| 171 | |
| 172 | template<typename T> |
| 173 | __device__ void load_shared_image(CParam<T> in, T *local_image, unsigned ix, |
| 174 | unsigned iy, unsigned bx, unsigned by, |
| 175 | unsigned x, unsigned y, unsigned lx, |
| 176 | unsigned ly, const unsigned edge) { |
| 177 | // Copy an image patch to shared memory, with a 3-pixel edge |
| 178 | if (ix < lx && iy < ly && x - 3 < in.dims[0] && y - 3 < in.dims[1]) { |
| 179 | local_image[(ix) + (bx + 6) * (iy)] = |
| 180 | in.ptr[(x - 3) + in.dims[0] * (y - 3)]; |
| 181 | if (x + lx - 3 < in.dims[0]) |
| 182 | local_image[(ix + lx) + (bx + 6) * (iy)] = |
| 183 | in.ptr[(x + lx - 3) + in.dims[0] * (y - 3)]; |
| 184 | if (y + ly - 3 < in.dims[1]) |
| 185 | local_image[(ix) + (bx + 6) * (iy + ly)] = |
| 186 | in.ptr[(x - 3) + in.dims[0] * (y + ly - 3)]; |
| 187 | if (x + lx - 3 < in.dims[0] && y + ly - 3 < in.dims[1]) |
| 188 | local_image[(ix + lx) + (bx + 6) * (iy + ly)] = |
| 189 | in.ptr[(x + lx - 3) + in.dims[0] * (y + ly - 3)]; |
| 190 | } |
| 191 | } |
| 192 | |
| 193 | template<typename T, int arc_length> |
| 194 | __global__ void locate_features(CParam<T> in, float *score, const float thr, |