Re: vectorization issue
In my application, I have the following pattern occurring at many places:
loop_over_i(i)
{
do_something(i); // common for all loops
do_remaining_things(i);
}
The do_something part is common for most loops and so I am trying to extract it out with a class. However, the loop no longer vectorizes. A sample code is given below:
#include <iostream>
using namespace std;
class Loop
{
public:
Loop(int max) : max_(max), count_(0), sqr_(0)
{ }
void more()
{
++count_ ;
sqr_ = count_ * count_;
}
bool done() const
{
return count_ >= max_;
}
operator int() const
{
return count_;
}
int sqr() const
{
return sqr_;
}
private:
int count_, max_, sqr_;
};
int main()
{
int square[50];
// this loop vectorizes
for (int i = 0; i < 50; ++i)
{
square[i] = i*i;
}
// this loop doesn't vectorize
for (Loop loop(50); !loop.done(); loop.more())
{
square[loop] = loop.sqr();
}
}
How can I implement what I need in a way that the loop vectorizes?
When indexing square[i], elements i and i+1, i+2, ... are in adjacent memory locations, thus permitting vectorization. The compiler vectorization code will recognize loop syntax such as i<50 and ++i and can analyse that for candidate of vectorization.
When you declare the class Loop, and declare the member functions, the compiler does not have the necessary information in order to vectorize the loop. The compiler _potentially_ could vectorize the code if you inline the member functions. But having the !loop.done() may present a problem.
Also note that your more() function is computing the result for the element after the last element. Consider converting the more() to return a bool and remove the done function.
for(Loop loop(50); loop.more();)
where
inline bool more()
{
if(count_ >= max_) return false;
sqr_ = count_ * count_;
++count_;
return true;
}
This syntax will not necessarily vectorize. It would depend on how aggressive the the compiler was at vectorization.
Jim Dempsey
Blog: The Parallel Void
www.quickthreadprogramming.com | 