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

use crate::internal::*;
use num_traits::AsPrimitive;
use std::iter::Sum;

use crate::ops::cnn::pools::{ConcretePoolGeometry, PoolGeometry, PoolSpec};

#[derive(Debug, Clone, new, Hash)]
pub struct SumPool {
    pub pool_spec: PoolSpec,
    pub count_include_pad: bool,
    pub normalize: bool,
}

impl_dyn_hash!(SumPool);

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

    fn info(&self) -> TractResult<Vec<String>> {
        Ok(self.pool_spec.info())
    }

    fn validation(&self) -> Validation {
        Validation::Rounding
    }

    op_core_mir!();
    op_as_typed_op!();
}

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

    fn eval(&self, inputs: TVec<Arc<Tensor>>) -> TractResult<TVec<Arc<Tensor>>> {
        let shape: TVec<TDim> = inputs[0].shape().iter().map(|d| d.to_dim()).collect();
        self.to_lir(&shape)?.eval(inputs)
    }
}

impl TypedOp for SumPool {
    fn output_facts(&self, inputs: &[&TypedFact]) -> TractResult<TVec<TypedFact>> {
        self.pool_spec.output_facts(inputs)
    }

    as_op!();
}

impl SumPool {
    fn to_lir(&self, input_shape: &[TDim]) -> TractResult<LirSumPool> {
        Ok(LirSumPool {
            pool_spec: self.pool_spec.clone(),
            count_include_pad: self.count_include_pad,
            normalize: self.normalize,
            geometry: self.pool_spec.compute_geo(&input_shape)?,
        })
    }
}

#[derive(Debug, Clone, new, Hash)]
pub struct LirSumPool {
    pub pool_spec: PoolSpec,
    pub count_include_pad: bool,
    pub normalize: bool,
    pub geometry: PoolGeometry,
}

impl_dyn_hash!(LirSumPool);

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

    fn info(&self) -> TractResult<Vec<String>> {
        Ok(self.pool_spec.info())
    }

    fn validation(&self) -> Validation {
        Validation::Rounding
    }

    op_core_lir!();
    op_as_typed_op!();
}

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

    fn eval(&self, mut inputs: TVec<Arc<Tensor>>) -> TractResult<TVec<Arc<Tensor>>> {
        let input = args_1!(inputs);
        let geo = self.geometry.to_concrete(input.shape())?;
        let values = if input.datum_type().is_float() {
            let mut values =
                unsafe { Tensor::uninitialized_dt(input.datum_type(), &*geo.output_shape.shape)? };
            dispatch_floatlike!(Self::eval_t(input.datum_type())(
                self,
                &*input,
                values.as_ptr_mut()?,
                geo.as_ref()
            ))?;
            values
        } else {
            let mut values =
                unsafe { Tensor::uninitialized_dt(DatumType::F32, &*geo.output_shape.shape)? };
            let input_f32 = input.cast_to_dt(DatumType::F32)?;
            self.eval_t::<f32>(input_f32.as_ref(), values.as_ptr_mut()?, geo.as_ref())?;
            values.cast_to_dt(input.datum_type())?.into_owned()
        };

        Ok(tvec!(values.into_arc_tensor()))
    }
}

impl TypedOp for LirSumPool {
    fn output_facts(&self, inputs: &[&TypedFact]) -> TractResult<TVec<TypedFact>> {
        self.pool_spec.output_facts(inputs)
    }

    as_op!();
}

impl LirSumPool {
    fn eval_t<T: Copy + Datum + Sum + num_traits::Float>(
        &self,
        input: &Tensor,
        values_ptr: *mut T,
        geo: &ConcretePoolGeometry,
    ) -> TractResult<()>
    where
        usize: AsPrimitive<T>,
    {
        let input_ptr = input.as_ptr::<T>()?;

        let n = *geo.input_shape.n().unwrap_or(&1);
        let n_stride_i = geo.input_shape.n_stride().unwrap_or(&0);
        let n_stride_o = geo.output_shape.n_stride().unwrap_or(&0);
        unsafe {
            geo.patch.visit_output(|visitor| {
                let div: Option<T> = if self.normalize {
                    Some(
                        if self.count_include_pad {
                            geo.patch.standard_layout_data_field.len().as_()
                        } else {
                            visitor.valid_count().as_()
                        }
                        .recip(),
                    )
                } else {
                    None
                };
                for n in 0..n {
                    let input_offset = n * n_stride_i;
                    let output_offset = n * n_stride_o;
                    for c in 0..*geo.input_shape.c() {
                        let input_offset = input_offset + geo.input_shape.c_stride() * c;
                        let output_offset = output_offset + geo.output_shape.c_stride() * c;
                        let sum = visitor
                            .valid_offsets()
                            .map(|v| *input_ptr.offset(v + input_offset as isize))
                            .sum::<T>();

                        if let Some(div) = div {
                            *values_ptr.offset(output_offset as isize + visitor.output_offset) =
                                sum * div;
                        }
                    }
                }
            });
        }
        Ok(())
    }
}