[Feature Request] [Pass] DMA-aware PipelinePlanning/InjectSoftwarePipeline #37
Description
Required prerequisites
- I have searched the Issue Tracker that this hasn't already been reported. (comment there if it has.)
Motivation
TileLang's current pipeline system (PipelinePlanning and InjectSoftwarePipeline) is designed around GPU thread/warp models. For the Sunmmio ZPU with DMA but no threads/warps, we need to adapt or create new passes to handle DMA-compute overlap.
We may need to:
Modify or Replace PipelinePlanning Pass:
- Replace thread/warp concepts with DMA channel scheduling
- Model DMA transfer latency and compute overlap
Modify or Replace InjectSoftwarePipeline:
- Generate DMA start/wait operations instead of async scopes
- Adapt buffer multi-versioning for DMA patterns
- Remove GPU-specific thread synchronization
Solution
No response
Alternatives
No response
Additional context
No response