utils

C++ Converter Utilities

This module provides utility functions to convert numpy arrays and binary files to C++ format.

Functions:

• array_dump –

  Generate C array of values from flattened numpy array.

• xxd_c_dump –

  Generate C array of values from binary file.

Functions

array_dump

array_dump(data: npt.NDArray, dst_path: os.PathLike, var_name: str = 'test_stimulus', var_dtype: str | None = None, row_len: int = 12, is_header: bool = False)

Generate C array of values from flattened numpy array.

Parameters:

• data (NDArray) –

  Data array

• dst_path (PathLike) –

  C file destination path

• var_name (str, default: 'test_stimulus' ) –

  C variable name. Defaults to "test_stimulus".

• var_dtype (str | None, default: None ) –

  C variable type. Defaults to None.

• row_len (int, default: 12 ) –

  Elements to write per row. Defaults to 12.

• is_header (bool, default: False ) –

  Write as header or source C file. Defaults to source.

Example:

data = np.random.rand(10, 10).astype(np.float32)
dst_path = "test_stimulus.h"
nse.converters.cpp.array_dump(
    data,
    dst_path,
    var_name="test_stimulus",
    var_dtype="float32_t",
    row_len=12,
    is_header=True
)

Source code in neuralspot_edge/converters/cpp/utils.py

def array_dump(
    data: npt.NDArray,
    dst_path: os.PathLike,
    var_name: str = "test_stimulus",
    var_dtype: str | None = None,
    row_len: int = 12,
    is_header: bool = False,
):
    """Generate C array of values from flattened numpy array.

    Args:
        data (npt.NDArray): Data array
        dst_path (PathLike): C file destination path
        var_name (str, optional): C variable name. Defaults to "test_stimulus".
        var_dtype (str | None, optional): C variable type. Defaults to None.
        row_len (int, optional): Elements to write per row. Defaults to 12.
        is_header (bool): Write as header or source C file. Defaults to source.

    Example:

    ```python
    data = np.random.rand(10, 10).astype(np.float32)
    dst_path = "test_stimulus.h"
    nse.converters.cpp.array_dump(
        data,
        dst_path,
        var_name="test_stimulus",
        var_dtype="float32_t",
        row_len=12,
        is_header=True
    )
    ```
    """
    data = data.flatten()

    if var_dtype is not None:
        pass
    elif isinstance(data[0], np.floating):
        var_dtype = "float"
    elif isinstance(data[0], np.int8):
        var_dtype = "int8_t"
    elif isinstance(data[0], np.int16):
        var_dtype = "int16_t"
    elif isinstance(data[0], np.integer):
        var_dtype = "int32_t"
    else:
        raise ValueError("Unsupported dtype")

    with open(dst_path, "w", encoding="UTF-8") as wfp:
        if is_header:
            wfp.write(f"#ifndef __{var_name.upper()}_H{os.linesep}")
            wfp.write(f"#define __{var_name.upper()}_H{os.linesep}")

        wfp.write(f"#include <cstdint>{os.linesep}{os.linesep}")

        wfp.write(f"const {var_dtype} {var_name}[] = {{{os.linesep}")
        for row in range(0, len(data), row_len):
            wfp.write("  " + ", ".join((str(val) for val in data[row : row + row_len])) + f", {os.linesep}")
        # END FOR
        wfp.write(f"}};{os.linesep}")
        wfp.write(f"const unsigned int {var_name}_len = {len(data)};{os.linesep}")
        if is_header:
            wfp.write(f"#endif // __{var_name.upper()}_H{os.linesep}")

xxd_c_dump

xxd_c_dump(src_path: os.PathLike, dst_path: os.PathLike, var_name: str = 'model', chunk_len: int = 12, is_header: bool = False)

Generate C like char array of hex values from binary source. Equivalent to xxd -i src_path > dst_path but with added features to provide # columns and variable name.

Parameters:

• src_path (PathLike) –

  Binary file source path

• dst_path (PathLike) –

  C file destination path

• var_name (str, default: 'model' ) –

  C variable name. Defaults to 'model'.

• chunk_len (int, default: 12 ) –

  Number of elements per row. Defaults to 12.

• is_header (bool, default: False ) –

  Write as header or source C file. Defaults to source.

Example:

src_path = "model.tflite"
dst_path = "model.h"
nse.converters.cpp.xxd_c_dump(
    src_path,
    dst_path,
    var_name="model",
    chunk_len=12,
    is_header=True
)

Source code in neuralspot_edge/converters/cpp/utils.py

def xxd_c_dump(
    src_path: os.PathLike,
    dst_path: os.PathLike,
    var_name: str = "model",
    chunk_len: int = 12,
    is_header: bool = False,
):
    """Generate C like char array of hex values from binary source. Equivalent to `xxd -i src_path > dst_path`
        but with added features to provide # columns and variable name.

    Args:
        src_path (PathLike): Binary file source path
        dst_path (PathLike): C file destination path
        var_name (str, optional): C variable name. Defaults to 'model'.
        chunk_len (int, optional): Number of elements per row. Defaults to 12.
        is_header (bool): Write as header or source C file. Defaults to source.

    Example:

    ```python
    src_path = "model.tflite"
    dst_path = "model.h"
    nse.converters.cpp.xxd_c_dump(
        src_path,
        dst_path,
        var_name="model",
        chunk_len=12,
        is_header=True
    )
    ```
    """
    var_len = 0
    with open(src_path, "rb", encoding=None) as rfp, open(dst_path, "w", encoding="UTF-8") as wfp:
        if is_header:
            wfp.write(f"#ifndef __{var_name.upper()}_H{os.linesep}")
            wfp.write(f"#define __{var_name.upper()}_H{os.linesep}")

        wfp.write(f"const unsigned char {var_name}[] = {{{os.linesep}")
        for chunk in iter(lambda: rfp.read(chunk_len), b""):
            wfp.write("  " + ", ".join((f"0x{c:02x}" for c in chunk)) + f", {os.linesep}")
            var_len += len(chunk)
        # END FOR
        wfp.write(f"}};{os.linesep}")
        wfp.write(f"const unsigned int {var_name}_len = {var_len};{os.linesep}")
        if is_header:
            wfp.write(f"#endif // __{var_name.upper()}_H{os.linesep}")