numpy float: 10x slower than builtin in arithmetic operations?

by

CPython floats are allocated in chunks

The key problem with comparing numpy scalar allocations to the float type is that CPython always allocates the memory for float and int objects in blocks of size N.

Internally, CPython maintains a linked list of blocks each large enough to hold N float objects. When you call float(1) CPython checks if there is space available in the current block; if not it allocates a new block. Once it has space in the current block it simply initializes that space and returns a pointer to it.

On my machine each block can hold 41 float objects, so there is some overhead for the first float(1) call but the next 40 run much faster as the memory is allocated and ready.

Slow numpy.float32 vs. numpy.float64

It appears that numpy has 2 paths it can take when creating a scalar type: fast and slow. This depends on whether the scalar type has a Python base class to which it can defer for argument conversion.

For some reason numpy.float32 is hard-coded to take the slower path (defined by the _WORK0 macro), while numpy.float64 gets a chance to take the faster path (defined by the _WORK1 macro). Note that scalartypes.c.src is a template which generates scalartypes.c at build time.

You can visualize this in Cachegrind. I’ve included screen captures showing how many more calls are made to construct a float32 vs. float64:

float64 takes the fast path

float64 takes the fast path

float32 takes the slow path

float32 takes the slow path

Updated – Which type takes the slow/fast path may depend on whether the OS is 32-bit vs 64-bit. On my test system, Ubuntu Lucid 64-bit, the float64 type is 10 times faster than float32.

More Related Contents:

  1. Roll rows of a matrix independently
  2. Link ATLAS/MKL to an installed Numpy
  3. Bin elements per row – Vectorized 2D Bincount for NumPy
  4. Fastest way to grow a numpy numeric array
  5. Vectorized searchsorted numpy
  6. Implement MATLAB’s im2col ‘sliding’ in Python
  7. Python’s sum vs. NumPy’s numpy.sum
  8. Why “numpy.any” has no short-circuit mechanism?
  9. Vectorizing Haversine distance calculation in Python
  10. Numpy and line intersections
  11. Why is numpy.array so slow?
  12. Efficient evaluation of a function at every cell of a NumPy array
  13. why isn’t numpy.mean multithreaded?
  14. Why list comprehension is much faster than numpy for multiplying arrays?
  15. Why is `np.sum(range(N))` very slow?
  16. Why is a `for` over a Python list faster than over a Numpy array?
  17. Performance: Matlab vs Python
  18. Are NumPy’s math functions faster than Python’s?
  19. Max in a sliding window in NumPy array
  20. No speedup when summing uint16 vs uint64 arrays with NumPy?
  21. Efficiently replace elements in array based on dictionary – NumPy / Python
  22. Why is Numpy much faster at creating a Zero array compared to replacing the values of an existing array with zeros?
  23. Function to determine if two numbers are nearly equal when rounded to n significant decimal digits
  24. Multiple matrix multiplication
  25. What is the most efficient way to check if a value exists in a NumPy array?
  26. Efficient dot products of large memory-mapped arrays
  27. mean from pandas and numpy differ
  28. Creating a numpy array of 3D coordinates from three 1D arrays
  29. Split Python sequence (time series/array) into subsequences with overlap
  30. numpy faster than numba and cython , how to improve numba code

Leave a Comment