Properly record number of samples between tags during frequency sweeps

[jcfitzpatrick12]

In this GNU Radio issue, Marcus Mueller pointed out that our implementation of Batched File Sink is broken by design. We're currently recording the number of samples per tag as single-precision floats. However, if the number of samples between tags is greater than the maximum integer you can store with single-precision floats we get undefined behaviour. I think this has been causing the mysterious bug faced with longer dwell time frequency sweeps.

In this issue, we should first confirm that our use of single-precision floats to record the number of samples per tag causes unwanted errors in frequency sweeps with longer dwell times. Then, we can redesign and reimplement the block to properly handle this.

[marcusmueller]

Easiest solution might simply to convert the taglist from vector<float>(2*N) to vector<pair<uint64_t, double>>(N) and to d_tags_buffer[d_n_buffered_tags++] = {offset, frequency}.

(you'll notice that sizeof(std::pair<uint64_t, float>)==16==sizeof(std::pair<uint64_t, double>) due to alignment constraints on all practical platforms (clang and GCC targetting ARM32/64, RISC-V32/64, x86_64, but not i386) so store things as double, solving potential frequency resolution issues on the way)

[jcfitzpatrick12]

@marcusmueller, Appreciate you taking the time to look! That's a neat way to handle it.

It shouldn't be too much trouble to adapt the existing code that reads these new metadata files either (see here). Though problematic, it is currently very convenient to read when it is all one data type. But I'm sure there's a nice NumPy way to handle interleaved values of different types.

[marcusmueller]

Yep, you can define your own interleaved dtype in numpy!

On 15 February 2026 11:58:14 CET, Jimmy Fitzpatrick ***@***.***> wrote: jcfitzpatrick12 left a comment (spectregrams/spectre#209) @marcusmueller, Appreciate you taking the time to look! That's a neat way to handle it. It shouldn't be too much trouble to adapt the existing code that reads these new metadata files either (see [here](https://github.com/jcfitzpatrick12/spectre-core/blob/v3.0.0/src/spectre_core/batches/_iq_stream.py#L80-L100)). Though problematic, it is currently very convenient to read when it is all one data type. But I'm sure there's a nice NumPy way to handle interleaved values of different types. -- Reply to this email directly or view it on GitHub: #209 (comment) You are receiving this because you were mentioned. Message ID: ***@***.***>

-- Sent from my Android device with K-9 Mail. Please excuse my brevity.