The fp32 rational approximation coefficients cannot be represented in
half precision: alpha_9/11/13 underflow to zero and alpha_7 becomes
subnormal, so tanh_ps_f16 saturates at ~1.044 and fast gelu returns
+0.022*|x| for large negative inputs. Transformer models with large
activations (SigLIP2, CLIP ViT) produce input-independent features under
use_fp16_arithmetic.
Refit the rational approximation for fp16 on [0, 4.5], return +/-1
beyond that range, clamp the result to 1, and add large-magnitude test
cases for GELU and TanH.
* runtime cpu dispatch
* force thread one
* disable openmp for coverage
* simplify test layer
* print NCNN_TARGET_ARCH
* less ci build variants
* weight fp16 storage option
* test convdw int8
* apple a12 a13
* ncnn_add_layer ncnn_add_shader cmake macro
nihui