Python with HighPerformance

最近在讀 high performance python, 底下就隨手記錄一下吧...XD

• test env
• MacBook 2.0GHz with 4GB RAM. The GPU is a 9400M
• 
  
 
• pypy(JIT) > pure python (~6times)
• cython(move math(complex) part to c) > pure python (1.5~30times)
• numpy(matrix vector) > pure python(30times) 
• shedskin(python 2 c++ compile) > pure_python(30times)

• PyCUDA(real hardware multi cores(DSP)) > numpy(CPU 50times)  

• using cprofile and dis to trace back where is the performance bottleneck 
• python -cprofile  
• @profie decorader in each proc call 
• python dis.dis(function call) 
• trace byte code linse, macro blocks, where to improve 
• serialize code
• data store in cache > store in memory   
• RunSnakeRun
• GUI tool(cprofile results)

• Multiprocessing: parallel process to multi cores(CPUS)
• pypy: JIT(just in time) compiler, llvm, byte code optimization, rpython(.Net), cython 
• numpy: N dimension array vector to one dimension vector array(serialize memory access)
• cpython: pre-compile python to having object type(int,char...) 
• PyCUDA: hardware speed up (DSPs)

tips

• a = {} >  a = dict()
• "".join(st) > st = "a" + "b" + "c"
• [i.upper() for i in test] > for i in test: arr.append(i.upper())
• def test() > global test ...
• init dict > without init dict
• pre load(import modules in header) > current load(import modules in middle)
• reduce the call back counts 
• xrange > range
• remap func without recursive loops
• hash(heap map) > loop searching  

ref:
EuroPython2011_HighPerformanceComputing
performance tips
timecomplex