Differences in results when using the same function in CTFE and Runtime

[Timon Gehr]

On 8/15/24 18:50, Abdulhaq wrote:
> On Thursday, 15 August 2024 at 16:21:35 UTC, Abdulhaq wrote:
>> On Thursday, 15 August 2024 at 09:13:31 UTC, Carsten Schlote
> 
> To clarify a bit more, I'm not just talking about single isolated 
> computations, I'm talking about e.g. matrix multiplication. Different 
> compilers, even LDC vs DMD for example, could optimise the calculation 
> in a different way, loop unrolling, step elimination etc. even if the 
> rounding algorithms etc. at the chip level are the same, the way the 
> code is compiled and calculations sequenced will change the error in the 
> final answer.
> ...

LDC disables -ffast-math by default.

> Then, variations in pipelining and caching at the processor level could 
> also affect the answer.
> ...

No.

> And if you move on to different computing paradigms such as quantum 
> computing and other as yet undiscovered techniques, again the way 
> operations and rounding etc is compounded will cause divergences in 
> computations.
> ...

Yes, if you move on to an analog computing paradigm with imperfect error 
correction, full reproducibility will go out the window. Floating point 
is not that though.

> Now, we could insist that we somehow legislate for the way compound 
> calculations are conducted. But that would cripple the speed of 
> calculations for some processor architecture / paradigms for a goal 
> (reproduceability) which is worthy, but for 99% of usages not 
> sufficiently beneficial to pay the big price in performance.
> 
> 

It's really not that expensive, changing the result via optimizations is 
disabled by default in LDC, and actually, how do you know that the 
compiler does not pessimize your hand-optimized compound operations.

I am not even against people being able to pass -ffast-math, it should 
just not destroy the correctness and reproducibility of everyone else's 
computations.