- Timestamp:
- Dec 29, 2021, 5:51:49 PM (3 years ago)
- Branches:
- feature/cnn, feature/crepe
- Children:
- b5b0ddd
- Parents:
- c97f7ed
- git-author:
- Paul Brossier <piem@piem.org> (01/08/19 15:49:30)
- git-committer:
- Paul Brossier <piem@piem.org> (12/29/21 17:51:49)
- File:
-
- 1 edited
Legend:
- Unmodified
- Added
- Removed
-
src/ai/conv1d.c
rc97f7ed rd91fa83 275 275 // blas implementation 276 276 // 277 // uses _sdoton the padded input to compute each output elements at once277 // uses gemv on the padded input to compute each output elements at once 278 278 // 279 279 // TODO 280 // - switch to sgemv to factorise over activations->shape[j]281 280 // - avoid copy when padding_start == 0 282 281 // - optimize copying using tensor helpers … … 286 285 { 287 286 uint_t i, j; 288 smpl_t bias, acc;289 287 290 288 uint_t sdot_size = c->kernel->shape[0] * c->kernel->shape[1]; 291 uint_t input_stride = c->stride_shape * c->padded_input->shape[1] 289 uint_t input_stride = c->stride_shape * c->padded_input->shape[1]; 292 290 293 291 AUBIO_ASSERT(c && input_tensor && activations);
Note: See TracChangeset
for help on using the changeset viewer.