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

use core::alloc::Layout;

use const_type_layout::TypeGraphLayout;

#[expect(clippy::module_name_repetitions)]
#[repr(transparent)]
pub struct ThreadBlockSharedSlice<T: 'static + TypeGraphLayout> {
    shared: *mut [T],
}

impl<T: 'static + TypeGraphLayout> ThreadBlockSharedSlice<T> {
    #[cfg(feature = "host")]
    #[must_use]
    pub fn new_uninit_with_len(len: usize) -> Self {
        Self {
            shared: Self::dangling_slice_with_len(len),
        }
    }

    #[cfg(feature = "host")]
    #[must_use]
    pub fn with_len(mut self, len: usize) -> Self {
        self.shared = Self::dangling_slice_with_len(len);
        self
    }

    #[cfg(feature = "host")]
    #[must_use]
    pub fn with_len_mut(&mut self, len: usize) -> &mut Self {
        self.shared = Self::dangling_slice_with_len(len);
        self
    }

    #[cfg(feature = "host")]
    fn dangling_slice_with_len(len: usize) -> *mut [T] {
        core::ptr::slice_from_raw_parts_mut(core::ptr::NonNull::dangling().as_ptr(), len)
    }

    #[must_use]
    pub fn len(&self) -> usize {
        core::ptr::metadata(self.shared)
    }

    #[must_use]
    pub fn is_empty(&self) -> bool {
        self.len() == 0
    }

    #[must_use]
    pub fn layout(&self) -> Layout {
        // Safety: the length of self.shared is always initialised
        unsafe { Layout::for_value_raw(self.shared) }
    }

    #[cfg(feature = "device")]
    #[must_use]
    pub const fn as_mut_ptr(&self) -> *mut T {
        self.shared.cast()
    }

    #[cfg(feature = "device")]
    #[must_use]
    pub const fn as_mut_slice_ptr(&self) -> *mut [T] {
        self.shared
    }

    #[cfg(feature = "device")]
    /// # Safety
    ///
    /// The provided `index` must not be out of bounds.
    #[inline]
    #[must_use]
    pub unsafe fn index_mut_unchecked<I: core::slice::SliceIndex<[T]>>(
        &self,
        index: I,
    ) -> *mut <I as core::slice::SliceIndex<[T]>>::Output {
        self.shared.get_unchecked_mut(index)
    }
}

#[cfg(feature = "device")]
impl<T: 'static + TypeGraphLayout> ThreadBlockSharedSlice<T> {
    /// # Safety
    ///
    /// Exposing the [`ThreadBlockSharedSlice`] must be preceded by exactly one
    /// call to [`init`].
    pub(crate) unsafe fn with_uninit_for_len<F: FnOnce(&mut Self) -> Q, Q>(
        len: usize,
        inner: F,
    ) -> Q {
        let base: *mut u8;

        unsafe {
            core::arch::asm!(
                "mov.u64    {base}, %rust_cuda_dynamic_shared;",
                base = out(reg64) base,
            );
        }

        let aligned_base = base.byte_add(base.align_offset(core::mem::align_of::<T>()));

        let data: *mut T = aligned_base.cast();

        let new_base = data.add(len).cast::<u8>();

        unsafe {
            core::arch::asm!(
                "mov.u64    %rust_cuda_dynamic_shared, {new_base};",
                new_base = in(reg64) new_base,
            );
        }

        let shared = core::ptr::slice_from_raw_parts_mut(data, len);

        inner(&mut Self { shared })
    }
}

#[cfg(feature = "device")]
/// # Safety
///
/// The thread-block shared dynamic memory must be initialised once and
/// only once per kernel.
pub unsafe fn init() {
    #[expect(clippy::multiple_unsafe_ops_per_block)]
    unsafe {
        core::arch::asm!(".reg .u64    %rust_cuda_dynamic_shared;");
        core::arch::asm!(
            "cvta.shared.u64    %rust_cuda_dynamic_shared, rust_cuda_dynamic_shared_base;",
        );
    }
}

#[cfg(feature = "device")]
core::arch::global_asm!(".extern .shared .align 8 .b8 rust_cuda_dynamic_shared_base[];");

#[cfg(feature = "host")]
pub struct SharedMemorySize {
    last_align: usize,
    total_size: usize,
}

#[cfg(feature = "host")]
impl SharedMemorySize {
    #[must_use]
    pub const fn new() -> Self {
        Self {
            // we allocate the shared memory with an alignment of 8
            last_align: 8,
            total_size: 0,
        }
    }

    pub fn add(&mut self, layout: core::alloc::Layout) {
        if layout.align() > self.last_align {
            // in the worst case, we are one element of the smaller alignment
            //  into the larger alignment, so we need to pad the entire rest
            let pessimistic_padding = layout.align() - self.last_align;

            self.total_size += pessimistic_padding;
        }

        self.last_align = layout.align();
        self.total_size += layout.size();
    }

    pub const fn total(self) -> usize {
        self.total_size
    }
}