A Compiler for Parallel Exeuction of Numerical Python Programs on ...

More documents

Recommendations

Info

$Formatting Instructions for Authors Using LaTeX - the Department of ...$

1 import numpy2 a r r = numpy . z e r o s ( ( 1 0 0 , 3 0 0 ) )3 v1 = a r r [ 5 : 9 7 : 3 , 3 : 3 0 0 : 2 ]4 a r r [ 5 , 3 ] = 1 . 05 print v1 [ 0 , 0 ]6 v1 [ 1 , 2 ] = 2 . 07 print a r r [ 8 , 7 ]8 print a r r . s t r i d e s9 print v1 . s t r i d e s10 v2 = a r r . swapaxes ( 0 , 1 )11 print v2 . s t r i d e s12 v3 = a r r . reshape (100∗300)13 print v3 [300∗5+3]14 v4 = v3 [ 3 : 1 0 0 ∗ 3 0 0 : 5 ]15 print v4 [ 3 0 0 ]Figure 3.3: Examples <strong>of</strong> array slicing in <strong>Python</strong>.The strides field can be changed directly without going through slicing or method callssuch as reshape. The strides <strong>of</strong> a NumPy array can also be changed through the NumPy CAPI.3.3 <strong>Python</strong>-C API<strong>Python</strong> combines the development flexibility <strong>of</strong> a dynamically-typed language with an API<strong>for</strong> efficient C implementation <strong>of</strong> per<strong>for</strong>mance critical portions — functions and classes —<strong>of</strong> an application. This <strong>Python</strong>-C API provides many utility functions <strong>for</strong> passing databack and <strong>for</strong>th between C and <strong>Python</strong>. <strong>Python</strong> exposes all <strong>Python</strong> objects as a PyObjectdatatype in C. The API provides various convenience functions and macros to convertbetween PyObject and C native types such as int, long and structures. Various other Cdatatypes are also provided <strong>for</strong> <strong>Python</strong> types such as list and tuple. However, the base type<strong>of</strong> all these types is PyObject. NumPy’s C-API exposes NumPy arrays as PyArrayObjectdatatype in C and allows direct manipulation <strong>of</strong> all fields <strong>of</strong> the NumPy object.For implementing functions, the <strong>Python</strong>-C API requires that any function that is to beused as a <strong>Python</strong> function must have a predefined type signature and must pass and returnpointers to PyObjects because the <strong>Python</strong> interpreter only knows about PyObjects.<strong>Python</strong> modules implemented in C utilizing the <strong>Python</strong>-C API are called <strong>Python</strong> extensionmodules. To write an extension module foo in C, a programmer first writes the code<strong>for</strong> the functions to be implemented in C in one or more C files. The programmer thenwrites a special module initialization function, which must be named initfoo if the desiredmodule name is foo. The C code is then compiled into a single dynamically loadable library(DLL on windows or shared object files on linux) called foo.so on Linux or foo.dll on22
Windows. The module initialization function passes a symbol table mapping C strings toC pointers to functions or variables <strong>of</strong> suitable type to be exported to <strong>Python</strong>. When the<strong>Python</strong> interpreter encounters an import foo statement, it first looks <strong>for</strong> a file called foo.so.If foo.so is found, then the interpreter looks <strong>for</strong> the function initfoo to initialize the module.If the initfoo function is not found, then <strong>Python</strong> throws an exception. If the file foo.so isnot found, then <strong>Python</strong> looks <strong>for</strong> a <strong>Python</strong> source file called foo.py.For details, see <strong>Python</strong>’s <strong>of</strong>ficial C-API documentation[6].3.4 Programmer’s workflowThe typical development <strong>of</strong> applications containing per<strong>for</strong>mance-critical, numerical-intensive,sections <strong>of</strong> code in <strong>Python</strong> consists <strong>of</strong> writing a prototype in <strong>Python</strong>, pr<strong>of</strong>iling the code toidentify per<strong>for</strong>mance-critical sections, and then re-writing these sections in a compiled languagesuch as C or C++. The programmer also writes glue code in C/C++ using the<strong>Python</strong>-C API. Figure 3.4(a) presents a block diagram that illustrates this state-<strong>of</strong>-theartdevelopment process and emphasizes the code that has to be written by a developer.Rewriting the code in C/C++ is a tedious and error-prone process that reduces developer’sproductivity and results in a code base that is less maintainable. If a GPU version <strong>of</strong> thecode is also required, then the amount <strong>of</strong> code to be written by the programmer, and thecomplexity <strong>of</strong> the code base, is further increased.This thesis presents an alternative <strong>for</strong> the development <strong>of</strong> such applications in <strong>Python</strong>,as illustrated in Figure 3.4(b). It introduces a new compiler system <strong>for</strong> automatic generation<strong>of</strong> C++ code <strong>for</strong> the per<strong>for</strong>mance-critical <strong>Python</strong> sections. The programmer is not requiredto rewrite the code in C++. Instead, the programmer annotates the <strong>Python</strong> code with typedeclarations and parallel loop annotations.The compiler framework that supports this new programming model is described indetail in chapter 4. The compiler automatically generates the C++ code as well as therequired glue code <strong>for</strong> the annotated code. The programmer then invokes a standard C++compiler to compile the code into a DLL. This DLL is a drop-in replacement <strong>for</strong> the original<strong>Python</strong> module. To generate code <strong>for</strong> multi-core CPUs, the programmer only needs to addparallel-loop annotations, where possible, to the code be<strong>for</strong>e compilation. To take advantage<strong>of</strong> GPU acceleration, the programmer simply annotates certain parallel loops to be executedon the GPU and the compiler handles everything else.The implemented compiler can only deal with a subset <strong>of</strong> <strong>Python</strong> and requires typeannotation. However, since only a small portion <strong>of</strong> the application needs to be compiled toC++, only a small portion <strong>of</strong> the <strong>Python</strong> code needs to con<strong>for</strong>m to these restrictions. Therest <strong>of</strong> the application, which does not need to be compiled, remains as is and requires nochanges.23
Page 3 and 4: Examining CommitteeJose Nelson Amar
Page 5 and 6: AbstractModern Graphics Processing
Page 8 and 9: 5.3.3 Multidimensional RCSLMADs wit
Page 10 and 11: List of Figures2.1
Page 12 and 13: RAMRISCSIMDSPSPMDSSETEXTPUVDVLIWbDE
Page 14 and 15: much lower than equivalent C progra
Page 16 and 17: Chapter 2AMD RV770 architecture <st
Page 18 and 19: Figure 2.1: RV770 Block Diagram. So
Page 20 and 21: 2.2 AMD CAL Programming modelVariou
Page 22 and 23: a context. The predefined names are
Page 24 and 25: 2.3 Hardware execution model2.3.1 P
Page 26 and 27: on the texture unit to bring the da
Page 28 and 29: float4 temp1 = i1[i][threadId.y/4];
Page 30: Chapter 3Python an
Page 33: Unlike C, NumPy does not impose any
Page 37 and 38: 3.5.1 Type annotationsType declarat
Page 39 and 40: UnPython generates
Page 41 and 42: Python source code
Page 43 and 44: and optimizations detailed later in
Page 45 and 46: Chapter 5Array access analysisArray
Page 47 and 48: LMADs can represent a wide variety
Page 49 and 50: Figure 5.2: Visualization o
Page 51 and 52: Algorithm 1 solves the data transfe
Page 53 and 54: 0 ≤ j < 5 (5.26)In the example, t
Page 55 and 56: case, assume we had a C array start
Page 57 and 58: solve for unknown
Page 59 and 60: approach, the number of</st
Page 61 and 62: Chapter 6Loop transfor</str
Page 63 and 64: Algorithm 6 perfor
Page 65 and 66: with suitable GPU code annotations.
Page 67 and 68: Table 7.3: Execution time f
Page 69 and 70: time and the execution time on the
Page 71 and 72: suite. The serial version o
Page 73 and 74: However Shedskin does not support e
Page 75 and 76: thread. Device functions are writte
Page 77 and 78: Chapter 9ConclusionsThis thesis int
Page 79: [14] Francois Labonte, Peter Mattso

A Compiler for Parallel Exeuction of Numerical Python Programs on ...

Create successful ePaper yourself

Delete template?

Save as template?