| 125 | |
| 126 | template <typename T> |
| 127 | void GaussianBlurImpl::exec_internal(_megdnn_tensor_in src, _megdnn_tensor_out dst) { |
| 128 | auto N = src.layout.shape[0], IH = src.layout.shape[1], IW = src.layout.shape[2], |
| 129 | IC = src.layout.shape[3]; |
| 130 | |
| 131 | using namespace megcv; |
| 132 | |
| 133 | Size ksize = Size(param().kernel_height, param().kernel_width); |
| 134 | Mat<float> kx(1, ksize.cols(), 1); |
| 135 | Mat<float> ky(1, ksize.rows(), 1); |
| 136 | |
| 137 | gaussian_blur::createGaussianKernels<float>( |
| 138 | kx, ky, ksize, param().sigma_x, param().sigma_y); |
| 139 | |
| 140 | uint32_t kernel_height = ky.width(); |
| 141 | uint32_t kernel_width = kx.width(); |
| 142 | uint32_t half_h = kernel_height / 2; |
| 143 | uint32_t half_w = kernel_width / 2; |
| 144 | |
| 145 | rep(n, N) rep(h, IH) rep(w, IW) rep(c, IC) { |
| 146 | double val = 0; |
| 147 | rep(iy, kernel_height) { |
| 148 | int y = gaussian_blur::border_interpolate( |
| 149 | h + iy - half_h, IH, param().border_mode); |
| 150 | rep(ix, kernel_width) { |
| 151 | int x = gaussian_blur::border_interpolate( |
| 152 | w + ix - half_w, IW, param().border_mode); |
| 153 | |
| 154 | //! BORDER_CONSTANT or BORDER_TRANSPARENT |
| 155 | if (x != -1 && y != -1) { |
| 156 | val += kx.at(0, ix, 0) * ky.at(0, iy, 0) * |
| 157 | src.ptr<T>() |
| 158 | [n * src.layout.stride[0] + |
| 159 | y * src.layout.stride[1] + |
| 160 | x * src.layout.stride[2] + |
| 161 | c * src.layout.stride[3]]; |
| 162 | } |
| 163 | } |
| 164 | } |
| 165 | dst.ptr<T>() |
| 166 | [n * dst.layout.stride[0] + h * dst.layout.stride[1] + |
| 167 | w * dst.layout.stride[2] + c * dst.layout.stride[3]] = |
| 168 | static_cast<T>(val); |
| 169 | } |
| 170 | } |
| 171 | |
| 172 | void GaussianBlurImpl::exec( |
| 173 | _megdnn_tensor_in src, _megdnn_tensor_in dst, _megdnn_workspace /*workspace*/) { |