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
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
use std::{
ops::{Deref, DerefMut},
rc::Rc,
};
use crate::*;
/// A slice into the device memory.
#[derive(Clone)]
pub struct DevicePtr<'a> {
pub(crate) handle: Rc<Handle<'a>>,
pub(crate) inner: u64,
pub(crate) len: u64,
}
impl<'a> DevicePtr<'a> {
pub fn as_raw(&self) -> u64 {
self.inner
}
pub unsafe fn from_raw_parts(handle: Rc<Handle<'a>>, ptr: u64, len: u64) -> Self {
Self {
handle,
inner: ptr,
len,
}
}
/// Synchronously copies data from `self` to `target`. Panics if length is not equal.
pub fn copy_to<'b>(&self, target: &DevicePtr<'b>) -> CudaResult<()> {
if self.len > target.len {
panic!("overflow in DevicePtr::copy_to");
} else if self.len < target.len {
panic!("underflow in DevicePtr::copy_to");
}
if std::ptr::eq(self.handle.context, target.handle.context) {
cuda_error(unsafe { sys::cuMemcpy(target.inner, self.inner, self.len as sys::size_t) })
} else {
cuda_error(unsafe {
sys::cuMemcpyPeer(
target.inner,
target.handle.context.inner,
self.inner,
self.handle.context.inner,
self.len as sys::size_t,
)
})
}
}
/// Asynchronously copies data from `self` to `target`. Panics if length is not equal.
pub fn copy_to_stream<'b, 'c: 'b + 'a>(
&self,
target: &DevicePtr<'b>,
stream: &mut Stream<'c>,
) -> CudaResult<()>
where
'a: 'b,
{
if self.len > target.len {
panic!("overflow in DevicePtr::copy_to");
} else if self.len < target.len {
panic!("underflow in DevicePtr::copy_to");
}
if std::ptr::eq(self.handle.context, target.handle.context) {
cuda_error(unsafe {
sys::cuMemcpyAsync(
target.inner,
self.inner,
self.len as sys::size_t,
stream.inner,
)
})
} else {
cuda_error(unsafe {
sys::cuMemcpyPeerAsync(
target.inner,
target.handle.context.inner,
self.inner,
self.handle.context.inner,
self.len as sys::size_t,
stream.inner,
)
})
}
}
// pub fn copy_to_async<'b>(&self, target: &DevicePtr<'b>) -> CudaResult<CudaFuture<'a, ()>>
// where
// 'a: 'b,
// {
// let mut stream = self.handle.get_async_stream()?;
// unsafe { self.copy_to_stream(target, &mut stream) }?;
// Ok(CudaFuture::new(self.handle.clone(), stream))
// }
/// Synchronously copies data from `source` to `self`. Panics if length is not equal.
pub fn copy_from<'b>(&self, source: &DevicePtr<'b>) -> CudaResult<()> {
source.copy_to(self)
}
/// Asynchronously copies data from `source` to `self`. Panics if length is not equal.
pub fn copy_from_stream<'b: 'a, 'c: 'a + 'b>(
&self,
source: &DevicePtr<'b>,
stream: &mut Stream<'c>,
) -> CudaResult<()> {
source.copy_to_stream(self, stream)
}
/// Gets a subslice of this slice from `[from:to]`
pub fn subslice(&self, from: u64, to: u64) -> Self {
if from > self.len || from > to || to > self.len {
panic!("overflow in DevicePtr::subslice");
}
Self {
handle: self.handle.clone(),
inner: self.inner + from,
len: to - from,
}
}
/// Gets the length of this slice
pub fn len(&self) -> u64 {
self.len
}
/// Check if the slice's length is 0
pub fn is_empty(&self) -> bool {
self.len == 0
}
/// Synchronously loads the data from this slice into a local buffer
pub fn load(&self) -> CudaResult<Vec<u8>> {
let mut buf = Vec::with_capacity(self.len as usize);
cuda_error(unsafe {
sys::cuMemcpyDtoH_v2(buf.as_mut_ptr() as *mut _, self.inner, self.len)
})?;
unsafe { buf.set_len(self.len as usize) };
Ok(buf)
}
/// Asynchronously loads the data from this slice into a local buffer.
/// The contents of the buffer are undefined until `stream.sync` is called.
/// The output must not be dropped until the stream is synced.
pub unsafe fn load_stream(&self, stream: &mut Stream<'a>) -> CudaResult<Vec<u8>> {
let mut buf = Vec::with_capacity(self.len as usize);
cuda_error(sys::cuMemcpyDtoHAsync_v2(
buf.as_mut_ptr() as *mut _,
self.inner,
self.len,
stream.inner,
))?;
buf.set_len(self.len as usize);
Ok(buf)
}
/// Synchronously stores host data from `data` to `self`.
pub fn store(&self, data: &[u8]) -> CudaResult<()> {
if data.len() > self.len as usize {
panic!("overflow in DevicePtr::store");
} else if data.len() < self.len as usize {
panic!("underflow in DevicePtr::store");
}
cuda_error(unsafe {
sys::cuMemcpyHtoD_v2(self.inner, data.as_ptr() as *const _, self.len)
})?;
Ok(())
}
/// Asynchronously stores host data from `data` to `self`.
/// The `data` must not be dropped or mutated until `stream.sync` is called.
pub unsafe fn store_stream(&self, data: &[u8], stream: &mut Stream<'a>) -> CudaResult<()> {
if data.len() > self.len as usize {
panic!("overflow in DevicePtr::store");
} else if data.len() < self.len as usize {
panic!("underflow in DevicePtr::store");
}
cuda_error(sys::cuMemcpyHtoDAsync_v2(
self.inner,
data.as_ptr() as *const _,
self.len,
stream.inner,
))?;
Ok(())
}
/// Synchronously set the contents of `self` to `data` repeated to fill length
pub fn memset_d8(&self, data: u8) -> CudaResult<()> {
cuda_error(unsafe { sys::cuMemsetD8_v2(self.inner, data, self.len) })
}
/// Asynchronously set the contents of `self` to `data` repeated to fill length
pub fn memset_d8_stream(&self, data: u8, stream: &mut Stream<'a>) -> CudaResult<()> {
cuda_error(unsafe { sys::cuMemsetD8Async(self.inner, data, self.len, stream.inner) })
}
/// Synchronously set the contents of `self` to `data` repeated to fill length.
/// Panics if [`Self::len`] is not a multiple of 2.
pub fn memset_d16(&self, data: u16) -> CudaResult<()> {
if self.len % 2 != 0 {
panic!("alignment failure in DevicePtr::memset_d16");
}
cuda_error(unsafe { sys::cuMemsetD16_v2(self.inner, data, self.len / 2) })
}
/// Asynchronously set the contents of `self` to `data` repeated to fill length.
/// Panics if [`Self::len`] is not a multiple of 2.
pub fn memset_d16_stream(&self, data: u16, stream: &mut Stream<'a>) -> CudaResult<()> {
if self.len % 2 != 0 {
panic!("alignment failure in DevicePtr::memset_d16_stream");
}
cuda_error(unsafe { sys::cuMemsetD16Async(self.inner, data, self.len / 2, stream.inner) })
}
/// Synchronously set the contents of `self` to `data` repeated to fill length.
/// Panics if [`Self::len`] is not a multiple of 4.
pub fn memset_d32(&self, data: u32) -> CudaResult<()> {
if self.len % 4 != 0 {
panic!("alignment failure in DevicePtr::memset_d32");
}
cuda_error(unsafe { sys::cuMemsetD32_v2(self.inner, data, self.len / 4) })
}
/// Asynchronously set the contents of `self` to `data` repeated to fill length.
/// Panics if [`Self::len`] is not a multiple of 4.
pub fn memset_d32_stream(&self, data: u32, stream: &mut Stream<'a>) -> CudaResult<()> {
if self.len % 4 != 0 {
panic!("alignment failure in DevicePtr::memset_d32_stream");
}
cuda_error(unsafe { sys::cuMemsetD32Async(self.inner, data, self.len / 4, stream.inner) })
}
/// Gets a reference to the owning handle
pub fn handle(&self) -> &Rc<Handle<'a>> {
&self.handle
}
}
/// An owned device-allocated buffer
pub struct DeviceBox<'a> {
pub(crate) inner: DevicePtr<'a>,
}
impl<'a> DeviceBox<'a> {
/// Allocate an uninitialized buffer of size `size` on the device
pub fn alloc(handle: &Rc<Handle<'a>>, size: u64) -> CudaResult<Self> {
let mut out = 0u64;
cuda_error(unsafe { sys::cuMemAlloc_v2(&mut out as *mut u64, size) })?;
Ok(DeviceBox {
inner: DevicePtr {
handle: handle.clone(),
inner: out,
len: size,
},
})
}
/// Allocate a new initialized buffer on the device matching the size and content of `input`.
pub fn new(handle: &Rc<Handle<'a>>, input: &[u8]) -> CudaResult<Self> {
let buf = Self::alloc(handle, input.len() as u64)?;
buf.store(input)?;
Ok(buf)
}
/// Allocates a new uninitialized buffer on the device, then synchronously fills it with `input`.
/// `input` must not be dropped or mutated until `stream.sync` is called.
/// Does not allocate the memory asynchronously.
pub unsafe fn new_stream(
handle: &Rc<Handle<'a>>,
input: &[u8],
stream: &mut Stream<'a>,
) -> CudaResult<Self> {
let buf = Self::alloc(handle, input.len() as u64)?;
buf.store_stream(input, stream)?;
Ok(buf)
}
/// Allocates a new initialized buffer on the device matching the size and content of `input`.
/// Note that memory is directly copied, so [`T`] must be [`Sized`] should not contain any pointers, references, unsized types, or other non-FFI safe types.
pub fn new_ffi<T>(handle: &Rc<Handle<'a>>, input: &[T]) -> CudaResult<Self> {
let raw = unsafe {
std::slice::from_raw_parts(
input.as_ptr() as *const u8,
input.len() * std::mem::size_of::<T>(),
)
};
let buf = Self::alloc(handle, raw.len() as u64)?;
buf.store(raw)?;
Ok(buf)
}
/// Allocates a new uninitialized buffer on the device, then synchronously fills it with `input`.
/// Note that memory is directly copied, so [`T`] must be [`Sized`] should not contain any pointers, references, unsized types, or other non-FFI safe types.
/// `input` must not be dropped or mutated until `stream.sync` is called.
/// Does not allocate the memory asynchronously.
pub unsafe fn new_ffi_stream<T>(
handle: &Rc<Handle<'a>>,
input: &[T],
stream: &mut Stream<'a>,
) -> CudaResult<Self> {
let raw = std::slice::from_raw_parts(
input.as_ptr() as *const u8,
input.len() * std::mem::size_of::<T>(),
);
let buf = Self::alloc(handle, raw.len() as u64)?;
buf.store_stream(raw, stream)?;
Ok(buf)
}
/// Leaks the DeviceBox, similar to [`Box::leak`].
pub fn leak(self) {
std::mem::forget(self);
}
/// Constructs a [`DeviceBox`] from a device pointer.
pub unsafe fn from_raw(raw: DevicePtr<'a>) -> Self {
Self { inner: raw }
}
}
impl<'a> Drop for DeviceBox<'a> {
fn drop(&mut self) {
if let Err(e) = cuda_error(unsafe { sys::cuMemFree_v2(self.inner.inner) }) {
eprintln!("CUDA: failed freeing device buffer: {:?}", e);
}
}
}
impl<'a> AsRef<DevicePtr<'a>> for DeviceBox<'a> {
fn as_ref(&self) -> &DevicePtr<'a> {
&self.inner
}
}
impl<'a> Deref for DeviceBox<'a> {
type Target = DevicePtr<'a>;
fn deref(&self) -> &Self::Target {
&self.inner
}
}
impl<'a> DerefMut for DeviceBox<'a> {
fn deref_mut(&mut self) -> &mut Self::Target {
&mut self.inner
}
}