Input¶
The input to Shroud is a YAML formatted file. YAML is a human friendly data serialization standard. [yaml] Structure is shown through indentation (one or more spaces). Sequence items are denoted by a dash, and key value pairs within a map are separated by a colon:
library: Tutorial
declarations:
- decl: typedef int TypeID
- decl: void Function1()
- decl: class Class1
declarations:
- decl: void Method1()
Each decl
entry corresponds to a line of C or C++ code. The top
level declarations
field represents the source file while nested
declarations
fields corresponds to curly brace blocks.
The above YAML file represent the source file:
typedef int TypeID;
void Function1();
class Class1
{
void Method1();
}
A block
can be used to group a collection of decl
entires.
Any option
or format
fields will apply to all declarations in
the group:
declarations:
- block: True
options:
F_name_impl_template: {library}_{undescore_name}
format:
F_impl_filename: localfile.f
declarations:
- decl: void func1()
- decl: void func2()
Shroud use curly braces for format strings. If a string starts with a curly brace YAML will interpret it as a map/dictionary instead of as part of the string. To avoid this behavior, strings which start with a curly brace should be quoted:
name : "{fmt}"
Strings may be split across several lines by indenting the continued line:
- decl: void Sum(int len, int *values+dimension+intent(in),
int *result+intent(out))
Some values consist of blocks of code. The pipe, |
, is used to indicate that
the string will span several lines and that newlines should be preserved:
C_invalid_name: |
if (! isNameValid({cxx_var})) {{
return NULL;
}}
Note that to insert a literal {
, a double brace, {{
, is
required since single braces are used for variable expansion.
{cxx_var}
in this example.
Literal newlines, /n
, are respected. Format strings can use a
tab, /t
, to hint where it would be convenient to add a
continuation if necessary. A formfeed, /f
, will force a
continuation. Lines which start with 0
are not indented. This
can be used with labels. A trailing +
will indent then next line
a level and a leading -
will deindent. Line lengths are controlled
by the options C_line_length and F_line_length and default to
72.:
C_invalid_name: |
if (! isNameValid({cxx_var})) {{+
return NULL;
-}}
The only formatting option is to control output line lengths. This is
required for Fortran which has a maximum line length of 132 in free
form which is generated by shroud. If you care where curly braces go
in the C source then it is best to set C_line_length to a large
number then use an external formatting tool such as indent
or
uncrustify
.
Customizing Behavior in the YAML file¶
Fields¶
A field only applies to the type, enumeration, function, structure or class to which it belongs. It is not inherited. For example, cxx_header is a field which is used to define the header file for class Names. Likewise, setting library within a class does not change the library name.
library: testnames
declarations:
- decl: class Names
cxx_header: names.hpp
declarations:
- decl: void method1
Options¶
Options are used to customize the behavior of Shroud. They are defined in the YAML files as a dictionary. Options can be defined at the global, class, or function level. Each level creates a new scope which can access all upper level options. This allows the user to modify behavior for all functions or just a single one:
options:
option_a = false
option_b = false
option_c = false
declarations:
- class: class1
options:
# option_a = false # inherited
option_b = true
# option_c = false # inherited
declarations:
- decl: void function1
options:
# option_a = false # inherited
# option_b = true # inherited
option_c = true
Format¶
A format dictionary contains strings which can be inserted into generated code. Generated filenames are also entries in the format dictionary. Format dictionaries are also scoped like options. For example, setting a format in a class also effects all of the functions within the class.
How code is formatted¶
Format strings contain “replacement fields” surrounded by curly braces
{}
. Anything that is not contained in braces is considered literal
text, which is copied unchanged to the output. If you need to include
a brace character in the literal text, it can be escaped by doubling:
{{
and }}
. [Python_Format]
There are some metacharacters that are used for formatting the line:
\f
Add an explicit formfeed
\t
A tab is used to suggest a place to break the line for a continuation before it exceeds option C_line_length or F_line_length. Any whitespace after a tab will be trimmed if the line is actually split at the tab. If a continuation was not needed (there was enough space on the current line) then the tab has no effect:
arg1,\t arg2
+ -
Increase or decrease indention indention level. Used at the beginning or end of a line:
if (condition) {{+ do_one(); -}} else {{+ do_two(); -}}The double curly braces are replace by a single curly. This will be indented as:
if (condition) { do_one(); } else { do_two(); }
#
If the first character is a #, ignore indention and write in column 0. Useful for preprocessing directives.
^
If the first character is ^, ignore indention and write in column 0. Useful for comments or labels.
@
If the first character is @, treat the following character literally. Used to ignore a metacharacter:
struct aa = {{++ 0// set field to 0 @0, -}};Formatted as:
struct aa = { // set field to 0 0, };
Attributes¶
Annotations or attributes apply to specific arguments or results. They describe semantic behavior for an argument. An attribute may be set to true by listing its name or it may have a value in parens:
- decl: Class1() +name(new)
- decl: void Sum(int len, int *values+dimension+intent(in))
- decl: const std::string getName() +len(30)
Attributes may also be added external to decl:
- decl: void Sum(int len, int *values)
attrs:
values:
dimension: True
intent: in
- decl: const std::string getName()
fattrs:
len: 30
allocatable¶
Sometimes it is more convenient to have the wrapper allocate an
intent(out)
array before passing it to the C++ function. This can
be accomplished by adding the allocatable attribute. For example the
C++ function cos_doubles
takes the cosine of an intent(in)
argument and assigns it to an intent(out)
argument:
void cos_doubles(double *in, double *out, int size)
{
for(int i = 0; i < size; i++) {
out[i] = cos(in[i]);
}
}
This is wrapped as:
decl: void cos_doubles(double * in +intent(in) +dimension(:),
double * out +intent(out) +allocatable(mold=in),
int sizein +implied(size(in)))
The mold argument is similar to the mold argument in the Fortran
allocate
statement, it will allocate out
as the same shape as
in
. Also notice the use of the implied attribute on the
size
argument. This argument is not added to the Fortran API
since its value is implied to be the size of argument in
.
size
is the Fortran intrinsic which returns the number of items
allocated by its argument.
The Fortran wrapper produced is:
subroutine cos_doubles(in, out)
use iso_c_binding, only : C_DOUBLE, C_INT
real(C_DOUBLE), intent(IN) :: in(:)
real(C_DOUBLE), intent(OUT), allocatable :: out(:)
integer(C_INT) :: sizein
allocate(out, mold=in)
sizein = size(in)
call c_cos_doubles(in, out, sizein)
end subroutine cos_doubles
The mold argument was added to the Fortran 2008 standard. If the option F_standard is not 2008 then the allocate statement will be:
allocate(out(lbound(in,1):ubound(in,1)))
For Python, a similar NumPy array object will be constructed using
PyArray_NewLikeArray
.
assumedtype¶
When this attribute is applied to a void *
argument, the Fortran
assumed-type declaration, type(*)
, will be used. Since Fortran
defaults to pass-by-reference, the argument will be passed to C as a
void *
argument. The C function will need some other mechanism to
determine the type of the argument before dereferencing the pointer.
Note that assumed-type is part of Fortran 2018.
charlen¶
charlen is used to define the size of a char *arg+intent(out)
argument in the Python wrapper. This deals with the case where arg
is provided by the user and the function writes into the provided
space. This technique has the inherent risk of overwritting memory if
the supplied buffer is not long enough. For example, when used in C
the user would write:
#define API_CHARLEN
char buffer[API_CHARLEN];
fill_buffer(buffer);
The Python wrapper must know the assumed length before calling the
function. It will then be converted into a str object by
PyString_FromString
.
Fortran does not use this attribute since the buffer argument is supplied by the user. However, it is useful to provide the parameter by adding a splicer block in the YAML file:
splicer_code:
f:
module_top:
- "integer, parameter :: MAXNAME = 20"
Warning
Using charlen and dimension together is not currently supported.
default¶
Default value for C++ function argument. This value is implied by C++ default argument syntax.
deref¶
List how to dereference pointer arguments or function results. This is used in conjunction with dimension to create arrays.
scalar
Treat the pointee as a scalar. For Python, this will not create a NumPy object.
pointer
For Fortran, add
POINTER
attribute to argument and is associated with the argument usingc_f_pointer
. If owner(caller) is also defined, add an additional argument which is used to release the memory.For Python, create a NumPy array.
allocatable
For Fortran, add
ALLOCATABLE
attribute to argument. AnALLOCATE
is added and the contents of the C++ argument is copied. If owner(caller) is also defined, the C++ argument is released. The caller is responsible toDEALLOCATE
the array.For Python, create a NumPy array (same as pointer)
raw
For Fortran, return a
type(C_PTR)
.For Python, return a
PyCapsule
.
dimension¶
Sets the Fortran DIMENSION attribute.
Pointer argument should be passed through since it is an array.
value attribute must not be True.
If set without a value, it defaults to (*)
:
double *array +dimension
double *array +dimension(len)
external¶
This attribute is only valid with function pointers. It will ensure
that a Fortran wrapper is created which uses the external
statement for the argument. This will allow any function to be used
as the dummy argument for the function pointer.
free_pattern¶
A name in the patterns section which lists code to be used to
release memory. Used with function results.
It is used in the C_memory_dtor_function and will have the
variable void *ptr
available as the pointer to the memory
to be released.
See Memory Management for details.
implied¶
The value of an arguments to the C++ function may be implied by other arguments. If so the implied attribute can be used to assign the value to the argument and it will not be included in the wrapped API.
Used to compute value of argument to C++ based on argument to Fortran or Python wrapper. Useful with array sizes:
int Sum(int * array +intent(in), int len +implied(size(array))
Several functions will be converted to the corresponding code for
Python wrappers: size
, len
and len_trim
.
intent¶
Valid valid values are in
, out
, inout
.
If the argument is const
, the default is in
.
len¶
For a string argument, pass an additional argument to the
C wrapper with the result of the Fortran intrinsic len
.
If a value for the attribute is provided it will be the name
of the extra argument. If no value is provided then the
argument name defaults to option C_var_len_template.
When used with a function, it will be the length of the return value of the function using the declaration:
character(kind=C_CHAR, len={c_var_len}) :: {F_result}
len_trim¶
For a string argument, pass an additional argument to the
C wrapper with the result of the Fortran intrinsic len_trim
.
If a value for the attribute is provided it will be the name
of the extra argument. If no value is provided then the
argument name defaults to option C_var_trim_template.
name¶
Name of the method. Useful for constructor and destructor methods which have no names.
owner¶
Specifies who is responsible to release the memory associated with the argument/result.
The terms follow Python’s reference counting . [Python_Refcount] The default is set by option default_owner which is initialized to borrow.
caller
The memory belongs to the user who is responsible to delete it. A shadow class must have a destructor wrapped in order to delete the memory.
library
The memory belongs to the library and should not be deleted by the user. This is the default value.
value¶
If true, pass-by-value; else, pass-by-reference.
This attribute is implied when the argument is not a pointer or reference.
This will also default to intent(IN)
since there is no way to return
a value.
Note
The Fortran wrapper may use an intrinsic function for some attributes. For example, len, len_trim, and size. If there is an argument with the same name, the generated code may not compile.
Shroud preserves the names of the arguments since Fortran
allows them to be used in function calls - call worker(len=10)
Patterns¶
To address the issue of semantic differences between Fortran and C++, patterns may be used to insert additional code. A pattern is a code template which is inserted at a specific point in the wrapper. They are defined in the input YAML file:
declarations:
- decl: const string& getString2+len=30()
C_error_pattern: C_invalid_name
patterns:
C_invalid_name: |
if ({cxx_var}.empty()) {{
return NULL;
}}
The C_error_pattern will insert code after the call to the C++
function in the C wrapper and before any post_call sections from the
types. The bufferified version of a function will append
_buf
to the C_error_pattern value. The pattern is
formatted using the context of the return argument if present,
otherwise the context of the function is used. This means that
c_var and c_var_len refer to the argument which is added to
contain the function result for the _buf
pattern.
The function getString2
is returning a std::string
reference.
Since C and Fortran cannot deal with this directly, the empty string
is converted into a NULL
pointer::
will blank fill the result:
const char * STR_get_string2()
{
const std::string & SHCXX_rv = getString2();
// C_error_pattern
if (SHCXX_rv.empty()) {
return NULL;
}
const char * SHC_rv = SHCXX_rv.c_str();
return SHC_rv;
}
Splicers¶
No matter how many features are added to Shroud there will always exist cases that it does not handle. One of the weaknesses of generated code is that if the generated code is edited it becomes difficult to regenerate the code and preserve the edits. To deal with this situation each block of generated code is surrounded by ‘splicer’ comments:
const char * STR_get_char3()
{
// splicer begin function.get_char3
const char * SH_rv = getChar3();
return SH_rv;
// splicer end function.get_char3
}
These comments delineate a section of code which can be replaced by
the user. The splicer’s name, function.get_char3
in the example,
is used to determine where to insert the code.
There are two ways to define splicers in the YAML file. First add a list of files which contain the splicer text:
splicer:
f:
- fsplicer.f
c:
- csplicer.c
In the listed file, add the begin and end splicer comments,
then add the code which should be inserted into the wrapper inbetween the comments.
Multiple splicer can be added to an input file. Any text that is not within a
splicer block is ignored. Splicers must be sorted by language. If
the input file ends with .f
or .f90
it is processed as
splicers for the generated Fortran code. Code for the C wrappers must
end with any of .c
, .h
, .cpp
, .hpp
, .cxx
,
.hxx
, .cc
, .C
:
-- Lines outside blocks are ignore
// splicer begin function.get_char3
const char * SH_rv = getChar3();
SH_rv[0] = 'F'; // replace first character for Fortran
return SH_rv + 1;
// splicer end function.get_char3
This technique is useful when the splicers are very large or are generated by some other process.
The second method is to add the splicer code directly into the YAML file.
A splicer can be added after the decl
line.
- decl: bool isNameValid(const std::string& name)
splicer:
c:
- "return name != NULL;"
f:
- 'rv = name .ne. " "'
A splicer can be added in the splicer_code
section.
This can be used to add code to spliers which do not correspond
directly to a declaration.
Each level of splicer is a mapping and each line of text is an array entry:
splicer_code:
c:
function:
get_char3:
- const char * SH_rv = getChar3();
- SH_rv[0] = 'F'; // replace first character for Fortran
- return SH_rv + 1;
In addition to replacing code for a function wrapper, there are splicers that are generated which allow a user to insert additional code for helper functions or declarations:
! file_top
module {F_module_name}
! module_use
implicit none
! module_top
type class1
! class.{cxx_class}.component_part
contains
! class.{cxx_class}.generic.{F_name_generic}
! class.{cxx_class}.type_bound_procedure_part
end type class1
interface
! additional_interfaces
end interface
contains
! function.{F_name_function}
! {cxx_class}.method.{F_name_function}
! additional_functions
end module {F_module_name}
C header:
// class.{class_name}.CXX_declarations
extern "C" {
// class.{class_name}.C_declarations
}
C implementation:
// class.{class_name}.CXX_definitions
extern "C" {
// class.{class_name}.C_definitions
// function.{underscore_name}{function_suffix}
// class.{cxx_class}.method.{underscore_name}{function_suffix}
}
The splicer comments can be eliminated by setting the option show_splicer_comments to false. This may be useful to eliminate the clutter of the splicer comments.
Footnotes
[Python_Format] | https://docs.python.org/2/library/string.html#format-string-syntax |
[Python_Refcount] | https://docs.python.org/3/c-api/intro.html#reference-count-details |
[yaml] | yaml.org |