Deep dive into CuTeDSL swizzle layouts and M parameter logic
snowclipsed · x · 2026-09-02
The post discusses the reasoning behind composing a <3,3,3> swizzle with an (8,64) block in CuTeDSL. It references a blog explaining the calculation of B and S for a layout with 8 rows and 64 bf16 elements per row, aligned to 16 bytes. Addressing why M is 3 instead of 4 in the Ampere example, it clarifies that CuTeDSL reasons about M in layout offset units, not directly on byte addresses, meaning pointers are offset by elements of the layout's codomain.
More from Infra
- Infra optimization potential when controlling open model inference — _ScottCondron · 2026-09-02
- 69% chance any US state bans data centers this year — Polymarket · 2026-09-02
- Agent retried for 16 hours silently? Author introspects durable runtime with Postgres persistence — slateraligator · 2026-09-02
- Anthropic Cuts Cache Read Prices by 75%, Closing Gap with DeepSeek — Teknium · 2026-09-02
- Trading sector too small to justify massive AI capex, analyst argues — iaindunning · 2026-09-02
- Speed Wars: GPT 5.6 Hits 750 Tokens/s on Cerebras — DeepLearningAI · 2026-09-02