The RDataSource implementation for RNTuple.
It lets RDataFrame read RNTuple data.
An RDataFrame that reads RNTuple data can be constructed using FromRNTuple().
For each column containing an array or a collection, a corresponding column #colname is available to access colname.size() without reading and deserializing the collection values.
Definition at line 100 of file RNTupleDS.hxx.
Classes | |
| struct | REntryRangeDS |
| Page source information consumed while running the computation graph. More... | |
| struct | RFieldInfo |
| Holds useful information about fields added to the RNTupleDS. More... | |
Public Member Functions | |
| RNTupleDS (const RNTupleDS &)=delete | |
| RNTupleDS (RNTupleDS &&)=delete | |
| RNTupleDS (std::string_view ntupleName, const std::vector< std::string > &fileNames) | |
| RNTupleDS (std::string_view ntupleName, std::string_view fileName) | |
| ~RNTupleDS () final | |
| ROOT::RDF::RSampleInfo | CreateSampleInfo (unsigned int, const std::unordered_map< std::string, ROOT::RDF::Experimental::RSample * > &) const final |
| void | Finalize () final |
| Convenience method called after concluding an event-loop. | |
| void | FinalizeSlot (unsigned int slot) final |
| Convenience method called at the end of the data processing associated to a slot. | |
| const std::vector< std::string > & | GetColumnNames () const final |
| Returns a reference to the collection of the dataset's column names. | |
| template<typename T > | |
| std::vector< T ** > | GetColumnReaders (std::string_view columnName) |
| Called at most once per column by RDF. | |
| std::unique_ptr< ROOT::Detail::RDF::RColumnReaderBase > | GetColumnReaders (unsigned int, std::string_view, const std::type_info &) final |
| If the other GetColumnReaders overload returns an empty vector, this overload will be called instead. | |
| std::vector< std::pair< ULong64_t, ULong64_t > > | GetEntryRanges () final |
| Return ranges of entries to distribute to tasks. | |
| std::string | GetLabel () final |
| Return a string representation of the datasource type. | |
| std::size_t | GetNFiles () const final |
| Returns the number of files from which the dataset is constructed. | |
| const std::vector< std::string > & | GetTopLevelFieldNames () const final |
| std::string | GetTypeName (std::string_view colName) const final |
| Type of a column as a string, e.g. | |
| bool | HasColumn (std::string_view colName) const final |
| Checks if the dataset has a certain column. | |
| void | Initialize () final |
| Convenience method called before starting an event-loop. | |
| void | InitSlot (unsigned int slot, ULong64_t firstEntry) final |
| Convenience method called at the start of the data processing associated to a slot. | |
| RNTupleDS & | operator= (const RNTupleDS &)=delete |
| RNTupleDS & | operator= (RNTupleDS &&)=delete |
| bool | SetEntry (unsigned int, ULong64_t) final |
| Advance the "cursors" returned by GetColumnReaders to the selected entry for a particular slot. | |
| virtual void | SetGlobalEntryRange (std::pair< ULong64_t, ULong64_t > entryRange) |
| Restrict processing to a [begin, end) range of entries. | |
| void | SetNSlots (unsigned int nSlots) final |
| Inform RDataSource of the number of processing slots (i.e. | |
Protected Types | |
| using | Record_t = std::vector<void *> |
Protected Member Functions | |
| virtual std::string | AsString () |
| virtual std::unique_ptr< ROOT::Detail::RDF::RColumnReaderBase > | CreateColumnReader (unsigned int slot, std::string_view col, const std::type_info &tid, TTreeReader *) |
| Creates a column reader for the requested column. | |
| virtual std::string | DescribeDataset () |
| virtual const std::vector< std::string > & | GetColumnNamesNoDuplicates () const |
| Record_t | GetColumnReadersImpl (std::string_view name, const std::type_info &) final |
| type-erased vector of pointers to pointers to column values - one per slot | |
| virtual std::string | GetTypeNameWithOpts (std::string_view colName, bool) const |
| virtual void | InitializeWithOpts (const std::set< std::string > &) |
| virtual void | RunFinalChecks (bool) const |
Protected Attributes | |
| std::optional< std::pair< ULong64_t, ULong64_t > > | fGlobalEntryRange {} |
| unsigned int | fNSlots {} |
Private Member Functions | |
| RNTupleDS (std::string_view ntupleName, const std::vector< std::string > &fileNames, const std::pair< ULong64_t, ULong64_t > &range) | |
| RNTupleDS (std::unique_ptr< ROOT::Internal::RPageSource > pageSource) | |
| void | AddField (const ROOT::RNTupleDescriptor &desc, std::string_view colName, ROOT::DescriptorId_t fieldId, std::vector< RFieldInfo > fieldInfos, bool convertToRVec=true) |
| Provides the RDF column "colName" given the field identified by fieldID. | |
| std::vector< std::uint64_t > | ComputeGlobalEntryOffsets (ROOT::TThreadExecutor &pool) const |
| Computes global entry offsets for the dataset. | |
| void | ExecStaging () |
| The main function of the fThreadStaging background thread. | |
| ROOT::RFieldBase * | GetFieldWithTypeChecks (std::string_view fieldName, const std::type_info &tid) |
| void | InsertActiveEntryRange (unsigned int slot, std::string_view fileName, std::shared_ptr< ROOT::Internal::RPageSource > pageSource, std::uint64_t beginEntry, std::uint64_t endEntry, std::uint64_t offset) |
| void | ParallelForBody (std::shared_ptr< ROOT::Internal::RPageSource > pageSource, const ROOT::Internal::IMTUtils::RAbstractTBBRangeImpl &rangeImpl, ROOT::Detail::RDF::RLoopManager &lm, std::string_view fileName, ROOT::Internal::RSlotStack &slotStack, std::atomic< ULong64_t > &processedEntries, ULong64_t columnReaderOffset) |
| Apply the computation graph to an input range. | |
| void | PrepareNextRanges () |
| Populates fNextRanges with the next set of entry ranges. | |
| void | ProcessMT (ROOT::Detail::RDF::RLoopManager &lm) final |
| Run the computation graph by scheduling MT tasks. | |
| void | ScheduleTaskForPageSource (ROOT::TThreadExecutor &pool, std::unique_ptr< ROOT::Internal::RPageSource > pageSource, std::string_view fileName, std::span< const ROOT::Internal::RNTupleClusterBoundaries > clusterBoundaries, std::uint64_t nEntries, ROOT::Detail::RDF::RLoopManager &lm, ROOT::Internal::RSlotStack &slotStack, std::atomic< ULong64_t > &processedEntries, std::atomic< ULong64_t > &globalEntries) |
| Call parallel_for to schedule a splittable task for a page source. | |
| void | ScheduleTasksForFile (std::size_t fileIdx, ROOT::TThreadExecutor &pool, ROOT::Detail::RDF::RLoopManager &lm, ROOT::Internal::RSlotStack &slotStack, std::atomic< ULong64_t > &processedEntries, std::atomic< ULong64_t > &globalEntries) |
| Launch the task(s) for the input file index. | |
| void | ScheduleTasksForFileWithGlobalOffset (std::size_t fileIdx, const std::vector< std::uint64_t > &globalFileOffsets, ROOT::TThreadExecutor &pool, ROOT::Detail::RDF::RLoopManager &lm, ROOT::Internal::RSlotStack &slotStack, std::atomic< ULong64_t > &processedEntries, std::atomic< ULong64_t > &globalEntries) |
| Launch the task(s) for the input file index, taking into account dataset global entry offsets. | |
| void | StageNextSources () |
Starting from fNextFileIndex, opens the next fNSlots files. | |
Private Attributes | |
| std::vector< std::vector< ROOT::Internal::RDF::RNTupleColumnReader * > > | fActiveColumnReaders |
| List of column readers returned by GetColumnReaders() organized by slot. | |
| std::vector< REntryRangeDS > | fActiveRangesPerSlot |
| The page source and entry window seen by an active thread task in multi-threaded runs. | |
| std::unordered_map< std::size_t, std::vector< std::unique_ptr< ROOT::RFieldBase > > > | fAlternativeProtoFields |
| Columns may be requested with types other than with which they were initially added as proto fields. | |
| std::vector< std::string > | fColumnNames |
| std::vector< std::string > | fColumnTypes |
| std::vector< REntryRangeDS > | fCurrentRanges |
| Basis for the ranges returned by the last GetEntryRanges() call. | |
| std::condition_variable | fCvStaging |
| Signal for the state information of fIsReadyForStaging and fHasNextSources. | |
| std::unordered_map< ROOT::DescriptorId_t, std::string > | fFieldId2QualifiedName |
| Connects the IDs of active proto fields and their subfields to their fully qualified name (a.b.c.d). | |
| std::vector< std::string > | fFileNames |
| std::unordered_map< ULong64_t, std::size_t > | fFirstEntry2RangeIdx |
| Maps the first entries from the ranges of the last GetEntryRanges() call to their corresponding index in the fCurrentRanges vectors. | |
| bool | fHasNextSources = false |
| Is true when the staging thread has populated the next batch of files to fStagingArea. | |
| bool | fIsReadyForStaging = false |
| Is true when the staging thread should start working. | |
| std::mutex | fMutexStaging |
| Protects the shared state between the main thread and the I/O thread. | |
| std::size_t | fNextFileIndex = 0 |
| Index into fFileNames to the next file to process. | |
| std::vector< REntryRangeDS > | fNextRanges |
| Basis for the ranges populated by the PrepareNextRanges() call. | |
| std::string | fNTupleName |
| The data source may be constructed with an ntuple name and a list of files. | |
| std::vector< std::pair< ULong64_t, ULong64_t > > | fOriginalRanges |
| ROOT::RNTupleDescriptor | fPrincipalDescriptor |
| A clone of the first pages source's descriptor. | |
| std::vector< std::unique_ptr< ROOT::RFieldBase > > | fProtoFields |
| We prepare a prototype field for every column. | |
| ULong64_t | fSeenEntriesNoGlobalRange = 0 |
| The number of entries seen so far in GetEntryRanges() | |
| std::vector< std::shared_ptr< ROOT::Internal::RPageSource > > | fStagingArea |
| The staging area is relevant for chains of files, i.e. | |
| bool | fStagingThreadShouldTerminate = false |
| Is true when the I/O thread should quit. | |
| std::thread | fThreadStaging |
| The background thread that runs StageNextSources() | |
| std::vector< std::string > | fTopLevelFieldNames |
Friends | |
| ROOT::RDataFrame | ROOT::Internal::RDF::FromRNTuple (std::string_view ntupleName, const std::vector< std::string > &fileNames, const std::pair< ULong64_t, ULong64_t > &range) |
| std::vector< std::pair< std::uint64_t, std::uint64_t > > | ROOT::Internal::RDF::GetDatasetGlobalClusterBoundaries (const ROOT::RDF::RInterface< ROOT::Detail::RDF::RNodeBase > &node) |
| class | ROOT::Internal::RDF::RNTupleColumnReader |
#include <ROOT/RNTupleDS.hxx>
|
protectedinherited |
Definition at line 142 of file RDataSource.hxx.
|
explicitprivate |
Definition at line 310 of file RNTupleDS.cxx.
|
explicitprivate |
Definition at line 369 of file RNTupleDS.cxx.
| ROOT::RDF::RNTupleDS::RNTupleDS | ( | std::string_view | ntupleName, |
| std::string_view | fileName ) |
Definition at line 354 of file RNTupleDS.cxx.
| ROOT::RDF::RNTupleDS::RNTupleDS | ( | std::string_view | ntupleName, |
| const std::vector< std::string > & | fileNames ) |
Definition at line 361 of file RNTupleDS.cxx.
|
delete |
|
delete |
|
finaldefault |
|
private |
Provides the RDF column "colName" given the field identified by fieldID.
For records and collections, AddField recurses into the sub fields. The fieldInfos argument is a list of objects holding info about the fields of the outer collection(s) (w.r.t. fieldId). For instance, if fieldId refers to an std::vector<Jet>, with
AddField will recurse into Jet.pt and Jet.eta and provide the two inner fields as ROOT::VecOps::RVec<float> each.
In case the field is a collection of type ROOT::VecOps::RVec, std::vector or std::array, its corresponding column is added as a ROOT::VecOps::RVec. Otherwise, the collection field's on-disk type is used. Note, however, that inner record members of such collections will still be added as ROOT::VecOps::RVec (e.g., std::set<Jet> will be added as astd::set, butJet.[pt|eta] will be added as `ROOT::VecOps::RVec<float>).
Definition at line 153 of file RNTupleDS.cxx.
|
inlineprotectedvirtualinherited |
Reimplemented in ROOT::Internal::RDF::RVecDS< ColumnTypes >, ROOT::RDF::RCsvDS, ROOT::RDF::RLazyDS< ColumnTypes >, ROOT::Internal::RDF::RRootDS, ROOT::RDF::RTrivialDS, and ROOT::Internal::RDF::RTTreeDS.
Definition at line 145 of file RDataSource.hxx.
|
private |
Computes global entry offsets for the dataset.
Used in MT scheduling when the user selects a global entry range to restrict processing to.
Definition at line 1018 of file RNTupleDS.cxx.
|
inlineprotectedvirtualinherited |
Creates a column reader for the requested column.
In the general case, this is just a redirect to the right GetColumnReaders overload. The signature notably also has a TTreeReader * parameter. This is currently necessary to still allow the TTree-based MT scheduling via TTreeProcessorMT. We use the TTreeProcessorMT::Process method to launch the same kernel across all threads. In each thread task, TTreeProcessorMT creates a thread-local instance of a TTreeReader which is going to read the range of events assigned to that task. That TTreeReader instance is what is passed to this method whenever a column reader needs to be created in a thread task. In the future this method might be removed by either allowing to request a handle to the thread-local TTreeReader instance programmatically from the TTreeProcessorMT, or refactoring the TTreeProcessorMT scheduling into RTTreeDS altogether.
Reimplemented in ROOT::Internal::RDF::RTTreeDS.
Definition at line 195 of file RDataSource.hxx.
|
finalvirtual |
Reimplemented from ROOT::RDF::RDataSource.
Definition at line 795 of file RNTupleDS.cxx.
|
inlineprotectedvirtualinherited |
Reimplemented in ROOT::Internal::RDF::RTTreeDS.
Definition at line 165 of file RDataSource.hxx.
|
private |
The main function of the fThreadStaging background thread.
Definition at line 478 of file RNTupleDS.cxx.
|
finalvirtual |
Convenience method called after concluding an event-loop.
See Initialize for more details.
Reimplemented from ROOT::RDF::RDataSource.
Definition at line 751 of file RNTupleDS.cxx.
|
finalvirtual |
Convenience method called at the end of the data processing associated to a slot.
| [in] | slot | The data processing slot wihch needs to be finalized This method might be called multiple times per thread per event-loop. |
Reimplemented from ROOT::RDF::RDataSource.
Definition at line 698 of file RNTupleDS.cxx.
|
inlinefinalvirtual |
Returns a reference to the collection of the dataset's column names.
Implements ROOT::RDF::RDataSource.
Definition at line 301 of file RNTupleDS.hxx.
|
inlineprotectedvirtualinherited |
Reimplemented in ROOT::Internal::RDF::RTTreeDS.
Definition at line 159 of file RDataSource.hxx.
|
inlineinherited |
Called at most once per column by RDF.
Return vector of pointers to pointers to column values - one per slot.
| T | The type of the data stored in the column |
| [in] | columnName | The name of the column |
These pointers are veritable cursors: it's a responsibility of the RDataSource implementation that they point to the "right" memory region.
Definition at line 248 of file RDataSource.hxx.
|
finalvirtual |
If the other GetColumnReaders overload returns an empty vector, this overload will be called instead.
| [in] | slot | The data processing slot that needs to be considered |
| [in] | name | The name of the column for which a column reader needs to be returned |
| [in] | tid | A type_info At least one of the two must return a non-empty/non-null value. |
Reimplemented from ROOT::RDF::RDataSource.
Definition at line 460 of file RNTupleDS.cxx.
|
finalprotectedvirtual |
type-erased vector of pointers to pointers to column values - one per slot
Implements ROOT::RDF::RDataSource.
Definition at line 377 of file RNTupleDS.cxx.
|
finalvirtual |
Return ranges of entries to distribute to tasks.
They are required to be contiguous intervals with no entries skipped. Supposing a dataset with nEntries, the intervals must start at 0 and end at nEntries, e.g. [0-5],[5-10] for 10 entries. This function will be invoked repeatedly by RDataFrame as it needs additional entries to process. The same entry range should not be returned more than once. Returning an empty collection of ranges signals to RDataFrame that the processing can stop.
Implements ROOT::RDF::RDataSource.
Definition at line 552 of file RNTupleDS.cxx.
|
private |
Definition at line 383 of file RNTupleDS.cxx.
|
inlinefinalvirtual |
Return a string representation of the datasource type.
The returned string will be used by ROOT::RDF::SaveGraph() to represent the datasource in the visualization of the computation graph. Concrete datasources can override the default implementation.
Reimplemented from ROOT::RDF::RDataSource.
Definition at line 306 of file RNTupleDS.hxx.
|
inlinefinalvirtual |
Returns the number of files from which the dataset is constructed.
Reimplemented from ROOT::RDF::RDataSource.
Definition at line 300 of file RNTupleDS.hxx.
|
inlinefinalvirtual |
Reimplemented from ROOT::RDF::RDataSource.
Definition at line 302 of file RNTupleDS.hxx.
|
finalvirtual |
Type of a column as a string, e.g.
GetTypeName("x") == "double". Required for jitting e.g. df.Filter("x>0").
| [in] | colName | The name of the column |
Implements ROOT::RDF::RDataSource.
Definition at line 709 of file RNTupleDS.cxx.
|
inlineprotectedvirtualinherited |
Reimplemented in ROOT::Internal::RDF::RTTreeDS.
Definition at line 152 of file RDataSource.hxx.
|
finalvirtual |
Checks if the dataset has a certain column.
| [in] | colName | The name of the column |
Implements ROOT::RDF::RDataSource.
Definition at line 722 of file RNTupleDS.cxx.
|
finalvirtual |
Convenience method called before starting an event-loop.
This method might be called multiple times over the lifetime of a RDataSource, since users can run multiple event-loops with the same RDataFrame. Ideally, Initialize should set the state of the RDataSource so that multiple identical event-loops will produce identical results.
Reimplemented from ROOT::RDF::RDataSource.
Definition at line 727 of file RNTupleDS.cxx.
|
inlineprotectedvirtualinherited |
Reimplemented in ROOT::Internal::RDF::RTTreeDS.
Definition at line 162 of file RDataSource.hxx.
Convenience method called at the start of the data processing associated to a slot.
| [in] | slot | The data processing slot wihch needs to be initialized |
| [in] | firstEntry | The first entry of the range that the task will process. This method might be called multiple times per thread per event-loop. |
Reimplemented from ROOT::RDF::RDataSource.
Definition at line 684 of file RNTupleDS.cxx.
|
private |
Definition at line 910 of file RNTupleDS.cxx.
|
private |
Apply the computation graph to an input range.
| pageSource | The page source to read data from. |
| rangeImpl | A Range type carrying information on the range of entries to process. |
| lm | The RLoopManager which will run the computation graph. |
| fileName | Path to the storage location of the page source. |
| slotStack | A stack of slots, each thread needs to request one before running the computation graph. |
| processedEntries | Cumulating number of processed entries by this computation graph run. |
| columnReaderOffset | Offset value to pass to column readers used in this task. The column reader will request entries from the page source receiving a global entry number, this offset is subtracted so that the column reader requests the local entry from the page source. |
Definition at line 934 of file RNTupleDS.cxx.
|
private |
Populates fNextRanges with the next set of entry ranges.
Moves files from the staging area as necessary and aligns ranges with cluster boundaries for scheduling the tail of files. Upon return, the fNextRanges list is ordered. It has usually fNSlots elements; fewer if there is not enough work to give at least one cluster to every slot.
Definition at line 515 of file RNTupleDS.cxx.
|
finalprivatevirtual |
Run the computation graph by scheduling MT tasks.
The data source interacts with the loop manager and with a pool of threads (via TThreadExecutor) at the same time. The function schedules one MT task per input file, eventually calling the parallel_for overload taking two input arguments respectively respecting the Range and Body requirements. This means that each of the file-based tasks can be split into various sub-tasks, each processing a set of consecutive clusters of the same file. Before launching the file-based tasks, the function checks if the user has selected a global range of entries to restrict processing to. If so, the function first computes the global entry offsets of all the input files, so they can be matched against the user-provided global range. These will be used to adjust the list of clusters of the file before spawning the corresponding task.
Reimplemented from ROOT::RDF::RDataSource.
Definition at line 1145 of file RNTupleDS.cxx.
|
inlineprotectedvirtualinherited |
Reimplemented in ROOT::Internal::RDF::RTTreeDS.
Definition at line 174 of file RDataSource.hxx.
|
private |
Call parallel_for to schedule a splittable task for a page source.
| pool | The pool of threads used to schedule the task. |
| pageSource | The RNTuple page source for the task. |
| fileName | Path to the storage location of the page source. |
| clusterBoundaries | A list of cluster boundaries valid for the page source. This may be computed differently, see ProcessMT. |
| nEntries | Number of entries in the page source. |
| lm | The RLoopManager which will run the computation graph. |
| slotStack | A stack of slots, each thread needs to request one before running the computation graph. |
| processedEntries | Cumulating number of processed entries by this computation graph run. |
| globalEntries | Cumulating total number of entries in the input dataset. |
This function calls the parallel_for overload that takes two arguments, one respecting the Range and the other respecting the Body requirements. The Range argument is a RSplittableRangeOfClusterBoundaries, a struct with a range of N cluster boundaries to process (initialised to the whole list of clusters). Immediately when calling parallel_for, TBB will recursively ask the RSplittableRangeOfClusterBoundaries to split itself into two roughly equal sub-tasks, until there are only leaf tasks that cannot be split further. This helps balancing cases where some tasks may end up running much longer than others.
Definition at line 992 of file RNTupleDS.cxx.
|
private |
Launch the task(s) for the input file index.
This function is used in MT scheduling to launch a task processing one input file. It is selected when there is no global entry range. The task may be split into smaller tasks based on entry ranges aligned to cluster boundaries, see ScheduleTaskForPageSource.
Definition at line 1111 of file RNTupleDS.cxx.
|
private |
Launch the task(s) for the input file index, taking into account dataset global entry offsets.
This function is used in MT scheduling to launch a task processing one input file. It is selected when the user specifies a global entry range to restrict processing to. In this scenario we have precomputed the global entry offsets of each file upfront so we can then adjust the range of entries in each file accordingly. The task may be split into smaller tasks based on entry ranges aligned to cluster boundaries, see ScheduleTaskForPageSource.
Definition at line 1050 of file RNTupleDS.cxx.
Advance the "cursors" returned by GetColumnReaders to the selected entry for a particular slot.
| [in] | slot | The data processing slot that needs to be considered |
| [in] | entry | The entry which needs to be pointed to by the reader pointers Slots are adopted to accommodate parallel data processing. Different workers will loop over different ranges and will be labelled by different "slot" values. Returns true if the entry has to be processed, false otherwise. |
Implements ROOT::RDF::RDataSource.
Definition at line 321 of file RNTupleDS.hxx.
|
inlinevirtualinherited |
Restrict processing to a [begin, end) range of entries.
| entryRange | The range of entries to process. |
Definition at line 327 of file RDataSource.hxx.
|
finalvirtual |
Inform RDataSource of the number of processing slots (i.e.
worker threads) used by the associated RDataFrame. Slots numbers are used to simplify parallel execution: RDataFrame guarantees that different threads will always pass different slot values when calling methods concurrently.
Reimplemented from ROOT::RDF::RDataSource.
Definition at line 776 of file RNTupleDS.cxx.
|
private |
Starting from fNextFileIndex, opens the next fNSlots files.
Calls LoadStructure() on the opened files. The very first file is already available from the constructor.
Definition at line 496 of file RNTupleDS.cxx.
|
friend |
|
friend |
|
friend |
Definition at line 101 of file RNTupleDS.hxx.
|
private |
List of column readers returned by GetColumnReaders() organized by slot.
Used to reconnect readers to new page sources when the files in the chain change.
Definition at line 174 of file RNTupleDS.hxx.
|
private |
The page source and entry window seen by an active thread task in multi-threaded runs.
Definition at line 182 of file RNTupleDS.hxx.
|
private |
Columns may be requested with types other than with which they were initially added as proto fields.
For example, a column with a ROOT::RVec<float> proto field may instead be requested as a std::vector<float>. In case this happens, we create an alternative proto field and store it here, with the original index in fProtoFields as key. A single column can have more than one alternative proto fields.
Definition at line 164 of file RNTupleDS.hxx.
|
private |
Definition at line 169 of file RNTupleDS.hxx.
|
private |
Definition at line 170 of file RNTupleDS.hxx.
|
private |
Basis for the ranges returned by the last GetEntryRanges() call.
Definition at line 178 of file RNTupleDS.hxx.
|
private |
Signal for the state information of fIsReadyForStaging and fHasNextSources.
Definition at line 195 of file RNTupleDS.hxx.
|
private |
Connects the IDs of active proto fields and their subfields to their fully qualified name (a.b.c.d).
This enables the column reader to rewire the field IDs when the file changes (chain), using the fully qualified name as a search key in the descriptor of the other page sources.
Definition at line 168 of file RNTupleDS.hxx.
|
private |
Definition at line 135 of file RNTupleDS.hxx.
|
private |
Maps the first entries from the ranges of the last GetEntryRanges() call to their corresponding index in the fCurrentRanges vectors.
This is necessary because the returned ranges get distributed arbitrarily onto slots. In the InitSlot method, the column readers use this map to find the correct range to connect to.
Definition at line 186 of file RNTupleDS.hxx.
|
protectedinherited |
Definition at line 149 of file RDataSource.hxx.
|
private |
Is true when the staging thread has populated the next batch of files to fStagingArea.
Definition at line 199 of file RNTupleDS.hxx.
|
private |
Is true when the staging thread should start working.
Definition at line 197 of file RNTupleDS.hxx.
|
private |
Protects the shared state between the main thread and the I/O thread.
Definition at line 193 of file RNTupleDS.hxx.
|
private |
Index into fFileNames to the next file to process.
Definition at line 153 of file RNTupleDS.hxx.
|
private |
Basis for the ranges populated by the PrepareNextRanges() call.
Definition at line 179 of file RNTupleDS.hxx.
|
protectedinherited |
Definition at line 147 of file RDataSource.hxx.
|
private |
The data source may be constructed with an ntuple name and a list of files.
Definition at line 134 of file RNTupleDS.hxx.
Definition at line 188 of file RNTupleDS.hxx.
|
private |
A clone of the first pages source's descriptor.
Definition at line 131 of file RNTupleDS.hxx.
|
private |
We prepare a prototype field for every column.
If a column reader is actually requested in GetColumnReaders(), we move a clone of the field into a new column reader for RDataFrame. Only the clone connects to the backing page store and acquires I/O resources. The field IDs are set in the context of the first source and used as keys in fFieldId2QualifiedName.
Definition at line 159 of file RNTupleDS.hxx.
|
private |
The number of entries seen so far in GetEntryRanges()
Definition at line 176 of file RNTupleDS.hxx.
|
private |
The staging area is relevant for chains of files, i.e.
when fFileNames is not empty. In this case, files are opened in the background in batches of size fNSlots and kept in the staging area. The first file (chains or no chains) is always opened on construction in order to process the schema. For all subsequent files, the corresponding page sources in the staging area only executed LoadStructure(), i.e. they should have a compressed buffer of the meta-data available. Concretely:
Initialize(), we start the I/O background thread, which in turn opens the first batch of files.GetEntryRanges(), we a) wait for the I/O thread to finish, b) call PrepareNextRanges() in the main thread to move the page sources from the staging area into fNextRanges; this will also call Attach() on the page sources (i.e., deserialize the meta-data), and c) trigger staging of the next batch of files in the I/O background thread.Finalize(), the I/O background thread is stopped. Definition at line 152 of file RNTupleDS.hxx.
|
private |
Is true when the I/O thread should quit.
Definition at line 201 of file RNTupleDS.hxx.
|
private |
The background thread that runs StageNextSources()
Definition at line 191 of file RNTupleDS.hxx.
|
private |
Definition at line 171 of file RNTupleDS.hxx.