1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138

use crate::internal::*;

// FIXME: pulsification is very fragile

#[derive(Debug, Clone, new)]
pub struct PermuteAxes {
    pub axes: Option<Vec<usize>>,
}

impl PermuteAxes {
    fn compute_shape<D: DimLike>(&self, input: &[D]) -> TVec<D> {
        if let Some(ref axes) = self.axes {
            let mut new_shape = tvec![D::zero(); input.len()];
            for (ix, &d) in axes.iter().enumerate() {
                new_shape[ix] = input[d].clone();
            }
            new_shape
        } else {
            let mut new_shape: TVec<D> = input.iter().cloned().collect();
            new_shape.reverse();
            new_shape
        }
    }

    /// Evaluates the operation given the input tensors.
    fn eval_t<T: Datum>(&self, input: Arc<Tensor>) -> TractResult<TVec<Arc<Tensor>>> {
        if let Some(ref axes) = self.axes {
            Ok(tvec![input
                .into_tensor()
                .into_array::<T>()?
                .permuted_axes(&**axes)
                .into_arc_tensor()])
        } else {
            Ok(tvec![input.into_tensor().into_array::<T>()?.reversed_axes().into_arc_tensor()])
        }
    }
}

impl Op for PermuteAxes {
    fn name(&self) -> Cow<str> {
        "PermuteAxes".into()
    }

    canonic!();
    op_as_typed_op!();
    op_as_pulsed_op!();
}

impl StatelessOp for PermuteAxes {
    fn eval(&self, mut inputs: TVec<Arc<Tensor>>) -> TractResult<TVec<Arc<Tensor>>> {
        let input = args_1!(inputs);
        dispatch_datum!(Self::eval_t(input.datum_type())(self, input))
    }
}

impl InferenceRulesOp for PermuteAxes {
    fn rules<'r, 'p: 'r, 's: 'r>(
        &'s self,
        s: &mut Solver<'r>,
        inputs: &'p [TensorProxy],
        outputs: &'p [TensorProxy],
    ) -> InferenceResult {
        check_output_arity(&outputs, 1)?;
        s.equals(&outputs[0].datum_type, &inputs[0].datum_type)?;
        s.equals(&outputs[0].rank, &inputs[0].rank)?;
        s.given(&inputs[0].shape, move |s, shape| {
            let output_shape = self.compute_shape(&shape);
            s.equals(&outputs[0].shape, output_shape)
        })
    }

    inference_op_as_op!();
    to_typed!();
}

impl TypedOp for PermuteAxes {
    fn output_facts(&self, inputs: &[&TypedFact]) -> TractResult<TVec<TypedFact>> {
        Ok(tvec!(TypedFact::dt_shape(
            inputs[0].datum_type,
            self.compute_shape(&*inputs[0].shape.to_tvec()).as_ref(),
        )?))
    }

    fn axes_info(&self, model: &TypedModel, node: &TypedNode) -> TractResult<AxesInfo> {
        let permutation = if let Some(axes) = self.axes.clone() {
            axes
        } else {
            (0..model.outlet_fact(node.inputs[0])?.shape.rank()).rev().collect()
        };
        let mut infos = tvec!();
        for (from, to) in permutation.iter().enumerate() {
            infos.push(AxisInfo {
                inputs: tvec!(Some(from)),
                outputs: tvec!(Some(*to)),
                period: 1,
                disposable: true,
            })
        }
        Ok(infos.into())
    }

    fn pulsify(
        &self,
        _source: &NormalizedModel,
        node: &NormalizedNode,
        target: &mut PulsedModel,
        mapping: &HashMap<OutletId, OutletId>,
        _pulse: usize,
    ) -> TractResult<TVec<OutletId>> {
        let input = mapping[&node.inputs[0]];
        let mut fact = target.outlet_fact(input)?.clone();
        if let Some(axes) = &self.axes {
            fact.axis = axes.iter().position(|x| x == &fact.axis).ok_or_else(|| {
                format!("Could not find streaming axis {} if permute axes {:?}", fact.axis, axes)
            })?;
            fact.shape = axes.iter().map(|idx| fact.shape[*idx]).collect();
        }
        target.wire_node(&*node.name, self.clone(), &[input])
    }

    typed_op_as_op!();
}

impl PulsedOp for PermuteAxes {
    fn pulsed_output_facts(&self, inputs: &[&PulsedFact]) -> TractResult<TVec<PulsedFact>> {
        let mut fact = inputs[0].clone();
        fact.axis = if let Some(axes) = &self.axes {
            axes[fact.axis]
        } else {
            fact.shape.len() - 1 - fact.axis
        };
        fact.shape = self.compute_shape(&*inputs[0].shape);
        Ok(tvec!(fact))
    }

    pulsed_op_as_op!();
    pulsed_op_to_typed_op!();
}