tract_gpu/ops/
rotate_half.rs1use crate::tensor::{DeviceTensor, DeviceTensorExt};
2use derive_new::new;
3use tract_core::internal::*;
4
5pub type DispatchRotateHalfFn = fn(&DeviceTensor, &DeviceTensor) -> TractResult<()>;
6
7#[derive(Clone, new)]
8pub struct GpuRotateHalf {
9 pub backend_name: &'static str,
10 pub dispatch: DispatchRotateHalfFn,
11}
12
13impl std::fmt::Debug for GpuRotateHalf {
14 fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
15 write!(f, "{}RotateHalf", self.backend_name)
16 }
17}
18
19impl PartialEq for GpuRotateHalf {
20 fn eq(&self, other: &Self) -> bool {
21 self.backend_name == other.backend_name
22 }
23}
24
25impl Eq for GpuRotateHalf {}
26
27impl std::hash::Hash for GpuRotateHalf {
28 fn hash<H: std::hash::Hasher>(&self, state: &mut H) {
29 self.backend_name.hash(state);
30 }
31}
32
33impl Op for GpuRotateHalf {
34 fn name(&self) -> StaticName {
35 format!("{}RotateHalf", self.backend_name).into()
36 }
37
38 op_as_typed_op!();
39}
40
41impl EvalOp for GpuRotateHalf {
42 fn is_stateless(&self) -> bool {
43 true
44 }
45
46 fn eval_with_session(
47 &self,
48 node_id: usize,
49 session: &TurnState,
50 inputs: TVec<TValue>,
51 ) -> TractResult<TVec<TValue>> {
52 let input_value = args_1!(inputs);
53 let input = input_value.to_device_tensor()?;
54 let output = crate::session_handler::make_tensor_for_node(
55 session,
56 node_id,
57 input.datum_type(),
58 input.shape(),
59 )?;
60 (self.dispatch)(input, &output)?;
61 Ok(tvec!(output.into_tensor().into_tvalue()))
62 }
63}
64
65impl TypedOp for GpuRotateHalf {
66 fn output_facts(&self, inputs: &[&TypedFact]) -> TractResult<TVec<TypedFact>> {
67 crate::utils::facts_to_device_facts(inputs, |facts| {
68 let dt = facts[0].datum_type;
69 let fact = dt.fact(facts[0].shape.clone());
70 Ok(tvec!(fact))
71 })
72 .with_context(|| format!("Error while computing facts for {:?}", self.name()))
73 }
74
75 as_op!();
76}