Struct CudaConv 

pub struct CudaConv { /* private fields */ }

Trait Implementations

impl Clone for CudaConv

fn clone(&self) -> Self

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for CudaConv

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Eq for CudaConv

impl EvalOp for CudaConv

fn eval_out_of_plan( &self, _inputs: TVec<TValue>, ) -> TractResult<Option<TVec<TValue>>>

Evaluate with no plan around the node – const folding, shape inference, tests – or None when there is no answer without one, because the op reads the context or keeps state between turns. Required, with no default: an op answering Some has produced the value from inputs alone, by construction, so the claim and the act cannot disagree. Write it with op_out_of_plan!() or not_out_of_plan!().

fn eval( &self, ctx: &EvalContext<'_>, inputs: TVec<TValue>, ) -> TractResult<TVec<TValue>>

Evaluate the op. ctx says where and when: the turn’s symbols, the shared resources a handler installed, and (session, node_id) so an op can key whatever scratch it manages for itself. Ops carrying state that must survive from one turn to the next build it in EvalOp::state instead, and evaluate through OpState::eval.

fn drop_session(&self, session: SessionId, node_id: usize)

Release whatever the op manages for session, called for every node as a state is dropped. Ops keeping scratch keyed by (session, node_id) must implement it, or that scratch outlives the session that made it.

fn forwards_input(&self) -> Option<usize>

The input this op hands straight back as its output when it has nothing to do, if any: what reads the output then reads the producer’s memory, so an allocator pooling node outputs has to keep that region alive as long as this op’s own output. A wrapper delegates to what it wraps.

fn state( &self, ctx: &EvalContext<'_>, ) -> Result<Option<Box<dyn OpState>>, Error>

Build this node’s inter-turn state, or None when the op keeps nothing between turns. This is what decides whether the plan holds an OpState for the node; there is no separate predicate.

impl Op for CudaConv

fn name(&self) -> StaticName

fn info(&self) -> TractResult<Vec<String>>

Short (one-line) strings giving hints on internal implementation or important configuration details to be displayed in dumps.

fn as_typed(&self) -> Option<&dyn TypedOp>

fn validation(&self) -> Validation

The kind of accuracy check that should be performed on operation when testing them.

impl PartialEq for CudaConv

fn eq(&self, other: &Self) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for CudaConv

impl TypedOp for CudaConv

fn as_op(&self) -> &dyn Op

Reinterpret the TypedOp as an Op.

fn as_op_mut(&mut self) -> &mut dyn Op

Reinterpret the TypedOp as an Op, mutably.

fn output_facts(&self, inputs: &[&TypedFact]) -> TractResult<TVec<TypedFact>>

Deduce output facts from input facts.

fn axes_mapping( &self, inputs: &[&TypedFact], outputs: &[&TypedFact], ) -> Result<AxesMapping, Error>

fn fuse( &self, _model: &Graph<TypedFact, Box<dyn TypedOp>>, _node: &Node<TypedFact, Box<dyn TypedOp>>, ) -> Result<Option<ModelPatch<TypedFact, Box<dyn TypedOp>>>, Error>

Fuse op after codegen to deal with local optimisations.

fn declutter_with_session( &self, session: &mut OptimizerSession<'_>, model: &Graph<TypedFact, Box<dyn TypedOp>>, node: &Node<TypedFact, Box<dyn TypedOp>>, ) -> Result<Option<ModelPatch<TypedFact, Box<dyn TypedOp>>>, Error>

Declutter the op to the tract_core operator set as much as possible.

fn declutter( &self, model: &Graph<TypedFact, Box<dyn TypedOp>>, node: &Node<TypedFact, Box<dyn TypedOp>>, ) -> Result<Option<ModelPatch<TypedFact, Box<dyn TypedOp>>>, Error>

Declutter the op to the tract_core operator set as much as possible.

fn cost( &self, _inputs: &[&TypedFact], ) -> Result<SmallVec<[(Cost, TDim); 4]>, Error>

Computes a cost hint of the operation.

fn input_roi( &self, model: &Graph<TypedFact, Box<dyn TypedOp>>, node: &Node<TypedFact, Box<dyn TypedOp>>, ) -> Result<Option<SmallVec<[Option<TDim>; 4]>>, Error>

Derive ROI (region of interest) expressions for this node’s inputs. Called by the PropagateRoi pass. Default returns None (no propagation). Override to introduce ROIs or bubble them through.

fn suggested_axis_changes( &self, ) -> Result<SmallVec<[(InOut, AxisOp); 4]>, Error>

fn change_axes( &self, model: &Graph<TypedFact, Box<dyn TypedOp>>, node: &Node<TypedFact, Box<dyn TypedOp>>, io: InOut, change: &AxisOp, ) -> Result<Option<AxisChangeConsequence>, Error>

fn slice( &self, patch: &mut ModelPatch<TypedFact, Box<dyn TypedOp>>, model: &Graph<TypedFact, Box<dyn TypedOp>>, node: &Node<TypedFact, Box<dyn TypedOp>>, prefix: &str, inputs: &[OutletId], output_axis: usize, start: &TDim, end: &TDim, ) -> Result<Option<SmallVec<[OutletId; 4]>>, Error>

fn quantize( &self, model: &Graph<TypedFact, Box<dyn TypedOp>>, node: &Node<TypedFact, Box<dyn TypedOp>>, dt: DatumType, scale: f32, zero_point: i32, ) -> Result<Option<Box<dyn TypedOp>>, Error>

Transforms the op in an equivalent one, operating on dt (i8 or u8).

fn set_symbols( &self, source: &Graph<TypedFact, Box<dyn TypedOp>>, node: &Node<TypedFact, Box<dyn TypedOp>>, target: &mut Graph<TypedFact, Box<dyn TypedOp>>, mapping: &HashMap<OutletId, OutletId>, subs: &HashMap<Symbol, TDim>, ) -> Result<SmallVec<[OutletId; 4]>, Error>

Transform the op by substituting one or more symbols with TDim expressions (a concrete integer is TDim::Val(v); an expression can be any other TDim, including symbolic ones).

fn codegen( &self, model: &Graph<TypedFact, Box<dyn TypedOp>>, node: &Node<TypedFact, Box<dyn TypedOp>>, ) -> Result<Option<ModelPatch<TypedFact, Box<dyn TypedOp>>>, Error>

Translate the op into the most efficient form possible for execution.

fn nested_model_multipliers( &self, inputs: &[&TypedFact], ) -> Vec<(Cow<'static, str>, TDim)>

Nested model multipliers, with label (for profiling).