SIMT divergence: count active lanes without assuming lockstep

Imagine eight lanes taking a short branch and 24 taking a longer one. In a simple serialized masked-path model, both nonempty paths execute, each with its own active subset. The inactive lanes do not contribute useful lane work on that path. If all lanes take the same branch, the other path need not execute under this model. Divergence depends on how data is grouped into warps, not merely on the whole dataset having equal counts of two outcomes.

A compiler may predicate small regions or transform control flow, and hardware scheduling is more involved than the toy accounting. Do not label every conditional a measured slowdown or infer an exact penalty from source code alone. Grouping similar work can reduce long divergent regions, but preprocessing, reordered memory access, and load imbalance can offset the benefit. First preserve semantics, inspect generated behavior where useful, and measure the actual kernel with representative input distributions.

Starting with Volta, independent thread scheduling maintains finer-grained execution state, so old code that relied on implicit warp lockstep can be wrong. A shared-memory write by one lane is not automatically safe to read by another because both belong to the same warp. Use the documented synchronization primitive and participation mask for the actual communicating group. The relevant *_sync intrinsics define participation requirements, but not all of them imply a memory barrier; a warp shuffle and a shared-memory handoff have different contracts.

A block barrier must be reached as required by all non-exited block threads, and a warp collective must use a valid matching participating set. A mask captured inside one branch is not automatically proof of the full intended logical group. Reconvergence of control flow should not be treated as a universal memory-ordering primitive. Compute Sanitizer’s synccheck and racecheck can help find some violations, but their scopes differ and a clean run does not prove every possible schedule is race-free.

A branch outcome can correlate strongly with neighboring data, so independent Bernoulli outcomes are not a default property of a warp. A dataset sorted by category can produce mostly uniform warps even if half the total data belongs to each category. Conversely, alternating categories can split many warps. Build adversarial data layouts in tests, including partial final warps, and separate control-flow efficiency from memory coalescing. A layout change can improve one while worsening the other.

The sampling lab permits a deliberately hypothetical branch-label model. Reset, set success probability p=0.5 and sample size n=32, predict mean fraction 0.5 and variance 0.25/32, then Step. The bars are the exact binomial distribution, and playback highlights outcomes rather than executing random branches. It is not a warp scheduler or a divergence-timing simulator. Its independence assumption lets you reason about possible branch counts, while the actual instruction-path cost still needs a separate execution model.