Summary

Refactors ThreadwiseTensorSliceTransfer_v3r1 to use SpaceFillingCurve for index computation, replacing O(N^2) nested ComputeForwardSweep/ComputeDataIndex/ComputeMoveOnDim calls with O(1) precomputed lookup via SFC::GetIndex().

This reuses the existing SpaceFillingCurve infrastructure already used by v6r1, v7r2, etc. Changes are confined to a single file.

What changed

• RunRead(): replaced static_ford serpentine loop with static_for over SrcSFC::GetIndex()/GetForwardStep()
• RunWrite(): same pattern with DstSFC
• TransferDataFromSrcThreadScratchToDstThreadScratch(): OOB check path uses SrcSFC::GetIndex()
• Coordinate reset: uses ComputeSFCCoordinateResetStep instead of ComputeCoordinateResetStep

Results (example_grouped_conv_fwd_xdl_fp16, n=10, interleaved, -j1, -ftime-trace)

Negative controls (unrelated TUs — no significant change)

Assembly equivalence

GPU assembly compared with -save-temps=obj on example_grouped_conv_fwd_xdl_fp16 (gfx942). Result: 7656 instructions in both builds, identical instruction mix. Only difference is a register allocation swap (v[160:161] <-> v[162:163] in two buffer_load_dwordx2 / v_cndmask_b32_e64 pairs) — functionally identical, no runtime impact.

Methodology

• 10 interleaved runs (baseline1, new1, baseline2, new2, ...) on same node to eliminate ordering/warmup bias
• Wilcoxon signed-rank test (paired) and Mann-Whitney U test (unpaired)
• Built with patched clang (LLVM 22) on ctr2-alola-compile-12, -j1 for accurate per-TU timing
• Raw data: Slurm job 277012, assembly comparison: Slurm job 277013

Test plan

• Compile-time benchmark with statistical significance (p < 0.01)
• Negative controls confirm stable node
• GPU assembly equivalence verified
• Full CI

Tracking issue: #4229

Generated-by: Claude Code (claude-opus-4-6)