Each warp is tagged with a stage ID, and the thread-block launch metadata is extended to include the number of pipeline stages, per-stage register needs, and named queues between stages. This matters because once hardware knows which stage a warp belongs to, it can make better decisions about where to run the warp, how many registers to allocate, and how to prioritize it.
register file queues (RFQs)
WASP-TMA
Existing TMA-like support mainly accelerates coarse tile transfers between global memory and shared memory. WASP extends this idea so hardware can also help with fine-grained global-to-RFQ streaming and gather operations. That reduces the amount of address-generation and control code that producer warps need to execute, which in turn frees issue bandwidth and registers for useful work.
On a conventional GPU, warps are mapped to processing blocks in a round-robin way. WASP instead groups warps that belong to the same pipeline slice, reducing cases where one processing block gets mostly memory-stage warps while another gets mostly compute-stage warps.
Since earlier stages in the pipeline are often memory stages with long latency, WASP gives the scheduler awareness of stage ID and queue state. The scheduler can prioritize earlier producer stages or consumer stages whose queues are ready, improving overlap and keeping the pipeline full.