dlpark 0.7.0

dlpack Rust binding for Python
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

use cudarc::driver::{CudaContext, CudaSlice, CudaView, DevicePtr, DeviceSlice};
use snafu::ensure;

use crate::error::UnsupportedDeviceSnafu;
use crate::ffi::DeviceType;
use crate::prelude::*;
use crate::traits::{InferDataType, RowMajorCompactLayout, TensorLike};

impl<T> TensorLike<RowMajorCompactLayout> for CudaSlice<T>
where
    T: InferDataType,
{
    type Error = crate::Error;

    fn data_ptr(&self) -> *mut std::ffi::c_void {
        let stream = self.stream();
        let (ptr, _) = self.device_ptr(stream);
        ptr as *mut T as *mut _
    }

    fn memory_layout(&self) -> RowMajorCompactLayout {
        RowMajorCompactLayout::new(vec![self.len() as i64])
    }

    fn device(&self) -> crate::Result<crate::ffi::Device> {
        Ok(crate::ffi::Device::cuda(self.ordinal()))
    }

    fn data_type(&self) -> crate::Result<crate::ffi::DataType> {
        Ok(T::data_type())
    }

    fn byte_offset(&self) -> u64 {
        0
    }
}

impl<T> TensorLike<RowMajorCompactLayout> for CudaView<'_, T>
where
    T: InferDataType,
{
    type Error = crate::Error;

    fn data_ptr(&self) -> *mut std::ffi::c_void {
        let stream = self.stream();
        let (ptr, _) = self.device_ptr(stream);
        ptr as *mut T as *mut _
    }

    fn memory_layout(&self) -> RowMajorCompactLayout {
        RowMajorCompactLayout::new(vec![self.len() as i64])
    }

    fn device(&self) -> crate::Result<crate::ffi::Device> {
        Ok(crate::ffi::Device::cuda(self.stream().context().ordinal()))
    }

    fn data_type(&self) -> crate::Result<crate::ffi::DataType> {
        Ok(T::data_type())
    }

    fn byte_offset(&self) -> u64 {
        0
    }
}

impl<'a, T> TryFrom<&'a SafeManagedTensor> for CudaView<'a, T> {
    type Error = crate::Error;

    fn try_from(value: &'a SafeManagedTensor) -> Result<Self, Self::Error> {
        let device = value.device();
        ensure!(
            device.device_type == DeviceType::Cuda,
            UnsupportedDeviceSnafu {
                device: device.device_type
            }
        );

        let ctx = CudaContext::new(device.device_id as usize).unwrap();
        let stream = ctx.default_stream();
        unsafe {
            let s = stream.upgrade_device_ptr(value.data_ptr() as u64, value.num_elements());
            let temp_view: CudaView<'_, T> = s.slice(..);
            let view = std::mem::transmute::<CudaView<'_, T>, CudaView<'_, T>>(temp_view);
            s.leak();
            Ok(view)
        }
    }
}

impl<'a, T> TryFrom<&'a SafeManagedTensorVersioned> for CudaView<'a, T> {
    type Error = crate::Error;

    fn try_from(value: &'a SafeManagedTensorVersioned) -> Result<Self, Self::Error> {
        let device = value.device();
        ensure!(
            device.device_type == DeviceType::Cuda,
            UnsupportedDeviceSnafu {
                device: device.device_type
            }
        );

        let ctx = CudaContext::new(device.device_id as usize).unwrap();
        let stream = ctx.default_stream();
        unsafe {
            let s = stream.upgrade_device_ptr(value.data_ptr() as u64, value.num_elements());
            let temp_view: CudaView<'_, T> = s.slice(..);
            let view = std::mem::transmute::<CudaView<'_, T>, CudaView<'_, T>>(temp_view);
            s.leak();
            Ok(view)
        }
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_cuda() {
        let ctx = CudaContext::new(0).unwrap();
        let stream = ctx.default_stream();
        let slice = stream.alloc_zeros::<f32>(10).unwrap();
        let mt = SafeManagedTensor::new(slice).unwrap();
        assert_eq!(mt.num_elements(), 10);
        let view = CudaView::<f32>::try_from(&mt).unwrap();
        assert_eq!(view.len(), 10);
    }

    #[test]
    fn test_cuda_view() {
        let ctx = CudaContext::new(0).unwrap();
        let stream = ctx.default_stream();
        let slice = stream.alloc_zeros::<f32>(10).unwrap();
        let view = slice.slice(..);

        let mt = SafeManagedTensorVersioned::new(view).unwrap();
        assert_eq!(mt.num_elements(), 10);
        let view = CudaView::<f32>::try_from(&mt).unwrap();
        assert_eq!(view.len(), 10);
    }
}