Promote data within a complete resource budget

A global buffer can be accessed across a wider execution scope, while workgroup storage is shared within a defined block or workgroup. Private storage belongs to a smaller scope, but an IR memory-space label does not universally guarantee a hardware register for every value. Register allocation, spills, address-taking, and backend lowering affect the physical implementation. Use the actual target and generated resource report before claiming that all promoted values stay on chip or avoid memory traffic.

Copies into shared storage need a correct producer-consumer schedule. Threads must not read before the relevant copy completes, and storage must not be overwritten while earlier readers still use it. Barriers, asynchronous-copy waits, and memory visibility rules depend on the target and selected operations. Every participant required by a barrier must reach it under the allowed control flow. A pipeline that overlaps copies and arithmetic must preserve those stage lifetimes rather than merely adding a second buffer.

Larger tiles and more buffering increase live storage per workgroup. A capacity limit then restricts how many workgroups can reside together, even if each individual launch is valid. Threads, registers, scheduling granularity, and architectural limits can restrict concurrency further. Dividing total shared capacity by per-block bytes supplies only one upper bound. It is not a prediction of achieved occupancy or an argument that maximizing that number minimizes kernel time.

A promotion can also move the bottleneck. Reduced global traffic may expose arithmetic, synchronization, bank conflicts, or instruction overhead. An output already reused in cache can save less main-memory traffic than its logical size suggests. Compare actual boundary traffic and completion time under the same workload and numerical contract. If promotion adds a layout conversion, count its storage and execution too. Keep scalar and boundary cases in the correctness suite because incomplete tiles often have different synchronization paths.

Inspect shared/workgroup allocation, registers, spills, and synchronization after lowering, not only high-level tensor shapes. Test multiple pipeline depths and tile dimensions under target limits, retaining a correct fallback when a configuration is unsupported. Compare cold setup and steady execution if promotion requires initialization. Use race and memory tools where supported, and differential value checks for awkward tails. A benchmark that never exercises competing workgroups may hide the concurrency penalty of a resource-heavy schedule.

The Roofline lab is a synthetic boundary calculation. Reset and set both I=4, B=80, predict 320 GFLOP/s. Reset and re-set I=8, B=80 to predict 640 under the fixed 1000 GFLOP/s compute roof, then Play or Step. It does not allocate shared memory, synchronize copies, or model occupancy. Interpret the higher intensity only as a hypothetical reduction in bytes at the named boundary with fixed work, and identify the additional target resource evidence needed before expecting that ceiling to be approached.