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

pub mod mat_mat_mul;
pub mod mat_mul;

pub use self::mat_mul::MatMul;
use crate::internal::*;
use num_traits::{Float, Zero};

use super::binary::*;

bin_to_super_type!(add, Add,
        flip:commute,
        validation: Validation::Rounding,
     [f32, i8, i16, i32, i64, u8, u16, f16, f64, TDim] => |c, a, b| *c = a.clone() + b);
bin_to_super_type!(sub, Sub, flip:flip_sub,
     [f32, i8, i16, i32, i64, u8, u16, f16, f64, TDim] => |c, a, b| *c = a.clone() - b);
#[inline]
bin_to_super_type!(mul, Mul,
        cost: |dt| tvec!((Cost::FMA(dt), 1)),
        flip:commute,
     [f32, i8, i16, i32, i64, u8, u16, f16, f64, TDim] => |c, a, b| *c = a.clone() * b);
bin_to_super_type!(div, Div,
        cost: |dt| tvec!((Cost::Div(dt), 1)),
     [f32, i8, i16, i32, i64, u8, u16, f16, f64, TDim] => |c, a, b| *c = a.clone() / b);
bin_to_super_type!(rem, Rem,
     [f32, i8, i16, i32, i64, u8, u16, f16, f64, TDim] => |c, a, b| *c = a.clone() % b);
bin_to_super_type!(min, Min, flip:commute,
     [f32, f64] => |c,a,b| *c = a.min(*b),
     [i8, i16, i32, i64, u8, u16] => |c, a, b| *c = *a.min(b));
bin_to_super_type!(max, Max, flip:commute,
     [f32, f64] => |c,a,b| *c = a.max(*b),
     [i8, i16, i32, i64, u8, u16] => |c, a, b| *c = *a.max(b));
bin_to_super_type!(pow, Pow,
     [f32, f64] => |c,a,b| *c = a.powf(*b));

fn flip_sub(_op: &dyn BinMiniOp, t: &Arc<Tensor>) -> Option<UnaryOp> {
    let mut t = t.clone().into_tensor();
    fn negate<T: Datum + std::ops::Neg<Output = T>>(t: &mut Tensor) {
        t.as_slice_mut::<T>().unwrap().iter_mut().for_each(|p| *p = -p.clone());
    }
    (|t: &mut Tensor| -> TractResult<()> {
        dispatch_signed!(negate(t.datum_type())(t));
        Ok(())
    })(&mut t)
    .unwrap();
    Some(UnaryOp::new(Box::new(Add), Arc::new(t)))
}

element_wise!(abs, Abs, [f16, f32, i32] => |_, xs| {
    xs.iter_mut().for_each(|x| *x = x.abs());
    Ok(())
});

element_wise!(exp, Exp, [f16, f32, f64] => |_, xs| {
    xs.iter_mut().for_each(|x| *x = x.exp());
    Ok(())
};
    validation: Validation::Rounding
);

element_wise!(ln, Ln, [f16, f32, f64] => |_, xs| {
    xs.iter_mut().for_each(|x| *x = x.ln());
    Ok(())
};
    validation: Validation::Rounding
);

element_wise!(sqrt, Sqrt, [f16, f32, f64] => |_, xs| {
    xs.iter_mut().for_each(|x| *x = x.sqrt());
    Ok(())
};
    validation: Validation::Rounding
);

element_wise!(recip, Recip, [f16, f32, f64] => |_, xs| {
    xs.iter_mut().for_each(|x| *x = x.recip());
    Ok(())
};
    validation: Validation::Rounding
);

element_wise!(rsqrt, Rsqrt, [f16, f32, f64] => |_, xs| {
    xs.iter_mut().for_each(|x| *x = x.sqrt().recip());
    Ok(())
};
    validation: Validation::Rounding
);

element_wise!(ceil, Ceil, [f16, f32, f64] => |_, xs| {
    xs.iter_mut().for_each(|x| *x = x.ceil());
    Ok(())
});

element_wise!(floor, Floor, [f16, f32, f64] => |_, xs| {
    xs.iter_mut().for_each(|x| *x = x.floor());
    Ok(())
});

element_wise!(scalar_min_max, ScalarMinMax { min: Tensor, max: Tensor },
   [f32, f64] => |m, xs| {
        let max = m.max.cast_to_scalar()?;
        let min = m.min.cast_to_scalar()?;
        xs.iter_mut().for_each(|x| { *x = x.max(max).min(min) });
        Ok(())
});

element_wise!(scalar_min, ScalarMin { min: Tensor },
   [f32, f64] => |m, xs| {
        let min = m.min.cast_to_scalar()?;
        xs.iter_mut().for_each(|x| *x = x.min(min));
        Ok(())
});

element_wise!(scalar_max, ScalarMax { max: Tensor },
   [f32, f64] => |m, xs| {
        let max = m.max.cast_to_scalar()?;
        xs.iter_mut().for_each(|x| *x = x.max(max));
        Ok(())
});

element_wise!(cos, Cos, [f16, f32, f64] => |_, xs| {
    xs.iter_mut().for_each(|x| *x = x.cos());
    Ok(())
});

element_wise!(sin, Sin, [f16, f32, f64] => |_, xs| {
    xs.iter_mut().for_each(|x| *x = x.sin());
    Ok(())
});

element_wise!(tan, Tan, [f16, f32, f64] => |_, xs| {
    xs.iter_mut().for_each(|x| *x = x.tan());
    Ok(())
});

element_wise!(acos, Acos, [f16, f32, f64] => |_, xs| {
    xs.iter_mut().for_each(|x| *x = x.acos());
    Ok(())
});

element_wise!(asin, Asin, [f16, f32, f64] => |_, xs| {
    xs.iter_mut().for_each(|x| *x = x.asin());
    Ok(())
});

element_wise!(atan, Atan, [f16, f32, f64] => |_, xs| {
    xs.iter_mut().for_each(|x| *x = x.atan());
    Ok(())
});

element_wise!(cosh, Cosh, [f16, f32, f64] => |_, xs| {
    xs.iter_mut().for_each(|x| *x = x.cosh());
    Ok(())
});

element_wise!(sinh, Sinh, [f16, f32, f64] => |_, xs| {
    xs.iter_mut().for_each(|x| *x = x.sinh());
    Ok(())
});

element_wise!(tanh, Tanh,
   [f32] => |_, xs| { (tract_linalg::ops().stanh)().run(xs); Ok(()) },
   [f16, f64] => |_, xs| { xs.iter_mut().for_each(|x| *x = x.tanh()); Ok(()) };
   cost: |dt| {tvec!((Cost::FMA(dt), 11), (Cost::Div(dt), 1))}
);

element_wise!(acosh, Acosh, [f16, f32, f64] => |_, xs| { xs.iter_mut().for_each(|x| *x = x.acosh()); Ok(()) });
element_wise!(asinh, Asinh, [f16, f32, f64] => |_, xs| { xs.iter_mut().for_each(|x| *x = x.asinh()); Ok(()) });
element_wise!(atanh, Atanh, [f16, f32, f64] => |_, xs| { xs.iter_mut().for_each(|x| *x = x.atanh()); Ok(()) });

element_wise!(neg, Neg, [i8, i16, i32, i64, f16, f32, f64, TDim] => |_, xs| {
    xs.iter_mut().for_each(|x| *x = -x.clone());
    Ok(())
});

element_wise!(sign, Sign, [f16, f32, f64] => |_, xs| {
    xs.iter_mut().for_each(|x| *x = if x.is_zero() { *x } else { x.signum() });
    Ok(())
});

#[cfg(test)]
mod tests {
    use ndarray::arr2;
    #[test]
    fn mul() {
        let a = arr2(&[[1., 2.], [3., 4.]]);
        let b = arr2(&[[1., 0.], [0., 0.]]);
        assert_eq!(a * b, arr2(&[[1., 0.], [0., 0.]]));
    }
    #[test]
    fn dot() {
        let a = arr2(&[[1., 2.], [3., 4.]]);
        let b = arr2(&[[1., 0.], [0., 0.]]);
        assert_eq!(a.dot(&b), arr2(&[[1., 0.], [3., 0.]]));
    }
}