Basic Usage

This section describes how to customize a VE function.

Constructing Custom VE Library

There are two ways to construct a custom VE library:

A. Compiling and Loading From Source Code

1. Define source code as a Python string.

   >>> c_src=r'''
   ...     int ve_add(double *px, double *py, double *pz, int n) {
   ...         #pragma omp parallel for
   ...         for (int i = 0; i  < n; i++) pz[i] = px[i] + py[i];
   ...         return 0;
   ...     }
   ... '''
   

   注釈

   When using the tripple quotes """ or ''', always prepend them by r such that the content is interpreted as raw string. Otherwise the escaped characters will be interpreted and spoil the source code.

2. Compile and Load from the Python string.

   >>> ve_lib = nlcpy.jit.CustomVELibrary(code=c_src)
   

B. Loading From Pre-built Binary

>>> ve_lib = nlcpy.jit.CustomVELibrary(path='/path/to/existing/library.so')

If you need complicated build rules, we recommend using case B.

参考

nlcpy.jit.CustomVELibrary

Getting a Symbol of a Function

You need to call CustomVELibrary.get_function() to get a symbol of a function on VE.

>>> from nlcpy import ve_types
>>> ve_add = ve_lib.get_function(
...     've_add',
...     args_type=(ve_types.uint64, ve_types.uint64, ve_types.uint64, ve_types.int32),
...     ret_type=ve_types.int32
... )

ve_add is the instance of CustomVEKernel class.

The tuple of character string element or the tuple of constant element of the nlcpy.ve_types can be passed into the argument args_type. Also, the character string or the constant of the nlcpy.ve_types can be passed into the argument ret_type.

The supported data-types and its corresponding notations are as follows:

data-type description	character strings	constants of nlcpy.ve_types
void	'void'	nlcpy.ve_types.void
8-bit character	'char'	nlcpy.ve_types.char
32-bit signed integer	'int32_t'	nlcpy.ve_types.int32
64-bit signed integer	'int64_t'	nlcpy.ve_types.int64
32-bit unsigned integer	'uint32_t'	nlcpy.ve_types.uint32
64-bit unsigned integer	'uint64_t'	nlcpy.ve_types.uint64
32-bit floating-point real	'float'	nlcpy.ve_types.float32
64-bit floating-point real	'double'	nlcpy.ve_types.float64
void pointer	'void *'	nlcpy.ve_types.void_p

注釈

For the complex data-type, it is necessary to transfer data as nlcpy.ndarray or nlcpy.veo.OnStack. Please refer to the Advanced Topics.

参考

nlcpy.jit.CustomVELibrary.get_function()

参考

nlcpy.jit.CustomVEKernel

Execution

Here, you can execute the VE function.

>>> x = nlcpy.arange(10., dtype='f8')
>>> y = nlcpy.arange(10., dtype='f8')
>>> z = nlcpy.empty(10, dtype='f8')
>>> ret = ve_add(x.ve_adr, y.ve_adr, z.ve_adr, z.size, sync=True)
>>> z
array([ 0.,  2.,  4.,  6.,  8., 10., 12., 14., 16., 18.])
>>> ret
0

Just only pass the attribute ndarray.ve_adr into arguments, the VE function can get the pointer of the array.

注釈

If you pass the argument sync=False, the return value will be None.

注釈

You can invoke the VE function without recompiling by calling CustomVEKernel.__call__() repeatedly.

参考

nlcpy.jit.CustomVEKernel.__call__()

Finalization (Optional)

If needed, you can unload the shared library.

>>> nlcpy.jit.unload_library(ve_lib)

制限事項

Please avoid unloading the shared library linked with FTRACE. Otherwise, SIGSEGV may occur.

参考

nlcpy.jit.unload_library()