GPU Computing · Cluster
The SIMT Model
Warps, divergence, and occupancy.
- 01 Warp Granularity: Why Thirty-Two Threads Share One Program Counter How a block of threads is carved into warps, why the hardware tracks one instruction stream per group of thirty-two lanes, and what that granularity costs at the edges.
- 02 Divergence and the Reconvergence Stack How a warp executes both sides of a branch when its lanes disagree, what the per-warp mask stack records, and why reconvergence is not always where the source suggests.
- 03 What Occupancy Actually Counts Why occupancy is a ratio of resident warps to a hardware maximum rather than a utilization measure, and how registers, shared memory, and block limits each cap it.
- 04 Predication and Branch Selection When a compiler converts a conditional into predicated instructions or a select, and the cost comparison that decides whether branching or straight-line code is cheaper for a warp.
- 05 Active-Mask Accounting and Issue Slot Efficiency How to compute the fraction of lane capacity a warp actually delivers from its active masks, and why instruction issue rather than lane count is the resource that runs out.
- 06 Register Allocation and the Occupancy Ceiling How per-thread register counts are rounded into per-warp reservations, why a few extra registers can cost half the resident warps, and what forcing a lower count actually costs.
- 07 Latency Hiding and the Arithmetic of Warp Counts Deriving how many resident warps are needed to cover a memory stall from instruction count and latency, and where the Little's law model overstates what interleaving can deliver.
- 08 Per-Lane Program Counters and Independent Thread Scheduling What changes when each lane carries its own program counter and call stack, how divergence stops being a strict two-phase serialization, and why implicit reconvergence disappears.
- 09 Warp Specialization and Producer-Consumer Warps How assigning different code to different warps of a block removes divergence at the source, what the handoff between producer and consumer warps costs, and when it loses.
- 10 Divergence Amplification in Nested Control Flow Why nesting multiplies serialized path count rather than adding to it, how the mask stack depth bounds what the hardware can track, and which rearrangements collapse the product.