[RESOLVED] pypy performance pants! Why?

[Hove]

I need to improve the performance of my CPython (interpreted) code. I don't want to move to Cython due to the significant changes to my code it imposes. I tried pypy as its compiles to 'C' code just in time, and is supposed to make code run very much faster as a result. But it takes 17.1s (not including the JIT compile time) for my performance loop whereas the CPython takes 7.6 seconds.

This is the version of pypy which comes as part of the standard raspbian distro. According to the pypy web site, this version should be, on average, 6 times faster, but for me it's 2.25 times slower.

Any ideas why?

I have recompiled my GPIO and RPIO modules to run with pypy. The rest of my python code is using standard python libraries.

TIA

[RaTTuS]

post your code to see what it is doing

[Hove]

post your code to see what it is doing

Code is 2000 lines long so not really postable; it's up on GitHub at PiStuffing/Quadcopter/Quadcopter.py. To summarise it though it's structured like this:

GPIO.wait_for_edge() to detect hardware interrupt
smbus.i2c to read sensor data
various bits of basic trigonometry not using numpy
RPIO to produce a PWM output signal.

GPIO and RPIO are thin python wrappers around 'C' code, both of which I've recompiled / installed with 'sudo pypy ./setup.py install'

I have to assume it's one of GPIO, smbus or RPIO that are in some way incompatible, but that's purely speculative.

Does that give you further insight?

[joan]

I wouldn't expect your code to be significantly faster, most of the heavy lifting is already being done by C, how much pure Python remains?

However I wouldn't expect it to be slower either.

[Hove]

It's motion processing code that's taking the 3.5ms in addition to the <1ms for the sensor sampling.

Other than trigonometry, there is plenty of pure python arithmetic for

• calculating angles
• merging angles from accel and gyro
• rotation matrix processing
• butterworth filters (3 thereof)
• PID processing (7 thereof )

Each is small, but together, I'd expect pypy to get this down to <<1ms which is what I need. So that fact it's gone up to 3.5ms from the 1.3ms I get from CPython is dissappointing

[Hove]

From the horse's mouth of the PyPy team: it's a miracle the PyPy works at all with the CPython C API used for GPIO / RPIO. The fact it does, and is only as slow as it is amazes them. The correct way for PyPy to call into ''C" code is to use a CFFI module: https://cffi.readthedocs.org/en/latest/overview.html will be much faster.

So there you have it!