High Performance Python (from Training at EuroPython 2011) by Ian Ozsvald - HTML preview

CHAPTER

TEN

A (SLIGHTLY) FASTER CPYTHON IMPLEMENTATION

Having taken a look at bytecode, let’s make a small modification to the code. This modification is only necessary for CPython and PyPy - the C compiler options for us won’t need the modification.

All we’ll do is dereference the z[i] and q[i] calls once, rather than many times in the inner loops:

# \python\pure_python_2.py

for i in range(len(q)):

    zi = z[i]

    qi = q[i]

    ...

    for iteration in range(maxiter):

        zi = zi * zi + qi

        if abs(zi) > 2.0:

Now look at the kernprof.py output on our modified pure_python_2.py. We have the same number of function calls but they’re quicker - the big change being the cost of 2.6 seconds dropping to 2.2 seconds for the z = z * z + q line. If you’re curious about how the change is reflected in the underlying bytecode then I urge that you try the dis module on your modified code.

Here’s the improved bytecode:

>>> dis.dis(calculate_z_serial_purepython)

...

  22         129 LOAD_FAST             5 (zi)

             132 LOAD_FAST             5 (zi)

             135 BINARY_MULTIPLY

             136 LOAD_FAST             6 (qi)

             139 BINARY_ADD

             140 STORE_FAST            5 (zi)

  24         143 LOAD_GLOBAL           2 (abs)

             146 LOAD_FAST             5 (zi)

             149 CALL_FUNCTION         1

             152 LOAD_CONST            6 (2.0)

             155 COMPARE_OP            4 (>)

             158 POP_JUMP_IF_FALSE   123

...

You can see that we don’t have to keep loading z and i, so we execute fewer instructions (so things run faster).