Daniel Lemire

Currently, there is no external-memory roaring data structure in this project. If you'd like to contribute one, it would be great.

1. It is unsafe to assume that the builtin is fallback. On my laptop, it is arm64. It should not be necessary. 2. The `count_elements()` that the code keeps computing...

Thanks. The functions that have to do with simdjson don’t have to be declared in your header file, do they? Or, at least, they don’t have to be in a...

> I'm on my Windows machine right now, so I can't give much info. If you are compiling under Windows, I recommend using Clang if possible. See https://lemire.me/blog/2023/03/03/float-parsing-benchmark-regular-visual-studio-clangcl-and-linux-gcc/

> It has actually increased the runtime speed of the fastgltf cases by often fourfold with both master and this branch. I see something like a factor of two... but...

@spnda Having a look.

> This lambda function takes up 36% of the entire average runtime of the parseAccessors function. It essentially just reads a number from an array, and then checks if it's...

This part is important: I recommend including a standard benchmark as part of your project. If I could just do... ``` cmake --build build ./build/benchmark ``` That would tremendously useful....

Feel free to steal some of my benchmarking code: https://github.com/lemire/Code-used-on-Daniel-Lemire-s-blog/tree/master/2023/11/28/benchmarks

My recommendation is to rely more on experiments and repeated high-quality benchmarks, than on profiling. Profilers mislead us all the time. They may point at a potential direction, but that's...