cudarc 0.19.4

Safe and minimal CUDA bindings
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
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

use std::mem::MaybeUninit;

use super::sys;

/// Wrapper around [sys::CUfileOpError] and [std::io::Error]
#[derive(Debug)]
pub enum CufileError {
    Cufile(sys::CUfileOpError),
    IO(std::io::Error),
}

impl sys::CUfileOpError {
    pub fn result(self) -> Result<(), CufileError> {
        match self {
            sys::CUfileOpError::CU_FILE_SUCCESS => Ok(()),
            _ => Err(CufileError::Cufile(self)),
        }
    }
}

impl sys::CUfileError {
    pub fn result(self) -> Result<(), CufileError> {
        match self.err {
            sys::CUfileOpError::CU_FILE_SUCCESS => Ok(()),
            _ => Err(CufileError::Cufile(self.err)),
        }
    }
}

#[cfg(feature = "std")]
impl std::fmt::Display for CufileError {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        write!(f, "{self:?}")
    }
}

#[cfg(feature = "std")]
impl std::error::Error for CufileError {}

/// See [cuda docs](https://docs.nvidia.com/gpudirect-storage/api-reference-guide/index.html#cufiledriveropen)
pub fn driver_open() -> Result<(), CufileError> {
    unsafe { sys::cuFileDriverOpen() }.result()
}

/// See [cuda docs](https://docs.nvidia.com/gpudirect-storage/api-reference-guide/index.html#cufiledriverclose)
pub fn driver_close() -> Result<(), CufileError> {
    unsafe { sys::cuFileDriverClose() }.result()
}

/// See [cuda docs](https://docs.nvidia.com/gpudirect-storage/api-reference-guide/index.html#cufiledrivergetproperties)
pub fn driver_get_properties() -> Result<sys::CUfileDrvProps, CufileError> {
    let mut props = MaybeUninit::uninit();
    unsafe { sys::cuFileDriverGetProperties(props.as_mut_ptr()) }.result()?;
    Ok(unsafe { props.assume_init() })
}

/// See [cuda docs](https://docs.nvidia.com/gpudirect-storage/api-reference-guide/index.html#cufiledriversetpollmode-bool-poll-size-t-poll-threshold-size)
pub fn driver_set_poll_mode(poll: bool, poll_threshold_size: usize) -> Result<(), CufileError> {
    unsafe { sys::cuFileDriverSetPollMode(poll, poll_threshold_size) }.result()
}

/// See [cuda docs](https://docs.nvidia.com/gpudirect-storage/api-reference-guide/index.html#cufiledriversetmaxdirectiosize-size-t-max-direct-io-size)
pub fn driver_set_max_direct_io_size(max_direct_io_size: usize) -> Result<(), CufileError> {
    unsafe { sys::cuFileDriverSetMaxDirectIOSize(max_direct_io_size) }.result()
}

/// See [cuda docs](https://docs.nvidia.com/gpudirect-storage/api-reference-guide/index.html#size-t-max-cache-size)
pub fn driver_set_max_cache_size(max_cache_size: usize) -> Result<(), CufileError> {
    unsafe { sys::cuFileDriverSetMaxCacheSize(max_cache_size) }.result()
}

/// See [cuda docs](https://docs.nvidia.com/gpudirect-storage/api-reference-guide/index.html#cufiledriversetmaxpinnedmemsize-size-t-max-pinned-memory-size)
pub fn driver_set_max_pinned_mem_size(max_pinned_size: usize) -> Result<(), CufileError> {
    unsafe { sys::cuFileDriverSetMaxPinnedMemSize(max_pinned_size) }.result()
}

/// See [cuda docs](https://docs.nvidia.com/gpudirect-storage/api-reference-guide/index.html#cufilehandleregister)
pub fn handle_register(descr: &sys::CUfileDescr_t) -> Result<sys::CUfileHandle_t, CufileError> {
    let mut fh = MaybeUninit::uninit();
    unsafe { sys::cuFileHandleRegister(fh.as_mut_ptr(), descr as *const _ as _) }.result()?;
    Ok(unsafe { fh.assume_init() })
}

/// See [cuda docs](https://docs.nvidia.com/gpudirect-storage/api-reference-guide/index.html#cufilehandlederegister)
///
/// # Safety
/// `fh` must be valid (not deregistered already)
pub unsafe fn handle_deregister(fh: sys::CUfileHandle_t) -> Result<(), CufileError> {
    sys::cuFileHandleDeregister(fh);
    Ok(())
}

/// See [cuda docs](https://docs.nvidia.com/gpudirect-storage/api-reference-guide/index.html#cufileread)
///
/// # Safety
/// Ensure data ranges are valid & pointers are valid
pub unsafe fn read(
    fh: sys::CUfileHandle_t,
    buf_ptr_base: *mut ::core::ffi::c_void,
    size: usize,
    file_offset: i64,
    buf_ptr_offset: i64,
) -> Result<isize, CufileError> {
    let file_offset = sys::off_t::try_from(file_offset)
        .expect("Casting file_offset to platform specific size failed.");
    let buf_ptr_offset = sys::off_t::try_from(buf_ptr_offset)
        .expect("Casting buf_ptr_offset to platform specific size failed.");
    let bytes_read = sys::cuFileRead(fh, buf_ptr_base, size, file_offset, buf_ptr_offset);

    if bytes_read == -1 {
        Err(CufileError::IO(std::io::Error::last_os_error()))
    } else if bytes_read < -1 {
        let errno = -bytes_read;
        assert!(errno > 0);
        let errno = errno as u32;
        let err: sys::CUfileOpError = std::mem::transmute(errno);
        Err(CufileError::Cufile(err))
    } else {
        Ok(bytes_read)
    }
}

/// See [cuda docs](https://docs.nvidia.com/gpudirect-storage/api-reference-guide/index.html#cufilewrite)
/// # Safety
/// Ensure data rangefs are valid & pointers are valid
pub unsafe fn write(
    fh: sys::CUfileHandle_t,
    buf_ptr_base: *mut ::core::ffi::c_void,
    size: usize,
    file_offset: i64,
    buf_ptr_offset: i64,
) -> Result<isize, CufileError> {
    let file_offset = sys::off_t::try_from(file_offset)
        .expect("Casting file_offset to platform specific size failed.");
    let buf_ptr_offset = sys::off_t::try_from(buf_ptr_offset)
        .expect("Casting buf_ptr_offset to platform specific size failed.");

    let bytes_written = sys::cuFileWrite(fh, buf_ptr_base, size, file_offset, buf_ptr_offset);

    if bytes_written == -1 {
        Err(CufileError::IO(std::io::Error::last_os_error()))
    } else if bytes_written < -1 {
        let errno = -bytes_written;
        assert!(errno > 0);
        let errno = errno as u32;
        let err: sys::CUfileOpError = std::mem::transmute(errno);
        Err(CufileError::Cufile(err))
    } else {
        Ok(bytes_written)
    }
}

/// See [cuda docs](https://docs.nvidia.com/gpudirect-storage/api-reference-guide/index.html#cufilebufregister)
///
/// # Safety
/// Pointers & lengths must be valid
pub unsafe fn buf_register(
    buf_ptr_base: *const ::core::ffi::c_void,
    length: usize,
    flags: i32,
) -> Result<(), CufileError> {
    sys::cuFileBufRegister(buf_ptr_base, length, flags).result()
}

/// See [cuda docs](https://docs.nvidia.com/gpudirect-storage/api-reference-guide/index.html#cufilebufderegister)
///
/// # Safety
/// Pointers & lengths must be valid
pub unsafe fn buf_deregister(buf_ptr_base: *const ::core::ffi::c_void) -> Result<(), CufileError> {
    sys::cuFileBufDeregister(buf_ptr_base).result()
}

/// See [cuda docs](https://docs.nvidia.com/gpudirect-storage/api-reference-guide/index.html#cufilestreamregister)
///
/// # Safety
/// `stream` must be valid
#[cfg(not(any(
    feature = "cuda-11040",
    feature = "cuda-11050",
    feature = "cuda-11060",
    feature = "cuda-11070",
    feature = "cuda-11080",
    feature = "cuda-12000",
    feature = "cuda-12010",
)))]
pub unsafe fn stream_register(stream: sys::CUstream, flags: u32) -> Result<(), CufileError> {
    sys::cuFileStreamRegister(stream, flags).result()
}

/// See [cuda docs](https://docs.nvidia.com/gpudirect-storage/api-reference-guide/index.html#cufilestreamderegister)
///
/// # Safety
/// `stream` must be valid
#[cfg(not(any(
    feature = "cuda-11040",
    feature = "cuda-11050",
    feature = "cuda-11060",
    feature = "cuda-11070",
    feature = "cuda-11080",
    feature = "cuda-12000",
    feature = "cuda-12010",
)))]
pub unsafe fn stream_deregister(stream: sys::CUstream) -> Result<(), CufileError> {
    sys::cuFileStreamDeregister(stream).result()
}

/// See [cuda docs](https://docs.nvidia.com/gpudirect-storage/api-reference-guide/index.html#cufileread)
///
/// # Safety
/// Ensure data rangefs are valid & pointers are valid **until execution time**. __NOT__ submission time. This executes
/// asynchronusly wrt host, so utilize pins appropriately.
#[cfg(not(any(
    feature = "cuda-11040",
    feature = "cuda-11050",
    feature = "cuda-11060",
    feature = "cuda-11070",
    feature = "cuda-11080",
    feature = "cuda-12000",
    feature = "cuda-12010",
)))]
pub unsafe fn read_async(
    fh: sys::CUfileHandle_t,
    buf_ptr_base: *mut ::core::ffi::c_void,
    size: &usize,
    file_offset: &i64,
    buf_ptr_offset: &i64,
    bytes_read: &mut isize,
    stream: sys::CUstream,
) -> Result<(), CufileError> {
    sys::cuFileReadAsync(
        fh,
        buf_ptr_base,
        size as *const _ as _,
        file_offset as *const _ as _,
        buf_ptr_offset as *const _ as _,
        bytes_read as _,
        stream,
    )
    .result()
}

/// See [cuda docs](https://docs.nvidia.com/gpudirect-storage/api-reference-guide/index.html#cufilewriteasync)
/// # Safety
/// Ensure data rangefs are valid & pointers are valid **until execution time**. __NOT__ submission time. This executes
/// asynchronusly wrt host, so utilize pins appropriately.
#[cfg(not(any(
    feature = "cuda-11040",
    feature = "cuda-11050",
    feature = "cuda-11060",
    feature = "cuda-11070",
    feature = "cuda-11080",
    feature = "cuda-12000",
    feature = "cuda-12010",
)))]
pub unsafe fn write_async(
    fh: sys::CUfileHandle_t,
    buf_ptr_base: *mut ::core::ffi::c_void,
    size: &usize,
    file_offset: &i64,
    buf_ptr_offset: &i64,
    bytes_written: &mut isize,
    stream: sys::CUstream,
) -> Result<(), CufileError> {
    sys::cuFileWriteAsync(
        fh,
        buf_ptr_base,
        size as *const _ as _,
        file_offset as *const _ as _,
        buf_ptr_offset as *const _ as _,
        bytes_written as _,
        stream,
    )
    .result()
}

#[cfg(any(feature = "cuda-11040", feature = "cuda-11050"))]
mod batch_io {}

#[cfg(not(any(feature = "cuda-11040", feature = "cuda-11050")))]
mod batch_io {
    use super::*;

    /// See [cuda docs](https://docs.nvidia.com/gpudirect-storage/api-reference-guide/index.html#cufilebatchiosetup)
    pub fn batch_io_setup(max_nr: u32) -> Result<sys::CUfileBatchHandle_t, CufileError> {
        let mut batch_idp = MaybeUninit::uninit();
        unsafe { sys::cuFileBatchIOSetUp(batch_idp.as_mut_ptr(), max_nr) }.result()?;
        Ok(unsafe { batch_idp.assume_init() })
    }

    /// See [cuda docs](https://docs.nvidia.com/gpudirect-storage/api-reference-guide/index.html#cufilebatchiosubmit)
    /// # Safety
    /// `handle` must be valid (not destroyed)
    pub unsafe fn batch_io_submit(
        handle: sys::CUfileBatchHandle_t,
        items: &[sys::CUfileIOParams_t],
        flags: u32,
    ) -> Result<(), CufileError> {
        sys::cuFileBatchIOSubmit(handle, items.len() as u32, items.as_ptr() as _, flags).result()
    }

    /// See [cuda docs](https://docs.nvidia.com/gpudirect-storage/api-reference-guide/index.html#cufilebatchiogetstatus)
    ///
    /// # Safety
    /// A lot of weirdness in this api. Just check cuda docs
    pub unsafe fn batch_io_get_status(
        handle: sys::CUfileBatchHandle_t,
        min_nr: u32,
        nr: &mut u32,
        events: &mut [sys::CUfileIOEvents_t],
        timeout: &sys::timespec,
    ) -> Result<(), CufileError> {
        sys::cuFileBatchIOGetStatus(
            handle,
            min_nr,
            nr as _,
            events.as_mut_ptr(),
            timeout as *const _ as _,
        )
        .result()
    }

    /// See [cuda docs](https://docs.nvidia.com/gpudirect-storage/api-reference-guide/index.html#cufilebatchiocancel)
    ///
    /// # Safety
    /// handle must be valid
    pub unsafe fn batch_io_cancel(handle: sys::CUfileBatchHandle_t) -> Result<(), CufileError> {
        sys::cuFileBatchIOCancel(handle).result()
    }

    /// See [cuda docs](https://docs.nvidia.com/gpudirect-storage/api-reference-guide/index.html#cufilebatchiodestroy)
    ///
    /// # Safety
    /// Must not already be destroyed
    pub unsafe fn batch_io_destroy(handle: sys::CUfileBatchHandle_t) -> Result<(), CufileError> {
        sys::cuFileBatchIODestroy(handle);
        Ok(())
    }
}

#[allow(unused_imports)]
pub use self::batch_io::*;