osrm-backend/third_party/vtzero/doc/reading.md

# Reading vector tiles

To access the contents of vector tiles with vtzero create a `vector_tile`
object first with the data of the tile as first argument:

```cpp
#include <vtzero/vector_tile.hpp> // always needed when reading vector tiles

std::string vt_data = ...;
vtzero::vector_tile tile{vt_data};
```

Instead of a string, you can also initialize the `vector_tile` using a
`vtzero::data_view`. This class contains only a pointer and size referencing
some data similar to the C++17 `std::string_view` class. It is typedef'd from
the `protozero::data_view`. See [the protozero
doc](https://github.com/mapbox/protozero/blob/master/doc/advanced.md#protozero_use_view)
for more details.

```cpp
vtzero::data_view vt_data = ...;
vtzero::vector_tile tile{vt_data};
```

In both cases the `vector_tile` object contains references to the original
tile data. You have to make sure this data stays available through the whole
lifetime of the `vector_tile` object and all the other objects we'll create
in this tutorial for accessing parts of the vector tile. The data is **not**
copied by vtzero when accessing vector tiles.

You can think of the `vector_tile` class as a "proxy" class giving you access
to the decoded data, similarly the classes `layer`, `feature`, and
`property` described in the next chapters are "proxy" classes, too.

## Accessing layers

Vector tiles consist of a list of layers. The list of layers can be empty
in which case `tile.empty()` will return true.

The simplest and fasted way to access the layers is through the `next_layer()`
function:

```cpp
while (auto layer = tile.next_layer()) {
    ...
}
```

Note that this creates new layer objects on the fly referencing the layer you
are currently looking at. Once you have iterated over all the layers,
`next_layer()` will return the "invalid" (default constructed) layer object
which converts to false in an boolean context.

You can reset the layer iterator to the beginning again if you need to go
over the layers again:

```cpp
tile.reset_layer();
```

Instead of using this external iterator, you can use a different function with
an internal iterator that calls a function defined by you for each layer. Your
function must take a `layer&&` as parameter and return `true` if the iteration
should continue and `false` otherwise:

```cpp
tile.for_each_layer([&](layer&& l) {
    // do something with layer
    return true;
});
```

Both the external and internal iteration do basically the same and have the
same performance characteristics.

You can also access layers through their index or name:

```cpp
tile.get_layer(3);
```

will give you the 4th layer in the tile. With

```cpp
tile.get_layer_by_name("foobar");
```

you'll get the layer with the specified name. Both will return the invalid
layer if that layer doesn't exist.

Note that accessing layers by index or name is less efficient than iterating
over them using `next_layer()` if you are accessing several layers. So usually
you should only use those function if you want to access one specific layer
only.

If you need the number of layers, you can call `tile.count_layers()`. This
function still has to iterate over the layers internally decoding some of the
data, so it is not cheap.

## The layer

Once you have a layer as described in the previous chapter you can access the
metadata of this layer easily:

* The version is available with `layer.version()`. Only version 1 and 2 are
  currently supported by this library.
* The extent of the tile is available through `layer.extent()`. This is usually
  4096.
* The function `layer.name()` returns the name of the layer as `data_view`.
  This does **not** include a final 0-byte!
* The number of features is returned by the `layer.num_features()` function.
  If it doesn't contain any features `layer.empty()` will return true.
  (Different then the `vector_tile::count_layers()`, the `layer::num_features()`
  function is `O(1)`).

To access the features call the `next_feature()` function until it returns
the invalid (default constructed) feature:

```cpp
while (auto feature = layer.next_feature()) {
    ...
}
```

Use `reset_feature()` to restart the feature iterator from the beginning.

Instead of using this external iterator, you can use a different function with
an internal iterator that calls a function defined by you for each feature.
Your function must take a `feature&&` as parameter and return `true` if the
iteration should continue and `false` otherwise:

```cpp
layer.for_each_feature([&](feature&& f) {
    // do something with the feature
    return true;
});
```

Both the external and internal iteration do basically the same and have the
same performance characteristics.

If you know the ID of a feature you can get the feature using
`get_feature_by_id()`, but note that this will do a linear search through
all the features in the layer, decoding each one until it finds the right ID.
This is almost always **not** what you want.

Note that the feature returned by `next_feature()` or `get_feature_by_id()`
will internally contain a pointer to the layer it came from. The layer has to
stay valid as long as the feature is used.

## The feature

You get features from the layer as described in the previous chapter. The
`feature` class gives you access to the ID, the geometry and the properties
of the feature. Access the ID using the `id()` method which will return 0
if no ID is set. You can ask for the existence of the ID using `has_id()`:

```cpp
auto feature = layer...;
if (feature.has_id()) {
    cout << feature.id() << '\n';
}
```

The `geometry()` method returns an object of the `geometry` class. It contains
the geometry type and a reference to the (un-decoded) geometry data. See a
later chapter on the details of decoding this geometry. You can also directly
add this geometry to a new feature you are writing.

The number of properties in the feature is returned by the
`feature::num_properties()` function. If the feature doesn't contain any
properties `feature.empty()` will return true. (Different then the
`vector_tile::count_layers()`, the `feature::num_properties()` function is
`O(1)`).

To access the properties call the `next_property()` function until it returns
the invalid (default constructed) property:

```cpp
while (auto property = feature.next_property()) {
    ...
}
```

Use `reset_property()` to restart the property iterator from the beginning.

Instead of using this external iterator, you can use a different function with
an internal iterator that calls a function defined by you for each property.
Your function must take a `property&&` as parameter and return `true` if the
iteration should continue and `false` otherwise:

```cpp
feature.for_each_property([&](property&& p) {
    ...
    return true;
});
```

Both the external and internal iteration do basically the same and have the
same performance characteristics.

## The property

Each property you get from the feature is an object of the `property` class. It
contains a view of the property key and value. The key is always a string
encoded in a `vtzero::data_view`, the value can be of different types but is
always encapsulated in a `property_value` type, a variant type that can be
converted into whatever type the value really has.

```cpp
auto property = ...;
auto pkey = std::string(property.key()); // returns a vtzero::data_view which can
                                         // be converted to std::string
property_value pvalue = property.value();
```

To get the type of the property value, call `type()`:

```cpp
const auto type = pvalue.type();
```

If the property value is an int, for instance, you can get it like this:

```cpp
if (pvalue.type() == property_value_type::int_value)
    int64_t v = pvalue.int_value();
}
```

Instead of accessing the values this way, you'll often use the visitor
interface. Here is an example where the `print_visitor` struct is used to print
out the values. In this case one overload is used for all primitive types
(`double`, `float`, `int`, `uint`, `bool`), one overload is used for the `string_value`
type which is encoded in a `data_view`. You must make sure your visitor handles
all those types.

```cpp
struct print_visitor {

    template <typename T>
    void operator()(T value) {
        std::cout << value;
    }

    void operator()(vtzero::data_view value) {
        std::cout << std::string(value);
    }

};

vtzero::apply_visitor(print_visitor{}, pvalue));
```

All call operators of your visitor class have to return the same type. In the
case above this was `void`, but it can be something else. That return type
will be the return type of the `apply_visitor` function. This can be used,
for instance, to convert the values into one type:

```cpp
struct to_string_visitor {

    template <typename T>
    std::string operator()(T value) {
        reutrn std::to_string(value);
    }

    std::string operator()(vtzero::data_view value) {
        return std::string(value);
    }

};

std::string v = vtzero::apply_visitor(to_string_visitor{}, pvalue);
```

Sometimes you want to convert the `property_value` type into your own variant
type. You can use the `vtzero::convert_property_value()` free function for
this.

Lets say you are using `boost` and this is your variant:

```cpp
using variant_type = boost::variant<std::string, float, double, int64_t, uint64_t, bool>;
```

You can then use the following line to convert the data:
```cpp
variant_type v = vtzero::convert_property_value<variant_type>(pvalue);
```

Your variant type must be constructible from all the types `std::string`,
`float`, `double`, `int64_t`, `uint64_t`, and `bool`. If it is not, you can
define a mapping between those types and the types you use in your variant
class.

```cpp
using variant_type = boost::variant<mystring, double, int64_t, uint64_t, bool>;

struct mapping : vtzero::property_value_mapping {
    using string_type = mystring; // use your own string type which must be
                                  // convertible from data_view
    using float_type = double; // no float in variant, so convert to double
};

variant_type v = vtzero::convert_property_value<variant_type, mapping>(pvalue);
```

## Creating a properties map

This linear access to the properties with lazy decoding of each property only
when it is accessed saves memory allocations, especially if you are only
interested in very few properties. But sometimes it is easier to create a
mapping (based on `std::unordered_map` for instance) between keys and values. This is where
the `vtzero::create_properties_map()` templated free function comes in. It
needs the map type as template parameter:

```cpp
using key_type = std::string; // must be something that can be converted from data_view
using value_type = boost::variant<std::string, float, double, int64_t, uint64_t, bool>;
using map_type = std::map<key_type, value_type>;

auto feature = ...;
auto map = create_properties_map<map_type>(feature);
```

Both `std::map` and `std::unordered_map` are supported as map type, but this
should also work with any other map type that has an `emplace()` method.

## Geometries

Features must contain a geometry of type UNKNOWN, POINT, LINESTRING, or
POLYGON. The UNKNOWN type is not further specified by the vector tile spec,
this library doesn't allow you to do anything with this type. Note that
multipoint, multilinestring, and multipolygon geometries are also possible,
they don't have special types.

You can get the geometry type with `feature.geometry_type()`, but usually
you'll get the geometry with `feature.geometry()`. This will return an object
of type `vtzero::geometry` which contains the geometry type and a view of
the raw geometry data. To decode the data you have to call one of the decoder
free functions `decode_geometry()`, `decode_point_geometry()`,
`decode_linestring_geometry()`, or `decode_polygon_geometry()`. The first of
these functions can decode any point, linestring, or polygon geometry. The
others must be called with a geometry of the specified type and will only
decode that type.

For all the decoder functions the first parameter is the geometry (as returned
by `feature.geometry()`), the second parameter is a *handler* object that you
must implement. The decoder function will call certain callbacks on this object
that give you part of the geometry data which allows you to use this data in
any way you like.

The handler for `decode_point_geometry()` must implement the following
functions:

* `void points_begin(uint32_t count)`: This is called once at the beginning
  with the number of points. For a point geometry, this will be 1, for
  multipoint geometries this will be larger.
* `void points_point(vtzero::point point)`: This is called once for each
  point.
* `void points_end()`: This is called once at the end.

The handler for `decode_linestring_geometry()` must implement the following
functions:

* `void linestring_begin(uint32_t count)`: This is called at the beginning
  of each linestring with the number of points in this linestring. For a simple
  linestring this function will only be called once, for a multilinestring
  it will be called several times.
* `void linestring_point(vtzero::point point)`: This is called once for each
  point.
* `void linestring_end()`: This is called at the end of each linestring.

The handler for `decode_polygon_geometry` must implement the following
functions:

* `void ring_begin(uint32_t count)`: This is called at the beginning
  of each ring with the number of points in this ring. For a simple polygon
  with only one outer ring, this function will only be called once, if there
  are inner rings or if this is a multipolygon, it will be called several
  times.
* `void ring_point(vtzero::point point)`: This is called once for each
  point.
* `void ring_end(vtzero::ring_type)`: This is called at the end of each ring.
  The parameter tells you whether the ring is an outer or inner ring or whether
  the ring was invalid (if the area is 0).

The handler for `decode_geometry()` must implement all of the functions
mentioned above for the different types. It is guaranteed that only one
set of functions will be called depending on the geometry type.

If your handler implements the `result()` method, the decode functions will
have the return type of the `result()` method and will return whatever
result returns. If the `result()` method is not available, the decode functions
return void.

Here is a typical implementation of a linestring handler:

```cpp
struct linestring_handler {

    using linestring = std::vector<my_point_type>;

    linestring points;

    void linestring_begin(uint32_t count) {
        points.reserve(count);
    }

    void linestring_point(vtzero::point point) noexcept {
        points.push_back(convert_to_my_point(point));
    }

    void linestring_end() const noexcept {
    }

    linestring result() {
        return std::move(points);
    }

};
```

Note that the `count` given to the `linestring_begin()` method is used here to
reserve memory. This is potentially problematic if the count is large. Please
keep this in mind.

## Accessing the key/value lookup tables in a layer

Vector tile layers contain two tables with all the property keys and all
property values used in the features in that layer. Vtzero usually handles
those table lookups internally without you noticing. But sometimes it might
be necessary to access this data directly.

From the layer object you can get references to the tables:

```cpp
vtzero::layer layer = ...;
const auto& kt = layer.key_table();
const auto& vt = layer.value_table();
```

Instead you can also lookup keys and values using methods on the layer object:
```cpp
vtzero::layer layer = ...;
const vtzero::data_view k = layer.key(17);
const vtzero::property_value_view v = layer.value(42);
```

As usual in vtzero you only get views back, so you need to keep the layer
object around as long as you are accessing the results of those methods.

Note that the lookup tables are created on first access from the layer data. As
long as you are not accessing those tables directly or by looking up any
properties in a feature, the tables are not created and no extra memory is
used.