""" Cast Copy Tranpose is used in numpy LinearAlgebra.py to convert C ordered arrays to Fortran order arrays before calling Fortran functions. A couple of C implementations are provided here that show modest speed improvements. One is an "inplace" transpose that does an in memory transpose of an arrays elements. This is the fastest approach and is beneficial if you don't need to keep the original array. """ # C:\home\ej\wrk\scipy\compiler\examples>python cast_copy_transpose.py # Cast/Copy/Transposing (150,150)array 1 times # speed in python: 0.870999932289 # speed in c: 0.25 # speed up: 3.48 # inplace transpose c: 0.129999995232 # speed up: 6.70 from __future__ import absolute_import, print_function import numpy from numpy import * import sys sys.path.insert(0,'..') import scipy.weave.inline_tools as inline_tools import scipy.weave.c_spec as c_spec from scipy.weave.converters import blitz as cblitz def _cast_copy_transpose(type,a_2d): assert(len(shape(a_2d)) == 2) new_array = zeros(shape(a_2d),type) code = """ for(int i = 0; i < Na_2d[0]; i++) for(int j = 0; j < Na_2d[1]; j++) new_array(i,j) = a_2d(j,i); """ inline_tools.inline(code,['new_array','a_2d'], type_converters=cblitz, compiler='gcc', verbose=1) return new_array def _cast_copy_transpose2(type,a_2d): assert(len(shape(a_2d)) == 2) new_array = zeros(shape(a_2d),type) code = """ const int I = Na_2d[0]; const int J = Na_2d[1]; for(int i = 0; i < I; i++) { int new_off = i*J; int old_off = i; for(int j = 0; j < J; j++) { new_array[new_off++] = a_2d[old_off]; old_off += I; } } """ inline_tools.inline(code,['new_array','a_2d'],compiler='gcc',verbose=1) return new_array def _inplace_transpose(a_2d): assert(len(shape(a_2d)) == 2) numeric_type = c_spec.num_to_c_types[a_2d.dtype.char] code = """ %s temp; for(int i = 0; i < Na_2d[0]; i++) for(int j = 0; j < Na_2d[1]; j++) { temp = a_2d(i,j); a_2d(i,j) = a_2d(j,i); a_2d(j,i) = temp; } """ % numeric_type inline_tools.inline(code,['a_2d'], type_converters=cblitz, compiler='gcc', extra_compile_args=['-funroll-all-loops'], verbose=2) return a_2d #assert(len(shape(a_2d)) == 2) #type = a_2d.typecode() #new_array = zeros(shape(a_2d),type) ##trans_a_2d = transpose(a_2d) #numeric_type = c_spec.num_to_c_types[type] #code = """ # for(int i = 0; i < Na_2d[0]; i++) # for(int j = 0; j < Na_2d[1]; j++) # new_array(i,j) = (%s) a_2d(j,i); # """ % numeric_type #inline_tools.inline(code,['new_array','a_2d'], # type_converters = cblitz, # compiler='gcc', # verbose = 1) #return new_array def cast_copy_transpose(type,*arrays): results = [] for a in arrays: results.append(_cast_copy_transpose(type,a)) if len(results) == 1: return results[0] else: return results def cast_copy_transpose2(type,*arrays): results = [] for a in arrays: results.append(_cast_copy_transpose2(type,a)) if len(results) == 1: return results[0] else: return results def inplace_cast_copy_transpose(*arrays): results = [] for a in arrays: results.append(_inplace_transpose(a)) if len(results) == 1: return results[0] else: return results def _castCopyAndTranspose(type, *arrays): cast_arrays = () import copy for a in arrays: if a.dtype == numpy.dtype(type): cast_arrays = cast_arrays + (copy.copy(numpy.transpose(a)),) else: cast_arrays = cast_arrays + (copy.copy( numpy.transpose(a).astype(type)),) if len(cast_arrays) == 1: return cast_arrays[0] else: return cast_arrays import time def compare(m,n): a = ones((n,n),float64) type = float32 print('Cast/Copy/Transposing (%d,%d)array %d times' % (n,n,m)) t1 = time.time() for i in range(m): for i in range(n): b = _castCopyAndTranspose(type,a) t2 = time.time() py = (t2-t1) print(' speed in python:', (t2 - t1)/m) # load into cache b = cast_copy_transpose(type,a) t1 = time.time() for i in range(m): for i in range(n): b = cast_copy_transpose(type,a) t2 = time.time() print(' speed in c (blitz):',(t2 - t1) / m) print(' speed up (blitz): %3.2f' % (py/(t2-t1))) # load into cache b = cast_copy_transpose2(type,a) t1 = time.time() for i in range(m): for i in range(n): b = cast_copy_transpose2(type,a) t2 = time.time() print(' speed in c (pointers):',(t2 - t1) / m) print(' speed up (pointers): %3.2f' % (py/(t2-t1))) # inplace tranpose b = _inplace_transpose(a) t1 = time.time() for i in range(m): for i in range(n): b = _inplace_transpose(a) t2 = time.time() print(' inplace transpose c:',(t2 - t1) / m) print(' speed up: %3.2f' % (py/(t2-t1))) if __name__ == "__main__": m,n = 1,500 compare(m,n)