Java 8 Stream 10X Multiplication Matrix is ​​slower than for a loop?

Question

I created a module that performs matrix multiplication using threads. It can be found here: https://github.com/firefly-math/firefly-math-linear-real/

I tried to write a test to compare the implementation of a thread loop with the corresponding loop implementation in Apache Commons Math.

The control module is here: https://github.com/firefly-math/firefly-math-benchmark

And the real indicator here: https://github.com/firefly-math/firefly-math-benchmark/blob/master/src/main/java/com/fireflysemantics/benchmark/MultiplyBenchmark.java

When I run the test on matrices 100X100 and 1000X1000, it turns out that the Apache Commons Math (uses the for loop) is 10 times faster (approximately) than the corresponding stream implementation.

# Run complete. Total time: 00:14:10

Benchmark                              Mode  Cnt      Score     Error      Units
MultiplyBenchmark.multiplyCM1000_1000  avgt   30   1040.804 ±  11.796  ms/op
MultiplyBenchmark.multiplyCM100_100    avgt   30      0.790 ±   0.010  ms/op
MultiplyBenchmark.multiplyFM1000_1000  avgt   30  11981.228 ± 405.812  ms/op
MultiplyBenchmark.multiplyFM100_100    avgt   30      7.224 ±   0.685  ms/op

I did something wrong in the benchmark (I hope :))?

I add proven methods so that everyone can see what is being compared. This is the Apache Commons Math Array2DRowRealMatrix.multiply () method:

/**
 * Returns the result of postmultiplying {@code this} by {@code m}.
 *
 * @param m matrix to postmultiply by
 * @return {@code this * m}
 * @throws DimensionMismatchException if
 * {@code columnDimension(this) != rowDimension(m)}
 */
public Array2DRowRealMatrix multiply(final Array2DRowRealMatrix m)
    throws DimensionMismatchException {
    MatrixUtils.checkMultiplicationCompatible(this, m);

    final int nRows = this.getRowDimension();
    final int nCols = m.getColumnDimension();
    final int nSum = this.getColumnDimension();

    final double[][] outData = new double[nRows][nCols];
    // Will hold a column of "m".
    final double[] mCol = new double[nSum];
    final double[][] mData = m.data;

    // Multiply.
    for (int col = 0; col < nCols; col++) {
        // Copy all elements of column "col" of "m" so that
        // will be in contiguous memory.
        for (int mRow = 0; mRow < nSum; mRow++) {
            mCol[mRow] = mData[mRow][col];
        }

        for (int row = 0; row < nRows; row++) {
            final double[] dataRow = data[row];
            double sum = 0;
            for (int i = 0; i < nSum; i++) {
                sum += dataRow[i] * mCol[i];
            }
            outData[row][col] = sum;
        }
    }

    return new Array2DRowRealMatrix(outData, false);
}

And this is the corresponding stream implementation:

/**
 * Returns a {@link BinaryOperator} that multiplies {@link SimpleMatrix}
 * {@code m1} times {@link SimpleMatrix} {@code m2} (m1 X m2).
 * 
 * Example {@code multiply(true).apply(m1, m2);}
 * 
 * @param parallel
 *            Whether to perform the operation concurrently.
 * 
 * @throws MathException
 *             Of type {@code MATRIX_DIMENSION_MISMATCH__MULTIPLICATION} if
 *             {@code m} is not the same size as {@code this}.
 * 
 * @return the {@link BinaryOperator} that performs the operation.
 */
public static BinaryOperator<SimpleMatrix> multiply(boolean parallel) {

    return (m1, m2) -> {
        checkMultiplicationCompatible(m1, m2);

        double[][] a1 = m1.toArray();
        double[][] a2 = m2.toArray();

        Stream<double[]> stream = Arrays.stream(a1);
        stream = parallel ? stream.parallel() : stream;

        final double[][] result =
                stream.map(r -> range(0, a2[0].length)
                        .mapToDouble(i -> range(0, a2.length).mapToDouble(j -> r[j]
                                * a2[j][i]).sum())
                        .toArray()).toArray(double[][]::new);

        return new SimpleMatrix(result);
    };
}

TIA, Ole

Answer 1

Take a look at DoublePipeline.toArray:

public final double[] toArray() {
  return Nodes.flattenDouble((Node.OfDouble) evaluateToArrayNode(Double[]::new))
                    .asPrimitiveArray();
}

It seems that a boxed array is first created, which is then converted to a primitive array.