io.deephaven.engine.table.impl.SortHelpers Maven / Gradle / Ivy
Go to download
Show more of this group Show more artifacts with this name
Show all versions of deephaven-engine-table Show documentation
Show all versions of deephaven-engine-table Show documentation
Engine Table: Implementation and closely-coupled utilities
/**
* Copyright (c) 2016-2022 Deephaven Data Labs and Patent Pending
*/
package io.deephaven.engine.table.impl;
import io.deephaven.base.verify.Assert;
import io.deephaven.base.verify.Require;
import io.deephaven.chunk.attributes.Any;
import io.deephaven.chunk.attributes.ChunkLengths;
import io.deephaven.chunk.attributes.ChunkPositions;
import io.deephaven.chunk.attributes.Values;
import io.deephaven.configuration.Configuration;
import io.deephaven.datastructures.util.CollectionUtil;
import io.deephaven.engine.rowset.RowSequence;
import io.deephaven.engine.rowset.RowSet;
import io.deephaven.engine.rowset.TrackingRowSet;
import io.deephaven.engine.rowset.RowSequenceFactory;
import io.deephaven.engine.table.WritableColumnSource;
import io.deephaven.engine.table.impl.indexer.RowSetIndexer;
import io.deephaven.engine.table.Table;
import io.deephaven.engine.table.impl.sort.LongMegaMergeKernel;
import io.deephaven.engine.table.impl.sources.*;
import io.deephaven.engine.table.impl.util.*;
import io.deephaven.engine.rowset.chunkattributes.RowKeys;
import io.deephaven.util.QueryConstants;
import io.deephaven.util.datastructures.LongSizedDataStructure;
import io.deephaven.engine.table.impl.sort.findruns.FindRunsKernel;
import io.deephaven.engine.table.impl.sort.timsort.LongIntTimsortKernel;
import io.deephaven.engine.table.impl.sort.permute.PermuteKernel;
import io.deephaven.engine.table.impl.sort.LongSortKernel;
import io.deephaven.engine.table.ColumnSource;
import io.deephaven.engine.table.impl.sources.LongArraySource;
import io.deephaven.engine.table.impl.sources.LongSparseArraySource;
import io.deephaven.chunk.*;
import io.deephaven.engine.table.impl.sources.regioned.SymbolTableSource;
import io.deephaven.util.annotations.VisibleForTesting;
import io.deephaven.util.type.TypeUtils;
import org.apache.commons.lang3.mutable.MutableInt;
import org.jetbrains.annotations.NotNull;
import java.lang.reflect.Array;
import java.util.Arrays;
import java.util.Map;
import java.util.function.LongPredicate;
public class SortHelpers {
public static boolean sortBySymbolTable =
Configuration.getInstance().getBooleanWithDefault("QueryTable.sortBySymbolTable", true);
/**
* If we have more than this many entries per group, instead of creating a large flat row redirection, we create a
* row redirection that is composed of the group indices and an accumulated cardinality cache. This can save a
* significant amount of memory when the groups are large and storing them using our RowSet structure is more
* efficient.
*/
public static int groupedRedirectionThreshold =
Configuration.getInstance().getIntegerWithDefault("SortHelpers.groupedRedirectionThreshold", 32);
/**
* When the sort is greater than or equal to than megaSortSize, instead of sorting one large chunk, we will sort
* individual chunks of sortChunkSize and then merge them into ArrayBackedColumnSources with the
* LongMegaMergeKernel.
*
* There are some boundary conditions in the chunk sizing math that make Integer.MAX_VALUE fail; you could probably
* back off to Integer.MAX_VALUE - 32 safely. We're being very conservative with 1 << 30 instead.
*/
@VisibleForTesting
static int megaSortSize = Configuration.getInstance().getIntegerWithDefault("QueryTable.sortChunkSize", 1 << 30);
/**
* The size of each chunk of a sort in a mega-merge based sort.
*/
@VisibleForTesting
static int sortChunkSize = Configuration.getInstance().getIntegerWithDefault("QueryTable.sortChunkSize", 1 << 30);
interface SortMapping extends LongSizedDataStructure {
long size();
long[] getArrayMapping();
boolean forEachLong(LongPredicate consumer);
WritableRowRedirection makeHistoricalRowRedirection();
}
final static class ArraySortMapping implements SortMapping {
/**
* The total size of the mapping.
*/
final int size;
/**
* The mapping as an array.
*/
@NotNull
final long[] mapping;
private ArraySortMapping(long[] mapping) {
size = mapping.length;
this.mapping = mapping;
}
@Override
final public long size() {
return size;
}
@NotNull
final public long[] getArrayMapping() {
return Require.neqNull(mapping, "mapping");
}
@Override
final public boolean forEachLong(LongPredicate consumer) {
for (int ii = 0; ii < mapping.length; ++ii) {
if (!consumer.test(mapping[ii])) {
return false;
}
}
return true;
}
@Override
public WritableRowRedirection makeHistoricalRowRedirection() {
return new ContiguousWritableRowRedirection(mapping);
}
}
final static class ColumnSourceSortMapping implements SortMapping {
/**
* The total size of the mapping.
*/
final long size;
/**
* The mapping as a column source, in locations 0..size-1
*/
final LongArraySource columnSource;
private ColumnSourceSortMapping(LongArraySource columnSource, long size) {
this.size = size;
this.columnSource = columnSource;
}
@Override
final public long size() {
return size;
}
@NotNull
final public long[] getArrayMapping() {
throw new ArrayIndexOutOfBoundsException();
}
@Override
final public boolean forEachLong(LongPredicate consumer) {
assert null != columnSource;
for (long ii = 0; ii < size; ++ii) {
if (!consumer.test(columnSource.getLong(ii))) {
return false;
}
}
return true;
}
@Override
public WritableRowRedirection makeHistoricalRowRedirection() {
return new LongColumnSourceWritableRowRedirection(columnSource);
}
}
final static class GroupedSortMapping implements SortMapping {
private final long size;
private final long[] groupSize;
private final RowSet[] groups;
private GroupedSortMapping(long size, long[] groupSize, RowSet[] groups) {
this.size = size;
this.groupSize = groupSize;
this.groups = groups;
}
@Override
final public long size() {
return size;
}
@NotNull
final public long[] getArrayMapping() {
if (size <= Integer.MAX_VALUE) {
final long[] mapping = new long[(int) size];
final MutableInt pos = new MutableInt(0);
forEachLong(idx -> {
mapping[pos.getAndIncrement()] = idx;
return true;
});
return mapping;
} else {
throw new ArrayIndexOutOfBoundsException();
}
}
@Override
final public boolean forEachLong(LongPredicate consumer) {
for (int ii = 0; ii < groups.length; ++ii) {
if (!groups[ii].forEachRowKey(consumer::test)) {
return false;
}
}
return true;
}
@Override
public WritableRowRedirection makeHistoricalRowRedirection() {
return new GroupedWritableRowRedirection(size, groupSize, groups);
}
}
static private final SortMapping EMPTY_SORT_MAPPING = new ArraySortMapping(CollectionUtil.ZERO_LENGTH_LONG_ARRAY);
/**
* Note that if usePrev is true, then rowSetToSort is the previous RowSet; not the current RowSet, and we should not
* need to call copyPrev.
*/
static SortMapping getSortedKeys(SortingOrder[] order, ColumnSource>[] columnsToSortBy,
RowSet rowSetToSort, boolean usePrev) {
return getSortedKeys(order, columnsToSortBy, rowSetToSort, usePrev, sortBySymbolTable);
}
/**
* Note that if usePrev is true, then rowSetToSort is the previous RowSet; not the current RowSet, and we should not
* need to call copyPrev.
*/
static SortMapping getSortedKeys(SortingOrder[] order, ColumnSource>[] columnsToSortBy,
RowSet rowSetToSort, boolean usePrev, boolean allowSymbolTable) {
if (rowSetToSort.size() == 0) {
return EMPTY_SORT_MAPPING;
}
if (columnsToSortBy.length == 1) {
if (rowSetToSort.isTracking()
&& RowSetIndexer.of(rowSetToSort.trackingCast()).hasGrouping(columnsToSortBy[0])) {
if (!usePrev || columnsToSortBy[0].isImmutable()) {
return getSortMappingGrouped(order[0], columnsToSortBy[0], rowSetToSort.trackingCast());
} else {
return getSortMappingOne(order[0], columnsToSortBy[0], rowSetToSort, usePrev);
}
} else {
if (allowSymbolTable && columnsToSortBy[0] instanceof SymbolTableSource
&& ((SymbolTableSource>) columnsToSortBy[0]).hasSymbolTable(rowSetToSort)) {
return doSymbolTableMapping(order[0], columnsToSortBy[0], rowSetToSort, usePrev);
} else {
return getSortMappingOne(order[0], columnsToSortBy[0], rowSetToSort, usePrev);
}
}
}
return getSortMappingMulti(order, columnsToSortBy, rowSetToSort, usePrev);
}
private static class SparseSymbolMapping {
private final int maxMapping;
private final int minTrailing;
private final int[][] lookupTable;
private SparseSymbolMapping(int maxMapping, int minTrailing, int[][] lookupTable) {
this.maxMapping = maxMapping;
this.minTrailing = minTrailing;
this.lookupTable = lookupTable;
}
private int getMaxMapping() {
return maxMapping;
}
private byte lookupByte(long symTabId) {
if (symTabId == QueryConstants.NULL_LONG) {
return QueryConstants.NULL_BYTE;
}
return (byte) doIntLookup(symTabId);
}
private short lookupShort(long symTabId) {
if (symTabId == QueryConstants.NULL_LONG) {
return QueryConstants.NULL_SHORT;
}
return (short) doIntLookup(symTabId);
}
private int lookupInt(long symTabId) {
if (symTabId == QueryConstants.NULL_LONG) {
return QueryConstants.NULL_INT;
}
return doIntLookup(symTabId);
}
private int doIntLookup(long symTabId) {
final int region = (int) (symTabId >> (32 + minTrailing));
final int id = (int) symTabId;
return lookupTable[region][id];
}
private static SparseSymbolMapping createMapping(LongChunk originalSymbol, IntChunk mappedIndex) {
// figure out what the maximum region is, and determine how many bits of it there are
int maxUpperPart = 0;
int minTrailing = 32;
int maxSymbol = 0;
for (int ii = 0; ii < originalSymbol.size(); ++ii) {
final long symTabId = originalSymbol.get(ii);
final long upperPart = (symTabId >> 32);
maxUpperPart = Math.max(maxUpperPart, (int) upperPart);
minTrailing = Math.min(minTrailing, Integer.numberOfTrailingZeros((int) upperPart));
maxSymbol = Math.max(maxSymbol, (int) symTabId);
}
final int maxShiftedRegion = maxUpperPart >> minTrailing;
if (minTrailing == 32) {
// this means we only found a zero region, in which case we do not want to shift by 64, but rather by
// zero to just truncate the region entirely
Assert.eqZero(maxShiftedRegion, "maxShiftedRegion");
minTrailing = 0;
}
final int[][] lookupTable = new int[maxShiftedRegion + 1][maxSymbol + 1];
// maxMapping ends up being the number of unique string values that we have. We compute it so that we can
// map symbol IDs to these unique integers using the narrowest primitive sorting kernel possible.
int maxMapping = 0;
for (int ii = 0; ii < originalSymbol.size(); ++ii) {
final long symTabId = originalSymbol.get(ii);
final int region = (int) (symTabId >> (32 + minTrailing));
final int id = (int) symTabId;
final int mappedId = mappedIndex.get(ii);
maxMapping = Math.max(maxMapping, mappedId);
lookupTable[region][id] = mappedId;
}
return new SparseSymbolMapping(maxMapping, minTrailing, lookupTable);
}
}
private static final String SORTED_INDEX_COLUMN_NAME = "SortedIndex";
private static final String SORTED_INDEX_COLUMN_UPDATE = SORTED_INDEX_COLUMN_NAME + "=i";
private static SortMapping doSymbolTableMapping(SortingOrder order, ColumnSource> columnSource,
RowSet rowSet, boolean usePrev) {
final int sortSize = rowSet.intSize();
final ColumnSource reinterpreted = columnSource.reinterpret(long.class);
final Table symbolTable = ((SymbolTableSource) columnSource).getStaticSymbolTable(rowSet, true);
if (symbolTable.size() >= sortSize) {
// the very first thing we will do is sort the symbol table, using a regular sort; if it is larger than the
// actual table we care to sort, then it is wasteful to use the symbol table sorting
return getSortMappingOne(order, columnSource, rowSet, usePrev);
}
final Table idMapping = symbolTable.sort(SymbolTableSource.SYMBOL_COLUMN_NAME)
.groupBy(SymbolTableSource.SYMBOL_COLUMN_NAME).update(SORTED_INDEX_COLUMN_UPDATE).ungroup()
.view(SymbolTableSource.ID_COLUMN_NAME, SORTED_INDEX_COLUMN_NAME);
final int symbolEntries = idMapping.intSize();
final SparseSymbolMapping mapping;
try (final WritableLongChunk originalSymbol = WritableLongChunk.makeWritableChunk(symbolEntries);
final WritableIntChunk mappedIndex = WritableIntChunk.makeWritableChunk(symbolEntries)) {
final ColumnSource idSource = idMapping.getColumnSource(SymbolTableSource.ID_COLUMN_NAME);
try (final ColumnSource.FillContext idContext = idSource.makeFillContext(symbolEntries)) {
idSource.fillChunk(idContext, originalSymbol, idMapping.getRowSet());
}
final ColumnSource sortedRowSetSource = idMapping.getColumnSource(SORTED_INDEX_COLUMN_NAME);
try (final ColumnSource.FillContext sortedIndexContext =
sortedRowSetSource.makeFillContext(symbolEntries)) {
sortedRowSetSource.fillChunk(sortedIndexContext, mappedIndex, idMapping.getRowSet());
}
mapping = SparseSymbolMapping.createMapping(originalSymbol, mappedIndex);
}
// Read the symbol table values into the unmappedValues chunk. The reinterpreted source provides the region and
// the symbol ID within the region as a packed long, which we then unpack in the type-specific loops below.
try (final WritableLongChunk unmappedValues =
makeAndFillValues(usePrev, rowSet, reinterpreted).asWritableLongChunk()) {
// Mapped values generic is the chunk of unique integer keys for the sort operation, using the narrowest
// possible primitive type (byte, short, or int).
final WritableChunk mappedValuesGeneric;
final LongSortKernel sortContext;
if (mapping.getMaxMapping() <= Byte.MAX_VALUE) {
final WritableByteChunk mappedValues;
mappedValues = WritableByteChunk.makeWritableChunk(rowSet.intSize());
for (int ii = 0; ii < unmappedValues.size(); ++ii) {
// symTabId is the packed ID, nulls get converted to the right kind of null, other values are pulled
// apart into region and id; and then used as offsets in the lookup table to get the unique sorted
// id
// and the mapped value is added to the mappedValues chunk
mappedValues.set(ii, mapping.lookupByte(unmappedValues.get(ii)));
}
sortContext = LongSortKernel.makeContext(ChunkType.Byte, order, sortSize, false);
mappedValuesGeneric = mappedValues;
} else if (mapping.getMaxMapping() <= Short.MAX_VALUE) {
final WritableShortChunk mappedValues = WritableShortChunk.makeWritableChunk(rowSet.intSize());
for (int ii = 0; ii < unmappedValues.size(); ++ii) {
// symTabId is the packed ID, nulls get converted to the right kind of null, other values are pulled
// apart into region and id; and then used as offsets in the lookup table to get the unique sorted
// id
// and the mapped value is added to the mappedValues chunk
mappedValues.set(ii, mapping.lookupShort(unmappedValues.get(ii)));
}
sortContext = LongSortKernel.makeContext(ChunkType.Short, order, sortSize, false);
mappedValuesGeneric = mappedValues;
} else {
final WritableIntChunk mappedValues = WritableIntChunk.makeWritableChunk(rowSet.intSize());
for (int ii = 0; ii < unmappedValues.size(); ++ii) {
// symTabId is the packed ID, nulls get converted to the right kind of null, other values are pulled
// apart into region and id; and then used as offsets in the lookup table to get the unique sorted
// id
// and the mapped value is added to the mappedValues chunk
mappedValues.set(ii, mapping.lookupInt(unmappedValues.get(ii)));
}
sortContext = LongSortKernel.makeContext(ChunkType.Int, order, sortSize, false);
mappedValuesGeneric = mappedValues;
}
// Fill a chunk that is Writable, and does not have an ordered tag with the row keys that we are sorting,
// the RowSet would do something very similar inside
// io.deephaven.engine.table.impl.util.RowSequence.asRowKeyChunk; but provides a LongChunk
// as its return.
final long[] rowKeysArray = new long[sortSize];
final WritableLongChunk rowKeys = WritableLongChunk.writableChunkWrap(rowKeysArray);
rowSet.fillRowKeyChunk(rowKeys);
sortContext.sort(rowKeys, mappedValuesGeneric);
sortContext.close();
mappedValuesGeneric.close();
return new ArraySortMapping(rowKeysArray);
}
}
private static SortMapping getSortMappingOne(SortingOrder order, ColumnSource> columnSource,
RowSet rowSet, boolean usePrev) {
final long sortSize = rowSet.size();
if (sortSize >= megaSortSize) {
return doMegaSortOne(order, columnSource, rowSet, usePrev, sortSize);
} else {
return new ArraySortMapping(doChunkSortingOne(order, columnSource, rowSet, usePrev, (int) sortSize));
}
}
@NotNull
private static SortMapping doMegaSortOne(SortingOrder order, ColumnSource> columnSource,
RowSet rowSet, boolean usePrev, long sortSize) {
final LongArraySource resultIndices = new LongArraySource();
resultIndices.ensureCapacity(sortSize, false);
final WritableColumnSource valuesToMerge =
ArrayBackedColumnSource.getMemoryColumnSource(0, columnSource.getType());
valuesToMerge.ensureCapacity(sortSize, false);
long accumulatedSize = 0;
final LongMegaMergeKernel longMegaMergeKernel =
LongMegaMergeKernel.makeContext(columnSource.getChunkType(), order);
try (final LongSortKernel sortContext =
LongSortKernel.makeContext(columnSource.getChunkType(), order, sortChunkSize, true);
final RowSequence.Iterator rsIt = rowSet.getRowSequenceIterator()) {
while (rsIt.hasMore()) {
final RowSequence chunkOk = rsIt.getNextRowSequenceWithLength(sortChunkSize);
final int chunkSize = chunkOk.intSize();
try (final WritableChunk partialValues = makeAndFillValues(usePrev, chunkOk, columnSource)) {
final long[] partialKeysArray = new long[chunkSize];
final WritableLongChunk partialKeys =
WritableLongChunk.writableChunkWrap(partialKeysArray);
chunkOk.fillRowKeyChunk(partialKeys);
sortContext.sort(partialKeys, partialValues);
longMegaMergeKernel.merge(resultIndices, valuesToMerge, 0, accumulatedSize, partialKeys,
partialValues);
accumulatedSize += chunkSize;
}
}
}
return new ColumnSourceSortMapping(resultIndices, sortSize);
}
@NotNull
private static long[] doChunkSortingOne(SortingOrder order, ColumnSource> columnSource,
RowSequence rowSequence, boolean usePrev, int chunkSize) {
try (final WritableChunk values = makeAndFillValues(usePrev, rowSequence, columnSource)) {
final long[] rowKeysArray = new long[chunkSize];
final WritableLongChunk rowKeys = WritableLongChunk.writableChunkWrap(rowKeysArray);
rowSequence.fillRowKeyChunk(rowKeys);
try (final LongSortKernel sortContext =
LongSortKernel.makeContext(columnSource.getChunkType(), order, chunkSize, false)) {
sortContext.sort(rowKeys, values);
}
return rowKeysArray;
}
}
private static SortMapping getSortMappingGrouped(SortingOrder order, ColumnSource> columnSource,
TrackingRowSet rowSet) {
final Map groupToRange = RowSetIndexer.of(rowSet).getGrouping(columnSource);
final Object[] keys = groupToRange.keySet().toArray((Object[]) Array.newInstance(
io.deephaven.util.type.TypeUtils.getBoxedType(columnSource.getType()), groupToRange.size()));
Arrays.sort((Comparable[]) keys, order.getComparator());
final boolean fitsIntoArray = rowSet.size() < (long) Integer.MAX_VALUE;
final long elementsPerGroup = rowSet.size() / (long) groupToRange.size();
if (fitsIntoArray && (elementsPerGroup < groupedRedirectionThreshold)) {
final long[] rowKeysArray = new long[rowSet.intSize()];
final MutableInt outputIdx = new MutableInt(0);
for (final Object key : keys) {
final RowSet group = groupToRange.get(key).intersect(rowSet);
group.forAllRowKeys(rowKey -> {
rowKeysArray[outputIdx.intValue()] = rowKey;
outputIdx.increment();
});
}
return new ArraySortMapping(rowKeysArray);
} else {
// create a grouped row redirection
final long[] groupSize = new long[groupToRange.size()];
final RowSet[] groupRowSet = new RowSet[groupToRange.size()];
long outputSize = 0;
int ii = 0;
for (final Object key : keys) {
final RowSet group = groupToRange.get(key).intersect(rowSet);
outputSize += group.size();
groupSize[ii] = outputSize;
groupRowSet[ii++] = group;
}
return new GroupedSortMapping(outputSize, groupSize, groupRowSet);
}
}
private static SortMapping getSortMappingMulti(SortingOrder[] order, ColumnSource>[] columnSources,
RowSet rowSet, boolean usePrev) {
Assert.gt(columnSources.length, "columnSources.length", 1);
final int sortSize = rowSet.intSize();
final long[] rowKeysArray = new long[sortSize];
final WritableLongChunk rowKeys = WritableLongChunk.writableChunkWrap(rowKeysArray);
WritableIntChunk offsetsOut = WritableIntChunk.makeWritableChunk((sortSize + 1) / 2);
WritableIntChunk lengthsOut = WritableIntChunk.makeWritableChunk((sortSize + 1) / 2);
ColumnSource> columnSource = columnSources[0];
if (rowSet.isTracking() && RowSetIndexer.of(rowSet.trackingCast()).hasGrouping(columnSources[0])) {
final Map, RowSet> groupToRange = columnSource.getGroupToRange();
final Object[] keys = groupToRange.keySet().toArray(
(Object[]) Array.newInstance(TypeUtils.getBoxedType(columnSource.getType()), groupToRange.size()));
Arrays.sort((Comparable[]) keys, order[0].getComparator());
offsetsOut.setSize(0);
lengthsOut.setSize(0);
final MutableInt outputIdx = new MutableInt(0);
for (final Object key : keys) {
// noinspection SuspiciousMethodCalls
final RowSet group = groupToRange.get(key).intersect(rowSet);
if (group.size() > 1) {
offsetsOut.add(outputIdx.intValue());
lengthsOut.add(group.intSize());
}
group.forAllRowKeys(rowKey -> {
rowKeysArray[outputIdx.intValue()] = rowKey;
outputIdx.increment();
});
}
} else {
rowSet.fillRowKeyChunk(rowKeys);
final ChunkType chunkType = columnSource.getChunkType();
final WritableChunk values = makeAndFillValues(usePrev, rowSet, columnSource);
try (final LongSortKernel sortContext =
LongSortKernel.makeContext(chunkType, order[0], sortSize, true)) {
sortContext.sort(rowKeys, values);
}
final FindRunsKernel findRunsKernel = FindRunsKernel.getInstance(chunkType);
findRunsKernel.findRuns(values, offsetsOut, lengthsOut);
values.close();
}
if (offsetsOut.size() == 0) {
lengthsOut.close();
offsetsOut.close();
return new ArraySortMapping(rowKeysArray);
}
final int totalRunLength = sumChunk(lengthsOut);
final WritableLongChunk indicesToFetch = WritableLongChunk.makeWritableChunk(totalRunLength);
final WritableIntChunk originalPositions = WritableIntChunk.makeWritableChunk(totalRunLength);
final LongIntTimsortKernel.LongIntSortKernelContext sortIndexContext =
LongIntTimsortKernel.createContext(totalRunLength);
ChunkType chunkType = columnSources[1].getChunkType();
int maximumSecondarySize =
computeIndicesToFetch(rowKeys, offsetsOut, lengthsOut, indicesToFetch, originalPositions);
WritableChunk values = fetchSecondaryValues(usePrev, columnSources[1], indicesToFetch,
originalPositions, sortIndexContext, maximumSecondarySize);
try (final LongSortKernel sortContext =
LongSortKernel.makeContext(chunkType, order[1], indicesToFetch.size(), columnSources.length != 2)) {
// and we can sort the stuff within the run now
sortContext.sort(rowKeys, values, offsetsOut, lengthsOut);
}
WritableIntChunk offsetsIn = offsetsOut;
WritableIntChunk lengthsIn = lengthsOut;
if (columnSources.length > 2) {
offsetsOut = WritableIntChunk.makeWritableChunk((totalRunLength + 1) / 2);
lengthsOut = WritableIntChunk.makeWritableChunk((totalRunLength + 1) / 2);
for (int columnIndex = 2; columnIndex < columnSources.length; ++columnIndex) {
columnSource = columnSources[columnIndex];
final FindRunsKernel findRunsKernel = FindRunsKernel.getInstance(chunkType);
findRunsKernel.findRuns(values, offsetsIn, lengthsIn, offsetsOut, lengthsOut);
if (offsetsOut.size() == 0) {
break;
}
chunkType = columnSource.getChunkType();
maximumSecondarySize =
computeIndicesToFetch(rowKeys, offsetsOut, lengthsOut, indicesToFetch, originalPositions);
values.close();
values = fetchSecondaryValues(usePrev, columnSources[columnIndex], indicesToFetch, originalPositions,
sortIndexContext, maximumSecondarySize);
try (final LongSortKernel sortContext = LongSortKernel.makeContext(chunkType,
order[columnIndex], indicesToFetch.size(), columnIndex != columnSources.length - 1)) {
// and we can sort the stuff within the run now
sortContext.sort(rowKeys, values, offsetsOut, lengthsOut);
}
final WritableIntChunk tempOff = offsetsIn;
final WritableIntChunk tempLen = lengthsIn;
offsetsIn = offsetsOut;
lengthsIn = lengthsOut;
offsetsOut = tempOff;
lengthsOut = tempLen;
}
offsetsIn.close();
lengthsIn.close();
}
values.close();
lengthsOut.close();
offsetsOut.close();
sortIndexContext.close();
originalPositions.close();
indicesToFetch.close();
return new ArraySortMapping(rowKeysArray);
}
private static WritableChunk fetchSecondaryValues(boolean usePrev, ColumnSource columnSource,
WritableLongChunk indicesToFetch, WritableIntChunk originalPositions,
LongIntTimsortKernel.LongIntSortKernelContext sortIndexContext,
int maximumSecondarySize) {
sortIndexContext.sort(originalPositions, indicesToFetch);
try (final WritableChunk secondaryValues = makeAndFillValues(usePrev,
RowSequenceFactory.wrapRowKeysChunkAsRowSequence(WritableLongChunk.downcast(indicesToFetch)),
columnSource)) {
final ChunkType chunkType = columnSource.getChunkType();
// make the big chunk that can hold all the relevant secondary values, in their desired position
final WritableChunk values = chunkType.makeWritableChunk(maximumSecondarySize);
final PermuteKernel permuteKernel = PermuteKernel.makePermuteKernel(chunkType);
permuteKernel.permute(secondaryValues, originalPositions, values);
return values;
}
}
private static int computeIndicesToFetch(WritableLongChunk rowKeys,
WritableIntChunk offsetsOut, WritableIntChunk lengthsOut,
WritableLongChunk indicesToFetch, WritableIntChunk originalPositions) {
indicesToFetch.setSize(0);
originalPositions.setSize(0);
int maximumSecondarySize = 0;
for (int ii = 0; ii < offsetsOut.size(); ++ii) {
final int runStart = offsetsOut.get(ii);
final int runLength = lengthsOut.get(ii);
maximumSecondarySize = Math.max(maximumSecondarySize, runStart + runLength);
for (int jj = 0; jj < runLength; ++jj) {
indicesToFetch.add(rowKeys.get(runStart + jj));
originalPositions.add(runStart + jj);
}
}
return maximumSecondarySize;
}
private static int sumChunk(IntChunk lengthsOut) {
int sum = 0;
for (int ii = 0; ii < lengthsOut.size(); ++ii) {
sum += lengthsOut.get(ii);
}
return sum;
}
@NotNull
private static WritableChunk makeAndFillValues(boolean usePrev, RowSequence ok,
ColumnSource columnSource) {
final int sortSize = LongSizedDataStructure.intSize("SortHelper.makeAndFillValues", ok.size());
final WritableChunk values = columnSource.getChunkType().makeWritableChunk(sortSize);
try (final ColumnSource.FillContext primaryColumnSourceContext = columnSource.makeFillContext(sortSize)) {
if (usePrev) {
columnSource.fillPrevChunk(primaryColumnSourceContext, values, ok);
} else {
columnSource.fillChunk(primaryColumnSourceContext, values, ok);
}
}
return values;
}
@NotNull
static WritableRowRedirection createSortRowRedirection() {
final WritableColumnSource sparseLongSource = new LongSparseArraySource();
return new LongColumnSourceWritableRowRedirection(sparseLongSource);
}
}
© 2015 - 2024 Weber Informatics LLC | Privacy Policy