permute_copy_f32

void permute_copy_f32(const float * in, float * out, const int & in_shape, const int & perm)

source edit

Copies a single-precision tensor into a permuted, densely-packed output buffer using the supplied axis permutation.

The implementation iterates the output in flat row-major order and, for each output flat index, back-computes its N-D coordinate and reads from the matching input position via the inverse permutation expressed through the input's C-order strides. No intermediate allocation beyond two stride vectors of size ndim is performed. This is an out-of-place layout transform — there is no in-place fast path.

Shape

out_shape[d] == in_shape[perm[d]] for every d.

See Also

permute_copy_f64, permute_copy_i32, permute_copy_i64 : Same operation for other element types.

Parameters

inconst float*

Source buffer laid out densely (C-order) according to in_shape.

outfloat*

Destination buffer. Caller must pre-allocate numel(in_shape) * sizeof(float) bytes. May not alias in.

in_shapeconst std::vector<int64_t>&

Shape of the source tensor. An empty vector is treated as a 0-D scalar and produces a single-element copy.

permconst std::vector<int>&

Axis permutation of length in_shape.size(). perm[d] names the source axis whose extent becomes output axis d. Must be a valid permutation of [0, ndim).

Notes

Cost is O(numel * ndim) due to per-element coordinate back-projection. For frequently used permutations (e.g. matrix transpose) consider a dimension-specialised kernel; this routine is the generic fallback used by CpuBackend and the GPU tensordot pre-layout path.

Examples

perm = {2, 0, 1} over in_shape = {H, W, C} yields a CHW output.
perm = {0, 2, 3, 1} over in_shape = {N, C, H, W} performs an NHWC
re-layout.