Methods and tools for system on chip retargetable parallel programming

Transcription

1 St. Petersburg State University of Aerospace Instrumentation Institute of High-Performance omputer and Network Technologies Methods and tools for system on chip retargetable parallel programming Alexey Syschikov SUAI , St. Petersburg Bolshaya Morskaya, No 67

2 Languages and architectures Traditional compilers Traditional three-phase compiler structure source ++ source Frontend Java Frontend # Frontend -based Optimizer -based x86 x86 x86 x64 x64 ARM ++-based x based Frontend Optimizer -based x64 -based source Frontend Optimizer based x based source Frontend Optimizer -based ARM -based source Frontend Optimizer ++ ++based ARM based source Frontend Optimizer FORTRAN FORTRAN Frontend Java source # source source Frontend ARM

3 Retargeting and cross-compilers Retargeting is an attribute of software development tools designed to generate for more than one computing platform. x86 x86 source Frontend based x86 x64 source Frontend Frontend x64 Java Java source Optimizer source Frontend Frontend ++ -based x64 ++ ARM source # # Frontend source Frontend ARM source Frontend x86 x86 x86 x64 x64 x64 ARM ARM Holy Grail for everybody in software or hardware development

4 omplexity Back to the real world Frontends Source language Language files parsing Abstract syntax tree Transformation from language to IR model rosscompiler IR Frontend for the real language requires: Develop a syntax parser Reveal correspondence between language and IR models Develop translator [Language model] => [IR model] More languages (frontends) more complex IR and IR model More complex IR harder to add new languages (frontends) Languages

5 Back to the real world Intermediate representations Syntax tree ontrol flow graph Data flow graph Dependency graph Acyclic dependency graph

6 Back to the real world Intermediate representations Syntax tree ontrol flow graph Data flow graph Dependency graph Acyclic dependency graph

7 Back to the real world Intermediate representations Syntax tree ontrol flow graph Data flow graph Dependency graph Acyclic dependency graph

8 Back to the real world Intermediate representations Syntax tree ontrol flow graph Data flow graph Dependency graph Acyclic dependency graph

9 Back to the real world Intermediate representations Syntax tree ontrol flow graph Data flow graph Dependency graph Acyclic dependency graph

10 Back to the real world Intermediate representations Syntax tree ontrol flow graph Data flow graph Dependency graph Acyclic dependency graph

11 Existing approach summary Main problems omplexity of Intermediate Representation model (IR): Hard to translate source language model to IR model; Hard to translate IR model to target platform model. Insufficient formalism: in most cases IR is not based on a formal model (model of computation, Mo): No ability for formal verification Optimizations correctness cannot be proved

12 Proposed approach Platformindependent optimizations Platformdependent optimizations Algorithm Visual frontend backend - Algorithm Textual frontend Model-based internal representation backend Formal verifications Simulation & Debugging Model-based internal representation IR model with limited complexity generators instead of backend compilers

13 Proposed approach Frontends We propose the So programming concept with the set of key thesis: Medium-grained parallel computations with Sequential processing inside blocks Visual approach to parallel programming Dynamics of parallel computations Objects: active program elements Links: connection of operators and data Objects count Name Type In data size Out data size Amount of comp. omment Meta-data 3 Block production Functional 1100 int 1100 int op Matrix blocks production function Matrix1 blocks Integer[550] (0/2) 3 5 Join 1 7 blocks 10 Blocks pool Integer[1100] (0/10) 4 Parallel processing block 12 Produced blocks Integer[1100] (0/10) 4 ollect result 8 Matrix2 blocks Integer[550] (0/2) 8 Three-treads parallel processing Matrix 2 blocks Blocks preparation for parallel processing Pool of blocks for processing 4 Blocks production 1 10 Need refactoring 4 Blocks for prod P 4 Blocks production Prod. result P 4 Blocks production 3 10 It s our vision but not a crucial point of our approach.

14 Proposed approach Formal model We use the AGP-model (Asynchronous Growing Processes) for IR-model. Key features: Is algorithmic complete: allows representing any type of computations Is parallel, asynchronous and decentralized: allows to describe truly parallel computations Limited complexity: has only tenth of base computational components

15 Proposed approach Formal correctness and model-based debugging Two key derivatives of ARP-model usage: Verification by design Finding deadlocks/livelocks Finding circularities Make conclusions about overall program execution basing on some execution trace Equivalent transformations Formally proven optimizations Automated optimizations for specific tasks and platforms Functional debugging of sequential version with results transition to the parallel version.

16 Proposed approach generators UDA Java VHDL ustom compiler ARM toolset UDA compiler Android toolset Synthesis flow toolset ustom core ARM core NVidia GPU core Android JVM Fabric platform x86 x86 Native platform compiler Machine Optimizer x64 ARM x64 ARM Optimizer Native /++ + MPI Java # /++ MPI parallel Java # Native platform compiler Native platform compiler Native platform compiler Native platform compiler Machine Machine Machine Machine Make love not war! Make not!

17 Retargetable approach Test cases results Approach evaluated on test cases: ANSI : sequential, 3-nd model level ++ & Microsoft agents: parallel, 3-d model level ++ & MPI: parallel, 2-nd model level VHDL: sequential, 1-d model level for M-24: parallel, 2-nd model level

18 Platformindependent optimizations Platformdependent optimizations Summary Frontends Algorithm Visual frontend backend - Algorithm Textual frontend Model-based internal representation backend Formal verifications Simulation & Debugging Algorithmic complete (describe any computation) Parallel (describe parallelism obviously) Limited complexity (easier frontends)

19 Platformindependent optimizations Summary Intermediate processing Platformdependent optimizations Algorithm Visual frontend backend - Algorithm Textual frontend Model-based internal representation backend Formal verifications Simulation & Debugging Verifications by design Formally proved optimizations Debugging extensible to any level of parallelism

20 Platformindependent optimizations Platformdependent optimizations Summary s Algorithm Visual frontend backend - Algorithm Textual frontend Model-based internal representation backend Formal verifications Simulation & Debugging Limited complexity (easier backends) Proven target have the same correctness as intermediate

21 The end