@Override
public void processInstruction(
Class<? extends MatrixValue> valueClass,
CachedValueMap cachedValues,
IndexedMatrixValue tempValue,
IndexedMatrixValue zeroInput,
int brlen,
int bclen)
throws DMLUnsupportedOperationException, DMLRuntimeException {
// get both inputs
IndexedMatrixValue left = cachedValues.getFirst(input1);
IndexedMatrixValue right = cachedValues.getFirst(input2);
// check non-existing block
if (left == null || right == null)
throw new DMLRuntimeException(
"Missing append input: isNull(left): "
+ (left == null)
+ ", isNull(right): "
+ (right == null));
// core append operation
MatrixBlock mbLeft = (MatrixBlock) left.getValue();
MatrixBlock mbRight = (MatrixBlock) right.getValue();
MatrixBlock ret = mbLeft.appendOperations(mbRight, new MatrixBlock(), _cbind);
// put result into cache
cachedValues.add(output, new IndexedMatrixValue(left.getIndexes(), ret));
}
@Override
public void processInstruction(
Class<? extends MatrixValue> valueClass,
CachedValueMap cachedValues,
IndexedMatrixValue tempValue,
IndexedMatrixValue zeroInput,
int blockRowFactor,
int blockColFactor)
throws DMLUnsupportedOperationException, DMLRuntimeException {
ArrayList<IndexedMatrixValue> blkList = cachedValues.get(input);
if (blkList != null)
for (IndexedMatrixValue in : blkList) {
if (in == null) continue;
// allocate space for the output value
IndexedMatrixValue out;
if (input == output) out = tempValue;
else out = cachedValues.holdPlace(output, valueClass);
MatrixIndexes inix = in.getIndexes();
// prune unnecessary blocks for trace
if ((((AggregateUnaryOperator) optr).indexFn instanceof ReduceDiag
&& inix.getColumnIndex() != inix.getRowIndex())) {
// do nothing (block not on diagonal); but reset
out.getValue().reset();
} else // general case
{
// process instruction
AggregateUnaryOperator auop = (AggregateUnaryOperator) optr;
OperationsOnMatrixValues.performAggregateUnary(
inix,
in.getValue(),
out.getIndexes(),
out.getValue(),
auop,
blockRowFactor,
blockColFactor);
if (_dropCorr)
((MatrixBlock) out.getValue()).dropLastRowsOrColums(auop.aggOp.correctionLocation);
}
// put the output value in the cache
if (out == tempValue) cachedValues.add(output, out);
}
}
@Override
public void processInstruction(
Class<? extends MatrixValue> valueClass,
CachedValueMap cachedValues,
IndexedMatrixValue tempValue,
IndexedMatrixValue zeroInput,
int blockRowFactor,
int blockColFactor)
throws DMLRuntimeException {
ArrayList<IndexedMatrixValue> blkList = cachedValues.get(input);
if (blkList == null) return;
for (IndexedMatrixValue in1 : blkList) {
if (in1 == null) continue;
MatrixIndexes inix = in1.getIndexes();
MatrixBlock blk = (MatrixBlock) in1.getValue();
long rixOffset = (inix.getRowIndex() - 1) * blockRowFactor;
boolean firstBlk = (inix.getRowIndex() == 1);
boolean lastBlk = (inix.getRowIndex() == _lastRowBlockIndex);
// introduce offsets w/ init value for first row
if (firstBlk) {
IndexedMatrixValue out = cachedValues.holdPlace(output, valueClass);
((MatrixBlock) out.getValue()).reset(1, blk.getNumColumns());
if (_initValue != 0) {
for (int j = 0; j < blk.getNumColumns(); j++)
((MatrixBlock) out.getValue()).appendValue(0, j, _initValue);
}
out.getIndexes().setIndexes(1, inix.getColumnIndex());
}
// output splitting (shift by one), preaggregated offset used by subsequent block
for (int i = 0; i < blk.getNumRows(); i++)
if (!(lastBlk && i == (blk.getNumRows() - 1))) // ignore last row
{
IndexedMatrixValue out = cachedValues.holdPlace(output, valueClass);
MatrixBlock tmpBlk = (MatrixBlock) out.getValue();
tmpBlk.reset(1, blk.getNumColumns());
blk.sliceOperations(i, i, 0, blk.getNumColumns() - 1, tmpBlk);
out.getIndexes().setIndexes(rixOffset + i + 2, inix.getColumnIndex());
}
}
}
@Override
public void processInstruction(
Class<? extends MatrixValue> valueClass,
CachedValueMap cachedValues,
IndexedMatrixValue tempValue,
IndexedMatrixValue zeroInput,
int blockRowFactor,
int blockColFactor)
throws DMLRuntimeException {
IndexedMatrixValue in1 = cachedValues.getFirst(input1);
IndexedMatrixValue in2 = cachedValues.getFirst(input2);
if (in1 == null && in2 == null) return;
// allocate space for the output value
// try to avoid coping as much as possible
IndexedMatrixValue out;
if ((output != input1 && output != input2)
|| (output == input1 && in1 == null)
|| (output == input2 && in2 == null)) out = cachedValues.holdPlace(output, valueClass);
else out = tempValue;
// if one of the inputs is null, then it is a all zero block
MatrixIndexes finalIndexes = null;
if (in1 == null) {
in1 = zeroInput;
in1.getValue().reset(in2.getValue().getNumRows(), in2.getValue().getNumColumns());
finalIndexes = in2.getIndexes();
} else finalIndexes = in1.getIndexes();
if (in2 == null) {
in2 = zeroInput;
in2.getValue().reset(in1.getValue().getNumRows(), in1.getValue().getNumColumns());
}
// process instruction
out.getIndexes().setIndexes(finalIndexes);
OperationsOnMatrixValues.performBinaryIgnoreIndexes(
in1.getValue(), in2.getValue(), out.getValue(), ((BinaryOperator) optr));
// put the output value in the cache
if (out == tempValue) cachedValues.add(output, out);
}
@Override
public void processInstruction(
Class<? extends MatrixValue> valueClass,
CachedValueMap cachedValues,
IndexedMatrixValue tempValue,
IndexedMatrixValue zeroInput,
int blockRowFactor,
int blockColFactor)
throws DMLRuntimeException {
IndexedMatrixValue in1 = cachedValues.getFirst(input1); // original data
IndexedMatrixValue in2 = cachedValues.getFirst(input2); // offset row vector
if (in1 == null || in2 == null)
throw new DMLRuntimeException(
"Unexpected empty input (left="
+ ((in1 == null) ? "null" : in1.getIndexes())
+ ", right="
+ ((in2 == null) ? "null" : in2.getIndexes())
+ ").");
// prepare inputs and outputs
IndexedMatrixValue out = cachedValues.holdPlace(output, valueClass);
MatrixBlock data = (MatrixBlock) in1.getValue();
MatrixBlock offset = (MatrixBlock) in2.getValue();
MatrixBlock blk = (MatrixBlock) out.getValue();
blk.reset(data.getNumRows(), data.getNumColumns());
// blockwise offset aggregation and prefix sum computation
MatrixBlock data2 = new MatrixBlock(data); // cp data
MatrixBlock fdata2 =
data2.sliceOperations(0, 0, 0, data2.getNumColumns() - 1, new MatrixBlock()); // 1-based
fdata2.binaryOperationsInPlace(_bop, offset); // sum offset to first row
data2.copy(0, 0, 0, data2.getNumColumns() - 1, fdata2, true); // 0-based
data2.unaryOperations(_uop, blk); // compute columnwise prefix sums/prod/min/max
// set output indexes
out.getIndexes().setIndexes(in1.getIndexes());
}
// Reused by both MR and Spark for performing zero out
public static IndexRange getSelectedRangeForZeroOut(
IndexedMatrixValue in, int blockRowFactor, int blockColFactor, IndexRange indexRange) {
IndexRange tempRange = new IndexRange(-1, -1, -1, -1);
long topBlockRowIndex = UtilFunctions.computeBlockIndex(indexRange.rowStart, blockRowFactor);
int topRowInTopBlock = UtilFunctions.computeCellInBlock(indexRange.rowStart, blockRowFactor);
long bottomBlockRowIndex = UtilFunctions.computeBlockIndex(indexRange.rowEnd, blockRowFactor);
int bottomRowInBottomBlock =
UtilFunctions.computeCellInBlock(indexRange.rowEnd, blockRowFactor);
long leftBlockColIndex = UtilFunctions.computeBlockIndex(indexRange.colStart, blockColFactor);
int leftColInLeftBlock = UtilFunctions.computeCellInBlock(indexRange.colStart, blockColFactor);
long rightBlockColIndex = UtilFunctions.computeBlockIndex(indexRange.colEnd, blockColFactor);
int rightColInRightBlock = UtilFunctions.computeCellInBlock(indexRange.colEnd, blockColFactor);
// no overlap
if (in.getIndexes().getRowIndex() < topBlockRowIndex
|| in.getIndexes().getRowIndex() > bottomBlockRowIndex
|| in.getIndexes().getColumnIndex() < leftBlockColIndex
|| in.getIndexes().getColumnIndex() > rightBlockColIndex) {
tempRange.set(-1, -1, -1, -1);
return tempRange;
}
// get the index range inside the block
tempRange.set(0, in.getValue().getNumRows() - 1, 0, in.getValue().getNumColumns() - 1);
if (topBlockRowIndex == in.getIndexes().getRowIndex()) tempRange.rowStart = topRowInTopBlock;
if (bottomBlockRowIndex == in.getIndexes().getRowIndex())
tempRange.rowEnd = bottomRowInBottomBlock;
if (leftBlockColIndex == in.getIndexes().getColumnIndex())
tempRange.colStart = leftColInLeftBlock;
if (rightBlockColIndex == in.getIndexes().getColumnIndex())
tempRange.colEnd = rightColInRightBlock;
return tempRange;
}
/**
* @param target
* @param groups
* @param brlen
* @param bclen
* @param outlist
* @throws DMLRuntimeException
*/
public static void performMapGroupedAggregate(
Operator op,
IndexedMatrixValue inTarget,
MatrixBlock groups,
int ngroups,
int brlen,
int bclen,
ArrayList<IndexedMatrixValue> outlist)
throws DMLRuntimeException {
MatrixIndexes ix = inTarget.getIndexes();
MatrixBlock target = (MatrixBlock) inTarget.getValue();
// execute grouped aggregate operations
MatrixBlock out = groups.groupedAggOperations(target, null, new MatrixBlock(), ngroups, op);
if (out.getNumRows() <= brlen && out.getNumColumns() <= bclen) {
// single output block
outlist.add(new IndexedMatrixValue(new MatrixIndexes(1, ix.getColumnIndex()), out));
} else {
// multiple output blocks (by op def, single column block )
for (int blockRow = 0;
blockRow < (int) Math.ceil(out.getNumRows() / (double) brlen);
blockRow++) {
int maxRow =
(blockRow * brlen + brlen < out.getNumRows())
? brlen
: out.getNumRows() - blockRow * brlen;
int row_offset = blockRow * brlen;
// copy submatrix to block
MatrixBlock tmp =
out.sliceOperations(
row_offset, row_offset + maxRow - 1, 0, out.getNumColumns() - 1, new MatrixBlock());
// append block to result cache
outlist.add(
new IndexedMatrixValue(new MatrixIndexes(blockRow + 1, ix.getColumnIndex()), tmp));
}
}
}
/**
* @param in
* @param ixrange
* @param brlen
* @param bclen
* @param rlen
* @param clen
* @param outlist
* @throws DMLRuntimeException
*/
public static void performShift(
IndexedMatrixValue in,
IndexRange ixrange,
int brlen,
int bclen,
long rlen,
long clen,
ArrayList<IndexedMatrixValue> outlist)
throws DMLRuntimeException {
MatrixIndexes ix = in.getIndexes();
MatrixBlock mb = (MatrixBlock) in.getValue();
long start_lhs_globalRowIndex = ixrange.rowStart + (ix.getRowIndex() - 1) * brlen;
long start_lhs_globalColIndex = ixrange.colStart + (ix.getColumnIndex() - 1) * bclen;
long end_lhs_globalRowIndex = start_lhs_globalRowIndex + mb.getNumRows() - 1;
long end_lhs_globalColIndex = start_lhs_globalColIndex + mb.getNumColumns() - 1;
long start_lhs_rowIndex = UtilFunctions.computeBlockIndex(start_lhs_globalRowIndex, brlen);
long end_lhs_rowIndex = UtilFunctions.computeBlockIndex(end_lhs_globalRowIndex, brlen);
long start_lhs_colIndex = UtilFunctions.computeBlockIndex(start_lhs_globalColIndex, bclen);
long end_lhs_colIndex = UtilFunctions.computeBlockIndex(end_lhs_globalColIndex, bclen);
for (long leftRowIndex = start_lhs_rowIndex; leftRowIndex <= end_lhs_rowIndex; leftRowIndex++) {
for (long leftColIndex = start_lhs_colIndex;
leftColIndex <= end_lhs_colIndex;
leftColIndex++) {
// Calculate global index of right hand side block
long lhs_rl = Math.max((leftRowIndex - 1) * brlen + 1, start_lhs_globalRowIndex);
long lhs_ru = Math.min(leftRowIndex * brlen, end_lhs_globalRowIndex);
long lhs_cl = Math.max((leftColIndex - 1) * bclen + 1, start_lhs_globalColIndex);
long lhs_cu = Math.min(leftColIndex * bclen, end_lhs_globalColIndex);
int lhs_lrl = UtilFunctions.computeCellInBlock(lhs_rl, brlen);
int lhs_lru = UtilFunctions.computeCellInBlock(lhs_ru, brlen);
int lhs_lcl = UtilFunctions.computeCellInBlock(lhs_cl, bclen);
int lhs_lcu = UtilFunctions.computeCellInBlock(lhs_cu, bclen);
long rhs_rl = lhs_rl - ixrange.rowStart + 1;
long rhs_ru = rhs_rl + (lhs_ru - lhs_rl);
long rhs_cl = lhs_cl - ixrange.colStart + 1;
long rhs_cu = rhs_cl + (lhs_cu - lhs_cl);
int rhs_lrl = UtilFunctions.computeCellInBlock(rhs_rl, brlen);
int rhs_lru = UtilFunctions.computeCellInBlock(rhs_ru, brlen);
int rhs_lcl = UtilFunctions.computeCellInBlock(rhs_cl, bclen);
int rhs_lcu = UtilFunctions.computeCellInBlock(rhs_cu, bclen);
MatrixBlock slicedRHSBlk =
mb.sliceOperations(rhs_lrl, rhs_lru, rhs_lcl, rhs_lcu, new MatrixBlock());
int lbrlen = UtilFunctions.computeBlockSize(rlen, leftRowIndex, brlen);
int lbclen = UtilFunctions.computeBlockSize(clen, leftColIndex, bclen);
MatrixBlock resultBlock = new MatrixBlock(lbrlen, lbclen, false);
resultBlock =
resultBlock.leftIndexingOperations(
slicedRHSBlk, lhs_lrl, lhs_lru, lhs_lcl, lhs_lcu, null, UpdateType.COPY);
outlist.add(
new IndexedMatrixValue(new MatrixIndexes(leftRowIndex, leftColIndex), resultBlock));
}
}
}
/**
* @param val
* @param range
* @param brlen
* @param bclen
* @param outlist
* @throws DMLRuntimeException
*/
public static void performSlice(
IndexedMatrixValue in,
IndexRange ixrange,
int brlen,
int bclen,
ArrayList<IndexedMatrixValue> outlist)
throws DMLRuntimeException {
long cellIndexTopRow = UtilFunctions.computeCellIndex(in.getIndexes().getRowIndex(), brlen, 0);
long cellIndexBottomRow =
UtilFunctions.computeCellIndex(
in.getIndexes().getRowIndex(), brlen, in.getValue().getNumRows() - 1);
long cellIndexLeftCol =
UtilFunctions.computeCellIndex(in.getIndexes().getColumnIndex(), bclen, 0);
long cellIndexRightCol =
UtilFunctions.computeCellIndex(
in.getIndexes().getColumnIndex(), bclen, in.getValue().getNumColumns() - 1);
long cellIndexOverlapTop = Math.max(cellIndexTopRow, ixrange.rowStart);
long cellIndexOverlapBottom = Math.min(cellIndexBottomRow, ixrange.rowEnd);
long cellIndexOverlapLeft = Math.max(cellIndexLeftCol, ixrange.colStart);
long cellIndexOverlapRight = Math.min(cellIndexRightCol, ixrange.colEnd);
// check if block is outside the indexing range
if (cellIndexOverlapTop > cellIndexOverlapBottom
|| cellIndexOverlapLeft > cellIndexOverlapRight) {
return;
}
IndexRange tmpRange =
new IndexRange(
UtilFunctions.computeCellInBlock(cellIndexOverlapTop, brlen),
UtilFunctions.computeCellInBlock(cellIndexOverlapBottom, brlen),
UtilFunctions.computeCellInBlock(cellIndexOverlapLeft, bclen),
UtilFunctions.computeCellInBlock(cellIndexOverlapRight, bclen));
int rowCut = UtilFunctions.computeCellInBlock(ixrange.rowStart, brlen);
int colCut = UtilFunctions.computeCellInBlock(ixrange.colStart, bclen);
int rowsInLastBlock = (int) ((ixrange.rowEnd - ixrange.rowStart + 1) % brlen);
if (rowsInLastBlock == 0) rowsInLastBlock = brlen;
int colsInLastBlock = (int) ((ixrange.colEnd - ixrange.colStart + 1) % bclen);
if (colsInLastBlock == 0) colsInLastBlock = bclen;
long resultBlockIndexTop =
UtilFunctions.computeBlockIndex(cellIndexOverlapTop - ixrange.rowStart + 1, brlen);
long resultBlockIndexBottom =
UtilFunctions.computeBlockIndex(cellIndexOverlapBottom - ixrange.rowStart + 1, brlen);
long resultBlockIndexLeft =
UtilFunctions.computeBlockIndex(cellIndexOverlapLeft - ixrange.colStart + 1, bclen);
long resultBlockIndexRight =
UtilFunctions.computeBlockIndex(cellIndexOverlapRight - ixrange.colStart + 1, bclen);
int boundaryRlen = brlen;
int boundaryClen = bclen;
long finalBlockIndexBottom =
UtilFunctions.computeBlockIndex(ixrange.rowEnd - ixrange.rowStart + 1, brlen);
long finalBlockIndexRight =
UtilFunctions.computeBlockIndex(ixrange.colEnd - ixrange.colStart + 1, bclen);
if (resultBlockIndexBottom == finalBlockIndexBottom) boundaryRlen = rowsInLastBlock;
if (resultBlockIndexRight == finalBlockIndexRight) boundaryClen = colsInLastBlock;
// allocate space for the output value
for (long r = resultBlockIndexTop; r <= resultBlockIndexBottom; r++)
for (long c = resultBlockIndexLeft; c <= resultBlockIndexRight; c++) {
IndexedMatrixValue out = new IndexedMatrixValue(new MatrixIndexes(), new MatrixBlock());
out.getIndexes().setIndexes(r, c);
outlist.add(out);
}
// execute actual slice operation
in.getValue()
.sliceOperations(
outlist, tmpRange, rowCut, colCut, brlen, bclen, boundaryRlen, boundaryClen);
}
@Override
public void processInstruction(
Class<? extends MatrixValue> valueClass,
CachedValueMap cachedValues,
IndexedMatrixValue tempValue,
IndexedMatrixValue zeroInput,
int blockRowFactor,
int blockColFactor)
throws DMLRuntimeException {
QuaternaryOperator qop = (QuaternaryOperator) optr;
ArrayList<IndexedMatrixValue> blkList = cachedValues.get(_input1);
if (blkList != null)
for (IndexedMatrixValue imv : blkList) {
// Step 1: prepare inputs and output
if (imv == null) continue;
MatrixIndexes inIx = imv.getIndexes();
MatrixValue inVal = imv.getValue();
// allocate space for the output value
IndexedMatrixValue iout = null;
if (output == _input1) iout = tempValue;
else iout = cachedValues.holdPlace(output, valueClass);
MatrixIndexes outIx = iout.getIndexes();
MatrixValue outVal = iout.getValue();
// Step 2: get remaining inputs: Wij, Ui, Vj
MatrixValue Xij = inVal;
// get Wij if existing (null of WeightsType.NONE or WSigmoid any type)
IndexedMatrixValue iWij = (_input4 != -1) ? cachedValues.getFirst(_input4) : null;
MatrixValue Wij = (iWij != null) ? iWij.getValue() : null;
if (null == Wij && qop.hasFourInputs()) {
MatrixBlock mb = new MatrixBlock(1, 1, false);
String[] parts = InstructionUtils.getInstructionParts(instString);
mb.quickSetValue(0, 0, Double.valueOf(parts[4]));
Wij = mb;
}
// get Ui and Vj, potentially through distributed cache
MatrixValue Ui =
(!_cacheU)
? cachedValues.getFirst(_input2).getValue() // U
: MRBaseForCommonInstructions.dcValues
.get(_input2)
.getDataBlock((int) inIx.getRowIndex(), 1)
.getValue();
MatrixValue Vj =
(!_cacheV)
? cachedValues.getFirst(_input3).getValue() // t(V)
: MRBaseForCommonInstructions.dcValues
.get(_input3)
.getDataBlock((int) inIx.getColumnIndex(), 1)
.getValue();
// handle special input case: //V through shuffle -> t(V)
if (Ui.getNumColumns() != Vj.getNumColumns()) {
Vj =
LibMatrixReorg.reorg(
(MatrixBlock) Vj,
new MatrixBlock(Vj.getNumColumns(), Vj.getNumRows(), Vj.isInSparseFormat()),
new ReorgOperator(SwapIndex.getSwapIndexFnObject()));
}
// Step 3: process instruction
Xij.quaternaryOperations(qop, Ui, Vj, Wij, outVal);
// set output indexes
if (qop.wtype1 != null || qop.wtype4 != null) outIx.setIndexes(1, 1); // wsloss
else if (qop.wtype2 != null
|| qop.wtype5 != null
|| qop.wtype3 != null && qop.wtype3.isBasic())
outIx.setIndexes(inIx); // wsigmoid/wdivmm-basic
else { // wdivmm
boolean left = qop.wtype3.isLeft();
outIx.setIndexes(left ? inIx.getColumnIndex() : inIx.getRowIndex(), 1);
}
// put the output value in the cache
if (iout == tempValue) cachedValues.add(output, iout);
}
}