TFF n-d array

Version 0.1

General-purpose n-dimensional array for C++, with many features useful for handling dense data. Developed as part of timed flow framework (TFF), but independent. Encapsulates low-level issues such as indexing and alignment in dealing with raw data, allowing for readable and efficient application code. Extensively tested.

Some features:

• ndarray<Dim, T> type inspired by Numpy's ndarray. C++ Container with random-access iterators. Sectioning and slicing giving non-owning read-only or read-write ndarray_view<Dim, T> with same interface. Byte-level strided data. Axis can be reversed. Convenient slicing syntaxes. Optimized copying and comparing for POD element types T. Row-major, column-major, or any other data ordering. Timed variant for each view, where absolute time index is associated to first dimension.

• Intra-element slicing for appropriate element types. For example ndarray_view<2, std::array<int, 3>> can be accessed as ndarray_view<3, int>. Standard-layout tuple type elem_tuple, allowing access to array of single component in packed multi-component n-d array.

• Wrap-around view of a smaller ndarray_view, where coordinate that cross the boundaries of the original view are mapped in a circular fashion. Same interface as normal views, including iteration and further sectioning.

• Opaque view where elements are replaced by frames or runtime-determined size and structure. Casting from and back to concrete ndarray_view, rendering given number of inner dimensions opaque. Provides type erasure and removes unnecessary templatization in application code operating on arbitrary data. For example ndarray_view<3, int> may be casted to ndarray_opaque_view<1, ndarray_frame>, and then back to ndarray_view<3, int>. Runtime verification of type. For POD-data (even strided), assignment and comparison possible in opaque state.

• Opaque array which is created and allocated in opaque state. Frames of may be of application-defined structure, instead of being casted ndarrays. Runtime-determined alignment requirement of frames is respected. Integrates application-defined frame handle class for convenient access to frame content.

• Any combination of these. For example 1D opaque array used for type-generic TFF ring buffers. Timed wraparound view to section of the array for readable and writable segment. For frames containing n-d arrays of given data type T, the section view is casted to a concrete ndarray_wraparound_view<1 + Frame_dim, T>, which is passed to application code.

Possible future features:

• Element-wise arithmetic or other operations on ndarray_view, possibly parallelized execution.
• More features from Numpy ndarray.
• Convolution operations with n-dimensional kernel. Masked arrays.
• Arrays with non-contiguous memory, possibly partially offloaded to secondary storage.
• Oblique slices, for example sloped line in 2D image. Iteration using Bresenham's line algorithm or similar.
• Virtual memory mapping optimization for wrap-around, allowing contiguous memory access across border.
• Helper functions for passing data to and from OpenCL or other accelerator API.
• Subset of library useable from inside C++ OpenCL kernel code.