def __init__(self, ecut, gd, dtype=None, kd=None,
fftwflags=fftw.ESTIMATE):
assert gd.pbc_c.all()
assert gd.comm.size == 1
self.ecut = ecut
self.gd = gd
N_c = gd.N_c
self.comm = gd.comm
assert ((gd.h_cv**2).sum(1) <= 0.5 * pi**2 / ecut).all()
if dtype is None:
if kd is None or kd.gamma:
dtype = float
else:
dtype = complex
self.dtype = dtype
if dtype == float:
Nr_c = N_c.copy()
Nr_c[2] = N_c[2] // 2 + 1
i_Qc = np.indices(Nr_c).transpose((1, 2, 3, 0))
i_Qc[..., :2] += N_c[:2] // 2
i_Qc[..., :2] %= N_c[:2]
i_Qc[..., :2] -= N_c[:2] // 2
self.tmp_Q = fftw.empty(Nr_c, complex)
self.tmp_R = self.tmp_Q.view(float)[:, :, :N_c[2]]
else:
i_Qc = np.indices(N_c).transpose((1, 2, 3, 0))
i_Qc += N_c // 2
i_Qc %= N_c
i_Qc -= N_c // 2
self.tmp_Q = fftw.empty(N_c, complex)
self.tmp_R = self.tmp_Q
self.nbytes = self.tmp_R.nbytes
self.fftplan = fftw.FFTPlan(self.tmp_R, self.tmp_Q, -1, fftwflags)
self.ifftplan = fftw.FFTPlan(self.tmp_Q, self.tmp_R, 1, fftwflags)
# Calculate reciprocal lattice vectors:
B_cv = 2.0 * pi * gd.icell_cv
i_Qc.shape = (-1, 3)
self.G_Qv = np.dot(i_Qc, B_cv)
self.nbytes += self.G_Qv.nbytes
self.kd = kd
if kd is None:
self.K_qv = np.zeros((1, 3))
else:
self.K_qv = np.dot(kd.ibzk_qc, B_cv)
# Map from vectors inside sphere to fft grid:
self.Q_qG = []
self.G2_qG = []
Q_Q = np.arange(len(i_Qc), dtype=np.int32)
self.ngmin = 100000000
self.ngmax = 0
for q, K_v in enumerate(self.K_qv):
G2_Q = ((self.G_Qv + K_v)**2).sum(axis=1)
mask_Q = (G2_Q <= 2 * ecut)
if self.dtype == float:
mask_Q &= ((i_Qc[:, 2] > 0) |
(i_Qc[:, 1] > 0) |
((i_Qc[:, 0] >= 0) & (i_Qc[:, 1] == 0)))
Q_G = Q_Q[mask_Q]
self.Q_qG.append(Q_G)
self.G2_qG.append(G2_Q[Q_G])
ng = len(Q_G)
self.ngmin = min(ng, self.ngmin)
self.ngmax = max(ng, self.ngmax)
self.nbytes += Q_G.nbytes + self.G2_qG[q].nbytes
if kd is not None:
self.ngmin = kd.comm.min(self.ngmin)
self.ngmax = kd.comm.max(self.ngmax)
self.n_c = np.array([self.ngmax]) # used by hs_operators.py XXX
def __init__(self, ecut, gd, dtype=None, kd=None,
fftwflags=fftw.ESTIMATE):
assert gd.pbc_c.all()
assert gd.comm.size == 1
self.ecut = ecut
self.gd = gd
self.fftwflags = fftwflags
N_c = gd.N_c
self.comm = gd.comm
assert ((gd.h_cv**2).sum(1) <= 0.5 * pi**2 / ecut).all()
if dtype is None:
if kd is None or kd.gamma:
dtype = float
else:
dtype = complex
self.dtype = dtype
if dtype == float:
Nr_c = N_c.copy()
Nr_c[2] = N_c[2] // 2 + 1
i_Qc = np.indices(Nr_c).transpose((1, 2, 3, 0))
i_Qc[..., :2] += N_c[:2] // 2
i_Qc[..., :2] %= N_c[:2]
i_Qc[..., :2] -= N_c[:2] // 2
self.tmp_Q = fftw.empty(Nr_c, complex)
self.tmp_R = self.tmp_Q.view(float)[:, :, :N_c[2]]
else:
i_Qc = np.indices(N_c).transpose((1, 2, 3, 0))
i_Qc += N_c // 2
i_Qc %= N_c
i_Qc -= N_c // 2
self.tmp_Q = fftw.empty(N_c, complex)
self.tmp_R = self.tmp_Q
self.nbytes = self.tmp_R.nbytes
self.fftplan = fftw.FFTPlan(self.tmp_R, self.tmp_Q, -1, fftwflags)
self.ifftplan = fftw.FFTPlan(self.tmp_Q, self.tmp_R, 1, fftwflags)
# Calculate reciprocal lattice vectors:
B_cv = 2.0 * pi * gd.icell_cv
i_Qc.shape = (-1, 3)
self.G_Qv = np.dot(i_Qc, B_cv)
self.nbytes += self.G_Qv.nbytes
self.kd = kd
if kd is None:
self.K_qv = np.zeros((1, 3))
else:
self.K_qv = np.dot(kd.ibzk_qc, B_cv)
# Map from vectors inside sphere to fft grid:
self.Q_qG = []
self.G2_qG = []
Q_Q = np.arange(len(i_Qc), dtype=np.int32)
self.ngmin = 100000000
self.ngmax = 0
for q, K_v in enumerate(self.K_qv):
G2_Q = ((self.G_Qv + K_v)**2).sum(axis=1)
mask_Q = (G2_Q <= 2 * ecut)
if self.dtype == float:
mask_Q &= ((i_Qc[:, 2] > 0) |
(i_Qc[:, 1] > 0) |
((i_Qc[:, 0] >= 0) & (i_Qc[:, 1] == 0)))
Q_G = Q_Q[mask_Q]
self.Q_qG.append(Q_G)
self.G2_qG.append(G2_Q[Q_G])
ng = len(Q_G)
self.ngmin = min(ng, self.ngmin)
self.ngmax = max(ng, self.ngmax)
self.nbytes += Q_G.nbytes + self.G2_qG[q].nbytes
if kd is not None:
self.ngmin = kd.comm.min(self.ngmin)
self.ngmax = kd.comm.max(self.ngmax)
self.n_c = np.array([self.ngmax]) # used by hs_operators.py XXX
def test(Plan, flags, input, output, sign):
t0 = time.time()
plan = Plan(input, output, sign, flags)
t1 = time.time()
t = 0.0
for i in range(100):
input[:] = 1.3
t2 = time.time()
plan.execute()
t3 = time.time()
t += t3 - t2
return t1 - t0, t / 100
if __name__ == '__main__':
a1 = fftw.empty((32, 28, 128), complex)
a2 = fftw.empty((32, 28, 128), complex)
b = fftw.empty((32, 28, 65), complex)
c1 = b.view(dtype=float)[:, :, :128]
c2 = fftw.empty((32, 28, 64), complex).view(dtype=float)
for input, output, sign in [
(a1, a1, -1),
(a1, a2, -1),
(b, c1, 1),
(b, c2, 1),
(c1, b, -1),
(c2, b, -1)]:
for Plan, flags in [(fftw.NumpyFFTPlan, 117),
(fftw.FFTWPlan, fftw.ESTIMATE),
(fftw.FFTWPlan, fftw.MEASURE),
(fftw.FFTWPlan, fftw.PATIENT),
def test(Plan, flags, input, output, sign):
t0 = time.time()
plan = Plan(input, output, sign, flags)
t1 = time.time()
t = 0.0
for i in range(100):
input[:] = 1.3
t2 = time.time()
plan.execute()
t3 = time.time()
t += t3 - t2
return t1 - t0, t / 100
if __name__ == '__main__':
a1 = fftw.empty((32, 28, 128), complex)
a2 = fftw.empty((32, 28, 128), complex)
b = fftw.empty((32, 28, 65), complex)
c1 = b.view(dtype=float)[:, :, :128]
c2 = fftw.empty((32, 28, 64), complex).view(dtype=float)
for input, output, sign in [(a1, a1, -1), (a1, a2, -1), (b, c1, 1),
(b, c2, 1), (c1, b, -1), (c2, b, -1)]:
for Plan, flags in [(fftw.NumpyFFTPlan, 117),
(fftw.FFTWPlan, fftw.ESTIMATE),
(fftw.FFTWPlan, fftw.MEASURE),
(fftw.FFTWPlan, fftw.PATIENT),
(fftw.FFTWPlan, fftw.EXHAUSTIVE)]:
tplan, tfft = test(Plan, flags, input, output, sign)
print(('%-12s %3d %10.6f %10.6f' %
(Plan.__name__, flags, tplan, tfft)))