[2026秋季][Task10] jinger861-git

[jinger861-git]

Summary

• Optimize Task10 Grid for Cambricon MLU.
• Fuse grid-index computation, clamping, multi-dimensional Horner ID construction, and inverse unique mapping into one Triton kernel.
• Statically unroll the dimension loop and vectorize over points.
• Use a small-N on-chip O(N^2) inverse-index mapping for N <= 32, with a PyTorch fallback for larger inputs.

Accuracy

PASS with the official KS benchmark.

• atol: 1e-2
• rtol: 1e-2
• Additional regression covered optional start/end, negative coordinates, float coordinates, and 3D coordinates; all inverse indices matched exactly.

Performance

• v0: 0.080147 ms
• v1: 0.004001 ms
• speedup: 20.032x

Changed file

• dlblas/kernels/ks_competition/QY2026_Autumn_jinger861-git_Task10Grid_MLU.py

[CLAassistant]

All committers have signed the CLA.