gam_gpu/
memory.rs

1
2#[derive(Clone, Debug)]
3pub struct DeviceBuffer<T> {
4    host_shadow: Vec<T>,
5}
6
7impl<T> DeviceBuffer<T> {
8    pub const fn from_host_shadow(host_shadow: Vec<T>) -> Self {
9        Self { host_shadow }
10    }
11
12    pub const fn len(&self) -> usize {
13        self.host_shadow.len()
14    }
15
16    pub const fn is_empty(&self) -> bool {
17        self.host_shadow.len() == 0
18    }
19
20    pub fn host_shadow(&self) -> &[T] {
21        &self.host_shadow
22    }
23}
24
25#[derive(Clone, Debug)]
26pub struct DeviceVector {
27    pub len: usize,
28    pub data: DeviceBuffer<f64>,
29}
30
31#[derive(Clone, Debug)]
32pub struct DeviceMatrix {
33    pub rows: usize,
34    pub cols: usize,
35    pub data: DeviceBuffer<f64>,
36    pub column_major: bool,
37}
38
39impl DeviceMatrix {
40
41    pub const fn bytes(&self) -> usize {
42        self.rows
43            .saturating_mul(self.cols)
44            .saturating_mul(std::mem::size_of::<f64>())
45    }
46}
47
48#[derive(Clone, Debug)]
49pub struct DeviceCsrMatrix {
50    pub rows: usize,
51    pub cols: usize,
52    pub rowptr: DeviceBuffer<i32>,
53    pub colidx: DeviceBuffer<i32>,
54    pub values: DeviceBuffer<f64>,
55}
56
57impl DeviceCsrMatrix {
58    /// Construct a CSR matrix, enforcing the structural invariant that `rowptr`
59    /// holds exactly `rows + 1` entries.
60    ///
61    /// A CSR row-pointer array must have one slot per row plus a trailing slot
62    /// equal to `nnz`. If the supplied `rowptr` violates this (too short or too
63    /// long), it is canonicalized to `rows + 1` monotone entries: a short
64    /// `rowptr` is padded with its final value (marking the remaining rows as
65    /// empty) and an over-long `rowptr` is truncated. This prevents downstream
66    /// row-slice and deallocation paths from indexing `rowptr[row + 1]` out of
67    /// bounds, which would be an invalid-free / out-of-bounds deallocation
68    /// hazard.
69    pub fn new(
70        rows: usize,
71        cols: usize,
72        rowptr: DeviceBuffer<i32>,
73        colidx: DeviceBuffer<i32>,
74        values: DeviceBuffer<f64>,
75    ) -> Self {
76        let expected = rows + 1;
77        let mut ptr = rowptr.host_shadow().to_vec();
78        if ptr.len() != expected {
79            let fill = ptr.last().copied().unwrap_or(0);
80            ptr.resize(expected, fill);
81        }
82        Self {
83            rows,
84            cols,
85            rowptr: DeviceBuffer::from_host_shadow(ptr),
86            colidx,
87            values,
88        }
89    }
90
91    pub const fn nnz(&self) -> usize {
92        self.values.len()
93    }
94}
gam_gpu/memory.rs

gam_gpu/
memory.rs
