What is PyCUDA?
Why Python PyCUDA
PyCUDA: short tutorial
Glib Ivashkevych
A.I. Akhiezer Institute of Theoretical Physics, NSC KIPT
Kharkov, Ukraine
Glib Ivashkevych PyCUDA: short tutorial
What is PyCUDA? Why Python PyCUDA
What is PyCUDA?
A simple way to compute on GPU from Python:
I complete access to CUDA features
I automatic resources management
I error checking and reporting
I high level abstractions: GPUArray, for example
I integration with NumPy
I documentation
Glib Ivashkevych PyCUDA: short tutorial
What is PyCUDA? Why Python PyCUDA
But before PyCUDA: why Python at all?
I general purpose
I interpreted
I simple to learn and use
I extensible and embeddable: Python C API
I science oriented too: NumPy, SciPy, SymPy, mpi4py,
MatPlotLib
I very well documented
NumPy
I flexible and effective arrays creation and manipulation
I FFT’s, signal processing, effective I/O and more
SciPy
I ODE’s, special functions, linear algebra, root finding and more
Glib Ivashkevych PyCUDA: short tutorial
What is PyCUDA? Why Python PyCUDA
Python goodies for scientific computing: NumPy arrays
Example:
>>> import numpy a s np
>>> a = np . a r a n g e ( 1 . , 2 . , 0 . 1 )
>>> a
array ([1. , 1.1 , 1.2 , 1.3 , 1.4 , 1.5 , 1.6 , 1.7 , 1.8 , 1.9])
>>> a [ 0 : 3 ]
array ([1. , 1.1 , 1.2])
>>> a [ : : 3 ]
array ([1. , 1.3 , 1.6 , 1.9])
>>> a [ a > 1 . 4 5 ]
array ([1.5 , 1.6 , 1.7 , 1.8 , 1.9])
Glib Ivashkevych PyCUDA: short tutorial
What is PyCUDA? Why Python PyCUDA
Python goodies for scientific computing: SciPy
How could we use C/C++ from Python?
I Python C API – a hard way
I SWIG, Boost::Python – simplier, but not simple enough
I scipy.weave.inline – just pass your C code as a string
Example:
>>> import numpy a s np
>>> from s c i p y . weave import i n l i n e
>>> b = np . n d a r r a y ( s h a p e =(10) , d t y p e=f l o a t ) ; b. f i l l (1.)
>>> b
array ([ 1. , 1. , 1. , 1. , 1. , 1. , 1. , 1. , 1. , 1.])
>>> inum = b . s i z e
>>> C c o d e = ’ f o r ( i n t i =0; i <inum ; i ++) b [ i ] ∗= 2 . ; ’
>>> i n l i n e ( C code , [ ’ b ’ , ’ inum ’ ] )
>>> b
array ([ 2. , 2. , 2. , 2. , 2. , 2. , 2. , 2. , 2. , 2.])
Glib Ivashkevych PyCUDA: short tutorial
What is PyCUDA? Why Python PyCUDA
Why is it so simple to use C/C++ from Python?
Python has C under the hood
I Python objects are in fact C structs (roughly speaking)
I ... and NumPy arrays too
Look under the hood: NumPy C API
typedef struct PyArrayObject {
PyObject HEAD
char ∗ data ;
i n t nd ;
npy intp ∗ dimensions ;
npy intp ∗ s t r i d e s ;
PyObject ∗ base ;
PyArray Descr ∗ descr ;
int flags ;
PyObject ∗ w e a k r e f l i s t ;
} PyArrayObject ;
Glib Ivashkevych PyCUDA: short tutorial
What is PyCUDA? Why Python PyCUDA
The same simplicity with CUDA in Python? Yes!
Image courtesy: Andreas Klöckner.
Glib Ivashkevych PyCUDA: short tutorial
What is PyCUDA? Why Python PyCUDA
Concept behind PyCUDA: code is
not a compile–time constant → metaprogramming
Image courtesy: Andreas Klöckner.
Glib Ivashkevych PyCUDA: short tutorial
What is PyCUDA? Why Python PyCUDA
Metaprogramming = runtime generation of code from
templates or code snippets
Example 1: simple metaprogramming (simple mprog.py)
...
from j i n j a 2 import Template
...
m o d u l e t p l = Template ( ”””
global void
k e r n e l ( { { t y p e }} ∗a , {{ t y p e }} ∗b , {{ t y p e }} ∗ r e s )
{
i n t i d x = t h r e a d I d x . x + {{ t p e r b l k }} ∗ b l o c k I d x . x ;
r e s [ i d x ] = a [ i d x ] {{ op }} b [ i d x ] ;
}
””” )
...
m o d u l e c o d e = m o d u l e t p l . r e n d e r ( t y p e=” f l o a t ” ,
op=”+” , t p e r b l k=k )
module = pycuda . c o m p i l e r . S o u r c e M o d u l e ( m o d u l e c o d e )
k e r n e l f u n c = module . g e t f u n c t i o n ( ” k e r n e l ” )
...
Glib Ivashkevych PyCUDA: short tutorial
What is PyCUDA? Why Python PyCUDA
PyCUDA abstractions: GPUArray
I you could do GPU programming that way (kinda C style) or ...
I ... make use of high level PyCUDA abstractions
Example 2: GPUArray class (add arrays.py)
import numpy a s np
import pycuda . a u t o i n i t
import pycuda . g p u a r r a y a s g a r r a y
a = np . a r a n g e ( 0 . , 1 2 8 . ∗ 1 2 8 . , 1 . , d t y p e = np . f l o a t 3 2 )
b = np . a r a n g e ( 1 2 8 . ∗ 1 2 8 . , 0 . , −1. , d t y p e = np . f l o a t 3 2 )
a gpu = g a r r a y . to gpu ( a )
b gpu = g a r r a y . to gpu ( b )
c = ( a gpu + b gpu ) . get ( )
p r i n t np . amax ( c − a − b )
Glib Ivashkevych PyCUDA: short tutorial
What is PyCUDA? Why Python PyCUDA
PyCUDA abstractions: GPUArray
Example 3: GPUArray class (double array.py)
...
a = np . n d a r r a y ( s h a p e =(128 , 1 2 8 ) , d t y p e = np . f l o a t 3 2 )
a . f i l l (1.)
a gpu = gpuarray . to gpu ( a )
b = (2 ∗ a gpu ) . get ( )
p r i n t np . amax ( b − 2∗ a )
GPUArray is handy: no need to ...
I ...allocate and free memory
I ...copy data between Host and Device
I ...write kernels for (at least) simple operations
Glib Ivashkevych PyCUDA: short tutorial
What is PyCUDA? Why Python PyCUDA
PyCUDA abstractions: ReductionKernel
Z ∞
2
Example: I = e −x dx
−∞
Example 4: ReductionKernel (gaussian integral.py)
...
G a u s s i a n I n t = RednKer ( np . f l o a t 3 2 , n e u t r a l=” 0 . ” ,
r e d u c e e x p r=” a+b” , m a p e x p r=” x ∗ y [ i ] ” ,
a r g u m e n t s=” f l o a t x , f l o a t ∗ y ” )
...
ReductionKernel for ...
I ...scalar products
I ...integrals
I ...even n-body and more
Glib Ivashkevych PyCUDA: short tutorial
What is PyCUDA? Why Python PyCUDA
And even more:
I handy GPUArray creation routines (mimics NumPy)
I pycuda.cumath & ElementwiseKernel & Prefix Sum
I FFT: PyFFT – designed to work with GPUArray’s, by Bogdan
Opanchuk
PyOpenCL
Pretty much the same in concept, but for OpenCL: platform
independent.
Glib Ivashkevych PyCUDA: short tutorial
What is PyCUDA? Why Python PyCUDA
Links:
I PyCUDA documentation:
http://documen.tician.de/pycuda/
I GTC 2010 presentations archive:
http://www.nvidia.com/object/gtc2010-presentation-
archive.html
I PASI screencasts (including 4 lectures on OpenCL &
PyOpenCL by Andreas):
http://www.bu.edu/pasi/materials/
I Scientific and numerical packages for Python:
http://wiki.python.org/moin/NumericAndScientific
Glib Ivashkevych PyCUDA: short tutorial
What is PyCUDA? Why Python PyCUDA
Thanks
To our directors for opportunity to present this tutorial
To Andreas Klöckner for PyCUDA, useful hints and images
Glib Ivashkevych PyCUDA: short tutorial
