Efficient vector operators in C++ / References to temporary objects - c++

I am trying to write a C++ vector class that stores an array of data and allows performing mathematical operations on an element-by-element basis. I want to implement this in such a way that an expression a = b + c + d should loop over all elements only once and directly write the sum b[i] + c[i] + d[i] to a[i] without creating intermediate vectors.
I was writing something like this:
template<class T, int N>
class VectorExpression {
public:
virtual T operator[] (int i) const = 0;
virtual ~VectorExpression() {}
}
template<class T, int N>
class MyVector : public VectorExpression<T, N> {
T data[N];
public:
T& operator[] (int i) { return data[i]; }
T& const operator[] (int i) const { return data[i]; }
MyVector<T,N>& operator=(const VectorExpression<T,N> &rhs) {
for (int i = 0; i < N; ++i)
data[i] = rhs[i];
return *this;
}
}
template<class T, int N>
class VectorSum : public VectorExpression<T, N> {
VectorExpression<T,N> &a, &b;
public:
VectorSum(VectorExpression<T,N> &aa, VectorExpression<T,N> &bb)
: a(aa), b(bb) {}
T operator[] (int i) const { return a[i] + b[i]; }
}
template<class T, int N>
VectorSum<T,N> operator+(const VectorExpression<T,N> &a,
const VectorExpression<T,N> &b)
{
return VectorSum<T,N>(a, b);
}
int main() {
MyVector<double,10> a, b, c, d;
// Initialize b, c, d here
a = b + c + d;
return 0;
}
Probably this functionality is provided by the valarray class but that's because I tried to strip it down to a minimal example.
I made operator[] virtual because this allows nesting all kinds of expressions (e.g. a = !(-b*c + d)) provided I would define all the operators and the corresponding classes similar to VectorSum.
I use references because ordinary variables aren't polymorphic and pointers don't work with operator overloading.
Now my questions about this are:
In the statement a = b + c + d;, two temporary VectorSum<double,10> objects will be created to store b + c and (b+c) + d respectively. Will they live long enough to make the polymorphic behavior work? More specifically, (b+c) + d will store a reference to b + c, but will that object still exist when operator= is called? According to this post all temporaries should exist until operator= returns, but does this also hold for older versions of C++?
If not, then how is this done? The only alternative I see would be to allocate the VectorSum objects using new, return them by reference and then delete them in the operator= functions, but that seems a little cumbersome, and probably a lot less efficient. I'm also not sure if it is always safe.
(Minor question) Is it okay to override the return type T of VectorExpression::operator[] by T& const in MyVector?
EDIT
I had wrong argument types in operator+: changed them from VectorSum to VectorExpression.

Well here's what I came up with:
#include <iostream>
#include <initializer_list>
#include <algorithm>
template<class T, int N>
class VectorExpression {
public:
virtual T operator[] (int i) = 0;
virtual const T operator[] (int i) const = 0;
virtual ~VectorExpression() {}
};
template<class T, int N>
class MyVector : public VectorExpression<T, N> {
T data[N];
public:
MyVector() {
// initialize zero
std::fill(std::begin(data), std::end(data), T());
}
MyVector(const std::initializer_list<T>& values) {
// initialize from array initializer_list
std::copy(std::begin(values), std::end(values), data);
}
MyVector(const VectorExpression<T,N>& rhs) {
for (int i = 0; i < N; ++i)
data[i] = rhs[i];
}
MyVector<T,N>& operator=(const VectorExpression<T,N>& rhs) {
for (int i = 0; i < N; ++i)
data[i] = rhs[i];
return *this;
}
T operator[] (int i) { return data[i]; }
const T operator[] (int i) const { return data[i]; }
friend std::ostream& operator<<(std::ostream& stream, MyVector& obj) {
stream << "[";
for (int i = 0; i < N; ++i) {
stream << obj.data[i] << ", ";
}
stream << "]";
return stream;
}
};
template<class T, int N>
class VectorSum : public VectorExpression<T, N> {
const MyVector<T,N> &a, &b;
public:
VectorSum(const MyVector<T,N>& aa, const MyVector<T,N>& bb):
a(aa), b(bb) {
}
T operator[] (int i) { return return a[i] + b[i]; }
const T operator[] (int i) const { return a[i] + b[i]; }
};
template<class T, int N>
MyVector<T,N> operator+(const MyVector<T,N>& a, const MyVector<T,N>& b) {
return VectorSum<T,N>(a, b);
}
int main() {
MyVector<double,3> a, b({1,2,3}), c({3,4,5}), d({4,5,6});
a = b + c + d;
std::cout << b << std::endl;
std::cout << c << std::endl;
std::cout << d << std::endl;
std::cout << "Result:\n" << a << std::endl;
return 0;
}
Output:
[1, 2, 3, ]
[3, 4, 5, ]
[4, 5, 6, ]
Result:
[8, 11, 14, ]
I've added an initializer_list (C++11) constructor and ostream operators purely for convenience/illustration purposes.
Since you've defined the operator[] as return by value, I was unable to set items in the data array for testing (since error: lvalue required as left operand of assignment); typically this operator should be by reference - but then, in your case, VectorSum::operator[] wouldn't work because that would fail compilation because of returning a reference to a temporary.
I also added a copy constructor because ...
// this calls MyVector's copy constructor when assigned to 'main::a'
template<class T, int N>
MyVector<T,N> operator+(const MyVector<T,N>& a, const MyVector<T,N>& b) {
return VectorSum<T,N>(a, b); // implicit MyVector::copy constructor
}
// this also calls MyVector's copy constructor (unless the copy constructor is defined explicit)
template<class T, int N>
MyVector<T,N> operator+(const MyVector<T,N>& a, const MyVector<T,N>& b) {
MyVector<T,N> res = VectorSum<T,N>(a, b);
return res;
}
// but this would call MyVector's assignment operator
template<class T, int N>
MyVector<T,N> operator+(const MyVector<T,N>& a, const MyVector<T,N>& b) {
MyVector<T,N> res;
res = VectorSum<T,N>(a, b);
return res;
}
In answer to your questions:
Yes - how would it behave if you explicitly defined the variable and
returned that? It's the same behaviour for temporaries, except there
is no variable declaration;
n/a
I touched on this above - you can't
use reference because of 'returning reference to temporary' error;
However there's no reason why you can add T& operator[] to MyVector( ie. not overriding).
EDIT: answers to comments:
The function specification must be identical when overriding including return type. Since you have defined it return by value in VectorExpression it must be return by value in MyVector. If you try to change it to reference in the child class you will get a compile error: conflicting return type specified. So no you can't override operator const with a version that returns const T& instead of T. Furthermore it must return by value since MyVectorSum returns { a[i] + b[i] } which would be a temporary and you can't return a reference to a temporary.
Sorry my mistake, fixed above.
because:
MyVector isn't a subtype of VectorSum - compile error ‘MyVector’ is not derived from ‘const VectorSum’
I've also tried with VectorExpression but compile error: 'cannot allocate an object of abstract type' - because it's trying to return by value
I chose MyVector since that's the type of your expected result. Yes it does those all those for loops but I can't see a way round that : there's three different array 'data' variables each of which need to be iterated into order to be accumulated. At some point in the code you will have to do the for loops.
Understood, yes I got confused. removed from post.

I didn't think of this at first but having a virtual operator[] method probably kills the efficiency I was trying to achieve by avoiding the 3 for-loops and intermediate storage of vector-sized temporaries. Making a method virtual prevents it from being inlined, which means it needs to be actually called as a function everytime an element is accessed.
Based on links I got from people who commented to my question and from Google, I ended up with the following solution which avoids needing any virtual methods.
template<class T, int N, class V>
class VectorExpressionBase {
V ref;
protected:
explicit VectorExpressionBase(const V *ref)
: ref(const_cast<V*>(ref)) {}
public:
T operator[] (int i) const { return ref[i]; }
T& operator[] (int i) { return ref[i]; }
};
template<class T, int N>
class VectorExpressionBase<T,N,void> {
T data[N];
protected:
explicit VectorExpressionBase(const void*) {
// Argument is unused but accepted to have uniform
// calling syntax
}
public:
T operator[] (int i) const { return data[i]; }
T& operator[] (int i) { return data[i]; }
};
template<class T, int N, class V>
class VectorExpression : public VectorExpressionBase<T,N,V> {
public:
template<class V1>
VectorExpression<T,N,V>& operator= (
const VectorExpression<T,N,V1> &rhs)
{
for (int i = 0; i < N; ++i)
data[i] = rhs[i];
return *this;
}
explicit VectorExpression(const V *ref = 0)
: VectorExpressionBase<T,N,V>(ref) {}
// Can define all kinds of operators and functions here such as
// +=, *=, unary + and -, max(), min(), sin(), cos(), ...
// They would automatically apply to MyVector objects and to the
// results of other operators and functions
};
template<class T, int N>
class MyVector : public VectorExpression<T,N,void> {
};
template<class T, int N, class VA, class VB>
class VectorSum {
VectorExpression<T,N,VA> &a;
VectorExpression<T,N,VB> &b;
public:
VectorSum(VectorExpression<T,N,VA> &aa, VectorExpression<T,N,VB> &bb)
: a(aa), b(bb) {}
T operator[] (int i) const { return a[i] + b[i]; }
};
template<class T, int N, class VA, class VB>
VectorExpression<T,N,VectorSum<T,N,VA,VB> >
operator+(const VectorExpression<T,N,VA> &a,
const VectorExpression<T,N,VB> &b)
{
VectorSum<T,N,VA,VB> sum(a, b);
return VectorExpression<T,N,VectorSum<T,N,VA,VB> >(sum);
}
Class VectorExpression now just wraps the class that does the work (in this case VectorSum). This allows defining all kinds of functions and operators for VectorExpression only, rather than having to overload them for VectorSum, VectorProduct, etc.
MyVector derives from a special case of VectorExpression which has a specialized base class; this isn't really necessary but it's nice because it makes all functions and operators defined for VectorExpression also available for MyVector. By using a simple base class VectorExpressionBase that only deals with storage and the [] operator, all other operators and methods don't need to be duplicated in the specialization for V = void.
Users would only need to know about classes MyVector<T,N> (for storing data) and possibly about VectorExpression<T,N,V> if they want to define additional functions and operators. VectorExpressionBase and the classes like VectorSum don't need to be visible to the outside world.
I find my original solution somewhat cleaner conceptually because the meaning of each class is more clear and because it doesn't require the template parameter V, but this one is more efficient because it doesn't require any virtual functions, which can probably make a large difference in certain cases.
Thanks for pointing me to the right links!
P.S. Surely most / all of this isn't new but I thought it would be nice to summarize and explain it a bit. I hope it can help others.
EDIT
I changed the type of data member VectorExpressionBase<T,N,V>::ref from V& to V. This is needed since the temporary V object the reference was pointing at may no longer exist at the time the VectorExpression is evaluated. For example, the temporary VectorSum object stops existing when the operator+ function returns, making the returned VectorExpression object useless.
I also completed the code with some constructors and corrected the operator+ function.

Related

How correct is auto & variadic template to interface underlying data containers methods

Here is some context :
I need to overload some operators to act on std::vector operands. I decided against implementing directly the operator overloads in the global scope, as I believe it would be a bad practice due to possible conflicts. As a result, I decided to implement a templated class which :
Behaves as much as possible like a std::vector
Allows the overloading of the operator in a transparent way
The initial idea was to inherit from std::vector but I quickly realized it is not recommended. I therefore decided to create a class which holds a std::vector data member and implements an interface that makes it as transparent to the user as possible.
Here is a snippet of what I came up with :
template<typename T>
class Vector {
public:
Vector() = default;
~Vector() = default;
Vector(std::initializer_list<T> list) :data(list) {}
template<typename... Args>
Vector(Args... args) : data(args...){}
/* All std::vector functions with parameters and return value will look like this*/
template<typename...Args>
auto reserve(Args... args) {
return data.reserve(args...);
}
/* All std::vector functions with parameters and void return value will look like this*/
template<typename... Args>
void resize(Args... args) {
data.resize(args...);
}
/* All std::vector functions without parameter and void return value will look like this*/
void clear() {
data.clear();
}
/* All std::vector functions with return value will look like this*/
Vector& operator=(Vector const & other) {
data = other.getData();
return *this;
}
size_t size() const {
return data.size();
}
std::vector<T> & getData() {
return data;
}
std::vector<T> const & getData() const {
return data;
}
T & operator[](size_t i) {
return data[i];
}
T const & operator[](size_t i) const {
return data[i];
}
/* All std::vector functions without args and with return value will look like this*/
auto begin() {
return data.begin();
}
private:
std::vector<T> data;
};
And the overloaded operators would be written like :
template <typename T>
void operator -= (Vector<T>& a, const Vector<T>& b) {
if (a.size() != b.size()) {
throw std::length_error("void operator -= (Vector<T>& a, const Vector<T>& b)\nBoth Vector<T> operands should have the same size");
}
for (int i = 0; i < a.size(); ++i) {
a[i] -= b[i];
}
}
My question is the following : how viable is this strategy ? Is there something inherently wrong with that ?
I do feel that there is quite a big amount of boilerplate code (all functions with arguments and with return value will look in a certain way, and only the name of the functions that will be passed to the data member will change), do you see a potential improvement possible there ?
The simplest and easiest way to achieve what you want is to implement a free function :
template <class T>
void minusEqual(std::vector<T>& target, const std::vector<T>& param)
{
if (a.size() != b.size()) {
throw std::length_error("void operator -= (Vector<T>& a, const Vector<T>& b)\nBoth Vector<T> operands should have the same size");
}
for (int i = 0; i < a.size(); ++i) {
a[i] -= b[i];
}
}
Of course, in order to prevent other contributors from duplicating it, you need to store it in a clear way so that they will know where to find it.
Although it is usually discouraged, you can also inherit from std::vector. This question and its answer will give you useful information about that.

overload multidimensional array operator for vector

hi i am trying to create matrix class and i want to assign like matrix[0][2]=3
i checked form and found a solution with array but i want to do with vector and could not understand why does not it works ?
template<class T>
class Matrix
{
public:
Matrix(int a, int b)
{
vector<vector<T> > vec( a , vector<T> (b, 0));
matrixData = vec;
}
class Array1D
{
public:
Array1D(vector<T> a):temp(a) {}
T& operator[](int a)
{
return temp[a];
}
vector<T> temp;
};
vector<vector<T> > matrixData;
Array1D operator[] (int a)
{
return Array1D(matrixData[a]);
}
};
int main()
{
Matrix<int> n(3,5);
n[0][2] = 123; //assign
cout<<n[0][2]; // wrong output getting 0
}
You have to change Array1D implementation to:
class Array1D
{
public:
Array1D(vector<T>& a):temp(a) {}
T& operator[](int a)
{
return temp[a];
}
vector<T>& temp;
};
without this, every time you call operator[] on matrix you access temporary vector. Hence, every call n[0] works on different vector. So any previous modifications cannot be saved, and you always see 0 as result.
With the above change, you access original vector of matrix by Array1D proxy class.
Demo
You are returning the wrong type from the Matrix::operator[]. You need to return a nested vector by reference, so that you can chain the next [] to it:
vector<T>& operator[] (int a)
{
return matrixData[a];
}
In fact, you don't need the internal Array1D class at all for this purpose, since vector already has an operator[], and you can remove it entirely.
Here's a working demo.

c++ - cannot use overloaded operator defined before in template class member function

I defined a Point<T, Size> template class, which has a std::array as its member variable. I overload [] operator to conveniently access the array, however, I found that in some following member functions I can't use [] operator and some others otherwise.
I was thinking if this is a template-class-related issue?
Note: Vector is also a template class similar to Point and I think its definition is not related to the error. I'll post its code here if further needed.
template<typename T, unsigned Size>
class Point {
private:
std::array<T, Size> _p;
public:
static const unsigned size = Size;
inline T &operator[](unsigned i) { return _p[i]; }
inline Point operator+(const Point &other) {
Point ret;
for (int i = 0; i < Size; i++) {
ret._p[i] = _p[i] + other[i]; // OK to directly access _p with [] operator
}
return ret;
}
inline Vector<T, Size> operator-(const Point &other) {
Vector<T, Size> ret;
for (int i = 0; i < Size; i++) ret[i] = _p[i] - other._p[i]; // compiler error if use other[i] directly
return ret;
}
};
The only difference between the two commented functions, I think, is the return type. Can anyone tell me why the usage of [] operator is different?
error C2678: binary '[': no operator found which takes a left-hand operand of type 'const Point<float,3>' (or there is no acceptable conversion)
The problem is that your operator [] isn't available for const Points. Usually, when providing operator [] you want to provide it in pairs:
inline T &operator[](unsigned i) { return _p[i]; }
inline T const &operator[](unsigned i) const { return _p[i]; }
As to why this is working in your operator + member function, you're probably never using it.

Inheritance of overloaded + operator

I have a problem in inheriting overloaded + operator.
Let me make an example.
class Data{
protected:
int data[3];
public:
Data(){
data[0] = data[1] = data[2] = 0;
}
Data operator+(const Data& other)
{
Data temp = *this;
for(int i=0;i<3;i++){
temp.data[i] += other.data[i]
}
return temp;
}
};
class DataInterited:public Data{
public:
};
/******************Main*****************/
DataInterited d1,d2,d3;
d3 = d1 + d2; //=> This is compile error
This code generate compile error saying,
no match for ‘operator=’ (operand types are ‘DataInterited’ and ‘Data’)
I think I have to implement operator+ for DataInherited so that it return DataInherited instance. But in this way, I cannot avoid code duplication.
Is there any way to make d3=d1+d2; line correct while avoiding duplicating the + operator implementation?
There are a couple of things you need to know.
First, always implement operator+ as a free function in terms of operator+=. It saves code duplication and is optimally efficient.
Second, you had no constructor in DataInherited that could take a Data as its argument. This is important because the result of Data::operator+ is a Data, not a DataInherited.
corrected code:
#include <iostream>
#include <algorithm>
class Data{
protected:
int data[3];
public:
Data(){
data[0] = data[1] = data[2] = 0;
}
Data(const Data& other)
{
std::copy(std::begin(other.data), std::end(other.data), data);
}
Data& operator=(const Data& other)
{
std::copy(std::begin(other.data), std::end(other.data), data);
return *this;
}
Data& operator+=(const Data& other)
{
for(int i=0;i<3;i++){
data[i] += other.data[i];
}
return *this;
}
};
Data operator+(Data left, const Data& right)
{
return left += right;
}
class DataInterited:public Data{
public:
DataInterited(Data d = {})
: Data(std::move(d))
{}
};
using namespace std;
auto main() -> int
{
DataInterited d1,d2,d3;
d3 = d1 + d2; //=> This is no longer a compile error
return 0;
}
Koenig operator forwarding to an increment_by function.
Derived classes can implement their own increment_by overloads if they want different behavior.
SFINAE stuff skipped, so bad types will give hard errors.
class Data{
public:
template<class D, class Rhs>
friend D operator+=(D&& lhs, Rhs&& rhs){
increment_by(lhs,std::forward<Rhs>(rhs));
return std::forward<D>(lhs);
}
template<class Lhs, class Rhs>
friend Lhs operator+(Lhs lhs, Rhs&& rhs){
lhs+=std::forward<Rhs>(rhs);
return std::move(lhs);
}
friend void increment_by(Data& self, Data const&other){
for(int i=0;i<6;i++){
self.data[i] += other.data[i];
}
}
};
Both + and += are template friends and hence the types passed can be derived classes. So type isn't lost,
increment_by needs overiding if derived type needs new behaviour. If not, leave it alone.
live example.
Do not leave the type needlessly. Converting from base to derived basically throws out the point of the derived type.

templated code not working while non-templated code works!

I spent several hours, completely stuck when I realized that only the templated version of my code has a bug.
In the following code, when pushing_back elements in the myMap, the original vectors myVec1 and myVec2 are modified and contain garbage at the end of the execution. If I un-template everything, just replacing template<T> by double, then the code works fine as I would expect (the original arrays are untouched).
The funny thing is if I put a cout in the copy constructor, it does not get called if the code is templated. But it gets called if I replace the copy constructor with Vector<T2> by the original type Vector<T>, and then everything work fine.
Why wouldn't the compiler know that T2==T since I only use double?
(note, the code has been made as short as possible so as to show the errors - I thus removed accessors, made everything public etc.).
#include <vector>
#include <map>
template<class T>
class Vector{
public:
Vector():n(0),data(0){};
Vector(int N):n(N),data(new T[N]){};
Vector(T x, T y):n(2),data(new T[2]){data[0]=x; data[1]=y;};
template<class T2> Vector(const Vector<T2>& rhs):n(rhs.n), data(new T[n])
{
for (int i=0; i<n; i++)
data[i] = T(rhs.data[i]);
}
~Vector(){delete[] data;}
Vector& operator=(const Vector& rhs)
{
if (rhs.n != n)
{
if (data)
delete[] data;
data = new T[rhs.n];
}
n = rhs.n;
memcpy(data, rhs.data, n*sizeof(T));
return *this;
}
T& operator[](int i){return data[i];}
const T& operator[](int i) const {return data[i];}
int n;
T* data;
};
typedef Vector<double> Vectord;
template <class T> inline bool operator<(const Vector<T>& v1, const Vector<T>& v2)
{
for (int i=0; i<v1.n; i++)
{
if (v1[i]<v2[i]) return true;
if (v1[i]>v2[i]) return false;
}
return false;
}
int main(int argc, char** argv)
{
std::vector<Vectord> myVec1(3);
myVec1[0] = Vectord(1.,3.);
myVec1[1] = Vectord(3.,3.);
myVec1[2] = Vectord(1.,5.);
std::vector<Vectord> myVec2(3);
myVec2[0] = Vectord(4.,1.);
myVec2[1] = Vectord(2.,5.);
myVec2[2] = Vectord(6.,5.);
std::map<Vectord, std::vector<Vectord> > myMap;
for (int i=0; i<3; i++)
{
myMap[myVec1[i]].push_back(myVec2[i]);
}
return 0;
}
A templated constructor is never a copy constructor.
So your class is using the automatically generated copy constructor.
Cheers & hth.,