1 #ifndef EIGEN_NEON_SPECIALFUNCTIONS_H 2 #define EIGEN_NEON_SPECIALFUNCTIONS_H 3 4 namespace Eigen { 5 namespace internal { 6 7 #if EIGEN_HAS_ARM64_FP16_VECTOR_ARITHMETIC 8 9 #define NEON_HALF_TO_FLOAT_FUNCTIONS(METHOD) \ 10 template <> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE \ 11 Packet8hf METHOD<Packet8hf>(const Packet8hf& x) { \ 12 const Packet4f lo = METHOD<Packet4f>(vcvt_f32_f16(vget_low_f16(x))); \ 13 const Packet4f hi = METHOD<Packet4f>(vcvt_f32_f16(vget_high_f16(x))); \ 14 return vcombine_f16(vcvt_f16_f32(lo), vcvt_f16_f32(hi)); \ 15 } \ 16 \ 17 template <> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE \ 18 Packet4hf METHOD<Packet4hf>(const Packet4hf& x) { \ 19 return vcvt_f16_f32(METHOD<Packet4f>(vcvt_f32_f16(x))); \ 20 } 21 22 NEON_HALF_TO_FLOAT_FUNCTIONS(perf) 23 NEON_HALF_TO_FLOAT_FUNCTIONS(pndtri) 24 25 #undef NEON_HALF_TO_FLOAT_FUNCTIONS 26 #endif 27 28 BF16_PACKET_FUNCTION(Packet4f, Packet4bf, perf) 29 BF16_PACKET_FUNCTION(Packet4f, Packet4bf, pndtri) 30 31 } // namespace internal 32 } // namespace Eigen 33 34 #endif // EIGEN_NEON_SPECIALFUNCTIONS_H 35