C++20 Branch

The original C++17 has been deprecated, as requires and concepts are just too important for readability and the ability to conditionally enable code.

CPU

Compiler

It was tested with Clang++-18 [Ubuntu clang version 18.1.6 (++20240518023229+1118c2e05e67-1~exp1~20240518143321.130)] . As well as g++-12 [gcc version 12.3.0 (Ubuntu 12.3.0-1ubuntu1~22.04)].

Flags

LLVM

OPT

-Os -std=c++20 -march=native -fconstexpr-steps=100000000

Warnings/Info

-Wall -Wpedantic -Winline -fstack-usage

pybind

-shared -fPIC $(python3 -m pybind11 --includes)

GCC

The new version is not compadiable with -Os anymore as for some reason functions are not inlined anymore correctly. Furthermore, its not compadable with the AXV2 MAC operations implemented here.

OPT

-O3 -std=c++20 -march=native

Warnings/Info

-Wall -Wpedantic -fstack-usage -Wno-inline -fdump-tree-optimized

pybind

-shared -fPIC $(python3 -m pybind11 --includes)

ARM

Compiler

We used STM32CubeIDE 1.15.0 with the Tool chain: 'GNU Tools for STM32 (12.3.rel1)'.

We're using C++20

Flags

We're compiling with O3 or Ofast. Os does not always inline ´__attribute__((always_inline))´ functions, which breaks design choises of this framework. Most data access is behind 3-4 of such functions, and only if they are inlined they can be fully resolved at compiletime leading to zero overhead and essentailly raw pointer operations.

GCC Compiler

-flto -ffast-math -funsafe-math-optimizations -fno-math-errno -fno-trapping-math -frename-registers -fprefetch-loop-arrays -fpredictive-commoning -fgcse-after-reload -falign-loops=4 -ffunction-sections -fdata-sections

G++ Compiler

-flto -Winline -fstack-usage -ffast-math -funsafe-math-optimizations -fno-math-errno -fno-trapping-math -fno-exceptions -fno-rtti -frename-registers -fprefetch-loop-arrays -fpredictive-commoning -fgcse-after-reload -falign-loops=4 -ftree-vectorize -fvect-cost-model=cheap -fsched-pressure -ffunction-sections -fdata-sections -fno-threadsafe-statics -fno-cxa-atexit -nostdlib++

Linker

-flto -nostdlib++

Perfrormance

SEdge

Tiny has slightly inreased in temp. memory compared to the C++17 variation, as it now also has a full copy of the outputs (float[35]) present, even though it does not use it.

Name S-Edge-	downsampling factors	Network Size [Byte]	Perm. Memory [Byte]	Temp. Memory [Byte]	Total FLOPS [MFLOP]	MCU Time [s]	Speed [MFLOP/s]	Acc. (Cpp on PC) [%]
Full RZOH	[1, 1, 1, 1, 1, 1]	563528	3456	768	4406.95	87.0000	50.65	95.19
Full RZOH	[1, 1, 1, 2, 4, 12]	563528	3456	768	2309.48	47.0000	49.14	95.22
Full RZOH	[1, 1, 4, 8, 16, 32]	563528	3456	768	1198.31	23.5652	50.85	94.99
Full RZOH	[1, 2, 14, 14, 42, 42]	563528	3456	768	573.51	11.0938	51.70	94.08
L RZOH	[1, 1, 1, 1, 1, 1]	224592	2944	480	1721.41	35.1360	48.99	94.11
L RZOH	[1, 1, 1, 2, 2, 38]	224592	2944	480	964.38	19.4856	49.49	94.19
L RZOH	[1, 2, 2, 8, 8, 48]	224592	2944	480	413.89	8.4339	49.07	93.99
L RZOH	[1, 3, 3, 6, 30, 60]	224592	2944	480	295.64	6.0680	48.72	93.09
S no RZOH	[1, 1, 1, 1, 1, 1]	80984	1104	472	572.50	11.6736	49.04	93.37
S no RZOH	[1, 1, 1, 1, 2, 22]	80984	1104	472	237.21	5.0897	46.61	93.47
S no RZOH	[1, 1, 3, 6, 6, 42]	80984	1104	472	93.71	2.0427	45.88	93.02
S RZOH	[1, 1, 1, 1, 1, 1]	80984	1104	472	572.50	11.6736	49.04	93.08
S RZOH	[1, 1, 1, 1, 9, 63]	80984	1104	472	183.72	4.0569	45.28	93.18
S RZOH	[1, 2, 2, 6, 18, 54]	80984	1104	472	71.21	1.5736	45.25	93.05
S RZOH	[1, 2, 4, 16, 32, 64]	80984	1104	472	47.29	1.1149	42.42	92.02
S RZOH	[1, 4, 4, 16, 32, 64]	80984	1104	472	39.25	0.8689	45.17	91.06
S RZOH	[1, 4, 4, 16, 64, 64]	80984	1104	472	37.44	0.8374	44.71	90.19
Tiny no RZOH	[1, 1, 1]	33424	704	396	193.61	4.1833	46.28	90.50
Tiny no RZOH	[1, 1, 5]	33424	704	396	65.29	1.3880	47.04	90.62
Tiny no RZOH	[1, 2, 22]	33424	704	396	26.98	0.6175	43.69	90.18
Tiny no RZOH	[1, 4, 28]	33424	704	396	18.65	0.4111	45.36	88.96
Tiny RZOH	[1, 1, 1]	33424	704	396	193.61	4.1833	46.28	89.41
Tiny RZOH	[1, 2, 6]	33424	704	396	46.42	1.0290	45.11	89.51
Tiny RZOH	[1, 5, 20]	33424	704	396	19.58	0.4307	45.46	89.12
Tiny RZOH	[1, 9, 54]	33424	704	396	12.13	0.2660	45.62	88.01
Tiny RZOH	[1, 11, 55]	33424	704	396	11.53	0.2518	45.78	87.11
Tiny RZOH	[2, 8, 64]	33424	704	396	8.97	0.2092	42.85	80.35
Tiny RZOH	[3, 12, 48]	33424	704	396	7.66	0.1816	42.17	70.32

cursive means they where measured with a smartphone stop watch, and are therefore highly inaccurate

bold are optimization points, the configurations that where hand tuned for maximum performance, which are then applied for all other configurations of that network.

Citation

Please use the following when citing our work:

@article{bittner2025sedge,
  author       = {Matthias Bittner and
                  Daniel Schnöll and
                  Matthias Wess and
                  Axel Jantsch},
  title        = {Efficient and interpretable raw audio classification with diagonal
                  state space models},
  journal      = {Mach. Learn.},
  volume       = {114},
  number       = {8},
  pages        = {175},
  year         = {2025},
  url          = {https://doi.org/10.1007/s10994-025-06807-z},
  doi          = {10.1007/S10994-025-06807-Z},
  timestamp    = {Sun, 29 Jun 2025 21:59:59 +0200},
  biburl       = {https://dblp.org/rec/journals/ml/BittnerSWJ25.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}