Second iteration crash - order irrelevant - c++

Question

To save on global memory transfers, and because all of the steps of the code work individually, I have tried to combine all of the kernals into a single kernal, with the first 2 (of 3) steps being done as device calls rather than global calls.
This is failing in the second half of the first step.
There is a function that I need to call twice, to calculate the 2 halves of an image. Regardless of the order the image is calculated in, it crashes on the second iteration.
After examining the code as well as I could, and running it multiple times with different return points, I have found what makes it crash.
__device__
void IntersectCone( float* ModDistance,
float* ModIntensity,
float3 ray,
int threadID,
modParam param )
{
bool ignore = false;
float3 normal = make_float3(0.0f,0.0f,0.0f);
float3 result = make_float3(0.0f,0.0f,0.0f);
float normDist = 0.0f;
float intensity = 0.0f;
float check = abs( Dot(param.position, Cross(param.direction,ray) ) );
if(check > param.r1 && check > param.r2)
ignore = true;
float tran = param.length / (param.r2/param.r1 - 1);
float length = tran + param.length;
float Lsq = length * length;
float cosSqr = Lsq / (Lsq + param.r2 * param.r2);
//Changes the centre position?
float3 position = param.position - tran * param.direction;
float aDd = Dot(param.direction, ray);
float3 e = position * -1.0f;
float aDe = Dot(param.direction, e);
float dDe = Dot(ray, e);
float eDe = Dot(e, e);
float c2 = aDd * aDd - cosSqr;
float c1 = aDd * aDe - cosSqr * dDe;
float c0 = aDe * aDe - cosSqr * eDe;
float discr = c1 * c1 - c0 * c2;
if(discr <= 0.0f)
ignore = true;
if(!ignore)
{
float root = sqrt(discr);
float sign;
if(c1 > 0.0f)
sign = 1.0f;
else
sign = -1.0f;
//Try opposite sign....?
float3 result = (-c1 + sign * root) * ray / c2;
e = result - position;
float dot = Dot(e, param.direction);
float3 s1 = Cross(e, param.direction);
float3 normal = Cross(e, s1);
if( (dot > tran) || (dot < length) )
{
if(Dot(normal,ray) <= 0)
{
normal = Norm(normal); //This stuff (1)
normDist = Magnitude(result);
intensity = -IntensAt1m * Dot(ray, normal) / (normDist * normDist);
}
}
}
ModDistance[threadID] = normDist; and this stuff (2)
ModIntensity[threadID] = intensity;
}
There are two things I can do to to make this not crash, both off which negate the point of the function: If I do not try to write to ModDistance[] and ModIntensity[], or if I do not write to normDist and intensity.
First chance exceptions are thrown by the code above, but not if either of the blocks commented out.
Also, The program only crashes the second time this routine is called.
Have been trying to figure this out all day, any help would be fantastic.
The code that calls it is:
int subrow = threadIdx.y + Mod_Height/2;
int threadID = subrow * (Mod_Width+1) + threadIdx.x;
int obsY = windowY + subrow;
float3 ray = CalculateRay(obsX,obsY);
if( !IntersectSphere(ModDistance, ModIntensity, ray, threadID, param) )
{
IntersectCone(ModDistance, ModIntensity, ray, threadID, param);
}
subrow = threadIdx.y;
threadID = subrow * (Mod_Width+1) + threadIdx.x;
obsY = windowY + subrow;
ray = CalculateRay(obsX,obsY);
if( !IntersectSphere(ModDistance, ModIntensity, ray, threadID, param) )
{
IntersectCone(ModDistance, ModIntensity, ray, threadID, param);
}

Answer 1

The kernel is running out of resources. As posted in the comments, it was giving the error CudaErrorLaunchOutOfResources.
To avoid this, you should use a __launch_bounds__ specifier to specify the block dimensions you want for your kernel. This will force the compiler to ensure there are enough resources. See the CUDA programming guide for details on __launch_bounds__.