An Architecture Exploration Framework for Embedded DSP Systems

1
An Architecture Exploration Framework for
Embedded DSP Systems

• Presenter Ahmed Elhossini
• ENG6530- Reconfigurable Computing
• School of Engineering, University of Guelph

2
Outlines

• Motivations Problem Definition
• Background
• Framework Overview
• Searching the design Space
• Evaluation Validation
• Architecture Exploration Flow
• Conclusions
• Present Work

3
Motivations

• The emerge of the concept of System-On-Chip
  introduced many challenges for the designer.
• A tool that help the designer to make an
  estimation of the design requirements is required.

4
Problem Definition

• Given a software application, explore the design
  space formed by the application and a core
  library to define
• Hardware architecture
• Computation (processors , cores, etc.)
• Communication (Bus interfaces, FIFOs, etc.)
• Partitioning/Mapping

5
Problem Definition

• Find Mi, Ci , Bi j to meat Ai of Si
• The problem is formulated as a multi-objective
  optimization problem.

6
Sample SOE Architecture
7
Architecture Exploration
8
Outlines

• Motivations Problem Definition
• Background
• Framework Overview
• Searching the design Space
• Evaluation Validation
• Architecture Exploration Flow
• Conclusions
• Present Work

9
Methodologies

• Orthogonalization of concerns is an important
  concept in architecture exploration.
• Separation of concerns deals with the following
• Separating function and architecture.
• Separating communication and computation.

10
The Y Chart
11
Requirements Challenges of Architecture
Exploration

• Modeling
• Application Modeling
• Architecture Modeling
• Mapping
• Performance Measurement
• System Update (Searching the Design Space)
• Software
• Architecture
• Mapping

12
Architecture Exploration
13
Architecture Exploration Support

• Both architecture and application need to be
  represented in some form to model different
  characteristics of both.
• Different representations at different
  abstraction levels are used to model applications
  and architectures.
• Co-simulation of both models is used to evaluate
  different implementations.

14
Architecture Modelling

• Architecture Description Languages (ADL) are used
  to model architecture at a high level of
  abstraction.
• Hardware Description Languages (HDL) are used to
  model the architecture at different levels of
  abstractions.
• Instruction Set Models are used to model the
  behaviour of the processor.
• When a micro-architecture templates is used as
  the implementation platform models and tools for
  simulation are available.

15
Application Modelling

• Applications are usually specified in high level
  language. Compilers and interpreters are
  required.
• Kahn Process Networks (KPN) model the application
  through concurrent processes communication using
  FIFO, unbounded, uni-directional, point to point
  channels.
• Ptolemy framework provide a good environment to
  model different types of application with
  different computation models.
• Directed Task Graph.

16
Evaluation Techniques
17
Evaluation Techniques

• Accurate Simulation of the System
• Give accurate evaluation with the cost of long
  evaluation time.
• Trace Driven Simulation
• An initial program run and extracts all memory
  accesses and store them in a trace which is used
  for performance estimation.
• It is more efficient in the estimation of the
  performance of the memory sub-system.
• Statistical Simulation
• The statistical simulator does not execute the
  program in a precise order, but instead it
  simulate a statistical profile of the program.

18
Evaluation Techniques

• Analytical Evaluation
• Analytical model is developed for different
  components. This model are used for performance
  estimation.
• The accuracy of analytical evaluation depends on
  the accuracy of the developed analytical model.
• Intelligent Approaches
• Some research is directed toward the use of
  intelligent approaches for the evaluation of
  embedded system.
• Fuzzy Logic and Neural Networks are an example of
  such methods.

19
Searching the Design Space
20
Searching The Design Space

• Exhaustive search covers all possible solutions
  in the design space.
• It has the cost of long search time.
• Local Search covers a portion of the design space
  to find a local sub-optimal solution.
• Advantage short search time
• Disadvantage high possibility of falling into
  local minima.

21
Mathematical Meta-Heuristics

• A meta-heuristic approach based on genetic
  algorithms and particle swarm optimization was
  developed.
• This approach is designed for multi-objective
  optimization.
• This approach was verified using a set of
  multi-objective problems found in the literature.
• Results was prepared for publication in
  evolutionary computation journal.

22
Outlines

• Motivations Problem Definition
• Background
• Framework Overview
• Searching the design Space
• Evaluation Validation
• Architecture Exploration Flow
• Conclusions
• Present Work

23
Framework Overview
24
Architecture Exploration Flow

• Input Application Model
• Output Near Optimal Hardware Architecture to
  implement the given application.
• Target
• Maximize the performance
• Minimize the power consumption
• Minimize the area (resources)
• Maximize the flexibility

25
Application Model (Directed Task Graph)
26
Chromosome Representation

• The candidate architecture is represented as a
  vector of integers.

27
Outlines

• Motivations Problem Definition
• Background
• Framework Overview
• Searching the design Space
• Evaluation Validation
• Architecture Exploration Flow
• Conclusions
• Present Work

28
Genetic Algorithms

• A meta-heuristic approach.
• Inspired by the natural selection and evolution
  of living species.
• Based on three basic operations
• Selection
• Crossover
• Mutation
• Problem Representation - Individuals

29
Genetic Algorithms
30
Genetic Algorithms

• Good exploration capabilities.
• Exploiting capabilities are low.
• The selection operation depends on the fitness of
  each individual.

31
Particle Swarm Optimization (PSO)

• PSO is a robust stochastic optimization technique
  based on the movement and intelligence of swarms.
• PSO applies the concept of social interaction to
  problem solving.
• It was developed in 1995 by James Kennedy
  (social-psychologist) and Russell Eberhart
  (electrical engineer).
• It uses a number of agents (particles) that
  constitute a swarm moving around in the search
  space looking for the best solution.
• Each particle is treated as a point in a
  N-dimensional space which adjusts its flying
  according to its own flying experience as well as
  the flying experience of other particles.

32
Particle Swarm Optimization (PSO)
y
x
sk current searching point.

sk1 modified searching point.

vk current velocity.

vk1 modified
velocity.

vpbest velocity based on pbest.

vgbest velocity based on gbest
33
Initialization. Positions and velocities
http//www.cems.uwe.ac.uk/jsmith/ci/pso/PSO20min
i20tutorial.ppt
34
Neighbourhoods
geographical
social
http//www.cems.uwe.ac.uk/jsmith/ci/pso/PSO20min
i20tutorial.ppt
35
Particle Swarm Optimization

• Good Exploiting Capabilities
• Exploration of PSO depends on the particle
  movement parameters and the design space.
• The selection of gbest, and pbest depends on the
  fitness of the different particles.

36
Hybrid GA-PSO

• Combines the advantages of both PSO and GA.
• Three Hybrid forms are introduced and tested.
• A common representation was used for both
  algorithms.

37
Hybrid Representation
38
Hybrid GA-PSO-1

• In this algorithm both PSO and GA operations are
  used in the same time.
• This combination allows gaining the advantages of
  both algorithms.
• PSO operations can be used to perform local
  search at each solution.

39
Hybrid GA-PSO-2

• In this combination GA is used to explore the
  design space in the first phase of the search.
• PSO is then used to exploit the solutions found
  by GA.

40
Hybrid GA-PSO-3

• PSO is used for exploration.
• GA is used to exploit the search.

41
Multi-objective optimization

• Finding the optimum solution that meet different
  conflicting constraints, by varying different
  parameters.
• Exact solution could be found by exhaustive
  search but this could be very time consuming.
• Exhaustive search speeded up by the study of
  parameters dependency and clustering of
  parameters.

42
Multi-Objective Optimization (Pareto front)

• Defines the set of solutions that covers all the
  trade-off of the different objectives.
• A solution is said to be pareto optimal if it
  part of the pareto front.

43
Multi-objective optimization

• Heuristic methods could be used to find a pareto
  optimal solution.
• Several heuristic methods exist for
  multi-objective optimization.
• Random Search Pareto (RSP).
• Pareto Simulated Annealing (PSA).
• Pareto Reactive Tabu Search (PRTS).
• Genetic Algorithms.

44
Multi-objective Genetic Algorithms

• Two widely used Multi-objective genetic
  algorithms
• Strength Pareto Evolutionary Algorithm (SPEA).
• Non-dominant Sorting Genetic Algorithm (NSGA).
• SPEA uses a fine grained but and expensive
  computation schema.
• NSGA uses a coarse grained and less expensive
  computation schema.

45
Multi-Objective Optimization
46
Strength Pareto Algorithm

• Fitness value is assigned to each individual.
• This value represent the strength of the
  individual in the Multi-Objective space
• How many individuals it dominates.
• The density of solutions near that individual.
• External Archive is used to store the best
  individuals found so far.
• Strength Pareto algorithm is used for both GA and
  PSO

47
Strength Pareto Algorithm
48
Testing The Meta-Heuristics Search Engine

• Standard Benchmarks for Multi-objective
  optimization are used.
• The search engine was also used to solve the
  scheduling problem (John Huisman)
• Task Scheduler used to estimate the number of
  cycles required for a specific application.
• The search engine used to search the design space
  for an optimal schedule.
• The Task Scheduler is used for performance
  estimation

49
Scheduling for Embedded Systems
50
Testing The search Engine
51
Outlines

• Motivations Problem Definition
• Background
• Framework Overview
• Searching the Design Space
• Evaluation Validation
• Architecture Exploration Flow
• Conclusions
• Present Work

52
Core Library

• The core library contains information about
• processing cores
• communication channels and buses
• Components are defined using the following model

53
Performance Estimation

• Performance estimation is the task of estimating
  the performance measures for a given
  architecture.
• This estimation is based on the existence of
  different models of the cores forming up the
  architecture (power, performance , ext.)
• Models provided by core vendors are based on
  different standards.
• Each vendor provides a different estimation tool
  at a different level of abstraction.

54
Performance Estimation
55
Modeling of Embedded DSP cores

• Each architecture is formed from a different set
  of cores and configurations.
• Different benchmarks are available.
• Each benchmark/architecture combination will be
  simulated (Cycle accurate/RTL) and performance
  measures will be extracted for each combination.
• For each core its own performance measures for
  each architecture will be extracted.
• For each core, the extracted performance measures
  will be used to create a model for that core.
• ANN will be used to build the model from the
  extracted information.

56
Analytical Model
57
Modeling of Embedded DSP cores
58
Artificial neurons
Neurons work by processing information. They
receive and provide information in form of spikes.
x1 x2 x3 xn-1 xn
w1
Output
w2
Inputs
y
w3
.
.
.
wn-1
wn
The McCullogh-Pitts model
59
Artificial neurons
Nonlinear generalization of the McCullogh-Pitts
neuron
y is the neurons output, x is the vector of
inputs, and w is the vector of synaptic
weights. Examples
sigmoidal neuron Gaussian neuron
60
Artificial neural networks
Hidden Layer
Inputs Layer
Output Layer
An artificial neural network is composed of many
artificial neurons that are linked together
according to a specific network architecture. The
objective of the neural network is to transform
the inputs into meaningful outputs.
61
Neural network tasks

• control
• classification
• prediction
• approximation

These can be reformulated in general as FUNCTION
APPROXIMATION tasks.
Approximation given a set of values of a
function g(x) build a neural network that
approximates the g(x) values for any input x.
62
ANN Modeling Initial Results (Linear)
63
ANN Modeling Initial Results (Linear)
64
Modeling Example
65
Outlines

• Motivations Problem Definition
• Background
• Framework Overview
• Searching the Design Space
• Evaluation Validation
• Architecture Exploration Flow
• Conclusions
• Present Work

66
Architecture Exploration Flow

• Manually create a call graph from profiling
  results.
• This graph contains information about different
  software blocks.
• It also contains the relation between different
  blocks.

67
VTune Profiling Results
68
Architecture Exploration Flow

• Using a Graph Editor call graphs could be
  modified and user constraints could be added.
• Graph information are stored into two different
  formats.

69
Graph Editor
70
Graph Editor
71
Architecture Exploration Flow

• The Library Editor module enables adding,
  removing and modifying different components in
  the working library.
• Library information are stored into two different
  files.

72
Library Editor
73
Architecture Exploration Flow

• The GA-PSA modules used the graph information,
  core library, and a parameter file to search the
  design space of the problem.

74
Results File
75
Architecture Exploration Flow

• The results obtained by searching the design
  space is decoded using the graph and library
  information.
• Display Result modules is used to display the
  results.

76
Generate Different Results
77
Generate Different Results
78
Outlines

• Motivations Problem Definition
• Background
• Framework Overview
• Searching the Design Space
• Evaluation Validation
• Architecture Exploration Flow
• Conclusions
• Present Work

79
Conclusions

• Architecture Exploration is Important with the
  new trends of Embedded Systems Design.
• Many challenges face the designer during
  architecture exploration
• Modeling of both architecture and application
• Searching the design space
• Evaluation of generated architectures.
• Meta-Heuristic methods are efficient for
  searching the design space in architecture
  exploration.
• Analytical and statistical methods for the
  evaluation of embedded systems give a good
  performance estimation in a reasonable time.

80
Conclusions

• A framework is introduced for architecture
  exploration.
• The flow starts with modeling a DSP application
  and ends with a set of architecture to implement
  the application.
• The quality of the framework was verified as
  follows
• The performance of the search engine was verified
  by testing known benchmarks that covers different
  level of complexities.
• The use of ANN to model different cores ensures
  the accuracy of the performance estimation.

81
Conclusions

• A hybrid PSO-GA algorithm based on the strength
  pareto algorithm is used to search the design
  space.
• The evaluation of the solutions was improved
  using different techniques
• The estimation of the performance is now based on
  the scheduling of tasks between different
  processors.
• ANN based power estimator is being developed.
• Analytical model is combined with the two
  previous techniques.
• A GUI has been developed to enhance the
  user-framework interaction.

82
Thank You!!!

• Questions ?