search

VA00 / Generalized-Fermi-Dirac-integrals

Provide reliable, accurate down to several ULP's and fast numerical framework to compute generalized Fermi-Dirac integrals aiming at full floating-point parameter coverage, generality, precision control and speed.
MIT License
0 stars 0 forks source link

Ffermi_derivatives_m_n_quad precision drop for eta<-64 #4

Open VA00 opened 2 years ago

VA00 commented 2 years ago

For eta=-64 and smaller, Ffermi_derivatives_m_n_quad produces huge error sqrt(DBL_EPSILON) and is very slow. Same behavior for gcc and icc. Changing PRECISION_GOAL_QUAD 128*FLT128_EPSILON to DBL_EPSILON did not help. Strangely, Ffermi_derivatives_m_n, do not show this error, possible _float128 coding bug.

VA00 commented 2 years ago

This is apparently caused by logistic_sigmoid_quad which have used tanh instead of exp.