| 52 | // transpose a matrix |
| 53 | template <typename T> |
| 54 | void symBatched(T* out, const T* in, size_s mat_size, size_s batch_size) { |
| 55 | dim3 blockDim(32, 32); |
| 56 | dim3 gridDim((mat_size + blockDim.x - 1) / blockDim.x, (mat_size * batch_size + blockDim.y - 1) / blockDim.y); |
| 57 | symBatchedKernel<<<gridDim, blockDim>>>(out, in, mat_size, batch_size); |
| 58 | } |
| 59 | |
| 60 | #endif // DENSE_TRANSPOSE_H |