* Yuval Lifshitz <yuva...@gmail.com> [180430 02:02]: > It seems like go and C++ are doing something different regarding floating > point arithmetic, but I cannot say one is better than the other. It is just > the that C++ consistently overshoots (hence truncation work), and go > consistently undershoots (this is why rounding is needed). For example, in > C++: > > F(59) = 956722026041.002 > > And in go, it is: > > F(59) = 956722026040.999878
/* aside - this doesn't seem relevant to this problem except perhaps * the implementation of math.Pow, but... The 8087 and successor FPUs (e.g. amd64 class processors) have an 80-bit extended precision format which is used for all calculations. More than a couple decades ago, when I was working on implementation of floating point in an APL interpreter (written in C, C++, asm), the IEEE 754 standard did not specify how compilers should handle precision of intermediate results, and the compiler we were using would only convert from 80-bit to 64-bit when it needed to store a result back into memory. This could easily give different results depending on how the floating point expression was broken up. E.g. a = (b + c) - d could give a different result than a = b + c; a = a - d. The standard has, since then, attempted to address this, but I have not studied this at all, so I don't know what is expected of C or Go. */ Back to your specific problem: Go: psi = -0.61803398874989490 C++: psi = -0.61803398874989479 Go: math.Pow(math.Phi, float64(33)) = 7.88119600000012666e+06 C++: pow(phi, 33) = 7.88119600000013597e+06 And using hand-typed constants: Go: math.Pow(1.61803398874989490, 33.0) = 7.88119600000012666e+06 C++: pow(1.61803398874989490, 33.0) = 7.88119600000013597e+06 So indeed there is a difference between math.Pow in Go and pow in C++. I would say to file an issue. I have no idea which answer is closer to correct, but using math/big and multiplying 1.61803398874989490 by itself 33 times with a result precision of 80, I get 7.88119600000013562e+06. Note: in Go, math.Phi is an untyped constant with precision greater than float64. Adding var phi float64 = math.Phi and replacing math.Phi everywhere else with phi did not change anything except the value of psi (which then matched the C++ value). The C++ version of psi is definitely farther from the "true" value, but as Michael said, the error in pow(psi, n) is swamped by the value of pow(phi, n). ...Marvin -- You received this message because you are subscribed to the Google Groups "golang-nuts" group. To unsubscribe from this group and stop receiving emails from it, send an email to golang-nuts+unsubscr...@googlegroups.com. For more options, visit https://groups.google.com/d/optout.
