EDM4hep - The common event data model

EDM4hep is the common and shared Event Data Model (EDM) of the Key4hep project. Here we will give a brief introduction to EDM4hep as well as some of the technicalities behind it. We will also guide you towards documentation and try to give you the knowledge to make sense of it.

Important resources

EDM4hep doxygen API reference page: edm4hep.web.cern.ch
EDM4hep github repository: github.com/key4hep/EDM4hep
podio documentation page (including API reference): key4hep.web.cern.ch/podio
podio github repository: github.com/AIDASoft/podio

Reading EDM4hep files

EDM4hep is implemented using the podio toolkit (see below for a brief introduction). Hence, we will also be using some podio functionality to read files.

For more technical information and pointers to how to read and understand the documentation see below.

Declaring a (generic) reader

podio offers generic readers that will try to dynamically dispatch to the correct low level reader by checking the file contents to figure out which actual reader to instantiate. The available low level readers are

podio::ROOTReader (default) for reading ROOT files that are stored using TTrees
podio::RNTupleReader for reading ROOT files that are stored using the new RNTuple format
podio::SIOReader for reading SIO files written via podio

Note

The support for the RNTuple and SIO format needs to be configured for podio during its build. In Key4hep both are generally available

The following snippets show the necessary includes / imports and how to instantiate the generic reader.

C++

#include <podio/Reader.h>      // makeReader
#include <podio/ROOTReader.h>  // for ROOTReader (similar for the others)

// The generic reader is constructed via makeReader and a (list of) filename(s)
auto reader = podio::makeReader("<input-file-name>");

// Low level readers are instantiated first and then call openFile(s)
auto rootReader = podio::ROOTReader();
rootReader.openFile("<input-file-name>"); // openFiles also exists!

Make sure to use podio::podioIO in your target_link_libraries if you build an executable via CMake.

Python

from podio.reading import get_reader  # for the generic reader
from podio.root_io import Reader  # for TTree (i.e. ROOTReader)

# The generic reader is obtained from get reader and a (list of) filename(s)
reader = get_reader("<input-file-name>")

# The low level Reader is explicitly instantiated with a (list of) filename(s)
rootReader = Reader("<your-file-name>");

Looping over events in a file

Once you have a (generic) reader you can start looping over the events and obtaining collections. In the following we will be using MCParticles, but conceptual all EDM4hep types work the same.

C++

#include <edm4hep/MCParticleCollection.h>

#include <podio/Frame.h>

// generic reader
for (size_t i = 0; i < reader.getEvents(); ++i) {
  auto event = reader.readEvent(i);
  const auto& mcParticles = event.get<edm4hep::MCParticleCollection>("MCParticles");
  // ...
}

// low level reader
for (size_t i = 0; i < rootReader.getEntries(podio::Category::Event); ++i) {
  auto event = podio::Frame(rootReader.readEntry(podio::Category::Event, i));
  const auto& mcParticles = event.get<edm4hep::MCParticleCollection>("MCParticles");
  // ...
}

Note

For the low level reader we have to explicitly create a podio::Frame and choose the events category, which is done automatically for us for the generic reader via readEvent. The generic reader can still read other categories via readEntry.

Python

# generic reader
for event in reader.get("events"):
    mcParticles = event.get("MCParticles")
    
# low level reader
for event in rootReader.get("events"):
    mcParticles = event.get("MCParticles")

Note

For python both are the exact same as get_reader simply returns an instance of one of the low level readers.

Doxygen API documentation

We start with having a look at the EDM4hep doxygen API reference page:

The overview diagram

You see a diagrammatic overview of EDM4hep with all the available data types, broadly organized into different categories. The arrows depict two ways data types can be related / linked with each other

“Relations” (black arrows)
“Links” (purple-ish arrows)

Relations

These are relations defined within the data types, and which are directly accessible from the data types. They come in two flavors, depending on the multiplicity of the relation

OneToOneRelations
OneToManyRelations

Data types can relate to other instances of the same type (e.g. MCParticles usually form a hierarchy of mothers/daughters). Relations are directed, i.e. it is possible to go from one object to a related object, but vice versa this does usually not hold. For example, a ReconstructedParticle can point to multiple Tracks or Clusters, but those do not point to a ReconstructedParticle.

Links

These are relations that are in a sense “external” to the data model definition. They are currently mainly used to connect MC and RECO information, as a direct link via a relation is not desirable as it would mix the two worlds. In contrast to relations, links are not directed, i.e. it is possible to access both involved objects from the link.

The table of available types

Just below the diagram is an overview table of all the types that are defined in EDM4hep. Here they are organized into

Components - Very simple types, that are used throughout the Datatypes
Datatypes - The data types that are defined in EDM4hep
Links - The available links between different data types
Generator related (meta-)data - Data types related to generator metadata
Interfaces - Abstractions for accessing different types by a shared set of properties

Clicking on any of these links will take you to the edm4hep.yaml definition file of EDM4hep, jumping directly to the definition of the respective datatype or component. For more information on this file check out the section about podio. In principle it is possible to have very educated guesses on how the interface of the classes will look like from this.

Navigating the doxygen reference page

To see all the available classes simply click on Classes -> Class Index or on Classes -> Class List. Doing the latter and expanding the edm4hep namespace gives you something like this

Clicking on any of the links in this list will take you to the reference page for that class, e.g. for the ReconstructedParticle

Why are there so many classes and do I need all of them?

If you look at the list you will realize there are many classes that are all named very similar, e.g.

CaloHitContribution
CaloHitContributionCollection
CaloHitContributionCollectionData
CaloHitContributionCollectionIterator
CaloHitContributionData
CaloHitContributionMutableCollectionIterator
CaloHitContributionObj
CaloHitContributionSIOBlock
MutableCaloHitContribution

From all of these classes only the ones marked bold are truly “visible” and intended for use. The others are internal classes or simple helper types that you will most likely only ever see in compiler errors, especially if you follow the Almost Always Auto style of writing c++ code (see Herb Sutter’s original blog post or a slightly easier to digest summary). To understand why these classes exist and what their purpose is, we have to make a slight detour to podio.

Some utility functionality

EDM4hep also brings a bit of utility functionality. You can find it in the edm4hep::utils namespace (click on Namespaces -> Namespace List, then expand the edm4hep namespace and then click on utils to arrive at this link).

podio - The technical infrastructure on which things run

podio is an EDM toolkit that is used by and developed further in the Key4hep context. The main purpose is to have an efficiently implemented, thread safe EDM starting from a high level description. For more (gory) details have a look at the github repository.

Here we will describe the code generation, and its implications for EDM4hep. A bit further down we will describe how to read and write podio (root) files and the podio-dump tool to inspect files without having to open them.

podio code generation

The podio code generator is a python script that reads in the EDM definition in yaml format, does a few basic validation checks on the definition, and then generates all the necessary code via the Jinja2 template engine.

The generated code should (among other things)

be efficient,
offer an easy to use interface,
offer performant I/O.

Having automated code generation has a few advantages:

Freeing the user from the repetitive task of implementing all the types themselves
Freeing the user from having to deal with all the details of how to do things efficiently
Making it very easy to roll out improved implementations (or bug fixes) via simply regenerating the code

The three layers of podio

To achieve the goals stated above podio favors composition over inheritance and uses plain-old-data (POD) types wherever possible. To achieve this podio employs a layered design, which makes it possible to have an efficient memory layout and performant I/O implementation, while still offering an easy to use interface

The User Layer is the top most layer and it offers the full functionality and is the only layer with which users interact directly. It consists mainly of the collections and lightweight handle classes, i.e.
- XYZCollection
- XYZ
- MutableXYZ
The Object Layer consists of the XYZObj classes, that take care of all resource management and which also enable the relations between different objects.
The POD Layer at the very bottom is where all the actual data lives in simple XYZData POD structs. These are the things that are actually stored in, e.g. root files that are written by podio.

Basics of generated code - value semantics

The generated c++ code offers so called value semantics. The exact details of what this actually means are not very important, the main point is that you can treat all objects as values and you don’t have to worry about inefficient copies or managing resources:

C++

auto recos = edm4hep::ReconstructedParticleCollection();

// ... fill, e.g. via
auto p = recos.create();
// or via
auto p2 = edm4hep::MutableReconstructedParticle();
recos.push_back(p2); 

// Loop over a collection
for (auto reco : recos) {
  auto vtx = reco.getDecayVertex();
  // do something with the vertex
  
  // loop over related tracks
  for (auto track : reco.getTracks()) {
    // do something with this track
  }
}

Python

recos = edm4hep.ReconstructedParticleCollection()

# ... fill, e.g. via
p = recos.create()
# or via
p2 = edm4hep.MutableReconstructedParticle()
recos.push_back(p2)

# Loop over a collection
for reco in recos:
    vtx = reco.getDecayVertex()
    # do something with the vertex
  
    # loop over related tracks
    for track in reco.getTracks():
        # do something with the tracks

The python interface is functionally equivalent to the one c++ interface, since that is implemented via PyROOT. There are some additions that make the python interface more pythonic, e.g. len(recos) is equivalent to recos.size(). Nevertheless, the doxygen reference is valid for both interfaces.

Guessing the interface from the yaml definition

Since all code is generated, it is usually pretty straight forward to guess how the interface will look like just from looking at the definition in the yaml file. For EDM4hep the general rule is to get a Member variable, a OneToOneRelation, a OneToManyRelation or a VectorMember is to simply stick a get in front of the name in the yaml file and to capitalize the first letter., e.g.

Members:
  - edm4hep::Vector3f momentum // the momentum in [GeV]

will turn into something like

const edm4hep::Vector3f& getMomentum() const;

Similar, but in slightly more nuanced rules apply for the methods that are generated for setting a value. For Member variables and OneToOneRelations the general rule is to stick a set in front of the name in the yaml file and to capitalize the first letter, e.g. (continuing from above)

void setMomentum(edm4hep::Vector3f value);

For the OneToManyRelations and VectorMembers the rule is to stick a addTo in front of the name in the yaml file and to capitalize the first letter, e.g.

OneToManyRelation:
  - MCParticle daughters // the daughters of this particle

will be generated to

void addToDaughters(MCParticle daughter);

Why is there a `XYZ` and a `MutableXYX`?

The underlying technical reasons are rather complex, dive quite deepish into c++ nuances, and definitely far beyond the scope of this tutorial. In short: We need two different handle classes in order to control whether users are allowed to modify things or not. As one of the main goals of podio generated EDMs is to be thread safe the default generated class for each data type allows only for immutable read access, i.e. it provides only the get methods. Only the Mutable classes actually have the set methods, and can hence be used to actually modify objects. The most important implication of this is the following: Everything that you read from file, or that you get from the Gaudi TES is immutable. I.e. there is no way for you to change or update the values that you read. The only way to “update” values (or collections) is to actually copy the contents and then store the updated values back. Independent copies of objects can be obtained with the clone` method.

Writing function interfaces

The Mutable objects implicitly convert to an instance of a default class. Hence, always use the default classes when specifying function interfaces (obviously this only works if you only need read access in the function). There is no implicit conversion from the default, immutable objects to the Mutable objects!

As an example

void printE(edm4hep::MCParticle particle) {
  std::cout << particle.getEnergy() << '\n';
}

void printEMutable(edm4hep::MutableMCParticle particle) {
  std::cout << particle.getEnergy() << '\n';
}

int main() {
  auto mutP = edm4hep::MutableMCParticle();
  p.setEnergy(3.14);
  
  printE(mutP);  // Works due to implicit conversion
  printEMutable(mutP);  // Obviously also works
  
  // Now we create an immutable object
  auto P = edm4hep::MCParticle();
  
  printE(P);  // Obviously works
  printEMutable(P);  // BREAKS: No conversion from default to Mutable

  return 0;
}

Subset collections

Similar to LCIO, podio generated EDMs offer a subset collection functionality. This allows to create collections of objects, that are actually part of another collection, e.g. to simply collect all the muons that are present in a larger collection of reconstructed particles:

To create a subset collection, simply do

auto muons = edm4hep::ReconstructedParticleCollection();
muons.setSubsetCollection();

// You can now add objects that are part 
// of another collection to this one via push_back
muons.push_back(recos[0]);

Reading a subset collection works exactly the same as reading a normal collection. This is handled in a transparent way, such that you usually don’t even realize that you are operating on a subset collection.

The `podio::Frame` container

The podio::Frame is a generalized event. It is a container that aggregates all relevant data (and some meta data). It also defines an implicit interval of validity (but that is less relevant for this tutorial). It provides a thread safe interface for data access

Immutable read access only for collections that are stored inside the a Frame
All data that is inside a Frame is owned by it, and this is also reflected in its interface.

Here we will just briefly introduce the main functionality, for more details see the documentation in podio.

Getting collections from a `Frame`

Assuming that event is a podio::Frame in the following code examples, getting a collection can be done via (c++)

C++

auto& mcParticles = event.get<edm4hep::MCParticleCollection>("MCParticles"); 

Python

mcParticles = event.get("MCParticles")

This retrieves the collection that is stored under the name MCParticles with type edm4hep::MCParticleCollection. If no such collection exists, it will simply return an empty collection of the desired type. As you can see, the type is automatically inferred in python. Note that get returns a const&, so it is required to actually put the & behind auto in c++, otherwise there will be a compilation error complaining about a copy-constructor being marked delete.

Putting a collection into a `Frame`

When putting a collection into a Frame you give up ownership of this collection. To signal this to the users, it is necessary to move the collection into a Frame. Again assuming event is a podio::Frame in the following examples, this looks like this

auto recos = edm4hep::ReconstructedParticleCollection();
event.put(std::move(recos), "ReconstructedParticles");

Note the requirement to explicitly use std::move in this case. At this point recos is moved into the event, and you are left with an object in a valid but unspecified state that you should under normal circumstances no longer use after this point. (Technically we do enough that you still can use this, but don’t expect the results to match your expectations).

Reading EDM4hep files

EDM4hep files are read with tools provided by podio. As podio supports multiple different backends there are several, low level readers that support all the necessary functionality. You can obviously use these readers directly, but we recommend to use the Reader class and the makeReader function that will dispatch to the correct low level reader automatically.

#include <podio/Reader.h>

#include <edm4hep/MCParticleCollection.h>

int main() {
  auto reader = podio::makeReader("some_file_containing_edm4hep.data.root");

  // Loop over all events
  for (size_t i = 0; i < reader.getEvents(); ++i) {
    auto event = reader.readNextEvent();
    auto& mcParticles = event.get<edm4hep::MCParticleCollection>("MCParticles");

    // do more stuff with this event
  }

  return 0;
}

The available low level readers

ROOTReader - The default reader for TTree based files
RNTupleReader - A reader for RNTuple based files
SIOReader - The reader for reading files using the SIO backend

As podio is a rather low level tool, also the interface of these readers feel somewhat low level. This is mostly visible in the fact, that you have to provide a category (name) when getting the number of entries, or when reading the next entry. This is because in principle podio can handle multiple different categories of Frames in one file. For the purpose of this tutorial and also for the majority of use cases, simply use "events" as category name. Readers in podio do not return a podio::Frame directly, rather they just return some frame data from which a podio::Frame can be constructed. Putting all of these things together, a simple event loop looks like this in c++:

C++

#include "podio/ROOTReader.h"
#include "podio/Frame.h"

#include "edm4hep/MCParticleCollection.h"

int main() {
  auto reader = podio::ROOTReader();
  reader.openFile("some_file_containing_edm4hep_data.root");
  
  // Loop over all events
  for (size_t i = 0; i < reader.getEntries("events"); ++i) {
    auto event = podio::Frame(reader.readNextEntry("events"));
    auto& mcParticles = event.get<edm4hep::MCParticleCollection>("MCParticles");
    
    // do more stuff with this event
  }

  return 0;
}

Python

from podio import root_io

reader = root_io.Reader("some_file_containing_edm4hep_data.root")

for event in reader.get("events"):
    mcParticles = event.get("MCParticles")
    # do more stuff with this event

ROOT file layout of podio generated EDMs

podio generated EDMs, i.e. also EDM4hep, use ROOT as their default I/O backend. Since everything is based on PODs, the produced root files are pretty straight forward to read and interpret (with some caveats). They are already almost flat ntuples.

`podio-dump`

The podio-dump utility allows to inspect EDM4hep files from the command line. The synopsis looks like this

$podio-dump --help
usage: podio-dump [-h] [-c CATEGORY] [-e ENTRIES] [-d] [--dump-edm DUMP_EDM] [--version] inputfile

Dump contents of a podio file to stdout

positional arguments:
  inputfile             Name of the file to dump content from

options:
  -h, --help            show this help message and exit
  -c CATEGORY, --category CATEGORY
                        Which Frame category to dump
  -e ENTRIES, --entries ENTRIES
                        Which entries to print. A single number, comma separated list of numbers or "first:last" for an inclusive range of entries. Defaults to the first entry.
  -d, --detailed        Dump the full contents not just the collection info
  --dump-edm DUMP_EDM   Dump the specified EDM definition from the file in yaml format
  --version             show program's version number and exit

By default it prints how many events are present in the file and also a summary of the contents of the first event. This overview consists of the names, data types and number of elements of the collections that are stored in this event. Using the --detailed flag, podio-dump will print the complete contents of all collections in ASCII format. This can be quite a bit of information. Using the --entries flag it is possible to choose which events to look at. The --categories flag is an advanced feature and not necessary for this tutorial.

podio-dump will also tell you whether the file that is passed to it is a legacy file in which case you will need the ROOTLegacyReader or the SIOLegacyReader to read it.

EDM4hep - The common event data model

Important resources

Reading EDM4hep files

Declaring a (generic) reader

Looping over events in a file

Doxygen API documentation

The overview diagram

Relations

Links

The table of available types

Navigating the doxygen reference page

Why are there so many classes and do I need all of them?

Some utility functionality

podio - The technical infrastructure on which things run

podio code generation

The three layers of podio

Basics of generated code - value semantics

Guessing the interface from the yaml definition

Why is there a XYZ and a MutableXYX?

Writing function interfaces

Subset collections

The podio::Frame container

Getting collections from a Frame

Putting a collection into a Frame

Reading EDM4hep files

The available low level readers

ROOT file layout of podio generated EDMs

podio-dump

Why is there a `XYZ` and a `MutableXYX`?

The `podio::Frame` container

Getting collections from a `Frame`

Putting a collection into a `Frame`

`podio-dump`