on the integration of opencl into a software defined radio...

On the Integration of OpenCL into a

Software Defined RadioMichael L Dickens and J Nicholas Laneman

!""#$#%& '(#)*+,#-. )#,'%& #/*(-#-. ,-%(/%+/,— "*0+'%+. 1223

GPUTC’12’US 2012-May-17

[email protected] [email protected]

Software Defined Radio

2 / 23 GPUTC’12’US 2012-May-17

➡ Real-time signal processing, using a class of radios that can be reprogrammed and reconfigured primarily via software

➡ Digitize as close to the antenna as possible

➡ Still significantly dependent on hardware

SDR Runtime

3 / 23 GPUTC’12’US 2012-May-17

SDRRuntime

Processing Devices

(CPU, GPU, DSP)

WaveformGraph

QuadratureDemodulator

Source

1:L

Demux

Sink

Subfilter #1

Subfilter #L

L-WaySync

Adder...

...

...

Mapping

Buffering

Scheduling

Processing

➡ Surfer Overview

➡ Key Features

➡ NBFM Example

➡ OFDM Demo

➡ Conclusions

➡ Future Work

Overview

4 / 23 GPUTC’12’US 2012-May-17

SDR Layers Model

5 / 23

1

Shark / naCLSalt / naCL

SalineSQuid

Application

23456

Operating System

7

Hardware

Block ProcessingRuntime / Block

ScriptGraphical

Application

Operating SystemHardware

Spf / naCL


Surfer Overview

6 / 23

Hardware

Salt: Surfer Operating System and Hardware Abstraction LayernaCL: OpenCL Abstraction Layer

Host Hardware Abstraction Layer

Operating System API, ABI, and Kernel

Surfer

Runtime Kernel(Shark)

SupervisorAPI

Script API(Saline)

Signal Processing

(Spf)Application

API

Graphics API(SQuid)


naCL Class Hierarchy

7 / 23 GPUTC’12’US 2012-May-17

naCLPlatform

...

CommandQueue

Context

...

...

......

ProgramSource ProgramSource

MemoryObject

Event ...

Platform

DeviceDevice

Context Context...

CommandQueue

CommandQueue

Program

Program

MemoryObject

Context

Program

...

MemoryObject

MemoryMap ... MemoryMap...CommandQueue

Device

DeviceDevice

Device

Legend...Owns

References

Vector

Class

Platform

Context

Device ... DeviceKernelKernel

Event

1 Signal processing flavors, providing a new form of dynamic runtime reconfiguration and heterogeneous processing

2 Use of thresholds for scheduling, to allow direct control over throughput and latency

2 Use of runtime statistics for modifying system runtime behavior, using a Supervisor

Key Features

8 / 23 GPUTC’12’US 2012-May-17

NBFM Example

9 / 23

Quadrature Demodulator

RateLimiter

Sink Downsampler

Source

➡ Demodulate narrowband FM signal and downsample to typical audio rate


Flavors

10 / 23

1 Signal processing flavors, providing a new form of dynamic runtime reconfiguration

Block

FlavorTable

Processing Flavors

GPUVariant 1

GPUVariant 2

State

Glue

Local CPU Generic

PPC CPUOptimized

Intel CPUOptimized

ARM CPUOptimized


Thresholds

11 / 23

2 Use of thresholds for scheduling, to allow explicit control over throughput and latency

Block 1

Threshold

Block 2

Threshold

Buffer

To iT


Throughput / Latency

12 / 23

1000 2000 3000 40006

7

8

9

10

Proc

essi

ng T

hrou

ghpu

t (M

S/s)

Proc

essi

ng L

aten

cy (m

s)

2000 3000 40000

0.1

0.2

0.3

0.4

Threshold (Items)1000


Supervisor / Statistics

13 / 23

3 Use of runtime statistics for modifying system runtime behavior, using a Supervisor

➡ Load

• Host Processor

• Surfer

• Any Surfer Thread

➡ Block Timing

• Queued

• Dequeued

• Processing Start

• Processing End

• Finished

➡ User Defined

• Energy UseGPUTC’12’US 2012-May-17

Data Flow by Thread

14 / 23

Queued X

Legend

Preprocessing

Processing

Postprocessing

Sleeping

Notification o

0

GPU

CPU

0.40.350.3Relative Time (ms)

0.20.150.10.05 0.25

Thread

0

1


Data Flow by Thread

14 / 23

Queued X

Legend

Preprocessing

Processing

Postprocessing

Sleeping

Notification o

0

GPU

CPU

0.40.350.3Relative Time (ms)

0.20.150.10.05 0.25

Source to Sink LatencyThread

0

1


Data Flow by Block

15 / 23

Legend

Preprocessing

Processing

Postprocessing

Waiting

Notification o

0

Sink

DS

QD

Source0.40.35

Relative Time (ms)0.250.20.150.10.05 0.3

0

1

0

1

Thread


Data Flow by Block

15 / 23

Legend

Preprocessing

Processing

Postprocessing

Waiting

Notification o

0

Sink

DS

QD

Source0.40.35

Relative Time (ms)0.250.20.150.10.05 0.3

Source to Sink Latency

Throughput

Latency 331 µs

3 MS/s

0

1

0

1

Thread


OFDM Demo

16 / 23

➡ Extended audio loopback

➡ Keyboard-based Supervisor runtime system control

• Rotate through flavors

• Increase and decrease processing threshold


OFDM Demo Tx

17 / 23

GNU Radio Style Packet

Encoder

UHD Tx

Surfer Style Packet

EncoderAudio

SourceGNU Radio

Style MapperGNU Radio

Style Preamble Inserter

Inverse FFT

GainCyclic Prefix

Data

Flag

FlavorTable

FFTW3 Optimized

Local CPU Optimized

GPU Variant 1 Optimized

FlavorTable

GPU Generic

Local CPU Generic

FlavorTable

GPU Generic

Local CPU Generic

FFT ShiftFlavorTable

GPU Generic

Local CPU Generic

Local CPU Optimized

GPU Generic

Local CPU Optimized

Local CPU Optimized


OFDM Demo Rx

18 / 23 GPUTC’12’US 2012-May-17

Default Settings

19 / 23

Legend

Preprocessing

Processing

Postprocessing

Waiting

Notification o


8.8 MS/sInstantaneous

Throughput

0

Sink

CP

Gain

iFFT

FFT Shift

GR Preamble

GR Mapper

S Packet

Relative Time (ms)

Source 1.41.210.80.60.40.2

GR Packet Latency 1.16 ms


GR Packet Encoder

20 / 23

Legend

Preprocessing

Processing

Postprocessing

Waiting

Notification o


InstantaneousThroughput

Latency

1.4 MS/s

0

Sink

CP

Gain

iFFT

FFT Shift

GR Preamble

GR Mapper

S Packet

Relative Time (ms)

Source 14012010080604020

GR Packet 109 ms


FFT Shift

21 / 23

Legend

Preprocessing

Processing

Postprocessing

Waiting

Notification o

1.2 MS/sInstantaneous

Throughput

Latency 1 ms

0

Sink

CP

Gain

iFFT

FFT Shift

GR Preamble

GR Mapper

S Packet

Relative Time (ms)

Source 12108642

GR Packet



➡ Evolving software defined radio runtime to exploit heterogeneous processing

➡ Key Features: Flavors, Thresholds, Statistics and Supervisor

➡ Supervisor can use statistics to set thresholds and flavors, providing runtime system control of throughput / latency, and mapping

Conclusions

22 / 23 GPUTC’12’US 2012-May-17

➡ GUI using Qt (SQuid)

➡ Flavors for CUDA and other processors

➡ Better optimized GPU flavors

➡ Interval / cluster mapping

➡ QoS-based adaptive runtime

➡ Using more embedded graph information

Future Work

23 / 23 GPUTC’12’US 2012-May-17

Thank you!

Questions?


on the integration of opencl into a software defined radio...

Documents