| 337 | |
| 338 | template <class T> |
| 339 | void array_copy_linear_to_strided( |
| 340 | void* dst, |
| 341 | int64_t offset, |
| 342 | void* src, |
| 343 | const std::vector<int64_t>& shape, |
| 344 | const std::vector<int64_t>& stride) { |
| 345 | auto dst_t = reinterpret_cast<T*>(dst); |
| 346 | auto src_t = reinterpret_cast<T*>(src); |
| 347 | |
| 348 | // Deal with scalar arrays |
| 349 | if (shape.size() == 0) { |
| 350 | dst_t[offset] = src_t[0]; |
| 351 | return; |
| 352 | } |
| 353 | |
| 354 | int ndim = shape.size(); |
| 355 | while (ndim > 1 && (shape[ndim - 1] == 1)) { |
| 356 | ndim--; |
| 357 | } |
| 358 | // note: isize/istride considers the istride-contiguous portion |
| 359 | // of the last dimensions (not only the last one) |
| 360 | int64_t isize = shape[ndim - 1]; |
| 361 | int64_t istride = stride[ndim - 1]; |
| 362 | while ((ndim > 2) && (stride[ndim - 2] == isize * istride)) { |
| 363 | isize *= shape[ndim - 2]; |
| 364 | ndim--; |
| 365 | } |
| 366 | int64_t osize = 1; |
| 367 | for (int dim = 0; dim < ndim - 1; dim++) { |
| 368 | osize *= shape[dim]; |
| 369 | } |
| 370 | int64_t oprod = 1; |
| 371 | for (int dim = 1; dim < ndim - 1; dim++) { |
| 372 | oprod *= shape[dim]; |
| 373 | } |
| 374 | if (istride == 1) { |
| 375 | for (int64_t idx = 0; idx < osize; idx++) { |
| 376 | int64_t roffset = offset; |
| 377 | int64_t r = idx; |
| 378 | int64_t rprod = oprod; |
| 379 | for (int dim = 0; dim < ndim - 1; dim++) { |
| 380 | int64_t rshape = r / rprod; |
| 381 | r = r % rprod; |
| 382 | rprod /= shape[dim + 1]; |
| 383 | roffset += stride[dim] * rshape; |
| 384 | } |
| 385 | std::memcpy(dst_t + roffset, src_t + idx * isize, isize * sizeof(T)); |
| 386 | } |
| 387 | } else { |
| 388 | for (int64_t idx = 0; idx < osize; idx++) { |
| 389 | int64_t roffset = offset; |
| 390 | int64_t r = idx; |
| 391 | int64_t rprod = oprod; |
| 392 | for (int dim = 0; dim < ndim - 1; dim++) { |
| 393 | int64_t rshape = r / rprod; |
| 394 | r = r % rprod; |
| 395 | rprod /= shape[dim + 1]; |
| 396 | roffset += stride[dim] * rshape; |