add write Assembler code

SSE/saxpy_sse_origin.py

@@ -8,16 +8,34 @@ X = numpy.random.rand(sizeX).astype(numpy.float32)
 Y = numpy.empty(sizeX).astype(numpy.float32)
 
 
-def BenchmarkCode(name, code,X,Y):
+
+def get_assembly(code):
+	return inline.cxx2asm(code, compiler_extra_args=['-msse2','-fopenmp', '-lstdc++']) 
+
+
+def BenchmarkCode(name, code, X, Y, SSE=True, OMP=True):
 	# init
 	Y[:]=0.5
 	X[:]=1.0
 
 	# compile the code
-	lib=inline.cxx(code, compiler_extra_args=['-msse2','-fopenmp'], link_extra_args= ['-msse2','-fopenmp']) 
+	compiler_extra_args = ['-g0','-lstdc++']
+	link_extra_args = ['-lstdc++']
+	if SSE :
+		compiler_extra_args+=['-msse2']
+		link_extra_args += ['-msse2']
+	if OMP:
+		compiler_extra_args+=['-fopenmp']
+		link_extra_args += ['-fopenmp']
+	lib=inline.cxx(code, compiler_extra_args= compiler_extra_args, link_extra_args= link_extra_args) 
 	p_float= numpy.ctypeslib.ndpointer(dtype=numpy.float32) 	
 	lib.compute.argtypes = [ctypes.c_int, ctypes.c_int, p_float, p_float] 
 
+	# print assembler code 
+	asm_code=inline.cxx2asm(code, compiler_extra_args= compiler_extra_args) 
+	with open(name+'.s','w') as f:
+		f.write(asm_code)
+
 	# start chronometer
 	start_time = time()
 	# run the code
@@ -53,7 +71,7 @@ void compute(int numberIterations, int sizeX, float *X, float *Y )
 }
 """
 
-referenceTime=BenchmarkCode('Reference', referenceCode,X,Y)
+referenceTime=BenchmarkCode('Reference', referenceCode, X, Y)
 
 SSECode="""
 #line 59 "saxpy.py" // helpful for debug