Logo ROOT  
Reference Guide
 
Loading...
Searching...
No Matches
RDataSource.hxx
Go to the documentation of this file.
1// Author: Enrico Guiraud, Danilo Piparo CERN 09/2017
2
3/*************************************************************************
4 * Copyright (C) 1995-2018, Rene Brun and Fons Rademakers. *
5 * All rights reserved. *
6 * *
7 * For the licensing terms see $ROOTSYS/LICENSE. *
8 * For the list of contributors see $ROOTSYS/README/CREDITS. *
9 *************************************************************************/
10
11#ifndef ROOT_RDATASOURCE
12#define ROOT_RDATASOURCE
13
15#include <string_view>
16#include "RtypesCore.h" // ULong64_t
17#include "TString.h"
18
19#include <algorithm> // std::transform
20#include <string>
21#include <typeinfo>
22#include <vector>
23
24namespace ROOT {
25namespace RDF {
26class RDataSource;
27}
28}
29
30/// Print a RDataSource at the prompt
31namespace cling {
32std::string printValue(ROOT::RDF::RDataSource *ds);
33} // namespace cling
34
35namespace ROOT {
36
37namespace Internal {
38namespace TDS {
39
40/// Mother class of TTypedPointerHolder. The instances
41/// of this class can be put in a container. Upon destruction,
42/// the correct deletion of the pointer is performed in the
43/// derived class.
45protected:
46 void *fPointer{nullptr};
47
48public:
49 TPointerHolder(void *ptr) : fPointer(ptr) {}
50 void *GetPointer() { return fPointer; }
51 void *GetPointerAddr() { return &fPointer; }
53 virtual ~TPointerHolder(){};
54};
55
56/// Class to wrap a pointer and delete the memory associated to it
57/// correctly
58template <typename T>
59class TTypedPointerHolder final : public TPointerHolder {
60public:
61 TTypedPointerHolder(T *ptr) : TPointerHolder((void *)ptr) {}
62
64 {
65 const auto typedPtr = static_cast<T *>(fPointer);
66 return new TTypedPointerHolder(new T(*typedPtr));
67 }
68
69 ~TTypedPointerHolder() { delete static_cast<T *>(fPointer); }
70};
71
72} // ns TDS
73} // ns Internal
74
75namespace RDF {
76
77// clang-format off
78/**
79\class ROOT::RDF::RDataSource
80\ingroup dataframe
81\brief RDataSource defines an API that RDataFrame can use to read arbitrary data formats.
82
83A concrete RDataSource implementation (i.e. a class that inherits from RDataSource and implements all of its pure
84methods) provides an adaptor that RDataFrame can leverage to read any kind of tabular data formats.
85RDataFrame calls into RDataSource to retrieve information about the data, retrieve (thread-local) readers or "cursors"
86for selected columns and to advance the readers to the desired data entry.
87
88The sequence of calls that RDataFrame (or any other client of a RDataSource) performs is the following:
89
90 - SetNSlots() : inform RDataSource of the desired level of parallelism
91 - GetColumnReaders() : retrieve from RDataSource per-thread readers for the desired columns
92 - Initialize() : inform RDataSource that an event-loop is about to start
93 - GetEntryRanges() : retrieve from RDataSource a set of ranges of entries that can be processed concurrently
94 - InitSlot() : inform RDataSource that a certain thread is about to start working on a certain range of entries
95 - SetEntry() : inform RDataSource that a certain thread is about to start working on a certain entry
96 - FinalizeSlot() : inform RDataSource that a certain thread finished working on a certain range of entries
97 - Finalize() : inform RDataSource that an event-loop finished
98
99RDataSource implementations must support running multiple event-loops consecutively (although sequentially) on the same dataset.
100 - \b SetNSlots() is called once per RDataSource object, typically when it is associated to a RDataFrame.
101 - \b GetColumnReaders() can be called several times, potentially with the same arguments, also in-between event-loops, but not during an event-loop.
102 - \b GetEntryRanges() will be called several times, including during an event loop, as additional ranges are needed. It will not be called concurrently.
103 - \b Initialize() and \b Finalize() are called once per event-loop, right before starting and right after finishing.
104 - \b InitSlot(), \b SetEntry(), and \b FinalizeSlot() can be called concurrently from multiple threads, multiple times per event-loop.
105
106 Advanced users that plan to implement a custom RDataSource can check out existing implementations, e.g. RCsvDS or RNTupleDS.
107 See the inheritance diagram below for the full list of existing concrete implementations.
108*/
110 // clang-format on
111protected:
112 using Record_t = std::vector<void *>;
113 friend std::string cling::printValue(::ROOT::RDF::RDataSource *);
114
115 virtual std::string AsString() { return "generic data source"; };
116
117public:
118 virtual ~RDataSource() = default;
119
120 // clang-format off
121 /// \brief Inform RDataSource of the number of processing slots (i.e. worker threads) used by the associated RDataFrame.
122 /// Slots numbers are used to simplify parallel execution: RDataFrame guarantees that different threads will always
123 /// pass different slot values when calling methods concurrently.
124 // clang-format on
125 virtual void SetNSlots(unsigned int nSlots) = 0;
126
127 // clang-format off
128 /// \brief Returns a reference to the collection of the dataset's column names
129 // clang-format on
130 virtual const std::vector<std::string> &GetColumnNames() const = 0;
131
132 /// \brief Checks if the dataset has a certain column
133 /// \param[in] colName The name of the column
134 virtual bool HasColumn(std::string_view colName) const = 0;
135
136 // clang-format off
137 /// \brief Type of a column as a string, e.g. `GetTypeName("x") == "double"`. Required for jitting e.g. `df.Filter("x>0")`.
138 /// \param[in] colName The name of the column
139 // clang-format on
140 virtual std::string GetTypeName(std::string_view colName) const = 0;
141
142 // clang-format off
143 /// Called at most once per column by RDF. Return vector of pointers to pointers to column values - one per slot.
144 /// \tparam T The type of the data stored in the column
145 /// \param[in] columnName The name of the column
146 ///
147 /// These pointers are veritable cursors: it's a responsibility of the RDataSource implementation that they point to
148 /// the "right" memory region.
149 // clang-format on
150 template <typename T>
151 std::vector<T **> GetColumnReaders(std::string_view columnName)
152 {
153 auto typeErasedVec = GetColumnReadersImpl(columnName, typeid(T));
154 std::vector<T **> typedVec(typeErasedVec.size());
155 std::transform(typeErasedVec.begin(), typeErasedVec.end(), typedVec.begin(),
156 [](void *p) { return static_cast<T **>(p); });
157 return typedVec;
158 }
159
160 /// If the other GetColumnReaders overload returns an empty vector, this overload will be called instead.
161 /// \param[in] slot The data processing slot that needs to be considered
162 /// \param[in] name The name of the column for which a column reader needs to be returned
163 /// \param[in] tid A type_info
164 /// At least one of the two must return a non-empty/non-null value.
165 virtual std::unique_ptr<ROOT::Detail::RDF::RColumnReaderBase>
166 GetColumnReaders(unsigned int /*slot*/, std::string_view /*name*/, const std::type_info &)
167 {
168 return {};
169 }
170
171 // clang-format off
172 /// \brief Return ranges of entries to distribute to tasks.
173 /// They are required to be contiguous intervals with no entries skipped. Supposing a dataset with nEntries, the
174 /// intervals must start at 0 and end at nEntries, e.g. [0-5],[5-10] for 10 entries.
175 /// This function will be invoked repeatedly by RDataFrame as it needs additional entries to process.
176 /// The same entry range should not be returned more than once.
177 /// Returning an empty collection of ranges signals to RDataFrame that the processing can stop.
178 // clang-format on
179 virtual std::vector<std::pair<ULong64_t, ULong64_t>> GetEntryRanges() = 0;
180
181 // clang-format off
182 /// \brief Advance the "cursors" returned by GetColumnReaders to the selected entry for a particular slot.
183 /// \param[in] slot The data processing slot that needs to be considered
184 /// \param[in] entry The entry which needs to be pointed to by the reader pointers
185 /// Slots are adopted to accommodate parallel data processing.
186 /// Different workers will loop over different ranges and
187 /// will be labelled by different "slot" values.
188 /// Returns *true* if the entry has to be processed, *false* otherwise.
189 // clang-format on
190 virtual bool SetEntry(unsigned int slot, ULong64_t entry) = 0;
191
192 // clang-format off
193 /// \brief Convenience method called before starting an event-loop.
194 /// This method might be called multiple times over the lifetime of a RDataSource, since
195 /// users can run multiple event-loops with the same RDataFrame.
196 /// Ideally, `Initialize` should set the state of the RDataSource so that multiple identical event-loops
197 /// will produce identical results.
198 // clang-format on
199 virtual void Initialize() {}
200
201 // clang-format off
202 /// \brief Convenience method called at the start of the data processing associated to a slot.
203 /// \param[in] slot The data processing slot wihch needs to be initialized
204 /// \param[in] firstEntry The first entry of the range that the task will process.
205 /// This method might be called multiple times per thread per event-loop.
206 // clang-format on
207 virtual void InitSlot(unsigned int /*slot*/, ULong64_t /*firstEntry*/) {}
208
209 // clang-format off
210 /// \brief Convenience method called at the end of the data processing associated to a slot.
211 /// \param[in] slot The data processing slot wihch needs to be finalized
212 /// This method might be called multiple times per thread per event-loop.
213 // clang-format on
214 virtual void FinalizeSlot(unsigned int /*slot*/) {}
215
216 // clang-format off
217 /// \brief Convenience method called after concluding an event-loop.
218 /// See Initialize for more details.
219 // clang-format on
220 virtual void Finalize() {}
221
222 /// \brief Return a string representation of the datasource type.
223 /// The returned string will be used by ROOT::RDF::SaveGraph() to represent
224 /// the datasource in the visualization of the computation graph.
225 /// Concrete datasources can override the default implementation.
226 virtual std::string GetLabel() { return "Custom Datasource"; }
227
228protected:
229 /// type-erased vector of pointers to pointers to column values - one per slot
230 virtual Record_t GetColumnReadersImpl(std::string_view name, const std::type_info &) = 0;
231};
232
233} // ns RDF
234
235} // ns ROOT
236
237/// Print a RDataSource at the prompt
238namespace cling {
239inline std::string printValue(ROOT::RDF::RDataSource *ds)
240{
241 return ds->AsString();
242}
243} // namespace cling
244
245#endif // ROOT_TDATASOURCE
unsigned long long ULong64_t
Definition RtypesCore.h:81
winID h TVirtualViewer3D TVirtualGLPainter p
char name[80]
Definition TGX11.cxx:110
Mother class of TTypedPointerHolder.
virtual TPointerHolder * GetDeepCopy()=0
Class to wrap a pointer and delete the memory associated to it correctly.
RDataSource defines an API that RDataFrame can use to read arbitrary data formats.
virtual bool HasColumn(std::string_view colName) const =0
Checks if the dataset has a certain column.
virtual void Finalize()
Convenience method called after concluding an event-loop.
virtual void InitSlot(unsigned int, ULong64_t)
Convenience method called at the start of the data processing associated to a slot.
virtual void FinalizeSlot(unsigned int)
Convenience method called at the end of the data processing associated to a slot.
virtual ~RDataSource()=default
virtual std::string AsString()
virtual bool SetEntry(unsigned int slot, ULong64_t entry)=0
Advance the "cursors" returned by GetColumnReaders to the selected entry for a particular slot.
std::vector< void * > Record_t
virtual std::string GetLabel()
Return a string representation of the datasource type.
virtual void SetNSlots(unsigned int nSlots)=0
Inform RDataSource of the number of processing slots (i.e.
virtual const std::vector< std::string > & GetColumnNames() const =0
Returns a reference to the collection of the dataset's column names.
virtual std::vector< std::pair< ULong64_t, ULong64_t > > GetEntryRanges()=0
Return ranges of entries to distribute to tasks.
virtual Record_t GetColumnReadersImpl(std::string_view name, const std::type_info &)=0
type-erased vector of pointers to pointers to column values - one per slot
virtual std::string GetTypeName(std::string_view colName) const =0
Type of a column as a string, e.g.
std::vector< T ** > GetColumnReaders(std::string_view columnName)
Called at most once per column by RDF.
virtual std::unique_ptr< ROOT::Detail::RDF::RColumnReaderBase > GetColumnReaders(unsigned int, std::string_view, const std::type_info &)
If the other GetColumnReaders overload returns an empty vector, this overload will be called instead.
virtual void Initialize()
Convenience method called before starting an event-loop.
This file contains a specialised ROOT message handler to test for diagnostic in unit tests.