| 21 | |
| 22 | template<typename T> |
| 23 | Array<T> transpose(const Array<T> &in, const bool conjugate) { |
| 24 | const dim4 &inDims = in.dims(); |
| 25 | dim4 outDims = dim4(inDims[1], inDims[0], inDims[2], inDims[3]); |
| 26 | Array<T> out = createEmptyArray<T>(outDims); |
| 27 | |
| 28 | const bool is32multiple = |
| 29 | inDims[0] % kernel::TILE_DIM == 0 && inDims[1] % kernel::TILE_DIM == 0; |
| 30 | |
| 31 | kernel::transpose<T>(out, in, getQueue(), conjugate, is32multiple); |
| 32 | |
| 33 | return out; |
| 34 | } |
| 35 | |
| 36 | #define INSTANTIATE(T) \ |
| 37 | template Array<T> transpose(const Array<T> &in, const bool conjugate); |