Amount of time to demux an audio file about demucs.cpp

[husnoo]

Thanks! I was able to run it on a 3090, it's much faster now!

…

On Tue, 9 Jan 2024, 12:23 Sevag H, ***@***.***> wrote: On my computer (AMD Ryzen 5950X processor), it takes ~34 seconds to demix a 5-second test track (with OMP_NUM_THREADS=1). That seems to line up with your measurement, 7x demixing time compared to track length Note that demucs.cpp was designed to use the demucs segment length of 7.8 seconds (that might explain where the 7 number comes from): - https://github.com/sevagh/demucs.cpp/blob/main/src/model.hpp#L669 - https://github.com/sevagh/demucs.cpp/blob/main/src/model.hpp#L19 — Reply to this email directly, view it on GitHub <#2 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAWJWSXGYZH4NF3KTAW5FTLYNUZCFAVCNFSM6AAAAABBNKZPPKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQOBSHE3TMNJTG4> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[husnoo]

Yeah it was the Facebook Demucs with pytorch. I'd prefer the cpp one but the torch GPU version was about 37x faster, so I couldn't ignore it haha!

…

On Thu, 11 Jan 2024, 14:46 Sevag H, ***@***.***> wrote: Nice, how? Did you use NVBLAS/cuBLAS? Or the proper PyTorch version of Demucs? I also have a 3090 but I don't target any GPU acceleration in the demucs.cpp code (the goal is to be simple and CPU-only, for now). — Reply to this email directly, view it on GitHub <#2 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAWJWSUUPUUXVPQAH6WVUE3YN73KZAVCNFSM6AAAAABBNKZPPKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQOBXGMZTONZRHA> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[sevagh]

On my computer (AMD Ryzen 5950X processor), it takes ~34 seconds to demix a 5-second test track (with OMP_NUM_THREADS=1). That seems to line up with your measurement, 7x demixing time compared to track length

Note that demucs.cpp was designed to use the demucs segment length of 7.8 seconds (that might explain where the 7 number comes from):

• https://github.com/sevagh/demucs.cpp/blob/main/src/model.hpp#L669
• https://github.com/sevagh/demucs.cpp/blob/main/src/model.hpp#L19

[sevagh]

Nice, how? Did you use NVBLAS/cuBLAS? Or the proper PyTorch version of Demucs?

I also have a 3090 but I don't target any GPU acceleration in the demucs.cpp code (the goal is to be simple and CPU-only, for now).