@Override
protected ExecType optFindExecType() throws HopsException {
checkAndSetForcedPlatform();
ExecType REMOTE = OptimizerUtils.isSparkExecutionMode() ? ExecType.SPARK : ExecType.MR;
if (_etypeForced != null) {
_etype = _etypeForced;
} else {
if (OptimizerUtils.isMemoryBasedOptLevel()) {
_etype = findExecTypeByMemEstimate();
} else if (getInput().get(0).areDimsBelowThreshold()) {
_etype = ExecType.CP;
} else {
_etype = REMOTE;
}
// check for valid CP dimensions and matrix size
checkAndSetInvalidCPDimsAndSize();
}
// mark for recompile (forever)
if (OptimizerUtils.ALLOW_DYN_RECOMPILATION && !dimsKnown(true) && _etype == REMOTE)
setRequiresRecompile();
return _etype;
}
@Override
protected ExecType optFindExecType() throws HopsException {
checkAndSetForcedPlatform();
ExecType REMOTE = OptimizerUtils.isSparkExecutionMode() ? ExecType.SPARK : ExecType.MR;
if (_etypeForced != null) {
_etype = _etypeForced;
} else {
if (OptimizerUtils.isMemoryBasedOptLevel()) {
_etype = findExecTypeByMemEstimate();
}
// Choose CP, if the input dimensions are below threshold or if the input is a vector
// Also, matrix inverse is currently implemented only in CP (through commons math)
else if (getInput().get(0).areDimsBelowThreshold()
|| getInput().get(0).isVector()
|| isInMemoryOperation()) {
_etype = ExecType.CP;
} else {
_etype = REMOTE;
}
// check for valid CP dimensions and matrix size
checkAndSetInvalidCPDimsAndSize();
}
// mark for recompile (forever)
if (OptimizerUtils.ALLOW_DYN_RECOMPILATION && !dimsKnown(true) && _etype == REMOTE)
setRequiresRecompile();
if (_op == OpOp1.PRINT || _op == OpOp1.STOP || _op == OpOp1.INVERSE) _etype = ExecType.CP;
return _etype;
}
@Override
protected ExecType optFindExecType() throws HopsException {
checkAndSetForcedPlatform();
ExecType REMOTE = OptimizerUtils.isSparkExecutionMode() ? ExecType.SPARK : ExecType.MR;
if (_etypeForced != null) {
_etype = _etypeForced;
} else {
if (OptimizerUtils.isMemoryBasedOptLevel()) {
_etype = findExecTypeByMemEstimate();
}
// Choose CP, if the input dimensions are below threshold or if the input is a vector
// Also, matrix inverse is currently implemented only in CP (through commons math)
else if (getInput().get(0).areDimsBelowThreshold()
|| getInput().get(0).isVector()
|| isInMemoryOperation()) {
_etype = ExecType.CP;
} else {
_etype = REMOTE;
}
// check for valid CP dimensions and matrix size
checkAndSetInvalidCPDimsAndSize();
}
// spark-specific decision refinement (execute unary w/ spark input and
// single parent also in spark because it's likely cheap and reduces intermediates)
if (_etype == ExecType.CP
&& _etypeForced != ExecType.CP
&& getInput().get(0).optFindExecType() == ExecType.SPARK
&& getDataType().isMatrix()
&& !isCumulativeUnaryOperation()
&& !isCastUnaryOperation()
&& _op != OpOp1.MEDIAN
&& _op != OpOp1.IQM
&& !(getInput().get(0) instanceof DataOp) // input is not checkpoint
&& getInput().get(0).getParent().size() == 1) // unary is only parent
{
// pull unary operation into spark
_etype = ExecType.SPARK;
}
// mark for recompile (forever)
if (ConfigurationManager.isDynamicRecompilation() && !dimsKnown(true) && _etype == REMOTE)
setRequiresRecompile();
// ensure cp exec type for single-node operations
if (_op == OpOp1.PRINT
|| _op == OpOp1.STOP
|| _op == OpOp1.INVERSE
|| _op == OpOp1.EIGEN
|| _op == OpOp1.CHOLESKY) {
_etype = ExecType.CP;
}
return _etype;
}
@Override
protected double computeIntermediateMemEstimate(long dim1, long dim2, long nnz) {
// default: no additional memory required
double val = 0;
double sparsity = OptimizerUtils.getSparsity(dim1, dim2, nnz);
switch (_op) // see MatrixAggLib for runtime operations
{
case MAX:
case MIN:
// worst-case: column-wise, sparse (temp int count arrays)
if (_direction == Direction.Col) val = dim2 * OptimizerUtils.INT_SIZE;
break;
case SUM:
case SUM_SQ:
// worst-case correction LASTROW / LASTCOLUMN
if (_direction == Direction.Col) // (potentially sparse)
val = OptimizerUtils.estimateSizeExactSparsity(1, dim2, sparsity);
else if (_direction == Direction.Row) // (always dense)
val = OptimizerUtils.estimateSizeExactSparsity(dim1, 1, 1.0);
break;
case MEAN:
// worst-case correction LASTTWOROWS / LASTTWOCOLUMNS
if (_direction == Direction.Col) // (potentially sparse)
val = OptimizerUtils.estimateSizeExactSparsity(2, dim2, sparsity);
else if (_direction == Direction.Row) // (always dense)
val = OptimizerUtils.estimateSizeExactSparsity(dim1, 2, 1.0);
break;
case VAR:
// worst-case correction LASTFOURROWS / LASTFOURCOLUMNS
if (_direction == Direction.Col) // (potentially sparse)
val = OptimizerUtils.estimateSizeExactSparsity(4, dim2, sparsity);
else if (_direction == Direction.Row) // (always dense)
val = OptimizerUtils.estimateSizeExactSparsity(dim1, 4, 1.0);
break;
case MAXINDEX:
case MININDEX:
Hop hop = getInput().get(0);
if (isUnaryAggregateOuterCPRewriteApplicable())
val = 3 * OptimizerUtils.estimateSizeExactSparsity(1, hop._dim2, 1.0);
else
// worst-case correction LASTCOLUMN
val = OptimizerUtils.estimateSizeExactSparsity(dim1, 1, 1.0);
break;
default:
// no intermediate memory consumption
val = 0;
}
return val;
}
@Override
protected ExecType optFindExecType() throws HopsException {
checkAndSetForcedPlatform();
ExecType REMOTE = OptimizerUtils.isSparkExecutionMode() ? ExecType.SPARK : ExecType.MR;
// forced / memory-based / threshold-based decision
if (_etypeForced != null) {
_etype = _etypeForced;
} else {
if (OptimizerUtils.isMemoryBasedOptLevel()) {
_etype = findExecTypeByMemEstimate();
}
// Choose CP, if the input dimensions are below threshold or if the input is a vector
else if (getInput().get(0).areDimsBelowThreshold() || getInput().get(0).isVector()) {
_etype = ExecType.CP;
} else {
_etype = REMOTE;
}
// check for valid CP dimensions and matrix size
checkAndSetInvalidCPDimsAndSize();
}
// spark-specific decision refinement (execute unary aggregate w/ spark input and
// single parent also in spark because it's likely cheap and reduces data transfer)
if (_etype == ExecType.CP
&& _etypeForced != ExecType.CP
&& !(getInput().get(0) instanceof DataOp) // input is not checkpoint
&& getInput().get(0).getParent().size() == 1 // uagg is only parent
&& getInput().get(0).optFindExecType() == ExecType.SPARK) {
// pull unary aggregate into spark
_etype = ExecType.SPARK;
}
// mark for recompile (forever)
if (ConfigurationManager.isDynamicRecompilation() && !dimsKnown(true) && _etype == REMOTE) {
setRequiresRecompile();
}
return _etype;
}
private boolean isUnaryAggregateOuterRewriteApplicable() {
boolean ret = false;
Hop input = getInput().get(0);
if (input instanceof BinaryOp && ((BinaryOp) input).isOuterVectorOperator()) {
// for special cases, we need to hold the broadcast twice in order to allow for
// an efficient binary search over a plain java array
double factor =
(isCompareOperator(((BinaryOp) input).getOp())
&& (_direction == Direction.Row
|| _direction == Direction.Col
|| _direction == Direction.RowCol)
&& (_op == AggOp.SUM))
? 2.0
: 1.0;
factor +=
(isCompareOperator(((BinaryOp) input).getOp())
&& (_direction == Direction.Row || _direction == Direction.Col)
&& (_op == AggOp.MAXINDEX || _op == AggOp.MININDEX))
? 1.0
: 0.0;
// note: memory constraint only needs to take the rhs into account because the output
// is guaranteed to be an aggregate of <=16KB
Hop right = input.getInput().get(1);
if ((right.dimsKnown()
&& factor * OptimizerUtils.estimateSize(right.getDim1(), right.getDim2())
< OptimizerUtils.getRemoteMemBudgetMap(true)) // dims known and estimate fits
|| (!right.dimsKnown()
&& factor * right.getOutputMemEstimate()
< OptimizerUtils.getRemoteMemBudgetMap(
true))) // dims unknown but worst-case estimate fits
{
ret = true;
}
}
return ret;
}
/**
* This will check if there is sufficient memory locally (twice the size of second matrix, for
* original and sort data), and remotely (size of second matrix (sorted data)).
*
* @return true if sufficient memory
*/
private boolean isUnaryAggregateOuterSPRewriteApplicable() {
boolean ret = false;
Hop input = getInput().get(0);
if (input instanceof BinaryOp && ((BinaryOp) input).isOuterVectorOperator()) {
// note: both cases (partitioned matrix, and sorted double array), require to
// fit the broadcast twice into the local memory budget. Also, the memory
// constraint only needs to take the rhs into account because the output is
// guaranteed to be an aggregate of <=16KB
Hop right = input.getInput().get(1);
double size =
right.dimsKnown()
? OptimizerUtils.estimateSize(right.getDim1(), right.getDim2())
: // dims known and estimate fits
right.getOutputMemEstimate(); // dims unknown but worst-case estimate fits
if (_op == AggOp.MAXINDEX || _op == AggOp.MININDEX) {
double memBudgetExec = SparkExecutionContext.getBroadcastMemoryBudget();
double memBudgetLocal = OptimizerUtils.getLocalMemBudget();
// basic requirement: the broadcast needs to to fit twice in the remote broadcast memory
// and local memory budget because we have to create a partitioned broadcast
// memory and hand it over to the spark context as in-memory object
ret = (2 * size < memBudgetExec && 2 * size < memBudgetLocal);
} else {
if (OptimizerUtils.checkSparkBroadcastMemoryBudget(size)) {
ret = true;
}
}
}
return ret;
}
private Lop constructLopsTernaryAggregateRewrite(ExecType et)
throws HopsException, LopsException {
Hop input1 = getInput().get(0);
Hop input11 = input1.getInput().get(0);
Hop input12 = input1.getInput().get(1);
Lop ret = null;
Lop in1 = null;
Lop in2 = null;
Lop in3 = null;
if (input11 instanceof BinaryOp && ((BinaryOp) input11).getOp() == OpOp2.MULT) {
in1 = input11.getInput().get(0).constructLops();
in2 = input11.getInput().get(1).constructLops();
in3 = input12.constructLops();
} else if (input12 instanceof BinaryOp && ((BinaryOp) input12).getOp() == OpOp2.MULT) {
in1 = input11.constructLops();
in2 = input12.getInput().get(0).constructLops();
in3 = input12.getInput().get(1).constructLops();
} else {
in1 = input11.constructLops();
in2 = input12.constructLops();
in3 = new LiteralOp(1).constructLops();
}
// create new ternary aggregate operator
int k = OptimizerUtils.getConstrainedNumThreads(_maxNumThreads);
// The execution type of a unary aggregate instruction should depend on the execution type of
// inputs to avoid OOM
// Since we only support matrix-vector and not vector-matrix, checking the execution type of
// input1 should suffice.
ExecType et_input = input1.optFindExecType();
ret =
new TernaryAggregate(
in1,
in2,
in3,
Aggregate.OperationTypes.KahanSum,
Binary.OperationTypes.MULTIPLY,
DataType.SCALAR,
ValueType.DOUBLE,
et_input,
k);
return ret;
}
@Override
public Lop constructLops() throws HopsException, LopsException {
// return already created lops
if (getLops() != null) return getLops();
try {
ExecType et = optFindExecType();
Hop input = getInput().get(0);
if (et == ExecType.CP) {
Lop agg1 = null;
if (isTernaryAggregateRewriteApplicable()) {
agg1 = constructLopsTernaryAggregateRewrite(et);
} else if (isUnaryAggregateOuterCPRewriteApplicable()) {
OperationTypes op = HopsAgg2Lops.get(_op);
DirectionTypes dir = HopsDirection2Lops.get(_direction);
BinaryOp binput = (BinaryOp) getInput().get(0);
agg1 =
new UAggOuterChain(
binput.getInput().get(0).constructLops(),
binput.getInput().get(1).constructLops(),
op,
dir,
HopsOpOp2LopsB.get(binput.getOp()),
DataType.MATRIX,
getValueType(),
ExecType.CP);
PartialAggregate.setDimensionsBasedOnDirection(
agg1, getDim1(), getDim2(), input.getRowsInBlock(), input.getColsInBlock(), dir);
if (getDataType() == DataType.SCALAR) {
UnaryCP unary1 =
new UnaryCP(
agg1, HopsOpOp1LopsUS.get(OpOp1.CAST_AS_SCALAR), getDataType(), getValueType());
unary1.getOutputParameters().setDimensions(0, 0, 0, 0, -1);
setLineNumbers(unary1);
setLops(unary1);
}
} else { // general case
int k = OptimizerUtils.getConstrainedNumThreads(_maxNumThreads);
if (DMLScript.USE_ACCELERATOR
&& (DMLScript.FORCE_ACCELERATOR
|| getMemEstimate() < OptimizerUtils.GPU_MEMORY_BUDGET)
&& (_op == AggOp.SUM)) {
et = ExecType.GPU;
k = 1;
}
agg1 =
new PartialAggregate(
input.constructLops(),
HopsAgg2Lops.get(_op),
HopsDirection2Lops.get(_direction),
getDataType(),
getValueType(),
et,
k);
}
setOutputDimensions(agg1);
setLineNumbers(agg1);
setLops(agg1);
if (getDataType() == DataType.SCALAR) {
agg1.getOutputParameters()
.setDimensions(1, 1, getRowsInBlock(), getColsInBlock(), getNnz());
}
} else if (et == ExecType.MR) {
OperationTypes op = HopsAgg2Lops.get(_op);
DirectionTypes dir = HopsDirection2Lops.get(_direction);
// unary aggregate operation
Lop transform1 = null;
if (isUnaryAggregateOuterRewriteApplicable()) {
BinaryOp binput = (BinaryOp) getInput().get(0);
transform1 =
new UAggOuterChain(
binput.getInput().get(0).constructLops(),
binput.getInput().get(1).constructLops(),
op,
dir,
HopsOpOp2LopsB.get(binput.getOp()),
DataType.MATRIX,
getValueType(),
ExecType.MR);
PartialAggregate.setDimensionsBasedOnDirection(
transform1,
getDim1(),
getDim2(),
input.getRowsInBlock(),
input.getColsInBlock(),
dir);
} else // default
{
transform1 =
new PartialAggregate(input.constructLops(), op, dir, DataType.MATRIX, getValueType());
((PartialAggregate) transform1)
.setDimensionsBasedOnDirection(
getDim1(), getDim2(), input.getRowsInBlock(), input.getColsInBlock());
}
setLineNumbers(transform1);
// aggregation if required
Lop aggregate = null;
Group group1 = null;
Aggregate agg1 = null;
if (requiresAggregation(input, _direction) || transform1 instanceof UAggOuterChain) {
group1 =
new Group(transform1, Group.OperationTypes.Sort, DataType.MATRIX, getValueType());
group1
.getOutputParameters()
.setDimensions(
getDim1(), getDim2(), input.getRowsInBlock(), input.getColsInBlock(), getNnz());
setLineNumbers(group1);
agg1 = new Aggregate(group1, HopsAgg2Lops.get(_op), DataType.MATRIX, getValueType(), et);
agg1.getOutputParameters()
.setDimensions(
getDim1(), getDim2(), input.getRowsInBlock(), input.getColsInBlock(), getNnz());
agg1.setupCorrectionLocation(PartialAggregate.getCorrectionLocation(op, dir));
setLineNumbers(agg1);
aggregate = agg1;
} else {
((PartialAggregate) transform1).setDropCorrection();
aggregate = transform1;
}
setLops(aggregate);
// cast if required
if (getDataType() == DataType.SCALAR) {
// Set the dimensions of PartialAggregate LOP based on the
// direction in which aggregation is performed
PartialAggregate.setDimensionsBasedOnDirection(
transform1,
input.getDim1(),
input.getDim2(),
input.getRowsInBlock(),
input.getColsInBlock(),
dir);
if (group1 != null && agg1 != null) { // if aggregation required
group1
.getOutputParameters()
.setDimensions(
input.getDim1(),
input.getDim2(),
input.getRowsInBlock(),
input.getColsInBlock(),
getNnz());
agg1.getOutputParameters()
.setDimensions(1, 1, input.getRowsInBlock(), input.getColsInBlock(), getNnz());
}
UnaryCP unary1 =
new UnaryCP(
aggregate,
HopsOpOp1LopsUS.get(OpOp1.CAST_AS_SCALAR),
getDataType(),
getValueType());
unary1.getOutputParameters().setDimensions(0, 0, 0, 0, -1);
setLineNumbers(unary1);
setLops(unary1);
}
} else if (et == ExecType.SPARK) {
OperationTypes op = HopsAgg2Lops.get(_op);
DirectionTypes dir = HopsDirection2Lops.get(_direction);
// unary aggregate
if (isTernaryAggregateRewriteApplicable()) {
Lop aggregate = constructLopsTernaryAggregateRewrite(et);
setOutputDimensions(aggregate); // 0x0 (scalar)
setLineNumbers(aggregate);
setLops(aggregate);
} else if (isUnaryAggregateOuterSPRewriteApplicable()) {
BinaryOp binput = (BinaryOp) getInput().get(0);
Lop transform1 =
new UAggOuterChain(
binput.getInput().get(0).constructLops(),
binput.getInput().get(1).constructLops(),
op,
dir,
HopsOpOp2LopsB.get(binput.getOp()),
DataType.MATRIX,
getValueType(),
ExecType.SPARK);
PartialAggregate.setDimensionsBasedOnDirection(
transform1,
getDim1(),
getDim2(),
input.getRowsInBlock(),
input.getColsInBlock(),
dir);
setLineNumbers(transform1);
setLops(transform1);
if (getDataType() == DataType.SCALAR) {
UnaryCP unary1 =
new UnaryCP(
transform1,
HopsOpOp1LopsUS.get(OpOp1.CAST_AS_SCALAR),
getDataType(),
getValueType());
unary1.getOutputParameters().setDimensions(0, 0, 0, 0, -1);
setLineNumbers(unary1);
setLops(unary1);
}
} else // default
{
boolean needAgg = requiresAggregation(input, _direction);
SparkAggType aggtype = getSparkUnaryAggregationType(needAgg);
PartialAggregate aggregate =
new PartialAggregate(
input.constructLops(),
HopsAgg2Lops.get(_op),
HopsDirection2Lops.get(_direction),
DataType.MATRIX,
getValueType(),
aggtype,
et);
aggregate.setDimensionsBasedOnDirection(
getDim1(), getDim2(), input.getRowsInBlock(), input.getColsInBlock());
setLineNumbers(aggregate);
setLops(aggregate);
if (getDataType() == DataType.SCALAR) {
UnaryCP unary1 =
new UnaryCP(
aggregate,
HopsOpOp1LopsUS.get(OpOp1.CAST_AS_SCALAR),
getDataType(),
getValueType());
unary1.getOutputParameters().setDimensions(0, 0, 0, 0, -1);
setLineNumbers(unary1);
setLops(unary1);
}
}
}
} catch (Exception e) {
throw new HopsException(
this.printErrorLocation() + "In AggUnary Hop, error constructing Lops ", e);
}
// add reblock/checkpoint lops if necessary
constructAndSetLopsDataFlowProperties();
// return created lops
return getLops();
}
@Override
protected double computeOutputMemEstimate(long dim1, long dim2, long nnz) {
double sparsity = OptimizerUtils.getSparsity(dim1, dim2, nnz);
return OptimizerUtils.estimateSizeExactSparsity(dim1, dim2, sparsity);
}
/**
* @return
* @throws HopsException
* @throws LopsException
*/
private Lop constructLopsSparkCumulativeUnary() throws HopsException, LopsException {
Hop input = getInput().get(0);
long rlen = input.getDim1();
long clen = input.getDim2();
long brlen = input.getRowsInBlock();
long bclen = input.getColsInBlock();
boolean force = !dimsKnown() || _etypeForced == ExecType.SPARK;
OperationTypes aggtype = getCumulativeAggType();
Lop X = input.constructLops();
Lop TEMP = X;
ArrayList<Lop> DATA = new ArrayList<Lop>();
int level = 0;
// recursive preaggregation until aggregates fit into CP memory budget
while (((2 * OptimizerUtils.estimateSize(TEMP.getOutputParameters().getNumRows(), clen)
+ OptimizerUtils.estimateSize(1, clen))
> OptimizerUtils.getLocalMemBudget()
&& TEMP.getOutputParameters().getNumRows() > 1)
|| force) {
DATA.add(TEMP);
// preaggregation per block (for spark, the CumulativePartialAggregate subsumes both
// the preaggregation and subsequent block aggregation)
long rlenAgg = (long) Math.ceil((double) TEMP.getOutputParameters().getNumRows() / brlen);
Lop preagg =
new CumulativePartialAggregate(
TEMP, DataType.MATRIX, ValueType.DOUBLE, aggtype, ExecType.SPARK);
preagg.getOutputParameters().setDimensions(rlenAgg, clen, brlen, bclen, -1);
setLineNumbers(preagg);
TEMP = preagg;
level++;
force = false; // in case of unknowns, generate one level
}
// in-memory cum sum (of partial aggregates)
if (TEMP.getOutputParameters().getNumRows() != 1) {
int k = OptimizerUtils.getConstrainedNumThreads(_maxNumThreads);
Unary unary1 =
new Unary(
TEMP, HopsOpOp1LopsU.get(_op), DataType.MATRIX, ValueType.DOUBLE, ExecType.CP, k);
unary1
.getOutputParameters()
.setDimensions(TEMP.getOutputParameters().getNumRows(), clen, brlen, bclen, -1);
setLineNumbers(unary1);
TEMP = unary1;
}
// split, group and mr cumsum
while (level-- > 0) {
// (for spark, the CumulativeOffsetBinary subsumes both the split aggregate and
// the subsequent offset binary apply of split aggregates against the original data)
double initValue = getCumulativeInitValue();
CumulativeOffsetBinary binary =
new CumulativeOffsetBinary(
DATA.get(level),
TEMP,
DataType.MATRIX,
ValueType.DOUBLE,
initValue,
aggtype,
ExecType.SPARK);
binary.getOutputParameters().setDimensions(rlen, clen, brlen, bclen, -1);
setLineNumbers(binary);
TEMP = binary;
}
return TEMP;
}
/**
* MR Cumsum is currently based on a multipass algorithm of (1) preaggregation and (2) subsequent
* offsetting. Note that we currently support one robust physical operator but many alternative
* realizations are possible for specific scenarios (e.g., when the preaggregated intermediate fit
* into the map task memory budget) or by creating custom job types.
*
* @return
* @throws HopsException
* @throws LopsException
*/
private Lop constructLopsMRCumulativeUnary() throws HopsException, LopsException {
Hop input = getInput().get(0);
long rlen = input.getDim1();
long clen = input.getDim2();
long brlen = input.getRowsInBlock();
long bclen = input.getColsInBlock();
boolean force = !dimsKnown() || _etypeForced == ExecType.MR;
OperationTypes aggtype = getCumulativeAggType();
Lop X = input.constructLops();
Lop TEMP = X;
ArrayList<Lop> DATA = new ArrayList<Lop>();
int level = 0;
// recursive preaggregation until aggregates fit into CP memory budget
while (((2 * OptimizerUtils.estimateSize(TEMP.getOutputParameters().getNumRows(), clen)
+ OptimizerUtils.estimateSize(1, clen))
> OptimizerUtils.getLocalMemBudget()
&& TEMP.getOutputParameters().getNumRows() > 1)
|| force) {
DATA.add(TEMP);
// preaggregation per block
long rlenAgg = (long) Math.ceil((double) TEMP.getOutputParameters().getNumRows() / brlen);
Lop preagg =
new CumulativePartialAggregate(
TEMP, DataType.MATRIX, ValueType.DOUBLE, aggtype, ExecType.MR);
preagg.getOutputParameters().setDimensions(rlenAgg, clen, brlen, bclen, -1);
setLineNumbers(preagg);
Group group = new Group(preagg, Group.OperationTypes.Sort, DataType.MATRIX, ValueType.DOUBLE);
group.getOutputParameters().setDimensions(rlenAgg, clen, brlen, bclen, -1);
setLineNumbers(group);
Aggregate agg =
new Aggregate(
group, HopsAgg2Lops.get(AggOp.SUM), getDataType(), getValueType(), ExecType.MR);
agg.getOutputParameters().setDimensions(rlenAgg, clen, brlen, bclen, -1);
agg.setupCorrectionLocation(
CorrectionLocationType
.NONE); // aggregation uses kahanSum but the inputs do not have correction values
setLineNumbers(agg);
TEMP = agg;
level++;
force = false; // in case of unknowns, generate one level
}
// in-memory cum sum (of partial aggregates)
if (TEMP.getOutputParameters().getNumRows() != 1) {
int k = OptimizerUtils.getConstrainedNumThreads(_maxNumThreads);
Unary unary1 =
new Unary(
TEMP, HopsOpOp1LopsU.get(_op), DataType.MATRIX, ValueType.DOUBLE, ExecType.CP, k);
unary1
.getOutputParameters()
.setDimensions(TEMP.getOutputParameters().getNumRows(), clen, brlen, bclen, -1);
setLineNumbers(unary1);
TEMP = unary1;
}
// split, group and mr cumsum
while (level-- > 0) {
double init = getCumulativeInitValue();
CumulativeSplitAggregate split =
new CumulativeSplitAggregate(TEMP, DataType.MATRIX, ValueType.DOUBLE, init);
split.getOutputParameters().setDimensions(rlen, clen, brlen, bclen, -1);
setLineNumbers(split);
Group group1 =
new Group(DATA.get(level), Group.OperationTypes.Sort, DataType.MATRIX, ValueType.DOUBLE);
group1.getOutputParameters().setDimensions(rlen, clen, brlen, bclen, -1);
setLineNumbers(group1);
Group group2 = new Group(split, Group.OperationTypes.Sort, DataType.MATRIX, ValueType.DOUBLE);
group2.getOutputParameters().setDimensions(rlen, clen, brlen, bclen, -1);
setLineNumbers(group2);
CumulativeOffsetBinary binary =
new CumulativeOffsetBinary(
group1, group2, DataType.MATRIX, ValueType.DOUBLE, aggtype, ExecType.MR);
binary.getOutputParameters().setDimensions(rlen, clen, brlen, bclen, -1);
setLineNumbers(binary);
TEMP = binary;
}
return TEMP;
}
@Override
public Lop constructLops() throws HopsException, LopsException {
// reuse existing lop
if (getLops() != null) return getLops();
try {
Hop input = getInput().get(0);
if (getDataType() == DataType.SCALAR // value type casts or matrix to scalar
|| (_op == OpOp1.CAST_AS_MATRIX && getInput().get(0).getDataType() == DataType.SCALAR)
|| (_op == OpOp1.CAST_AS_FRAME && getInput().get(0).getDataType() == DataType.SCALAR)) {
if (_op == Hop.OpOp1.IQM) // special handling IQM
{
Lop iqmLop = constructLopsIQM();
setLops(iqmLop);
} else if (_op == Hop.OpOp1.MEDIAN) {
Lop medianLop = constructLopsMedian();
setLops(medianLop);
} else // general case SCALAR/CAST (always in CP)
{
UnaryCP.OperationTypes optype = HopsOpOp1LopsUS.get(_op);
if (optype == null)
throw new HopsException(
"Unknown UnaryCP lop type for UnaryOp operation type '" + _op + "'");
UnaryCP unary1 =
new UnaryCP(input.constructLops(), optype, getDataType(), getValueType());
setOutputDimensions(unary1);
setLineNumbers(unary1);
setLops(unary1);
}
} else // general case MATRIX
{
ExecType et = optFindExecType();
// special handling cumsum/cumprod/cummin/cumsum
if (isCumulativeUnaryOperation() && et != ExecType.CP) {
// TODO additional physical operation if offsets fit in memory
Lop cumsumLop = null;
if (et == ExecType.MR) cumsumLop = constructLopsMRCumulativeUnary();
else cumsumLop = constructLopsSparkCumulativeUnary();
setLops(cumsumLop);
} else // default unary
{
int k =
isCumulativeUnaryOperation()
? OptimizerUtils.getConstrainedNumThreads(_maxNumThreads)
: 1;
Unary unary1 =
new Unary(
input.constructLops(),
HopsOpOp1LopsU.get(_op),
getDataType(),
getValueType(),
et,
k);
setOutputDimensions(unary1);
setLineNumbers(unary1);
setLops(unary1);
}
}
} catch (Exception e) {
throw new HopsException(
this.printErrorLocation() + "error constructing Lops for UnaryOp Hop -- \n ", e);
}
// add reblock/checkpoint lops if necessary
constructAndSetLopsDataFlowProperties();
return getLops();
}
