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

use super::{ElementWise, ElementWiseState};
use crate::{
    element::JitElement, fusion::ElementWiseBuilder, tensor::JitTensor, JitBackend, Runtime,
};
use burn_compute::client::ComputeClient;
use burn_fusion::{client::MutexFusionClient, FusionBackend};
use burn_tensor::Shape;
use core::marker::PhantomData;
use serde::{Deserialize, Serialize};

/// Fusion optimization type for JIT.
///
/// More optimization variants should be added here.
pub enum JitOptimization<R: Runtime> {
    /// Element wise optimization.
    ElementWise(ElementWise<R>),
}

/// Fusion optimization state type for JIT.
///
/// More optimization variants should be added here.
#[derive(Serialize, Deserialize)]
pub enum JitOptimizationState {
    /// Element wise state.
    ElementWise(ElementWiseState),
}

impl<R: Runtime> burn_fusion::Optimization<JitBackend<R>> for JitOptimization<R> {
    fn execute(&mut self, context: &mut burn_fusion::stream::Context<'_, JitBackend<R>>) {
        match self {
            Self::ElementWise(op) => op.execute(context),
        }
    }

    fn len(&self) -> usize {
        match self {
            Self::ElementWise(op) => op.len(),
        }
    }

    fn to_state(&self) -> JitOptimizationState {
        match self {
            Self::ElementWise(value) => JitOptimizationState::ElementWise(value.to_state()),
        }
    }

    fn from_state(device: &R::Device, state: JitOptimizationState) -> Self {
        match state {
            JitOptimizationState::ElementWise(state) => {
                Self::ElementWise(ElementWise::from_state(device, state))
            }
        }
    }
}

impl<R: Runtime> FusionBackend for JitBackend<R> {
    type OptimizationState = JitOptimizationState;
    type Optimization = JitOptimization<R>;
    type FusionDevice = R::Device;
    type Handle = JitFusionHandle<R>;
    type FusionClient = MutexFusionClient<Self>;

    fn optimizations(
        device: R::Device,
    ) -> Vec<Box<dyn burn_fusion::OptimizationBuilder<Self::Optimization>>> {
        vec![Box::new(ElementWiseBuilder::new(device))]
    }

    fn float_tensor<const D: usize>(
        handle: Self::Handle,
        shape: Shape<D>,
    ) -> Self::FloatTensorPrimitive<D> {
        handle.into_tensor(shape)
    }

    fn int_tensor<const D: usize>(
        handle: Self::Handle,
        shape: Shape<D>,
    ) -> Self::IntTensorPrimitive<D> {
        handle.into_tensor(shape)
    }

    fn bool_tensor<const D: usize>(
        handle: Self::Handle,
        shape: Shape<D>,
    ) -> Self::BoolTensorPrimitive<D> {
        handle.into_tensor(shape)
    }

    fn float_tensor_handle<const D: usize>(tensor: Self::FloatTensorPrimitive<D>) -> Self::Handle {
        tensor.into()
    }

    fn int_tensor_handle<const D: usize>(tensor: Self::IntTensorPrimitive<D>) -> Self::Handle {
        tensor.into()
    }

    fn bool_tensor_handle<const D: usize>(tensor: Self::BoolTensorPrimitive<D>) -> Self::Handle {
        tensor.into()
    }
}

pub fn strides_dyn_rank(shape: &[usize]) -> Vec<usize> {
    let mut strides = vec![0; shape.len()];

    let mut current = 1;
    shape.iter().enumerate().rev().for_each(|(index, val)| {
        strides[index] = current;
        current *= val;
    });

    strides
}

/// Handle to be used when fusing operations.
pub struct JitFusionHandle<R: Runtime> {
    /// Compute client for jit.
    pub client: ComputeClient<R::Server, R::Channel>,
    /// The buffer where the data are stored.
    pub handle: burn_compute::server::Handle<R::Server>,
    /// The device of the current tensor.
    pub device: R::Device,
    pub(crate) strides: Vec<usize>,
}

impl<R: Runtime> core::fmt::Debug for JitFusionHandle<R> {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        f.write_fmt(format_args!(
            "JitFusionHandle {{ device: {:?}, runtime: {}}}",
            self.device,
            R::name(),
        ))
    }
}

impl<R: Runtime> Clone for JitFusionHandle<R> {
    fn clone(&self) -> Self {
        Self {
            client: self.client.clone(),
            handle: self.handle.clone(),
            device: self.device.clone(),
            strides: self.strides.clone(),
        }
    }
}

unsafe impl<R: Runtime> Send for JitFusionHandle<R> {}
unsafe impl<R: Runtime> Sync for JitFusionHandle<R> {}

impl<R: Runtime> JitFusionHandle<R> {
    pub(crate) fn into_tensor<const D: usize, E: JitElement>(
        self,
        shape: Shape<D>,
    ) -> JitTensor<R, E, D> {
        JitTensor {
            client: self.client,
            handle: self.handle,
            device: self.device,
            shape,
            strides: self.strides.try_into().expect("Wrong dimension"),
            elem: PhantomData,
        }
    }
}

impl<R: Runtime, E: JitElement, const D: usize> From<JitTensor<R, E, D>> for JitFusionHandle<R> {
    fn from(value: JitTensor<R, E, D>) -> Self {
        Self {
            client: value.client,
            handle: value.handle,
            device: value.device,
            strides: value.strides.into(),
        }
    }
}