Skip to content

numpy-performance.md

Latest commit

 

History

History
79 lines (60 loc) · 2.88 KB

numpy-performance.md

File metadata and controls

79 lines (60 loc) · 2.88 KB
Raw

NumPy API measurements

Apple M1 Pro; Public NumPy calls include dispatch and any unpacking; np.asarray(ndarray) may share data, while tightarray always copies. Deterministic periodic input; no isolation or cold-cache control.

These timings include NumPy dispatch and unpacking, separately from direct native methods.

count_nonzero(ndarray) needs no equality-mask allocation, unlike the value-count baseline in the native benchmark.

Median microseconds per public NumPy call; lower is better.

65,536 elements, 2 bits

NumPy call ndarray Packed Word-aligned
asarray 0.053 8.961 8.930
array_equal 4.272 2.698 2.738
count_nonzero 2.464 2.561 2.576
take 1.144 2.373 2.393
add 1.572 15.620 15.298

65,536 elements, 5 bits

NumPy call ndarray Packed Word-aligned
asarray 0.053 12.802 11.822
array_equal 4.256 3.411 3.580
count_nonzero 2.420 5.038 3.672
take 1.147 2.472 2.408
add 1.423 19.465 19.422

65,536 elements, 7 bits

NumPy call ndarray Packed Word-aligned
asarray 0.053 15.962 14.417
array_equal 4.248 4.070 4.124
count_nonzero 2.468 5.187 4.246
take 1.130 2.465 2.432
add 1.644 22.871 21.329

1,048,576 elements, 2 bits

NumPy call ndarray Packed Word-aligned
asarray 0.053 121.519 119.237
array_equal 49.545 9.259 9.159
count_nonzero 36.121 12.724 12.759
take 1.148 2.334 2.374
add 22.833 146.390 144.589

1,048,576 elements, 5 bits

NumPy call ndarray Packed Word-aligned
asarray 0.052 179.711 164.987
array_equal 49.510 18.854 19.941
count_nonzero 36.097 53.946 30.160
take 1.141 2.528 2.373
add 22.928 207.891 190.232

1,048,576 elements, 7 bits

NumPy call ndarray Packed Word-aligned
asarray 0.052 370.960 201.815
array_equal 49.490 25.578 25.669
count_nonzero 36.123 53.952 39.466
take 1.144 2.492 2.477
add 22.743 254.609 229.810

Interpretation

  • Whole-array equality retains a packed-kernel advantage at these sizes.
  • NumPy take dispatch costs more than direct .gather() and is slower than take on an existing ndarray here.
  • Ufunc arithmetic unpacks first; it is interoperability, not an accelerated arithmetic kernel.
  • np.asarray(ndarray) can share storage; tightarray always creates an independent writable array.
  • Result retention and traced peak memory are recorded per method. Conversion peaks include intermediate bytes and a writable bytearray.

Raw samples and source hashes · Timing and memory CSV