Skill Claude CodeCodex
Build, run, and debug the one-level rmsnormbinary kernel (R-split RMSNorm) with SuperNPUBench + runop.py + gfrun/gfsim precision checks. Use when editing rmsnormbinary.hpp, rmsnormbinary tests, workspace/GetCacheId reduce, TADD cross-tile sum, or verifying [1,8192] fp16 binary RMSNorm. Shape dims are A (outer) and R…
Skill Claude CodeCodex
Explains binary-tree accumulation (二分累加) cache addressing via GetCacheId / ctz(idx+1), and how buffer[cacheId] stores partial sums of length 2^k. Use when writing or debugging AscendC/CANN reduce kernels, GetCacheId, ScalarGetCountOfValue, binary accumulation, or multi-level sum buffers.
Skill Claude CodeCodex
Use when building the Linx compiler toolchain (linx-toolchain-build), compiling SuperNPUBench operators to ELF binaries and disassembly, building the SuperScalarModel simulator (gfrun/gfsim), or running simulation tests on compiled ELF files. Trigger keywords include "linx-toolchain-build", "SuperNPUBench"…