Error: an illegal memory access was encountered - c++

Question

I had asked doubt error: calling a __host__ function from a __global__ function is not allowed and i got the ans . accordingly i have modified my code bt i am unable to access d_point[i]. how can i access that.
__global__ void densefun(int *d_counters,float2 *d_points,int d_x_max,int d_y_max,int width,int height, int min_distance,int size)
{
int i = blockDim.x * blockIdx.x + threadIdx.x;
if(i <= size)
{
float2 point = (d_points)[i];
int x = floorf(point.x);
int y = floorf(point.y);
printf(" ( %d %d )",x,y);
if(x < d_x_max && y < d_y_max)
{
x /= min_distance;
y /= min_distance;
(d_counters)[y*width+x]++;
__syncthreads();
}
}
}
void DenseSample(const Mat& grey, std::vector<Point2f>& points, const double quality, const int min_distance)
{
int width = grey.cols/min_distance;
int height = grey.rows/min_distance;
Mat eig;
cornerMinEigenVal(grey, eig, 3, 3);
double maxVal = 0;
minMaxLoc(eig, 0, &maxVal);
const double threshold = maxVal*quality;
std::vector<int> counters(width*height);
int x_max = min_distance*width;
int y_max = min_distance*height;
printf("in descriptor size:%ld ",points.size());
int *d_counters;
float2 *d_points;
cudaMalloc(&d_counters,counters.size()*width*height*sizeof(int));
printf("in cuda point size:%d ",points.size());
cudaMalloc(&d_points,points.size()*sizeof(float2));
cout<<"points.size() : "<<points.size()<<endl;
cudaMemcpy(d_points, &points, points.size()*sizeof(float2), cudaMemcpyHostToDevice);
int blk=cvFloor(points.size()/1024)+1;
cout<<"blk : "<<blk<<endl;
if(points.size()>0)
{
densefun<<<blk,1024>>>(d_counters,d_points,x_max,y_max,width,height,min_distance, points.size());
cudaError_t err = cudaGetLastError();
if (err != cudaSuccess)
printf("Error: %s\n", cudaGetErrorString(err));
cudaMemcpy(&counters, d_counters, counters.size()* width*height*sizeof(int), cudaMemcpyDeviceToHost);
}
cudaFree(d_counters);
cudaFree(d_points);
points.clear();
int index = 0;
int offset = min_distance/2;
for(int i = 0; i < height; i++)
for(int j = 0; j < width; j++, index++)
{
if(counters[index] <= 0)
{
int x = j*min_distance+offset;
int y = i*min_distance+offset;
if(eig.at<float>(y, x) > threshold)
points.push_back(Point2f(float(x), float(y)));
}
}
}
output is:
in descriptor size:1605 in cuda point size:1605 points.size() : 1605
blk : 2
Error: an illegal memory access was encountered
in descriptor size:918 in cuda point size:918 points.size() : 918
blk : 1
Error: an illegal memory access was encountered

Answer 1

You create a thread gird with block length 1024 and grid length equal to
int blk=cvFloor(points.size()/1024)+1;
Which basically means that the number of threads will be multiple of 1024 greater than points.size(). In this case using:
int i = blockDim.x * blockIdx.x + threadIdx.x;
float2 point = (d_points)[i];
cannot be successful, because you can be almost certain that you will get out of bounds memory access. Add some conditional to ensure that it won't happen.
__global__ void densefun(int *d_counters,float2 *d_points,int d_x_max,int d_y_max,int width, int height, int min_distance)
{
int i = blockDim.x * blockIdx.x + threadIdx.x;
if(i < width * height)
{
//rest of the code
}
}
Also, you don't allocate enugh memory for d_points:
float2 *d_points;
cudaMalloc(&d_points,points.size()*sizeof(float));
If you want to allocate array of float2 (or copy to it) you need to use sizeof(float2).