org.apache.sysml.runtime.instructions.spark.AppendGSPInstruction Maven / Gradle / Ivy
Go to download
Show more of this group Show more artifacts with this name
Show all versions of systemml Show documentation
Show all versions of systemml Show documentation
Declarative Machine Learning
/*
* Licensed to the Apache Software Foundation (ASF) under one
* or more contributor license agreements. See the NOTICE file
* distributed with this work for additional information
* regarding copyright ownership. The ASF licenses this file
* to you under the Apache License, Version 2.0 (the
* "License"); you may not use this file except in compliance
* with the License. You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing,
* software distributed under the License is distributed on an
* "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
* KIND, either express or implied. See the License for the
* specific language governing permissions and limitations
* under the License.
*/
package org.apache.sysml.runtime.instructions.spark;
import java.util.ArrayList;
import java.util.Iterator;
import org.apache.spark.api.java.JavaPairRDD;
import org.apache.spark.api.java.function.PairFlatMapFunction;
import org.apache.spark.api.java.function.PairFunction;
import scala.Tuple2;
import org.apache.sysml.runtime.DMLRuntimeException;
import org.apache.sysml.runtime.DMLUnsupportedOperationException;
import org.apache.sysml.runtime.controlprogram.context.ExecutionContext;
import org.apache.sysml.runtime.controlprogram.context.SparkExecutionContext;
import org.apache.sysml.runtime.functionobjects.OffsetColumnIndex;
import org.apache.sysml.runtime.instructions.InstructionUtils;
import org.apache.sysml.runtime.instructions.cp.CPOperand;
import org.apache.sysml.runtime.matrix.MatrixCharacteristics;
import org.apache.sysml.runtime.matrix.data.MatrixBlock;
import org.apache.sysml.runtime.matrix.data.MatrixIndexes;
import org.apache.sysml.runtime.matrix.operators.Operator;
import org.apache.sysml.runtime.matrix.operators.ReorgOperator;
import org.apache.sysml.runtime.util.UtilFunctions;
public class AppendGSPInstruction extends BinarySPInstruction
{
private boolean _cbind = true;
public AppendGSPInstruction(Operator op, CPOperand in1, CPOperand in2, CPOperand offset, CPOperand offset2, CPOperand out, boolean cbind, String opcode, String istr)
{
super(op, in1, in2, out, opcode, istr);
_sptype = SPINSTRUCTION_TYPE.GAppend;
_cbind = cbind;
}
public static AppendGSPInstruction parseInstruction ( String str )
throws DMLRuntimeException
{
String[] parts = InstructionUtils.getInstructionPartsWithValueType(str);
InstructionUtils.checkNumFields(parts, 6);
String opcode = parts[0];
CPOperand in1 = new CPOperand(parts[1]);
CPOperand in2 = new CPOperand(parts[2]);
CPOperand in3 = new CPOperand(parts[3]);
CPOperand in4 = new CPOperand(parts[4]);
CPOperand out = new CPOperand(parts[5]);
boolean cbind = Boolean.parseBoolean(parts[6]);
if(!opcode.equalsIgnoreCase("gappend"))
throw new DMLRuntimeException("Unknown opcode while parsing a AppendGSPInstruction: " + str);
return new AppendGSPInstruction(
new ReorgOperator(OffsetColumnIndex.getOffsetColumnIndexFnObject(-1)),
in1, in2, in3, in4, out, cbind, opcode, str);
}
@Override
public void processInstruction(ExecutionContext ec)
throws DMLUnsupportedOperationException, DMLRuntimeException
{
// general case append (map-extend, aggregate)
SparkExecutionContext sec = (SparkExecutionContext)ec;
checkBinaryAppendInputCharacteristics(sec, _cbind, false, false);
MatrixCharacteristics mc1 = sec.getMatrixCharacteristics(input1.getName());
MatrixCharacteristics mc2 = sec.getMatrixCharacteristics(input2.getName());
JavaPairRDD in1 = sec.getBinaryBlockRDDHandleForVariable( input1.getName() );
JavaPairRDD in2 = sec.getBinaryBlockRDDHandleForVariable( input2.getName() );
JavaPairRDD out = null;
// General case: This one needs shifting and merging and hence has huge performance hit.
JavaPairRDD shifted_in2 = in2
.flatMapToPair(new ShiftMatrix(mc1, mc2, _cbind));
out = in1.cogroup(shifted_in2)
.mapToPair(new MergeWithShiftedBlocks(mc1, mc2, _cbind));
//put output RDD handle into symbol table
updateBinaryAppendOutputMatrixCharacteristics(sec, _cbind);
sec.setRDDHandleForVariable(output.getName(), out);
sec.addLineageRDD(output.getName(), input1.getName());
sec.addLineageRDD(output.getName(), input2.getName());
}
/**
*
*/
public static class MergeWithShiftedBlocks implements PairFunction,Iterable>>, MatrixIndexes, MatrixBlock>
{
private static final long serialVersionUID = 848955582909209400L;
private boolean _cbind;
private long _lastIxLeft;
private int _blen;
public MergeWithShiftedBlocks(MatrixCharacteristics mc1, MatrixCharacteristics mc2, boolean cbind)
{
_cbind = cbind;
_blen = cbind ? mc1.getColsPerBlock() : mc1.getRowsPerBlock();
_lastIxLeft = (long) Math.ceil((double)(cbind ? mc1.getCols():mc1.getRows()) / _blen);
}
@Override
public Tuple2 call(Tuple2, Iterable>> kv)
throws Exception
{
Iterator iterLeft = kv._2._1.iterator();
Iterator iterRight = kv._2._2.iterator();
//handle single left/right block input
if( !iterLeft.hasNext() ) {
MatrixBlock tmp = iterRight.next();
if( iterRight.hasNext() )
tmp.merge(iterRight.next(), false);
return new Tuple2(kv._1, tmp);
}
else if ( !iterRight.hasNext() ) {
return new Tuple2(kv._1, iterLeft.next());
}
MatrixBlock firstBlk = iterLeft.next();
MatrixBlock secondBlk = iterRight.next();
long ix = _cbind ? kv._1.getColumnIndex() : kv._1.getRowIndex();
// Since merge requires the dimensions matching
if( ix == _lastIxLeft && (_cbind && firstBlk.getNumColumns() < secondBlk.getNumColumns()
|| !_cbind && firstBlk.getNumRows()(kv._1, firstBlk);
}
}
/**
*
*/
public static class ShiftMatrix implements PairFlatMapFunction, MatrixIndexes,MatrixBlock>
{
private static final long serialVersionUID = 3524189212798209172L;
private boolean _cbind;
private long _startIx;
private int _shiftBy;
private int _blen;
private long _outlen;
public ShiftMatrix(MatrixCharacteristics mc1, MatrixCharacteristics mc2, boolean cbind)
{
_cbind = cbind;
_startIx = cbind ? UtilFunctions.computeBlockIndex(mc1.getCols(), mc1.getColsPerBlock()) :
UtilFunctions.computeBlockIndex(mc1.getRows(), mc1.getRowsPerBlock());
_blen = (int) (cbind ? mc1.getColsPerBlock() : mc1.getRowsPerBlock());
_shiftBy = (int) (cbind ? mc1.getCols()%_blen : mc1.getRows()%_blen);
_outlen = cbind ? mc1.getCols()+mc2.getCols() : mc1.getRows()+mc2.getRows();
}
@Override
public Iterable> call(Tuple2 kv)
throws Exception
{
//common preparation
ArrayList> retVal = new ArrayList>();
MatrixIndexes ix = kv._1();
MatrixBlock in = kv._2();
int cutAt = _blen - _shiftBy;
if( _cbind )
{
MatrixIndexes firstIndex = new MatrixIndexes(ix.getRowIndex(), ix.getColumnIndex()+_startIx-1);
MatrixIndexes secondIndex = new MatrixIndexes(ix.getRowIndex(), ix.getColumnIndex()+_startIx);
int lblen1 = UtilFunctions.computeBlockSize(_outlen, firstIndex.getColumnIndex(), _blen);
if(cutAt >= in.getNumColumns()) {
// The block is too small to be cut
MatrixBlock firstBlk = new MatrixBlock(in.getNumRows(), lblen1, true);
firstBlk = firstBlk.leftIndexingOperations(in, 0, in.getNumRows()-1, lblen1-in.getNumColumns(), lblen1-1, new MatrixBlock(), true);
retVal.add(new Tuple2(firstIndex, firstBlk));
}
else {
// Since merge requires the dimensions matching, shifting = slicing + left indexing
MatrixBlock firstSlicedBlk = in.sliceOperations(0, in.getNumRows()-1, 0, cutAt-1, new MatrixBlock());
MatrixBlock firstBlk = new MatrixBlock(in.getNumRows(), lblen1, true);
firstBlk = firstBlk.leftIndexingOperations(firstSlicedBlk, 0, in.getNumRows()-1, _shiftBy, _blen-1, new MatrixBlock(), true);
MatrixBlock secondSlicedBlk = in.sliceOperations(0, in.getNumRows()-1, cutAt, in.getNumColumns()-1, new MatrixBlock());
int llen2 = UtilFunctions.computeBlockSize(_outlen, secondIndex.getColumnIndex(), _blen);
MatrixBlock secondBlk = new MatrixBlock(in.getNumRows(), llen2, true);
secondBlk = secondBlk.leftIndexingOperations(secondSlicedBlk, 0, in.getNumRows()-1, 0, secondSlicedBlk.getNumColumns()-1, new MatrixBlock(), true);
retVal.add(new Tuple2(firstIndex, firstBlk));
retVal.add(new Tuple2(secondIndex, secondBlk));
}
}
else //rbind
{
MatrixIndexes firstIndex = new MatrixIndexes(ix.getRowIndex()+_startIx-1, ix.getColumnIndex());
MatrixIndexes secondIndex = new MatrixIndexes(ix.getRowIndex()+_startIx, ix.getColumnIndex());
int lblen1 = UtilFunctions.computeBlockSize(_outlen, firstIndex.getRowIndex(), _blen);
if(cutAt >= in.getNumRows()) {
// The block is too small to be cut
MatrixBlock firstBlk = new MatrixBlock(lblen1, in.getNumColumns(), true);
firstBlk = firstBlk.leftIndexingOperations(in, lblen1-in.getNumRows(), lblen1-1, 0, in.getNumColumns()-1, new MatrixBlock(), true);
retVal.add(new Tuple2(firstIndex, firstBlk));
}
else {
// Since merge requires the dimensions matching, shifting = slicing + left indexing
MatrixBlock firstSlicedBlk = in.sliceOperations(0, cutAt-1, 0, in.getNumColumns()-1, new MatrixBlock());
MatrixBlock firstBlk = new MatrixBlock(lblen1, in.getNumColumns(), true);
firstBlk = firstBlk.leftIndexingOperations(firstSlicedBlk, _shiftBy, _blen-1, 0, in.getNumColumns()-1, new MatrixBlock(), true);
MatrixBlock secondSlicedBlk = in.sliceOperations(cutAt, in.getNumRows()-1, 0, in.getNumColumns()-1, new MatrixBlock());
int lblen2 = UtilFunctions.computeBlockSize(_outlen, secondIndex.getRowIndex(), _blen);
MatrixBlock secondBlk = new MatrixBlock(lblen2, in.getNumColumns(), true);
secondBlk = secondBlk.leftIndexingOperations(secondSlicedBlk, 0, secondSlicedBlk.getNumRows()-1, 0, in.getNumColumns()-1, new MatrixBlock(), true);
retVal.add(new Tuple2(firstIndex, firstBlk));
retVal.add(new Tuple2(secondIndex, secondBlk));
}
}
return retVal;
}
}
}