libsndfile
Multi threading
Experimental multi-thread implementation. (disabled by default with BUILD_MULTI_THREADING build option.)
Related issue: Parallelize resampling (with high delay) #176
compiled and tested on
FreeBSD clang version 14.0.5 (https://github.com/llvm/llvm-project.git llvmorg-14.0.5-0-gc12386ae247c) Target: x86_64-unknown-freebsd13.2 Thread model: posix
@SoapGentoo , @arthurt , what do you think? I guess we can add it when it's ready.
Hi. @evpobr
Sorry for that I did some commits after you had reviewed the code once. The change may boost the performance when an integer equivalant src_ratio is in use. I hope it works ok since I quickly merged it from my local branch (which supports S32 interface; if you are also interested: https://github.com/ss3git/libsamplerate/tree/S32Interface).
I have no more changes to this branch planned, so I'm leaving this up to you (or, maintainers) unless I find bugs.
One thing I'm aware of is that if you are actually releasing it, adding converter types like "SRC_SINC_BEST_QUALITY_MT" may be an idea to have both ST and MT versions switchable in one library.
Hello @luzpaz.
This PR works well enough at least on my machine. Maybe waiting for more testers other than me?
@ss3git , are the results bitwise identical to the reference implementation?
Hi, @evpobr .
I've already forgotten a lot of the details though, the core calculation part is intended to be bitwise identical to the reference implementation (if you mean the single-thread version).
Since the single-thread implementation seems to have input/output processing variance depending on the number of channels processed (e.g. https://github.com/libsndfile/libsamplerate/issues/206), I am not hundred percent sure if the outputs "always" match.
One thing I'm aware of is that if you are actually releasing it, adding converter types like "SRC_SINC_BEST_QUALITY_MT" may be an idea to have both ST and MT versions switchable in one library.
Good idea.