onnx-runtime-ep-cpu 0.1.0-dev.5

//! `Transpose`: permute axes by moving raw fixed-width element bytes. The `perm` attribute gives the axis
//! order; it defaults to reversing all axes (`docs/ORT2.md` §4.4).

use onnx_runtime_ep_api::{
    EpError, Kernel, KernelFactory, Result, TensorMut, TensorView, ViewOutput,
};
use onnx_runtime_ir::{Node, compute_contiguous_strides};

use super::{check_arity, elem_size, to_dense_bytes, write_dense_bytes};
use crate::strided::{next_index, numel};

/// Dtype-generic Transpose kernel carrying the resolved `perm`.
pub struct TransposeKernel {
    /// Axis permutation; `None` means reverse all axes.
    perm: Option<Vec<usize>>,
}

/// Factory reading the `perm` attribute from the node.
pub struct TransposeFactory;

impl KernelFactory for TransposeFactory {
    fn create(&self, node: &Node, _input_shapes: &[Vec<usize>]) -> Result<Box<dyn Kernel>> {
        let perm = node
            .attr("perm")
            .and_then(|a| a.as_ints())
            .map(|ints| ints.iter().map(|&v| v as usize).collect::<Vec<_>>());
        Ok(Box::new(TransposeKernel { perm }))
    }
}

impl Kernel for TransposeKernel {
    fn execute(&self, inputs: &[TensorView], outputs: &mut [TensorMut]) -> Result<()> {
        check_arity("Transpose", inputs, outputs, 1, 1, 1)?;
        let in_shape = inputs[0].shape.to_vec();
        let rank = in_shape.len();
        let perm = match &self.perm {
            Some(p) => {
                if p.len() != rank {
                    return Err(EpError::KernelFailed(format!(
                        "Transpose: perm rank {} != input rank {rank}",
                        p.len()
                    )));
                }
                p.clone()
            }
            None => (0..rank).rev().collect(),
        };

        if outputs[0].dtype != inputs[0].dtype {
            return Err(EpError::KernelFailed(format!(
                "Transpose: output dtype {:?} must match input dtype {:?}",
                outputs[0].dtype, inputs[0].dtype
            )));
        }
        let esize = elem_size(inputs[0].dtype)?;
        let din = to_dense_bytes(&inputs[0])?;
        let in_strides = compute_contiguous_strides(&in_shape);
        // Output axis i corresponds to input axis perm[i].
        let out_shape: Vec<usize> = perm.iter().map(|&p| in_shape[p]).collect();
        let mut out = vec![0u8; numel(&out_shape) * esize];

        if !out.is_empty() {
            let mut oidx = vec![0usize; rank];
            let mut flat = 0usize;
            loop {
                // input index: in[perm[i]] = out[i]
                let mut in_flat = 0i64;
                for (i, &p) in perm.iter().enumerate() {
                    in_flat += in_strides[p] * oidx[i] as i64;
                }
                let src = in_flat as usize * esize;
                let dst = flat * esize;
                out[dst..dst + esize].copy_from_slice(&din[src..src + esize]);
                flat += 1;
                if !next_index(&out_shape, &mut oidx) {
                    break;
                }
            }
        }
        write_dense_bytes(&mut outputs[0], &out)
    }

    fn supports_strided_input(&self, _input_idx: usize) -> bool {
        true
    }

    fn view_outputs(&self, inputs: &[TensorView], num_outputs: usize) -> Option<Vec<ViewOutput>> {
        if num_outputs != 1 || inputs.len() != 1 || inputs[0].dtype.byte_size() == 0 {
            return None;
        }
        let input = &inputs[0];
        let rank = input.shape.len();
        let perm = self
            .perm
            .clone()
            .unwrap_or_else(|| (0..rank).rev().collect());
        if perm.len() != rank || perm.iter().any(|&axis| axis >= rank) || {
            let mut sorted = perm.clone();
            sorted.sort_unstable();
            sorted != (0..rank).collect::<Vec<_>>()
        } {
            return None;
        }
        Some(vec![ViewOutput {
            input_index: 0,
            shape: perm.iter().map(|&axis| input.shape[axis]).collect(),
            strides: perm.iter().map(|&axis| input.strides[axis]).collect(),
            byte_offset: input.byte_offset,
        }])
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::kernels::testutil::Owned;

    fn run(perm: Option<Vec<usize>>, input: &Owned, out: &mut Owned) {
        let k = TransposeKernel { perm };
        k.execute(&[input.view()], &mut [out.view_mut()]).unwrap();
    }

    #[test]
    fn transpose_2d_default_reverses() {
        let a = Owned::f32(&[2, 3], &[1., 2., 3., 4., 5., 6.]);
        let mut out = Owned::zeros_f32(&[3, 2]);
        run(None, &a, &mut out);
        // [[1,4],[2,5],[3,6]]
        assert_eq!(out.to_f32(), vec![1., 4., 2., 5., 3., 6.]);
    }

    #[test]
    fn transpose_3d_perm() {
        // shape [2,1,3], perm [1,0,2] -> [1,2,3]
        let a = Owned::f32(&[2, 1, 3], &[1., 2., 3., 4., 5., 6.]);
        let mut out = Owned::zeros_f32(&[1, 2, 3]);
        run(Some(vec![1, 0, 2]), &a, &mut out);
        assert_eq!(out.to_f32(), vec![1., 2., 3., 4., 5., 6.]);
    }

    #[test]
    fn transpose_3d_swap_last_two() {
        // shape [1,2,3], perm [0,2,1] -> [1,3,2]
        let a = Owned::f32(&[1, 2, 3], &[1., 2., 3., 4., 5., 6.]);
        let mut out = Owned::zeros_f32(&[1, 3, 2]);
        run(Some(vec![0, 2, 1]), &a, &mut out);
        // rows [1,2,3],[4,5,6] transposed -> [1,4],[2,5],[3,6]
        assert_eq!(out.to_f32(), vec![1., 4., 2., 5., 3., 6.]);
    }

    #[test]
    fn transpose_is_a_zero_copy_strided_view() {
        let a = Owned::f32(&[2, 3], &[1., 2., 3., 4., 5., 6.]);
        let view = TransposeKernel {
            perm: Some(vec![1, 0]),
        }
        .view_outputs(&[a.view()], 1)
        .unwrap()
        .pop()
        .unwrap();
        assert_eq!(view.shape, [3, 2]);
        assert_eq!(view.strides, [1, 3]);
        assert_eq!(view.byte_offset, 0);
    }
    #[test]
    fn transpose_bf16_preserves_element_bits() {
        let x = Owned::bf16(&[2, 2], &[1., -2., 3., 4.]);
        let mut out = Owned::zeros(onnx_runtime_ir::DataType::BFloat16, &[2, 2]);
        run(None, &x, &mut out);
        assert_eq!(
            out.to_u16_bits(),
            vec![
                x.to_u16_bits()[0],
                x.to_u16_bits()[2],
                x.to_u16_bits()[1],
                x.to_u16_bits()[3]
            ]
        );
    }
}