Primary traits template for unified compile-time operation dispatch. More...

Detailed Description

template<OperationType TOp, DeviceType TDeviceType, TensorDataType TPrecision, typename TPolicy = void>
struct Mila::Dnn::Compute::OperationTraits< TOp, TDeviceType, TPrecision, TPolicy >

Primary traits template for unified compile-time operation dispatch.

Each specialization provides either a nested type alias naming the concrete operation class, or (for functor-templated ops such as ElementwiseActivationOp) a nested op_for<Functor> alias, for the given (OperationType, DeviceType, precision, policy) combination.

The primary template is intentionally left undefined. An unsupported tuple therefore names an incomplete type, and the compiler reports a single-line "use of undefined type OperationTraits<Op, Device, Precision, Policy>" naming the exact tuple – a readable diagnostic, not a multi-hundred-line constraint cascade. (A kernel that only supports FP32 must not advertise a BF16 row: the honest failure is a missing specialization here, not a poisoned row whose op fails a float || half kernel constraint deep in dispatch.)

To branch on availability at compile time – e.g. a multi-precision typed test skipping the precisions an op does not implement – use the SFINAE-safe OperationSupported<...> predicate below rather than instantiating this template. The primary must stay undefined for that predicate to work: a static_assert in the primary body would fire during the predicate's own probe, turning the detectable "false" back into a hard error.

Template Parameters

TOp	Operation identifier from the OperationType enum.
TDeviceType	Target device (Cpu, Cuda, ...).
TPrecision	Compute and activation precision.
TPolicy	Optional policy type. Defaults to void for policy-free ops.

The documentation for this struct was generated from the following file:

Mila/Src/Dnn/Compute/Operations/OperationTraits.Template.ixx