cubecl_std/tensor/
handle.rs

1use core::marker::PhantomData;
2use cubecl_core::ir::StorageType;
3use cubecl_core::tensor_line_size_parallel;
4use cubecl_core::{Runtime, server};
5use cubecl_core::{calculate_cube_count_elemwise, server::Allocation};
6use cubecl_core::{prelude::*, server::CopyDescriptor};
7use cubecl_runtime::server::Handle;
8
9/// Tensor representation containing a [server handle](Handle) as well as basic tensor metadata.,
10pub struct TensorHandle<R>
11where
12    R: Runtime,
13{
14    /// The buffer where the data are stored.
15    pub handle: server::Handle,
16    pub shape: Vec<usize>,
17    pub strides: Vec<usize>,
18    /// The type used as storage.
19    pub dtype: StorageType,
20    runtime: PhantomData<R>,
21}
22
23impl<R> core::fmt::Debug for TensorHandle<R>
24where
25    R: Runtime,
26{
27    fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result {
28        f.write_fmt(format_args!(
29            "Tensor {{ shape: {:?}, strides: {:?}, dtype: {}}}",
30            self.shape, self.strides, self.dtype,
31        ))
32    }
33}
34
35impl<R> Clone for TensorHandle<R>
36where
37    R: Runtime,
38{
39    fn clone(&self) -> Self {
40        Self {
41            handle: self.handle.clone(),
42            shape: self.shape.clone(),
43            strides: self.strides.clone(),
44            dtype: self.dtype,
45            runtime: PhantomData,
46        }
47    }
48}
49
50impl<R> TensorHandle<R>
51where
52    R: Runtime,
53{
54    /// Create a new tensor.
55    pub fn new(
56        handle: server::Handle,
57        shape: Vec<usize>,
58        strides: Vec<usize>,
59        storage: StorageType,
60    ) -> Self {
61        Self {
62            handle,
63            shape,
64            strides,
65            dtype: storage,
66            runtime: PhantomData,
67        }
68    }
69
70    pub fn empty(client: &ComputeClient<R>, shape: Vec<usize>, storage: StorageType) -> Self {
71        let elem_size = storage.size();
72        let Allocation { handle, strides } = client.empty_tensor(&shape, elem_size);
73
74        Self::new(handle, shape, strides, storage)
75    }
76
77    /// Create a new tensor.
78    pub fn from_ref(handle: &TensorHandleRef<'_, R>, storage: StorageType) -> Self {
79        Self {
80            handle: handle.handle.clone(),
81            shape: handle.shape.to_vec(),
82            strides: handle.strides.to_vec(),
83            dtype: storage,
84            runtime: PhantomData,
85        }
86    }
87
88    /// Create a new tensor with a contiguous memory layout.
89    pub fn new_contiguous(shape: Vec<usize>, handle: Handle, storage: StorageType) -> Self {
90        let strides = Self::contiguous_strides(&shape);
91
92        Self {
93            handle,
94            shape,
95            strides,
96            dtype: storage,
97            runtime: PhantomData,
98        }
99    }
100
101    /// Check if the tensor is safe to mutate.
102    pub fn can_mut(&self) -> bool {
103        self.handle.can_mut()
104    }
105
106    pub fn as_ref(&self) -> TensorHandleRef<'_, R> {
107        unsafe {
108            TensorHandleRef::from_raw_parts(
109                &self.handle,
110                &self.strides,
111                &self.shape,
112                self.dtype.size(),
113            )
114        }
115    }
116
117    /// Return the reference to a tensor argument.
118    pub fn as_arg<'a>(&'a self, vectorisation: u8) -> TensorArg<'a, R> {
119        let handle: TensorHandleRef<'a, R> = self.as_ref();
120
121        unsafe {
122            TensorArg::from_raw_parts_and_size(
123                handle.handle,
124                handle.strides,
125                handle.shape,
126                vectorisation,
127                handle.elem_size,
128            )
129        }
130    }
131
132    pub fn as_copy_descriptor<'a>(&'a self) -> CopyDescriptor<'a> {
133        CopyDescriptor {
134            binding: self.handle.clone().binding(),
135            shape: &self.shape,
136            strides: &self.strides,
137            elem_size: self.dtype.size(),
138        }
139    }
140
141    fn contiguous_strides(shape: &[usize]) -> Vec<usize> {
142        let mut strides = Vec::with_capacity(shape.len());
143
144        let mut current = 1;
145        shape.iter().enumerate().rev().for_each(|(_, val)| {
146            strides.push(current);
147            current *= val;
148        });
149        strides.reverse();
150        strides
151    }
152}
153impl<R> TensorHandle<R>
154where
155    R: Runtime,
156{
157    pub fn zeros(client: &ComputeClient<R>, shape: Vec<usize>, dtype: StorageType) -> Self {
158        let num_elements: usize = shape.iter().product();
159        let rank = shape.len();
160        let output = Self::empty(client, shape, dtype);
161
162        let line_size = tensor_line_size_parallel(
163            R::supported_line_sizes().iter().cloned(),
164            &output.shape,
165            &output.strides,
166            rank - 1,
167        );
168
169        let cube_dim = CubeDim::default();
170        let cube_count = calculate_cube_count_elemwise(num_elements / line_size as usize, cube_dim);
171        let array_len = output.handle.size() as usize / dtype.size();
172
173        unsafe {
174            init::zeros_array::launch_unchecked(
175                client,
176                cube_count,
177                cube_dim,
178                ArrayArg::from_raw_parts_and_size(
179                    &output.handle,
180                    array_len,
181                    line_size,
182                    dtype.size(),
183                ),
184                dtype,
185            )
186            .expect("Should be able to launch the kernel all the time")
187        };
188
189        output
190    }
191}
192
193pub(crate) mod init {
194    use cubecl::prelude::*;
195    use cubecl_core::{self as cubecl, ir::StorageType};
196
197    #[cube(launch_unchecked)]
198    pub fn zeros_array<C: Numeric>(output: &mut Array<Line<C>>, #[define(C)] _elem: StorageType) {
199        if ABSOLUTE_POS < output.len() {
200            output[ABSOLUTE_POS] = Line::cast_from(C::from_int(0));
201        }
202    }
203}
cubecl_std/tensor/handle.rs

cubecl_std/tensor/
handle.rs
