Include `decimal_rs` into benchmarks

[xamgore]

Hey, thanks for such a great library, I really like your approach which is contrary to decimal_rs where a lot of unwrapping and Results are needed. The crates are similar, and fastnum has a certain flexibility, as you can switch from 128 → 256 and exceptions handling.

Though, I don't quite understand the reason you have decided to exclude decimal_rs from the benchmarks.

[neogenie]

Thank you! @xamgore

The main reason it's not included in the comparison at the moment is that decimal_rs is not arbitrary precision and is limited to using only 64 and 128-bit containers. But I will definitely add it to the benchmarks soon.

[delta4chat]

can also add num-bigint, rug and crypto-bigint to benchmarks to compare their integer performance?

maybe num-bigint is (very little) faster than bigdecimal as it doesn't handle scale and fractional parts?

rug actually uses the GMP binding. it also provides float numbers, but it is not accurate (0.1 + 0.2 != 0.3), but it can be added to the benchmark because it doesn't care about correctness.

[xamgore]

@delta4chat hey, if you are familiar with them, you could probably help with the benchmarks?

[delta4chat]

maybe tomorrow, but I am working for others thing, such as avoid unnecessary copy in most math operations (by using reference instead of take ownership to methods like add, sub, mul, div, pow, cos, tan, etc.), if do this, some performance improvement is expected, since the size of fastnum is usually larger than the CPU pointer width (64 or 32), so using references instead of implicit copy should result in better performance, especially for huge types such as D8192. However, this is limited to Decimal and UnsignedDecimal types, and if a similar performance improvement is needed for Int, it will involve a PR to the upstream bnum library.

[neogenie]

It's not that simple :)

1. First, using pass-by-value instead of pass-by-reference is already an optimization that avoids indirect addressing, which in practice is usually significantly more expensive than copying in the stack or registers due to the CPU cache.
2. Second, most functions are inline, so the way arguments are passed in most cases doesn't really matter.
3. The main focus now is on optimizing 128, 256, and 512-bit numbers, since types like D8192 are rather exotic.

[delta4chat]

Well, maybe copying on the stack is cheaper, so much so that it's possible not to use references? Or does using references lead to lower or higher performance?

It's still an interesting topic though, and I'll still try it out and compare the performance differences and learn from it.