search

jiixyj / libebur128

A library implementing the EBU R128 loudness standard.
MIT License
402 stars 76 forks source link

fix overflow when calculating number of frames in a window for dsd512… #130

Open bcoburn3 opened 2 months ago

bcoburn3 commented 2 months ago

This fixes an integer overflow that results in libebur128 calculating 0.0 for the loudness range for DSD512 files on Windows.

The sample rate for these files (2822400 samples/sec) multiplied by the window size (3000) overflows a 32 bit integer, resulting in a value for audio_data_frames being half what it should be. This PR re-orders the arithmetic to reduce the size of the intermediate values.

This doesn't happen on Linux/Mac because on those platforms unsigned long is 64 bits, but for Windows unsigned long is 32 bits on 64 bit systems.

jiixyj commented 2 months ago

Thanks a lot for the fix! If possible I'd avoid reordering the operations though, because it results in different values. What do you think about just casting the samplerate to unsigned long long before doing the arithmetic? This should work as well and the resulting values stay the same:

  st->d->audio_data_frames = (unsigned long long) st->samplerate * st->d->window / 1000;
bcoburn3 commented 2 months ago

I was sure I'd tried that cast and seen it not work, but I tried it and you're correct that re-ordering changes the result and that cast fixes the problem. Thanks for catching the change in results and suggesting this :)

for reference, here's what I did to check: https://godbolt.org/z/T35hGvGPW