Optimize raw backend PCM conversion path in audioread/rawread.py

[AsymptotaX]

What changed

• byteswap(s):

  • replaced per-chunk Python struct.pack/unpack loop with array('h').byteswap()
  • this reduces Python-level overhead and allocations

• read_data(self, block_samples=1024):

  • precompute needs_conversion and needs_byteswap outside the loop
  • skip audioop.lin2lin(...) when old_width == TARGET_WIDTH
  • keep byteswap only when required (big-endian / non-sowt)

Environment

• Machine: Mac mini M4, 16GB RAM
• Backend used: rawread.RawAudioFile
• Inputs: 4 WAV + 4 AIFF files
• Repeats: 5 runs per file

Results

file	size	samples	Old mean(ms)	New mean(ms)	Speedup
300s.wav	50.47 MiB	13230000	31.95	11.78	2.71x
60s.wav	10.09 MiB	2646000	14.83	2.64	5.62x
20s.wav	3.36 MiB	882000	2.83	0.99	2.86x
5s.wav	861.4 KiB	220500	0.97	0.22	4.41x
300s.aiff	50.47 MiB	13230000	3355.95	25.01	134.18x
60s.aiff	10.09 MiB	2646000	626.81	4.50	139.29x
20s.aiff	3.36 MiB	882000	209.07	1.37	152.61x
5s.aiff	861.4 KiB	220500	52.30	0.36	145.28x

[semohr]

LGTM. Sorry for the long delay. Was just investigating beetbox/pyacoustid#88 and found we already handle byteswaps correctly in here.

Also thought that we should no do the swap in a python loop 😄