use of org.apache.sysml.hops.AggBinaryOp in project incubator-systemml by apache.
the class PlanSelectionFuseCostBased method rGetComputeCosts.
private static void rGetComputeCosts(Hop current, HashSet<Long> partition, HashMap<Long, Double> computeCosts) {
if (computeCosts.containsKey(current.getHopID()))
return;
// recursively process children
for (Hop c : current.getInput()) rGetComputeCosts(c, partition, computeCosts);
// get costs for given hop
double costs = 1;
if (current instanceof UnaryOp) {
switch(((UnaryOp) current).getOp()) {
case ABS:
case ROUND:
case CEIL:
case FLOOR:
case SIGN:
costs = 1;
break;
case SPROP:
case SQRT:
costs = 2;
break;
case EXP:
costs = 18;
break;
case SIGMOID:
costs = 21;
break;
case LOG:
case LOG_NZ:
costs = 32;
break;
case NCOL:
case NROW:
case PRINT:
case ASSERT:
case CAST_AS_BOOLEAN:
case CAST_AS_DOUBLE:
case CAST_AS_INT:
case CAST_AS_MATRIX:
case CAST_AS_SCALAR:
costs = 1;
break;
case SIN:
costs = 18;
break;
case COS:
costs = 22;
break;
case TAN:
costs = 42;
break;
case ASIN:
costs = 93;
break;
case ACOS:
costs = 103;
break;
case ATAN:
costs = 40;
break;
// TODO:
case SINH:
costs = 93;
break;
case COSH:
costs = 103;
break;
case TANH:
costs = 40;
break;
case CUMSUM:
case CUMMIN:
case CUMMAX:
case CUMPROD:
costs = 1;
break;
default:
LOG.warn("Cost model not " + "implemented yet for: " + ((UnaryOp) current).getOp());
}
} else if (current instanceof BinaryOp) {
switch(((BinaryOp) current).getOp()) {
case MULT:
case PLUS:
case MINUS:
case MIN:
case MAX:
case AND:
case OR:
case EQUAL:
case NOTEQUAL:
case LESS:
case LESSEQUAL:
case GREATER:
case GREATEREQUAL:
case CBIND:
case RBIND:
costs = 1;
break;
case INTDIV:
costs = 6;
break;
case MODULUS:
costs = 8;
break;
case DIV:
costs = 22;
break;
case LOG:
case LOG_NZ:
costs = 32;
break;
case POW:
costs = (HopRewriteUtils.isLiteralOfValue(current.getInput().get(1), 2) ? 1 : 16);
break;
case MINUS_NZ:
case MINUS1_MULT:
costs = 2;
break;
case CENTRALMOMENT:
int type = (int) (current.getInput().get(1) instanceof LiteralOp ? HopRewriteUtils.getIntValueSafe((LiteralOp) current.getInput().get(1)) : 2);
switch(type) {
// count
case 0:
costs = 1;
break;
// mean
case 1:
costs = 8;
break;
// cm2
case 2:
costs = 16;
break;
// cm3
case 3:
costs = 31;
break;
// cm4
case 4:
costs = 51;
break;
// variance
case 5:
costs = 16;
break;
}
break;
case COVARIANCE:
costs = 23;
break;
default:
LOG.warn("Cost model not " + "implemented yet for: " + ((BinaryOp) current).getOp());
}
} else if (current instanceof TernaryOp) {
switch(((TernaryOp) current).getOp()) {
case PLUS_MULT:
case MINUS_MULT:
costs = 2;
break;
case CTABLE:
costs = 3;
break;
case CENTRALMOMENT:
int type = (int) (current.getInput().get(1) instanceof LiteralOp ? HopRewriteUtils.getIntValueSafe((LiteralOp) current.getInput().get(1)) : 2);
switch(type) {
// count
case 0:
costs = 2;
break;
// mean
case 1:
costs = 9;
break;
// cm2
case 2:
costs = 17;
break;
// cm3
case 3:
costs = 32;
break;
// cm4
case 4:
costs = 52;
break;
// variance
case 5:
costs = 17;
break;
}
break;
case COVARIANCE:
costs = 23;
break;
default:
LOG.warn("Cost model not " + "implemented yet for: " + ((TernaryOp) current).getOp());
}
} else if (current instanceof ParameterizedBuiltinOp) {
costs = 1;
} else if (current instanceof IndexingOp) {
costs = 1;
} else if (current instanceof ReorgOp) {
costs = 1;
} else if (current instanceof AggBinaryOp) {
// matrix vector
costs = 2;
} else if (current instanceof AggUnaryOp) {
switch(((AggUnaryOp) current).getOp()) {
case SUM:
costs = 4;
break;
case SUM_SQ:
costs = 5;
break;
case MIN:
case MAX:
costs = 1;
break;
default:
LOG.warn("Cost model not " + "implemented yet for: " + ((AggUnaryOp) current).getOp());
}
}
computeCosts.put(current.getHopID(), costs);
}
use of org.apache.sysml.hops.AggBinaryOp in project incubator-systemml by apache.
the class PlanSelectionFuseCostBasedV2 method pruneInvalidAndSpecialCasePlans.
private static void pruneInvalidAndSpecialCasePlans(CPlanMemoTable memo, PlanPartition part) {
// prune invalid row entries w/ violated blocksize constraint
if (OptimizerUtils.isSparkExecutionMode()) {
for (Long hopID : part.getPartition()) {
if (!memo.contains(hopID, TemplateType.ROW))
continue;
Hop hop = memo.getHopRefs().get(hopID);
boolean isSpark = DMLScript.rtplatform == RUNTIME_PLATFORM.SPARK || OptimizerUtils.getTotalMemEstimate(hop.getInput().toArray(new Hop[0]), hop, true) > OptimizerUtils.getLocalMemBudget();
boolean validNcol = hop.getDataType().isScalar() || (HopRewriteUtils.isTransposeOperation(hop) ? hop.getDim1() <= hop.getRowsInBlock() : hop.getDim2() <= hop.getColsInBlock());
for (Hop in : hop.getInput()) validNcol &= in.getDataType().isScalar() || (in.getDim2() <= in.getColsInBlock()) || (hop instanceof AggBinaryOp && in.getDim1() <= in.getRowsInBlock() && HopRewriteUtils.isTransposeOperation(in));
if (isSpark && !validNcol) {
List<MemoTableEntry> blacklist = memo.get(hopID, TemplateType.ROW);
memo.remove(memo.getHopRefs().get(hopID), TemplateType.ROW);
memo.removeAllRefTo(hopID, TemplateType.ROW);
if (LOG.isTraceEnabled()) {
LOG.trace("Removed row memo table entries w/ violated blocksize constraint (" + hopID + "): " + Arrays.toString(blacklist.toArray(new MemoTableEntry[0])));
}
}
}
}
// prune row aggregates with pure cellwise operations
// (we determine a blacklist of all operators in a partition that either
// depend upon row aggregates or on which row aggregates depend)
HashSet<Long> blacklist = collectIrreplaceableRowOps(memo, part);
for (Long hopID : part.getPartition()) {
if (blacklist.contains(hopID))
continue;
MemoTableEntry me = memo.getBest(hopID, TemplateType.ROW);
if (me != null && me.type == TemplateType.ROW && memo.hasOnlyExactMatches(hopID, TemplateType.ROW, TemplateType.CELL)) {
List<MemoTableEntry> rmList = memo.get(hopID, TemplateType.ROW);
memo.remove(memo.getHopRefs().get(hopID), new HashSet<>(rmList));
if (LOG.isTraceEnabled()) {
LOG.trace("Removed row memo table entries w/o aggregation: " + Arrays.toString(rmList.toArray(new MemoTableEntry[0])));
}
}
}
// we'd prune sum(X*(U%*%t(V))*Z), Z=U2%*%t(V2) because this would unnecessarily destroy a fusion pattern.
for (Long hopID : part.getPartition()) {
if (memo.countEntries(hopID, TemplateType.OUTER) == 2) {
List<MemoTableEntry> entries = memo.get(hopID, TemplateType.OUTER);
MemoTableEntry me1 = entries.get(0);
MemoTableEntry me2 = entries.get(1);
MemoTableEntry rmEntry = TemplateOuterProduct.dropAlternativePlan(memo, me1, me2);
if (rmEntry != null) {
memo.remove(memo.getHopRefs().get(hopID), Collections.singleton(rmEntry));
memo.getPlansBlacklisted().remove(rmEntry.input(rmEntry.getPlanRefIndex()));
if (LOG.isTraceEnabled())
LOG.trace("Removed dominated outer product memo table entry: " + rmEntry);
}
}
}
}
use of org.apache.sysml.hops.AggBinaryOp in project incubator-systemml by apache.
the class TemplateRow method rConstructCplan.
private void rConstructCplan(Hop hop, CPlanMemoTable memo, HashMap<Long, CNode> tmp, HashSet<Hop> inHops, HashMap<String, Hop> inHops2, boolean compileLiterals) {
// memoization for common subexpression elimination and to avoid redundant work
if (tmp.containsKey(hop.getHopID()))
return;
// recursively process required childs
MemoTableEntry me = memo.getBest(hop.getHopID(), TemplateType.ROW, TemplateType.CELL);
for (int i = 0; i < hop.getInput().size(); i++) {
Hop c = hop.getInput().get(i);
if (me != null && me.isPlanRef(i))
rConstructCplan(c, memo, tmp, inHops, inHops2, compileLiterals);
else {
CNodeData cdata = TemplateUtils.createCNodeData(c, compileLiterals);
tmp.put(c.getHopID(), cdata);
inHops.add(c);
}
}
// construct cnode for current hop
CNode out = null;
if (hop instanceof AggUnaryOp) {
CNode cdata1 = tmp.get(hop.getInput().get(0).getHopID());
if (((AggUnaryOp) hop).getDirection() == Direction.Row && HopRewriteUtils.isAggUnaryOp(hop, SUPPORTED_ROW_AGG)) {
if (hop.getInput().get(0).getDim2() == 1)
out = (cdata1.getDataType() == DataType.SCALAR) ? cdata1 : new CNodeUnary(cdata1, UnaryType.LOOKUP_R);
else {
String opcode = "ROW_" + ((AggUnaryOp) hop).getOp().name().toUpperCase() + "S";
out = new CNodeUnary(cdata1, UnaryType.valueOf(opcode));
if (cdata1 instanceof CNodeData && !inHops2.containsKey("X"))
inHops2.put("X", hop.getInput().get(0));
}
} else if (((AggUnaryOp) hop).getDirection() == Direction.Col && ((AggUnaryOp) hop).getOp() == AggOp.SUM) {
// vector add without temporary copy
if (cdata1 instanceof CNodeBinary && ((CNodeBinary) cdata1).getType().isVectorScalarPrimitive())
out = new CNodeBinary(cdata1.getInput().get(0), cdata1.getInput().get(1), ((CNodeBinary) cdata1).getType().getVectorAddPrimitive());
else
out = cdata1;
} else if (((AggUnaryOp) hop).getDirection() == Direction.RowCol && ((AggUnaryOp) hop).getOp() == AggOp.SUM) {
out = (cdata1.getDataType().isMatrix()) ? new CNodeUnary(cdata1, UnaryType.ROW_SUMS) : cdata1;
}
} else if (hop instanceof AggBinaryOp) {
CNode cdata1 = tmp.get(hop.getInput().get(0).getHopID());
CNode cdata2 = tmp.get(hop.getInput().get(1).getHopID());
if (HopRewriteUtils.isTransposeOperation(hop.getInput().get(0))) {
// correct input under transpose
cdata1 = TemplateUtils.skipTranspose(cdata1, hop.getInput().get(0), tmp, compileLiterals);
inHops.remove(hop.getInput().get(0));
if (cdata1 instanceof CNodeData)
inHops.add(hop.getInput().get(0).getInput().get(0));
// note: vectorMultAdd applicable to vector-scalar, and vector-vector
if (hop.getInput().get(1).getDim2() == 1)
out = new CNodeBinary(cdata1, cdata2, BinType.VECT_MULT_ADD);
else {
out = new CNodeBinary(cdata1, cdata2, BinType.VECT_OUTERMULT_ADD);
if (!inHops2.containsKey("B1")) {
// incl modification of X for consistency
if (cdata1 instanceof CNodeData)
inHops2.put("X", hop.getInput().get(0).getInput().get(0));
inHops2.put("B1", hop.getInput().get(1));
}
}
if (!inHops2.containsKey("X"))
inHops2.put("X", hop.getInput().get(0).getInput().get(0));
} else {
if (hop.getInput().get(0).getDim2() == 1 && hop.getInput().get(1).getDim2() == 1)
out = new CNodeBinary((cdata1.getDataType() == DataType.SCALAR) ? cdata1 : new CNodeUnary(cdata1, UnaryType.LOOKUP0), (cdata2.getDataType() == DataType.SCALAR) ? cdata2 : new CNodeUnary(cdata2, UnaryType.LOOKUP0), BinType.MULT);
else if (hop.getInput().get(1).getDim2() == 1) {
out = new CNodeBinary(cdata1, cdata2, BinType.DOT_PRODUCT);
inHops2.put("X", hop.getInput().get(0));
} else {
out = new CNodeBinary(cdata1, cdata2, BinType.VECT_MATRIXMULT);
inHops2.put("X", hop.getInput().get(0));
inHops2.put("B1", hop.getInput().get(1));
}
}
} else if (HopRewriteUtils.isTransposeOperation(hop)) {
out = TemplateUtils.skipTranspose(tmp.get(hop.getHopID()), hop, tmp, compileLiterals);
if (out instanceof CNodeData && !inHops.contains(hop.getInput().get(0)))
inHops.add(hop.getInput().get(0));
} else if (hop instanceof UnaryOp) {
CNode cdata1 = tmp.get(hop.getInput().get(0).getHopID());
// if one input is a matrix then we need to do vector by scalar operations
if (hop.getInput().get(0).getDim1() >= 1 && hop.getInput().get(0).getDim2() > 1 || (!hop.dimsKnown() && cdata1.getDataType() == DataType.MATRIX)) {
if (HopRewriteUtils.isUnary(hop, SUPPORTED_VECT_UNARY)) {
String opname = "VECT_" + ((UnaryOp) hop).getOp().name();
out = new CNodeUnary(cdata1, UnaryType.valueOf(opname));
if (cdata1 instanceof CNodeData && !inHops2.containsKey("X"))
inHops2.put("X", hop.getInput().get(0));
} else
throw new RuntimeException("Unsupported unary matrix " + "operation: " + ((UnaryOp) hop).getOp().name());
} else // general scalar case
{
cdata1 = TemplateUtils.wrapLookupIfNecessary(cdata1, hop.getInput().get(0));
String primitiveOpName = ((UnaryOp) hop).getOp().toString();
out = new CNodeUnary(cdata1, UnaryType.valueOf(primitiveOpName));
}
} else if (HopRewriteUtils.isBinary(hop, OpOp2.CBIND)) {
// special case for cbind with zeros
CNode cdata1 = tmp.get(hop.getInput().get(0).getHopID());
CNode cdata2 = null;
if (HopRewriteUtils.isDataGenOpWithConstantValue(hop.getInput().get(1))) {
cdata2 = TemplateUtils.createCNodeData(HopRewriteUtils.getDataGenOpConstantValue(hop.getInput().get(1)), true);
// rm 0-matrix
inHops.remove(hop.getInput().get(1));
} else {
cdata2 = tmp.get(hop.getInput().get(1).getHopID());
cdata2 = TemplateUtils.wrapLookupIfNecessary(cdata2, hop.getInput().get(1));
}
out = new CNodeBinary(cdata1, cdata2, BinType.VECT_CBIND);
if (cdata1 instanceof CNodeData && !inHops2.containsKey("X"))
inHops2.put("X", hop.getInput().get(0));
} else if (hop instanceof BinaryOp) {
CNode cdata1 = tmp.get(hop.getInput().get(0).getHopID());
CNode cdata2 = tmp.get(hop.getInput().get(1).getHopID());
// if one input is a matrix then we need to do vector by scalar operations
if ((hop.getInput().get(0).getDim1() >= 1 && hop.getInput().get(0).getDim2() > 1) || (hop.getInput().get(1).getDim1() >= 1 && hop.getInput().get(1).getDim2() > 1) || (!(hop.dimsKnown() && hop.getInput().get(0).dimsKnown() && hop.getInput().get(1).dimsKnown()) && // not a known vector output
(hop.getDim2() != 1) && (cdata1.getDataType().isMatrix() || cdata2.getDataType().isMatrix()))) {
if (HopRewriteUtils.isBinary(hop, SUPPORTED_VECT_BINARY)) {
if (TemplateUtils.isMatrix(cdata1) && (TemplateUtils.isMatrix(cdata2) || TemplateUtils.isRowVector(cdata2))) {
String opname = "VECT_" + ((BinaryOp) hop).getOp().name();
out = new CNodeBinary(cdata1, cdata2, BinType.valueOf(opname));
} else {
String opname = "VECT_" + ((BinaryOp) hop).getOp().name() + "_SCALAR";
if (TemplateUtils.isColVector(cdata1))
cdata1 = new CNodeUnary(cdata1, UnaryType.LOOKUP_R);
if (TemplateUtils.isColVector(cdata2))
cdata2 = new CNodeUnary(cdata2, UnaryType.LOOKUP_R);
out = new CNodeBinary(cdata1, cdata2, BinType.valueOf(opname));
}
if (cdata1 instanceof CNodeData && !inHops2.containsKey("X") && !(cdata1.getDataType() == DataType.SCALAR)) {
inHops2.put("X", hop.getInput().get(0));
}
} else
throw new RuntimeException("Unsupported binary matrix " + "operation: " + ((BinaryOp) hop).getOp().name());
} else // one input is a vector/scalar other is a scalar
{
String primitiveOpName = ((BinaryOp) hop).getOp().toString();
if (TemplateUtils.isColVector(cdata1))
cdata1 = new CNodeUnary(cdata1, UnaryType.LOOKUP_R);
if (// vector or vector can be inferred from lhs
TemplateUtils.isColVector(cdata2) || (TemplateUtils.isColVector(hop.getInput().get(0)) && cdata2 instanceof CNodeData && hop.getInput().get(1).getDataType().isMatrix()))
cdata2 = new CNodeUnary(cdata2, UnaryType.LOOKUP_R);
out = new CNodeBinary(cdata1, cdata2, BinType.valueOf(primitiveOpName));
}
} else if (hop instanceof TernaryOp) {
TernaryOp top = (TernaryOp) hop;
CNode cdata1 = tmp.get(hop.getInput().get(0).getHopID());
CNode cdata2 = tmp.get(hop.getInput().get(1).getHopID());
CNode cdata3 = tmp.get(hop.getInput().get(2).getHopID());
// add lookups if required
cdata1 = TemplateUtils.wrapLookupIfNecessary(cdata1, hop.getInput().get(0));
cdata3 = TemplateUtils.wrapLookupIfNecessary(cdata3, hop.getInput().get(2));
// construct ternary cnode, primitive operation derived from OpOp3
out = new CNodeTernary(cdata1, cdata2, cdata3, TernaryType.valueOf(top.getOp().toString()));
} else if (HopRewriteUtils.isNary(hop, OpOpN.CBIND)) {
CNode[] inputs = new CNode[hop.getInput().size()];
for (int i = 0; i < hop.getInput().size(); i++) {
Hop c = hop.getInput().get(i);
CNode cdata = tmp.get(c.getHopID());
if (TemplateUtils.isColVector(cdata) || TemplateUtils.isRowVector(cdata))
cdata = TemplateUtils.wrapLookupIfNecessary(cdata, c);
inputs[i] = cdata;
if (i == 0 && cdata instanceof CNodeData && !inHops2.containsKey("X"))
inHops2.put("X", c);
}
out = new CNodeNary(inputs, NaryType.VECT_CBIND);
} else if (hop instanceof ParameterizedBuiltinOp) {
CNode cdata1 = tmp.get(((ParameterizedBuiltinOp) hop).getTargetHop().getHopID());
cdata1 = TemplateUtils.wrapLookupIfNecessary(cdata1, hop.getInput().get(0));
CNode cdata2 = tmp.get(((ParameterizedBuiltinOp) hop).getParameterHop("pattern").getHopID());
CNode cdata3 = tmp.get(((ParameterizedBuiltinOp) hop).getParameterHop("replacement").getHopID());
TernaryType ttype = (cdata2.isLiteral() && cdata2.getVarname().equals("Double.NaN")) ? TernaryType.REPLACE_NAN : TernaryType.REPLACE;
out = new CNodeTernary(cdata1, cdata2, cdata3, ttype);
} else if (hop instanceof IndexingOp) {
CNode cdata1 = tmp.get(hop.getInput().get(0).getHopID());
out = new CNodeTernary(cdata1, TemplateUtils.createCNodeData(new LiteralOp(hop.getInput().get(0).getDim2()), true), TemplateUtils.createCNodeData(hop.getInput().get(4), true), (hop.getDim2() != 1) ? TernaryType.LOOKUP_RVECT1 : TernaryType.LOOKUP_RC1);
}
if (out == null) {
throw new RuntimeException(hop.getHopID() + " " + hop.getOpString());
}
if (out.getDataType().isMatrix()) {
out.setNumRows(hop.getDim1());
out.setNumCols(hop.getDim2());
}
tmp.put(hop.getHopID(), out);
}
use of org.apache.sysml.hops.AggBinaryOp in project incubator-systemml by apache.
the class RewriteCompressedReblock method rAnalyzeHopDag.
private static void rAnalyzeHopDag(Hop current, ProbeStatus status) {
if (current.isVisited())
return;
// process children recursively
for (Hop input : current.getInput()) rAnalyzeHopDag(input, status);
// handle source persistent read
if (current.getHopID() == status.startHopID) {
status.compMtx.add(getTmpName(current));
status.foundStart = true;
}
// a) handle function calls
if (current instanceof FunctionOp && hasCompressedInput(current, status)) {
// TODO handle of functions in a more fine-grained manner
// to cover special cases multiple calls where compressed
// inputs might occur for different input parameters
FunctionOp fop = (FunctionOp) current;
String fkey = fop.getFunctionKey();
if (!status.procFn.contains(fkey)) {
// memoization to avoid redundant analysis and recursive calls
status.procFn.add(fkey);
// map inputs to function inputs
FunctionStatementBlock fsb = status.prog.getFunctionStatementBlock(fkey);
FunctionStatement fstmt = (FunctionStatement) fsb.getStatement(0);
ProbeStatus status2 = new ProbeStatus(status);
for (int i = 0; i < fop.getInput().size(); i++) if (status.compMtx.contains(getTmpName(fop.getInput().get(i))))
status2.compMtx.add(fstmt.getInputParams().get(i).getName());
// analyze function and merge meta info
rAnalyzeProgram(fsb, status2);
status.foundStart |= status2.foundStart;
status.usedInLoop |= status2.usedInLoop;
status.condUpdate |= status2.condUpdate;
status.nonApplicable |= status2.nonApplicable;
// map function outputs to outputs
String[] outputs = fop.getOutputVariableNames();
for (int i = 0; i < outputs.length; i++) if (status2.compMtx.contains(fstmt.getOutputParams().get(i).getName()))
status.compMtx.add(outputs[i]);
}
} else // b) handle transient reads and writes (name mapping)
if (HopRewriteUtils.isData(current, DataOpTypes.TRANSIENTWRITE) && status.compMtx.contains(getTmpName(current.getInput().get(0))))
status.compMtx.add(current.getName());
else if (HopRewriteUtils.isData(current, DataOpTypes.TRANSIENTREAD) && status.compMtx.contains(current.getName()))
status.compMtx.add(getTmpName(current));
else // c) handle applicable operations
if (hasCompressedInput(current, status)) {
// valid with uncompressed outputs
boolean compUCOut = (// tsmm
current instanceof AggBinaryOp && current.getDim2() <= current.getColsInBlock() && ((AggBinaryOp) current).checkTransposeSelf() == MMTSJType.LEFT) || // mvmm
(current instanceof AggBinaryOp && (current.getDim1() == 1 || current.getDim2() == 1)) || (HopRewriteUtils.isTransposeOperation(current) && current.getParent().size() == 1 && current.getParent().get(0) instanceof AggBinaryOp && (current.getParent().get(0).getDim1() == 1 || current.getParent().get(0).getDim2() == 1)) || HopRewriteUtils.isAggUnaryOp(current, AggOp.SUM, AggOp.SUM_SQ, AggOp.MIN, AggOp.MAX);
// valid with compressed outputs
boolean compCOut = HopRewriteUtils.isBinaryMatrixScalarOperation(current) || HopRewriteUtils.isBinary(current, OpOp2.CBIND);
boolean metaOp = HopRewriteUtils.isUnary(current, OpOp1.NROW, OpOp1.NCOL);
status.nonApplicable |= !(compUCOut || compCOut || metaOp);
if (compCOut)
status.compMtx.add(getTmpName(current));
}
current.setVisited();
}
use of org.apache.sysml.hops.AggBinaryOp in project incubator-systemml by apache.
the class RewriteSplitDagDataDependentOperators method rCollectDataDependentOperators.
private void rCollectDataDependentOperators(Hop hop, ArrayList<Hop> cand) {
if (hop.isVisited())
return;
// prevent unnecessary dag split (dims known or no consumer operations)
boolean noSplitRequired = (hop.dimsKnown() || HopRewriteUtils.hasOnlyWriteParents(hop, true, true));
boolean investigateChilds = true;
// #1 removeEmpty
if (hop instanceof ParameterizedBuiltinOp && ((ParameterizedBuiltinOp) hop).getOp() == ParamBuiltinOp.RMEMPTY && !noSplitRequired && !(hop.getParent().size() == 1 && hop.getParent().get(0) instanceof TernaryOp && ((TernaryOp) hop.getParent().get(0)).isMatrixIgnoreZeroRewriteApplicable())) {
ParameterizedBuiltinOp pbhop = (ParameterizedBuiltinOp) hop;
cand.add(pbhop);
investigateChilds = false;
// keep interesting consumer information, flag hops accordingly
boolean noEmptyBlocks = true;
boolean onlyPMM = true;
boolean diagInput = pbhop.isTargetDiagInput();
for (Hop p : hop.getParent()) {
// list of operators without need for empty blocks to be extended as needed
noEmptyBlocks &= (p instanceof AggBinaryOp && hop == p.getInput().get(0) || HopRewriteUtils.isUnary(p, OpOp1.NROW));
onlyPMM &= (p instanceof AggBinaryOp && hop == p.getInput().get(0));
}
pbhop.setOutputEmptyBlocks(!noEmptyBlocks);
if (onlyPMM && diagInput) {
if (ConfigurationManager.isDynamicRecompilation())
pbhop.setOutputPermutationMatrix(true);
for (Hop p : hop.getParent()) ((AggBinaryOp) p).setHasLeftPMInput(true);
}
}
// #2 ctable with unknown dims
if (HopRewriteUtils.isTernary(hop, OpOp3.CTABLE) && // dims not provided
hop.getInput().size() < 4 && !noSplitRequired) {
cand.add(hop);
investigateChilds = false;
// keep interesting consumer information, flag hops accordingly
boolean onlyPMM = true;
for (Hop p : hop.getParent()) {
onlyPMM &= (p instanceof AggBinaryOp && hop == p.getInput().get(0));
}
if (onlyPMM && HopRewriteUtils.isBasic1NSequence(hop.getInput().get(0)))
hop.setOutputEmptyBlocks(false);
}
// #3 orderby childs computed in same DAG
if (HopRewriteUtils.isReorg(hop, ReOrgOp.SORT)) {
// params 'decreasing' / 'indexreturn'
for (int i = 2; i <= 3; i++) {
Hop c = hop.getInput().get(i);
if (!(c instanceof LiteralOp || c instanceof DataOp)) {
cand.add(c);
c.setVisited();
investigateChilds = false;
}
}
}
// #4 second-order eval function
if (HopRewriteUtils.isNary(hop, OpOpN.EVAL) && !noSplitRequired) {
cand.add(hop);
investigateChilds = false;
}
// otherwise, processed by recursive rule application)
if (investigateChilds && hop.getInput() != null)
for (Hop c : hop.getInput()) rCollectDataDependentOperators(c, cand);
hop.setVisited();
}
Aggregations