Benchmarks
Opened this issue · 5 comments
Do you have benchmarks for measuring the performance of double/float against fpm? Or plan to add this.
P.S is the fpm can be automatically vectorized by the compilers
Hi, https://github.com/MikeLankamp/fpm/blob/master/docs/performance.md (linked from the README) should contain what you need.
Code using FPM may be vectorizable depending on its use and the compiler; I can't make blanket statement.
Does this answer your questions?
Yep, just looking for an fixed <48,16> or fixed<32,16> and its performance. But I am not sure if the library supports this. Is it supported? I saw the closed issue but this is unclear whether or not it is supported now.
@MikeLankamp this one #5 (is this issue with the support of fixed<48,16> is abandoned?)
@NikitaMishin As mentioned in issue #5, this library already supports fixed<48, 16>. I can confirm as just started using fixed<32, 32> today
On topic: I'd also like +1 for benchmarks using larger type size. In fact, I'd love a benchmark for fixed<std::int64_t, boost::multiprecision::int128_t, 32> specifically, but that's because that's exactly what I'm using and won't be benchmarking myself for a long time coming =)
In case anyone is looking for some quick numbers, I did a quick test comparing standard fixed_16_16 vs fixed<48, 16> with int64_t and boost's int128_t in the context of my (very early phase) game.
- Context:
- What's being used: Using my game library/simulation layer which primarily consists of a custom, not-optimized physics engine with some ECS and simple physics-based gameplay. This "sim layer" library is then used on top of vanilla UE4, with the goal being to achieve full determinism
- Metric: How long does a single sim library frame update take to process
- Method: Just logging metric, playing for a few seconds, then scrolling through all the numbers. All numbers below are just a quick and rough estimate
- Data
| "Float" type | Range for frame update time | Est. mean |
|---|---|---|
| pre-fpm raw floats | ~80-400us | ~150us |
| fixed_16_16 | ~400-750us | ~500us |
| fixed<48, 16> w/ boost | ~4-7ms | ~5ms |
That's about a 10x increase in time per frame update from switching to fixed<std::int64_t, boost::multiprecision::int128_t, 32> in my specific case... which is way too slow for my needs
Edit:
Tested fpm::fixed<std::int64_t, std::int64_t, 16> and it has about the same performance as fixed_16_16 in my context