permute

permute#

Permute the dimensions of an operator

Added in version 0.3.0.

template<typename T>
__MATX_INLINE__ auto matx::permute(const T &op, const int32_t (&dims)[T::Rank()])#

Operator to permute the dimensions of a tensor or operator.

Each dimension must appear in the dims array once, otherwise matxInvalidDim is thrown.

This operator can appear as an rvalue or lvalue.

Template Parameters:

T – Input operator/tensor type

Parameters:
  • op – Input operator

  • dims – the reordered dimensions of the operator.

Returns:

permuted operator

template<typename T>
__MATX_INLINE__ auto matx::permute(const T &op, const cuda::std::array<int32_t, T::Rank()> &dims)#

Examples#

// Permute from dims {0, 1, 2} to {2, 0, 1}
auto op = permute(A, {2, 0, 1});

Adjacent permutations of expressions are composed into a single permutation. Inverse permutations therefore restore the original mapping, including for in-place element-wise operations with unsafe alias detection enabled. A non-identity combined permutation that reads the destination remains unsafe. Each permutation’s axes are validated before composition.