📄 http:^^www.cs.rice.edu^~mpal^sc95^
字号:
Date: Tue, 14 Jan 1997 21:25:51 GMTServer: NCSA/1.4.2Content-type: text/htmlLast-modified: Mon, 04 Dec 1995 01:05:01 GMTContent-length: 10777<!-- --><HTML><HEAD><TITLE>Fortran D95 Compiler Overview</TITLE></HEAD><! color == rrggbb ><! BODY BGCOLOR=#color : for background TEXT =#color : for text LINK =#color : for normal links VLINK =#color : for recently followed links ALINK =#color : for ?><BODY><HR><HR><A NAME=1><H1><P ALIGN=CENTER> Fortran D95 Compiler Overview</P></H1><P ALIGN=CENTER><BR><b>Project Leaders:</b><BR>John Mellor-Crummey and Vikram Adve<BR><BR><b>Participants:</b><BR>Zoran Budimlic, Alan Carle,Kevin Cureton, Gil Hansen, Ken Kennedy,Charles Koelbel, Bo Lu, Collin McCurdy,Nat McIntosh,Dejan Mircevski,Nenad Nedeljkovic, Mike Paleczny, Ajay Sethi, Yoshiki Seo,Lisa Thomas, Lei Zhou</P><BR><H3><P ALIGN=CENTER><!WA0><A HREF="http://softlib.rice.edu/fortran-tools/fortran-tools.html">Parallel Compiler and Tools Group</A><BR><!WA1><A HREF="http://www.cs.rice.edu/CRPC.html"><i>Center for Research on Parallel Computation</i></A><BR><!WA2><A HREF="http://www.rice.edu/"><i>Rice University</i></A><BR></P></H3><BR><b>Group Mission: </b><BR>Develop integrated compiler and tools to support effective machine-independent parallel programming<BR><HR><HR><BR><A NAME=HEADING16><A NAME=16><b>Contents</b>: <UL> <LI> <!WA3><A HREF="#HEADING15">Compiler Goals</A> <LI> <!WA4><A HREF="#HEADING2">Fortran D95 Language</A> <LI> <!WA5><A HREF="#HEADING3">Fortran D95 Compiler Organization</A> <LI> <!WA6><A HREF="#HEADING17">Compiled Examples</A> </UL><P ><A NAME=HEADING15><A NAME=15><HR><H2> <P ALIGN=CENTER> Compiler Goals</P></H2><P>Serve as an extensible platform for experimental research on compilationtechniques and programming tools for full applications, including:<UL><LI> unified treatment of regular and irregular problems<LI> global strategies for computation partitioning<LI> parallelism-enhancing and latency-hiding transformations<LI> whole-program compilation and interprocedural transformations<LI> code generation strategies that approach hand-tuned performance<LI> architecture-independent compilation based on machine models<LI> message-passing, shared-memory, and hybrid communication models<LI> optimization in the presence of resource constraints<LI> programming tools that fully support abstract, high-level, programming models</UL><P><P><A NAME=HEADING2><A NAME=2><HR><H2> <P ALIGN=CENTER> Fortran D95 Language</P></H2><P>Fortran D95 is designed to support research ondata-parallel programming in High Performance Fortran (HPF) and toexplore extensions that would broaden HPF's applicability or enhanceperformance.<P> <BR><b>Features</b><UL><LI> Fortran 77 + Fortran 90 array syntax + FORALL + ALLOCATE<LI> High Performance Fortran (HPF) data mapping directives for regular problems<UL><LI> ALIGN, DISTRIBUTE, REALIGN, REDISTRIBUTE, TEMPLATE, PROCESSORS</UL><LI> INDEPENDENT, and ON_HOME<LI> value-based data-mapping directives to support irregular problems<LI> experimental support (under development) for<UL><LI> parallel input/output, including out-of-core array management<LI> complex data structures <LI> structured use of task parallelism</UL></UL><P><BR><BR><BR><HR><A NAME=HEADING3><A NAME=3><HR><H2> <P ALIGN=CENTER> Fortran D95 Compiler Organization</P></H2><pre> <!WA7><A HREF="#4">Front End</A><BR> <!WA8><IMG SRC="http://www.cs.rice.edu/~mpal/SC95/downa.gif" ALT="v"> Parallelism <!WA9><A HREF="#5">Preliminary Communication Placement</A><BR> <!WA10><IMG SRC="http://www.cs.rice.edu/~mpal/SC95/downa.gif" ALT="v"> and <!WA11><A HREF="#6">Computation Partitioning</A><BR>Communication <!WA12><IMG SRC="http://www.cs.rice.edu/~mpal/SC95/downa.gif" ALT="v"> <!WA13><IMG SRC="http://www.cs.rice.edu/~mpal/SC95/upa.gif" ALT="^"> Placement <!WA14><A HREF="#7">Communication Refinement</A><BR> <!WA15><IMG SRC="http://www.cs.rice.edu/~mpal/SC95/downa.gif" ALT="v"> <!WA16><A HREF="#8">Code Generation</A><BR></pre><BR><P><A NAME=HEADING4><A NAME=4><HR><H2> <P ALIGN=CENTER> Front End</P></H2><P><b>Purpose:</b> interpret HPF directives and compute directives affecting each statement and reference<P><b>Directive Processing</b><UL><LI> semantic analysis of directives in program<LI> infer canonical synthetic layout directives for all program variables unmentioned in program directives<LI> intraprocedural flow-sensitive propagation of (RE)ALIGN, (RE)DISTRIBUTE to statements and array references</UL><P><b>Limitations (November 1995)</b><UL><LI> no interprocedural propagation of layout information</UL><P><BR><P><A NAME=HEADING5><A NAME=5><HR><H2> <P ALIGN=CENTER> Preliminary Communication Placement</P></H2><P><b>Purpose:</b>provide feedback to the computation partitioner about where(conservatively) communication might be needed<P><b>Strategy</b> <UL><LI> conservatively assume all references to non-replicated variablesmay need communication <LI>hoist communication for a reference to the outermost loop level possible whilerespecting data dependences on the reference or its subscripts<LI> conservatively prevent communication from being hoisted out of ``non-do-loop'' iterative constructs </UL><P><b>Limitations (November 1995)</b><UL><LI> placement independent of resource constraints<LI> no support for pipelining communication to achieve partial parallelism <LI> lacks dataflow placement optimization to<UL><LI> eliminate partial redundancies <LI> hide communication latency</UL><LI> no inspector placement for irregular data accesses</UL><P><BR><P><A NAME=HEADING6><A NAME=6><HR><H2> <P ALIGN=CENTER> Computation Partitioning Selection</P> </H2><P><b>Purpose:</b>a framework to evaluate and select from several computation partitioningalternatives, not restricted to the owner-computes rule.<P><b>Approach:</b> explicitly enumerate candidate partitioningchoices and use explicit cost estimation to select the best partitioning.<P><UL><LI> enumerate candidate CP choices for a loop nest (or set of loop nests)<!WA17><A HREF="http://www.cs.rice.edu/~mpal/SC95/comp-part.gif">[example]</A><LI> refine communication information for each candidate CP<LI> estimate performance of each candidate CP: <UL><LI> load-balance (unimplemented) <LI> communication overhead </UL><LI> propagate computation partitionings to DO, IF, statements, and computations involving only privatizable variables.</UL><P><b>Limitations (November 1995)</b><UL><LI> load balance is not considered<LI> ignores message coalescing across loop nests<LI> communication cost estimates are very simplistic<LI> requires constant loop bounds (but simplistic handling of symbolics will be straightforward)</UL><P><BR><P><A NAME=HEADING7><A NAME=7><HR><H2> <P ALIGN=CENTER> Communication Refinement</P></H2><P><b>Purpose:</b>given a computation partition choice, CP, compute a projection ofthe conservatively placed communication w.r.t. CP: <!WA18><A HREF="http://www.cs.rice.edu/~mpal/SC95/comm-place.gif">[example]</A><UL><LI> eliminate communication for references local to that CP<LI> eliminate redundant communication by coalescing<LI> determine communication pattern<LI> perform message coalescing optimization</UL><P><b>Limitations (November 1995)</b><UL><LI> assume one single reaching layout per reference<LI> conservative unless single processors statement,perfect alignment, and same number of distributed dimensions<LI> communication pattern recognition is somewhat limited<LI> dataflow analysis for eliminating partial redundancies andlatency hiding not yet fully in place</UL><P><BR><P><HR><A NAME=HEADING8><A NAME=8><H2> <P ALIGN=CENTER> Code Generation </P></H2><b>Principal Functionality</b><!WA19><A HREF="http://www.cs.rice.edu/~mpal/SC95/examples/src/simpletest.html">[source for running example]</A><UL><LI>Computation partitioning transformations: <UL> <LI> reduce loop bounds and insert guards where necessary<!WA20><A HREF="http://www.cs.rice.edu/~mpal/SC95/examples/generated-code/fragments/reduced.html">[example]</A> <LI> separate loop iterations that might access non-local values to minimize overhead from runtime locality checks<!WA21><A HREF="http://www.cs.rice.edu/~mpal/SC95/examples/generated-code/prelin.12.2.95/simpletest-fragments/local-nonlocal.html">[example]</A> </UL><LI>Communication generation and storage management: <UL> <LI> compute data sets to send/recv between processor pairs <LI> generate code to pack/unpack buffers and send/recv data<!WA22><A HREF="http://www.cs.rice.edu/~mpal/SC95/examples/generated-code/prelin.12.2.95/simpletest-fragments/recv-unpack.html">[example]</A> <LI> generate run-time dynamic storage management to cope with dynamic layouts<!WA23><A HREF="http://www.cs.rice.edu/~mpal/SC95/examples/generated-code/prelin.12.2.95/simpletest-fragments/storage-mgmt.html">[example]</A> <LI> localize and linearize subscripts<!WA24><A HREF="http://www.cs.rice.edu/~mpal/SC95/examples/generated-code/fragments/loc-lin.html">[example]</A> </UL></UL><P><BR><b>Current Strategy</b><P>Except for storage management, all the code generation tasks requireheavy manipulation of integer sets, especially for compiling regularapplications for distributed-memory machines. Examples:<UL><LI> data to send or recieve for a particular reference, given itscomputation partition<LI> a processor's loop iterations that access local or non-local data</UL>Current implementation uses the Omega library (University of Maryland):<DL COMPACT><DT><b>(+)</b><DD> arbitrary integer sets<DT><b>(+)</b><DD> rich language for mappings between sets<DT><b>(+)</b><DD> almost complete set of operations on sets and mappings: (union, intersection, difference, inverse, composition)<DT><b>(+)</b><DD> good code-generation and optimization<P><DT><b>(-)</b><DD> code generation is slow<DT><b>(-)</b><DD> limited support for symbolics<P></DL><b>Limitations (November 1995)</b> <UL><LI> run-time resolution guards currently handle only one or all processors per dynamic statement instance<LI> lack library support for dynamic remapping<LI> current localization and linearization strategy producesgeneral, but slow code.</UL><P><BR><P><HR><A NAME=HEADING17><A NAME=17><H2> <P ALIGN=CENTER> Compiled Examples</P></H2><UL><LI> simple 1D shift kernel<!WA25><A HREF="http://www.cs.rice.edu/~mpal/SC95/examples/src/simpletest.html">[HPF]</A><!WA26><A HREF="http://www.cs.rice.edu/~mpal/SC95/examples/generated-code/final.12.2.95/simpletest.html">[F77+MPI]</A><LI> Jacobi iteration <!WA27><A HREF="http://www.cs.rice.edu/~mpal/SC95/examples/src/jacobi.html">[HPF]</A><!WA28><A HREF="http://www.cs.rice.edu/~mpal/SC95/examples/generated-code/final.12.2.95/jacobi.html">[F77+MPI]</A><LI> Livermore 18 explicit hydrodynamics kernel <!WA29><A HREF="http://www.cs.rice.edu/~mpal/SC95/examples/src/liv18.html">[HPF]</A><!WA30><A HREF="http://www.cs.rice.edu/~mpal/SC95/examples/generated-code/final.12.2.95/expl.html">[F77+MPI]</A><UL><!WA31><AHREF="http://www.cs.rice.edu/~mpal/SC95/examples/generated-code/final.12.2.95/expl-non-owned.html"><LI>Non owner-computes partitioning fragment</A></UL></UL><P><BR><HR><HR><P><ADDRESS>http://www.cs.rice.edu/~mpal/SC95/index.html</ADDRESS><P><HR><HR></BODY></BODY></HTML>⌨️ 快捷键说明
复制代码
Ctrl + C
搜索代码
Ctrl + F
全屏模式
F11
切换主题
Ctrl + Shift + D
显示快捷键
?
增大字号
Ctrl + =
减小字号
Ctrl + -