Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Original file line number Diff line number Diff line change
Expand Up @@ -63,6 +63,8 @@ private COV() {
public Data execute(Data in1, double u, double v, double w2)
{
CmCovObject cov1=(CmCovObject) in1;
if(w2 == 0)
return cov1;
Comment on lines +66 to +67

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why is this change necessary?

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

its a reliability case where during the weighted mean computation, if its a sparse matrix, we run into the divide by zero error but a zero weight has no impact on the original aggregation so I just returned the original covariance itself.

if(cov1.isCOVAllZeros())
{
cov1.w=w2;
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -29,9 +29,10 @@
import org.apache.sysds.runtime.instructions.ooc.BinaryOOCInstruction;
import org.apache.sysds.runtime.instructions.ooc.CSVReblockOOCInstruction;
import org.apache.sysds.runtime.instructions.ooc.CentralMomentOOCInstruction;
import org.apache.sysds.runtime.instructions.ooc.CovarianceOOCInstruction;
import org.apache.sysds.runtime.instructions.ooc.CtableOOCInstruction;
import org.apache.sysds.runtime.instructions.ooc.IndexingOOCInstruction;
import org.apache.sysds.runtime.instructions.ooc.DataGenOOCInstruction;
import org.apache.sysds.runtime.instructions.ooc.IndexingOOCInstruction;
import org.apache.sysds.runtime.instructions.ooc.OOCInstruction;
import org.apache.sysds.runtime.instructions.ooc.ParameterizedBuiltinOOCInstruction;
import org.apache.sysds.runtime.instructions.ooc.ReblockOOCInstruction;
Expand Down Expand Up @@ -103,6 +104,8 @@ else if(parts.length == 4)
return TeeOOCInstruction.parseInstruction(str);
case CentralMoment:
return CentralMomentOOCInstruction.parseInstruction(str);
case Covariance:
return CovarianceOOCInstruction.parseInstruction(str);
case Ctable:
return CtableOOCInstruction.parseInstruction(str);
case ParameterizedBuiltin:
Expand Down
Original file line number Diff line number Diff line change
@@ -0,0 +1,123 @@
/*
* 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.sysds.runtime.instructions.ooc;

import java.util.List;

import org.apache.sysds.common.Opcodes;
import org.apache.sysds.runtime.DMLRuntimeException;
import org.apache.sysds.runtime.controlprogram.caching.MatrixObject;
import org.apache.sysds.runtime.controlprogram.context.ExecutionContext;
import org.apache.sysds.runtime.controlprogram.parfor.LocalTaskQueue;
import org.apache.sysds.runtime.functionobjects.COV;
import org.apache.sysds.runtime.instructions.InstructionUtils;
import org.apache.sysds.runtime.instructions.cp.CPOperand;
import org.apache.sysds.runtime.instructions.cp.CmCovObject;
import org.apache.sysds.runtime.instructions.cp.DoubleObject;
import org.apache.sysds.runtime.instructions.spark.data.IndexedMatrixValue;
import org.apache.sysds.runtime.matrix.data.MatrixBlock;
import org.apache.sysds.runtime.matrix.operators.COVOperator;
import org.apache.sysds.runtime.meta.DataCharacteristics;

public class CovarianceOOCInstruction extends ComputationOOCInstruction {

private CovarianceOOCInstruction(COVOperator cov, CPOperand in1, CPOperand in2, CPOperand in3, CPOperand out,
String opcode, String str) {
super(OOCType.COV, cov, in1, in2, in3, out, opcode, str);
}

public static CovarianceOOCInstruction parseInstruction(String str) {
String[] parts = InstructionUtils.getInstructionPartsWithValueType(str);
String opcode = parts[0];

if(!opcode.equalsIgnoreCase(Opcodes.COV.toString()))
throw new DMLRuntimeException("CovarianceOOCInstruction.parseInstruction():: Unknown opcode " + opcode);

// the OOC instruction string matches the Spark format,

COVOperator cov = new COVOperator(COV.getCOMFnObject());
if(parts.length == 4) { // this is the case for unweighted cov.A.B.out
CPOperand in1 = new CPOperand(parts[1]);
CPOperand in2 = new CPOperand(parts[2]);
CPOperand out = new CPOperand(parts[3]);
return new CovarianceOOCInstruction(cov, in1, in2, null, out, opcode, str);
}
else if(parts.length == 5) {// this is the case for weighted cov.A.B.W.out
CPOperand in1 = new CPOperand(parts[1]);
CPOperand in2 = new CPOperand(parts[2]);
CPOperand in3 = new CPOperand(parts[3]);
CPOperand out = new CPOperand(parts[4]);
return new CovarianceOOCInstruction(cov, in1, in2, in3, out, opcode, str);
}
else {
throw new DMLRuntimeException("Invalid number of arguments in Instruction: " + str);
}
}

@Override
public void processInstruction(ExecutionContext ec) {
COVOperator cov_op = (COVOperator) _optr;

MatrixObject mo1 = ec.getMatrixObject(input1.getName());
MatrixObject mo2 = ec.getMatrixObject(input2.getName());

OOCStream<IndexedMatrixValue> q1 = mo1.getStreamHandle();
OOCStream<IndexedMatrixValue> q2 = mo2.getStreamHandle();

OOCStream<CmCovObject> covObjs = createWritableStream();

if(input3 == null) {
// unweighted covariance join the two tile streams by block index
joinOOC(q1, q2, covObjs,
(a, b) -> ((MatrixBlock) a.getValue()).covOperations(cov_op, (MatrixBlock) b.getValue()),
IndexedMatrixValue::getIndexes);
}
else {
// weighted covariance additionally join the weights tile stream
MatrixObject mo3 = ec.getMatrixObject(input3.getName());

DataCharacteristics dc1 = ec.getDataCharacteristics(input1.getName());
DataCharacteristics dc2 = ec.getDataCharacteristics(input2.getName());
DataCharacteristics dc3 = ec.getDataCharacteristics(input3.getName());
if(dc1.getBlocksize() != dc2.getBlocksize() || dc1.getBlocksize() != dc3.getBlocksize())
throw new DMLRuntimeException("Different block sizes are not yet supported");

OOCStream<IndexedMatrixValue> q3 = mo3.getStreamHandle();

joinOOC(List.of(q1, q2, q3), covObjs,
tiles -> ((MatrixBlock) tiles.get(0).getValue()).covOperations(cov_op,
(MatrixBlock) tiles.get(1).getValue(), (MatrixBlock) tiles.get(2).getValue()),
IndexedMatrixValue::getIndexes);
}

try {
CmCovObject agg = covObjs.dequeue();
CmCovObject next;

while((next = covObjs.dequeue()) != LocalTaskQueue.NO_MORE_TASKS)
agg = (CmCovObject) cov_op.fn.execute(agg, next);

ec.setScalarOutput(output.getName(), new DoubleObject(agg.getRequiredResult(cov_op)));
}
catch(Exception ex) {
throw new DMLRuntimeException(ex);
}
}
}
Original file line number Diff line number Diff line change
Expand Up @@ -80,7 +80,7 @@ public abstract class OOCInstruction extends Instruction {

public enum OOCType {
Reblock, Tee, Binary, Ternary, Unary, AggregateUnary, AggregateBinary, AggregateTernary, MAPMM, MMTSJ,
MAPMMCHAIN, Reorg, CM, Ctable, MatrixIndexing, ParameterizedBuiltin, Rand, Append, Quaternary, Reshape
MAPMMCHAIN, Reorg, CM, COV, Ctable, MatrixIndexing, ParameterizedBuiltin, Rand, Append, Quaternary, Reshape
}

protected final OOCInstruction.OOCType _ooctype;
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -19,18 +19,23 @@

package org.apache.sysds.runtime.instructions.ooc;

import java.util.List;
import java.util.concurrent.CompletableFuture;

import org.apache.sysds.common.Opcodes;
import org.apache.sysds.lops.MMTSJ;
import org.apache.sysds.lops.MMTSJ.MMTSJType;
import org.apache.sysds.runtime.DMLRuntimeException;
import org.apache.sysds.runtime.controlprogram.caching.MatrixObject;
import org.apache.sysds.runtime.controlprogram.context.ExecutionContext;
import org.apache.sysds.runtime.controlprogram.parfor.LocalTaskQueue;
import org.apache.sysds.runtime.functionobjects.Multiply;
import org.apache.sysds.runtime.functionobjects.Plus;
import org.apache.sysds.runtime.instructions.InstructionUtils;
import org.apache.sysds.runtime.instructions.cp.CPOperand;
import org.apache.sysds.runtime.instructions.spark.data.IndexedMatrixValue;
import org.apache.sysds.runtime.matrix.data.LibMatrixReorg;
import org.apache.sysds.runtime.matrix.data.MatrixBlock;
import org.apache.sysds.runtime.matrix.data.MatrixIndexes;
import org.apache.sysds.runtime.matrix.operators.AggregateBinaryOperator;
import org.apache.sysds.runtime.matrix.operators.AggregateOperator;
import org.apache.sysds.runtime.matrix.operators.BinaryOperator;
Expand All @@ -48,7 +53,7 @@ public static TSMMOOCInstruction parseInstruction(String str) {
String[] parts = InstructionUtils.getInstructionPartsWithValueType(str);
InstructionUtils.checkNumFields(parts, 3);
String opcode = parts[0];
CPOperand in1 = new CPOperand(parts[1]); // the large matrix (streamed), columns <= blocksize
CPOperand in1 = new CPOperand(parts[1]); // the large matrix (streamed)
CPOperand out = new CPOperand(parts[2]);
MMTSJ.MMTSJType mmtsjType = MMTSJ.MMTSJType.valueOf(parts[3]);

Expand All @@ -59,39 +64,90 @@ public static TSMMOOCInstruction parseInstruction(String str) {
}

@Override
public void processInstruction( ExecutionContext ec ) {
public void processInstruction(ExecutionContext ec) {
MatrixObject min = ec.getMatrixObject(input1);
int nRows = (int) min.getDataCharacteristics().getRows();
int nCols = (int) min.getDataCharacteristics().getCols();
int bLen = min.getDataCharacteristics().getBlocksize();

OOCStream<IndexedMatrixValue> qIn = min.getStreamHandle();
int numRowBlocks = Math.toIntExact(min.getDataCharacteristics().getNumRowBlocks());
int numColBlocks = Math.toIntExact(min.getDataCharacteristics().getNumColBlocks());
int blocksPerJoinGroup = _type.isLeft() ? numColBlocks : numRowBlocks;
int partialsPerOutput = _type.isLeft() ? numRowBlocks : numColBlocks;

OOCStreamable<IndexedMatrixValue> inputStreamable = min.getStreamable();
final boolean createdCache = !inputStreamable.hasStreamCache();
final CachingStream inputCache = createdCache ? new CachingStream(min.getStreamHandle())
: inputStreamable.getStreamCache();

OOCStream<List<IndexedMatrixValue>> groupedPartials = createWritableStream();
OOCStream<IndexedMatrixValue> partials = createWritableStream();
OOCStream<IndexedMatrixValue> out = createWritableStream();
addOutStream(out);
ec.getMatrixObject(output).setStreamHandle(out);

CompletableFuture<Void> joinFuture = joinManyOOC(inputCache.getReadStream(), inputCache.getReadStream(), groupedPartials,
this::createPartialOutputTiles, this::getJoinIndex, this::getJoinIndex,
blocksPerJoinGroup, blocksPerJoinGroup);
CompletableFuture<Void> expandFuture = expandOOC(groupedPartials, partials, values -> values);

BinaryOperator plus = InstructionUtils.parseBinaryOperator(Opcodes.PLUS.toString());
CompletableFuture<Void> outFuture = groupedReduceOOC(partials, out, (left, right) -> {
MatrixBlock result = ((MatrixBlock) left.getValue()).binaryOperations(plus, right.getValue());
left.setValue(result);
return left;
}, partialsPerOutput);

propagateFailuresToOutput(out, List.of(joinFuture, expandFuture, outFuture));

outFuture.whenComplete((result, error) -> {
if(createdCache)
inputCache.scheduleDeletion();
});
}

private long getJoinIndex(IndexedMatrixValue value) {
return _type.isLeft() ? value.getIndexes().getRowIndex() : value.getIndexes().getColumnIndex();
}

//validation check TODO extend compiler to not create OOC otherwise
if( (_type.isLeft() && nCols > bLen)
|| (_type.isRight() && nRows > bLen) )
{
throw new UnsupportedOperationException();
private long getOutputIndex(IndexedMatrixValue value) {
return _type.isLeft() ? value.getIndexes().getColumnIndex() : value.getIndexes().getRowIndex();
}

private List<IndexedMatrixValue> createPartialOutputTiles(IndexedMatrixValue left, IndexedMatrixValue right) {
long leftIndex = getOutputIndex(left);
long rightIndex = getOutputIndex(right);
if(leftIndex > rightIndex)
return List.of();

MatrixBlock leftBlock = (MatrixBlock) left.getValue();
MatrixBlock rightBlock = (MatrixBlock) right.getValue();
if(leftIndex == rightIndex) {
MatrixBlock diagonal = leftBlock.transposeSelfMatrixMultOperations(new MatrixBlock(), _type);
return List.of(new IndexedMatrixValue(new MatrixIndexes(leftIndex, rightIndex), diagonal));
}

//int dim = _type.isLeft() ? nCols : nRows;
MatrixBlock resultBlock = null;

OOCStream<MatrixBlock> tmpStream = createWritableStream();

mapOOC(qIn, tmpStream,
tmp -> ((MatrixBlock) tmp.getValue())
.transposeSelfMatrixMultOperations(new MatrixBlock(), _type));

MatrixBlock tmp;
while ((tmp = tmpStream.dequeue()) != LocalTaskQueue.NO_MORE_TASKS) {
if (resultBlock == null)
resultBlock = tmp;
else
resultBlock.binaryOperationsInPlace(plus, tmp);
Comment on lines -62 to -92

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

These changes look good for the general case. However, the more general execution path is more expensive. I'd like you to keep the special case where we can avoid creating a CachingStream and using the heavier primitives (when only a single output tile is produced).


MatrixBlock partial = multiplyOffDiagonal(leftBlock, rightBlock);
MatrixBlock mirror = LibMatrixReorg.transpose(partial);
return List.of(
new IndexedMatrixValue(new MatrixIndexes(leftIndex, rightIndex), partial),
new IndexedMatrixValue(new MatrixIndexes(rightIndex, leftIndex), mirror));
}

private MatrixBlock multiplyOffDiagonal(MatrixBlock leftBlock, MatrixBlock rightBlock) {
if(_type.isLeft()) {
MatrixBlock leftTranspose = LibMatrixReorg.transpose(leftBlock);
return leftTranspose.aggregateBinaryOperations(leftTranspose, rightBlock, new MatrixBlock(),
(AggregateBinaryOperator) _optr);
}

ec.setMatrixOutput(output.getName(), resultBlock);
MatrixBlock rightTranspose = LibMatrixReorg.transpose(rightBlock);
return leftBlock.aggregateBinaryOperations(leftBlock, rightTranspose, new MatrixBlock(),
(AggregateBinaryOperator) _optr);
}

private static void propagateFailuresToOutput(OOCStream<?> out, List<CompletableFuture<Void>> futures) {
for(CompletableFuture<Void> future : futures) {
future.exceptionally(error -> {
out.propagateFailure(DMLRuntimeException.of(error));
return null;
});
}
}
}
Loading
Loading