Operator overloading, C++ performance crappiness

[Jojo]

Is there any way to get to the left-hand side of an operator? Consider
the following (this is not meant to be perfect code, just an example of
the problem):

class Matrix
{
public:
int data[1024];

Matrix() {}

Matrix(int value)
{
for (unsigned i = 0; i < sizeof(data)/sizeof(int); i++)
data = value;
}

void add(const Matrix& obj, Matrix* output)
{
for (unsigned i = 0; i < sizeof(data)/sizeof(int); i++)
output->data = data + obj.data;
}

Matrix operator +(const Matrix& obj)
{
Matrix temp; // "unnecessary" creation of temp variable

for (unsigned i = 0; i < sizeof(data)/sizeof(int); i++)
temp.data = data + obj.data;

return temp; // "unnecessary" extra copy of output
}
};

For nice looking syntax you _really_ want to use the operator+ like:
matrix3 = matrix1 + matrix2;

However, that is some 50% slower than the _much_ uglier:
matrix1.add(matrix2, &matrix3);

If only there were a way to get to the left-hand argument of the
operator+ then it could be fast and easy to use. Consider the following
code which is not valid C++ and will not compile for this example:

Matrix as M
operator+(const Matrix& obj)
{
for (unsigned i = 0; i < sizeof(data)/sizeof(int); i++)
M.data = data + obj.data;
}

That would be fast and clean to use. Is there any way to accomplish
this? Otherwise the situation is just ugly and there is no point in
using operator overloading for these types of situations (which really
defeats the purpose of operator overloading in the first place).

Thanks! Jo

 

[Christian Meier]

Is there any way to get to the left-hand side of an operator? Consider
the following (this is not meant to be perfect code, just an example of
the problem):

class Matrix
{
public:
int data[1024];

Matrix() {}

Matrix(int value)
{
for (unsigned i = 0; i < sizeof(data)/sizeof(int); i++)
data = value;
}

void add(const Matrix& obj, Matrix* output)
{
for (unsigned i = 0; i < sizeof(data)/sizeof(int); i++)
output->data = data + obj.data;
}

Matrix operator +(const Matrix& obj)
{
Matrix temp; // "unnecessary" creation of temp variable

for (unsigned i = 0; i < sizeof(data)/sizeof(int); i++)
temp.data = data + obj.data;

return temp; // "unnecessary" extra copy of output
}
};

For nice looking syntax you _really_ want to use the operator+ like:
matrix3 = matrix1 + matrix2;

However, that is some 50% slower than the _much_ uglier:
matrix1.add(matrix2, &matrix3);

If only there were a way to get to the left-hand argument of the
operator+ then it could be fast and easy to use. Consider the following
code which is not valid C++ and will not compile for this example:

Matrix as M
operator+(const Matrix& obj)
{
for (unsigned i = 0; i < sizeof(data)/sizeof(int); i++)
M.data = data + obj.data;
}

That would be fast and clean to use. Is there any way to accomplish
this? Otherwise the situation is just ugly and there is no point in
using operator overloading for these types of situations (which really
defeats the purpose of operator overloading in the first place).

Thanks! Jo

You could just use operator+=
matrix1 += matrix2;
matrix1 += matrix3;

or rewrite add()
matrix1.add(matrix2);
matrix1.add(matrix3);

or you could write the function add using ellipses:
add(...);

matrix1.add(&matrix2, &matrix3, &matrix4, &matrix5);

Greets Chris

 

[Jojo]

You could just use operator+=
matrix1 += matrix2;
matrix1 += matrix3;

or rewrite add()
matrix1.add(matrix2);
matrix1.add(matrix3);

or you could write the function add using ellipses:
add(...);

matrix1.add(&matrix2, &matrix3, &matrix4, &matrix5);

Greets Chris

"+=" does not accomplish the same thing, and neither does add(Matrix).
There is a third variable involved which is the result of adding two
other variables that you don't want to modify.

As I mentioned the "matrix.add()" syntax certainly works and it is fast
but it is extremely awkward to use and makes for some nasty looking code.

Jo

 

[benben]

Is there any way to get to the left-hand side of an operator? Consider

the following (this is not meant to be perfect code, just an example of
the problem):

class Matrix
{
public:
int data[1024];

Matrix() {}

Matrix(int value)
{
for (unsigned i = 0; i < sizeof(data)/sizeof(int); i++)
data = value;
}

void add(const Matrix& obj, Matrix* output)
{
for (unsigned i = 0; i < sizeof(data)/sizeof(int); i++)
output->data = data + obj.data;
}

Matrix operator +(const Matrix& obj)
{
Matrix temp; // "unnecessary" creation of temp variable

for (unsigned i = 0; i < sizeof(data)/sizeof(int); i++)
temp.data = data + obj.data;

return temp; // "unnecessary" extra copy of output
}
};

For nice looking syntax you _really_ want to use the operator+ like:
matrix3 = matrix1 + matrix2;

However, that is some 50% slower than the _much_ uglier:
matrix1.add(matrix2, &matrix3);

If only there were a way to get to the left-hand argument of the operator+
then it could be fast and easy to use. Consider the following code which
is not valid C++ and will not compile for this example:

Matrix as M
operator+(const Matrix& obj)
{
for (unsigned i = 0; i < sizeof(data)/sizeof(int); i++)
M.data = data + obj.data;
}

That would be fast and clean to use. Is there any way to accomplish this?
Otherwise the situation is just ugly and there is no point in using
operator overloading for these types of situations (which really defeats
the purpose of operator overloading in the first place).

Thanks! Jo

If that code is slow, that's because it is not well written. In production
code you probably don't want to return a full-blown matrix object from
operator + because that'd make a temporary. Instead, a small object of some
other type is returned that records the operation and operands. The
evaluation happens eventually in the assignment operation.

class MatrixOp;

MatrixOp operator + (const Matrix&, const Matrix);
Matrix operator = (Matrix&, const MatrixOp);

Ben

 

[Christian Meier]

"+=" does not accomplish the same thing, and neither does add(Matrix).
There is a third variable involved which is the result of adding two
other variables that you don't want to modify.

Why not? After your statements an object has the value of the sum of two
others.
obj1 = obj2 + obj3;
obj1 += obj2;
obj1 += obj3;

If you need to reset obj1 first, you can write a reset() function. Or assign
0 before the +=operator calls.

As I mentioned the "matrix.add()" syntax certainly works and it is fast
but it is extremely awkward to use and makes for some nasty looking code.

Jo

Is sizeof(data)/sizeof(int) not the same for these three objects?
If not, then you have problems with your function add():

void add(const Matrix& obj, Matrix* output)
{
for (unsigned i = 0; i < sizeof(data)/sizeof(int); i++)
output->data = data + obj.data;
}

You do no range checking here for output->data and obj.data.

Greets Chris

 

[Jojo]

If that code is slow, that's because it is not well written. In production
code you probably don't want to return a full-blown matrix object from
operator + because that'd make a temporary. Instead, a small object of some
other type is returned that records the operation and operands. The
evaluation happens eventually in the assignment operation.

class MatrixOp;

MatrixOp operator + (const Matrix&, const Matrix);
Matrix operator = (Matrix&, const MatrixOp);

Ben

You are still going to incur a performance penalty from copying data
into the MatrixOp variable. This penalty is certainly much smaller than
that of copying the objects in my example but as I said in the beginning
my example is to illustrate the problem and not be an example of perfect
code used in production.

My very last pseudo code for operator+ would still be faster than using
a MatrixOp temp variable and the "Matrix.add(const Matrix&, Matrix*
output)" is faster as well (although extemely ugly as I already mentioned).

Jo

 

[Kyle]

Is there any way to get to the left-hand side of an operator? Consider
the following (this is not meant to be perfect code, just an example of
the problem):

class Matrix
{
public:
int data[1024];

Matrix() {}

Matrix(int value)
{
for (unsigned i = 0; i < sizeof(data)/sizeof(int); i++)
data = value;
}

void add(const Matrix& obj, Matrix* output)
{
for (unsigned i = 0; i < sizeof(data)/sizeof(int); i++)
output->data = data + obj.data;
}

Matrix operator +(const Matrix& obj)
{
Matrix temp; // "unnecessary" creation of temp variable

for (unsigned i = 0; i < sizeof(data)/sizeof(int); i++)
temp.data = data + obj.data;

return temp; // "unnecessary" extra copy of output
}
};

For nice looking syntax you _really_ want to use the operator+ like:
matrix3 = matrix1 + matrix2;

However, that is some 50% slower than the _much_ uglier:
matrix1.add(matrix2, &matrix3);

If only there were a way to get to the left-hand argument of the
operator+ then it could be fast and easy to use.

you would want to get to left hand argument of operator=, but then how
do you know its there ?

what about:

foo( a + b );

operator+ would need to know about the context its used in than ...

Consider the following
code which is not valid C++ and will not compile for this example:

Matrix as M
operator+(const Matrix& obj)
{
for (unsigned i = 0; i < sizeof(data)/sizeof(int); i++)
M.data = data + obj.data;
}

That would be fast and clean to use. Is there any way to accomplish
this? Otherwise the situation is just ugly and there is no point in
using operator overloading for these types of situations (which really
defeats the purpose of operator overloading in the first place).

Thanks! Jo

 

[benben]

You are still going to incur a performance penalty from copying data into

the MatrixOp variable. This penalty is certainly much smaller than that
of copying the objects in my example but as I said in the beginning my
example is to illustrate the problem and not be an example of perfect code
used in production.

MatrixOp can be constructed without copying any matrix at all. A reference
or pointer will do.


Ben

 

[Jojo]

MatrixOp can be constructed without copying any matrix at all. A reference
or pointer will do.


Ben

I didn't say anything to the contrary. The "add(const Object& obj,
Object* output)" method in my example is still faster. MatrixOp adds
overhead no matter how you look at it. Like I said before, it's a
smaller overhead than copying a large object but it _does_ add overhead.

Consider if we're working with millions of "Vector" objects that only
have two or three float members. MatrixOp would have about the same
complexity as the Vector object itself. There would be no benefit to
using it and it would be slower than the ever so ugly:

void add(const Vector& obj, Vector* output)
{
*output = obj;
}

Jo

 

[Jojo]

Why not? After your statements an object has the value of the sum of two
others.
obj1 = obj2 + obj3;
obj1 += obj2;
obj1 += obj3;

That is not going to be any faster than the plain slow operator+ in my
example. First you "add" (copy) obj2 into obj1, then you add obj3.
This is slower than just adding two variables straight out into a third.

If you need to reset obj1 first, you can write a reset() function. Or assign
0 before the +=operator calls.

Again, another performance hit that is going to be slower than just
adding two objects straight out into a (uninitialized) third.

Jo

 

[Christian Meier]

obj1 += obj2;

That is not going to be any faster than the plain slow operator+ in my
example. First you "add" (copy) obj2 into obj1, then you add obj3.
This is slower than just adding two variables straight out into a third.

Who mentioned the word copy?
Matrix& Matrix:perator+=(const Matrix& rMatrix)
{}

No copy is created.

 

[Jojo]

you would want to get to left hand argument of operator=, but then how
do you know its there ?

Good point but the compiler can handle that. If no left-hand variable
is present then the compiler would generate a transient temp variable of
the return type.

what about:

foo( a + b );

operator+ would need to know about the context its used in than ...

That is still a lot uglier than "var3 = var1 + var2".

Jo

 

[Maxim Yegorushkin]

Jojo wrote:

[]

That would be fast and clean to use. Is there any way to accomplish
this? Otherwise the situation is just ugly and there is no point in
using operator overloading for these types of situations (which really
defeats the purpose of operator overloading in the first place).

Use expression templates. Refer to
http://osl.iu.edu/~tveldhui/papers/Expression-Templates/exprtmpl.html

 

[Jojo]

Who mentioned the word copy?
Matrix& Matrix:perator+=(const Matrix& rMatrix)
{}

No copy is created.

LOL... That's why I said "add (copy)". The first addition is equivalent
to a copy. See:

1. obj1 is in "emtpy" state (this has overhead because it would need to
be initialied to all 0).

2. obj1 += obj2. This is essientially copying obj2 into obj1 because
obj1 now has the value of obj2. Overhead again.

3. obj1 += obj3. Now you finally do the operation you want where you
add obj2 (remember obj1==obj2 at this point) to obj3. This is really
were all the work is and the only required overhead.

That is slower than:

1. obj1 in completely uninitialized or otherwise unknown state (no overhead)

2. Directly set obj1 to the value of obj2 + obj3

Jo

 

[Karl Heinz Buchegger]

Is there any way to get to the left-hand side of an operator? Consider
the following (this is not meant to be perfect code, just an example of
the problem):

You must have a compiler which is really bad at optimization.
In the following code, the timing is as follows:

operator+: 180 clock ticks
function add(): 240 clock ticks

So (thanks to the optimizer, which optimizes away the total overhead
of the temporary) the code using operator+ is actually *faster* then
your specialized function.

#include <iostream>
#include <ctime>

using namespace std;

class Matrix
{
public:
int data[1024];

Matrix() {}
Matrix(int value)
{
for (unsigned i = 0; i < sizeof(data)/sizeof(int); i++)
data = value;
}

void add(const Matrix& obj, Matrix& output)
{
for (unsigned i = 0; i < sizeof(data)/sizeof(int); i++)
output.data = data + obj.data;
}

friend Matrix operator+( const Matrix& lhs, const Matrix& rhs );
};

inline Matrix operator +( const Matrix& lhs, const Matrix& rhs )
{
Matrix temp; // "unnecessary" creation of temp variable

for (unsigned i = 0; i < sizeof(lhs.data)/sizeof(int); i++)
temp.data = lhs.data + rhs.data;

return temp;
}

int main()
{
Matrix a(2), b(3), c;
time_t start, end;

start = clock();
for( int j = 0; j < 100000; ++j )
c = a + b;
end = clock();

cout << end - start << endl;

start = clock();
for( int j = 0; j < 100000; ++j )
a.add( b, c );
end = clock();

cout << end - start << endl;
}

class Matrix
{
public:
int data[1024];

Matrix() {}

Matrix(int value)
{
for (unsigned i = 0; i < sizeof(data)/sizeof(int); i++)
data = value;
}

void add(const Matrix& obj, Matrix* output)
{
for (unsigned i = 0; i < sizeof(data)/sizeof(int); i++)
output->data = data + obj.data;
}

Matrix operator +(const Matrix& obj)
{
Matrix temp; // "unnecessary" creation of temp variable

for (unsigned i = 0; i < sizeof(data)/sizeof(int); i++)
temp.data = data + obj.data;

return temp; // "unnecessary" extra copy of output
}
};

For nice looking syntax you _really_ want to use the operator+ like:
matrix3 = matrix1 + matrix2;

However, that is some 50% slower than the _much_ uglier:
matrix1.add(matrix2, &matrix3);

If only there were a way to get to the left-hand argument of the
operator+ then it could be fast and easy to use. Consider the following
code which is not valid C++ and will not compile for this example:

Matrix as M
operator+(const Matrix& obj)
{
for (unsigned i = 0; i < sizeof(data)/sizeof(int); i++)
M.data = data + obj.data;
}

That would be fast and clean to use. Is there any way to accomplish
this? Otherwise the situation is just ugly and there is no point in
using operator overloading for these types of situations (which really
defeats the purpose of operator overloading in the first place).

Thanks! Jo

--
Karl Heinz Buchegger, GASCAD GmbH
Teichstrasse 2
A-4595 Waldneukirchen
Tel ++43/7258/7545-0 Fax ++43/7258/7545-99
email: (e-mail address removed) Web: www.gascad.com

Fuer sehr grosse Werte von 2 gilt: 2 + 2 = 5

 

[Christian Meier]

LOL... That's why I said "add (copy)". The first addition is equivalent
to a copy. See:

1. obj1 is in "emtpy" state (this has overhead because it would need to
be initialied to all 0).

2. obj1 += obj2. This is essientially copying obj2 into obj1 because
obj1 now has the value of obj2. Overhead again.

3. obj1 += obj3. Now you finally do the operation you want where you
add obj2 (remember obj1==obj2 at this point) to obj3. This is really
were all the work is and the only required overhead.

That is slower than:

1. obj1 in completely uninitialized or otherwise unknown state (no overhead)

2. Directly set obj1 to the value of obj2 + obj3

Jo

aaah, now I start to understand your "copy"
OK, back to the beginning.
Your function add() is not totally bad. But the name "add" is a bit
confusing with your two parameters. What about something like "setToSum"?
Now to your parameters. Why don't use use *this for the object to be set? Is
there a reason why *this has to be unchanged?

void setToSum(const Matrix& first, const Matrix& second)
{
for (unsigned i = 0; i < sizeof(data)/sizeof(int); ++i)
data = first.data + second.data;
}

This is perhaps a bit cleaner.

Greets Chris

 

[Jojo]

You must have a compiler which is really bad at optimization.
In the following code, the timing is as follows:

operator+: 180 clock ticks
function add(): 240 clock ticks

So (thanks to the optimizer, which optimizes away the total overhead
of the temporary) the code using operator+ is actually *faster* then
your specialized function.

Yes a good compiler could do that. I have not found one than can.

So the big question is what compiler are you using?

I'm using GCC (I tried versions 3.3 and 4.0).

Jo

 

[Jojo]

aaah, now I start to understand your "copy"
OK, back to the beginning.
Your function add() is not totally bad. But the name "add" is a bit
confusing with your two parameters. What about something like "setToSum"?
Now to your parameters. Why don't use use *this for the object to be set? Is
there a reason why *this has to be unchanged?

void setToSum(const Matrix& first, const Matrix& second)
{
for (unsigned i = 0; i < sizeof(data)/sizeof(int); ++i)
data = first.data + second.data;
}

This is perhaps a bit cleaner.

Greets Chris

True, but you can name it anything and the code will still look really
bad when things get complicated.

Nothing works as clean as "var3 = var1 + var2". There just doesn't seem
to be a way to do that efficiently in C++ (of course unless you have a
really good compiler as mentioned in another post; show me that compiler!).

Jo

 

[Christian Meier]

aaah, now I start to understand your "copy"
OK, back to the beginning.
Your function add() is not totally bad. But the name "add" is a bit
confusing with your two parameters. What about something like "setToSum"?
Now to your parameters. Why don't use use *this for the object to be set? Is
there a reason why *this has to be unchanged?

void setToSum(const Matrix& first, const Matrix& second)
{
for (unsigned i = 0; i < sizeof(data)/sizeof(int); ++i)
data = first.data + second.data;
}

This is perhaps a bit cleaner.

Greets Chris

True, but you can name it anything and the code will still look really
bad when things get complicated.

Also true, but a bit better than add(const Matrix&, Matrix*);

Nothing works as clean as "var3 = var1 + var2". There just doesn't seem
to be a way to do that efficiently in C++ (of course unless you have a
really good compiler as mentioned in another post; show me that

compiler!).

IMHO optimize away a temporary return value is support by many compilers....

 

[Jojo]

inline Matrix operator +( const Matrix& lhs, const Matrix& rhs )
{
Matrix temp; // "unnecessary" creation of temp variable

for (unsigned i = 0; i < sizeof(lhs.data)/sizeof(int); i++)
temp.data = lhs.data + rhs.data;

return temp;
}

OK, _this_ is the answer I think I have been looking for. This method
is not the method I posted and this is indeed faster in GCC.

My method was:
Matrix operator+(const Matrix&)

Your method is:
operator+(const Matrix&, const Matrix&)

I did not know there was a two parameter version of operator+.

Thanks! Jo