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
|
use crate::{CpuStorage, DType, Shape};
use candle_kernels as kernels;
use cudarc::driver::{CudaFunction, CudaSlice, LaunchAsync, LaunchConfig};
/// cudarc related errors
#[derive(thiserror::Error, Debug)]
pub enum CudaError {
#[error(transparent)]
Cuda(#[from] cudarc::driver::DriverError),
#[error(transparent)]
Compiler(#[from] cudarc::nvrtc::CompileError),
#[error("{op} only supports contiguous tensors")]
RequiresContiguous { op: &'static str },
#[error("missing kernel '{module_name}'")]
MissingKernel { module_name: &'static str },
}
type Result<T> = std::result::Result<T, CudaError>;
#[derive(Debug, Clone)]
pub struct CudaDevice(std::sync::Arc<cudarc::driver::CudaDevice>);
impl CudaDevice {
pub(crate) fn new(ordinal: usize) -> Result<Self> {
let device = cudarc::driver::CudaDevice::new(ordinal)?;
Ok(Self(device))
}
pub(crate) fn ordinal(&self) -> usize {
self.0.ordinal()
}
pub(crate) fn zeros_impl(&self, shape: &Shape, dtype: DType) -> Result<CudaStorage> {
let elem_count = shape.elem_count();
match dtype {
DType::F32 => {
let data = self.0.alloc_zeros::<f32>(elem_count)?;
Ok(CudaStorage::F32(data))
}
DType::F64 => {
let data = self.0.alloc_zeros::<f64>(elem_count)?;
Ok(CudaStorage::F64(data))
}
}
}
pub(crate) fn const_impl(&self, v: f64, shape: &Shape, dtype: DType) -> Result<CudaStorage> {
let elem_count = shape.elem_count();
let cfg = LaunchConfig::for_num_elems(elem_count as u32);
let dev = &self.0;
match dtype {
DType::F32 => {
// SAFETY: Set later by running the fill kernel.
let data = unsafe { dev.alloc::<f32>(elem_count) }?;
let func = self.get_or_load_func("fill_f32", kernels::FILL)?;
let params = (&data, v as f32, elem_count);
unsafe { func.launch(cfg, params) }?;
Ok(CudaStorage::F32(data))
}
DType::F64 => {
// SAFETY: Set later by running the fill kernel.
let data = unsafe { dev.alloc::<f64>(elem_count) }?;
let func = self.get_or_load_func("fill_f64", kernels::FILL)?;
let params = (&data, v, elem_count);
unsafe { func.launch(cfg, params) }?;
Ok(CudaStorage::F64(data))
}
}
}
pub(crate) fn ones_impl(&self, shape: &Shape, dtype: DType) -> Result<CudaStorage> {
self.const_impl(1., shape, dtype)
}
pub(crate) fn cuda_from_cpu_storage(&self, storage: &CpuStorage) -> Result<CudaStorage> {
match storage {
CpuStorage::F32(storage) => {
let data = self.0.htod_sync_copy(storage)?;
Ok(CudaStorage::F32(data))
}
CpuStorage::F64(storage) => {
let data = self.0.htod_sync_copy(storage)?;
Ok(CudaStorage::F64(data))
}
}
}
fn get_or_load_func(
&self,
module_name: &'static str,
ptx: &'static str,
) -> Result<CudaFunction> {
let dev = &self.0;
if !dev.has_func(module_name, module_name) {
dev.load_ptx(ptx.into(), module_name, &[module_name])?;
}
dev.get_func(module_name, module_name)
// Clippy recommends this `ok_or` rather than `ok_or_else` so hopefully the compiler is
// able to only build the error value if needed.
.ok_or(CudaError::MissingKernel { module_name })
}
}
#[derive(Debug, Clone)]
pub enum CudaStorage {
F32(CudaSlice<f32>),
F64(CudaSlice<f64>),
}
impl CudaStorage {
pub fn dtype(&self) -> DType {
match self {
Self::F32(_) => DType::F32,
Self::F64(_) => DType::F64,
}
}
pub fn device(&self) -> CudaDevice {
match self {
Self::F32(slice) => CudaDevice(slice.device()),
Self::F64(slice) => CudaDevice(slice.device()),
}
}
pub(crate) fn affine_impl(
&self,
shape: &Shape,
stride: &[usize],
mul: f64,
add: f64,
) -> Result<Self> {
if !shape.is_contiguous(stride) {
return Err(CudaError::RequiresContiguous { op: "affine" });
}
let elem_count = shape.elem_count();
let cfg = LaunchConfig::for_num_elems(elem_count as u32);
let dev = self.device();
match self {
Self::F32(arg) => {
let func = dev.get_or_load_func("affine_f32", kernels::AFFINE)?;
// SAFETY: if this function returns Ok(..), the kernel has been applied
// and has set the initially unset memory.
let out = unsafe { dev.0.alloc::<f32>(elem_count) }?;
let params = (elem_count, arg, &out, mul as f32, add as f32);
// SAFETY: well, well, well...
unsafe { func.launch(cfg, params) }?;
Ok(Self::F32(out))
}
Self::F64(arg) => {
let func = dev.get_or_load_func("affine_f64", kernels::AFFINE)?;
// SAFETY: if this function returns Ok(..), the kernel has been applied
// and has set the initially unset memory.
let out = unsafe { dev.0.alloc::<f64>(elem_count) }?;
let params = (elem_count, arg, &out, mul, add);
// SAFETY: well, well, well...
unsafe { func.launch(cfg, params) }?;
Ok(Self::F64(out))
}
}
}
pub(crate) fn to_cpu_storage(&self) -> Result<CpuStorage> {
match self {
Self::F32(slice) => {
let dev = slice.device();
let cpu_storage = dev.dtoh_sync_copy(slice)?;
Ok(CpuStorage::F32(cpu_storage))
}
Self::F64(slice) => {
let dev = slice.device();
let cpu_storage = dev.dtoh_sync_copy(slice)?;
Ok(CpuStorage::F64(cpu_storage))
}
}
}
}
|
