tract_gpu/ops/
resize.rs

1use crate::tensor::{DeviceTensor, DeviceTensorExt, IntoDevice};
2use derive_new::new;
3use tract_core::internal::*;
4
5/// Resamples one axis: `output[.., x, ..] = sum_k weights[x, k] * input[.., indices[x, k], ..]`,
6/// with `indices` already clamped into the axis by the host-built plan.
7pub type DispatchResizeAxisFn = fn(
8    input: &DeviceTensor,
9    axis: usize,
10    indices: &DeviceTensor,
11    weights: &DeviceTensor,
12    window: usize,
13    output: &DeviceTensor,
14) -> TractResult<()>;
15
16/// Resize against a plan baked at translation time, one dispatch per resampled
17/// axis. The plan makes the op independent of the interpolator: nearest, linear
18/// and cubic differ only in window size and weights, so translation is limited
19/// to nodes whose shapes and scales are known then. The scales/sizes input is
20/// kept for arity but no longer read.
21#[derive(Clone, new)]
22pub struct GpuResize {
23    pub axes: TVec<usize>,
24    pub windows: TVec<usize>,
25    pub plans: TVec<(Arc<Tensor>, Arc<Tensor>)>,
26    pub output_shape: TVec<usize>,
27    pub backend_name: &'static str,
28    pub dispatch: DispatchResizeAxisFn,
29}
30
31impl std::fmt::Debug for GpuResize {
32    fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
33        write!(f, "{}Resize", self.backend_name)
34    }
35}
36
37impl PartialEq for GpuResize {
38    fn eq(&self, other: &Self) -> bool {
39        self.backend_name == other.backend_name
40            && self.axes == other.axes
41            && self.windows == other.windows
42            && self.output_shape == other.output_shape
43    }
44}
45impl Eq for GpuResize {}
46
47impl std::hash::Hash for GpuResize {
48    fn hash<H: std::hash::Hasher>(&self, state: &mut H) {
49        self.backend_name.hash(state);
50        self.axes.hash(state);
51        self.windows.hash(state);
52        self.output_shape.hash(state);
53    }
54}
55
56impl Op for GpuResize {
57    fn name(&self) -> StaticName {
58        format!("{}Resize", self.backend_name).into()
59    }
60    fn info(&self) -> TractResult<Vec<String>> {
61        Ok(vec![format!("axes={:?} windows={:?}", self.axes, self.windows)])
62    }
63    op_as_typed_op!();
64}
65
66impl EvalOp for GpuResize {
67    fn is_stateless(&self) -> bool {
68        true
69    }
70
71    fn eval_with_session(
72        &self,
73        node_id: usize,
74        session: &TurnState,
75        inputs: TVec<TValue>,
76    ) -> TractResult<TVec<TValue>> {
77        let data = inputs[0].to_device_tensor()?;
78        let dt = data.datum_type();
79        let mut shape: TVec<usize> = data.shape().into();
80        let mut current = data.clone();
81        for (step, (&axis, &window)) in self.axes.iter().zip(&self.windows).enumerate() {
82            let (indices, weights) = &self.plans[step];
83            let indices = indices.as_ref().clone().into_device()?;
84            let weights = weights.as_ref().clone().into_device()?;
85            shape[axis] = self.output_shape[axis];
86            let last = step + 1 == self.axes.len();
87            let output = if last {
88                crate::session_handler::make_tensor_for_node(session, node_id, dt, &shape)?
89            } else {
90                DeviceTensor::uninitialized_dt(dt, &shape)?
91            };
92            (self.dispatch)(&current, axis, &indices, &weights, window, &output)?;
93            current = output;
94        }
95        Ok(tvec!(current.into_tensor().into_tvalue()))
96    }
97}
98
99impl TypedOp for GpuResize {
100    fn output_facts(&self, inputs: &[&TypedFact]) -> TractResult<TVec<TypedFact>> {
101        crate::utils::facts_to_device_facts(inputs, |facts| {
102            ensure!(facts.len() == 1);
103            let shape: TVec<TDim> = self.output_shape.iter().map(|d| d.to_dim()).collect();
104            Ok(tvec!(facts[0].datum_type.fact(&shape)))
105        })
106        .with_context(|| format!("Error while computing facts for {:?}", self.name()))
107    }
108    as_op!();
109}
tract_gpu/ops/resize.rs

tract_gpu/ops/
resize.rs
