IBM pseries Compiler Roadmap

Transcription

1 IBM pseries Compiler Roadmap Roch Archambault IBM Toronto Laboratory SciComp 10 August 13, 2004

2 Agenda The pseries Compiler Products Roadmaps XL Fortran XL C/C++ Performance Multiple compiler installation Performance Results Q&A

3 The pseries Compiler Products Latest Versions (All POWER4 enabled except Mac OS X) VisualAge C++ Version 6.0 for AIX VisualAge C++ Version 6.0 for Linux on pseries XL C/C++ Advanced Edition Version 6.0 for Mac OS X (PPC970 only) C for AIX V6.0 XL Fortran Version 8.1 for AIX (latest level is version 8.1.1) XL Fortran Version 8.1 for Linux on pseries (latest level is version 8.1.1) XL Fortran Advanced Edition Version 8.1 for Mac OS X (PPC970 only) Upcoming Versions (All POWER4, POWER5 and PPC970 enabled) XL C/C++ Enterprise Edition V7.0 for AIX, Linux XL C/C++ Advanced Edition V7.0 for Mac OS X (PPC970 only) XL Fortran Enterprise Edition V9.1 for AIX, Linux XL Fortran Advanced Edition V9.1 for Mac OS X (PPC970 only)

4 XL Fortran Roadmap: Strategic Priorities Premium Customer Service Continue to work closely with key ISVs and customers in scientific and technical computing industries Compliance to Language Standards and Industry Specifications OpenMP Fortran API V2.0 Fortran 77, 90 and 95 standards Emerging 2003 Standard Exploitation of Hardware Committed to maximum performance on POWER4, PPC970, POWER5 and successors Continue to work very closely with processor design teams

5 XL Fortran Version 8.1 for AIX Fully compliant Fortran 77/90/95 compiler Partial Fortran 2003 support Allocatable components, IEEE module, INTENT of F90 pointers Many industry extensions 128-bit float, 64-bit integer, Cray pointers, structure record, union map, BYTE, STATIC, AUTOMATIC, asynchronous I/O, SIZE intrinsic 32- and 64-bit support Optimized OpenMP Fortran V2.0 support Symbolic Debugging support TotalView and dbx/pdbx Full support for debugging of OpenMP programs Partial support for debugging of optimized code Portfolio of Optimizing Transformations Comprehensive path length reduction Whole program analysis Loop optimization for parallelism, locality and instruction scheduling Tuned support for all pseries processors (including POWER4) More info:

6 XL Fortran V8.1.1 for AIX (Update to V8.1) Fortran 2003 Stream I/O VALUE and PROTECTED attributes and statements Performance Customer benchmarks, SPECCPU2000 improvements, scalability for SPECOMP 64-bit TPO enablement (compiler component) Stream_unroll and Unroll_and_fuse directives Unroll directive extended to outer loops Intrinsic Performance improvements Customer Requirements -qsuppress suboption -qextname suboption Improved Debug support with TotalView Installation and operation on OS/400 PASE on iseries

7 XL Fortran Version 8.1 for Linux on pseries Based on XL Fortran Version 8.1 for AIX Product Leverage proven industry leading performance compiler technology Fully compliant Fortran 77/90/95 compiler Partial Fortran 2003 support Allocatable components, IEEE module, INTENT on F90 pointers Many industry extensions 128-bit float, 64-bit integer, Cray pointers, structure record, union map, BYTE, STATIC, AUTOMATIC, asynchronous I/O, SIZE intrinsic 32- and 64-bit support Optimized OpenMP Fortran V2.0 support Symbolic Debugging support with gdb Portfolio of Optimizing Transformations (same as XLF V8.1 for AIX) Supports SUSE Linux Enterprise Server 8 (SLES 8) Supports RedHat Enterprise Linux 3.0 (RHEL 3) Support for TurboLinux (TLES 8) and Conectiva

8 XL Fortran V9.1 And Beyond XL Fortran Enterprise Edition V9.1 for AIX (2004) XL Fortran Enterprise Edition V9.1 for Linux (2004) XL Fortran Advanced Edition V9.1 for Mac OS X Continued rollout of Fortran 2003 POWER5 support PPC970 and PPC970FX support including SIMD vectorization (VMX) Include MASS library in compiler products Improved performance Improved compile time and reduce space usage Improved debugging New directives and intrinsics New customer requirements Supported on SLES9 and RHEL 3 update 3 (Linux only) Supported on OS/400 (V5R2) and i5/os PASE (V5R3) All information subject to change without notice

9 Tentative rollout of Fortran 2003 (XLF V9.1) Available in XL Fortran V9.1 (2004) BIND attribute and statement, BIND on derived types, common blocks, and procedures PROCEDURE, IMPORT and FLUSH statements PUBLIC/PRIVATE on components of a type ISO_FORTRAN_ENV and ISO_C_BINDING intrinsic modules ASSOCIATE Construct Command Line Argument and NEW_LINE intrinsics Intrinsic attribute on USE IOMSG= specifier All information subject to change without notice

10 Prioritized rollout of Fortran 2003 (XLF V10.1) 1. OO extensions 2. ENUM/ENUMERATOR 3. Derived type I/O 4. Asynchronous I/O additions 5. Allow character on MAX/MIN family of intrinsics 6. Procedure pointers 7. I/O specifiers (DECIMAL=, SIGN=, ROUND=, and PAD=) 8. Complex literal 9. Derived type parameters 10. Abstract interface 11. Enhance structure constructors 12. Enhanced array constructors 13.Pointer assignment enhancement (specify bounds) 14. Support for International Usage All information subject to change without notice

11 C/C++ Roadmap: Strategic Priorities Premium Customer Service Compliance to Language Standards and Industry Specifications ANSI / ISO C and C++ Standards OpenMP C/C++ API V2.0 Exploitation of Hardware Committed to maximum performance on POWER4, PPC970, POWER5 and successors Continue to work very closely with processor design teams Exploitation of OS and Middleware Synergies with operating system and middleware ISVs (performance, specialized function) Committed to AIX Linux affinity strategy and to Linux on pseries Reduced Emphasis on proprietary Tooling Affinity with GNU toolchain

12 VisualAge C++ Version 6.0 for AIX C for AIX Version 6.0 Fully compliant ISO C 1999 and ISO C standards AIX 5.2 required for ISO C 1999 complex runtime support Partial GNU C/C++ language and options compatibility 32- and 64-bit support Interprocedural Analysis (IPA) including specialized C++ optimizations for templates, exception handling and virtual functions Automatic Parallelization for C/C++ Optimized OpenMP C/C++ API V1.0 support Template instantiation improvements Reuse of template instantiations to reduce code size Parsing templates at instantiation time Improved automatic template instantiations Symbolic Debugging Support TotalView, IBM Distributed Debugger and dbx/pdbx Full support for debugging of OpenMP programs Partial support for debugging of optimized code Runtime memory debug support

13 VisualAge C++ Version 6.0 for AIX (continued) C for AIX Version 6.0 (continued) Portfolio of optimizing transformations Comprehensive path length reduction Whole program analysis Loop optimization for parallelism, locality and instruction scheduling Tuned support for all pseries processors (including POWER4) More info:

14 VisualAge C++ Version 6.0 for Linux on pseries Based on VisualAge C++ Version 6.0 for AIX Product Leverage proven industry leading performance compiler technology Fully compliant ISO C 1999 and ISO C standards Partial GNU C/C++ language and options compatibility 32- and 64-bit support Automatic Parallelization for C/C++ Optimized OpenMP C/C++ API V1.0 support Portfolio of Optimizing Transformations (same as C/C++ V6.0 for AIX) Supports SUSE Linux Enterprise Server 8 (SLES 8) Supports RedHat Enterprise Linux 3.0 (RHEL 3) Support for TurboLinux (TLES 3) and Conectiva

15 XL C/C++ V7.0 And Beyond XL C/C++ Enterprise Edition Version 7.0 for AIX (2004) XL C/C++ Enterprise Edition Version 7.0 for Linux (2004) XL C/C++ Advanced Edition Version 7.0 for Mac OS X Addenda to ISO C/C++ standards (C is 1999, C++ is 1998 with TC1) Support for some C99 features in C++ compiler: Complex, Hex Float and Variable Length Arrays, Compound Literal Further GNU C/C++ compatibility features OpenMP C/C++ API V2.0 POWER5 support PPC970 and PPC970FX support including SIMD vectorization (VMX) Improved performance 64-bit TPO enablement (compiler component) on AIX and Linux Boost and STLport libraries (as they become better accepted by the standard committee) All information subject to change without notice

16 XL C/C++ V7.0 And Beyond (Continued) XL C/C++ Enterprise Edition Version 7.0 for AIX (2004) XL C/C++ Enterprise Edition Version 7.0 for Linux (2004) XL C/C++ Advanced Edition Version 7.0 for Mac OS X IOStream Performance improvements New pragmas / options Xcode enabled and support for Frameworks (Mac OS X only) Supported on SLES9 and RHEL 3 update 3 (Linux only) Supported on OS/400 (V5R2) and i5/os PASE (V5R3) New compiler drivers (gxlc and gxlc) for compatibility with gcc and g++ Support for UPC language (AIX only) Support for Objective-C (Mac OS X only) Include MASS library in compiler products All information subject to change without notice

17 Tentative GNU C/C++ Compatibility Enhancements Labels as values / computed goto Nested functions Naming types Conditionals with omitted operands Zero length arrays Labeled elements (C only) Case ranges (C only) Cast to union (C only) Function Attributes Support Noinline, always_inline, format, format_arg, section Accept and ignore used All information subject to change without notice

18 Tentative GNU C/C++ Compatibility Enhancements (Continued) Variable Attributes Support Nocommon, transparent_union Type Attributes Support Aligned, packed Accept and ignore Transparent_union extension Incomplete enums Function names as strings Partial Asm support All information subject to change without notice

19 Performance Improvements Delivered in 2003 Included in XLF V8.1.1 release OpenMP Much faster barrier implementation (over 5x faster on 32-way p690) Much faster uncontended atomic (over 4x faster on 32-way p690) Improved parallel region startup (25%) POWER4 Expanded scope and precision of instruction scheduling More precise loop unrolling for pipelining and parallelization of reductions Use of dcbz to optimize store streams Loop unrolling for prefetch stream utilization Loop Optimization Aggressive loop fusion to exploit data reuse Index set splitting and gather/scatter to move branches out of loops Loop peeling for automatic parallelization Improved loop interchange for automatic parallelization Strip-mining of vectorized loops Vector calls to specialized routines for POWER4 (vsqrt, vrsqrt, vtan, vlog, vexp)

20 Expected Performance Improvements in 2004 Included in XLF V9.1 and XL C/C++ V7 on all platforms (AIX, Linux, Mac OS X) POWER5 Modified scheduling machine model Usage of improved prefetch facilities Usage of new instructions PPC970 Automatic generation of VMX code on Linux and Mac OS X (SIMD vectorization) Interprocedural pointer alignment propagation OpenMP Tuned support for 64-way SMP Continued improvements in overhead reduction Intrinsic functions (Fortran Only) MATMUL, TRANSFER, INDEX, TRANSPOSE

21 Expected Performance Improvements in 2004 Included in XLF V9.1 and XL C/C++ V7 on all platforms (AIX, Linux, Mac OS X) Tuning assists : BLOCK_LOOP and LOOPID directives to specify which set of loops to tile, interchange or strip-mine NOVECTOR and NOSIMD directive to tell compiler not to vectorize or simdize loop Builtin functions for generating software divides (full double precision on POWER5) Thread binding (set via XLSMPOPTS env variable) Environment variable (intrinthds) to control number of threads used by MATMUL and RANDOM_NUMBER (added in XLF V8.1.1 PTF) View and manipulate information gathered by profile directed feedback (-qpdf1/-qpdf2) via showpdf and mergepdf tools New prefetch directives for POWER5

22 Expected Performance Improvements in 2004 Included in XLF V9.1 and XL C/C++ V7 on all platforms (AIX, Linux, Mac OS X) Loop Optimizations : Modulo scheduling of loops which contain branches Further improvements to loop fusion for data reuse (e.g. loop alignment) Perform vectorization on all platforms (including Linux and Mac OS X) Enhancement of vectorization (additional functions, loop versioning, vector merging) Tiling for BLAS-like and streaming loop nests Predictive Commoning (common subexpression elimination across loop iterations) Improved data dependence analysis Automatic generation of software divides on POWER5 Automatic generation of stream prefetch instructions on POWER5

23 Expected Performance Improvements in 2004 Included in XLF V9.1 and XL C/C++ V7 on all platforms (AIX, Linux, Mac OS X) Other Optimizations : Interprocedural Strength reduction Interprocedural Register Allocation Split array of structures into multiple arrays for better exploitation of hardware streams and smaller d-cache footprint Use Profile Directed Feedback (PDF) information to: Specialize calls to malloc/calloc to use pools of small objects Specialize memcpy and memset with small lengths Specialize integer divide and modulo Form Super blocks for better instruction scheduling

24 Expected Performance Improvements in 2005 Loop Optimizations: Sparse Vectorization Recognize BLAS loop and replace with call to ESSL Exploitation of SMT threads in parallel programs Improved dependence analysis Array Data flow analysis for array privatization and contraction Automatic array padding Improved automatic parallelization (wavefront, multidimensional reductions) Other Optimizations: Outline cold fields of data structures for smaller d-cache footprint (using pdf) Outline cold regions of code for smaller i-cache footprint (using pdf) All information subject to change without notice

25 Installation of Multiple Compiler Versions Installation of multiple compiler versions is supported The vacppndi and xlfndi scripts shipped with VisualAge C and XL Fortran 8.1 allow the installation of a given compiler release or update into a non-default directory The configuration file can be used to direct compilation to a specific version of the compiler Example: xlf_v8r1 c foo.f May direct compilation to use components in a non-default directory Care must be taken when multiple runtimes are installed on the same machine (details on next slide)

26 Coexistence of Multiple Compiler Runtimes Backward compatibility C, C++ and Fortran runtimes support backward compatibility. Executables generated by an earlier release of a compiler will work with a later version of the run-time environment. Concurrent installation Multiple versions of a compiler and run-time environment can be installed on the same machine Full support in xlfndi and vacppndi scripts is now available Limited support for coexistence LIBPATH must be used to ensure that a compatible runtime version is used with a given executable Only one runtime version can be used in a given process. Renaming a compiler library is not allowed. Take care in statically linking compiler libraries or in the use of dlopen or load. Details in the compiler FAQ

27 Documentation An information center containing the documentation for the current versions of the AIX compilers (both Fortran and C/C++) is available at: Similar documentation for the current version of Mac OS X compilers (both Fortran and C/C++) is available at: This information center contains all the html documentation shipped with the compilers. It is completely searchable. Please send any comments or suggestions on this information center or about the existing C, C++ or Fortran documentation shipped with the products to compinfo@ca.ibm.com.

28 Software Divide Improvements On POWER5 % Imrovement vs hardware divide SP SWDIV SP SWDIV_NOCHK Software Divide Instrinsics DP SWDIV DP SWDIV -qnostrict DP SWDIV_NOCHK DP SWDIV_NOCHK -qnostrict

29 SPEC FP Base Improvements From Compiler On POWER4 XLF V9.1 & XLC V7 Versus XLF V8.1.1 & XLC V6 % improvement SPECFP Benchmarks 168.wupwise 171.swim 172.mgrid 173.applu 177.mesa 178.galgel 179.art 183.equake 187.facerec 188.ammp 189.lucas 191.fma3d 200.sixtrack 301.apsi SPEC FP + 18%

30 SPEC FP Base Improvements From Compiler On POWER5 XLF V9.1 & XLC V7 Versus XLF V8.1.1 & XLC V6 % Improvement SPECFP Benchmarks 168.wupwise 171.swim 172.mgrid 173.applu 177.mesa 178.galgel 179.art 183.equake 187.facerec 188.ammp 189.lucas 191.fma3d 200.sixtrack 301.apsi SPEC FP + 20%

31 SPEC FP Base Auto-Parallelization (2 CPUs, POWER4) XLF V9.1 & XLC V7 2-CPU versus 1-CPU % Improvement SPECFP Benchmarks 168.wupwise 171.swim 172.mgrid 173.applu 177.mesa 178.galgel 179.art 183.equake 187.facerec 188.ammp 189.lucas 191.fma3d 200.sixtrack 301.apsi SPEC FP + 8%

32 90 70 SPEC FP Peak Differences Between IT2 (1.5 GHZ) and POWER5 P5 / IT2 % difference GHZ 1.9 GHZ 168.wupwise 171.swim 172.mgrid 173.applu 177.mesa 178.galgel 179.art 183.equake 187.facerec 188.ammp 189.lucas 191.fma3d 200.sixtrack 301.apsi SPECFP Peak

33 32-way EPCC results on AIX 5.2 p690 system (1.1 GHZ) XLF XLF PAR DO PDO BAR SING CRIT LOCK ORD ATM REDC Time in micro-seconds - lower is better

34 SPECOMP Base Improvements From Compiler On POWER4 (32-way) XLF V9.1 & XLC V7 Versus XLF V8.1.1 & XLC V6 % improvement SPECOMP Benchmarks 310.wupwise_m 312.swim_m 314.mgrid_m 316.applu_m 318.galgel_m 320.equake_m 324.apsi_m 326.gafort_m 328.fma3d_m 330.art_m 332.ammp_m SPECOMP + 8%

35 SPECOMP Scalability On POWER4 (16 vs 32 CPUs) Scalability Factor 1.0 implies perfect scaling SPECOMP Benchmarks 310.wupwise_m 312.swim_m 314.mgrid_m 316.applu_m 318.galgel_m 320.equake_m 324.apsi_m 326.gafort_m 328.fma3d_m 330.art_m 332.ammp_m

36 SPEC OMPM2001 Base Versus Competition 40 Thousands IBM 32x1.7 HP-I 32x1.5 SGI-I 32x1.5 HP-A 64x1.15 HP-I 64x1.5 SGI-I 64x1.5 FUJ 64x1.3 IBM P *1.9 0 SPECMARK

37 Himeno Benchmark Improvement on POWER XLF V8.1.1 XLF V CPU 2-CPU 4-CPU 8-CPU Small Data Problem Measured in MFLOPS on 1.1GHZ p690