编译HPL (hpl-2.0_FERMI_v08.tar) 过程简记

HPL: A Portable Implementation of the High-Performance Linpack Benchmark for Distributed-Memory Computers

安装HPL for GPU (hpl-2.0_FERMI_v08.tar) 之前，机器里要预装编译器，并行环境MPI、基本线性代数子方程（BLAS）或矢量图形信号处理器（VSIPL）两者之一。

我装的是BLAS，也装了CBLAS。不记得是不是必须，我还装了一个LAPACK（Linear Algebra PACKage，http://www.netlib.org/lapack/）。

1. BLAS

    比较简单，make即可

2. CLBAS

    make文件中有个BLLIB的路径指示为../../librefblas.a，修改为BLAS的blas.LINUX.a实际位置

3. LAPCK

    make文件中指定对BLAS的位置就行了。成功后大概有个下面的结果。

                        -->   LAPACK TESTING SUMMARY  <--
                Processing LAPACK Testing output found in the TESTING direcory
SUMMARY                 nb test run     numerical error         other error
================        ===========     =================       ================
REAL                    1064911         39      (0.004%)        0       (0.000%)
DOUBLE PRECISION        1052315         203     (0.019%)        0       (0.000%)
COMPLEX                 508588          2       (0.000%)        0       (0.000%)
COMPLEX16               530862          28      (0.005%)        0       (0.000%)

--> ALL PRECISIONS      3156676         272     (0.009%)        0       (0.000%)

4.  HPL for GPU

1.     关键改对make中这两个文件的位置改对，其他的保持默认。
       LAdir：  CBLAS 库或 VSIPL 库所在的目录
       LAlib：CBLAS 库或 VSIPL 库头文件、库文件
       编译完成后，生成可执行文件 xhpl (在hpl/<arch>/bin 目录下)。我的 <arch> 用的是默认的 CUDA_pinned
       可以借鉴 setup 目录里的各种平台下的编译文件模板，我用的是 Make.Linux_PII_CBLAS
      
      
   
2.     中间有几次提示找不到 libhpl.a, needed by 'dexe.grd'  我将Make.CUDA_pinned 中定义 libhpl.a 的地址写成绝对地址后，错误消失
       原始为：HPLlib       = $(LIBdir)/libhpl.a
       修改为：HPLlib       = /home/michaelchen/SHOC/hpl-2.0_FERMI_v08/lib/CUDA_pinned/libhpl.a
   
   
3.     MPI的地址也要写对。我用的是openmpi，lib用的是 lib/libmpi.so
   
   
4.     Make.CUDA_pinned 中 LINKER 用的是g77，这个在我这里没有，已经更新为 gfortran
   
   
5.     遇到Make.inc access permission denied  错误，多半是链接文件原始是root开头，需要修改，链接到我们自己的Make.CUDA_pinned文件内容
       使用 ln -sf source/Make.CUDA_pinned Make.inc
   
   
6. 
   
       这里附上我的编译文件： Make.CUDA_pinned
       粗体部分为我修改过的地方。
   # ----------------------------------------------------------------------
   # - shell --------------------------------------------------------------
   # ----------------------------------------------------------------------
   #
   SHELL        = /bin/sh
   #
   CD           = cd
   CP           = cp
   LN_S         = ln -s
   MKDIR        = mkdir
   RM           = /bin/rm -f
   TOUCH        = touch
   #
   # ----------------------------------------------------------------------
   # - Platform identifier ------------------------------------------------
   # ----------------------------------------------------------------------
   #
   ARCH         = CUDA_pinned
   #
   # ----------------------------------------------------------------------
   # - HPL Directory Structure / HPL library ------------------------------
   # ----------------------------------------------------------------------
   #
   TOPdir       = /home/michaelchen/SHOC/hpl-2.0_FERMI_v08
   INCdir       = $(TOPdir)/include
   BINdir       = $(TOPdir)/bin/$(ARCH)
   LIBdir       = $(TOPdir)/lib/$(ARCH)
   #
   HPLlib       = /home/michaelchen/SHOC/hpl-2.0_FERMI_v08/lib/CUDA_pinned/libhpl.a
   #
   # ----------------------------------------------------------------------
   # - Message Passing library (MPI) --------------------------------------
   # ----------------------------------------------------------------------
   # MPinc tells the  C  compiler where to find the Message Passing library
   # header files,  MPlib  is defined  to be the name of  the library to be
   # used. The variable MPdir is only used for defining MPinc and MPlib.
   #
   MPdir        = /opt/openmpi-1.4.3
   MPinc        = -I$(MPdir)/include
   MPlib        = $(MPdir)/lib/libmpi.so
   #
   # ----------------------------------------------------------------------
   # - Linear Algebra library (BLAS or VSIPL) -----------------------------
   # ----------------------------------------------------------------------
   # LAinc tells the  C  compiler where to find the Linear Algebra  library
   # header files,  LAlib  is defined  to be the name of  the library to be
   # used. The variable LAdir is only used for defining LAinc and LAlib.
   #
   LAdir        = $(HOME)/SHOC/CBLAS
   LAinc        =
   LAlib        = $(LAdir)/lib/cblas_LINUX.a
   #
   # ----------------------------------------------------------------------
   # - F77 / C interface --------------------------------------------------
   # ----------------------------------------------------------------------
   # You can skip this section  if and only if  you are not planning to use
   # a  BLAS  library featuring a Fortran 77 interface.  Otherwise,  it  is
   # necessary  to  fill out the  F2CDEFS  variable  with  the  appropriate
   # options.  **One and only one**  option should be chosen in **each** of
   # the 3 following categories:
   #
   # 1) name space (How C calls a Fortran 77 routine)
   #
   # -DAdd_              : all lower case and a suffixed underscore  (Suns,
   #                       Intel, ...),                           [default]
   # -DNoChange          : all lower case (IBM RS6000),
   # -DUpCase            : all upper case (Cray),
   # -DAdd__             : the FORTRAN compiler in use is f2c.
   #
   # 2) C and Fortran 77 integer mapping
   #
   # -DF77_INTEGER=int   : Fortran 77 INTEGER is a C int,         [default]
   # -DF77_INTEGER=long  : Fortran 77 INTEGER is a C long,
   # -DF77_INTEGER=short : Fortran 77 INTEGER is a C short.
   #
   # 3) Fortran 77 string handling
   #
   # -DStringSunStyle    : The string address is passed at the string loca-
   #                       tion on the stack, and the string length is then
   #                       passed as  an  F77_INTEGER  after  all  explicit
   #                       stack arguments,                       [default]
   # -DStringStructPtr   : The address  of  a  structure  is  passed  by  a
   #                       Fortran 77  string,  and the structure is of the
   #                       form: struct {char *cp; F77_INTEGER len;},
   # -DStringStructVal   : A structure is passed by value for each  Fortran
   #                       77 string,  and  the  structure is  of the form:
   #                       struct {char *cp; F77_INTEGER len;},
   # -DStringCrayStyle   : Special option for  Cray  machines,  which  uses
   #                       Cray  fcd  (fortran  character  descriptor)  for
   #                       interoperation.
   #
   F2CDEFS      =
   #
   # ----------------------------------------------------------------------
   # - HPL includes / libraries / specifics -------------------------------
   # ----------------------------------------------------------------------
   #
   HPL_INCLUDES = -I$(INCdir) -I$(INCdir)/$(ARCH) $(LAinc) $(MPinc)
   HPL_LIBS     = $(HPLlib) $(LAlib) $(MPlib)
   #
   # - Compile time options -----------------------------------------------
   #
   # -DHPL_COPY_L           force the copy of the panel L before bcast;
   # -DHPL_CALL_CBLAS       call the cblas interface;
   # -DHPL_CALL_VSIPL       call the vsip  library;
   # -DHPL_DETAILED_TIMING  enable detailed timers;
   #
   # By default HPL will:
   #    *) not copy L before broadcast,
   #    *) call the BLAS Fortran 77 interface,
   #    *) not display detailed timing information.
   #
   HPL_OPTS     = -DHPL_CALL_CBLAS
   #
   # ----------------------------------------------------------------------
   #
   HPL_DEFS     = $(F2CDEFS) $(HPL_OPTS) $(HPL_INCLUDES)
   #
   # ----------------------------------------------------------------------
   # - Compilers / linkers - Optimization flags ---------------------------
   # ----------------------------------------------------------------------
   #
   CC           = /usr/bin/gcc
   CCNOOPT      = $(HPL_DEFS)
   CCFLAGS      = $(HPL_DEFS) -fomit-frame-pointer -O3 -funroll-loops
   #
   # On some platforms,  it is necessary  to use the Fortran linker to find
   # the Fortran internals used in the BLAS library.
   #
   LINKER       = /usr/bin/gfortran
   LINKFLAGS    = $(CCFLAGS)
   #
   ARCHIVER     = ar
   ARFLAGS      = r
   RANLIB       = echo
   #
   # ----------------------------------------------------------------------
   
   
7. 运行时好几次说有些.so, .a 文件找不到等等，这些都可以在同目录下的path文件指定清楚，记住source使其生效。
   或者直接拷贝/链接对应的文件到目录中。
   比如我用的是CUDA 4.0，但是HPL for CUDA是建立在CUDA3.0上的。所以当它寻找 usr/local/cuda/lib64 目录中的 libcublas.so.3 等文件时会提示找不到，只有so.4。
   我们可以用上面的方法解决。
   
   
8. 运行时还可能会有一个错误：
   hpl-2.0_FERMI_v08/bin/CUDA_pinned/xhpl: error while loading shared libraries: libcublas.so.3: wrong ELF class: ELFCLASS32
   这是因为CUDA中用错了lib的版本，不应该用 usr/local/cuda/lib ， 而应该用 usr/local/cuda/lib64 中的 so 文件
   

    运行mpirun -np 1 xi03 成功。

    如果出现“Error allocating scratch space 2048.00 MB”之类的问题，在对应的src/cuda中找到源文件（比如cuda_dgemm.c第229行中的‘2048’），进行相应修改（比如改为1024），调小一些即可。
    eg：err1=cudaMalloc((void**)&dev_scratch[0], (size_t)(1024.0*1024.0*1024.0) );

    具体如何微调 HPL.dat 中的值，可以参考HPL 根目录下的 TUNE 文件。

Ref:  http://blog.sina.com.cn/s/blog_442806280100mxbu.html