Skip to content

Latest commit

 

History

History
366 lines (258 loc) · 12.9 KB

operator.rst

File metadata and controls

366 lines (258 loc) · 12.9 KB

Operators and functions

Arithmetic operators

xtensor provides overloads of traditional arithmetic operators for :cpp:type:`xt::xexpression` objects:

All these operators are element-wise operators and apply the lazy broadcasting rules explained in a previous section.

#incude "xtensor/xarray.hpp"

xt::xarray<int> a = {{1, 2}, {3, 4}};
xt::xarray<int> b = {1, 2};

xt::xarray<int> res = 2 * (a + b);
// => res = {{4, 8}, {8, 12}}

Logical operators

xtensor also provides overloads of the logical operators:

Like arithmetic operators, these logical operators are element-wise operators and apply the lazy broadcasting rules. In addition to these element-wise logical operators, xtensor provides two reducing boolean functions:

and an element-wise ternary function (similar to the : ? ternary operator):

#include <xtensor/xarray.hpp>

xt::xarray<bool> b = { false, true, true, false };
xt::xarray<int> a1 = { 1,   2,  3,  4 };
xt::xarray<int> a2 = { 11, 12, 13, 14 };

xt::xarray<int> res = xt::where(b, a1, a2);
// => res = { 11, 2, 3, 14 }

Unlike in :any:`numpy.where`, :cpp:func:`xt::where` takes full advantage of the lazyness of xtensor.

Comparison operators

xtensor provides overloads of the inequality operators:

These overloads of inequality operators are quite different from the standard C++ inequality operators: they are element-wise operators returning boolean :cpp:type:`xexpression`:

#include <xtensor/xarray.hpp>

xt::xarray<int> a1 = {  1, 12,  3, 14 };
xt::xarray<int> a2 = { 11,  2, 13, 4  };
xt::xarray<bool> comp = a1 < a2;
// => comp = { true, false, true, false }

However, equality operators are similar to the traditional ones in C++:

Element-wise equality comparison can be achieved through the :cpp:func:`xt::equal` function.

#include <xtensor/xarray.hpp>

xt::xarray<int> a1 = {  1,  2, 3, 4};
xt::xarray<int> a2 = { 11, 12, 3, 4};

bool res = (a1 == a2);
// => res = false

xt::xarray<bool> re = xt::equal(a1, a2);
// => re = { false, false, true, true }

Bitwise operators

xtensor also contains the following bitwise operators:

Mathematical functions

xtensor provides overloads for many of the standard mathematical functions:

See the API reference for a comprehensive list of available functions. Like operators, the mathematical functions are element-wise functions and apply the lazy broadcasting rules.

Casting

xtensor will implicitly promote and/or cast tensor expression elements as needed, which suffices for most use-cases. But explicit casting can be performed via :cpp:func:`xt::cast`, which performs an element-wise static_cast.

#include <xtensor/xarray.hpp>

xt::xarray<int> a = { 3, 5, 7 };

auto res = a / 2;
// => res = { 1, 2, 3 }

auto res2 = xt::cast<double>(a) / 2;
// => res2 = { 1.5, 2.5, 3.5 }

Reducers

xtensor provides reducers, that is, means for accumulating values of tensor expressions over prescribed axes. The return value of a reducer is an :cpp:type:`xt::xexpression` with the same shape as the input expression, with the specified axes removed.

#include <xtensor/xarray.hpp>
#include <xtensor/xmath.hpp>

xt::xarray<double> a = xt::ones<double>({3, 2, 4, 6, 5});
xt::xarray<double> res = xt::sum(a, {1, 3});
// => res.shape() = { 3, 4, 5 };
// => res(0, 0, 0) = 12

You can also call the :cpp:func:`xt::reduce` generator with your own reducing function:

#include <xtensor/xarray.hpp>
#include <xtensor/xreducer.hpp>

xt::xarray<double> arr = some_init_function({3, 2, 4, 6, 5});
xt::xarray<double> res = xt::reduce([](double a, double b) { return a*a + b*b; },
                                    arr,
                                    {1, 3});

The reduce generator also accepts a :cpp:type:`xt::xreducer_functors` object, a tuple of three functions (one for reducing, one for initialization and one for merging). A generator is provided to build the :cpp:type:`xt::xreducer_functors` object, the last function can be omitted:

#include <xtensor/xarray.hpp>
#include <xtensor/xreducer.hpp>

xt::xarray<double> arr = some_init_function({3, 2, 4, 6, 5});
xt::xarray<double> res = xt::reduce(xt::make_xreducer_functor([](double a, double b) { return a*a + b*b; },
                                                              [](double a) { return a * 2; })
                                    arr,
                                    {1, 3});

If no axes are provided, the reduction is performed over all the axes, and the result is a 0-D expression. Since xtensor's expressions are lazy evaluated, you need to explicitely call the access operator to trigger the evaluation and get the result:

#include <xtensor/xarray.hpp>
#include <xtensor/xreducer.hpp>

xt::xarray<double> arr = some_init_function({3, 2, 4, 6, 5});
double res = xt::reduce([](double a, double b) { return a*a + b*b; }, arr)();

The value_type of a reducer is the traditional result type of the reducing operation. For instance, the value_type of the reducer for the sum is:

  • int if the underlying expression holds int values
  • int if the underlying expression holds short values, because short + short = int

You can pass a template argument to the reducer functions to specify the type of the initial value of the reduction. This allows you to "promote" the value type of the reducer and limit overflows in computation:

#include <xtensor/xarray.hpp>
#include <xtensor/xreducer.hpp>

xt::xarray<int> arr = some_init_function({3, 2, 4, 6, 5});
auto s1 = xt::sum<short>(arr); // No effect, short + int = int
auto s2 = xt::sum<long int>(arr); // The value_type of s2 is long int

When you write generic code and you want to limit overflows, you can use :cpp:any:`xt::big_promote_value_type_t` as shown below:

#include <xtensor/xarray.hpp>
#include <xtensor/xreducer.hpp>

template <class E>
void my_computation(E&& e)
{
    auto s = xt::sum<xt::big_promote_value_type_t<E>>(e);
}

Accumulators

Similar to reducers, xtensor provides accumulators which are used to implement cumulative functions such as :cpp:func:`xt::cumsum` or :cpp:func:`xt::cumprod`. Accumulators can currently only work on a single axis. Additionally, the accumulators are not lazy and do not return an xexpression, but rather an evaluated :cpp:type:`xt::xarray` or :cpp:type:`xt::xtensor`.

#include <xtensor/xarray.hpp>
#include <xtensor/xmath.hpp>

xt::xarray<double> a = xt::ones<double>({5, 8, 3});
xt::xarray<double> res = xt::cumsum(a, 1);
// => res.shape() = {5, 8, 3};
// => res(0, 0, 0) = 1
// => res(0, 7, 0) = 8

You can also call the :cpp:func:`xt::accumulate` generator with your own accumulating function. For example, the implementation of cumsum is as follows:

#include <xtensor/xarray.hpp>
#include <xtensor/xaccumulator.hpp>

xt::xarray<double> arr = some_init_function({5, 5, 5});
xt::xarray<double> res = xt::accumulate([](double a, double b) { return a + b; },
                                        arr,
                                        1);

Like reducers, accumulators accept a template parameter to specify the value_type of the initial value of the accumulation. The value_type of the result is computed with the same rules as those for reducers:

#include <xtensor/xarray.hpp>
#include <xtensor/xaccumulator.hpp>

xt::xarray<int> arr = some_init_function({5, 5, 5});
auto r1 = xt::cumsum<short>(a, 1);
// r1 holds int values
auto r2 = xt::cumsum<long int>(a, 1);
// r2 hols long int values

Evaluation strategy

Generally, xtensor implements a :ref:`lazy execution model <lazy-evaluation>`, but under certain circumstances, a greedy execution model with immediate execution can be favorable. For example, reusing (and recomputing) the same values of a reducer over and over again if you use them in a loop can cost a lot of CPU cycles. Additionally, greedy execution can benefit from SIMD acceleration over reduction axes and is faster when the entire result needs to be computed.

Therefore, xtensor allows to select an :cpp:enum:`xt::evaluation_strategy`. Currently, two evaluation strategies are implemented: :cpp:enumerator:`xt::evaluation_strategy::immediate` and :cpp:enumerator:`xt::evaluation_strategy::lazy`. When :cpp:enumerator:`~xt::evaluation_strategy::immediate` evaluation is selected, the return value is not an xexpression, but an in-memory datastructure such as a xarray or xtensor (depending on the input values).

Choosing an evaluation_strategy is straightforward. For reducers:

#include <xtensor/xarray.hpp>
#include <xtensor/xreducer.hpp>

xt::xarray<double> a = xt::ones<double>({3, 2, 4, 6, 5});
auto res = xt::sum(a, {1, 3}, xt::evaluation_strategy::immediate);
// or select the default:
// auto res = xt::sum(a, {1, 3}, xt::evaluation_strategy::lazy);

Note: for accumulators, only the :cpp:enumerator:`~xt::evaluation_strategy::immediate` evaluation strategy is currently implemented.

Universal functions and vectorization

xtensor provides utilities to vectorize any scalar function (taking multiple scalar arguments) into a function that will perform on :cpp:type:`xt::xexpression` s, applying the lazy broadcasting rules which we described in a previous section. These functions are called :cpp:type:`xt::xfunction` s. They are xtensor's counterpart to numpy's universal functions.

Actually, all arithmetic and logical operators, inequality operator and mathematical functions we described before are :cpp:type:`xt::xfunction` s.

The following snippet shows how to vectorize a scalar function taking two arguments:

#include <xtensor/xarray.hpp>
#include <xtensor/xvectorize.hpp>

int f(int a, int b)
{
    return a + 2 * b;
}

auto vecf = xt::vectorize(f);
xt::xarray<int> a = { 11, 12, 13 };
xt::xarray<int> b = {  1,  2,  3 };
xt::xarray<int> res = vecf(a, b);
// => res = { 13, 16, 19 }