| 63 | } |
| 64 | |
| 65 | bool RelayoutForwardImpl::Param::try_copy_contig() { |
| 66 | auto &&lsrc = m_src.layout, &&ldst = m_dst.layout; |
| 67 | if (lsrc.ndim != 1 || ldst.ndim != 1) |
| 68 | return false; |
| 69 | if (lsrc.stride[0] != 1 || ldst.stride[0] != 1) |
| 70 | return false; |
| 71 | size_t copy_size = ldst.span().dist_byte(); |
| 72 | |
| 73 | cuda_check(cudaMemcpyAsync( |
| 74 | m_dst.raw_ptr(), m_src.raw_ptr(), copy_size, cudaMemcpyDeviceToDevice, |
| 75 | m_opr->stream())); |
| 76 | return true; |
| 77 | } |
| 78 | |
| 79 | bool RelayoutForwardImpl::expand_dim2(TensorLayout& dst, const TensorLayout& src) { |
| 80 | megdnn_assert(src.ndim == 2 && dst.ndim == 1); |