[FEATURE REQUEST] Support for bincount?

[kwagyeman]

https://numpy.org/doc/stable/reference/generated/numpy.bincount.html?

It solves a very challenging problem that otherwise has to be done in Python:

        # Reset the frequency histogram.
        freq_histogram[:, :, :] = 0

        # Flatten 3D index: (y, x, f) → idx
        flat_idx = (y_bins * X_BINS + x_bins) * FREQ_BINS + freq_indices

        # Update the frequency histogram.
        for i in range(len(target_events)):
            freq_histogram[flat_idx[i]] += gain[i]

[v923z]

Yes, this could be done with little effort. Here are a couple of comments/questions:

1. bincount is for 1-D arrays, so you would still need to flatten your array. We could, in principle, treat higher-dimensional tensors, but that would be breaking with numpy's convention.
2. numpy always returns an array of uint64 type, except, when you supply the weights keyword argument. We would return uint16, as that is the largest integer type in our case.
3. Do you also need support for the weights keyword?
4. Do you need minlength?

[kwagyeman]

It would be used like this:

# Keep ON or OFF events (use OR, not AND)
mask = (events_slice[:, 0] == csi.PIX_ON_EVENT) | (events_slice[:, 0] == csi.PIX_OFF_EVENT)
if np.any(mask):
    target_events = events_slice[mask]

    # +1 for ON, -1 for OFF (use ±64 if you want by multiplying at the end)
    gain = np.where(target_events[:, 0] == csi.PIX_ON_EVENT, 1, -1)

    # Timestamps → µs
    ts = target_events[:, 1] * 1000000 + target_events[:, 2] * 1000 + target_events[:, 3]

    # Frequency bin per event
    freq_indices = (ts * FREQ_BINS // 1000000) % FREQ_BINS

    # Spatial bins
    x_bins = (target_events[:, 4] * X_BINS) // 320
    y_bins = (target_events[:, 5] * Y_BINS) // 320

    # Flatten 3D index: (y, x, f) → idx
    flat_idx = (y_bins * X_BINS + x_bins) * FREQ_BINS + freq_indices

    # Zero the histogram quickly
    freq_histogram[:] = 0

    # Accumulate in flat space
    flat_len = Y_BINS * X_BINS * FREQ_BINS
    updates = np.bincount(flat_idx, weights=gain, minlength=flat_len)

    # Write back into 3D histogram
    freq_histogram.reshape((Y_BINS * X_BINS * FREQ_BINS,))[:] = updates
    freq_histogram.reshape((Y_BINS, X_BINS, FREQ_BINS))

1. Okay
2. Okay
3. Yes
4. Yes

[kwagyeman]

One issue I ran into with uint16s though is that for a 32x32 x/y bin image with 256 deep FFTs you overflowed the uint16 indices... not sure how to overcome this with the above.

I had to do full 3d indices to get it working.

        # Update the frequency histogram.
        for i in range(len(target_events)):
            freq_histogram_3d[y_bins[i], x_bins[i], freq_indices[i]] += gain[i]

[v923z]

One issue I ran into with uint16s though is that for a 32x32 x/y bin image with 256 deep FFTs you overflowed the uint16 indices... not sure how to overcome this with the above.

I'm not, in principle, against adding other dtypes, but that's really a Pandora's box, and would require a lot of consideration. Other people bring this up from time to time, so if you want to go down that path, then I would suggest that we open a separate thread in https://github.com/v923z/micropython-ulab/discussions.

[v923z]

@kwagyeman Basic functionality is there (#738), I'll add the keyword parsing tomorrow.

[v923z]

This has been implemented in #738, so closing now.