| 257 | // ------------------------------------------------------------------------------------ |
| 258 | |
| 259 | void add( |
| 260 | float beta, |
| 261 | tensor& dest, |
| 262 | float alpha, |
| 263 | const tensor& src |
| 264 | ) |
| 265 | { |
| 266 | DLIB_CASSERT( |
| 267 | (have_same_dimensions(src, dest) || |
| 268 | (src.num_samples()==1 && src.k()==dest.k() && src.nr()==1 && src.nc()==1) || |
| 269 | (src.num_samples()==1 && src.k()==dest.k() && src.nr()==dest.nr() && src.nc()==dest.nc()) || |
| 270 | (src.num_samples()==1 && src.k()==1 && src.nr()==dest.nr() && src.nc()==dest.nc()) || |
| 271 | (src.num_samples()==dest.num_samples() && src.k()==1 && src.nr()==1 && src.nc()==1)) && |
| 272 | is_same_object(src,dest) == false , |
| 273 | "\n\t dest.num_samples(): " << dest.num_samples() |
| 274 | <<"\n\t dest.k(): " << dest.k() |
| 275 | <<"\n\t dest.nr(): " << dest.nr() |
| 276 | <<"\n\t dest.nc(): " << dest.nc() |
| 277 | <<"\n\t src.num_samples(): " << src.num_samples() |
| 278 | <<"\n\t src.k(): " << src.k() |
| 279 | <<"\n\t src.nr(): " << src.nr() |
| 280 | <<"\n\t src.nc(): " << src.nc() |
| 281 | ); |
| 282 | |
| 283 | if (dest.size() == src.size() && beta == 1) |
| 284 | { |
| 285 | // Call the dlib function in this case since it's faster than the one that |
| 286 | // comes with cuDNN (at least as of cuDNN v4). |
| 287 | add_scaled(dest, alpha, src); |
| 288 | return; |
| 289 | } |
| 290 | else if (src.num_samples()==dest.num_samples() && src.k()==1 && src.nr()==1 && src.nc()==1) |
| 291 | { |
| 292 | add_cv_to_all_columns(beta, dest, alpha, src); |
| 293 | return; |
| 294 | } |
| 295 | |
| 296 | CHECK_CUDNN(cudnnAddTensor(context(), |
| 297 | &alpha, |
| 298 | descriptor(src), |
| 299 | src.device(), |
| 300 | &beta, |
| 301 | descriptor(dest), |
| 302 | dest.device())); |
| 303 | } |
| 304 | |
| 305 | void assign_conv_bias_gradient ( |
| 306 | tensor& grad, |
no test coverage detected