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
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208

use crate::infer::*;
use crate::internal::*;

use tract_core::ops as mir;
pub use tract_core::ops::binary::wire_cast;
pub use tract_core::ops::binary::wire_rank_broadcast;
use tract_core::ops::binary::BinMiniOp;

#[derive(Debug, Clone)]
pub struct InferenceBinOp(pub Box<dyn BinMiniOp>);

impl Expansion for InferenceBinOp {
    fn name(&self) -> Cow<str> {
        self.0.name().into()
    }

    fn validation(&self) -> Validation {
        self.0.validation()
    }

    fn rules<'r, 'p: 'r, 's: 'r>(
        &'s self,
        s: &mut Solver<'r>,
        inputs: &'p [TensorProxy],
        outputs: &'p [TensorProxy],
    ) -> InferenceResult {
        rules(s, inputs, outputs, move |typa, typb| self.0.result_datum_type(typa, typb))
    }

    fn wire(
        &self,
        prefix: &str,
        target: &mut TypedModel,
        inputs: &[OutletId],
    ) -> TractResult<TVec<OutletId>> {
        let operating_datum_type = self.0.operating_datum_type(
            target.outlet_fact(inputs[0])?.datum_type,
            target.outlet_fact(inputs[1])?.datum_type,
        )?;
        let wires = wire_rank_broadcast(prefix, target, inputs)?;
        let wires = wire_cast(prefix, target, &wires, operating_datum_type)?;
        target.wire_node(prefix, mir::binary::TypedBinOp(self.0.clone()), &wires)
    }
}

pub fn rules<'r, 'p: 'r, 's: 'r, DT: Fn(DatumType, DatumType) -> TractResult<DatumType> + 'p>(
    s: &mut Solver<'r>,
    inputs: &'p [TensorProxy],
    outputs: &'p [TensorProxy],
    dt: DT,
) -> InferenceResult {
    check_input_arity(inputs, 2)?;
    check_output_arity(outputs, 1)?;

    s.with(&inputs[0].shape, move |s, a_shape| {
        s.with(&inputs[1].shape, move |s, b_shape| {
            if let Some(c_shape) =
                crate::infer::helpers::infer_shape_broadcasting(&[&a_shape, &b_shape])
                    .with_context(|| {
                        format!(
                            "Matching {a_shape:?} and {b_shape:?} with numpy/onnx broadcast rules"
                        )
                    })?
            {
                s.equals(&outputs[0].shape, c_shape)?;
            }
            Ok(())
        })
    })?;
    s.given_2(&inputs[0].datum_type, &inputs[1].datum_type, move |s, typa, typb| {
        s.equals(&outputs[0].datum_type, dt(typa, typb)?)
    })?;
    Ok(())
}

pub trait BinIntoHir {
    fn into_hir(self) -> Box<dyn InferenceOp>;
}

impl<B: BinMiniOp> BinIntoHir for B {
    fn into_hir(self) -> Box<dyn InferenceOp> {
        expand(InferenceBinOp(Box::new(self) as _))
    }
}

#[derive(Debug, Clone)]
pub struct Nary(pub Box<dyn mir::binary::BinMiniOp>, pub bool);

impl Nary {
    fn normalize_t<T>(t: &mut Tensor, n: usize) -> TractResult<()>
    where
        T: Datum + std::ops::DivAssign<T> + Copy,
        usize: tract_num_traits::AsPrimitive<T>,
    {
        use tract_num_traits::AsPrimitive;
        let mut t = t.to_array_view_mut::<T>()?;
        let n: T = n.as_();
        t /= &tract_ndarray::arr0(n);
        Ok(())
    }
}

impl Op for Nary {
    fn name(&self) -> Cow<str> {
        format!("{}Nary", self.0.name()).into()
    }

    fn validation(&self) -> Validation {
        self.0.validation()
    }

    not_a_typed_op!();
}

impl EvalOp for Nary {
    fn is_stateless(&self) -> bool {
        true
    }

    fn eval(&self, inputs: TVec<TValue>) -> TractResult<TVec<TValue>> {
        let mut t = inputs[0].clone().into_tensor();
        for i in inputs[1..].iter() {
            let mut i = i.clone().into_tensor();
            let operating_datum_type =
                self.0.operating_datum_type(t.datum_type(), i.datum_type())?;
            if i.datum_type() != operating_datum_type {
                i = i.cast_to_dt(operating_datum_type)?.into_owned();
            }
            if t.datum_type() != operating_datum_type {
                t = t.cast_to_dt(operating_datum_type)?.into_owned();
            }
            t = self.0.eval(t.into_tvalue(), i.into_tvalue())?;
        }
        if self.1 {
            dispatch_numbers!(Self::normalize_t(t.datum_type())(&mut t, inputs.len()))?;
        }
        Ok(tvec!(t.into_tvalue()))
    }
}

impl InferenceRulesOp for Nary {
    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)?;
        let n = inputs.len();
        s.given_all(
            (0..n).map(|i| (&inputs[i].datum_type).bex()),
            move |s, types: Vec<DatumType>| {
                let dt = DatumType::super_type_for(&types)
                    .with_context(|| format!("No super type for {types:?}"))?;
                let dt = self.0.operating_datum_type(dt, dt)?;
                let result = self.0.result_datum_type(dt, dt)?;
                s.equals(&outputs[0].datum_type, result)
            },
        )?;
        s.given_all(inputs.iter().map(|i| &i.shape), move |s, shapes: Vec<TVec<TDim>>| {
            let out = tract_core::broadcast::multi_broadcast(&shapes)
                .with_context(|| format!("Failed to broadcast {:?}", &shapes))?;
            s.equals(&outputs[0].shape, ShapeFactoid::from(out))
        })
    }

    fn to_typed(
        &self,
        _source: &InferenceModel,
        node: &InferenceNode,
        target: &mut TypedModel,
        mapping: &HashMap<OutletId, OutletId>,
    ) -> TractResult<TVec<OutletId>> {
        let inputs = node.inputs.iter().map(|i| mapping[i]).collect::<Vec<_>>();
        let types = inputs
            .iter()
            .map(|i| Ok(target.outlet_fact(*i)?.datum_type))
            .collect::<TractResult<Vec<_>>>()?;
        let dt = DatumType::super_type_for(&types)
            .with_context(|| format!("No super type for {types:?}"))?;
        let operating = self.0.operating_datum_type(dt, dt)?;
        let inputs = wire_cast(&node.name, target, &inputs, operating)?;
        let mut wire = inputs[0];
        for (ix, i) in inputs[1..].iter().enumerate() {
            let wires = wire_rank_broadcast(&format!("{}.{}", node.name, ix), target, &[wire, *i])?;
            wire = target.wire_node(
                format!("{}.{}", node.name, ix),
                mir::binary::TypedBinOp(self.0.clone()),
                &wires,
            )?[0];
        }
        if self.1 {
            let n = tensor0(inputs.len() as i32)
                .cast_to_dt(node.outputs[0].fact.datum_type.concretize().unwrap())?
                .into_owned()
                .broadcast_into_rank(target.outlet_fact(inputs[0])?.rank())?;
            let n = target.add_const(format!("{}.n", node.name), n.into_arc_tensor())?;
            wire = target.wire_node(
                format!("{}.norm", node.name),
                crate::ops::math::div(),
                &[wire, n],
            )?[0];
        }
        Ok(tvec!(wire))
    }

    as_op!();
}