SeqAn3 3.4.0-rc.1
The Modern C++ library for sequence analysis.
|
This tutorial introduces the notion of ranges, a C++20 feature that SeqAn3 makes strong use of.
Difficulty | Moderate |
---|---|
Duration | 90 min |
Prerequisite tutorials | C++ Concepts |
Recommended reading |
Traditionally most generic algorithms in the C++ standard library, like std::sort, take a pair of iterators (e.g. the object returned by begin()
). If you want to sort a std::vector v
, you have to call std::sort(v.begin(), v.end())
and not std::sort(v)
. Why was this design with iterators chosen? It is more flexible, because it allows e.g.:
std::sort(v.begin() + 5, v.end())
std::sort(v.rbegin() + 5, v.rend())
(sorts in reverse order)But this interface is less intuitive than just calling std::sort on the entity that you wish to sort and it allows for more mistakes, e.g. mixing two incompatible iterators. C++20 introduces the notion of ranges and provides algorithms that accept such in the namespace std::ranges::
, e.g. std::ranges::sort(v)
now works if v
is range – and vectors are ranges!
What about the two examples that suggest superiority of the iterator-based approach? In C++20 you can do the following:
std::ranges::sort(std::views::drop(v, 5))
std::ranges::sort(std::views::reverse(v))
We will discuss later what std::views::reverse(v)
does. For now, it is enough to understand that it returns something that appears like a container and that std::ranges::sort can sort it. Later we will see that this approach offers even more flexibility than working with iterators.
Ranges are an abstraction of "a collection of items", or "something iterable". The most basic definition requires only the existence of begin()
and end()
on the range.
There are different ways to classify ranges, one way is through the capabilities of its iterator.
Ranges are typically either input ranges (they can be read from) or output ranges (they can be written to) or both. E.g. a std::vector<int>
is both, but a std::vector<int> const
would only be an input range.
Input ranges have different strengths that are realised through more refined concepts (i.e. types that model a stronger concept, always also model the weaker one):
Concept | Description |
---|---|
std::ranges::input_range | can be iterated from beginning to end at least once |
std::ranges::forward_range | can be iterated from beginning to end multiple times |
std::ranges::bidirectional_range | iterator can also move backwards with -- |
std::ranges::random_access_range | you can jump to elements in constant-time [] |
std::ranges::contiguous_range | elements are always stored consecutively in memory |
For the well-known containers from the standard library this matrix shows which concepts they model:
std::forward_list | std::list | std::deque | std::array | std::vector | |
---|---|---|---|---|---|
std::ranges::input_range | ✅ | ✅ | ✅ | ✅ | ✅ |
std::ranges::forward_range | ✅ | ✅ | ✅ | ✅ | ✅ |
std::ranges::bidirectional_range | ✅ | ✅ | ✅ | ✅ | |
std::ranges::random_access_range | ✅ | ✅ | ✅ | ||
std::ranges::contiguous_range | ✅ | ✅ |
There are also range concepts that are independent of input or output or one of the above concepts, e.g. std::ranges::sized_range which requires that the size of a range can be computed and in constant time.
Containers are the ranges most well known, they own their elements. SeqAn3 makes use of standard STL containers like std::vector
, but also implements some custom containers.
Decorators are ranges that are always defined on another range and decorate/annotate the underlying range with additional information. They do not own the underlying range but can contain member data of their own.
Views are ranges that are usually defined on another range and transform the underlying range via some algorithm or operation. Views do not own any data beyond their algorithm and the time it takes to construct, destruct or copy them should not depend on the number of elements they represent. The algorithm is required to be lazy-evaluated so it is feasible to combine multiple views. More on this below.
If you are confused about decorators vs views, think of decorators as "underlying range + data" and views as "underlying range + algorithm".
The storage behaviour is orthogonal to the range concepts defined by the iterators mentioned above, i.e. you can have a container that satisfies std::ranges::random_access_range (e.g. std::vector
does, but std::list
does not) and you can have views or decorators that do so or don't. For some combinations of iterator capabilities and storage behaviour there are extra concept definitions, e.g. seqan3::random_access_container.
As mentioned above, views are a specific kind of range. They are incredibly useful and you will find them throughout the library.
A key feature of views is that whatever transformation they apply, they do so at the moment you request an element, not when the view is created.
Here v
is a view; creating it neither changes vec
, nor does v
store any elements. The time it takes to construct v
and its size in memory is independent of the size of vec
.
This will print "6", but the important thing is that resolving the first element of v
to the last element of vec
happens on-demand. This guarantees that views can be used as flexibly as iterators, but it also means that if the view performs an expensive transformation, it will have to do so repeatedly if the same element is requested multiple times.
You may have wondered why we wrote
and not
That's because std::views::reverse
is not the view itself, it's an adaptor that takes the underlying range (in our case the vector) and returns a view object over the vector. The exact type of this view is hidden behind the auto
statement. This has the advantage that we don't need to worry about the template arguments of the view type. The adaptor has another, very useful feature: it can be chained with other adaptors!
What will this print?
In the above example, the vector is "piped" (similar to the unix command line) into the reverse adaptor and then into the drop adaptor and a combined view object is returned. Note that accessing the 0th element of the view is still lazy, determining which element it maps to happens at the time of access.
What does this imply for argument types and return types of the lambda functions?
Task: Create a view on std::vector vec{1, 2, 3, 4, 5, 6};
that filters out all uneven numbers and squares the remaining (even) values, i.e.
Views are a specific kind of range that is formalised in the std::ranges::view concept. Every view returned by a view adaptor models this concept, but which other range concepts are modeled by a view?
It depends on the underlying range and also the view itself. With few exceptions, views don't model more/stronger range concepts than their underlying range (other than std::ranges::view) and they try to preserve as much of the underlying range's concepts as possible. For instance the view returned by std::views::reverse
models std::ranges::random_access_range (and weaker concepts) if the underlying range also models the respective concept. It never models std::ranges::contiguous_range, because the third element of the view is not located immediately after the second in memory (but instead before the second).
Perhaps surprising to some, many views also model std::ranges::output_range if the underlying range does, i.e. views are not read-only:
v
models?
Concept | yes/no? |
---|---|
std::ranges::input_range | |
std::ranges::forward_range | |
std::ranges::bidirectional_range | |
std::ranges::random_access_range | |
std::ranges::contiguous_range | |
std::ranges::view | |
std::ranges::sized_range | |
std::ranges::output_range |
Concept | yes/no? |
---|---|
std::ranges::input_range | ✅ |
std::ranges::forward_range | ✅ |
std::ranges::bidirectional_range | ✅ |
std::ranges::random_access_range | |
std::ranges::contiguous_range | |
std::ranges::view | ✅ |
std::ranges::sized_range | |
std::ranges::output_range |
Surprised? Let's have a closer look at the std::views::filter view. The filter view only returns the value of the underlying range for which the given predicate evaluates to true
. To know which value is an element of the filter view, the view has to look at each of them. Thus, it must scan the underlying range value-by-value and cannot jump to an arbitrary location in constant time since it cannot know how many elements it had to skip without looking at them. Accordingly, the std::views::filter preserves only std::ranges::bidirectional_range, because it can scan the text in reverse order as well. Since the view cannot guarantee that the values lie in contiguous memory, it can also not preserve std::ranges::contiguous_range. Similarly, the view cannot model std::ranges::sized_range as it cannot determine the number of values not filtered out in constant time.
The transform on the other hand produces a new element on every access (the result of the multiplication), therefore v
is not a std::ranges::output_range, you cannot assign values to its elements. Note that this prevents modelling the std::ranges::contiguous_range as well because values are created on-demand and are not stored in memory at all.
We provide overview tables for all our view adaptors that document which concepts are modelled by the views they return.
The standard library in C++20 provides a number of useful views and SeqAn provides many views, as well. Most views provided by SeqAn3 are specific to biological operations, like seqan3::views::trim_quality which trims sequences based on the quality or seqan3::views::complement which generates the complement of a nucleotide sequence. But SeqAn3 also provides some general purpose views.
Have a look at the views-submodule to get an overview of SeqAn's views and also read through the detailed description on that page now that you had a more gentle introduction.
Use views to implement steps 2.-4.
Containers are ranges that own their data. SeqAn3 uses the standard library containers, like std::vector and std::list to store elements. For certain use-cases we have introduced our own containers, though.
All standard library containers model std::ranges::forward_range (see above), but we have introduced container concepts that encompass more of a containers interface. Have a look at the API documentation of seqan3::container and unfold the inheritance diagram. What can you learn about the different refinements and their relation to the range concepts?
If you followed the alphabet tutorial closely, you will know that seqan3::dna4 needs only two bits to represent its state. However, single objects are always at least a byte (eight bits) big in C++. To store sequences of small alphabets more space-efficiently, we have developed seqan3::bitpacked_sequence.
Open the API documentation of seqan3::bitpacked_sequence, display the inheritance diagram and read through the interface overview and the detailed description.
Measure and compare the amount of main memory that your program uses depending on the vector implementation. On Linux based systems use /usr/bin/time -v <program> <args>
and look for "Maximum resident set size". (Not to be confused with the built-in Bash time command! So use the full path /usr/bin/time
)
On macOS and BSD use /usr/bin/time -l <program> <args>
and look for "maximum resident set size".