Cuda returns wrong value from __device__ function when cos is used - c

Question

I'm trying to run this code on nvidia GPU and it returns strange values. It consist of two modules main.cu and exmodul.cu (listed bellow). For building I'm using:
nvcc -dc -arch sm_35 main.cu
nvcc -dc -arch sm_35 exmodul.cu
nvcc -arch sm_35 -lcudart -o main main.o exmodul.o
If I run that I obtained strange last line!!! gd must be 1.
result=0
result=0
result=0
result=0
gd=-0.5
When I change 1.0 in exmodul.cu to number greater than
1.000000953 or bellow 0.999999999999999945, it return proper
result.
When I change 1.1 in exmodul.cu it also fails except value 1.0.
Behavior doesn't depend on constant 2.0 in the same module.
When I use another function instead of cos like sin or exp it works properly.
Use of double q = cos(1.1); has no effect.
When I copy function extFunc() to module main.cu it works properly.
If I uncomment *gd=1.0; in main.cu it returns correct 1.0.
Tested on Nvidia GT750M and GeForce GTX TITAN Black. (on GT750M returns different value gd=6.1232329394368592e-17 but still wrong). OS: Debian Jessie.
$ nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2013 NVIDIA Corporation
Built on Thu_Mar_13_11:58:58_PDT_2014
Cuda compilation tools, release 6.0, V6.0.1
Have you got any idea what is wrong?
Thanks, Lukas
main.cu
#include <stdio.h> // printf
#include "exmodul.h" // extFunc
__global__ void mykernel(double*gd);
void deviceCheck();
int main(int argc, char *argv[])
{
double gd, *d_gd;
cudaMalloc(&d_gd, sizeof(double)); deviceCheck();
mykernel<<<1,1>>>(d_gd); deviceCheck();
cudaMemcpy(&gd, d_gd, sizeof(double), cudaMemcpyDeviceToHost);
deviceCheck();
cudaFree(d_gd); deviceCheck();
fprintf(stderr,"gd=%.17g\n",gd);
return 0;
}
void deviceCheck()
{
cudaError_t result = cudaSuccess;
cudaDeviceSynchronize();
result = cudaGetLastError();
fprintf(stderr,"result=%d\n",result); fflush(stderr);
}
__global__ void mykernel(double *gd)
{
*gd = extFunc();
//*gd=1.0;
__syncthreads();
return;
}
exmodul.cu
#include "exmodul.h"
__device__ double extFunc()
{
double q = 1.1;
q = cos(q);
if(q<2.0) { q = 1.0; }
return q;
}
exmodul.h
__device__ double extFunc();

Answer 1

I was able to reproduce problematic behavior on a supported config (CUDA 6.5, CentOS 6.2, K40).
When I switched from CUDA 6.5 to CUDA 7 RC, the problem went away.
The problem also did not appear to be reproducible on an older config (CUDA 5.5, CentOS 6.2, M2070)
I suggest switching to CUDA 7 RC to address this issue. I suspect an underlying bug in the compilation process that has been fixed already.