Python Scientific and Numerical Computing (fast with pypy?)

Five minute lightning talk given at the pymntos (www.pymntos.org) monthly meeting. This talk reviews: if pypy is useful for numerical and scientific computing.

Transcript

1. Python Scientific and Numerical Computing Is it fast with pypy?
2. Context • Hardware design • HDL development and simulation • synthesizable descriptions and verification environments • Signal / data analysis (lab data, field data) • Before Python • HDL (Verilog/SV and VHDL), Matlab, and C/C++ • Ad hoc growth into a Python user
3. pypy • All the rave - PyCon 2012 • JIT compiler for Python • impressive results (benchmarks) • www.pypy.org
4. Is pypy the run-time environment for scientific / numerical computing?
5. ... for most people NO
6. Then why are we talking about it?
7. cpython gfortran romberg-integration* 25x faster 2x faster akima-interpolation* 20x faster 5x slower float 7x faster meteor-contest 2x faster chaos 25x faster ai 4x faster lfsr** 19.5x faster randgen** 12.3x faster longdiv** 10.9x faster findmax** 7.8x faster image processing simulation 2x faster -- FFT 8.6x faster -- above all pure python implementations (except the referenced fortran) pypy compared to cpython and fortran * pypy mailing-list ** HDL (hardware description language) simulation Because pypy is fast (simple fast)
8. Nfft, Nloops 4096, 1024 262144, 256 avg increase cpython pypy numpy* 1.000 1.000 0.126 0.136 7.7x 0.003 0.003 259.2x FFT Comparison ... wait there is more Not really a fair comparison ... let’s look a little closer ... the FFT available with the numpy package much faster def _fftr(x): n = len(x) if (n == 1): return x # even and odd parts X,Y = (x[::2], x[1::2],) Xr,Yr = (_fftr(X), _fftr(Y),) w = [(cos(2*pi*k/float(n)) + 1.j*sin(2*pi*k/float(n))) for k in range(n)] m = n/2; b = [Xr[(k%m)]+w[k]*Yr[(k%m)] for k in range(n)] return b Simple recursive FFT (python only) * The numpy FFT, optimized FFTW code @ http://www.dsprelated.com/showcode/16.php or www.bitbucket.org/cfelton/examples/src/tip/rfft
9. FFT Comparison ... wait there is more The FFT is used in a simulation but more than just the FFT in the simulation. What is the overall performance? This simulation uses the FFT but the FFT speed up alone was not enough to affect the overall performance! Nfft, Nloops 4096, 1024 4096, 256 avg increase cpython pypy numpy 1.000 1.000 0.010 0.073 13.8x 1.005 0.987 1.0x 0 0.275 0.550 0.825 1.100 4096,1024 4096,256 4069,32 4096,8 128,128 Simulation 1 cpython pypy numpy
10. Doesn’t that mean pypy would be good for scientific / numerical?
11. yes, fast and Python programming == nice!
12. NO, because scientist/engineers need (require) a large and stable collection of libraries
13. it is great for hardware development (EDA) See MyHDL digital circuit simulations: www.myhdl.org/performance
14. Filling the Library Gap • rewrite of sci/num libraries in Python (golden models) • new C and fortran interfaces to existing packages (cffi and ???) • key portions written in rpython (this is what the pypy team is currently doing with numpypy)
15. Transmitting Mixed Signals • pypy current state • No - analysis, classical signals and systems, etc • Yes - simulation: HDL simulation and Simulink like simulations • projects of interest (where applicable) • gnuradio for fpga (krypton) • large correlators for radio astronomy
16. • python : 2.7.5 • pypy : 2.0.2 • numpy : 1.6.2