Codecov Report

❌ Patch coverage is 50.00000% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 61.85%. Comparing base (14b372c) to head (8ba90ce).
⚠️ Report is 1 commits behind head on main.

@@            Coverage Diff             @@
##             main    #3345      +/-   ##
==========================================
- Coverage   61.86%   61.85%   -0.01%     
==========================================
  Files          85       85              
  Lines       10111    10112       +1     
==========================================
  Hits         6255     6255              
- Misses       3856     3857       +1     

to test this, could we perhaps patch np.full and np.zeros to raise exceptions, that you would then catch in the test? I'm thinking there would be 2 test functions, 1 testing the various inputs that should hit the np.full branch, and another testing the various inputs that should hit the np.zeros branch. This could also work for a single test function that internally checks which numpy routine the input scalar should trigger

I've done some digging into this topic, I'll post in a bit but I would like to hold off on merging for a bit.

Ok apologies, I know that Tom's PR supercedes this in many ways but for anyone curious, my worry was about how memory is allocated.

I learned that zeros is only faster than full because it doesn't allocate memory when requested, lazily doing so later when needed (see https://stackoverflow.com/questions/70055063/how-is-memory-handled-once-touched-for-the-first-time-in-numpy-zeros/70056188#70056188). In benchmarks, I see the times of subsequent operations on the matrix, whether something like sum that should touch all elements or __setitem__ to be identical independent of initialization so I definitely think this should be merged!

import numpy as np
buf = np.arange(100_000, dtype=np.float64)
# cell 1
%%timeit full = np.full((1_000_000,), 0, dtype=np.float64)

full[900_000:] = buf
# cell 2
%%timeit zeros = np.zeros((1_000_000,), dtype=np.float64)
    
zeros[900_000:] = buf
# cell 3
%%timeit full = np.full((1_000_000,), 0, dtype=np.float64)

full.sum()
# cell 4
%%timeit zeros = np.zeros((1_000_000,), dtype=np.float64)

zeros.sum()