Quickstart

How do I build and read one array?

#include <aether/aether.h>, construct an aether::Array<T, Es...> with however many samples you need, and read/write it through the View its .hostView() hands back — indexed (component, sample), the same shape you would use from NumPy with the axes swapped.

The umbrella header

aether/aether.h pulls in the whole library: layout, views, expression templates, arrays, random numbers, residency — everything this site’s pages cover. There is no separate install step or Python package to reach it from (see Interoperability for how aether reaches Python at all) — it is a header-only C++23 library, so a consumer just includes it and links nothing extra beyond the compiler’s own runtime.

aether::Array

An aether::Array<T, Es...> (aether/array/Array.h) is the OWNING convenience type: it allocates its own memory and follows the symmetry rule — Array<T, Es...> always appends one more, dynamic dimension (the number of samples) on top of the Es... you name, so Array<double, 3> is an array of N 3-vectors, not a single one. .samples() reports how many there are; .size() reports the total scalar count (samples() * 3 for a 3-vector array).

On a build with a CUDA backend, an Array actually owns TWO buffers — a host copy and a device copy — and .upload()/.download() move data between them (see Memory ownership: Chunk, Array, View). In a CPU-only build there is only one buffer, so those calls are harmless no-ops; the code above compiles and behaves identically either way, which is the whole point of the dual-mode design this library is built on.

A runnable example

TEST_F(ArrayTest, SamplesAndSize)
{
    aether::Array<double, 3> arr(10);
    EXPECT_EQ(arr.samples(), 10u);
    EXPECT_EQ(arr.size(), 30u);
}

Ten 3-component samples, so samples() == 10 and size() == 30. From here, Views and Items covers how to actually read and write the elements through a View, and Memory ownership: Chunk, Array, View covers what Array is doing underneath.