Fabrizio Ferrandi

1
hArtes SW Partitioning overview

• Fabrizio Ferrandi
• Politecnico di Milano
• CASTNESS'08 15-18 January 2008 - ROMA - Italy

2
Summary

• Objectives
• Task partitioning
• C based descriptions
• The PandA framework
• Preliminary results
• Conclusions

3
hArtes partitioning objectives

• An important step in the hArtes design space
  exploration process is the identification of a
  good partitioning of the given application
• identifying fragments of parallel code on which
  transformations and metric analysis can be
  performed
• starting from C based application coming from
  advanced audio and video systems that support the
  next-generation of communication and
  entertainment facilities
• targeting a combination of embedded processors,
  digital signal processing and reconfigurable
  hardware.

4
hArtes concurrency model

• hArtes applications characteristics
• Data intensive
• Control intensive
• Parallelism granularity
• Fine grain (basic blocks)
• Coarse grain (tasks)
• Specifications
• Sequential programs
• Cuncurrency extracted
• Explicit Concurrency with split/joint barriers

5
Detection of clusters/tasks

• The detection of clusters of operations connected
  by data exchanges
• Uses dataflow analysis based on Control and data
  Dependence Graphs (PDG)
• Two strategies
• Top-down
• starting from high-level clusters of operations
  specified by the system designers (the set of
  functions) decompose each function in sub-tasks
  that are connected by minimal sets of data
  exchanges.
• Bottom-up
• starting from minimum size clusters of operations
  (i.e., the individual instructions as specified
  by the C intermediate representation), clusterise
  instructions along the heavier communications.

6
Task granularity

• The actual size of the tasks can be controlled
• exploiting replication of the operations to
  obtain more parallelism
• exploiting loop based transformations
  fusion/fission, loop unrolling, .

7
Task mapping goal

• Goal identify the best trade-off in terms of
  hardware and software tasks to satisfy designers
  constraints
• Parameters to be considered
• Platform architecture
• Performance required by the application
• Reconfigurable area available
• Profiling information starting from performance
  costs evaluated on each specific platform
  component

8
Task mapping how

• The evaluation of the allocation of the modules
  onto the different components of the platform is
  performed as a step-by-step process.
• Here the focus is either
• on traditional reinforcement learning algorithms
  based on dynamic programming, such as Q-Learning,
  TD (lambda),
• or on more advanced techniques that exploit both
  reinforcement learning and evolutionary
  computation, such as learning classifier systems,
  which provide more sophisticated generalization
  capabilities

9
Cost estimations and metrics

• Metric evaluation for
• partitioning
• mapping
• Input
• C based annotated description
• Application constraints Maximum size of task,
  bandwidth, performance
• Task decomposition (expressed as annotations or
  external file)
• Profiling information (expressed as annotations
  or external file)
• Target architecture constraints and description
• Tool extracts
• Inter-process synchronization and communication
• Inter-procedural control and data flow dependency
• Output
• C based annotated description
• Control and data characterization of task
  exchanges
• Behavior similarity
• Closeness between tasks
• Structural relationship between application and
  target architecture

10
Task partitioning interactions
11
C based descriptions

• C does not have any parallelism based instruction
• Task described by C function
• A notation to express parallelism is required
• Many different notations are possible
• Pragma based
• Comment based
• XML based
• The notation must be powerful enough to express
  that
• two or more tasks can run in parallel (fork
  operations)
• the execution flow must wait for the termination
  of one or more tasks (join operation)

12
Motivations for OpenMP

• OpenMP is a collection of pre-processor
  directives (pragmas) used to express parallelism
  in C programs
• Purposely created for a fork/join model
• It is an open and widely adopted standard
• supported in next release of GNU/gcc compiler
  (gcc 4.2)
• platform independent
• mapping independent
• easy functional verification of partitioned
  programs on host machines

13
Supported Constructs in Source Code

• We already support a large number of C constructs
  (e.g. nested struct, union, pointer arithmetic,
  )
• Selected a meaningful subset of the GNU/GCC
  Torture Testsuite composed by 834 benchmarks.
  Currently, we cover 828/834 benchmarks (99.3).
• Covering means
• parsing
• building internal representation
• dumping back the C code
• compiling the produced code
• executing it without errors (most of them are
  software fault tolerant to detect execution
  errors)
• Ongoing works to support var_args, computed goto,
  not reducible loops.

14
The PandA framework
Integrated with GCC infrastructure The output of
the task partitioning is an OpenMP compliant C
code Group operations into tasks in an efficient
way, in order to meet performance
requirements Initial mapping on the platform
based on metrics under development
15
Preliminary results
ADPCM
JPEG
Results on ADPCM and JPEG algorithms (1.7-1.4 max
speedup) Correctly analyzed all the hArtes
applications
16
Questions?

• ? Hopefully on Friday directly to Fabrizio