double → float Migration Incomplete

[ankitkumarrain]

Issue: v1.0.3 fixed "most" double-precision ops in SpeexLib — but "most" implies some remain. On Cortex-M targets without an FPU supporting double, this causes soft-float fallback and significant performance penalties.

Reason : AudioMark targets embedded MCUs (e.g., Cortex-M4/M33) that may lack a double-precision FPU. Any double usage causes the compiler to emit soft-float fallback code, significantly degrading performance and contradicting the benchmark's stated goal of single-precision operation.

Files Requiring changes are : lib/speexdsp/libspeexdsp/resample.c, lib/speexdsp/libspeexdsp/preprocess.c, lib/speexdsp/libspeexdsp/kiss_fft.c, lib/speexdsp/libspeexdsp/kiss_fftr.c

[ankitkumarrain]

Hii @joseph-yiu, @llefaucheur Should i propose Pull Request on this bug.

[llefaucheur]

Using double precision on Cortex-M4 is a bad idea when you are looking at speed.
Do not worry since AudioMark do not use those libraries (when compiled for ARM processors), even if the source code comes with the full speexdsp package.

[ankitkumarrain]

@llefaucheur Thank you for the clarification.

[ankitkumarrain]

Hi @llefaucheur ,
I’d like to continue contributing — is there a specific issue or component you’d recommend I work on next?

[llefaucheur]

Hi @ankitkumarrain , you write (in "https://github.com/ankitkumarrain/RISC-V_Audiomark") the code must be bit-exact. This is not true. I arranged DSP vendors with fixed-point multipliers to pass the end of computation criterion (it is 4 years back now but on top of my mind you only need accuracy with average error at -70dB below full-scale, or 11bits if you prefer).

As told in #95 this benchmark was used as a proof of concept for more generic stream based processing (see here https://github.com/llefaucheur/NanoGraph/blob/main/doc/NanoGraph_presentation.pdf and in other repositories) for video, motion sensor, etc.. with the ultimate idea of building a public repository of computing nodes (imagine a node for beamforming, another for speech recognition, NN models, etc..). You will see some proof of concepts using Renesas-RA8 and others to come. >> I will appreciate to check your interest in low/no-code approach for tiny devices << . Write to (laurent.lefaucheur) at (arm.com) or LinkedIn. Regards. Laurent.