I feel like this project is a classic example of the README missing a "Why?" section. Not to justify the project, it just makes it hard to evaluate without understanding why they choose implementing a transpiler rather than an embedded WASM VM. Or why not transpile to native assembly? I'm sure they have good reasons but they aren't listed here.
i think your questions are implicitly answered in the top page of the readme, but by showing rather than telling
esp32 and stm32 run three different native assembly instruction sets (though this only supports two of them, the others listed in the first line of the readme are all arm)
the coremark results seem like an adequate answer to why you'd use a compiler rather than an interpreter
i do think it would be clearer, though, to explain up front that the module being built is not made of python code, but rather callable from python code
There's no scenario where a bespoke WASM interpreter is slower than MicroPython - though this isn't really Python.
WASM is an almost "ideal" bytecode target for embedded AOT native compilation as well, though yeah, you'd have to implement all of the backend targets.
maybe if you don't have enough space in flash for both the μpython interpreter and the wasm interpreter? but, yeah, this isn't being compiled to python but to c (for invocation from python)
> i do think it would be clearer, though, to explain up front that the module being built is not made of python code, but rather callable from python code
I know nothing about micropython. Do the modules contain bytecode?
In this case, the modules contain native code compiled for the target architecture. Micropython has something approximating a dynamic linker to load them.
tools/mpy_ld.py: https://github.com/micropython/micropython/blob/master/tools...
tools/mpy-tool.py lists opcodes: https://github.com/micropython/micropython/blob/master/tools...
Can the same be done with .pyc files; what are the advantages of MicroPython native modules?
Why does it need wasm2c?
I don't think .pyc files can contain native code, but .mpy can! And that's exactly what I'm using here:
- compile any language to wasm
- translate wasm to C99
- use the native target toolchain to compile C99 to .mpy
- run that on the device
If they're both Turing-complete opcodes instruction sets, Church-Turing says it is possible with just .pyc, too.
But that doesn't solve for WASM on a MicroPython device.
And that is why smart contracts without costed opcodes could not be built on Python.
.mpy modules can contain MicroPython bytecode and/or native machine code. In this case, WASM is compiled (via C) to native code. So the performance is very good, much better than interpreting either MicroPython bytecode or WASM bytecode.
The conventional way of creating native modules for MicroPython is to write them in C. This work allows to use any language that supports WASM output target.
Oh, I didn't know MicroPython was fully AOT. I'm surprised, since vanilla Python is typically interpreted with maybe a bytecode cache, right?
It isn’t fully AOT, but support for AOT modules is excellent.
Thanks for the explanation this project makes so much more sense now. If people want a little more detail, they have a brief how it works section in readme https://github.com/vshymanskyy/wasm2mpy?tab=readme-ov-file#h...
Thanks for the feedback. I'll improve the README based on your inputs. I was focused on the actual research. Overall, this allows writing code in statically compiled languages and running it (fast) on embedded systems with MicroPython. MicroPython itself is comparatively slow, and this provides tools to deliver more demanding software (AI, signal processing, etc).
There exists tooling already to compile for wasm from static languages (rust, C++, and so on) and there exists tooling to run this WASM on a raspberry pi (wasmer, etc).
That makes me feel the end goal here is not what's described ("make wasm so it can run on a raspberry"), but rather "make wasm run in micropython".
Am I missing something?
Yes, we're talking about Raspberry Pi Pico (RP2040) here. Not the Linux-based SBCs (Single Board Computers)
Curious if you've seen the WAMR runtime which is explicitly designed to be lightweight enough for freestanding embedded systems on some higher end microcontrollers (Cortex-M4F).
Yes, Wasm3 also targets Cortex-M4 and even Cortex-M0 :)
I see, thanks for the clarification!
Probably at least one of "it's cool" or "I'm already using MicroPython"
What would be the recommendation to run on ESP32?
https://github.com/wasm3/wasm3? https://github.com/espressif/esp-wasmachine ? https://github.com/bytecodealliance/wasm-micro-runtime/tree/... ? https://github.com/TOPLLab/WARDuino ?
For esp32 I'd recommend https://github.com/wasm3/embedded-wasm-apps
It obviously depends on your needs/goals.
P.S. I'm also the author of wasm3
Actually in my case I need a VM, because I want users to be able to upload "extension plugins" in real-time, without reuploading firmware
In that case which approach / library would you recommend to embedd a WASM VM?
This is exactly what wasm2mpy does for MicroPython: Dynamically loaded, AOT-compiled WASM modules
Okay great! But I'm not using MicroPython, just compiling directly for the ESP32, would that mean that to use wasm2mpy I would need to add a MicroPython VM in my code?
would you say that using wasm2mpy is better suited for my needs than having a wasm runtime like the one of bytecodealliance? What would be the advantage in this case?
Yes, for these microcontroller-type chips, ahead-of-time compilation would be definitely more suitable. WAMR probably supports it, but you can also to it via WASM2C, easily. I prefer not depending on WAMR, it is very bloated
thanks for all those answers! Slava Ukraini
Geroyam Slava!
thanks for your answer!! Can you tell me a bit more about why this suggestion might be more appropriate than the others?
Is it in terms of maturity? Speed? Memory footprint? Safety?
Wasm3 happily runs on ESP32, it provides a fully dynamic wasm runtime. But it's much slower than AOT compiled native code, obviously.
Can we have python compiled to wasm compiled to python compiled to wasm compiled to python - just for obfuscation reasons?
Jokes aside, when I first saw this I also assumed that WASM was being transpiled into a subset of micropython as some sort of homage to WASM's asm.js roots. The explanations in this thread make much more sense.
With this approach, wasm is compiled to native code. Translating that to Python is in theory possible, but very impractical.
Best add a LLM into the loop for real speed.
Author here, ready to answer any questions.
Like several others I'm confused about the tagline for this. If I understand correctly, it's not compiling "WASM to micropython", but to native binary code with generated micropython bindings?
Correct. I didn't create the tagline, sorry
A great way to allow code created in another language to be compiled to a MicroPython module and then run in a MicroPython environment.
These .mpy files are dynamically loaded into MicroPython using a regular python import.
[dead]