Can't get modf() to work in multithread - c++

Question

I have implemented two different functions to round a double figure to integer.
Here is the first function
static inline int round_v1(double value)
{
int t;
__asm
{
fld value;
fistp t;
}
return t;
}
Here is the second function
static inline int round_v2(double value)
{
double intpart, fractpart;
fractpart = modf(value, &intpart);
if ((fabs(fractpart) != 0.5) || ((((int)intpart) % 2) != 0))
return (int)(value + (value >= 0 ? 0.5 : -0.5));
else
return (int)intpart;
}
Both functions can work well in single thread, but the second one cannot work int multi-thread (using openMP). The program just crash when I use the second one.
Here is the main code where the round_v1 or round_v2 function is called.
void
BilateralFilter_Invoker::doFilter() const
{
if (!src || !dst) return;
int i, j;
int src_width = width + (radius << 1);
omp_set_num_threads(2);
#pragma omp parallel for
for (i = 0; i < height; ++i)
{
unsigned char* pSrc = src + (i+radius)*src_step + radius;
unsigned char* pDst = dst + i*dst_step;
for (j = 0; j < width; ++j)
{
float sum = 0.f, wsum = 0.f;
int val0 = pSrc[j];
for (int k = 0; k < maxk; ++k)
{
int val = pSrc[j + space_offset[k]];
float w = space_weight[k] * color_weight[std::abs(val-val0)];
sum += val * w;
wsum += w;
}
//pDst[j] = (unsigned char)round_v2(sum / wsum);
pDst[j] = (unsigned char)round_v1(sum / wsum);
}
}
}
the variables src, dst, height, width, src_step, dst_step, radius, maxk, space_offset, space_weight, color_weight are member variables of class BilateralFilter_Invoker.
I respectively call round_v1 and round_v2 for test and program crashes only when round_v2 was called. I wonder whether the modf(double, double*) function may cause this problem. For further test, I comment this line
fractpart = modf(value, &intpart);
and replace it by
fractpart = intpart = value;
I run the program again and it did not crash again. I have no idea whether modf(double, double*) causes this problem. Or maybe there is something wrong in my code causes the problem rather than the modf(double, double*) function.
Notice that The operating system I use is Windows7 and the compiler is VC10.

Answer 1

You have made the most common mistake with OpenMP on SO. The iterator of your inner loop needs to be made private. You can either do
#pragma omp parallel for private(j)
or use loop initial declarations
for (int j = 0; j < width; ++j)
In fact, since you never use i or j outside of the loops they apply to there is no reason to declare them C89 style outside of the loops.