NumPy / buffer views

Nuwa SDK wraps existing Python objects that export PEP 3118 (NumPy ndarrays, many memoryviews). It does not allocate or return ndarrays. To give Python a new array, copy into a seq (or build an array in Python).

Read and write

import nuwa_sdk

proc sumInt64(arr: PyObject): int64 {.nuwa_export.} =
  var view = asNumpyArray(arr, int64)
  result = 0
  for x in view:
    result += x

proc scaleInPlace(arr: PyObject, s: float64) {.nuwa_export.} =
  var view = asNumpyArrayWrite(arr, float64)
  for x in mitems(view):
    x = x * s

asStridedArray is the same constructor under another name.

Types

Use a Nim element type that matches the array dtype: int8–int64, uint8–uint64, float32, float64, bool, plus platform int / uint. Wrong dtype raises TypeError. Byteswapped (non-native endian) multi-byte dtypes are rejected.

Complex, datetime, and object arrays are not supported.

Layout

API	Meaning
isContiguous	C-order (row-major) only
isFortranContiguous	Column-major
data	Pointer to the contiguous block if C- or Fortran-contiguous. Fortran data[i] is memory (column) order.
items	Logical C-order walk, including for Fortran arrays
arr[i]	1D only. 2D+ uses arr[i, j, ...]

Non-contiguous views (slices, some transposes) use strided indexing. data raises LayoutError; copy in Python with np.ascontiguousarray / np.asfortranarray if you need a pointer.

GIL

Acquire the view with the GIL held, then take data / len and run pure Nim inside withNogil. Do not call Python or nimpy inside that block.

Cleanup

The wrapper releases the Py_buffer when it goes out of scope. close() is optional and needs a var binding.

Read-only NumPy arrays (setflags(write=False)) work with asNumpyArray and fail with asNumpyArrayWrite.