use of org.apache.sysml.runtime.matrix.mapred.IndexedMatrixValue in project incubator-systemml by apache.
the class AggregateBinaryInstruction method processInstruction.
@Override
public void processInstruction(Class<? extends MatrixValue> valueClass, CachedValueMap cachedValues, IndexedMatrixValue tempValue, IndexedMatrixValue zeroInput, int blockRowFactor, int blockColFactor) {
IndexedMatrixValue in1 = cachedValues.getFirst(input1);
IndexedMatrixValue in2 = cachedValues.getFirst(input2);
if (_opcode.equals(MapMult.OPCODE)) {
// check empty inputs (data for different instructions)
if (_cacheType.isRight() ? in1 == null : in2 == null)
return;
// one of the input is from distributed cache.
processMapMultInstruction(valueClass, cachedValues, in1, in2, blockRowFactor, blockColFactor);
} else // generic matrix mult
{
// check empty inputs (data for different instructions)
if (in1 == null || in2 == null)
return;
// allocate space for the output value
IndexedMatrixValue out;
if (output == input1 || output == input2)
out = tempValue;
else
out = cachedValues.holdPlace(output, valueClass);
// process instruction
OperationsOnMatrixValues.performAggregateBinary(in1.getIndexes(), (MatrixBlock) in1.getValue(), in2.getIndexes(), (MatrixBlock) in2.getValue(), out.getIndexes(), (MatrixBlock) out.getValue(), ((AggregateBinaryOperator) optr));
// put the output value in the cache
if (out == tempValue)
cachedValues.add(output, out);
}
}
use of org.apache.sysml.runtime.matrix.mapred.IndexedMatrixValue in project incubator-systemml by apache.
the class AggregateBinaryInstruction method processMapMultInstruction.
/**
* Helper function to perform map-side matrix-matrix multiplication.
*
* @param valueClass matrix value class
* @param cachedValues cached value map
* @param in1 indexed matrix value 1
* @param in2 indexed matrix value 2
* @param blockRowFactor ?
* @param blockColFactor ?
*/
private void processMapMultInstruction(Class<? extends MatrixValue> valueClass, CachedValueMap cachedValues, IndexedMatrixValue in1, IndexedMatrixValue in2, int blockRowFactor, int blockColFactor) {
boolean removeOutput = true;
if (_cacheType.isRight()) {
DistributedCacheInput dcInput = MRBaseForCommonInstructions.dcValues.get(input2);
long in2_cols = dcInput.getNumCols();
long in2_colBlocks = (long) Math.ceil(((double) in2_cols) / dcInput.getNumColsPerBlock());
for (int bidx = 1; bidx <= in2_colBlocks; bidx++) {
// Matrix multiply A[i,k] %*% B[k,bid]
// Setup input2 block
IndexedMatrixValue in2Block = dcInput.getDataBlock((int) in1.getIndexes().getColumnIndex(), bidx);
MatrixValue in2BlockValue = in2Block.getValue();
MatrixIndexes in2BlockIndex = in2Block.getIndexes();
// allocate space for the output value
IndexedMatrixValue out = cachedValues.holdPlace(output, valueClass);
// process instruction
OperationsOnMatrixValues.performAggregateBinary(in1.getIndexes(), (MatrixBlock) in1.getValue(), in2BlockIndex, (MatrixBlock) in2BlockValue, out.getIndexes(), (MatrixBlock) out.getValue(), ((AggregateBinaryOperator) optr));
removeOutput &= (!_outputEmptyBlocks && out.getValue().isEmpty());
}
} else {
DistributedCacheInput dcInput = MRBaseForCommonInstructions.dcValues.get(input1);
long in1_rows = dcInput.getNumRows();
long in1_rowsBlocks = (long) Math.ceil(((double) in1_rows) / dcInput.getNumRowsPerBlock());
for (int bidx = 1; bidx <= in1_rowsBlocks; bidx++) {
// Matrix multiply A[i,k] %*% B[k,bid]
// Setup input2 block
IndexedMatrixValue in1Block = dcInput.getDataBlock(bidx, (int) in2.getIndexes().getRowIndex());
MatrixValue in1BlockValue = in1Block.getValue();
MatrixIndexes in1BlockIndex = in1Block.getIndexes();
// allocate space for the output value
IndexedMatrixValue out = cachedValues.holdPlace(output, valueClass);
// process instruction
OperationsOnMatrixValues.performAggregateBinary(in1BlockIndex, (MatrixBlock) in1BlockValue, in2.getIndexes(), (MatrixBlock) in2.getValue(), out.getIndexes(), (MatrixBlock) out.getValue(), ((AggregateBinaryOperator) optr));
removeOutput &= (!_outputEmptyBlocks && out.getValue().isEmpty());
}
}
// empty block output filter (enabled by compiler consumer operation is in CP)
if (removeOutput)
cachedValues.remove(output);
}
use of org.apache.sysml.runtime.matrix.mapred.IndexedMatrixValue in project incubator-systemml by apache.
the class AppendGInstruction method processInstruction.
@Override
public void processInstruction(Class<? extends MatrixValue> valueClass, CachedValueMap cachedValues, IndexedMatrixValue tempValue, IndexedMatrixValue zeroInput, int brlen, int bclen) {
// setup basic meta data
int blen = _cbind ? bclen : brlen;
// Step 1: handle first input (forward blocks, change dim of last block)
ArrayList<IndexedMatrixValue> blkList1 = cachedValues.get(input1);
if (blkList1 != null)
for (IndexedMatrixValue in1 : blkList1) {
if (in1 == null)
continue;
if (_offset % blen == 0) {
// special case: forward only
cachedValues.add(output, in1);
} else // general case: change dims and forward
{
MatrixIndexes tmpix = in1.getIndexes();
// always block
MatrixBlock tmpval = (MatrixBlock) in1.getValue();
if (// border block
_cbind && _offset / blen + 1 == tmpix.getColumnIndex() || !_cbind && _offset / blen + 1 == tmpix.getRowIndex()) {
IndexedMatrixValue data = cachedValues.holdPlace(output, valueClass);
// always block
MatrixBlock tmpvalNew = (MatrixBlock) data.getValue();
int lrlen = _cbind ? tmpval.getNumRows() : Math.min(blen, (int) (_len - (tmpix.getRowIndex() - 1) * blen));
int lclen = _cbind ? Math.min(blen, (int) (_len - (tmpix.getColumnIndex() - 1) * blen)) : tmpval.getNumColumns();
tmpvalNew.reset(lrlen, lclen);
tmpvalNew.copy(0, tmpval.getNumRows() - 1, 0, tmpval.getNumColumns() - 1, tmpval, true);
data.getIndexes().setIndexes(tmpix);
} else // inner block
{
cachedValues.add(output, in1);
}
}
}
// Step 2: handle second input (split/forward blocks with new index)
ArrayList<IndexedMatrixValue> blkList2 = cachedValues.get(input2);
if (blkList2 != null)
for (IndexedMatrixValue in2 : blkList2) {
if (in2 == null)
continue;
MatrixIndexes tmpix = in2.getIndexes();
// always block
MatrixBlock tmpval = (MatrixBlock) in2.getValue();
if (// special case no split
_offset % bclen == 0) {
IndexedMatrixValue data = cachedValues.holdPlace(output, valueClass);
MatrixIndexes ix1 = data.getIndexes();
long rix = _cbind ? tmpix.getRowIndex() : _offset / blen + tmpix.getRowIndex();
long cix = _cbind ? _offset / blen + tmpix.getColumnIndex() : tmpix.getColumnIndex();
ix1.setIndexes(rix, cix);
data.set(ix1, in2.getValue());
} else // general case: split and forward
{
IndexedMatrixValue data1 = cachedValues.holdPlace(output, valueClass);
MatrixIndexes ix1 = data1.getIndexes();
// always block
MatrixBlock tmpvalNew = (MatrixBlock) data1.getValue();
if (_cbind) {
// first half
int cix1 = (int) (_offset / blen + tmpix.getColumnIndex());
int cols1 = Math.min(blen, (int) (_len - (long) (cix1 - 1) * blen));
ix1.setIndexes(tmpix.getRowIndex(), cix1);
tmpvalNew.reset(tmpval.getNumRows(), cols1);
tmpvalNew.copy(0, tmpval.getNumRows() - 1, (int) ((_offset + 1) % blen) - 1, cols1 - 1, tmpval.slice(0, tmpval.getNumRows() - 1, 0, (int) (cols1 - ((_offset) % blen) - 1), new MatrixBlock()), true);
data1.getIndexes().setIndexes(ix1);
if (cols1 - ((_offset) % blen) < tmpval.getNumColumns()) {
// second half (if required)
IndexedMatrixValue data2 = cachedValues.holdPlace(output, valueClass);
MatrixIndexes ix2 = data2.getIndexes();
// always block
MatrixBlock tmpvalNew2 = (MatrixBlock) data2.getValue();
int cix2 = (int) (_offset / blen + 1 + tmpix.getColumnIndex());
int cols2 = Math.min(blen, (int) (_len - (long) (cix2 - 1) * blen));
ix2.setIndexes(tmpix.getRowIndex(), cix2);
tmpvalNew2.reset(tmpval.getNumRows(), cols2);
tmpvalNew2.copy(0, tmpval.getNumRows() - 1, 0, cols2 - 1, tmpval.slice(0, tmpval.getNumRows() - 1, (int) (cols1 - ((_offset) % blen)), tmpval.getNumColumns() - 1, new MatrixBlock()), true);
data2.getIndexes().setIndexes(ix2);
}
} else // rbind
{
// first half
int rix1 = (int) (_offset / blen + tmpix.getRowIndex());
int rows1 = Math.min(blen, (int) (_len - (long) (rix1 - 1) * blen));
ix1.setIndexes(rix1, tmpix.getColumnIndex());
tmpvalNew.reset(rows1, tmpval.getNumColumns());
tmpvalNew.copy((int) ((_offset + 1) % blen) - 1, rows1 - 1, 0, tmpval.getNumColumns() - 1, tmpval.slice(0, (int) (rows1 - ((_offset) % blen) - 1), 0, tmpval.getNumColumns() - 1, new MatrixBlock()), true);
data1.getIndexes().setIndexes(ix1);
if (rows1 - ((_offset) % blen) < tmpval.getNumRows()) {
// second half (if required)
IndexedMatrixValue data2 = cachedValues.holdPlace(output, valueClass);
MatrixIndexes ix2 = data2.getIndexes();
// always block
MatrixBlock tmpvalNew2 = (MatrixBlock) data2.getValue();
int rix2 = (int) (_offset / blen + 1 + tmpix.getRowIndex());
int rows2 = Math.min(blen, (int) (_len - (long) (rix2 - 1) * blen));
ix2.setIndexes(rix2, tmpix.getColumnIndex());
tmpvalNew2.reset(rows2, tmpval.getNumColumns());
tmpvalNew2.copy(0, rows2 - 1, 0, tmpval.getNumColumns() - 1, tmpval.slice((int) (rows1 - ((_offset) % blen)), tmpval.getNumRows() - 1, 0, tmpval.getNumColumns() - 1, new MatrixBlock()), true);
data2.getIndexes().setIndexes(ix2);
}
}
}
}
}
use of org.apache.sysml.runtime.matrix.mapred.IndexedMatrixValue in project incubator-systemml by apache.
the class BinaryMInstruction method processInstruction.
@Override
public void processInstruction(Class<? extends MatrixValue> valueClass, CachedValueMap cachedValues, IndexedMatrixValue tempValue, IndexedMatrixValue zeroInput, int blockRowFactor, int blockColFactor) {
ArrayList<IndexedMatrixValue> blkList = cachedValues.get(input1);
if (blkList == null)
return;
for (IndexedMatrixValue in1 : blkList) {
// allocate space for the output value
// try to avoid coping as much as possible
IndexedMatrixValue out;
if ((output != input1 && output != input2))
out = cachedValues.holdPlace(output, valueClass);
else
out = tempValue;
// get second
DistributedCacheInput dcInput = MRBaseForCommonInstructions.dcValues.get(input2);
IndexedMatrixValue in2 = null;
if (_vectorType == VectorType.COL_VECTOR)
in2 = dcInput.getDataBlock((int) in1.getIndexes().getRowIndex(), 1);
else
// _vectorType == VectorType.ROW_VECTOR
in2 = dcInput.getDataBlock(1, (int) in1.getIndexes().getColumnIndex());
// process instruction
out.getIndexes().setIndexes(in1.getIndexes());
OperationsOnMatrixValues.performBinaryIgnoreIndexes(in1.getValue(), in2.getValue(), out.getValue(), ((BinaryOperator) optr));
// put the output value in the cache
if (out == tempValue)
cachedValues.add(output, out);
}
}
use of org.apache.sysml.runtime.matrix.mapred.IndexedMatrixValue in project incubator-systemml by apache.
the class QuaternaryInstruction method processInstruction.
@Override
public void processInstruction(Class<? extends MatrixValue> valueClass, CachedValueMap cachedValues, IndexedMatrixValue tempValue, IndexedMatrixValue zeroInput, int blockRowFactor, int blockColFactor) {
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();
MatrixBlock inVal = (MatrixBlock) 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
MatrixBlock 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 = // U
(!_cacheU) ? // U
cachedValues.getFirst(_input2).getValue() : MRBaseForCommonInstructions.dcValues.get(_input2).getDataBlock((int) inIx.getRowIndex(), 1).getValue();
MatrixValue Vj = // t(V)
(!_cacheV) ? // t(V)
cachedValues.getFirst(_input3).getValue() : 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, (MatrixBlock) Ui, (MatrixBlock) Vj, (MatrixBlock) Wij, (MatrixBlock) outVal);
if (qop.wtype1 != null || qop.wtype4 != null)
// wsloss
outIx.setIndexes(1, 1);
else if (qop.wtype2 != null || qop.wtype5 != null || qop.wtype3 != null && qop.wtype3.isBasic())
// wsigmoid/wdivmm-basic
outIx.setIndexes(inIx);
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);
}
}
Aggregations