Trouble using thrust complex(double) vector with sinus function - cuda

Question

I have some problems using the sin function on a Thrust complex double vector on the device and on the host: it's seem like computation is done in float.
With thrust::device_vector< thrust::complex<double> > and thrust::host_vector< thrust::complex<double> >, I obtain:
sin( 1+0i ) == (0.8414709568023682,0)
with std::complex<double>:
sin( 1+0i ) == (0.8414709848078965,0)
and std::complex<float> :
sin( 1+0i ) == (0.8414709568023682,0)
What mistake did I do in my code? In the compilation process I used
nvcc test.cu -o test
Here is the full code:
#include <thrust/host_vector.h>
#include <thrust/device_vector.h>
#include <thrust/complex.h>
#include <iostream>
#include <iomanip>
#include <complex>
#include <cmath>
template <typename Vector>
void Print(Vector &V){
for (int i=0;i<V.size();i++)
std::cout << V[i] << " ";
std::cout << "\n";
}
template <typename T>
struct sin_functor : public thrust::unary_function< T , T >
{
__host__ __device__
T operator()( T x) const
{
return sin( x );
}
};
template <typename Vector>
void ThrustComputation(){
typedef typename Vector::value_type Tvec;
Vector A(2);
A[0]=Tvec(1.,0.);A[1]=Tvec(1.,1.);
std::cout << "A: " << std::endl;
std::cout << " ";Print<Vector>(A);
Vector B(A.size());
thrust::transform(A.begin(),A.end(),B.begin(), sin_functor<Tvec>());
std::cout << "B =sin(A): " << std::endl;
std::cout << " ";Print<Vector>(B);
}
template <typename T>
void stdComputation(){
std::complex<T> sA[2];
sA[0]=std::complex<T>(1.,0.);
sA[1]=std::complex<T>(1.,1.);
std::cout << "sA: " << std::endl;
std::cout << " " << sA[0] << " " << sA[1] << std::endl;
std::cout << "sin(sA): " << std::endl;
std::cout << " " << sin(sA[0]) << " " << sin(sA[1]) << std::endl;
}
int main(int argc, char **argv)
{
std::cout << std::setprecision(16);
std::cout << "Thrust: Computation on GPU device (double)\n";
ThrustComputation<thrust::device_vector< thrust::complex<double> > >();
std::cout << "Thrust: Computation on host (double)\n";
ThrustComputation<thrust::host_vector< thrust::complex<double> > >();
std::cout << "std: Computation (double)\n";
stdComputation<double>();
std::cout << "std: Computation (float)\n";
stdComputation<float>();
return 0;
}
The output on my computer (Ubuntu 14.04 LTS, cuda 7.5) is :
Thrust: Computation on GPU device (double)
A:
(1,0) (1,1)
B =sin(A):
(0.8414709568023682,0) (1.298457622528076,0.6349639296531677)
Thrust: Computation on host (double)
A:
(1,0) (1,1)
B =sin(A):
(0.8414709568023682,0) (1.298457622528076,0.6349639296531677)
std: Computation (double)
sA:
(1,0) (1,1)
sin(sA):
(0.8414709848078965,0) (1.298457581415977,0.6349639147847361)
std: Computation (float)
sA:
(1,0) (1,1)
sin(sA):
(0.8414709568023682,0) (1.298457503318787,0.6349638700485229)

Answer 1

This would appear to be a genuine bug in the thrust library. A quick scan of the code on github led me to this, which might be the culprit. It seems that thrust's double precision csinh function, upon which complex sin relies, has an accidental intermediate cast to float, which is probably causing the loss of precision you have observed. As suggested in comments, you should report this as a bug.