multicore digital signal processing - etsist.upm.es · crc/ turbo coding channel coding symbol...

1

Multicore DSPs – Karol Desnos ([email protected])

MULTICORE DIGITAL

SIGNAL PROCESSING

Karol Desnos – [email protected]

Slides from M. Pelcat, K. Desnos,

J.-F. Nezan, D. Ménard,

M. Raulet, J. Gorin

2


Previously

on MDSPs

3


Previously on MDSPs

Source: ATEME

HEVC Encoder Bloc description

4


Previously on MDSPs

HEVC Decoder Bloc description

VLC

Decoding

Dequantization

Inverse Transform

Filtering

Source: Hervé Yviquel

5


Application is naturally described with blocs

Applications for MDSPs

OFDMA

Encoding

Downlink Data Encoding

data

(bits)

Multi-Antenna

PrecodingModulation

Interleaving/

Scrambling

Rate

Matching

CRC/Turbo

Coding

Channel Coding Symbol Processing

Turbo

Decoding/CRC

Uplink Data Decoding

data

(bits)

Rate

Dematching

/HARQ

Descrambling/

Deinterleaving

Channel Decoding

DemodulationSC-FDMA Decoding/

Multi-Antenna Equalization

Symbol Processing

Channel Estimation

6


Grail of Heterogeneous MPSoCs Programming

Previously on MDSPs

Multicore Compiler

Simulator

+ Debugger

+ Profiler

Algorithm

Architecture

Portable Multicore Program

PE

Main

Proc.

Main

Proc.

Main

Proc.

Main

Proc.

PE PE PE PE

Peripherals

Main

Memory

Multicore Runtime

7


• Lecture 1 – Maxime Pelcat

• Introduction to the course

• Applications for MDSPs

• Lecture 2 – Karol Desnos

• Languages and MoCs

• Programming MPSoCs

• Dataflow MoCs

• Lecture 3 – Maxime Pelcat

• Hardware Architectures

Course Outline

8



• Theoretical Bounds

• Mapping/Scheduling Strategies


• Lab Session

Course Outline

9


Languages and Models of

Computations

10


What is a Model of Computation (MoC) ?

• A set of operational elements that can be composed to

describe the behavior of an application.

Semantics of the MoC

• Interface between: • Computer science

• Mathematical domain

• A MoC is not a language !

Languages and MoCs

11


What is a language ?

• A set of textual/graphical symbols that can be assembled

respecting a well defined grammar to specify the behavior

of a program

Syntax of the language

• Interface between: • Programmer

• Machine (through a compiler)

• A language implements one or several MoCs

Languages and MoCs

12


Semantics and Syntax

• Among the 10 most used languages,

• all 10 are imperative,

• 7 are object-oriented.

• Other semantics exist. A MoC specifies semantics

independently from a language syntax

Languages and MoCs

13


Semantics and Syntax

• UML implements object-oriented semantics

• C++ implements object-oriented semantics

• They share semantics but not syntax

Languages and MoCs

14


A few MoCs

• Finite State Machine MoCs • Semantics

• States

• Transitions based on conditions

• MoC of imperative languages (C/C++/Java…)

• Computation is executed on transitions

• Representing the behavior of a Turing machine

Languages and MoCs

Start

End

pts = @TAB

(pts) = 0

pts = pts + 1

End

TAB ?

yes

no

15


A few MoCs

• Discrete Event MoCs • Modules react to events by producing events

• Events tagged in time, i.e. the time at which events are consumed

and produced is essential and is used to model system behavior

• Modules share a clock

• Used to model HDL behavior

• Functional MoCs and lambda calculus • No state, a program returns the result of composed mathematical

functions: result = f ס g ס h(inputs)

• Based on lambda calculus

• Haskell, Caml, Scheme, XSLT

• Functional languages are examples of declarative languages

Languages and MoCs

16


A few MoCs

• Petri Nets • Close to state machines but

• able to represent parallelism

• Operations are done on transitions

• Synchronous MoCs • Like in Discrete Events, modules react to events

• by producing new events

• Contrary to Discrete Events, time is not explicit and

• only the simultaneity of events and causality are important

• Language examples: Signal, Esterel, Lustre…

Languages and MoCs

Pro

cessin

g th

read 1

Pro

cessin

g th

read 2

OP1 OP2

Loo

p

Loo

p

17


A few MoCs

• Process Network MoCs

• concurrent and independent modules (processes) communicate

ordered tokens (data quanta) through First-In, First-Out (FIFO)

channels

• Include dataflow process networks

• Untimed models: the time is abstracted

What we naturally used to describe MPEG HEVC and 3GPP LTE

processing

Languages and MoCs

18


Programming MPSoCs

19


• Parallelisms

• Programming Coarse-Grain Parallelism

Programming MPSoCs

20


Parallelisms

• Granularity of parallelism

• Amount of computation of an atomic operation relatively to

the overhead to manage/execute several operations concurrently

Programming MPSoCs

Operation size

Overhead

small

large

small large

Perfect world

21


Parallelisms

• Single-Instruction Multiple-Data (SIMD)

• Primitive operations (+ - * / >> min max …)

• Low/No cost when code is suitable (data dependencies)

• Explicitly specified with intrinsic / implicitly supported by HW

Programming MPSoCs

Overhead

Operation size

small

large

small large

Perfect world SIMD

22


Parallelisms

• Software Pipelining

• Loop optimization: start a new iteration before the previous end

• Low cost when code is suitable (data dependencies)

• Automatically performed by compilers (-O3)

Programming MPSoCs

Overhead

Operation size

small

large

small large

Perfect world SIMD

SW Pipe.

23


Parallelisms

• Instruction Level Parallelism

• Out-of-order execution

• SIMD

• Register Renaming

Programming MPSoCs

Overhead

Operation size

small

large

small large

Perfect world

ILP

SIMD

SW Pipe.

• Branch prediction

• SW pipelining

• …

24


Parallelisms

• Threads

• “Independent” sequence of instructions with private context

• Moderate to high cost => quasi-total hardware abstraction

• Complex to program: deadlocks, preemption, …

Programming MPSoCs

Overhead

Operation size

small

large

small large

Perfect world

ILP

SIMD

SW Pipe.

Threads

25


Parallelisms

• Cores + on-chip interconnect

• Each core runs a bare-metal program (no OS).

• HW-supported means of communication (Interrupt, DMA, Packet, ...)

Programming MPSoCs

Overhead

Operation size

small

large

small large

Perfect world

ILP

SIMD

SW Pipe.

Cores + interconnect

Threads

26


Parallelisms

• Chips + LAN or WAN

• Independent servers/computers communicating through a network

• Very high cost (communication, distributed memory)

• High Performance Computing

Programming MPSoCs

Overhead

Operation size

small

large

small large

Perfect world

ILP

SIMD

SW Pipe.


Chips + L/WAN

Threads

27


Parallelisms

• For MPSoCs: Fine and Coarse-grain parallelisms

• ILP and SW Pipelining: Compiler (& crazy developers)

• Threads and Interconnect: Developers and/or optimization tools

Programming MPSoCs

Overhead

Operation size

small

large

small large

Perfect world

ILP

SIMD

SW Pipe.


Chips + L/WAN

Threads

28


• Parallelisms

• Programming Coarse-Grain Parallelism

Programming MPSoCs

29


Programming Coarse-Grain Parallelism

• Manually, using threads

• “Independent” sequence of instructions with private context

• Usually preemptible

• Number of thread ≈ number of cores • Otherwise, context switching rapidly decreases application performance

=> Test in Laboratories

• Solutions: • posix thread (linux/windows)

• win32 API

• Boost Threads

• « Idle » threads on SysBios (TI DSP)

• …

Programming MPSoCs

30



• Manually, using tasks

• Short independent sequence of instructions with private context

• Not preemptible

• Number of tasks >> number of cores • Finer granularity than threads.

• Using many tasks ensures load-balancing

• Solutions: • Multicore Task Management API (MTAPI)

• OpenEvent Machine

• StarSs

• …

Programming MPSoCs

31



• Semi-automated, using OpenMP

• Add #pragma omp instruction to parallelize sections of C code

• Compiler optimization automatically create threads/tasks

• Shared-Memory targets !

• Supported on some MPSoCs (including c6x DSPs !!)

• Examples:

• Split loop iterations among threads #pragma omp parallel for

for(i=0; i< L; i++) { … }

• Execution barrier for several threads #pragma omp parallel for

for(i=0; i< L; i++) {

tab[i] = 0; // init

#pragma omp barrier

tab[i] = … // work on the tab.

}

Programming MPSoCs

32



And so many other solutions:

• Message-Passing Interface (MPI), Multicore

Communication API (MCAPI)

• Communication protocol for parallel computing abstracting architecture

concerns.

• OpenCL, CUDA

• GPGPU oriented

• Cilk, Go Language, Threading Building Blocks, …

• Threads without headaches

• Dataflow programming

Programming MPSoCs

33


Dataflow MoCs

34


• Dataflow Models of Computations

• Properties of Dataflow MoCs

• Dataflow Frameworks

• Laboratories Example

Dataflow MoCs

35


Dataflow MoCs

• Compilation for block diagrams

• Accessible to persons with no knowledge of programming languages

• 1961: BLODI compiler • 30 basic DSP blocks (Adder, Filter, Quantizer, …)

• 1 Punched card for each block.

Dataflow MoCs

Source: J.L. Kelly, C. Lochbaum, and V.A. Vyssotsky. A block diagram compiler. Bell System Technical Journal, 1961

36


Dataflow MoCs

• Kahn Process Network (KPN)

• Actors and Data Ports

• First-In, First-Out (FIFO) Queues

Dataflow MoCs

B

A C D E

37


Dataflow MoCs


• Production/consumption rates of an process can change at

each firing, but:

• Reading from a FIFO is a blocking operation

• This ensures the determinism of the application

i.e. Same input stream => Same output stream

Dataflow MoCs

B

A C D E

38


Dataflow MoCs


• Process example

• Process Sort separates odd and even values.

Process sort(in int i, out int even, out int odd) {

int value = i.read(); // Blocking

if( value % 2 == 0)

even.write(value);

else

odd.write(value);

}

Dataflow MoCs

3 2 6 5 1 Sort

i even odd

39


Dataflow MoCs


• Determinism example

• Process Interleave output value alternatively taken from its inputs.

Process inteleave(in int a, in int b, out int o) {

static bool = true;

int value = (bool)? a.read() : b.read();

bool = !bool;

o.write(value);

}

Dataflow MoCs

Sort i even

odd 3

2 6

5 1

blocked Interleave a o b

40


Dataflow MoCs

• Dataflow Process Network (DPN)

• Firing rules of an actor can change at each firing

• Firing rules depend on: • State of actors

• Time / Randomness…

• Number/value of available tokens

Dataflow MoCs

B

A C D E

Source: E.A. Lee and T.M. Parks. Dataflow process networks. Proceedings of the IEEE, 1995.

x2

Model is non-deterministic

41


Dataflow MoCs


• Non-Determinism example

• Process Interleave output value alternatively taken from its inputs.

• Time-out on FIFO read !


static bool = true;


bool = !bool;

if(no_timeout) o.write(value);

}

Dataflow MoCs

3

2 6

5 1

Timeout Interleave a o b

42


Dataflow MoCs


• Non-Determinism example

• Same inputs with different arrival time produces different output


static bool = true;


bool = !bool;

if(no_timeout) o.write(value);

}

Dataflow MoCs

3

2 6

5 1

Timeout Interleave a o b

43






Dataflow MoCs

44


Properties

• DPN and KPN are Dynamic Dataflow MoCs

• It is not possible to find a fixed sequence of actor firing that can be

repeatedly executed for any input sequence.

• It is not possible to guarantee that an application will not deadlock

for any input sequence.

• It is not possible to guarantee that an application will always require

a finite amount of memory for its execution (tokens can accumulate

indefinitely in unbounded FIFOs)

Dataflow MoCs

45


Properties

• Synchronous Dataflow (SDF)

• Stateless actors and data ports with fixed exchange rates

• First-In, First-Out (FIFO) Queues

• Delays

Dataflow MoCs

2

2

1

1

1 2 1 1

B

A C D E

Source: E. Lee and D. Messerschmitt, “Synchronous data flow”, Proceedings of the IEEE, 1987.

x2

46


Properties


• SDF is a static model: It cannot be used to model all applications

• Production/consumption rates are fixed

• It is possible to find a fixed sequence of actor firing that

can be repeatedly executed for any input sequence, with a

finite memory footprint and no deadlock.

Dataflow MoCs

47


Properties


• Data-driven execution: An actor is fired when its input FIFOs contain

enough data-tokens.

Dataflow MoCs

B

A C D E

Core1 A B C C D E

2

2

1

1

1 2 1 1


48


Properties


• Parallelisms: / / /

Dataflow MoCs


2

2

1

1

1 2 1 1

B

A C D E

Core1

Core2

Core3

x2

A B C C E A C +1 +1

B C +1 +1

D E +1 +1

D

Task parallelism Data parallelism Pipeline parallelism Internal parallelism

Data Pipeline Internal Task

49


Properties

• There exists many other dataflow MoCs.

• What is the difference between them ?

• Decidability • Application schedulability can be guaranteed at compile-time

• Determinism • Same inputs => Same outputs (regardless of time, randomness, or

implementation)

• Compositionality • Properties of a graph (topology, firing rules of actors, repetition

vector, …) are independent from the internal implementation of the

actors that compose it.

Dataflow MoCs

50


Properties

• Reconfigurability • Firing rules of actors can change dynamically during the execution of

the application, depending on inputs. DPN and KPN are

reconfigurable, SDF is not.

• Predictability (related to reconfigurability) • Amount of time between the reconfiguration of firing rules of an actor

and its actual firing. (Relative)

• Conciseness (or succinctness)* • Ability of a MoC to model an application with a small number of

actors. (Relative)

Dataflow MoCs

51


Properties

• Analyzability • Availability of analysis techniques to characterize an application

graph (eg. worst-case latency, memory requirements). (Relative)

• Expressivity • Complexity of application behavior that can be described.

Ex. DPN and KPN are Turing-Complete

SDF is not.

Dataflow MoCs

52


Properties • About compositionality: A non-compositional hierarchical model

Dataflow MoCs

1 3 h B 3 1 A

1 N C 1 1 D 1 1

B

C2

C3

C4

C5

C6

C1 D1

D2

D3

A1

With N=2

A2

h1

h2

h3

With N=1

B C2

C3

C1 D1

D2

D3

A

53


Properties • About compositionality: The IBSDF compositional model

Dataflow MoCs

1 3 h B 3 1 A

1 N C 1 1 D 1 1

B

C2

C3

C4

C5

C6

C1 D1

D2

D3

A

With N=2

h1

h2

h3

h1

h2

h3

With N=1

B C2

C3

C1 D1

D2

D3

A

54


Properties • About predictability and reconfigurability:

The more dynamism, the less predictability

Dataflow MoCs

B 1 3 A 2 C 2 SDF

x1 x2 x3

DPN A C B x2 x1 x1 x2 x3 x1

PSDF

B 1 A 2 C 2

body

subinit Set_p

init

x2 x3 x1

p:=3

p

Compile Time

After Subinit

After execution

55


Properties

• There exists many other dataflow MoCs.

Dataflow MoCs

Feature SDF ADF IBSDF DSSF PSDF PiSDF SADF SPDF DPN KPN

Expressivity Low Med. Turing

Hierarchical X X X X

Compositional X X X

Reconfigurable X X X X X X

Statically schedulable X X X X

Decidable X X X X (X) (X) X (X)

Variable rates X X X X X X X

Non-determinism X X X

SDF: Synchronous Dataflow

ADF: Affine Dataflow

IBSDF: Interface-Based Dataflow

DSSF: Deterministic SDF with Shared Fifos

PSDF: Parameterized SDF

PiSDF Parameterized and Interfaced SDF

SADF: Scenario-Aware Dataflow

SPDF: Schedulable Parametric Dataflow

DPN: Dataflow Process Network

KPN: Kahn Process Network

56


Properties

• It is important to choose the appropriate MoC before

developing each part of an application.

• LTE example:

Dataflow MoCs

X Number of root sequences

DFT Filter

Select preamble subcarriers

Correlate with root sequence

IDFT Power

Noise floor estimation Peak Search

X Root Sequences X Reception Antennas X Preambles

Detected Users Timing Advance

X Reception Antennas X Preambles

Static

57


Properties

• It is important to choose the appropriate MoC before

developing each part of an application.

• LTE example:

X Users X Transport Blocks

Blocks

X Users X Code Blocks

X Users

X Reception Antennas

FFT Prepare

Demap Control/Data

Decode Control

X symbols (14) X Reception Antennas

Data of Each user

Estimate Channel

Equalize Multiple Antennas

Symbol and Bit Processing

Code Block Processing

Transport Block Processing

Static

Static

Param.

Param.

Param.

Dataflow MoCs

58






Dataflow MoCs

59


Dataflow MoCs

Dataflow Framework

• CAL and DDF

• CAL = CAL Actor Language, J. Ecker and J. Janneck

• Implements the DPN MoC

• RVC-CAL: Normalized in MPEG

actor decimate () I ==> O :

action I:[ a , b ] ==> O:[ a ] end

end Consommation Production

décimate

I O

60


Dataflow MoCs

Dataflow Framework

• CAL and DDF

• Via classification, a CAL actor can be transformed into SDF, CSDF in order to

gain predictability

• The Orcc compiler, developed at IETR, is a compiler for CAL

• From CAL, both HW and SW solutions can be generated

• Supports real-life applications such as HEVC !

• Examples

actor decimate () I ==> O:

a: action [ x ] ==> [ x ] end

b: action [ x ] ==> end

schedule fsm s1:

s1 ( a ) --> s2;

s2 ( b ) --> s1;

end

end

Consommation Production

actor posValue() I ==> O:

pos: action I: [ x ] ==> O: [ x ] guard x >= 0

end

neg: action I: [ x ] ==> O:

end

priority

pos > neg;

end

end

actor decimate () I ==> O:

int s := 0;

action [ x ] ==> [ x ] guard s = 0

do s := 1; end

action [ x ] ==> guard s = 1

do s := 0; end

end

61


Dataflow MoCs

Dataflow Framework

• Ptolemy (by UC Berkeley)

• First tool

• Support for composition of DPN, PSDF and SDF MoCs

• Lightweight Dataflow Environment (LIDE) (by U. of Maryland)

• First tool

• Support for composition of DPN, PSDF and SDF MoCs

• MAPS Compiler (by RWTH Aachen U.)

• KPN and C language

• Support C6678

• SynDEx (by INRIA)

• Custom Dataflow MoC (SDF-like) an C Code

• Similar to Preesm (but closed source)

62


Dataflow MoCs

Dataflow Framework

• OpenDF (by EPFL)

• CAL Code compiler

• Targeting reconfigurable hardware and multicore systems.

• SDF For Free (SDF3) (by Technische Universiteit Eindhoven)

• SDF, CSDF and SADF

• Focusing on analysis of SDF application and generation of random graphs

• Canals (Abo Akademi)

• A language and its compiler where schedulers are specified explicitely at

each level of hierarchy

• AccessCore Studio (by Kalray)

• Programming for Many-Core MPSoC with 256 cores.

63



And so many other solutions

Programming MPSoCs

Deduce parallelism from Models of Computation

Extract parallelism from code

High Performance Computing

Embedded Systems

Ptolemy II

PREESM

OpenMP

Matlab + Simulink

OpenCL

CAL

SystemC

Simulation oriented

Execution oriented

64






Dataflow MoCs

65


Dataflow MoCs

Laboratories Example

• PiMM Dataflow MoC • Use the IBSDF subset

+ static param

Channel Decoding

Converge

Config NbUE

Actors

Data Ports

FIFO Queues

SDF

Hierarchy

Interfaces

IBSDF

𝝅SDF

Parameters

Param. Ports

Dependencies

Dynamic Config.

PiMM

PUSCH

NbUE

mac

an

ten

na

sin

k

MaxCB

NbUE 1

NbUE*MaxCB MaxCB

Source: J. Piat, S.S. Bhattacharyya, and M. Raulet. Interface-based hierarchy for synchronous dataflow graphs. SiPS 2009

66


Dataflow MoCs


• Sobel Application

• Sequential version

Process

67


Dataflow MoCs



• Parallel version

Split

Process

Process

Process

Merge

Convolution of a 3x3 Matrix with pixels of the input image

Each slice has a 1-line overlap with the next and previous slices

68


Dataflow MoCs



• Demo time !

multicore digital signal processing - etsist.upm.es · crc/ turbo coding channel coding symbol...

Documents