use of org.apache.sysml.runtime.matrix.operators.ScalarOperator in project incubator-systemml by apache.
the class BinarySPInstruction method processMatrixScalarBinaryInstruction.
protected void processMatrixScalarBinaryInstruction(ExecutionContext ec) throws DMLRuntimeException {
SparkExecutionContext sec = (SparkExecutionContext) ec;
//get input RDD
String rddVar = (input1.getDataType() == DataType.MATRIX) ? input1.getName() : input2.getName();
JavaPairRDD<MatrixIndexes, MatrixBlock> in1 = sec.getBinaryBlockRDDHandleForVariable(rddVar);
//get operator and scalar
CPOperand scalar = (input1.getDataType() == DataType.MATRIX) ? input2 : input1;
ScalarObject constant = (ScalarObject) ec.getScalarInput(scalar.getName(), scalar.getValueType(), scalar.isLiteral());
ScalarOperator sc_op = (ScalarOperator) _optr;
sc_op.setConstant(constant.getDoubleValue());
//execute scalar matrix arithmetic instruction
JavaPairRDD<MatrixIndexes, MatrixBlock> out = in1.mapValues(new MatrixScalarUnaryFunction(sc_op));
//put output RDD handle into symbol table
updateUnaryOutputMatrixCharacteristics(sec, rddVar, output.getName());
sec.setRDDHandleForVariable(output.getName(), out);
sec.addLineageRDD(output.getName(), rddVar);
}
use of org.apache.sysml.runtime.matrix.operators.ScalarOperator in project incubator-systemml by apache.
the class ScalarMatrixRelationalCPInstruction method processInstruction.
@Override
public void processInstruction(ExecutionContext ec) throws DMLRuntimeException {
CPOperand mat = (input1.getDataType() == DataType.MATRIX) ? input1 : input2;
CPOperand scalar = (input1.getDataType() == DataType.MATRIX) ? input2 : input1;
MatrixBlock inBlock = ec.getMatrixInput(mat.getName());
ScalarObject constant = (ScalarObject) ec.getScalarInput(scalar.getName(), scalar.getValueType(), scalar.isLiteral());
ScalarOperator sc_op = (ScalarOperator) _optr;
sc_op.setConstant(constant.getDoubleValue());
MatrixBlock retBlock = (MatrixBlock) inBlock.scalarOperations(sc_op, new MatrixBlock());
ec.releaseMatrixInput(mat.getName());
// Ensure right dense/sparse output representation (guarded by released input memory)
if (checkGuardedRepresentationChange(inBlock, retBlock)) {
retBlock.examSparsity();
}
ec.setMatrixOutput(output.getName(), retBlock);
}
use of org.apache.sysml.runtime.matrix.operators.ScalarOperator in project incubator-systemml by apache.
the class LibMatrixCUDA method squareMatrix.
/**
* Helper method to square a matrix in GPU memory
* @param gCtx a valid {@link GPUContext}
* @param instName the invoking instruction's name for record {@link Statistics}.
* @param in input matrix on GPU
* @param out output matrix on GPU
* @param rlen row length
* @param clen column length
* @throws DMLRuntimeException if error
*/
private static void squareMatrix(GPUContext gCtx, String instName, Pointer in, Pointer out, int rlen, int clen) throws DMLRuntimeException {
ScalarOperator power2op = new RightScalarOperator(Power.getPowerFnObject(), 2);
matrixScalarOp(gCtx, instName, in, 2, rlen, clen, out, power2op);
}
use of org.apache.sysml.runtime.matrix.operators.ScalarOperator in project incubator-systemml by apache.
the class LibMatrixCUDA method unaryAggregate.
//********************************************************************/
//***************** END OF MATRIX MULTIPLY Functions *****************/
//********************************************************************/
//********************************************************************/
//**************** UNARY AGGREGATE Functions ************************/
//********************************************************************/
/**
* Entry point to perform Unary aggregate operations on the GPU.
* The execution context object is used to allocate memory for the GPU.
* @param ec Instance of {@link ExecutionContext}, from which the output variable will be allocated
* @param gCtx a valid {@link GPUContext}
* @param instName name of the invoking instruction to record{@link Statistics}.
* @param in1 input matrix
* @param output output matrix/scalar name
* @param op Instance of {@link AggregateUnaryOperator} which encapsulates the direction of reduction/aggregation and the reduction operation.
* @throws DMLRuntimeException if {@link DMLRuntimeException} occurs
*/
public static void unaryAggregate(ExecutionContext ec, GPUContext gCtx, String instName, MatrixObject in1, String output, AggregateUnaryOperator op) throws DMLRuntimeException {
if (ec.getGPUContext() != gCtx)
throw new DMLRuntimeException("GPU : Invalid internal state, the GPUContext set with the ExecutionContext is not the same used to run this LibMatrixCUDA function");
LOG.trace("GPU : unaryAggregate" + ", GPUContext=" + gCtx);
final int REDUCTION_ALL = 1;
final int REDUCTION_ROW = 2;
final int REDUCTION_COL = 3;
final int REDUCTION_DIAG = 4;
// A kahan sum implemention is not provided. is a "uak+" or other kahan operator is encountered,
// it just does regular summation reduction.
final int OP_PLUS = 1;
final int OP_PLUS_SQ = 2;
final int OP_MEAN = 3;
final int OP_VARIANCE = 4;
final int OP_MULTIPLY = 5;
final int OP_MAX = 6;
final int OP_MIN = 7;
final int OP_MAXINDEX = 8;
final int OP_MININDEX = 9;
// Sanity Checks
if (!in1.getGPUObject(gCtx).isAllocated())
throw new DMLRuntimeException("Internal Error - The input is not allocated for a GPU Aggregate Unary:" + in1.getGPUObject(gCtx).isAllocated());
boolean isSparse = in1.getGPUObject(gCtx).isSparse();
IndexFunction indexFn = op.indexFn;
AggregateOperator aggOp = op.aggOp;
// Convert Reduction direction to a number to pass to CUDA kernel
int reductionDirection = -1;
if (indexFn instanceof ReduceAll) {
reductionDirection = REDUCTION_ALL;
} else if (indexFn instanceof ReduceRow) {
reductionDirection = REDUCTION_ROW;
} else if (indexFn instanceof ReduceCol) {
reductionDirection = REDUCTION_COL;
} else if (indexFn instanceof ReduceDiag) {
reductionDirection = REDUCTION_DIAG;
} else {
throw new DMLRuntimeException("Internal Error - Invalid index function type, only reducing along rows, columns, diagonals or all elements is supported in Aggregate Unary operations");
}
assert reductionDirection != -1 : "Internal Error - Incorrect type of reduction direction set for aggregate unary GPU instruction";
// Convert function type to a number to pass to the CUDA Kernel
int opIndex = -1;
if (aggOp.increOp.fn instanceof KahanPlus) {
opIndex = OP_PLUS;
} else if (aggOp.increOp.fn instanceof KahanPlusSq) {
opIndex = OP_PLUS_SQ;
} else if (aggOp.increOp.fn instanceof Mean) {
opIndex = OP_MEAN;
} else if (aggOp.increOp.fn instanceof CM) {
assert ((CM) aggOp.increOp.fn).getAggOpType() == CMOperator.AggregateOperationTypes.VARIANCE : "Internal Error - Invalid Type of CM operator for Aggregate Unary operation on GPU";
opIndex = OP_VARIANCE;
} else if (aggOp.increOp.fn instanceof Plus) {
opIndex = OP_PLUS;
} else if (aggOp.increOp.fn instanceof Multiply) {
opIndex = OP_MULTIPLY;
} else if (aggOp.increOp.fn instanceof Builtin) {
Builtin b = (Builtin) aggOp.increOp.fn;
switch(b.bFunc) {
case MAX:
opIndex = OP_MAX;
break;
case MIN:
opIndex = OP_MIN;
break;
case MAXINDEX:
opIndex = OP_MAXINDEX;
break;
case MININDEX:
opIndex = OP_MININDEX;
break;
default:
new DMLRuntimeException("Internal Error - Unsupported Builtin Function for Aggregate unary being done on GPU");
}
} else {
throw new DMLRuntimeException("Internal Error - Aggregate operator has invalid Value function");
}
assert opIndex != -1 : "Internal Error - Incorrect type of operation set for aggregate unary GPU instruction";
int rlen = (int) in1.getNumRows();
int clen = (int) in1.getNumColumns();
if (isSparse) {
// The strategy for the time being is to convert sparse to dense
// until a sparse specific kernel is written.
in1.getGPUObject(gCtx).sparseToDense(instName);
// long nnz = in1.getNnz();
// assert nnz > 0 : "Internal Error - number of non zeroes set to " + nnz + " in Aggregate Binary for GPU";
// MatrixObject out = ec.getSparseMatrixOutputForGPUInstruction(output, nnz);
// throw new DMLRuntimeException("Internal Error - Not implemented");
}
Pointer out = null;
if (reductionDirection == REDUCTION_COL || reductionDirection == REDUCTION_ROW) {
// Matrix output
MatrixObject out1 = getDenseMatrixOutputForGPUInstruction(ec, instName, output);
out = getDensePointer(gCtx, out1, instName);
}
Pointer in = getDensePointer(gCtx, in1, instName);
int size = rlen * clen;
// For scalars, set the scalar output in the Execution Context object
switch(opIndex) {
case OP_PLUS:
{
switch(reductionDirection) {
case REDUCTION_ALL:
{
double result = reduceAll(gCtx, instName, "reduce_sum", in, size);
ec.setScalarOutput(output, new DoubleObject(result));
break;
}
case REDUCTION_COL:
{
// The names are a bit misleading, REDUCTION_COL refers to the direction (reduce all elements in a column)
reduceRow(gCtx, instName, "reduce_row_sum", in, out, rlen, clen);
break;
}
case REDUCTION_ROW:
{
reduceCol(gCtx, instName, "reduce_col_sum", in, out, rlen, clen);
break;
}
case REDUCTION_DIAG:
throw new DMLRuntimeException("Internal Error - Row, Column and Diag summation not implemented yet");
}
break;
}
case OP_PLUS_SQ:
{
// Calculate the squares in a temporary object tmp
Pointer tmp = gCtx.allocate(instName, size * Sizeof.DOUBLE);
squareMatrix(gCtx, instName, in, tmp, rlen, clen);
// Then do the sum on the temporary object and free it
switch(reductionDirection) {
case REDUCTION_ALL:
{
double result = reduceAll(gCtx, instName, "reduce_sum", tmp, size);
ec.setScalarOutput(output, new DoubleObject(result));
break;
}
case REDUCTION_COL:
{
// The names are a bit misleading, REDUCTION_COL refers to the direction (reduce all elements in a column)
reduceRow(gCtx, instName, "reduce_row_sum", tmp, out, rlen, clen);
break;
}
case REDUCTION_ROW:
{
reduceCol(gCtx, instName, "reduce_col_sum", tmp, out, rlen, clen);
break;
}
default:
throw new DMLRuntimeException("Internal Error - Unsupported reduction direction for summation squared");
}
gCtx.cudaFreeHelper(instName, tmp);
break;
}
case OP_MEAN:
{
switch(reductionDirection) {
case REDUCTION_ALL:
{
double result = reduceAll(gCtx, instName, "reduce_sum", in, size);
double mean = result / size;
ec.setScalarOutput(output, new DoubleObject(mean));
break;
}
case REDUCTION_COL:
{
reduceRow(gCtx, instName, "reduce_row_mean", in, out, rlen, clen);
break;
}
case REDUCTION_ROW:
{
reduceCol(gCtx, instName, "reduce_col_mean", in, out, rlen, clen);
break;
}
default:
throw new DMLRuntimeException("Internal Error - Unsupported reduction direction for mean");
}
break;
}
case OP_MULTIPLY:
{
switch(reductionDirection) {
case REDUCTION_ALL:
{
double result = reduceAll(gCtx, instName, "reduce_prod", in, size);
ec.setScalarOutput(output, new DoubleObject(result));
break;
}
default:
throw new DMLRuntimeException("Internal Error - Unsupported reduction direction for multiplication");
}
break;
}
case OP_MAX:
{
switch(reductionDirection) {
case REDUCTION_ALL:
{
double result = reduceAll(gCtx, instName, "reduce_max", in, size);
ec.setScalarOutput(output, new DoubleObject(result));
break;
}
case REDUCTION_COL:
{
reduceRow(gCtx, instName, "reduce_row_max", in, out, rlen, clen);
break;
}
case REDUCTION_ROW:
{
reduceCol(gCtx, instName, "reduce_col_max", in, out, rlen, clen);
break;
}
default:
throw new DMLRuntimeException("Internal Error - Unsupported reduction direction for max");
}
break;
}
case OP_MIN:
{
switch(reductionDirection) {
case REDUCTION_ALL:
{
double result = reduceAll(gCtx, instName, "reduce_min", in, size);
ec.setScalarOutput(output, new DoubleObject(result));
break;
}
case REDUCTION_COL:
{
reduceRow(gCtx, instName, "reduce_row_min", in, out, rlen, clen);
break;
}
case REDUCTION_ROW:
{
reduceCol(gCtx, instName, "reduce_col_min", in, out, rlen, clen);
break;
}
default:
throw new DMLRuntimeException("Internal Error - Unsupported reduction direction for min");
}
break;
}
case OP_VARIANCE:
{
// Temporary GPU array for
Pointer tmp = gCtx.allocate(instName, size * Sizeof.DOUBLE);
Pointer tmp2 = gCtx.allocate(instName, size * Sizeof.DOUBLE);
switch(reductionDirection) {
case REDUCTION_ALL:
{
double result = reduceAll(gCtx, instName, "reduce_sum", in, size);
double mean = result / size;
// Subtract mean from every element in the matrix
ScalarOperator minusOp = new RightScalarOperator(Minus.getMinusFnObject(), mean);
matrixScalarOp(gCtx, instName, in, mean, rlen, clen, tmp, minusOp);
squareMatrix(gCtx, instName, tmp, tmp2, rlen, clen);
double result2 = reduceAll(gCtx, instName, "reduce_sum", tmp2, size);
double variance = result2 / (size - 1);
ec.setScalarOutput(output, new DoubleObject(variance));
break;
}
case REDUCTION_COL:
{
reduceRow(gCtx, instName, "reduce_row_mean", in, out, rlen, clen);
// Subtract the row-wise mean from every element in the matrix
BinaryOperator minusOp = new BinaryOperator(Minus.getMinusFnObject());
matrixMatrixOp(gCtx, instName, in, out, rlen, clen, VectorShape.NONE.code(), VectorShape.COLUMN.code(), tmp, minusOp);
squareMatrix(gCtx, instName, tmp, tmp2, rlen, clen);
Pointer tmpRow = gCtx.allocate(instName, rlen * Sizeof.DOUBLE);
reduceRow(gCtx, instName, "reduce_row_sum", tmp2, tmpRow, rlen, clen);
ScalarOperator divideOp = new RightScalarOperator(Divide.getDivideFnObject(), clen - 1);
matrixScalarOp(gCtx, instName, tmpRow, clen - 1, rlen, 1, out, divideOp);
gCtx.cudaFreeHelper(instName, tmpRow);
break;
}
case REDUCTION_ROW:
{
reduceCol(gCtx, instName, "reduce_col_mean", in, out, rlen, clen);
// Subtract the columns-wise mean from every element in the matrix
BinaryOperator minusOp = new BinaryOperator(Minus.getMinusFnObject());
matrixMatrixOp(gCtx, instName, in, out, rlen, clen, VectorShape.NONE.code(), VectorShape.ROW.code(), tmp, minusOp);
squareMatrix(gCtx, instName, tmp, tmp2, rlen, clen);
Pointer tmpCol = gCtx.allocate(instName, clen * Sizeof.DOUBLE);
reduceCol(gCtx, instName, "reduce_col_sum", tmp2, tmpCol, rlen, clen);
ScalarOperator divideOp = new RightScalarOperator(Divide.getDivideFnObject(), rlen - 1);
matrixScalarOp(gCtx, instName, tmpCol, rlen - 1, 1, clen, out, divideOp);
gCtx.cudaFreeHelper(instName, tmpCol);
break;
}
default:
throw new DMLRuntimeException("Internal Error - Unsupported reduction direction for variance");
}
gCtx.cudaFreeHelper(instName, tmp);
gCtx.cudaFreeHelper(instName, tmp2);
break;
}
case OP_MAXINDEX:
{
switch(reductionDirection) {
case REDUCTION_COL:
throw new DMLRuntimeException("Internal Error - Column maxindex of matrix not implemented yet for GPU ");
default:
throw new DMLRuntimeException("Internal Error - Unsupported reduction direction for maxindex");
}
// break;
}
case OP_MININDEX:
{
switch(reductionDirection) {
case REDUCTION_COL:
throw new DMLRuntimeException("Internal Error - Column minindex of matrix not implemented yet for GPU ");
default:
throw new DMLRuntimeException("Internal Error - Unsupported reduction direction for minindex");
}
// break;
}
default:
throw new DMLRuntimeException("Internal Error - Invalid GPU Unary aggregate function!");
}
}
use of org.apache.sysml.runtime.matrix.operators.ScalarOperator in project incubator-systemml by apache.
the class BasicScalarOperationsTest method runScalarOperationsTest.
/**
*
* @param mb
*/
private void runScalarOperationsTest(SparsityType sptype, ValueType vtype, boolean compress) {
try {
//prepare sparsity for input data
double sparsity = -1;
switch(sptype) {
case DENSE:
sparsity = sparsity1;
break;
case SPARSE:
sparsity = sparsity2;
break;
case EMPTY:
sparsity = sparsity3;
break;
}
//generate input data
double min = (vtype == ValueType.CONST) ? 10 : -10;
double[][] input = TestUtils.generateTestMatrix(rows, cols, min, 10, sparsity, 7);
if (vtype == ValueType.RAND_ROUND_OLE || vtype == ValueType.RAND_ROUND_DDC) {
CompressedMatrixBlock.ALLOW_DDC_ENCODING = (vtype == ValueType.RAND_ROUND_DDC);
input = TestUtils.round(input);
}
MatrixBlock mb = DataConverter.convertToMatrixBlock(input);
//compress given matrix block
CompressedMatrixBlock cmb = new CompressedMatrixBlock(mb);
if (compress)
cmb.compress();
//matrix-scalar uncompressed
ScalarOperator sop = new RightScalarOperator(Multiply.getMultiplyFnObject(), 7);
MatrixBlock ret1 = (MatrixBlock) mb.scalarOperations(sop, new MatrixBlock());
//matrix-scalar compressed
MatrixBlock ret2 = (MatrixBlock) cmb.scalarOperations(sop, new MatrixBlock());
if (compress)
ret2 = ((CompressedMatrixBlock) ret2).decompress();
//compare result with input
double[][] d1 = DataConverter.convertToDoubleMatrix(ret1);
double[][] d2 = DataConverter.convertToDoubleMatrix(ret2);
TestUtils.compareMatrices(d1, d2, rows, cols, 0.0000001);
} catch (Exception ex) {
throw new RuntimeException(ex);
} finally {
CompressedMatrixBlock.ALLOW_DDC_ENCODING = true;
}
}
Aggregations