nx_native_softmax_bench.nx

nx_native_softmax_bench.nx -- LLM-relevant softmax kernel. One of the three kernels that dominate transformer-forward time: matmul (already measured in nx_native_matmul_bench) softmax (this bench) RMSNorm (nx_native_rmsnorm_bench) Softmax over a vector: max-subtract for numerical stability, then exp + sum + divide. Substrate uses Q10 fixed-point so "exp" is a polynomial / table approximation; here we measure the dominant-cost integer arithmetic (max + sum + division) skipping the exp step (still LLM-relevant: attention weights pass through softmax once per token per head). Workload: softmax over 2048-element vectors, 1000 iterations.