Memory Limitations

[Nucs]

Allocation currently supports up to 2^32 bytes due to using int and not IntPtr and long.

[Nucs]

Unit test

[TestMethod]
public void MyTestMethod()
{
    NDArray x = np.zeros(new Shape(600, 1000, 1000), np.float32);
}

[Nucs]

Ported UnmanagedMemoryBlock and ArraySlice to use long in commit 539683f .
Next is to port Shape to use long and all the algorithms with it. Mainly refactoring job

[Nucs]

Done via the int→long rewrite (#596, in 0.70.0 #628). The int/IntPtr allocation cap is gone: Shape size/dimensions/strides/offset are long (npy_intp parity) and the sort/select core is 64-bit, so arrays past int.MaxValue elements work:

var a = np.arange(6).reshape(2, 3);
Console.WriteLine(string.Join(",", a.strides));   // 24,8  (bytes per axis, npy_intp)

The >2 GB / >int.MaxValue paths are gated by the [HighMemory] NpyLargeFileTests and a >int.MaxValue byte oracle. Closing.