Modification of https://github.com/DlangScience/examples/blob/master/mean_of_columns.d

main.d

/**
* This code uses std.experimental.ndslice to take the mean of the columns in
* a 2d 100x1000 array and creates a benchmark of that code.
*
* My mean running time is about 1.2 ms.
*
* If we compare this code to the Numpy equivalent,
*
*     import numpy
*     data = numpy.arange(100000, dtype="int32").reshape((100, 1000))
*     means = numpy.mean(data, axis=0)
*
* and we benchmark the numpy.mean line using the following Python command,
*
*    python -m timeit \
*         -s 'import numpy; data = numpy.arange(100000, dtype="int32").reshape((100, 1000))' \
*         'means = numpy.mean(data, axis=0)'
*
* then we get a mean running time of about 65 µs. That means the Python version is about 20x
* as fast as the D version.
*
* D's slowness here is unexpected. Precomputing `means` as a `double` type does not fix this.
*
* If instead you do
*
*     .map!(r => sum(r, 0L) / r.length)        // D
*     numpy.mean(data, dtype="uint64", axis=0)  # Python
*
* I get about 35 µs and 85 µs instead, so D is roughly 2.4x as fast as the Python version.
*/


import std.range : iota;
import std.array : array;
import std.algorithm;
import std.datetime;
import std.conv : to;
import std.stdio;
import std.experimental.ndslice;

enum testCount = 10_000;
double[] means;
int[] data;

void f0() {
    means = data
        .sliced(100, 1000)
        .transposed
        .map!(r => sum(r, 0.0) / r.length)
        .array;
}

void main() {
    data = 100_000.iota.array;
    auto r = benchmark!(f0)(testCount);
    auto f0Result = to!Duration(r[0] / testCount);
    f0Result.writeln;
}