Examples with KahanPlus org.apache.sysml.runtime.functionobjects.KahanPlus used on opensource projects

Example 16 with KahanPlus

use of org.apache.sysml.runtime.functionobjects.KahanPlus in project incubator-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 17 with KahanPlus

use of org.apache.sysml.runtime.functionobjects.KahanPlus in project incubator-systemml by apache.

the class SpoofCellwise method execute.

@Override
public ScalarObject execute(ArrayList<MatrixBlock> inputs, ArrayList<ScalarObject> scalarObjects, int k) {
    // sanity check
    if (inputs == null || inputs.size() < 1)
        throw new RuntimeException("Invalid input arguments.");
    // input preparation
    MatrixBlock a = inputs.get(0);
    SideInput[] b = prepInputMatrices(inputs);
    double[] scalars = prepInputScalars(scalarObjects);
    final int m = a.getNumRows();
    final int n = a.getNumColumns();
    // sparse safe check
    boolean sparseSafe = isSparseSafe() || (b.length == 0 && genexec(0, b, scalars, m, n, 0, 0) == 0);
    long inputSize = sparseSafe ? getTotalInputNnz(inputs) : getTotalInputSize(inputs);
    if (inputSize < PAR_NUMCELL_THRESHOLD) {
        // serial execution
        k = 1;
    }
    double ret = 0;
    if (// SINGLE-THREADED
    k <= 1) {
        if (inputs.get(0) instanceof CompressedMatrixBlock)
            ret = executeCompressedAndAgg((CompressedMatrixBlock) a, b, scalars, m, n, sparseSafe, 0, m);
        else if (!inputs.get(0).isInSparseFormat())
            ret = executeDenseAndAgg(a.getDenseBlock(), b, scalars, m, n, sparseSafe, 0, m);
        else
            ret = executeSparseAndAgg(a.getSparseBlock(), b, scalars, m, n, sparseSafe, 0, m);
    } else // MULTI-THREADED
    {
        try {
            ExecutorService pool = CommonThreadPool.get(k);
            ArrayList<ParAggTask> tasks = new ArrayList<>();
            int nk = (a instanceof CompressedMatrixBlock) ? k : UtilFunctions.roundToNext(Math.min(8 * k, m / 32), k);
            int blklen = (int) (Math.ceil((double) m / nk));
            if (a instanceof CompressedMatrixBlock)
                blklen = BitmapEncoder.getAlignedBlocksize(blklen);
            for (int i = 0; i < nk & i * blklen < m; i++) tasks.add(new ParAggTask(a, b, scalars, m, n, sparseSafe, i * blklen, Math.min((i + 1) * blklen, m)));
            // execute tasks
            List<Future<Double>> taskret = pool.invokeAll(tasks);
            pool.shutdown();
            // aggregate partial results
            ValueFunction vfun = getAggFunction();
            if (vfun instanceof KahanFunction) {
                KahanObject kbuff = new KahanObject(0, 0);
                KahanPlus kplus = KahanPlus.getKahanPlusFnObject();
                for (Future<Double> task : taskret) kplus.execute2(kbuff, task.get());
                ret = kbuff._sum;
            } else {
                for (Future<Double> task : taskret) ret = vfun.execute(ret, task.get());
            }
        } catch (Exception ex) {
            throw new DMLRuntimeException(ex);
        }
    }
    // correction for min/max
    if ((_aggOp == AggOp.MIN || _aggOp == AggOp.MAX) && sparseSafe && a.getNonZeros() < a.getNumRows() * a.getNumColumns())
        // unseen 0 might be max or min value
        ret = getAggFunction().execute(ret, 0);
    return new DoubleObject(ret);
}

Example 18 with KahanPlus

use of org.apache.sysml.runtime.functionobjects.KahanPlus in project incubator-systemml by apache.

the class SpoofMultiAggregate method aggregatePartialResults.

public static void aggregatePartialResults(AggOp[] aggOps, MatrixBlock c, MatrixBlock b) {
    ValueFunction[] vfun = getAggFunctions(aggOps);
    for (int k = 0; k < aggOps.length; k++) {
        if (vfun[k] instanceof KahanFunction) {
            KahanObject kbuff = new KahanObject(c.quickGetValue(0, k), 0);
            KahanPlus kplus = KahanPlus.getKahanPlusFnObject();
            kplus.execute2(kbuff, b.quickGetValue(0, k));
            c.quickSetValue(0, k, kbuff._sum);
        } else {
            double cval = c.quickGetValue(0, k);
            double bval = b.quickGetValue(0, k);
            c.quickSetValue(0, k, vfun[k].execute(cval, bval));
        }
    }
}

Example 19 with KahanPlus

use of org.apache.sysml.runtime.functionobjects.KahanPlus in project incubator-systemml by apache.

the class SpoofMultiAggregate method aggregatePartialResults.

private void aggregatePartialResults(double[] c, ArrayList<double[]> pret) {
    ValueFunction[] vfun = getAggFunctions(_aggOps);
    for (int k = 0; k < _aggOps.length; k++) {
        if (vfun[k] instanceof KahanFunction) {
            KahanObject kbuff = new KahanObject(0, 0);
            KahanPlus kplus = KahanPlus.getKahanPlusFnObject();
            for (double[] tmp : pret) kplus.execute2(kbuff, tmp[k]);
            c[k] = kbuff._sum;
        } else {
            for (double[] tmp : pret) c[k] = vfun[k].execute(c[k], tmp[k]);
        }
    }
}

Example 20 with KahanPlus

use of org.apache.sysml.runtime.functionobjects.KahanPlus in project incubator-systemml by apache.

the class ColGroupDDC1 method computeRowSums.

public static void computeRowSums(ColGroupDDC1[] grps, MatrixBlock result, KahanFunction kplus, int rl, int ru) {
    // note: due to corrections the output might be a large dense block
    DenseBlock c = result.getDenseBlock();
    KahanObject kbuff = new KahanObject(0, 0);
    KahanPlus kplus2 = KahanPlus.getKahanPlusFnObject();
    // prepare distinct values once
    double[][] vals = new double[grps.length][];
    for (int i = 0; i < grps.length; i++) {
        // pre-aggregate all distinct values (guaranteed <=255)
        vals[i] = grps[i].sumAllValues(kplus, kbuff);
    }
    // cache-conscious row sums operations
    // iterative over codes of all groups and add to output
    // (use kahan plus not general KahanFunction for correctness in case of sqk+)
    // 16KB
    int blksz = 1024;
    double[] tmpAgg = new double[blksz];
    for (int bi = rl; bi < ru; bi += blksz) {
        Arrays.fill(tmpAgg, 0);
        // aggregate all groups
        for (int j = 0; j < grps.length; j++) {
            double[] valsj = vals[j];
            byte[] dataj = grps[j]._data;
            for (int i = bi; i < Math.min(bi + blksz, ru); i++) tmpAgg[i - bi] += valsj[dataj[i] & 0xFF];
        }
        // add partial results of all ddc groups
        for (int i = bi; i < Math.min(bi + blksz, ru); i++) {
            double[] cvals = c.values(i);
            int cix = c.pos(i);
            kbuff.set(cvals[cix], cvals[cix + 1]);
            kplus2.execute2(kbuff, tmpAgg[i - bi]);
            cvals[cix] = kbuff._sum;
            cvals[cix + 1] = kbuff._correction;
        }
    }
}