英雄帖：Investigate SIMD optimization for SoundTouch

[xen0n]

Repo: https://codeberg.org/soundtouch/soundtouch

This library is used by Firefox to handle audio time-stretching for <video> or <audio> elements. It is not bottle-necking video playback but some functions from it do show up in perf top measurements when I watch a video on Bilibili with 2x playback rate for example.

Example perf top output on 3A6000:

   PerfTop:    5738 irqs/sec  kernel:27.2%  exact:  0.0% lost: 0/0 drop: 0/0 [4000Hz cycles:P],  (all, 8 CPUs)
-----------------------------------------------------------------------------------------------------------------------------

     9.08%  liblgpllibs.so  [.] soundtouch::TDStretch::calcCrossCorrAccumulate
     4.73%  [kernel]        [k] finish_task_switch.isra.0
     3.09%  [kernel]        [k] __arch_cpu_idle
     1.56%  firefox         [.] arena_t::MallocSmall
     1.08%  libc.so.6       [.] memset
     1.08%  firefox         [.] arena_dalloc
     1.06%  libc.so.6       [.] _wordcopy_fwd_aligned
     1.05%  libc.so.6       [.] __GI___pthread_mutex_unlock_usercnt
     0.94%  libc.so.6       [.] pthread_mutex_lock@@GLIBC_2.36
     0.74%  liblgpllibs.so  [.] soundtouch::FIRFilter::evaluateFilterStereo
     0.69%  libxul.so       [.] nsIFrame::BuildDisplayListForChild
     0.66%  [kernel]        [k] try_to_wake_up
     0.61%  libc.so.6       [.] memcpy
     0.58%  libc.so.6       [.] memcmp
     0.51%  [kernel]        [k] syscall_enter_from_user_mode

Feel free to investigate whether hand-written SIMD optimization or generic auto-vectorization would help. (In my case the hottest function was not utilizing LoongArch SIMD but my compiler flag is just -O2 so I'm not excluding the possibility that a simple change to -O3 would solve it.)

[xry111]

Is --enable-openmp helping?