Skip to content

Dispatch and Kernels

BitCal documents dispatch as a support boundary, not as a personality trait. Readers need to know where implementation freedom begins, because that line defines which changes are architectural and which are merely backend work.

Dispatch boundary

Public contract versus implementation freedom
Public contractinclude seam · owner/view roles · algorithm semanticsdispatch boundarydispatch policyfeature detection, routing, fallback choicekernel familiesAVX2 primary path, scalar portability floormicro-tuningintrinsic selection, unrolling, alignment strategy
The contract line keeps user-visible roles and semantics above dispatch policy and kernel replacement.

Kernel families

Kernel familyCurrent documentation stanceWhy it is separated
Scalar floorRetained portability baselineGives the project a correctness floor even when no wide path is active.
AVX2 pathPrimary evidence and optimization targetCurrent benchmark posture is concentrated here.
Future x86 variantsMention only as design directionThey should not be overclaimed before retained evidence exists.

Support boundary

Dispatch and kernels must respect the documented support posture:

  • Linux / Windows x86-64 are the primary optimization and validation targets.
  • Secondary targets may remain buildable, but must not inherit stronger claims than the retained evidence supports.
  • Public documentation should mention backend detail only when it changes user-visible semantics, benchmark scope, or support boundaries.

What readers should carry forward

The dispatch layer is allowed to change aggressively as long as the public role model, algorithm semantics, and include seam remain coherent. That is why the next stop after this page is the Performance section: benchmark claims belong below this boundary, not above it.

Whitepaper-first technical documentation for BitCal vNext.