Accelerate Your Python* Code through Profiling, Tuning, and Compilation Part 2: Numba*

Accelerate Your Python* Code through Profiling, Tuning, and Compilation Part 2: Numba*

• 1.
  © 2016 ContinuumAnalytics - Confidential & Proprietary© 2016 Continuum Analytics - Confidential & Proprietary Numba: A Python Compiler Stan Seibert Continuum Analytics 2016-11-12
• 2.
  © 2016 ContinuumAnalytics - Confidential & Proprietary 2 Numba: A JIT Compiler for Python Functions ▪ An open-source, function-at-a-time compiler library for Python ▪ Compiler toolbox for different targets and execution models: – single-threaded CPU, multi-threaded CPU, GPU – regular functions, “universal functions” (array functions), GPU kernels ▪ Speedup: 2x (compared to basic NumPy code) to 200x (compared to pure Python) ▪ Combine ease of writing Python with speeds approaching FORTRAN ▪ Empowers data scientists who make tools for themselves and other data scientists
• 3.
  © 2016 ContinuumAnalytics - Confidential & Proprietary 3 How does Numba work? Python Function (bytecode) Bytecode Analysis Functions Arguments Numba IR Machine Code Execute! Type Inference LLVM/NVVM JIT LLVM IR Lowering Rewrite IR Cache @jit def do_math(a, b): … >>> do_math(x, y)
• 4.
  © 2016 ContinuumAnalytics - Confidential & Proprietary 4 Supported Platforms and Hardware OS HW SW Windows (7 and later) 32 and 64-bit x86 CPUs Python 2 and 3 OS X (10.9 and later) CUDA & HSA Capable GPUs NumPy 1.7 through 1.11 Linux (RHEL 5 and later) Experimental support for ARM, Xeon Phi, AMD Fiji GPUs
• 5.
  © 2016 ContinuumAnalytics - Confidential & Proprietary 5 Tutorial Acknowledgements ▪ These Numba tutorial materials are adapted from the Numba Tutorial at SciPy 2016 by Gil Forsyth and Lorena Barba ▪ I’ve made some adjustments and additions, and also had to skip quite a bit of material for time. ▪ Check out https://github.com/barbagroup/numba_tutorial_scipy2016 for more details.
• 6.
  © 2016 ContinuumAnalytics - Confidential & Proprietary 6 Notebook 1: Numba Basics
• 7.
  © 2016 ContinuumAnalytics - Confidential & Proprietary 7 Notebook 2: How Numba Works
• 8.
  © 2016 ContinuumAnalytics - Confidential & Proprietary 8 Ex01: Intro to JIT
• 9.
  © 2016 ContinuumAnalytics - Confidential & Proprietary 9 That’s it? ▪ Mostly, yes. ▪ The Secret of Numba is: ▪ If it doesn’t need to be fast, leave it alone. (See the profiler section of this tutorial.) ▪ Stick to the well-worn path: Numba works best on loop-heavy numerical algorithms. ▪ Choose the right data structures: Numba works best on NumPy arrays and scalars.
• 10.
  © 2016 ContinuumAnalytics - Confidential & Proprietary 10 Ex02: Direct Summation
• 11.
  © 2016 ContinuumAnalytics - Confidential & Proprietary 11 There is more, though. ▪ Numba can compile other kinds of functions: ▪ Universal function (ufuncs) apply a scalar function to elements of the input arrays according to the broadcast rules: numpy.add([1, 2, 3], 1) == [2, 3, 4] numpy.add([1, 2, 3], [10, 20, 30]) == [11, 12, 13]
• 12.
  © 2016 ContinuumAnalytics - Confidential & Proprietary 12 Notebook 3: Ufuncs
• 13.
  © 2016 ContinuumAnalytics - Confidential & Proprietary 13 More Advanced Topics ▪ Generalized ufuncs: ▪ Instead of broadcasting all dimensions into a scalar function, you can control how input dimensions are broadcast. ▪ Example: Writing a norm() function ▪ http://numba.pydata.org/numba-doc/0.29.0/user/vectorize.html#the-guvectorize-decorator ▪ Calling external code: ▪ Numba can call C code that has been wrapped with ctypes or CFFI ▪ http://numba.pydata.org/numba-doc/0.29.0/reference/pysupported.html#ctypes ▪ http://numba.pydata.org/numba-doc/0.29.0/reference/pysupported.html#cffi ▪ Ahead of time compilation: ▪ http://numba.pydata.org/numba-doc/dev/user/pycc.html