burn-jit 0.16.1

Generic backend that can be compiled just-in-time to any shader language target
Documentation 
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

use crate::kernel::{
    launch_binop, launch_scalar_binop, AddOp, DivOp, MulOp, PowOp, RemainderOp, SubOp,
};
use crate::{element::JitElement, tensor::JitTensor};
use crate::{FloatElement, JitRuntime};
use burn_tensor::{ElementConversion, Shape};
use cubecl::client::ComputeClient;
use cubecl::tensor_vectorization_factor;
use cubecl::{calculate_cube_count_elemwise, prelude::*};

pub fn full<R: JitRuntime, E: JitElement>(
    shape: Shape,
    device: &R::Device,
    value: E,
) -> JitTensor<R> {
    let client = R::client(device);

    full_device::<R, E>(client, shape, device.clone(), value)
}

pub fn full_device<R: JitRuntime, E: JitElement>(
    client: ComputeClient<R::Server, R::Channel>,
    shape: Shape,
    device: R::Device,
    value: E,
) -> JitTensor<R> {
    let ndims = shape.num_dims();
    let empty = empty_device::<R, E>(client, device, shape);

    #[cube(launch)]
    pub fn full_kernel<C: Numeric>(tensor: &mut Tensor<C>, value: C) {
        if ABSOLUTE_POS >= tensor.len() {
            return;
        }

        tensor[ABSOLUTE_POS] = value;
    }

    let num_elems = empty.shape.num_elements();
    let vectorization_factor =
        tensor_vectorization_factor(&[4, 2], &empty.shape.dims, &empty.strides, ndims - 1);

    let cube_dim = CubeDim::default();
    let cube_count =
        calculate_cube_count_elemwise(num_elems / vectorization_factor as usize, cube_dim);

    full_kernel::launch::<E, R>(
        &empty.client,
        cube_count,
        cube_dim,
        empty.as_tensor_arg::<E>(vectorization_factor),
        ScalarArg::new(value),
    );

    empty
}

pub fn zeros<R: JitRuntime, E: JitElement>(shape: Shape, device: &R::Device) -> JitTensor<R> {
    let client = R::client(device);

    zeros_device::<R, E>(client, device.clone(), shape)
}

pub fn zeros_device<R: JitRuntime, E: JitElement>(
    client: ComputeClient<R::Server, R::Channel>,
    device: R::Device,
    shape: Shape,
) -> JitTensor<R> {
    full_device::<R, E>(client, shape, device, 0.elem())
}

pub fn ones<R: JitRuntime, E: JitElement>(shape: Shape, device: &R::Device) -> JitTensor<R> {
    let client = R::client(device);

    ones_device::<R, E>(client, device.clone(), shape)
}

pub fn ones_device<R: JitRuntime, E: JitElement>(
    client: ComputeClient<R::Server, R::Channel>,
    device: R::Device,
    shape: Shape,
) -> JitTensor<R> {
    full_device::<R, E>(client, shape, device, 1.elem())
}

pub fn empty_device<R: JitRuntime, E: JitElement>(
    client: ComputeClient<R::Server, R::Channel>,
    device: R::Device,
    shape: Shape,
) -> JitTensor<R> {
    let buffer = client.empty(shape.num_elements() * core::mem::size_of::<E>());

    JitTensor::new_contiguous(client, device, shape, buffer, E::dtype())
}

pub fn add<R: JitRuntime, E: JitElement>(lhs: JitTensor<R>, rhs: JitTensor<R>) -> JitTensor<R> {
    launch_binop::<R, E, AddOp>(lhs, rhs)
}

pub fn add_scalar<R: JitRuntime, E: JitElement>(lhs: JitTensor<R>, rhs: E) -> JitTensor<R> {
    launch_scalar_binop::<R, E, AddOp>(lhs, rhs)
}

pub fn sub<R: JitRuntime, E: JitElement>(lhs: JitTensor<R>, rhs: JitTensor<R>) -> JitTensor<R> {
    launch_binop::<R, E, SubOp>(lhs, rhs)
}

pub fn sub_scalar<R: JitRuntime, E: JitElement>(lhs: JitTensor<R>, rhs: E) -> JitTensor<R> {
    launch_scalar_binop::<R, E, SubOp>(lhs, rhs)
}

pub fn mul<R: JitRuntime, E: JitElement>(lhs: JitTensor<R>, rhs: JitTensor<R>) -> JitTensor<R> {
    launch_binop::<R, E, MulOp>(lhs, rhs)
}

pub fn mul_scalar<R: JitRuntime, E: JitElement>(lhs: JitTensor<R>, rhs: E) -> JitTensor<R> {
    launch_scalar_binop::<R, E, MulOp>(lhs, rhs)
}

pub fn div<R: JitRuntime, E: JitElement>(lhs: JitTensor<R>, rhs: JitTensor<R>) -> JitTensor<R> {
    launch_binop::<R, E, DivOp>(lhs, rhs)
}

pub fn div_scalar<R: JitRuntime, E: JitElement>(lhs: JitTensor<R>, rhs: E) -> JitTensor<R> {
    launch_scalar_binop::<R, E, DivOp>(lhs, rhs)
}

pub fn remainder<R: JitRuntime, E: JitElement>(
    lhs: JitTensor<R>,
    rhs: JitTensor<R>,
) -> JitTensor<R> {
    launch_binop::<R, E, RemainderOp>(lhs, rhs)
}

pub fn remainder_scalar<R: JitRuntime, E: JitElement>(lhs: JitTensor<R>, rhs: E) -> JitTensor<R> {
    launch_scalar_binop::<R, E, RemainderOp>(lhs, rhs)
}

pub fn pow<R: JitRuntime, E: FloatElement>(lhs: JitTensor<R>, rhs: JitTensor<R>) -> JitTensor<R> {
    launch_binop::<R, E, PowOp<E>>(lhs, rhs)
}