3 ms·
I get what you mean and agree, and have seen almost traumatized rants against ffast-math from the very same people. After digging, I think this is the kind of
by FooBarBizBazz 2y ago
I get what you mean and agree, and have seen almost traumatized rants against ffast-math from the very same people.
After digging, I think this is the kind of thing I'm referring to:
https://people.eecs.berkeley.edu/~wkahan/JAVAhurt.pdf https://people.eecs.berkeley.edu/~wkahan/JAVAhurt.pdf
https://news.ycombinator.com/item?id=37028310 https://news.ycombinator.com/item?id=37028310
I've seen other course notes, I think also from Kahan, extolling 80-bit hardware.
Personally I am starting to think that, if I'm really thinking about precision, I had maybe better just use fixed point, but this again is just a "lean" that could prove wrong over time. Somehow we use floats everywhere and it seems to work pretty well, almost unreasonably so.
- stephencanon 2y agoYeah. Kahan was involved in the design of the 8087, so he’s always wanted to _have_ extended precision available. What he (and I, and most other numerical analysts) are opposed to is the fact that (a) language bindings historically had no mechanism to force rounding to float/double when necessary, and (b) compilers commonly spilled x87 intermediate results to the stack as doubles, leading to intermediate rounding that was extremely sensitive to optimization and subroutine calls, making debugging numerical issues harder than it should be. Modern floating-point is much more reproducible than fixed-point, FWIW, since it has an actual standard that’s widely adopted, and these excess-precision issues do not apply to SSE or ARM FPUs.