I am working on my first CUDA program and running into error using the nvcc compiler that I do not encounter if I compile with g++.
My code:
#include <iostream>
#include <cmath>
using namespace std;
double distance(double first, double second);
int main(){
double dis;
dis = distance(7.0, 1.0);
cout << "distance = " << dis << endl;
return 0;
}
double distance(double first, double second){
double diff;
diff = abs(first-second);
return diff;
}
If I compile with nvcc test.cu -o test, the result is:
/usr/include/c++/5/bits/stl_iterator_base_types.h(168): error: name followed by "::" must be a class or namespace name
detected during instantiation of class "std::iterator_traits<_Iterator> [with _Iterator=double]"
test.cu(11): here
/usr/include/c++/5/bits/stl_iterator_base_types.h(169): error: name followed by "::" must be a class or namespace name
detected during instantiation of class "std::iterator_traits<_Iterator> [with _Iterator=double]"
test.cu(11): here
/usr/include/c++/5/bits/stl_iterator_base_types.h(170): error: name followed by "::" must be a class or namespace name
detected during instantiation of class "std::iterator_traits<_Iterator> [with _Iterator=double]"
test.cu(11): here
/usr/include/c++/5/bits/stl_iterator_base_types.h(171): error: name followed by "::" must be a class or namespace name
detected during instantiation of class "std::iterator_traits<_Iterator> [with _Iterator=double]"
test.cu(11): here
/usr/include/c++/5/bits/stl_iterator_base_types.h(172): error: name followed by "::" must be a class or namespace name
detected during instantiation of class "std::iterator_traits<_Iterator> [with _Iterator=double]"
test.cu(11): here
When I change the file extension to .cpp and compile as follows, g++ test.cpp -o test, the code complies. If I then execute ./test, I get the result I am looking for:
distance = 6
Looking at this post inspired me to consider the possibility that I am invoking something from the wrong side of the host/device divide, however, I am not making any GPU calls yet.
Not sure what is going on, but so far the CUDA compiler seems extremely finicky.
