| 342 | } |
| 343 | |
| 344 | af_err af_convolve2_nn(af_array *out, const af_array signal, |
| 345 | const af_array filter, const unsigned stride_dims, |
| 346 | const dim_t *strides, const unsigned padding_dims, |
| 347 | const dim_t *paddings, const unsigned dilation_dims, |
| 348 | const dim_t *dilations) { |
| 349 | try { |
| 350 | const ArrayInfo &sInfo = getInfo(signal); |
| 351 | const ArrayInfo &fInfo = getInfo(filter); |
| 352 | |
| 353 | af::dim4 sDims = sInfo.dims(); |
| 354 | af::dim4 fDims = fInfo.dims(); |
| 355 | |
| 356 | const af_dtype signalType = sInfo.getType(); |
| 357 | |
| 358 | dim4 stride(stride_dims, strides); |
| 359 | dim4 padding(padding_dims, paddings); |
| 360 | dim4 dilation(dilation_dims, dilations); |
| 361 | |
| 362 | size_t stride_ndims = stride.ndims(); |
| 363 | size_t padding_ndims = padding.ndims(); |
| 364 | size_t dilation_ndims = dilation.ndims(); |
| 365 | ARG_ASSERT(3, stride_ndims > 0 && stride_ndims <= 2); |
| 366 | ARG_ASSERT(5, padding_ndims >= 0 && padding_ndims <= 2); |
| 367 | ARG_ASSERT(7, dilation_ndims > 0 && dilation_ndims <= 2); |
| 368 | |
| 369 | // assert number of features matches between signal and filter |
| 370 | DIM_ASSERT(1, sDims[2] == fDims[2]); |
| 371 | |
| 372 | af_array output; |
| 373 | switch (signalType) { |
| 374 | case f32: |
| 375 | output = convolve2Strided<float>(signal, filter, stride, |
| 376 | padding, dilation); |
| 377 | break; |
| 378 | case f64: |
| 379 | output = convolve2Strided<double>(signal, filter, stride, |
| 380 | padding, dilation); |
| 381 | break; |
| 382 | case f16: |
| 383 | output = convolve2Strided<half>(signal, filter, stride, padding, |
| 384 | dilation); |
| 385 | break; |
| 386 | default: TYPE_ERROR(1, signalType); |
| 387 | } |
| 388 | std::swap(*out, output); |
| 389 | } |
| 390 | CATCHALL; |
| 391 | return AF_SUCCESS; |
| 392 | } |
| 393 | |
| 394 | template<typename T> |
| 395 | af_array conv2GradCall(const af_array incoming_gradient, |
no test coverage detected