EBGeometry::TapeFeature Namespace Reference

Optional tape features, listed among a Tape's template arguments: Tape<T, TapeFeature::ExactOffsets>. More...

Classes
struct	ExactOffsets
	Ray queries trace an Offset or Annular of a primitive with closed-form level sets (or of an affine map of one) exactly, from its level sets (see RayDetail::LevelForm); without it they march it. The level passes through Scale, Translate and Rotate to the primitive, divided by the map's value scale, so an annulus of a scaled mesh is still traced exactly. More...
struct	Tag
	Base of every feature: a Tape template argument derived from it is a feature, not a primitive type. More...

Detailed Description

Optional tape features, listed among a Tape's template arguments: Tape<T, TapeFeature::ExactOffsets>.

A feature is code a tape compiles only when its type asks for it. A plain Tape<T> compiles none, so its kernels are not slowed by code it never runs. Without a feature, the tape still computes the same object; the operation the feature serves takes the general path (for a ray query, marching).

Why features exist: a tape's evaluator and ray code are compiled once per tape type, for any expression a tape of that type could hold, and on a device ptxas allocates registers and stack over all the code a kernel can reach, run or not. Reaching the exact-offset code from the ray code, behind a flag no plain tape sets, slowed plain float firstHit kernels by 10-16% (sm_86). So code that only some expressions need, such as a new exact ray form for an operation, is a feature, and a new feature follows these rules:

• The tape stays correct without it. Without the feature the compiler emits what it emitted before the feature existed (for a ray form, a RayOp::Marched clause), so a tape that does not list it computes the same object by the general path. A feature changes speed or exactness, never which object is computed.
• The compiler decides. TapeDetail::Compiler receives the features and emits the feature's clauses, flags or parameters only when it is listed; clause flags a feature sets are never set otherwise.
• The evaluator compiles the feature's code only under if constexpr (hasFeature<F>), at every place it is reached from (for ExactOffsets: rayLeaf(), rayMarchedWithin(), traceRay() and rayLeafHit()). A run-time test of a flag alone is not enough: the code behind it is still in every kernel.
• A plain tape compiles exactly the code it did before. Check that Tape<T>'s kernels are unchanged: the ptxas register, stack and spill figures (or the SASS) of the float tape kernels (signedDistance() once and sixteen times, and firstHit()) and the host evaluator's out-of-line function sizes (nm -C -S on Performance/Tape) match the base's, and the performance comparison shows no change in the tape cases.
• Tests and probes use both. The feature's tests use a tape that lists it, and at least one test checks that a tape without it gives the same answer by the general path. Probe cases that measure it use a tape that lists it; the existing tape cases keep plain tapes.
• Documented here, as a tag with a description, and in the user documentation's list of tape features.