tract-gpu 0.23.6

use crate::tensor::{DeviceTensor, DeviceTensorExt};
use derive_new::new;
use tract_core::internal::*;

pub type DispatchApplyRopeFn =
    fn(&DeviceTensor, &DeviceTensor, &DeviceTensor, &DeviceTensor) -> TractResult<()>;

#[derive(Clone, new)]
pub struct GpuApplyRope {
    pub backend_name: &'static str,
    pub dispatch: DispatchApplyRopeFn,
}

impl std::fmt::Debug for GpuApplyRope {
    fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
        write!(f, "{}ApplyRope", self.backend_name)
    }
}

impl PartialEq for GpuApplyRope {
    fn eq(&self, other: &Self) -> bool {
        self.backend_name == other.backend_name
    }
}
impl Eq for GpuApplyRope {}

impl std::hash::Hash for GpuApplyRope {
    fn hash<H: std::hash::Hasher>(&self, state: &mut H) {
        self.backend_name.hash(state);
    }
}

impl Op for GpuApplyRope {
    fn name(&self) -> StaticName {
        format!("{}ApplyRope", self.backend_name).into()
    }
    op_as_typed_op!();
}

impl EvalOp for GpuApplyRope {
    op_out_of_plan!();

    fn eval(&self, ctx: &EvalContext, inputs: TVec<TValue>) -> TractResult<TVec<TValue>> {
        let (input_val, cos_val, sin_val) = args_3!(inputs);
        let input = input_val.to_device_tensor()?;
        let cos = cos_val.to_device_tensor()?;
        let sin = sin_val.to_device_tensor()?;
        let output =
            crate::turn_handler::make_tensor_for_node(ctx, input.datum_type(), input.shape())?;
        (self.dispatch)(input, cos, sin, &output)?;
        Ok(tvec!(output.into_tensor().into_tvalue()))
    }
}

impl TypedOp for GpuApplyRope {
    fn output_facts(&self, inputs: &[&TypedFact]) -> TractResult<TVec<TypedFact>> {
        crate::utils::facts_to_device_facts(inputs, |facts| {
            let dt = facts[0].datum_type;
            let fact = dt.fact(facts[0].shape.clone());
            Ok(tvec!(fact))
        })
        .with_context(|| format!("Error while computing facts for {:?}", self.name()))
    }
    as_op!();
}