hpt 0.1.3

High Performance Tensor (HPT) - A fast, efficient, and user-friendly tensor computation library for Rust
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

use cudarc::driver::DeviceRepr;
use hpt_common::error::base::TensorError;
use hpt_cudakernels::PAD;
use hpt_traits::{
    ops::{
        advance::{AdvancedOps, HardMax, TensorWhere},
        creation::TensorCreator,
    },
    tensor::{CommonBounds, TensorInfo},
};
use hpt_types::{
    dtype::CudaType,
    into_scalar::Cast,
    type_promote::{Cmp, NormalOut},
};

use crate::{
    backend::Cuda,
    backends::cuda::cuda_utils::{compute_kernel_launch_config, load_ptx_and_get_data},
    tensor_base::_Tensor,
};
use cudarc::driver::LaunchAsync;
use hpt_allocator::traits::{Allocator, AllocatorOutputRetrive};
impl<T: CommonBounds + PartialOrd + DeviceRepr + CudaType, const DEVICE: usize, Al> AdvancedOps
    for _Tensor<T, Cuda, DEVICE, Al>
where
    T: NormalOut<bool, Output = T> + Cast<i64>,
    f64: Cast<T>,
    Al: Allocator,
    Al::Output: AllocatorOutputRetrive,
{
    type Meta = T;
    type Output = _Tensor<T, Cuda, DEVICE, Al>;

    type IndexOutput = _Tensor<i64, Cuda, DEVICE, Al>;

    fn pad(&self, pads: &[(i64, i64)], val: Self::Meta) -> Result<Self::Output, TensorError> {
        let res_shape = self
            .shape()
            .iter()
            .zip(pads.iter())
            .map(|(x, (a, b))| x + a + b)
            .collect::<Vec<_>>();

        let res = _Tensor::<T, Cuda, DEVICE, Al>::full(val, &res_shape)?;
        let mut pads = pads.to_vec();
        if pads.len() < self.ndim() {
            pads.resize(self.ndim(), (0, 0));
        }
        let pads_start = pads.iter().map(|(a, _)| *a).collect::<Vec<_>>();
        let pads_end = pads.iter().map(|(_, b)| *b).collect::<Vec<_>>();
        let mut cuda_pads_start = unsafe { self.device().alloc::<i64>(pads.len())? };
        self.device()
            .htod_copy_into(pads_start, &mut cuda_pads_start)?;
        let mut cuda_pads_end = unsafe { self.device().alloc::<i64>(pads.len())? };
        self.device().htod_copy_into(pads_end, &mut cuda_pads_end)?;

        let (kernel, reg_info) = load_ptx_and_get_data(
            "pad",
            &format!("pad_{}", T::STR),
            self.device(),
            self.device_cap(),
            &PAD,
        )
        .unwrap();

        let cfg = compute_kernel_launch_config(self.device(), &reg_info, res.size());

        let res_cuda_shape = res.cuda_shape()?;
        let res_cuda_strides = res.cuda_strides()?;
        let self_cuda_shape = self.cuda_shape()?;
        let self_cuda_strides = self.cuda_strides()?;
        unsafe {
            kernel.launch(
                cfg,
                (
                    res.cuda_slice(),
                    self.cuda_slice(),
                    val,
                    &res_cuda_shape,
                    &res_cuda_strides,
                    &self_cuda_shape,
                    &self_cuda_strides,
                    self.ndim(),
                    &cuda_pads_start,
                    &cuda_pads_end,
                    res.size(),
                ),
            )?
        };

        Ok(res)
    }

    fn topk(
        &self,
        _: i64,
        _: i64,
        _: bool,
        _: bool,
    ) -> Result<(Self::IndexOutput, Self::Output), TensorError> {
        unimplemented!()
    }

    fn onehot(
        &self,
        _: usize,
        _: i64,
        _: Self::Meta,
        _: Self::Meta,
    ) -> Result<Self::Output, TensorError> {
        unimplemented!()
    }

    fn scatter(
        &self,
        _: &Self::IndexOutput,
        _: i64,
        _: &Self::Output,
    ) -> Result<Self::Output, TensorError> {
        unimplemented!()
    }
}

impl<T, const DEVICE: usize, Al> HardMax<T> for _Tensor<T, Cuda, DEVICE, Al>
where
    T: CommonBounds + Cmp<Output = bool>,
    bool: NormalOut<T> + Cast<T>,
    Al: Allocator,
    Al::Output: AllocatorOutputRetrive,
{
    type Output = _Tensor<T, Cuda, DEVICE, Al>;
    fn hardmax(&self, _: i64) -> Result<Self::Output, TensorError> {
        unimplemented!()
    }
}

impl<T, const DEVICE: usize, Al> TensorWhere for _Tensor<T, Cuda, DEVICE, Al>
where
    T: CommonBounds,
    Al: Allocator,
    Al::Output: AllocatorOutputRetrive,
{
    type Output = _Tensor<T, Cuda, DEVICE, Al>;
    type Condition = _Tensor<bool, Cuda, DEVICE, Al>;
    fn tensor_where(
        _: &Self::Condition,
        _: &Self::Output,
        _: &Self::Output,
    ) -> Result<Self::Output, TensorError> {
        unimplemented!()
    }
}