| 33 | |
| 34 | template<typename T> |
| 35 | void gaussian1D(T* out, const int dim, double sigma = 0.0) { |
| 36 | if (!(sigma > 0)) sigma = 0.25 * dim; |
| 37 | |
| 38 | T sum = (T)0; |
| 39 | for (int i = 0; i < dim; i++) { |
| 40 | int x = i - (dim - 1) / 2; |
| 41 | T el = 1. / sqrt(2 * PI_VAL * sigma * sigma) * |
| 42 | exp(-((x * x) / (2 * (sigma * sigma)))); |
| 43 | out[i] = el; |
| 44 | sum += el; |
| 45 | } |
| 46 | |
| 47 | for (int k = 0; k < dim; k++) out[k] /= sum; |
| 48 | } |
| 49 | |
| 50 | inline __device__ float block_reduce_sum(float val) { |
| 51 | __shared__ float data[THREADS_X * THREADS_Y]; |