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
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
use crate::internal::*;
use crate::pulse::PulsedFact;
use ndarray::*;

#[derive(Debug, new, Clone)]
struct DelayState {
    buffer: Tensor,
}

impl DelayState {
    pub fn eval_t<T: Datum>(&mut self, op: &Delay, input: Arc<Tensor>) -> TractResult<Arc<Tensor>> {
        let axis = Axis(op.axis);
        let input = input.to_array_view::<T>()?;
        let mut buffer = self.buffer.to_array_view_mut::<T>()?;

        let buffered = op.delay + op.overlap;
        let input_pulse = input.shape()[op.axis];
        let output_pulse = input_pulse + op.overlap;
        let mut output_shape: TVec<usize> = input.shape().into();
        output_shape[op.axis] = output_pulse;
        // build output
        let output = if op.delay < input_pulse {
            let mut output = unsafe { Tensor::uninitialized::<T>(&*output_shape)? };
            let from_input = input_pulse - op.delay;
            let from_buffer = output_pulse - from_input;
            output
                .to_array_view_mut::<T>()?
                .slice_axis_mut(axis, Slice::from(..from_buffer))
                .assign(&buffer.slice_axis(axis, Slice::from(..from_buffer)));
            output
                .to_array_view_mut::<T>()?
                .slice_axis_mut(axis, Slice::from(from_buffer..))
                .assign(&input.slice_axis(axis, Slice::from(..from_input)));
            output
        } else {
            buffer.slice_axis(axis, Slice::from(..output_pulse)).to_owned().into_tensor()
        };
        // maintain buffer
        if buffered < input_pulse {
            buffer.assign(&input.slice_axis(axis, Slice::from((input_pulse - buffered)..)));
        } else {
            let stride = buffer.strides()[op.axis] as usize * input_pulse;
            buffer.as_slice_mut().unwrap().rotate_left(stride);
            buffer.slice_axis_mut(axis, Slice::from((buffered - input_pulse)..)).assign(&input);
        }
        let output = output.into_arc_tensor();
        Ok(output)
    }
}

impl OpState for DelayState {
    fn eval(
        &mut self,
        _state: &mut SessionState,
        op: &dyn Op,
        mut inputs: TVec<Arc<Tensor>>,
    ) -> TractResult<TVec<Arc<Tensor>>> {
        let input = args_1!(inputs);
        let op = op.downcast_ref::<Delay>().ok_or("Wrong Op type")?;
        Ok(tvec!(dispatch_datum!(Self::eval_t(input.datum_type())(self, op, input))?))
    }
}

#[derive(Clone, Debug, PartialEq, Hash)]
pub struct Delay {
    datum_type: DatumType,
    buffer_shape: TVec<usize>,
    axis: usize,
    delay: usize,
    overlap: usize,
}

tract_linalg::impl_dyn_hash!(Delay);

impl Delay {
    pub fn new(input_fact: &PulsedFact, delay: usize, overlap: usize) -> Delay {
        let axis = input_fact.axis;
        let mut buffer_shape = input_fact.shape.clone();
        buffer_shape[axis] = delay + overlap;
        Delay { datum_type: input_fact.datum_type, buffer_shape, axis, delay, overlap }
    }
}

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

    fn info(&self) -> TractResult<Vec<String>> {
        Ok(vec![
            format!("axis: {} delay: {} overlap: {}", self.axis, self.delay, self.overlap),
            format!("buffer: {:?} {:?}", self.buffer_shape, self.datum_type),
        ])
    }

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

impl StatefullOp for Delay {
    fn state(
        &self,
        _session: &mut SessionState,
        _node_id: usize,
    ) -> TractResult<Option<Box<dyn OpState>>> {
        let buffer = unsafe { Tensor::uninitialized_dt(self.datum_type, &*self.buffer_shape)? };
        Ok(Some(Box::new(DelayState { buffer })))
    }
}

impl TypedOp for Delay {
    as_op!();

    fn output_facts(&self, inputs: &[&TypedFact]) -> TractResult<TVec<TypedFact>> {
        let mut fact = inputs[0].clone();
        fact.shape.set_dim(self.axis, fact.shape.dim(self.axis) + self.overlap)?;
        Ok(tvec!(fact))
    }

    fn cost(&self, _inputs: &[&TypedFact]) -> TractResult<TVec<(Cost, TDim)>> {
        Ok(tvec!((
            Cost::Buffer(self.datum_type),
            self.buffer_shape.iter().product::<usize>().to_dim()
        )))
    }
}

impl PulsedOp for Delay {
    fn pulsed_output_facts(&self, inputs: &[&PulsedFact]) -> TractResult<TVec<PulsedFact>> {
        let mut fact = inputs[0].clone();
        fact.shape[self.axis] += self.overlap;
        fact.delay += self.delay + self.overlap;
        Ok(tvec!(fact))
    }

    as_op!();
    pulsed_op_to_typed_op!();
}

#[cfg(test)]
mod test {
    use super::*;
    use crate::*;

    fn test_pulse_delay_over(pulse: usize, delay: usize, overlap: usize) {
        let mut model = PulsedModel::default();
        let fact1 = PulsedFact {
            datum_type: u8::datum_type(),
            shape: tvec![pulse],
            axis: 0,
            dim: TDim::s(),
            delay: 0,
        };
        let source = model.add_source("source", fact1.clone()).unwrap();
        model.wire_node("delay", Delay::new(&fact1, delay, overlap), &[source]).unwrap();
        model.auto_outputs().unwrap();

        let plan = SimplePlan::new(model).unwrap();
        let mut state = crate::plan::SimpleState::new(plan).unwrap();

        for i in 0..5 {
            let input: Vec<u8> = (pulse * i..(pulse * (i + 1))).map(|a| a as u8).collect();
            let expect: Vec<u8> = (pulse * i..(pulse * (i + 1) + overlap))
                .map(|i| i.saturating_sub(delay + overlap) as u8)
                .collect();
            let output = state.run(tvec!(Tensor::from(arr1(&input)))).unwrap();
            let skip = (delay + overlap).saturating_sub(i * pulse).min(pulse + overlap);
            assert_eq!(&output[0].as_slice::<u8>().unwrap()[skip..], &expect[skip..]);
        }
    }

    #[test]
    fn sub_pulse() {
        test_pulse_delay_over(4, 1, 0);
    }

    #[test]
    fn supra_pulse() {
        test_pulse_delay_over(4, 5, 0);
    }

    #[test]
    fn sub_pulse_context() {
        test_pulse_delay_over(4, 0, 2);
    }

    #[test]
    fn supra_pulse_context() {
        test_pulse_delay_over(4, 0, 6);
    }

    #[test]
    fn test_two_delays() {
        let pulse = 4;
        let mut model = PulsedModel::default();
        let fact_0 = PulsedFact {
            datum_type: u8::datum_type(),
            shape: tvec![pulse],
            axis: 0,
            dim: TDim::s(),
            delay: 0,
        };
        let source = model.add_source("source", fact_0.clone()).unwrap();
        let delay_1 = model.wire_node("delay-1", Delay::new(&fact_0, 2, 0), &[source]).unwrap()[0];
        let fact_1 = model.outlet_fact(delay_1).unwrap().clone();
        let delay_2 = model.wire_node("delay-1", Delay::new(&fact_1, 2, 0), &[delay_1]).unwrap();
        model.set_output_outlets(&delay_2).unwrap();

        let plan = SimplePlan::new(model).unwrap();
        let mut state = crate::plan::SimpleState::new(plan).unwrap();

        for i in 0..5 {
            let input: Vec<u8> = (pulse * i..(pulse * (i + 1))).map(|a| a as u8).collect();
            let expect: Vec<u8> =
                (pulse * i..(pulse * (i + 1))).map(|i| i.saturating_sub(4) as u8).collect();
            let skip = 4usize.saturating_sub(i * pulse).min(pulse);
            let output = state.run(tvec!(Tensor::from(arr1(&input)))).unwrap();
            assert_eq!(&output[0].as_slice::<u8>().unwrap()[skip..], &expect[skip..]);
        }
    }
}