TorchLean API

FloatLib.Floats.Formats.BinaryInterchange.Configured.Plan.FirstOrder

Direct calls to selected structural kernels #

Word and pair storage can call a structural kernel without projecting a certificate closure. The guard below uses the same eligibility and cost comparison as the complete portfolio. Its selection proofs unfold that portfolio; they do not identify different algorithms through their common numerical specification.

@[inline]

Storage plans whose structural candidate can be called without a carrier-specific table or limb closure. This is a storage eligibility test, not a selection decision.

Instances For
    @[inline]

    Whether the structural candidate wins its comparison with the generic kernel.

    Instances For
      @[noinline]

      Whether a direct structural call executes the winner of the complete configured portfolio.

      Keep this decision out of the arithmetic body: closed format/plan/policy applications are cached by the compiler, while the selected structural branch remains available for specialization.

      Instances For
        @[always_inline]
        def FloatLib.Floats.Formats.BinaryInterchange.Configured.Plan.structuralSqrt {format : FloatFormat} {plan : StoragePlan format} {code : Type} [ExecFloat.ModelCodec plan (Model format) code] (value : ExecFloat (Family format code plan)) :
        ExecFloat (Family format code plan)

        The square-root function stored in the descriptor's structural candidate.

        Instances For
          @[always_inline]
          def FloatLib.Floats.Formats.BinaryInterchange.Configured.Plan.structuralFma {format : FloatFormat} {plan : StoragePlan format} {code : Type} [ExecFloat.ModelCodec plan (Model format) code] (left right third : ExecFloat (Family format code plan)) :
          ExecFloat (Family format code plan)

          The fused-multiply-add function stored in the descriptor's structural candidate.

          Instances For