use of org.apache.sysml.hops.IndexingOp 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.IndexingOp 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.IndexingOp in project incubator-systemml by apache.
the class RewriteForLoopVectorization method vectorizeElementwiseUnary.
private static StatementBlock vectorizeElementwiseUnary(StatementBlock sb, StatementBlock csb, Hop from, Hop to, Hop increment, String itervar) {
StatementBlock ret = sb;
// check supported increment values
if (!(increment instanceof LiteralOp && ((LiteralOp) increment).getDoubleValue() == 1.0)) {
return ret;
}
// check for applicability
boolean apply = false;
// row or col
boolean rowIx = false;
if (csb.getHops() != null && csb.getHops().size() == 1) {
Hop root = csb.getHops().get(0);
if (root.getDataType() == DataType.MATRIX && root.getInput().get(0) instanceof LeftIndexingOp) {
LeftIndexingOp lix = (LeftIndexingOp) root.getInput().get(0);
Hop lixlhs = lix.getInput().get(0);
Hop lixrhs = lix.getInput().get(1);
if (lixlhs instanceof DataOp && lixrhs instanceof UnaryOp && lixrhs.getInput().get(0) instanceof IndexingOp && lixrhs.getInput().get(0).getInput().get(0) instanceof DataOp) {
boolean[] tmp = checkLeftAndRightIndexing(lix, (IndexingOp) lixrhs.getInput().get(0), itervar);
apply = tmp[0];
rowIx = tmp[1];
}
}
}
// apply rewrite if possible
if (apply) {
Hop root = csb.getHops().get(0);
LeftIndexingOp lix = (LeftIndexingOp) root.getInput().get(0);
UnaryOp uop = (UnaryOp) lix.getInput().get(1);
IndexingOp rix = (IndexingOp) uop.getInput().get(0);
int index1 = rowIx ? 2 : 4;
int index2 = rowIx ? 3 : 5;
// modify left indexing bounds
HopRewriteUtils.replaceChildReference(lix, lix.getInput().get(index1), from, index1);
HopRewriteUtils.replaceChildReference(lix, lix.getInput().get(index2), to, index2);
// modify right indexing
HopRewriteUtils.replaceChildReference(rix, rix.getInput().get(index1 - 1), from, index1 - 1);
HopRewriteUtils.replaceChildReference(rix, rix.getInput().get(index2 - 1), to, index2 - 1);
updateLeftAndRightIndexingSizes(rowIx, lix, rix);
uop.refreshSizeInformation();
// after uop update
lix.refreshSizeInformation();
ret = csb;
LOG.debug("Applied vectorizeElementwiseUnaryForLoop.");
}
return ret;
}
use of org.apache.sysml.hops.IndexingOp in project incubator-systemml by apache.
the class RewriteForLoopVectorization method vectorizeIndexedCopy.
private static StatementBlock vectorizeIndexedCopy(StatementBlock sb, StatementBlock csb, Hop from, Hop to, Hop increment, String itervar) {
StatementBlock ret = sb;
// check supported increment values
if (!(increment instanceof LiteralOp && ((LiteralOp) increment).getDoubleValue() == 1.0)) {
return ret;
}
// check for applicability
boolean apply = false;
// row or col
boolean rowIx = false;
if (csb.getHops() != null && csb.getHops().size() == 1) {
Hop root = csb.getHops().get(0);
if (root.getDataType() == DataType.MATRIX && root.getInput().get(0) instanceof LeftIndexingOp) {
LeftIndexingOp lix = (LeftIndexingOp) root.getInput().get(0);
Hop lixlhs = lix.getInput().get(0);
Hop lixrhs = lix.getInput().get(1);
if (lixlhs instanceof DataOp && lixrhs instanceof IndexingOp && lixrhs.getInput().get(0) instanceof DataOp) {
boolean[] tmp = checkLeftAndRightIndexing(lix, (IndexingOp) lixrhs, itervar);
apply = tmp[0];
rowIx = tmp[1];
}
}
}
// apply rewrite if possible
if (apply) {
Hop root = csb.getHops().get(0);
LeftIndexingOp lix = (LeftIndexingOp) root.getInput().get(0);
IndexingOp rix = (IndexingOp) lix.getInput().get(1);
int index1 = rowIx ? 2 : 4;
int index2 = rowIx ? 3 : 5;
// modify left indexing bounds
HopRewriteUtils.replaceChildReference(lix, lix.getInput().get(index1), from, index1);
HopRewriteUtils.replaceChildReference(lix, lix.getInput().get(index2), to, index2);
// modify right indexing
HopRewriteUtils.replaceChildReference(rix, rix.getInput().get(index1 - 1), from, index1 - 1);
HopRewriteUtils.replaceChildReference(rix, rix.getInput().get(index2 - 1), to, index2 - 1);
updateLeftAndRightIndexingSizes(rowIx, lix, rix);
ret = csb;
LOG.debug("Applied vectorizeIndexedCopy.");
}
return ret;
}
use of org.apache.sysml.hops.IndexingOp in project incubator-systemml by apache.
the class RewriteForLoopVectorization method vectorizeScalarAggregate.
private static StatementBlock vectorizeScalarAggregate(StatementBlock sb, StatementBlock csb, Hop from, Hop to, Hop increment, String itervar) {
StatementBlock ret = sb;
// check missing and supported increment values
if (!(increment != null && increment instanceof LiteralOp && ((LiteralOp) increment).getDoubleValue() == 1.0)) {
return ret;
}
// check for applicability
boolean leftScalar = false;
boolean rightScalar = false;
// row or col
boolean rowIx = false;
if (csb.getHops() != null && csb.getHops().size() == 1) {
Hop root = csb.getHops().get(0);
if (root.getDataType() == DataType.SCALAR && root.getInput().get(0) instanceof BinaryOp) {
BinaryOp bop = (BinaryOp) root.getInput().get(0);
Hop left = bop.getInput().get(0);
Hop right = bop.getInput().get(1);
// check for left scalar plus
if (HopRewriteUtils.isValidOp(bop.getOp(), MAP_SCALAR_AGGREGATE_SOURCE_OPS) && left instanceof DataOp && left.getDataType() == DataType.SCALAR && root.getName().equals(left.getName()) && right instanceof UnaryOp && ((UnaryOp) right).getOp() == OpOp1.CAST_AS_SCALAR && right.getInput().get(0) instanceof IndexingOp) {
IndexingOp ix = (IndexingOp) right.getInput().get(0);
if (ix.isRowLowerEqualsUpper() && ix.getInput().get(1) instanceof DataOp && ix.getInput().get(1).getName().equals(itervar)) {
leftScalar = true;
rowIx = true;
} else if (ix.isColLowerEqualsUpper() && ix.getInput().get(3) instanceof DataOp && ix.getInput().get(3).getName().equals(itervar)) {
leftScalar = true;
rowIx = false;
}
} else // check for right scalar plus
if (HopRewriteUtils.isValidOp(bop.getOp(), MAP_SCALAR_AGGREGATE_SOURCE_OPS) && right instanceof DataOp && right.getDataType() == DataType.SCALAR && root.getName().equals(right.getName()) && left instanceof UnaryOp && ((UnaryOp) left).getOp() == OpOp1.CAST_AS_SCALAR && left.getInput().get(0) instanceof IndexingOp) {
IndexingOp ix = (IndexingOp) left.getInput().get(0);
if (ix.isRowLowerEqualsUpper() && ix.getInput().get(1) instanceof DataOp && ix.getInput().get(1).getName().equals(itervar)) {
rightScalar = true;
rowIx = true;
} else if (ix.isColLowerEqualsUpper() && ix.getInput().get(3) instanceof DataOp && ix.getInput().get(3).getName().equals(itervar)) {
rightScalar = true;
rowIx = false;
}
}
}
}
// apply rewrite if possible
if (leftScalar || rightScalar) {
Hop root = csb.getHops().get(0);
BinaryOp bop = (BinaryOp) root.getInput().get(0);
Hop cast = bop.getInput().get(leftScalar ? 1 : 0);
Hop ix = cast.getInput().get(0);
int aggOpPos = HopRewriteUtils.getValidOpPos(bop.getOp(), MAP_SCALAR_AGGREGATE_SOURCE_OPS);
AggOp aggOp = MAP_SCALAR_AGGREGATE_TARGET_OPS[aggOpPos];
// replace cast with sum
AggUnaryOp newSum = HopRewriteUtils.createAggUnaryOp(ix, aggOp, Direction.RowCol);
HopRewriteUtils.removeChildReference(cast, ix);
HopRewriteUtils.removeChildReference(bop, cast);
HopRewriteUtils.addChildReference(bop, newSum, leftScalar ? 1 : 0);
// modify indexing expression according to loop predicate from-to
// NOTE: any redundant index operations are removed via dynamic algebraic simplification rewrites
int index1 = rowIx ? 1 : 3;
int index2 = rowIx ? 2 : 4;
HopRewriteUtils.replaceChildReference(ix, ix.getInput().get(index1), from, index1);
HopRewriteUtils.replaceChildReference(ix, ix.getInput().get(index2), to, index2);
// update indexing size information
if (rowIx)
((IndexingOp) ix).setRowLowerEqualsUpper(false);
else
((IndexingOp) ix).setColLowerEqualsUpper(false);
ix.refreshSizeInformation();
ret = csb;
LOG.debug("Applied vectorizeScalarSumForLoop.");
}
return ret;
}
Aggregations