use crate::session_handler::make_tensor_for_node;
use crate::tensor::{DeviceTensor, DeviceTensorExt};
use tract_core::internal::*;
pub type DispatchCausalConv1dUpdateFn = fn(
&DeviceTensor,
&DeviceTensor,
&DeviceTensor,
&DeviceTensor,
&DeviceTensor,
) -> TractResult<()>;
#[derive(Clone, Debug)]
pub struct GpuCausalConv1dUpdate {
pub backend_name: &'static str,
pub dispatch: DispatchCausalConv1dUpdateFn,
}
impl PartialEq for GpuCausalConv1dUpdate {
fn eq(&self, other: &Self) -> bool {
self.backend_name == other.backend_name
}
}
impl Eq for GpuCausalConv1dUpdate {}
impl std::hash::Hash for GpuCausalConv1dUpdate {
fn hash<H: std::hash::Hasher>(&self, state: &mut H) {
self.backend_name.hash(state);
}
}
impl Op for GpuCausalConv1dUpdate {
fn name(&self) -> StaticName {
format!("{}CausalConv1dUpdate", self.backend_name).into()
}
op_as_typed_op!();
}
impl EvalOp for GpuCausalConv1dUpdate {
fn is_stateless(&self) -> bool {
true
}
fn eval_with_session(
&self,
node_id: usize,
session: &TurnState,
inputs: TVec<TValue>,
) -> TractResult<TVec<TValue>> {
let (input, weight, state) = args_3!(inputs);
let input = input.to_device_tensor()?;
let weight = weight.to_device_tensor()?;
let state = state.to_device_tensor()?;
let output = make_tensor_for_node(session, node_id, DatumType::F16, input.shape())?;
let final_state = DeviceTensor::uninitialized_dt(DatumType::F16, state.shape())?;
(self.dispatch)(input, weight, state, &output, &final_state)?;
Ok(tvec![output.into_tensor().into_tvalue(), final_state.into_tensor().into_tvalue()])
}
}
impl TypedOp for GpuCausalConv1dUpdate {
fn output_facts(&self, inputs: &[&TypedFact]) -> TractResult<TVec<TypedFact>> {
crate::utils::facts_to_device_facts(inputs, |facts| {
ensure!(facts.len() == 3);
ensure!(facts.iter().all(|fact| fact.datum_type == DatumType::F16));
Ok(tvec![facts[0].without_value(), facts[2].without_value()])
})
.with_context(|| format!("invalid facts for {}", self.name()))
}
as_op!();
}