[dlamd1dai]

Summary

• Stacked on feat(rccl-tests): GIN Anvil-SDMA ReduceScatter (-D 3) ROCm/rocm-systems#12332 (users/dondai/gin-sdma-rs-d3). This draft is on the fork so the diff is only the tier. Retarget it to ROCm/rocm-systems develop after feat(rccl-tests): GIN Anvil-SDMA ReduceScatter (-D 3) ROCm/rocm-systems#12332 merges.
• When NCCL_GIN_ANVIL_RS_CTAS pins the grid below 16, ReduceScatter -D 3 GIN-puts each per-dest slice into the owner's resource window and SM-reduces the local slots. The default 32/48-CTA launch stays on the LSA read-reduce.
• Every CTA waits for the inbound puts. The earlier single-waiter version corrupted output once the grid exceeded nRanks. Scratch requirements are thread_local, and scratch sizing rejects a wrapping multiply.

This is still a draft. The CTA pin is the selection policy, and there is no performance evidence that this path beats LSA. Do not merge it on correctness alone.

JIRA ID: AICOMRCCL-1803
https://amd-hub.atlassian.net/browse/AICOMRCCL-1803

Test plan

• Policy tests cover usesSdmaTier, the pin-below-16 grid, and scratch overflow rejection
• GPU unit test: device usesSdmaTier matches the host
• NCCL_GIN_ANVIL_RS_CTAS=4 datacheck at 1 MiB/rank and 8 MiB/rank, in-place and out-of-place, with a grid above nRanks
• Default unset pin still takes the LSA ladder and registers no scratch
• Compare bus bandwidth of the pin against the LSA ladder before promoting this off draft

Made with Cursor

[dlamd1dai]

Superseded by the ROCm#12747 stack. The low-CTA tier is now split the same way, on top of current develop:

1. Policy helpers and host unit test: feat(rccl-tests): GIN ReduceScatter host policy ROCm/rocm-systems#13302
2. Kernel and GPU usesSdmaTier test: feat(rccl-tests): GIN Anvil-SDMA ReduceScatter (-D 3) ROCm/rocm-systems#13303
3. Pytest pin and GIN CI case count (21 -> 25): test(rccl): GIN CI and pytest for ReduceScatter (-D 3) ROCm/rocm-systems#13304

This draft was one commit on top of 8e871459, two commits behind the closed monolithic branch, so it was rebased rather than retargeted.