128Bit reg • What does that do? pmaddwd xmm1, qword ptr [edi+ebp] // pairwise mult with TAPs paddd xmm0, xmm1 // accum results in xmm0 (4x32 Bit) add ebp, 2*8 cmp ebp, 2*64 // repeat for all 64 Taps • Is that optimal in any sense? jnz innerloop phaddd xmm0, xmm0 phaddd xmm0, xmm0 // accumulate four partial vals to one result psrad xmm0, 15 // scale to 16 Bit • What extension set is it using? movd qword ptr [eax], xmm0 // store result (only 16 Bit relevant) add esi, 2 add eax, 2 // increment pointer for input & output sub ecx, 1 // repeat for all input samples • Will it run on an Z3735F? On an ARMv7?

https://software.intel.com/en-us/articles/using-intel-streaming--extensions-and-intel-integrated-performance-primitives-to-accelerate-algorithms

74 GTRI Questions?

75 GTRI Part 4 Overview

• CPU extensions • GPGPU on embedded • Using the FPGA • Special • Conclusions and wrap-up

76 GTRI Optimization Path 1: CPU

77 GTRI How to Leverage CPU Extensions: Compiler

• Compiler intrinsics expose the CPU vector operations. Like writing assembly code in the middle of C++. - Portability problem across CPU types • Special data structures (vector types) + normal operations that the compiler can recognize and optimize - Portability problem across compilers • -O3 - It just works!...? No.

78 GTRI How to Leverage CPU Extensions: Compiler

• Demo 7: C++ code with assembly breakdown - Add two signals together sample by sample in a loop • You have to coax the compiler into optimizing, then look at the performance yourself and determine if the assembly matches your expectations • Even then it might be suboptimal and more performance tuning can be required

http://locklessinc.com/articles/vectorize/

79 GTRI How to Leverage CPU Extensions: ORC

• ORC, Optimized inner loop Runtime Compiler • Generic SIMD language that gets compiled to optimized instructions for the target processor • Inner loop code is optimized using vector instructions, across architectures • User drops in optimized implementations - Without having to coax the compiler into the optimization - Without hand writing the machine specific assembly (but ORC is assembly-like) • Supports ARM NEON, Intel MMX/SSE, and others (PowerPC Altivec) • Different modes - JIT compilation mode - Hardware specific machine code that gets baked into the executable.

80 GTRI How to Leverage CPU Extensions: VOLK

• Vector Optimized Library of Kernels • Signal processing functions that vectorize well across architectures • Protokernel – tuned implementation that runs on specific hardware • Different protokernels are profiled and what’s fastest is stored • Kernel -> profiler data -> dispatch - Empirically fastest implementation - Similar to the FFTW library (“wisdom”) • ARM NEON and Intel SSE/AVX • If no SIMD acceleration, plain C implementation is used

http://libvolk.org/ Image © Marcus Müller

81 GTRI How to Leverage CPU Extensions: OpenCL

• Open Computing Language • C-based language designed for optimizing math heavy operations • Some constructs are designed to map to SIMD structures on particular hardware - “float4” is a type with four single precision floating point numbers - Can map directly to Intel SSE XMM0-XMM7 • OpenCL code can be compiled for - CPU: uses SIMD / other extensions - GPU: uses architecture - Special OpenCL coprocessors: Intel Phi, Adapteva Epiphany

82 GTRI How to Leverage CPU Extensions: OpenCL

Hardware must have OpenCL implementation available • Intel ships OpenCL implementation for their CPUs • Official ARM CPU OpenCL using NEON was announced but unavailable? • ARM Mali GPU design does have an OpenCL implementation, more on this later • POCL - Portable Compute Language - OpenCL implementation for ARMv7, x86, MIPS32

83 GTRI How to Leverage CPU Extensions: OpenCL

• Demo 8: OpenCL Acceleration demo with gr-fosphor on Vensmile iPC002 () • Formerly had to do this on some distros: - https://bitbucket.org/snippets/rajb245/epGzn • But latest Intel OpenCL Runtime 16.1 supports RHEL, SLES, and Ubuntu - https://software.intel.com/en-us/articles/opencl-drivers

84 GTRI How to Leverage CPU Extensions: Performance Tuned Libraries Background – BLAS • Basic Linear Algebra Subprograms • API definition of common numerical operations • Level 1: Scalar/vector and vector/vector operations - Rotating 2-vectors, dot products, scalar times vector, vector norms • Level 2: Matrix/vector operations - Matrix/vector multiplication, triangular matrix system solver, vector outer product • Level 3: Matrix/matrix operations - Matrix multiplication - Sums of products of matrices • Do they map directly to signal processing?

85 GTRI How to Leverage CPU Extensions: Performance-Tuned Libraries • CPU vendors make performance-tuned BLAS (and the related LAPACK) implementations - AMD ACML – open sourced, BLIS is what we want for the CPU, runs on ARM and x86, very portable - Intel MKL – Intel and AMD x86, closed source • ATLAS project automatically tunes itself on the target hardware, works on ARM • OpenBLAS has comprehensive architecture support https://software.intel.com/en-us/intel-mkl

86 GTRI How to Leverage CPU Extensions: Performance-Tuned Libraries Open questions • Integration of high-performance BLAS into GNU Radio? REDHAWK?

87 GTRI Optimization Path 2: Embedded GPU

88 GTRI Embedded GPU Landscape

• Consumer electronics embedded processors often have specialized multimedia processing hardware / GPU (recall the killer app) • End device list not exhaustive IP Holder Brand Name Licensees / Manufacturers End Devices

Broadcom VideoCore Broadcom Raspberry Pi, Samsung Galaxy (various)

ARM Holdings Mali Allwinner, Amlogic, MediaTek, Samsung Gear/Galaxy/Chromebook (various), Rockchip, Samsung Google , ODROID (various) NVIDIA Tegra GPU NVIDIA NVIDIA Shield, Acer Chromebook, Google C, HTC Intel HD Graphics, GMA, Intel Dell Pro 11 2in1, Lenovo ThinkPad (various), Iris, Iris Pro Dell Venue 11 Pro Qualcomm Qualcomm Samsung Galaxy (various), Sony Xperia X, HTC 10 Imagination PowerVR Ti, Apple, Intel, Broadcom, iPhone (various), iPad (various), Apple Watch, Technologies Allwinner, Samsung, Rockchip PS Vita, Asus Zenfone 4, ODROID-XU Vivante (None) Marvell, Freescale, Ingenic Samsung Galaxy Tab 4, Chuwi V90, Corporation Semiconductor, Rockchip Hummingboard (various)

89 GTRI GPU Processing on Embedded

• Mali OpenCL is the only easy option right now - Signal processing with AMD clMath? - Integration into GNU Radio or REDHAWK? • PowerVR has beta access to OpenCL for developers • NVIDIA has CUDA - Library of functions that turn into highly parallelized code that runs on the GPU - Specific to one vendor • Some work has been done on Broadcom VideoCore - Assemblers for the VideoCore, including one where you write assembly in python (PyVideoCore) - PyVideoCore has a single example of a single BLAS/LAPACK function

90 GTRI GPU Processing on Embedded

• Demo 9: Mali GPU w/ OpenCL gr-fosphor

91 GTRI Optimization Path 3: FPGA

92 GTRI FPGA Processing

• Field-programmable gate array - Can be configured to be an arbitrary digital logic circuit (within a size constraint) • Many of the SDR hardware platforms have an on-board FPGA - Handles interpolation, decimation, fine frequency tuning, digital I/O buses (host and RF front-end) - HackRF*, Ettus B and N series, bladeRF, and others • Some embedded processors have an on-die FPGA† - Zynq - SoC

*Actually a CPLD, but the difference is in implementation not functionality †Intel Xeon+FPGA is hardly embedded and is so far 93 GTRI FPGA Processing

• Xilinx Zynq powers three popular SDR platforms - Ettus E310 - Red Pitaya - PicoZed SDR • Do most of the signal processing on the FPGA http://www.xilinx.com/products/boards-and-kits/1-askjht.html • Use the ARM for what’s left • Development Tools - LabVIEW FPGA Module* - MATLAB HDL Coder* - RFNoC https://www.ettus.com/product/details/E310-KIT *Not free, no public ARM release

94 GTRI FPGA Processing

• Demo 9: RFNoC fosphor on the X310 - (Martin Braun’s demo is much better, so we’ll skip this)

95 GTRI Optimization Path 4: Other Coprocessors

96 GTRI Special Coprocessors

• Xeon Phi - Not embedded yet; power draw is same as thirty SBCs • Adapteva Epiphany - Has not really caught-on - Used in one design, the Adapteva * Parallella-16 SBC

http://www.adapteva.com/epiphanyiii/

*http://www.intel.com/content/www/us/en/architecture-and-technology/many-integrated-core/intel-many-integrated-core-architecture.html

97 GTRI Conclusions and Summary (what did we learn)

98 GTRI Conclusions and Summary

• Embedded SDR runs the gamut of price and performance - Pi Zero + RTL2832U device - ODroid + Hack RF - Ettus E310 or PicoZed SDR - NVIDIA Tegra TX1 + Ettus X310 • GNU Radio is the dominant open-source SDR package that is popular for researchers and hobbyists - Others exist but there’s no annual “REDHAWK con”, i.e., the user community and support seems to be best for GNU Radio • Using available, COTS, embedded hardware and open-source software, any practical radio technology can be implemented - Even experimental technology can be realized - Massive MIMO - Spectrum sensing / cognitive radio

99 GTRI Conclusions and Summary

• You can’t pick up an embedded device and get optimal signal processing performance - Software work to be done to leverage the processing power of the GPUs and SIMD technologies - Should the VOLK project support GPU accelerated implementations? - Should everyone use OpenCL? - Or should there be accelerated BLAS/LAPACK, and other code can call those functions?

100 GTRI Conclusions and Summary

• It’s an exciting time in embedded SDR! • The unprecedented popularity of mobile, wireless, and TV markets has effectively subsidized the SBC and embedded computer market. • You can realistically learn SDR and signal processing with practical, hands-on experience for the price of a drink at a nice restaurant.

101 GTRI Acknowledgements

• Ben Riley of GTRI for funding IRAD

102 GTRI Thank You

103 GTRI Copyright Notice

• Copyright 2016, Georgia Tech Applied Research Corporation • Portions of this work are under the copyright of their respective, original rights holders. - No copyright claim is made on these portions. - Where possible, attribution has been given to the original rights holders. - Use of these portions has been for the purposes of comment, teaching, scholarship, and research; as such, the authors of this work assert that this usage constitutes a fair use of the copyrighted material.

104 GTRI 17 U.S. Code § 107 - Limitations on exclusive rights: Fair use • Notwithstanding the provisions of sections 106 and 106A, the fair use of a copyrighted work, including such use by reproduction in copies or phonorecords or by any other means specified by that section, for purposes such as criticism, comment, news reporting, teaching (including multiple copies for classroom use), scholarship, or research, is not an infringement of copyright. In determining whether the use made of a work in any particular case is a fair use the factors to be considered shall include— - (1)the purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes; - (2)the nature of the copyrighted work; - (3)the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and - (4)the effect of the use upon the potential market for or value of the copyrighted work.

105 GTRI

View publication stats