Soonhoi Ha, SNU 11 Feb. 14, 2006 HOPES: Embedded Software Development Environment for MPSoC Feb. 14,...

Soonhoi Ha, SNU 11Feb. 14, 2006

HOPESHOPES: Embedded Software Developme: Embedded Software Development Environment for MPSoCnt Environment for MPSoC

Feb. 14, 2006

Soonhoi HaSeoul National University


Contents

• Introduction• Proposed Design Flow• Key Techniques• Status• Conclusion


Deep Submicron Era

• Increased chip density– MPSoC (Multiprocessor System on Chip), NoC (Network on Ch

ip)– Increased chip density– Increased NRE cost– Increased manufacturing cost

• Platform based design– Platform: common (HW/SW) denominator over multiple applic

ations– Pre-built and verified (HW/SW) architecture– SW design and system verification are major challenges


SoC (System-on-a-Chip)

• Assembly of “prefabricated component”– Maximize VC(IP) reuse: over

90% – New economics:

fast and correct design > optimum design

• Design and Verification at the system level– interface between VCs– SW becomes more important

up memory

video unitgraphics

coms DSP custom

software

up


HOPES Project

• Project– Title: Development of Embedded software design and verificat

ion techniques for MPSoC, Period: 2005.3.1 - 2008.2.29,– Amount: $4,200,000 – Sponsor: Korean Ministry of Information and Communication,

• Mission– Develop an integrated embedded software development envir

onment based on model-based programming techniques and multi-stage validation techniques for MPSoC (Multi-Processor System on Chip)


Motivation

• Current SoC-related projects (in Korea) focus on hardware design and verification

• Software design on MPSoC– Embedded software with timing and resource constraints– Parallel programming for heterogeneous multiprocessors– Fast virtual prototyping


Current Practice (in Korea)

• Virtual prototyping– Coware ConvergenSC, ARM(Axys) MaxSim– Manual software and hardware design– TLM (transaction level modeling) simulation: 100K inst./sec

• Model-based programming– UML based tool: (ex) Telelogic TAU– Limited capability: documentation, code structure.

• No tool is available for parallel programming– Manual programming


Common Intermediate Code (CIC)

CIC (w/ API) translator

Generic API

HOPES Proposal

UMLModel-basedProgramming

(PeaCE model)

KPN

Target HW platform

SW Platform 1 2

API lib.Per-processor code

Static C code analysis

Perf. & power estimationVirtual

prototypingProcessor ISS


Key Techniques

• SW design techniques– Model-based specification– Partitioning and automatic code generation from

specification– Parallel programming (task parallelism, data parallelism)– Generic APIs (independent of OS and target

architecture)

• SW verification techniques: 3-phase verification– At specification level: static analysis of program

specification and functional simulation– After code generation: static analysis of C code– At simulation time: debugging on virtual prototyping

environment


Contents



Proposed design flow

• Model-based programming– PeaCE model (dataflow + FSM + task model)– ESUML (embedded system UML) model

• CIC (Common Intermediate Code)– OpenMP pragma + generic API

• Static Analysis– Buffer overrun, memory leak, null dereference, stack size

• Virtual prototyping – Performance and power estimation– with source-level debugging capability


Software Module Interface

OS API Translator

Verification OS

Code Synthesizer

PeaCE_Modeler

PeaCE_Partitioner

ARM processor simulation

ESUML_Modeler

Power analysis module

MPI Library

Common Intermediate Code (CIC): task_graph.cic

GUI SW Platform

Document generation

Performance analysys module

ESUML_Verifier

results

ESUML_Partitioner

Block library DB

Generic parallel API Generic API

{processorX.c}

virtual prototype system

proc. code with OS

OpenMP translatorparallel API

MPSoC {prototype system}

C-code static analyzer

HW component DBproc. code with generic API

Comm. API Lib.

proc. code without OS


PeaCE_Modeler

PeaCE_Partitioner

ESUML_Modeler

▼ 공동 1


GUI SW Platform

Document gen.

Application_graph.xmluser command

Task_graph.xmlsched.xml

ESUML_Verifier

resultsApp_classDiagram.xmlApp_seqDiagram.xmlAction semantics

{Application_graph.xml}user command

User_commandsched.xml

ESUML_Partitioner

Block perf. DB

TimeCost.xmlarchi.xml


OS API Translator

Verification OS

Code Synthesizer

ARM ISS

Power estimation

MPI Library


Perf. estimation

processor parameter

openMP pragma Generic API

{processorX.c}

processorX.elf

virtual prototype system

code w/ OS

OpenMP translatorparallel API

MPSoC {prototype system}Function symbol table

Cycle-level execution trace

C-code static analyzer

HW component DBproc. Code(generic API)

comm API Lib.

code w/o OS


Demo 1. HOPES Design Flow

• Model-based programming (ex: divx player)– PeaCE Model

• Dataflow + (FSM) + task model

– Fault simulation• sample rate inconsistency

(Y;U;V= 3:1:1)• deadlock error

– Functional simulation• Host machine


Demo 1. HOPES Design Flow (2)

• C code static analysis – Buffer overrun error detection– Fault simulation: wrong array siz

e in “AviFileReader task”

• Runtime verification– Simulation with Realview ARM d

ebugger– Fault simulation

• Logical error insertion


Demo 2. HOPES (single task)

• Single task example: H.263 decoder

• C-code static analysis• Performance and power

estimation – Custom-designed ARM

Simulator– Per-function estimation

• Runtime verification– Source level debugging– Fault simulation : logical error

insertion


HOPES flow for single processor

Manual for now

HOPES GUI

PeaCE code gen.- CIC(xml) - CIC(.cic)

Xml files (topology, property)

CIC code withGeneric API, OpenMP pragma

Generic API tranformer

OpenMP translator

Code Synthesizer- C code, Executable

CIC code With Generic API

CIC xml

C codesWith Generic API

ExecutableWith target API

C code staticanalysis

perf. And powerestimation

ARM ISS


Demo 3. UML 2.0 –based Modeling

• ESUML : Embedded Software modeling with UML 2.0

Analysis &Design

Verification& Validation

Code Gen. &Optimization

feedback

ESUML Modeler ESUML Statis analyzer ESUML Simulator

Simulation TraceSimulation Trace

Environment Interface Controller Heater

PushButton()

HeaterOn()

Heat()

Heat_OK

DisplayHeat()

Embedded S/W Needs

TargetMPSoC


ESUML

• Model – Driven Application Development

Requirements Capturing PIM Modeling

Model Verification

PSM Modeling

Code Generation

Use-Casemodel

Use-Casemodel

Structural model

Structural model

Behavioral model

Behavioral model

H/W architecture modelH/W architecture model

Static AnalysisStatic Analysis

Model SimulationModel Simulation

CIC codeCIC code

PIM : Platform Independent ModelPSM : Platform Specific Model

Physical modelPhysical model JavacodeJavacode


ESUML Characterisitcs

• Light-Weight Development Methodology

Real-Time Embedded Software

Interaction-based Behavior Modeling

Large-Scale System

Requirement Capturing

PIM and PSM modeling

Analysis

Code Generation

Use Case DiagramClass DiagramInteraction Overview DiagramSequence DiagramHardware Architecture Diagram

Event-Based Modeling

Hierarchical Decomposition

Modeling Guidelines and Tips

ESUML

Concepts Process

Notations Rules


UCDuse case 1

use case 2

use case 3

Scenarios

ESUML Characteristics

• Hierarchical Modeling• Interaction-based Behavior Modeling

IOD of UC3

SDs of X

Global IODs

A B

C

IOD of UC1X

IOD of UC2

Y Global behavioral view SDs of Z SDs of Y

SDs of X


PSM Modeling (To-Be)

• PSM Modeling : Physical Model

ref Door Opened

ref Door Closed

ref Cook Button Pressed

ref Cooking Complete

ref Waiting

DoorClosed

CookButtonPressed

CookButtonPressed

DoorOpened

[ Door. isOpen()==false]

OvenCtrl Beeper

beep()

sd Single Beep

[ Door. isOpen()==true]

TimerExpired

ref Cancel Button PressedCancelButtonPressed

ref DoorOpenedDoorOpened

sd Cook

Add60SecTimer

StartTimer

StopTimer

Add60SecTimer

P11_ActiveMain

P30_PostProcessing

P20_ActiveSub1

P12_ActiveSub2

ARM1 ARM2SRAM1 SRAM2S_SRAM1

AHB_1

AHBtoAHBBridge

AHB2

ARMSub S_SRAM2TFTLCD

Ctrl

DMACtrl

APB1

IntrCtrl AudCodec UARTCtrl

AHBtoAPBBrdige I2CCtrlGPIOCtrlDCTIDCT

RDRAMCtrlARM1 ARM2SRAM1 SRAM2S_SRAM1

AHB_1

AHBtoAHBBridge

AHB2

ARMSub S_SRAM2TFTLCD

Ctrl

DMACtrl

APB1

IntrCtrl AudCodec UARTCtrl

AHBtoAPBBrdige I2CCtrlGPIOCtrlDCTIDCT

RDRAMCtrl

Automatically Recommended Partition +ModelerEnforced Partition


ESUML Code Gen. (To-Be)

• Platform Independent Code Generation

PIM/PSM Models

C

C++

JAVACICCICCode Generator


ESUML Verification and Validation

• Rule-based consistency analyzer– Intra-model, inter-model, action-language

64Rules


ESUML Simulator

MicroOven.XMI File

SimMicroOven.txt

Define Scenario

Simulation result

(1) input

(2) input

Feedback to Modeler

(4) output

(5) save

(5) modify model (if needed)

SimulationEngine

Simulation TraceSimulation Trace

Environment Interface Controller Heater

PushButton()

HeaterOn()

Heat()

Heat_OK

DisplayHeat()

(3) output

CFGGeneration


Contents

• Introduction• Proposed Design Flow• Key Techniques

– PeaCE Modeling– CIC & Generic API– Static Analysis

• Status• Conclusion


PeaCE Model

• Top model: Task model

• Computation task model: Dataflow model– Extended SDF model: SPDF (Synchronous Piggybacked Data Fl

ow) and FRDF (Fractional Rate Data Flow)

• Control task model: FSM model– Hierarchical and concurrent FSM model: fFSM (flexible Finite

State Machine)


PeaCE Project

• PeaCE– Ptolemy extension as a Codesign Environment

based on Ptolemy classic– open-source research platform– Officially released in DAC 2005. (version 1.0)

• PeaCE home page– http://peace.snu.ac.kr/research/peace


Seamless HW/SW Codesign Flow

1 2 3

4 5 6 7 8

Func. Simulation

C-code

PE SelectionPartition & mapping

DB

sched.xml

PE Code generation

Comm. arch DSE

IF Code generation

OS API

Coverification & Prototyping

clustered.xml

Archi.xml

VHDL-codeC-code

Graph Analysis

+

Profiling33333

11

44

35335

36336

37337

36336 37337

38338

38338

TimeCost.xml.

Algorithm spec. Architecturetemplates

22

Block Lib.


Divx Player: Top Model


H.264 Decoder

if-then-else construct


SPDF Model

• Basic SDF Model– A node represents a function block (ex: FIR, DCT) – A block is fireable (executable when it receives all required nu

mber of input samples) – Statically scheduled at compile-time:

sample rate inconsistency, deadlock condition

• Our extensions– Generic message type data -> code generation only– FRDF: fractional rate data flow– SPDF: synchronous piggybacked dataflow


Fractional Rate Dataflow

Schedule : 99(ME,EN)

The macro block is 1/99 of the video frame in its size.

For each execution of the ME node, it consumes a macro-blockfrom the input frame, and produces a macro-block output.

ME1

EN1 1

1/99

1/99 16x16


FRDF (Fractional Rate DF)

Reference Code SDF FRDF

361KB 686KB 225KBBuffer size

5 8 3No. of frameType data

MotionEstimation

MacroBlock

Encoding

VariableLengthCoding

MacroBlock

Decoding

MotionCompensation

WriteTo

Device

ReadFrom

Device

11/99

1 1 1

1111/99

1/99

1 1/99

11/99


Synchronous Piggybacked Dataflow

• Global data structure with limited access• Each data sample is piggybacked with an SU

request• A block updates its internal parameters

depending on applicability of SU request

PD A B C

“gain”= 5

SU

data

piggyback

202nd iteration

“gain”= 10

C

“gain” 10

….

State name 1st iteration


SPDF (Synchronous Piggybacked DF)

• Piggybacking the control variable• Special super-block definition for the body of DC

PB

xc

A B

Mux

y

1 0MUX

A B

x

y

c


fFSM (flexible FSM) Model

• Hierarchical FSM

• Concurrent FSM

A

a

b/v

C Dc/u

B

A BC

BD

a

b/v

b/v b’c/u

AC BC

AD BD

ac’

bc’

a

b

a’c b’cbcac

A B

C D

a

b

c


fFSM + SPDF

start

divx

divxstop

scripts in divx {deliver ui filename divxread filename} {run divx}

stop==1/tostart=1

tostart==1

script in divxstop{stop divx}

text==4exitDivx==1

exit

1

script in exit state{get divx exit exitDivx}

divxrun

divxsuspend

script in divxrun{run divx}

script in divsuspend{suspend divx}

suspend==1

start==1

startsuspend

stoptext

AVIReader

H.263Decoder

MP3Decoder

divx

divxread

exit


Task Model

• Task execution semantics– periodic, sporadic, functional

• Port properties– data rate : static or dynamic – data size : static or variable– port type : queue or buffer

AviReader

H.263decoder

MP3decoder

periodic

Basic Task Block(SDF or FSM)

Task Wrapper

Execution Typeport semantic


CIC Code Generation

• Application tasks– Task in task-level specification model task_name.cic

• Generic API• Partitioning considering data parallelism

– openMP programming

void h263decoder_go (void) {...

l = MQ_RECEIVE("mq0", (char *)(ld_106->rdbfr), 2048); ... # pragma omp parallel for for(i=0; i<99; i++) { //thread_main()

.... } // display the decode frame dither(frame); }


CIC Format

Architecture Task Code

Constraints

Structure

Kernel code_init()_go()

_wrapup()

Generic API Parallel API

Hardware


DivX player CIC xml (Hardware)

<?xml version="1.0" ?> <CIC_XML> <hardware> <processor name="arm926ej-s0"> <index>0</index> <localMem name=“lmap0”> <addr>0x0</addr> <size>0x10000</size> // 64KB </localMem> <sharedMem name=“shmap0”> <addr>0x10000</size> <size>0x40000</size> // 256KB <sharedWith>1</sharedWith> </sharedMem> <OS> <support>TRUE</support> </OS> </processor>

<processor name="arm926ej-s1"> <index>0</index> <localMem name=“lmap1”> <addr>0x0</addr> <size>0x20000</size> // 128KB </localMem> <sharedMem name=“shmap1”> <addr>0x20000</size> <size>0x40000</size> // 256KB <sharedWith>0</sharedWith> </sharedMem> <OS> <support>TRUE</support> </OS> </processor> </hardware>


P1P1 P2P2LocalMem

Shared Mem

TCM0x0

0x10000

shared=NO

0x50000

shared=YES

0x0

0x20000

shared=NO

0x60000

shared=YES

Hardware Architectures

P1P1 P2P2 P3P3

SM1

SM2


DivX player CIC xml (Constraint)

<constraints> <memory>16MB</memory> <power>50mWatt</power> <mode name="default"> <task name="AviReaderI0"> <period>120000</period> <deadline>120000</deadline> <priority>0</priority> <subtask name="arm926ej-s0"> <execTime>186</execTime> </subtask> </task> <task name="H263FRDivxI3"> <period>120000</period> <deadline>120000</deadline>

<priority>0</priority> <subtask name="arm926ej-s0"> <execTime>5035</execTime> </subtask> </task> <task name="MADStreamI5"> <period>120000</period> <deadline>120000</deadline> <priority>0</priority> <subtask name="arm926ej-s0"> <execTime>13</execTime> </subtask> </task> </mode></constraints>


DivX player CIC xml (Structure)

<structure> <mode name="default"> <task name="AviReaderI0"> <subtask name="arm926ej-s0"> <procMap>0</procMap> <fileName>AviReaderI0_arm926ej_s0.cic</fileName> </subtask> </task> <task name="H263FRDivxI3"> <subtask name="arm926ej-s0"> <procMap>0</procMap> <fileName>H263FRDivxI3_arm926ej_s0.cic</fileName> </subtask> </task> <task name="MADStreamI5"> <subtask name="arm926ej-s0"> <procMap>0</procMap> <fileName>MADStreamI5_arm926ej_s0.cic</fileName> </subtask> </task> </mode>

<queue> <name>mq0</name> <src>AviReaderI0</src> <dst>H263FRDivxI3</dst> <size>30000</size> </queue> <queue> <name>mq1</name> <src>AviReaderI0</src> <dst>MADStreamI5</dst> <size>30000</size> </queue></structure></CIC_XML>


Task Code: Structure

• Task kernel code – _init(): before main loop– _go(): in the main loop – _wrapup(): after main loop

(example) h263dec.cic // header file // global declaration and definition // procedure definition h263dec_init() { … }

h263dec_go() { … } h263dec_wrapup() { … }


Generic API

• Generic API– OS-indepednent API– API translator to OS-specific APIs

• Generic API Definition– Abstraction from IEEE POSIX 1003.1-2004– OS services:

• Process, memory, synchronization, file system, networking, real-time, thread, I/O, timer, etc.

– C library functions based on use frequency• stdio.h, stdlib.h, time.h, signal.h


Generic API 정의function header total API Selected API Generic API

C Library Signal, I/O, Time

signal.h 26 2 2stdio.h 65 19 10stdlib.h 68 5 3time.h 30 3 3

OS Services

Processsched.h 8 3 3sys/wait.h 3 1 1unistd.h 79 13 13

Memory stdlib.h 68 5 3

Synchronizationsys/mman.h 14 4 4pthread.h 100 21 12

File systemfcntl.h 5 1 1sys/stat.h 9 2 2unistd.h 79 13 13

IPCmqueue.h 10 8 5semaphore.h 10 8 3sys/mman.h 14 4 4

Network sys/socket.h 18 11 9

Realtimemqueue.h 10 8 5semaphore.h 10 8 3sched.h 8 3 3

Thread pthread.h 100 21 12

I/Ostropts.h 8 1 1sys/select.h 6 1 1poll.h 1 1 1

Timertime.h 30 3 3sys/time.h 4 3 3

Total 　　 18 464 107 77


OS API Translator

• OS API Translator

OS API Translator

OS-Specific C Code

Code Analysis

Transformation

Symbol TableConstruction

CIC’s Task Code APIs, Patterns, and Rules

register

Generic APIs

modify

POSIX APIs

Non-POSIX APIs

TransformationRules

Translation pattern and rules

Pattern mapping - 1-to-1, 1-to-m, n-to-m - initialize/finalize - acquire/releaseparameters - type, number, valueVariable names - naming of additional variables


Airac5

• Static Analyzer for Detecting All Buffer Overrun Errors in C Programs

• Keywords– static program analysis: abstract interpretation – C: ANSI C + somje GNU C extensions– all: detects all buffer-overrun places – statically: at compile-time– always stops: finishes even for infinite-loop programs– modular: separate C files– correct: based on a firm theoretical framework

• SAS’05, BUGS’05• being in use inside Samsung and LG


Airac5’s Performance(1/3)

Linux kernel-2.6.4

Software #Lines Time(sec) #Alarms #Bugs

vmax301.c 246 7.42 1 1

cdc-acm.c 849 17.66 7 2

atkbd.c 944 4.20 3 2

eata-pio.c 984 15.45 16 1

ip6_output.c 1110 47.69 0 0

xfrm_user.c 1201 56.52 4 1

keyboard.c 1256 90.42 31 1

af_inet.c 1273 186.06 12 2

usb-midi.c 2206 53.98 12 4

mptbase.c 6158 206.04 12 1

on P4, 3.2GHz, 4GB



Program(LOC)

Airac5 SWAT#Real Bugs

#Time(sec)

#Alarms

#Alarms

vmax301.c(246) 7.42 1 1 1

xfrm_user.c(1201)

56.52 4 1 1

usb_midi.c(2206) 53.98 12 2 4

atkbd.c(944) 4.20 3 2 2

keyboard.c(1256) 90.42 31 1 1

af_inet.c(1273) 186.06

12 1 2

eata_pio.c(984) 15.45 16 1 1

cdc_acm.c(849) 17.66 7 1 2

ip6_output.c(1110)

47.69 0 1 0

mptbase.c(6158) 206.04

12 1 1

aty128fb.c(2466) 98.49 4 1 1

Comparison with Coverity’s SWAT


GNU applications

Software #Lines Time(sec)

#Alarms #Bugs

tar-1.13 20258 2003 76 1

bison-1.875 25907 7050 30 0

sed-4.0.8 6053 256 5 0

gzip-1.2.4a 7327 330 50 0

grep-2.5.1 9297 1314 24 0


• Statistical technique ranks the alarms• Linear regression statistical analysis• Users see truer alarms first• 4.10% of false alarms mixed up until 50% of true alarms seen


Airac5 Available

• free-trial version• performance data• on-line demo• etc informationqq

ropaswork.com/airac5


Contents



Status (1)

• CIC and Generic API translator– Translator for single processor– To-do

• openMP translator• support on more APIs

• Target specific API libraries– ARM multi-processor platform for now– To-do

• Libraries for a real MPSoC platform


Status (2)

• Virtual Prototyping– 1st version of ARM simulator and debugger – To-do

• Multiprocessor system simulator and source level debugger• OS porting on the ISS

• Performance and power estimation– Function-based estimation for ARM processor– To-do

• system-level estimation, other processors


Conclusion

• HOPES is a newly launched project to make a embedded software development environment for MPSoCs– Support of diverse models– Target independent environment + target specific libraries– 3-phase verification: model-level, C code static analysis, run-t

ime simulation– Integration of software modules at various stages

– http://peace.snu.ac.kr/hopes (sorry, only Korean for now)


Thank you !

Soonhoi Ha, SNU 11 Feb. 14, 2006 HOPES: Embedded Software Development Environment for MPSoC Feb. 14,...

Documents

Transcript of Soonhoi Ha, SNU 11 Feb. 14, 2006 HOPES: Embedded Software Development Environment for MPSoC Feb. 14,...