A FUSE FS that does on-the-fly sample type conversion.

Currently this is only a proof-of-concept that operates on a single file with i16 values and converts to f32.

The f32 file view can be memmaped (e.g. in NumPy) without needing to convert the entire file. This is useful for random-access analysis of very large SDR capture files. I have tested with 100GB files with no issues.

This will work with scalar or complex values.

Example

# Make a mount point
mkdir testmnt

# Mount the DSP FS
cargo run -- test.cs16 testmnt &

# Check it out
ls -alh testmnt
hexdump testmnt/test.cs16

# Unmount when done
fusermount -u testmnt

If you do this in NumPy:

a = np.memmap("testmnt/test.cs16",dtype=np.complex64, mode="r")
plt.plot(np.real(a),np.imag(a))

You should see this:

Notes

Building fuser will require something like:

sudo apt-get install libfuse-dev pkg-config

See: https://github.com/cberner/fuser

Ideas / TODO

Support more conversion types
Rename converted files? (test.cs16 to be seen as test.cf32 or test.cs16.cf32)
Q: To scale or not to scale? (optional?)
Allow writing? (opposite direction conversion)
Any need for output formats other than f32? (e.g. f64)

Multiple file support

There are a few different ways to do this:

Specify multiple files on the command line
Use a configuration file
Use an entire directory as a source

Things get a little tricky when files have different formats. How do we specify the format for each file? Using recognizable file extensions seems like the easiest choice.

Other DSP operations

In general, I think it is preferable to do most signal processing in the client application. Numpy or whatever.

But there are some cases that might make sense in the FUSE layer.

Channelization

While the normal operation allows for easy random access to different offsets (times), some capture files may be very wide bandwidth, and only a subsection of the frequency spectrum is desired.

Down-sampling without frequency translation would be a special case of the same general technique.

Filtering would be required for this operation, so there is no reason not to support general filtering without resampling or frequency translation.

Parameter specification with multiple file support could get complex and would likely require the use of a configuration file.

etschneider / fuse-dsp