Examples with RightScalarOperator org.apache.sysml.runtime.matrix.operators.RightScalarOperator used on opensource projects

Example 11 with RightScalarOperator

use of org.apache.sysml.runtime.matrix.operators.RightScalarOperator in project incubator-systemml by apache.

the class BasicScalarOperationsSparseUnsafeTest method runScalarOperationsTest.

private static 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(Plus.getPlusFnObject(), 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;
    }
}

Example 12 with RightScalarOperator

use of org.apache.sysml.runtime.matrix.operators.RightScalarOperator in project 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
 */
private static void squareMatrix(GPUContext gCtx, String instName, Pointer in, Pointer out, int rlen, int clen) {
    ScalarOperator power2op = new RightScalarOperator(Power.getPowerFnObject(), 2);
    matrixScalarOp(gCtx, instName, in, 2, rlen, clen, out, power2op);
}

Example 13 with RightScalarOperator

use of org.apache.sysml.runtime.matrix.operators.RightScalarOperator in project systemml by apache.

the class LibMatrixCUDA method unaryAggregate.

// ********************************************************************/
// ******** End of TRANSPOSE SELF 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.
 */
public static void unaryAggregate(ExecutionContext ec, GPUContext gCtx, String instName, MatrixObject in1, String output, AggregateUnaryOperator op) {
    if (ec.getGPUContext(0) != gCtx)
        throw new DMLRuntimeException("GPU : Invalid internal state, the GPUContext set with the ExecutionContext is not the same used to run this LibMatrixCUDA function");
    if (LOG.isTraceEnabled()) {
        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
    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");
    }
    if (reductionDirection == -1)
        throw new DMLRuntimeException("Internal Error - Incorrect type of reduction direction set for aggregate unary GPU instruction");
    // Convert function type to a number
    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) {
        if (((CM) aggOp.increOp.fn).getAggOpType() != CMOperator.AggregateOperationTypes.VARIANCE)
            throw new DMLRuntimeException("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");
    }
    if (opIndex == -1)
        throw new DMLRuntimeException("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");
    }
    long outRLen = -1;
    long outCLen = -1;
    if (indexFn instanceof ReduceRow) {
        // COL{SUM, MAX...}
        outRLen = 1;
        outCLen = clen;
    } else if (indexFn instanceof ReduceCol) {
        // ROW{SUM, MAX,...}
        outRLen = rlen;
        outCLen = 1;
    }
    Pointer out = null;
    if (reductionDirection == REDUCTION_COL || reductionDirection == REDUCTION_ROW) {
        // Matrix output
        MatrixObject out1 = getDenseMatrixOutputForGPUInstruction(ec, instName, output, outRLen, outCLen);
        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 * sizeOfDataType);
                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 * sizeOfDataType);
                Pointer tmp2 = gCtx.allocate(instName, size * sizeOfDataType);
                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 * sizeOfDataType);
                            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 * sizeOfDataType);
                            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!");
    }
}

Example 14 with RightScalarOperator

use of org.apache.sysml.runtime.matrix.operators.RightScalarOperator in project systemml by apache.

the class LibMatrixCUDA method matrixScalarArithmetic.

/**
 * Entry point to perform elementwise matrix-scalar arithmetic operation specified by op
 *
 * @param ec                execution context
 * @param gCtx              a valid {@link GPUContext}
 * @param instName          the invoking instruction's name for record {@link Statistics}.
 * @param in                input matrix
 * @param outputName        output matrix name
 * @param isInputTransposed true if input transposed
 * @param op                scalar operator
 */
public static void matrixScalarArithmetic(ExecutionContext ec, GPUContext gCtx, String instName, MatrixObject in, String outputName, boolean isInputTransposed, ScalarOperator op) {
    if (ec.getGPUContext(0) != gCtx)
        throw new DMLRuntimeException("GPU : Invalid internal state, the GPUContext set with the ExecutionContext is not the same used to run this LibMatrixCUDA function");
    double constant = op.getConstant();
    if (LOG.isTraceEnabled()) {
        LOG.trace("GPU : matrixScalarArithmetic, scalar: " + constant + ", GPUContext=" + gCtx);
    }
    int outRLen = isInputTransposed ? (int) in.getNumColumns() : (int) in.getNumRows();
    int outCLen = isInputTransposed ? (int) in.getNumRows() : (int) in.getNumColumns();
    // if(!isCUDALibAvailable) {
    if (constant == 0) {
        if (op.fn instanceof Plus || (op.fn instanceof Minus && op instanceof RightScalarOperator) || op.fn instanceof Or) {
            deviceCopy(ec, gCtx, instName, in, outputName, isInputTransposed);
        } else if (op.fn instanceof Multiply || op.fn instanceof And) {
            setOutputToConstant(ec, gCtx, instName, 0.0, outputName, outRLen, outCLen);
        } else if (op.fn instanceof Power) {
            setOutputToConstant(ec, gCtx, instName, 1.0, outputName, outRLen, outCLen);
        } else // TODO:
        // x/0.0 is either +Infinity or -Infinity according to Java.
        // In the context of a matrix, different elements of the matrix
        // could have different values.
        // If the IEEE 754 standard defines otherwise, this logic needs
        // to be re-enabled and the Java computation logic for divide by zero
        // needs to be revisited
        // else if(op.fn instanceof Divide && isSparseAndEmpty(gCtx, in)) {
        // setOutputToConstant(ec, gCtx, instName, Double.NaN, outputName);
        // }
        // else if(op.fn instanceof Divide) {
        // //For division, IEEE 754 defines x/0.0 as INFINITY and 0.0/0.0 as NaN.
        // compareAndSet(ec, gCtx, instName, in, outputName, 0.0, 1e-6, Double.NaN, Double.POSITIVE_INFINITY, Double.POSITIVE_INFINITY);
        // }
        {
            // TODO: Potential to optimize
            matrixScalarOp(ec, gCtx, instName, in, outputName, isInputTransposed, op);
        }
    } else if (constant == 1.0 && op.fn instanceof Or) {
        setOutputToConstant(ec, gCtx, instName, 1.0, outputName, outRLen, outCLen);
    } else if (constant == 1.0 && (op.fn instanceof And || op.fn instanceof Power)) {
        deviceCopy(ec, gCtx, instName, in, outputName, isInputTransposed);
    } else {
        matrixScalarOp(ec, gCtx, instName, in, outputName, isInputTransposed, op);
    }
// }
// else {
// double alpha = 0;
// if(op.fn instanceof Multiply) {
// alpha = op.getConstant();
// }
// else if(op.fn instanceof Divide && op instanceof RightScalarOperator) {
// alpha = Math.pow(op.getConstant(), -1.0);
// }
// else {
// throw new DMLRuntimeException("Unsupported op");
// }
// TODO: Performance optimization: Call cublasDaxpy if(in.getNumRows() == 1 || in.getNumColumns() == 1)
// C = alpha* op( A ) + beta* op ( B )
// dgeam(ec, gCtx, instName, in, in, outputName, isInputTransposed, isInputTransposed, alpha, 0.0);
// }
}

Example 15 with RightScalarOperator

use of org.apache.sysml.runtime.matrix.operators.RightScalarOperator in project systemml by apache.

the class LibMatrixCUDA method matrixMatrixOp.

/**
 * Utility to launch binary cellwise matrix-matrix operations CUDA kernel
 *
 * @param gCtx              a valid {@link GPUContext}
 * @param ec                execution context
 * @param instName          the invoking instruction's name for record {@link Statistics}.
 * @param in1               left input matrix
 * @param in2               right input matrix
 * @param outputName        output variable name
 * @param isLeftTransposed  true if left matrix is transposed
 * @param isRightTransposed true if right matrix is transposed
 * @param op                operator
 */
private static void matrixMatrixOp(ExecutionContext ec, GPUContext gCtx, String instName, MatrixObject in1, MatrixObject in2, String outputName, boolean isLeftTransposed, boolean isRightTransposed, BinaryOperator op) {
    if (ec.getGPUContext(0) != gCtx)
        throw new DMLRuntimeException("GPU : Invalid internal state, the GPUContext set with the ExecutionContext is not the same used to run this LibMatrixCUDA function");
    boolean isEmpty1 = isSparseAndEmpty(gCtx, in1);
    boolean isEmpty2 = isSparseAndEmpty(gCtx, in2);
    int rlenA = toInt(in1.getNumRows());
    int rlenB = toInt(in2.getNumRows());
    int clenA = toInt(in1.getNumColumns());
    int clenB = toInt(in2.getNumColumns());
    int vecStatusA = getVectorStatus(rlenA, clenA).code();
    int vecStatusB = getVectorStatus(rlenB, clenB).code();
    if (isLeftTransposed || isRightTransposed) {
        throw new DMLRuntimeException("Unsupported operator: GPU transposed binary op " + isLeftTransposed + " " + isRightTransposed);
    }
    long outRLen = Math.max(rlenA, rlenB);
    long outCLen = Math.max(clenA, clenB);
    if (isEmpty1 && isEmpty2) {
        MatrixObject out = ec.allocateGPUMatrixObject(outputName, outRLen, outCLen);
        // When both inputs are empty, the output is empty too (except in the case of division)
        if (op.fn instanceof Divide || op.fn instanceof IntegerDivide || op.fn instanceof Modulus) {
            out.getGPUObject(gCtx).allocateAndFillDense(Double.NaN);
        } else if (op.fn instanceof Minus1Multiply) {
            out.getGPUObject(gCtx).allocateAndFillDense(1.0);
        } else {
            out.getGPUObject(gCtx).allocateSparseAndEmpty();
        }
    } else // Check for M1 * M2 when M1 is empty; if M2 is a vector then fallback to general case
    if (isEmpty1 && clenB != 1 && rlenB != 1) {
        // C = empty_in1 op in2 ==> becomes ==> C = 0.0 op in2
        matrixScalarArithmetic(ec, gCtx, instName, in2, outputName, isRightTransposed, new LeftScalarOperator(op.fn, 0.0));
    } else // Check for M1 * M2 when M2 is empty; if M1 is a vector then fallback to general case
    if (isEmpty2 && clenA != 1 && rlenA != 1) {
        // C = in1 op empty_in2 ==> becomes ==> C = in1 op 0.0
        matrixScalarArithmetic(ec, gCtx, instName, in1, outputName, isLeftTransposed, new RightScalarOperator(op.fn, 0.0));
    } else {
        // TODO: FIXME: Implement sparse binCellSparseOp kernel
        Pointer A = getDensePointer(gCtx, in1, instName);
        // TODO: FIXME: Implement sparse binCellSparseOp kernel
        Pointer B = getDensePointer(gCtx, in2, instName);
        // Allocated the dense output matrix
        MatrixObject out = null;
        try {
            out = getDenseMatrixOutputForGPUInstruction(ec, instName, outputName, outRLen, outCLen);
        } catch (DMLRuntimeException e) {
            throw new DMLRuntimeException("Incorrect dimensions: dimA:[" + rlenA + "," + clenA + "]" + " dimB:[" + rlenB + "," + clenB + "] out:[" + outRLen + "," + outCLen + "]", e);
        }
        Pointer C = getDensePointer(gCtx, out, instName);
        int maxRlen = Math.max(rlenA, rlenB);
        int maxClen = Math.max(clenA, clenB);
        matrixMatrixOp(gCtx, instName, A, B, maxRlen, maxClen, vecStatusA, vecStatusB, C, op);
    }
}