reduced memory footprint (substantially) by switching to 32bit precision

This commit is contained in:
Philip Eisenlohr 2016-12-08 23:34:43 -05:00
parent 96669b0ebd
commit d67be0e3f3
1 changed files with 10 additions and 9 deletions

View File

@ -51,7 +51,7 @@ parser.set_defaults(d = 1,
options.immutable = map(int,options.immutable)
getInterfaceEnergy = lambda A,B: (A*B != 0)*(A != B)*1.0 # 1.0 if A & B are distinct & nonzero, 0.0 otherwise
getInterfaceEnergy = lambda A,B: np.float32((A*B != 0)*(A != B)*1.0) # 1.0 if A & B are distinct & nonzero, 0.0 otherwise
struc = ndimage.generate_binary_structure(3,1) # 3D von Neumann neighborhood
# --- loop over input files -----------------------------------------------------------------------
@ -99,14 +99,13 @@ for name in filenames:
X,Y,Z = np.mgrid[0:grid[0],0:grid[1],0:grid[2]]
# Calculates gaussian weights for simulating 3d diffusion
gauss = np.exp(-(X*X + Y*Y + Z*Z)/(2.0*options.d*options.d)) \
/math.pow(2.0*np.pi*options.d*options.d,(3.0 - np.count_nonzero(info['grid'] == 1))/2.)
gauss = np.exp(-(X*X + Y*Y + Z*Z)/(2.0*options.d*options.d),dtype=np.float32) \
/np.power(2.0*np.pi*options.d*options.d,(3.0 - np.count_nonzero(info['grid'] == 1))/2.,dtype=np.float32)
gauss[:,:,:grid[2]/2:-1] = gauss[:,:,1:(grid[2]+1)/2] # trying to cope with uneven (odd) grid size
gauss[:,:grid[1]/2:-1,:] = gauss[:,1:(grid[1]+1)/2,:]
gauss[:grid[0]/2:-1,:,:] = gauss[1:(grid[0]+1)/2,:,:]
gauss = np.fft.rfftn(gauss)
gauss = np.fft.rfftn(gauss).astype(np.complex64)
for smoothIter in range(options.N):
@ -118,7 +117,7 @@ for name in filenames:
index[1],
index[2]]
interfaceEnergy = np.zeros(microstructure.shape)
interfaceEnergy = np.zeros(microstructure.shape,dtype=np.float32)
for i in (-1,0,1):
for j in (-1,0,1):
for k in (-1,0,1):
@ -141,11 +140,12 @@ for name in filenames:
periodic_bulkEnergy = periodic_interfaceEnergy[index[0],
index[1],
index[2]].reshape(2*grid) # fill bulk with energy of nearest interface
if options.ndimage:
periodic_diffusedEnergy = ndimage.gaussian_filter(
np.where(ndimage.morphology.binary_dilation(periodic_interfaceEnergy > 0.,
structure = struc,
iterations = int(round(options.d*2.)), # fat boundary
iterations = int(round(options.d*2.))-1, # fat boundary
),
periodic_bulkEnergy, # ...and zero everywhere else
0.),
@ -159,7 +159,8 @@ for name in filenames:
periodic_bulkEnergy[grid[0]/2:-grid[0]/2, # retain filled energy on fat boundary...
grid[1]/2:-grid[1]/2,
grid[2]/2:-grid[2]/2], # ...and zero everywhere else
0.))*gauss)
0.)).astype(np.complex64) *
gauss).astype(np.float32)
periodic_diffusedEnergy = np.tile(diffusedEnergy,(3,3,3))[grid[0]/2:-grid[0]/2,
grid[1]/2:-grid[1]/2,
@ -181,7 +182,7 @@ for name in filenames:
grid[1]/2:-grid[1]/2,
grid[2]/2:-grid[2]/2] # extent grains into interface region
immutable = np.zeros(microstructure.shape, dtype=bool)
immutable = np.zeros(microstructure.shape, dtype=np.bool)
# find locations where immutable microstructures have been (or are now)
for micro in options.immutable:
immutable += microstructure_original == micro