Daniel Lemire
Daniel Lemire
It is likely that this PR is as good as it is going to guess on the short term. This is very good code and it is a very good...
We need documentation before merging this, but I am still opening up this issue for comments. If you use simdjson, please consider reviewing.
@the-moisrex I am not verifying your good results. See https://github.com/simdjson/simdjson/pull/2274
Indeed, we could use 16B aligned moves on x64 systems.
Reference https://ibraheem.ca/posts/128-bit-atomics/
We won’t change the default in the near term, as it would break existing code. However, we can take optional alternatives and if they provide significant, real performance benefits, advertize...
If threading is added, I think we need something like this in CMake: ```cmake option(SIMDUTF_ENABLE_THREADS "thread support" ON) if(SIMDUTF_ENABLE_THREADS) find_package(Threads REQUIRED) target_add_library(something_that_requires_threads PRIVATE Threads::Threads) target_compile_definitions(something_that_requires_threads PRIVATE SIMDUTF_THREADS_ENABLED=1) endif() ``` Right?...
> I guess it's quite promising, especially for our CI, which runs multitude of tests on so many platforms and compilers. We can parallelize the builds with an additional flag......
Agreed. Let us fix this.