wrap.py

a PMPI wrapper generator

by Todd Gamblin, tgamblin@llnl.gov, https://github.com/tgamblin/wrap

Usage: wrap.py [-fgd] [-i pmpi_init] [-c mpicc_name] [-o file] wrapper.w [...]
 Python script for creating PMPI wrappers. Roughly follows the syntax of
   the Argonne PMPI wrapper generator, with some enhancements.
 Options:"
   -d             Just dump function declarations parsed out of mpi.h
   -f             Generate fortran wrappers in addition to C wrappers.
   -g             Generate reentry guards around wrapper functions.
   -c exe         Provide name of MPI compiler (for parsing mpi.h).
                  Default is \'mpicc\'.
   -s             Skip writing #includes, #defines, and other
                  front-matter (for non-C output).
   -i pmpi_init   Specify proper binding for the fortran pmpi_init
                  function.  Default is \'pmpi_init_\'.  Wrappers
                  compiled for PIC will guess the right binding
                  automatically (use -DPIC when you compile dynamic
                  libs).
   -o file        Send output to a file instead of stdout.

Thanks to these people for their suggestions and contributions:

• David Lecomber, Allinea
• Barry Rountree, LLNL

Known Bugs:

• Certain fortran bindings need some bugfixes and may not work.

Tutorial

For a thorough tutorial, look at examples/tutorial.w! It walks you through the process of using wrap.py. It is also legal wrap.py code, so you can run wrap.py on it and see the output to better understand what's going on.

CMake Integration

wrap.py includes a WrapConfig.cmake file. You can use this in your CMake project to automatically generate rules to generate wrap.py code.

Here's an example. Suppose you put wrap.py in a subdirectory of your project called wrap, and your project looks like this:

project/
    CMakeLists.txt
    wrap/
        wrap.py
        WrapConfig.cmake

In your top-level CMakeLists.txt file, you can now do this:

# wrap.py setup -- grab the add_wrapped_file macro.
set(WRAP ${PROJECT_SOURCE_DIR}/wrap/wrap.py)
include(wrap/WrapConfig.cmake)

If you have a wrapped source file, you can use the wrapper auto-generation like this:

add_wrapped_file(wrappers.C wrappers.w)
add_library(tool_library wrappers.C)

The add_wrapped_file function takes care of the dependences and code generation for you. If you need fortran support, call it like this:

add_wrapped_file(wrappers.C wrappers.w -f)

And note that if you generate a header that your .C files depend on, you need to explicitly include it in a target's sources, unlike non-generated headers. e.g.:

add_wrapped_file(my-header.h my-header.w)
add_library(tool_library
    tool.C         # say that this includes my-header.h
    my-header.h)   # you need to add my-header.h here.

If you don't do this, then the header dependence won't be accounted for when tool.C is built.

Wrapper file syntax

Wrap syntax is a superset of the syntax defined in Appendix C of the MPE manual [1], but many commands from the original wrapper generator are now deprecated.

The following two macros generate skeleton wrappers and allow delegation via {{callfn}}:

• fn iterates over only the listed functions.

• fnall iterates over all functions minus the named functions.

  {{fnall ... }} // code here {{endfnall}}

  {{fn ... }} {{endfn}

  {{callfn}}

callfn expands to the call of the function being profiled.

fnall defines a wrapper to be used on all functions except the functions named. fn is identical to fnall except that it only generates wrappers for functions named explicitly.

{{fn FOO MPI_Abort}}
	// Do-nothing wrapper for {{FOO}}
{{endfn}}

generates (in part):

/* ================== C Wrappers for MPI_Abort ================== */
_EXTERN_C_ int PMPI_Abort(MPI_Comm arg_0, int arg_1);
_EXTERN_C_ int MPI_Abort(MPI_Comm arg_0, int arg_1) {
    int return_val = 0;

// Do-nothing wrapper for MPI_Abort
    return return_val;
}

foreachfn and forallfn are the counterparts of fn and fnall, but they don't generate the skeletons (and therefore you can't delegate with {{callfn}}). However, you can use things like fn_name (or foo) and argTypeList, retType, argList, etc.

They're not designed for making wrappers, but declarations of lots of variables and other things you need to declare per MPI function. e.g., say you wanted a static variable per MPI call for some flag.

{{forallfn <iterator variable name> <function A> <function B> ... }}
  // code here
{{endforallfn}

{foreachfn <iterator variable name> <function A> <function B> ... }}
  // code here
{{endforeachfn}}

The code between {{forallfn}} and {{endforallfn}} is copied once for every function profiled, except for the functions listed. For example:

{{forallfn fn_name}}
  static int {{fn_name}}_ncalls_{{fileno}};
{{endforallfn}}

might expand to:

static int MPI_Send_ncalls_1;
static int MPI_Recv_ncalls_1;
...

etc.

• {{get_arg <argnum>}} OR {{<argnum>}} Arguments to the function being profiled may be referenced by number, starting with 0 and increasing. e.g., in a wrapper file:

    void process_argc_and_argv(int *argc, char ***argv) {
    // do stuff to argc and argv.
    }
  
    {{fn fn_name MPI_Init}}
        process_argc_and_argv({{0}}, {{1}});
    {{callfn}}
    {{endfn}}
  

  Note that {{0}} is just a synonym for {{get_arg 0}}

• {{ret_val}} ReturnVal expands to the variable that is used to hold the return value of the function being profiled. (was: {{returnVal}})

• {{fn_num}} This is a number, starting from zero. It is incremented every time it is used.

• {{ret_type}} The return type of the function. (was: {{retType}})

• {{formals}} Essentially what would be in a formal declaration for the function. Can be used this with forallfn and foreachfn; these don't generate prototypes, they just iterate over the functions without making a skeleton. (was: {{argTypeList}})

• {{args}} Names of the arguments in a comma-separated list, e.g.: buf, type, count, comm

• {{argList}} Same as {{args}}, but with parentheses around the list, e.g.: (buf, type, count, comm)

• {{applyToType <type> <callable>}} This macro must be nested inside either a fn or fnall block. Within the functions being wrapped by fn or fnall, this macro will apply <callable> to any arguments of the function with type <type>. For example, you might write a wrapper file like this:

    #define my_macro(comm) do_something_to(comm);
    {{fn fn_name MPI_Send MPI_Isend MPI_Ibsend}}
        {{applyToType MPI_Comm my_macro}}
        {{callfn}}
    {{endfn}}
  

Now the generated wrappers to MPI_Send, MPI_Isend, and MPI_Ibsend will do something like this:

int MPI_Isend(void *buf, int count, MPI_Datatype datatype, int dest, int tag, MPI_Comm comm, MPI_Request *request) {
  int _wrap_py_return_val = 0;
  my_macro(comm);
  PMPI_Isend(buf, count, datatype, dest, tag, comm, request);
}

• {{sub <new_string> <old_string> <regexp> <substitution>}} Declares <new_string> in the current scope and gives it the value of <old_string> with all instances of <regexp> replaced with <substitution>. You may use any valid python regexp for <regexp> and any valid substitution value for <substitution>. The regexps follow the same syntax as Python's re.sub(), and they may be single or double quoted (though it's not necessary unless you use spaces in the expressions).

  Example:

    {{forallfn foo}}
        {{sub nqjfoo foo '^MPI_' NQJ_}}
        {{nqjfoo}}
    {{endforallfn}}
  

  This will print NQJ_xxx instead of MPI_xxx for each MPI function.

• {{fileno}} An integral index representing which wrapper file the macro came from. This is useful when decalring file-global variables to prevent name collisions. Identifiers declared outside functions should end with _{{fileno}}. For example:

    static double overhead_time_{{fileno}};
  

  might expand to

    static double overhead_time_0;
  

• {{vardecl <type> <arg> <arg> ...}} (not yet supported) Declare variables within a wrapper definition. Wrap will decorate the variable name to prevent collisions.

• {{<varname>}} (not yet supported) Access a variable declared by {{vardecl}}.

Notes on the fortran wrappers

#if (!defined(MPICH_HAS_C2F) && defined(MPICH_NAME) && (MPICH_NAME == 1))
    /* MPICH call */
    return_val = MPI_Abort((MPI_Comm)(*arg_0), *arg_1);
#else
    /* MPI-2 safe call */
    return_val = MPI_Abort(MPI_Comm_f2c(*arg_0), *arg_1);
#endif

This is the part of the wrapper that delegates from Fortran to C. There are two ways to do that. The MPI-2 way is to call the appropriate _f2c call on the handle and pass that to the C function. The f2c/c2f calls are also available in some versions of MPICH1, but not all of them (I believe they were backported), so you can do the MPI-2 thing if MPICH_HAS_C2F is defined.

If c2f functions are not around, then the script tries to figure out if it's dealing with MPICH1, where all the handles are ints. In that case, you can just pass the int through.

Right now, if it's not specifically MPICH1, wrap.py does the MPI-2 thing. From what Barry was telling me, your MPI environment might have int handles, but it is not MPICH1. So you could either define all the MPI_Foo_c2f/MPI_Foo_f2c calls to identity macros, e.g.:

#define MPI_File_c2f(x) (x)
#define MPI_File_f2c(x) (x)

or you could add something to wrap.py to force the int-passing behavior. I'm not sure if you have to care about this, but I thought I'd point it out.

-s, or 'structural' mode

If you use the -s option, this skips the includes and defines used for C wrapper functions. This is useful if you want to use wrap to generate non-C files, such as XML.

If you use -s, we recommend that you avoid using {{fn}} and {{fnall}}, as these generate proper wrapper functions that rely on some of the header information. Instead, use {{foreachfn}} and {{forallfn}}, as these do not generate wrappers around each iteration of the macro.

e.g. if you want to generate a simple XML file with descriptions of the MPI arguments, you might write this in a wrapper file:

{{forallfn fun}}
    <function name="{{fun}}" args="{{args}}"/>
{{endforallfn}}

We don't disallow {{fnall}} or {{fn}} with -s, but If you used {{fnall}} here, each XML tag would have a C wrapper function around it, which is probably NOT what you want.

1. Anthony Chan, William Gropp and Weing Lusk. User's Guide for MPE: Extensions for MPI Programs. ANL/MCS-TM-ANL-98/xx. ftp://ftp.mcs.anl.gov/pub/mpi/mpeman.pdf