I compared the running time to my PR and I do see about 10% reduction in time when the number of vectors is high when using your PR, so I disagree with the "no effect on performance" part ;)

In addition I received an issue when running your code:

   2622             heads = shafts[:, :1] - np.multiply.outer(arrow_dt, head_dirs)
   2623             # stack left and right head lines together
-> 2624             heads.shape = (len(arrow_dt), -1, 3)
   2625             # transpose to get a list of lines
   2626             heads = heads.swapaxes(0, 1)
AttributeError: incompatible shape for a non-contiguous array

This could be reproduced with this MRE:

x, y, z = [np.random.rand(10) for _ in range(3)]
X, Y, Z = np.meshgrid(x, y, z)
U, V, W = [np.random.rand(*X.shape) for _ in range(3)]
ax.quiver(X, Y, Z, U, V, W)

I believe I solved this with expand_dims before the matrix multiplication on my PR.

Can you confirm that you can reproduce this issue?

If you can reproduce this as well, that also means that the automatic checks missed it somehow and perhaps they need to be improved as well

I compared the running time to my PR and I do see about 10% reduction in time when the number of vectors is high when using your PR, so I disagree with the "no effect on performance" part ;)

Is my version faster or yours?

In addition I received an issue when running your code: (elided)

Did you miss https://github.com/matplotlib/matplotlib/pull/15373/files#diff-84224cb1c8cd1f13b7adc5930ee2fc8fR2625?

Did you miss https://github.com/matplotlib/matplotlib/pull/15373/files#diff-84224cb1c8cd1f13b7adc5930ee2fc8fR2625?

I indeed missed the reshape, thanks! Now there is no issue indeed :)

Is my version faster or yours?

I ran it again and it isn't faster now. I guess this happened because I only looked at the mean and not at the std as well?

rebased

Flake8:

./lib/mpl_toolkits/mplot3d/axes3d.py:2544:17: E128 continuation line under-indented for visual indent

fixed