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

use crate::op::{BinaryOpT, CmpOp, ReduceOp, UnaryOpT};
use crate::{CpuStorage, DType, Layout, Result, Shape};

pub trait BackendStorage: Sized {
    type Device: BackendDevice;

    fn try_clone(&self, _: &Layout) -> Result<Self>;

    fn dtype(&self) -> DType;

    fn device(&self) -> &Self::Device;

    // Maybe this should return a Cow instead so that no copy is done on the cpu case.
    fn to_cpu_storage(&self) -> Result<CpuStorage>;

    fn affine(&self, _: &Layout, _: f64, _: f64) -> Result<Self>;

    fn powf(&self, _: &Layout, _: f64) -> Result<Self>;

    fn elu(&self, _: &Layout, _: f64) -> Result<Self>;

    fn reduce_op(&self, _: ReduceOp, _: &Layout, _: &[usize]) -> Result<Self>;

    fn cmp(&self, _: CmpOp, _: &Self, _: &Layout, _: &Layout) -> Result<Self>;

    fn to_dtype(&self, _: &Layout, _: DType) -> Result<Self>;

    fn unary_impl<B: UnaryOpT>(&self, _: &Layout) -> Result<Self>;

    fn binary_impl<B: BinaryOpT>(&self, _: &Self, _: &Layout, _: &Layout) -> Result<Self>;

    fn where_cond(&self, _: &Layout, _: &Self, _: &Layout, _: &Self, _: &Layout) -> Result<Self>;

    fn conv1d(
        &self,
        _l: &Layout,
        _kernel: &Self,
        _kernel_l: &Layout,
        _params: &crate::conv::ParamsConv1D,
    ) -> Result<Self>;

    fn conv_transpose1d(
        &self,
        _l: &Layout,
        _kernel: &Self,
        _kernel_l: &Layout,
        _params: &crate::conv::ParamsConvTranspose1D,
    ) -> Result<Self>;

    fn conv2d(
        &self,
        _l: &Layout,
        _kernel: &Self,
        _kernel_l: &Layout,
        _params: &crate::conv::ParamsConv2D,
    ) -> Result<Self>;

    fn conv_transpose2d(
        &self,
        _l: &Layout,
        _kernel: &Self,
        _kernel_l: &Layout,
        _params: &crate::conv::ParamsConvTranspose2D,
    ) -> Result<Self>;

    fn avg_pool2d(&self, _: &Layout, _: (usize, usize), _: (usize, usize)) -> Result<Self>;
    fn max_pool2d(&self, _: &Layout, _: (usize, usize), _: (usize, usize)) -> Result<Self>;
    fn upsample_nearest1d(&self, _: &Layout, _: usize) -> Result<Self>;
    fn upsample_nearest2d(&self, _: &Layout, _: usize, _: usize) -> Result<Self>;

    fn gather(&self, _: &Layout, _: &Self, _: &Layout, _: usize) -> Result<Self>;
    fn scatter_add(
        &self,
        _: &Layout,
        _: &Self,
        _: &Layout,
        _: &Self,
        _: &Layout,
        _: usize,
    ) -> Result<Self>;
    fn index_select(&self, _: &Self, _: &Layout, _: &Layout, _: usize) -> Result<Self>;
    fn index_add(
        &self,
        _: &Layout,
        _: &Self,
        _: &Layout,
        _: &Self,
        _: &Layout,
        _: usize,
    ) -> Result<Self>;

    fn matmul(
        &self,
        _: &Self,
        _: (usize, usize, usize, usize),
        _: &Layout,
        _: &Layout,
    ) -> Result<Self>;

    fn copy_strided_src(&self, _: &mut Self, _: usize, _: &Layout) -> Result<()>;

    #[allow(clippy::too_many_arguments)]
    // Similar to cudaMemcpy2D, though values are in elements and not in bytes.
    fn copy2d(
        &self,
        _: &mut Self,
        _d1: usize,
        _d2: usize,
        _src_stride1: usize,
        _dst_stride1: usize,
        _src_offset: usize,
        _dst_offset: usize,
    ) -> Result<()>;
}

pub trait BackendDevice: Sized + std::fmt::Debug + Clone {
    type Storage: BackendStorage;

    // TODO: Make the usize generic and part of a generic DeviceLocation.
    fn new(_: usize) -> Result<Self>;

    fn location(&self) -> crate::DeviceLocation;

    fn same_device(&self, _: &Self) -> bool;

    fn zeros_impl(&self, _shape: &Shape, _dtype: DType) -> Result<Self::Storage>;

    fn ones_impl(&self, _shape: &Shape, _dtype: DType) -> Result<Self::Storage>;

    /// # Safety
    /// This function is unsafe as it doesn't initialize the underlying data store.
    /// The caller should ensure that the data is properly initialized as early as possible
    /// after this call.
    unsafe fn alloc_uninit(&self, _shape: &Shape, _dtype: DType) -> Result<Self::Storage>;

    fn storage_from_slice<T: crate::WithDType>(&self, _: &[T]) -> Result<Self::Storage>;

    fn storage_from_cpu_storage(&self, _: &CpuStorage) -> Result<Self::Storage>;

    fn storage_from_cpu_storage_owned(&self, _: CpuStorage) -> Result<Self::Storage>;

    fn rand_uniform(&self, _: &Shape, _: DType, _: f64, _: f64) -> Result<Self::Storage>;

    fn rand_normal(&self, _: &Shape, _: DType, _: f64, _: f64) -> Result<Self::Storage>;

    fn set_seed(&self, _: u64) -> Result<()>;

    /// Synchronize should block until all the operations on the device are completed.
    fn synchronize(&self) -> Result<()>;
}