Provide a method to temporarily suppress model prewarm behavior

[mikeoliphant]

It could be useful to have a way to temporarily suppress model prewarming if you want to take several actions (change slimmable size, set maximum buffer size, etc.) that would each individually prewarm the model.

Ideally, this would be possible to do before calling get_dsp(), but the model doesn't yet exist. Making it static would allow this, but might be undesirable.

[rerdavies]

Fix for the spurious prewarm() calls.

A two line fix.

(1) Remove the prewarm() call from

dsp.cpp: 101

void nam::DSP::Reset(const double sampleRate, const int maxBufferSize)

(2) Remove the Prewarm() call from

get_dsp.cpp:221

std::unique_ptr<DSP> create_dsp(std::unique_ptr<ModelConfig> config, std::vector<float> weights,
                                const ModelMetadata& metadata)

I have tested these changes, and they seem to work, at least superficially (I only have WavNet A2 slimmable models to test with). With these changes, prewarm() gets called once and only once. and only gets called when Prewarm() is called. As it turns out, prewarm() gets called in SetSlimmableWeight() only because it calls Reset.

This is admittedly a potentially breaking change in behaviour: Prewarm() now becomes a mandatory call after get_dsp(), SetSlimmableWeight, and Reset().

Proposed Mitigations For Risk

It's difficult to write tests for this, because prewarming state is completely hidden from public APIs. And even then, test cases are not going to catch breaking changes to shipping products.

A runtime check, however, is pretty straightforward. One could set a flag in prewarm() on each Model, and modify all Process calls to throw an exception if prewarmed is false. The advantage of doing so is that it would almost certainly catch any cases where somebody is about to ship a product that contains the proposed changes, but have not made the now-mandatory Prewarm call. The flag would be cleared in SetSlimmableWeight and Reset() too, since these must now be followed by a Prewarm() as well. So it would catch existing cases where Slimmable models have not been pre-warmed after a SetSlimmableSize call (which seems like a good thing).

The throw of the exception would need to be guarded by an #ifdef so that NAM Core can be built in environments where consumers of the library have chosen to disable C++ compiler exceptions. This does not need to be documented. The throw statement will generate an error if being compiled with exceptions disabled. And consumers will see a comment before the throw statement, warning them about Prewarm() now being a mandatory call.

If a decision is made not to throw an exception, the following code would at least guarantee that returned models are realtime-safe, and will not allocate on the realtime thread, even if they are not prewarmed:

    // Ensure all working buffers are allocated, and not re-allocate on a realtime thread.
    SetMaxBufferSize(NAM_DEFAULT_MAX_BUFFER_SIZE); 
    Do  a brief 64-sample prewarm() in order to pre-allocate working EIGEN matrices before hitting a realtime thread.

Would you like a Pull request?

I have the two line fix as a commit, tested and ready to be pulled, but it seems pretty silly to actually pull for a two line change. Adding the mitigation code is a little bit more work, which would make a sensible pull request. I can do that if you like.

[rerdavies]

OR... one more option, and then I'll be quiet.

A cmake option that makes the Prewarm() call mandatory. Once that's true, the fix becomes startlingly simple I think. Just conditionally #IFDEF out the two prewarm() calls given above.