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
|
use crate::{CpuStorage, DType, Result, Shape, Storage, WithDType};
/// A `DeviceLocation` represents a physical device whereas multiple `Device`
/// can live on the same location (typically for cuda devices).
#[derive(Debug, Copy, Clone, PartialEq, Eq, Hash)]
pub enum DeviceLocation {
Cpu,
Cuda { gpu_id: usize },
}
#[derive(Debug, Clone)]
pub enum Device {
Cpu,
Cuda(crate::CudaDevice),
}
// TODO: Should we back the cpu implementation using the NdArray crate or similar?
pub trait NdArray {
fn shape(&self) -> Result<Shape>;
fn to_cpu_storage(&self) -> CpuStorage;
}
impl<S: WithDType> NdArray for S {
fn shape(&self) -> Result<Shape> {
Ok(Shape::from(()))
}
fn to_cpu_storage(&self) -> CpuStorage {
S::to_cpu_storage(&[*self])
}
}
impl<S: WithDType, const N: usize> NdArray for &[S; N] {
fn shape(&self) -> Result<Shape> {
Ok(Shape::from(self.len()))
}
fn to_cpu_storage(&self) -> CpuStorage {
S::to_cpu_storage(self.as_slice())
}
}
impl<S: WithDType> NdArray for &[S] {
fn shape(&self) -> Result<Shape> {
Ok(Shape::from(self.len()))
}
fn to_cpu_storage(&self) -> CpuStorage {
S::to_cpu_storage(self)
}
}
impl<S: WithDType, const N: usize, const M: usize> NdArray for &[[S; N]; M] {
fn shape(&self) -> Result<Shape> {
Ok(Shape::from((M, N)))
}
fn to_cpu_storage(&self) -> CpuStorage {
S::to_cpu_storage_owned(self.concat())
}
}
impl<S: WithDType, const N1: usize, const N2: usize, const N3: usize> NdArray
for &[[[S; N3]; N2]; N1]
{
fn shape(&self) -> Result<Shape> {
Ok(Shape::from((N1, N2, N3)))
}
fn to_cpu_storage(&self) -> CpuStorage {
let mut vec = Vec::new();
vec.reserve(N1 * N2 * N3);
for i1 in 0..N1 {
for i2 in 0..N2 {
vec.extend(self[i1][i2])
}
}
S::to_cpu_storage_owned(vec)
}
}
impl Device {
pub fn new_cuda(ordinal: usize) -> Result<Self> {
Ok(Self::Cuda(crate::CudaDevice::new(ordinal)?))
}
pub fn same_id(&self, rhs: &Self) -> bool {
match (self, rhs) {
(Self::Cpu, Self::Cpu) => true,
(Self::Cuda(lhs), Self::Cuda(rhs)) => lhs.same_id(rhs),
_ => false,
}
}
pub fn location(&self) -> DeviceLocation {
match self {
Self::Cpu => DeviceLocation::Cpu,
Self::Cuda(device) => DeviceLocation::Cuda {
gpu_id: device.ordinal(),
},
}
}
pub fn is_cuda(&self) -> bool {
match self {
Self::Cpu => false,
Self::Cuda(_) => true,
}
}
pub(crate) fn ones(&self, shape: &Shape, dtype: DType) -> Result<Storage> {
match self {
Device::Cpu => {
let storage = CpuStorage::ones_impl(shape, dtype);
Ok(Storage::Cpu(storage))
}
Device::Cuda(device) => {
let storage = device.ones_impl(shape, dtype)?;
Ok(Storage::Cuda(storage))
}
}
}
pub(crate) fn zeros(&self, shape: &Shape, dtype: DType) -> Result<Storage> {
match self {
Device::Cpu => {
let storage = CpuStorage::zeros_impl(shape, dtype);
Ok(Storage::Cpu(storage))
}
Device::Cuda(device) => {
let storage = device.zeros_impl(shape, dtype)?;
Ok(Storage::Cuda(storage))
}
}
}
pub(crate) fn storage<A: NdArray>(&self, array: A) -> Result<Storage> {
match self {
Device::Cpu => Ok(Storage::Cpu(array.to_cpu_storage())),
Device::Cuda(device) => {
let storage = array.to_cpu_storage();
let storage = device.cuda_from_cpu_storage(&storage)?;
Ok(Storage::Cuda(storage))
}
}
}
pub(crate) fn storage_owned<S: WithDType>(&self, data: Vec<S>) -> Result<Storage> {
match self {
Device::Cpu => Ok(Storage::Cpu(S::to_cpu_storage_owned(data))),
Device::Cuda(device) => {
let storage = S::to_cpu_storage_owned(data);
let storage = device.cuda_from_cpu_storage(&storage)?;
Ok(Storage::Cuda(storage))
}
}
}
}
|
