Dillon

FAQ initial revision

1

Here's the first few things I thought to put in an "FAQ". It's written as a test (like readme.cpp).

When allocators are expensive to construct (e.g. `auto_allocator` uses a lot of stack space), the current `make_array` and similar helpers are problematic, because they use a default argument for the...

Remove copies from Einstein expression construction

1

Currently, it is not possible to construct an Einstein reduction expression without copying the operands. Most of these copies are small/lightweight (never arrays, only array_refs), but operands that are functions/lambdas...

Consider adding "flat iterators" for arrays

5

It would be great to have "flat iterators" for arrays. Among other things, this could enable the standard `` algorithms to work on arrays where it makes sense, and we...

whole-matrix einsum is slower than hand-written code

2

The current timing I see for the matrix tests is: ``` reduce_matrix time: 10.6356 ms einsum time: 13.4737 ms reduce_tiles time: 0.890935 ms einsum_tiles time: 0.929729 ms ``` The einsum_tiles...

When accessing a dim of a shape, ideally we would allow simply `auto x = s.dim()`. Unfortunately, the reality of template syntax issues in many situations means it has to...

shape allows construction from incompatible shapes (std::is_constructible and std::is_assignable are broken for std::tuple on various STLs)

4

Some compilers have a broken implementation of `std::is_constructible`, which makes it hard to prevent shapes from being constructed/assigned from incompatible shapes.

It would be better if shapes/dims and ranges were better differentiated. They are very similar, but differ in the following ways: 1. shapes/dims have strides, ranges do not. 2. Typically,...

Cropping via ranges results in an array with the same indices as the old array

1

In numpy and many other systems, `b = a(interval(5, 12))` results in an array where `b(x) == a(x + 5)` and b has an extent of 7 (and a min...