GPU Design · All levels

Warp Scheduling & Control Flow: Tricky Q&A

Senior interview and review questions for Warp Scheduling & Control Flow.

Section Q&A bank

Use these drills after completing all topics in Warp Scheduling & Control Flow. Answer with workload context, mechanism proof, artifact, owner, and release decision.

What indicates scheduler policy mismatch rather than insufficient occupancy?

diagram
[INT][GPU][WARP-SCHEDULING]

Q: What indicates scheduler policy mismatch rather than insufficient occupancy?

A:
Large ready-warp pools with persistent unfairness and tail stalls suggest arbitration strategy, not residency depth, is the bottleneck.

FOLLOW-UP TRAP: Adding more active warps without fairness diagnostics.

How do barriers become dominant even in high-throughput kernels?

diagram
[INT][GPU][WARP-SCHEDULING]

Q: How do barriers become dominant even in high-throughput kernels?

A:
Warp progress imbalance makes early-arriving warps idle at synchronization points, shrinking effective issue opportunity.

FOLLOW-UP TRAP: Assuming barriers are negligible in compute-heavy kernels.

When is predication better than branching in SIMT kernels?

diagram
[INT][GPU][WARP-SCHEDULING]

Q: When is predication better than branching in SIMT kernels?

A:
When branch paths are short and divergence overhead exceeds extra instruction work from masked predicated execution.

FOLLOW-UP TRAP: Applying predication blindly to long heavy code paths.

How should reconvergence behavior be validated in reviews?

diagram
[INT][GPU][WARP-SCHEDULING]

Q: How should reconvergence behavior be validated in reviews?

A:
Use lane-mask timelines and reconvergence delay counters tied to concrete branch regions in source or SASS.

FOLLOW-UP TRAP: Discussing divergence conceptually without trace-level evidence.

Q&A drill guide

diagram
WORKLOAD -> BOTTLENECK -> METRIC -> CHANGE -> REGRESSION

Sketch while answering

diagram
WARP SCHEDULING LOOP

ready warp? -> issue -> dependency wait -> reconverge -> issue

Key takeaways

  • Always pair micro-kernel metrics with end-to-end workload impact.

  • Lock toolchain, driver, and launch metadata before comparing performance results.

Common pitfalls

  • Optimizing occupancy without checking memory-system saturation.

  • Comparing profiler captures from different driver or compiler builds.

  • Declaring wins without reproducible accuracy and performance gates.