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

use std::fmt::Debug;

use kn_cuda_sys::bindings::cublasOperation_t;
use kn_cuda_sys::wrapper::group::MatMulOperand;
use kn_graph::dtype::DType;
use kn_graph::graph::SliceRange;

use crate::shape::{StridedShape, ViewError};

pub trait OffsetPtr: Debug + Clone {
    fn offset_bytes(self, offset: isize) -> Self;
}

/// A generic Tensor representation.
#[derive(Debug, Clone, Eq, PartialEq)]
pub struct PtrTensor<P> {
    ptr: P,
    dtype: DType,
    shape: StridedShape,
}

impl<P> PtrTensor<P> {
    pub fn from_parts(ptr: P, shape: StridedShape, dtype: DType) -> Self {
        PtrTensor { ptr, shape, dtype }
    }

    pub fn into_ptr(self) -> P {
        self.ptr
    }

    pub fn ptr(&self) -> &P {
        &self.ptr
    }

    pub fn strided_shape(&self) -> &StridedShape {
        &self.shape
    }

    pub fn dtype(&self) -> DType {
        self.dtype
    }

    pub fn dense_size_bytes(&self) -> usize {
        self.strided_shape().size() * self.dtype().size().bytes()
    }

    pub fn map_ptr<K>(self, f: impl FnOnce(P) -> K) -> PtrTensor<K> {
        PtrTensor::from_parts(f(self.ptr), self.shape, self.dtype)
    }
}

impl<P: OffsetPtr> PtrTensor<P> {
    fn offset(&self, offset_elem: isize, shape: StridedShape) -> Self {
        let offset_bytes = self.dtype.size().bytes() as isize * offset_elem;
        Self::from_parts(self.ptr.clone().offset_bytes(offset_bytes), shape, self.dtype)
    }

    pub fn permute(&self, permutation: &[usize]) -> Self {
        self.offset(0, self.shape.permute(permutation))
    }

    pub fn view(&self, new_shape: Vec<usize>) -> Result<Self, ViewError> {
        self.shape.view(new_shape).map(|shape| self.offset(0, shape))
    }

    pub fn broadcast(&self, new_shape: Vec<usize>) -> Self {
        self.offset(0, self.shape.broadcast(new_shape))
    }

    pub fn slice(&self, axis: usize, range: impl Into<SliceRange>) -> Self {
        let range = range.into();

        // use the new shape & strides (which only change along `axis`)
        let result_shape = self.shape.slice(axis, range);

        let offset = if result_shape.size() != 0 {
            // offset initial pointer to account for `start`
            self.strided_shape().strides()[axis] * range.start as isize
        } else {
            0
        };

        self.offset(offset, result_shape)
    }

    pub fn index(&self, axis: usize, index: usize) -> Self {
        let mut new_shape = self.shape.shape().to_vec();
        new_shape.remove(axis);

        self.slice(axis, SliceRange::simple(index, index + 1))
            .view(new_shape)
            .unwrap()
    }

    pub fn flip(&self, axis: usize) -> Self {
        // invert the axis stride
        let result_shape = self.shape.flip(axis);

        let axis_len = self.shape.shape()[axis];
        let offset = if self.shape.size() != 0 && axis_len != 0 {
            // offset so index 0 gets the last element along the axis
            (axis_len - 1) as isize * self.shape.strides()[axis]
        } else {
            0
        };

        self.offset(offset, result_shape)
    }

    pub fn repeat_unary(&self, axis: usize, count: usize) -> Self {
        let result_shape = self.shape.repeat_unary(axis, count);
        self.offset(0, result_shape)
    }
}

impl<P: Clone> PtrTensor<P> {
    //TODO move this somewhere else, this is pretty random
    pub fn to_mat_mul_arg(&self) -> MatMulOperand<P> {
        assert_eq!(self.strided_shape().rank(), 3);

        // prefer col-major in case of a tie, since cublas likes that more
        let (trans, lead_axis) = if self.shape.strides()[1] == 1 {
            (cublasOperation_t::CUBLAS_OP_N, 2)
        } else if self.shape.strides()[2] == 1 {
            (cublasOperation_t::CUBLAS_OP_T, 1)
        } else {
            panic!(
                "GPU matmul operand must be either col- or row-dense, got {:?}",
                self.shape
            )
        };

        MatMulOperand {
            ptr: self.ptr().clone(),
            trans,
            ld: self.shape.strides()[lead_axis] as i32,
            stride: self.strided_shape().strides()[0] as i64,
        }
    }
}