// 2. Transpose to [H, W, ..., G, filter_in_depth, out_depth / G].
      llvm::SmallVector<int64_t, 6> perm(num_dims + 1);
      std::iota(perm.begin(), perm.end(), 0);
      std::swap(perm[num_spatial_dims], perm[num_spatial_dims + 1]);
      std::swap(new_shape[num_spatial_dims], new_shape[num_spatial_dims + 1]);