{ "name": "softmax_fp32", "op_type": "softmax", "description": "1-D softmax, float32 in and out: y = exp(x - max(x)) / sum(exp(x - max(x))) over a vector of length N (e.g. one attention row or one vocabulary of logits). N varies per workload. Judged by SQNR against the baseline: outputs are about 1/N, below the default absolute tolerance. Baseline: KleidiAI kai_softmax_f32_f32_1d_sve_fexpa (SVE FEXPA exponential, any vector length).", "tags": [ "status:active", "baseline-solution:kleidiai", "correctness:sqnr" ], "axes": { "N": { "type": "var" } }, "inputs": { "x": { "shape": [ "N" ], "dtype": "float32" } }, "outputs": { "y": { "shape": [ "N" ], "dtype": "float32" } }, "reference": "import numpy as np\n\n\ndef run(x):\n x = x.astype(np.float64)\n e = np.exp(x - x.max())\n return (e / e.sum()).astype(np.float32)\n" }