def ao2mo_(casscf, mo):
t0 = (time.clock(), time.time())
log = logger.Logger(casscf.stdout, casscf.verbose)
# using dm=[], a hacky call to dfhf.get_jk, to generate casscf._cderi
dfhf.get_jk_(casscf, casscf.mol, [])
if log.verbose >= logger.DEBUG1:
t1 = log.timer('Generate density fitting integrals', *t0)
if hasattr(casscf._scf, '_tag_df') and casscf._scf._tag_df:
eris = _ERIS(casscf, mo)
else:
# Only approximate the orbital rotation, call the 4-center integral
# transformation. CASSCF is exact.
eris = mc_ao2mo._ERIS(casscf, mo, 'incore', level=2)
t0 = (time.clock(), time.time())
mo = numpy.asarray(mo, order='F')
nao, nmo = mo.shape
ncore = casscf.ncore
eris.j_pc = numpy.zeros((nmo,ncore))
k_cp = numpy.zeros((ncore,nmo))
fmmm = _ao2mo._fpointer('AO2MOmmm_nr_s2_iltj')
fdrv = _ao2mo.libao2mo.AO2MOnr_e2_drv
ftrans = _ao2mo._fpointer('AO2MOtranse2_nr_s2')
bufs1 = numpy.empty((dfhf.BLOCKDIM,nmo,nmo))
with df.load(casscf._cderi) as feri:
for b0, b1 in dfhf.prange(0, casscf._naoaux, dfhf.BLOCKDIM):
eri1 = numpy.asarray(feri[b0:b1], order='C')
buf = bufs1[:b1-b0]
if log.verbose >= logger.DEBUG1:
t1 = log.timer('load buf %d:%d'%(b0,b1), *t1)
fdrv(ftrans, fmmm,
buf.ctypes.data_as(ctypes.c_void_p),
eri1.ctypes.data_as(ctypes.c_void_p),
mo.ctypes.data_as(ctypes.c_void_p),
ctypes.c_int(b1-b0), ctypes.c_int(nao),
ctypes.c_int(0), ctypes.c_int(nmo),
ctypes.c_int(0), ctypes.c_int(nmo),
ctypes.c_void_p(0), ctypes.c_int(0))
if log.verbose >= logger.DEBUG1:
t1 = log.timer('transform [%d:%d]'%(b0,b1), *t1)
bufd = numpy.einsum('kii->ki', buf)
eris.j_pc += numpy.einsum('ki,kj->ij', bufd, bufd[:,:ncore])
k_cp += numpy.einsum('kij,kij->ij', buf[:,:ncore], buf[:,:ncore])
if log.verbose >= logger.DEBUG1:
t1 = log.timer('j_pc and k_pc', *t1)
eri1 = None
eris.k_pc = k_cp.T.copy()
log.timer('ao2mo density fit part', *t0)
return eris
def ao2mo(self, mo):
t0 = (time.clock(), time.time())
ncore = self.ncore
log = pyscf.lib.logger.Logger(self.stdout, self.verbose)
# using dm=[], a hacky call to dfhf.get_jk, to generate self._cderi
self.get_jk(self.mol, [])
if log.verbose >= pyscf.lib.logger.DEBUG1:
t1 = log.timer('Generate density fitting integrals', *t0)
eris = mc_ao2mo._ERIS(self, mo, 'incore', level=2)
t0 = (time.clock(), time.time())
mo = numpy.asarray(mo, order='F')
nao, nmo = mo.shape
eris.j_pc = numpy.zeros((nmo,ncore))
k_cp = numpy.zeros((ncore,nmo))
fmmm = _ao2mo._fpointer('AO2MOmmm_nr_s2_iltj')
fdrv = _ao2mo.libao2mo.AO2MOnr_e2_drv
ftrans = _ao2mo._fpointer('AO2MOtranse2_nr_s2kl')
bufs1 = numpy.empty((dfhf.BLOCKDIM,nmo,nmo))
with df.load(self._cderi) as feri:
for b0, b1 in dfhf.prange(0, self._naoaux, dfhf.BLOCKDIM):
eri1 = numpy.array(feri[b0:b1], copy=False)
buf = bufs1[:b1-b0]
if log.verbose >= pyscf.lib.logger.DEBUG1:
t1 = log.timer('load buf %d:%d'%(b0,b1), *t1)
fdrv(ftrans, fmmm,
buf.ctypes.data_as(ctypes.c_void_p),
eri1.ctypes.data_as(ctypes.c_void_p),
mo.ctypes.data_as(ctypes.c_void_p),
ctypes.c_int(b1-b0), ctypes.c_int(nao),
ctypes.c_int(0), ctypes.c_int(nmo),
ctypes.c_int(0), ctypes.c_int(nmo),
ctypes.c_void_p(0), ctypes.c_int(0))
if log.verbose >= pyscf.lib.logger.DEBUG1:
t1 = log.timer('transform [%d:%d]'%(b0,b1), *t1)
bufd = numpy.einsum('kii->ki', buf).copy()
#:eris.j_pc += numpy.einsum('ki,kj->ij', bufd, bufd[:,:ncore])
pyscf.lib.dot(bufd.T, numpy.asarray(bufd[:,:ncore],order='C'),
1, eris.j_pc, 1)
k_cp += numpy.einsum('kij,kij->ij', buf[:,:ncore], buf[:,:ncore])
if log.verbose >= pyscf.lib.logger.DEBUG1:
t1 = log.timer('j_pc and k_pc', *t1)
eri1 = None
eris.k_pc = k_cp.T.copy()
log.timer('ao2mo density fit part', *t0)
return eris
def ao2mo(self, mo_coeff):
log = logger.Logger(self.stdout, self.verbose)
# the exact integral transformation
eris = casscf_class.ao2mo(self, mo_coeff)
# using dm=[], a hacky call to dfhf.get_jk, to generate self._cderi
dfhf.get_jk_(self, self.mol, [])
# Add the approximate diagonal term for orbital hessian
t0 = (time.clock(), time.time())
mo = numpy.asarray(mo_coeff, order='F')
nao, nmo = mo.shape
ncore = self.ncore
eris.j_pc = numpy.zeros((nmo,ncore))
k_cp = numpy.zeros((ncore,nmo))
fmmm = _ao2mo._fpointer('AO2MOmmm_nr_s2_iltj')
fdrv = _ao2mo.libao2mo.AO2MOnr_e2_drv
ftrans = _ao2mo._fpointer('AO2MOtranse2_nr_s2')
bufs1 = numpy.empty((dfhf.BLOCKDIM,nmo,nmo))
with df.load(self._cderi) as feri:
for b0, b1 in dfhf.prange(0, self._naoaux, dfhf.BLOCKDIM):
eri1 = numpy.asarray(feri[b0:b1], order='C')
buf = bufs1[:b1-b0]
if log.verbose >= logger.DEBUG1:
t1 = log.timer('load buf %d:%d'%(b0,b1), *t1)
fdrv(ftrans, fmmm,
buf.ctypes.data_as(ctypes.c_void_p),
eri1.ctypes.data_as(ctypes.c_void_p),
mo.ctypes.data_as(ctypes.c_void_p),
ctypes.c_int(b1-b0), ctypes.c_int(nao),
ctypes.c_int(0), ctypes.c_int(nmo),
ctypes.c_int(0), ctypes.c_int(nmo),
ctypes.c_void_p(0), ctypes.c_int(0))
if log.verbose >= logger.DEBUG1:
t1 = log.timer('transform [%d:%d]'%(b0,b1), *t1)
bufd = numpy.einsum('kii->ki', buf)
eris.j_pc += numpy.einsum('ki,kj->ij', bufd, bufd[:,:ncore])
k_cp += numpy.einsum('kij,kij->ij', buf[:,:ncore], buf[:,:ncore])
if log.verbose >= logger.DEBUG1:
t1 = log.timer('j_pc and k_pc', *t1)
eri1 = None
eris.k_pc = k_cp.T.copy()
log.timer('ao2mo density fit part', *t0)
return eris
def ao2mo(self, mo):
ncore = self.ncore
#self._cderi = None # FIXME? leave as much memory as possible for mc_ao2mo
eris = mc_ao2mo._ERIS(self, mo, 'incore', 2)
# using dm=[], a hacky call to dfhf.get_jk, to generate self._cderi
t0 = (time.clock(), time.time())
log = pyscf.lib.logger.Logger(self.stdout, self.verbose)
self.get_jk(self.mol, [])
if log.verbose >= pyscf.lib.logger.DEBUG1:
t1 = log.timer('Generate density fitting integrals', *t0)
mo = numpy.asarray(mo, order='F')
nao, nmo = mo.shape
eris.j_cp = numpy.zeros((ncore,nmo))
eris.k_cp = numpy.zeros((ncore,nmo))
fmmm = _ao2mo._fpointer('AO2MOmmm_nr_s2_iltj')
fdrv = _ao2mo.libao2mo.AO2MOnr_e2_drv
ftrans = _ao2mo._fpointer('AO2MOtranse2_nr_s2kl')
with df.load(self._cderi) as feri:
for b0, b1 in dfhf.prange(0, self._naoaux, dfhf.BLOCKDIM):
eri1 = numpy.array(feri[b0:b1], copy=False)
buf = numpy.empty((b1-b0,nmo,nmo))
if log.verbose >= pyscf.lib.logger.DEBUG1:
t1 = log.timer('load buf %d:%d'%(b0,b1), *t1)
fdrv(ftrans, fmmm,
buf.ctypes.data_as(ctypes.c_void_p),
eri1.ctypes.data_as(ctypes.c_void_p),
mo.ctypes.data_as(ctypes.c_void_p),
ctypes.c_int(b1-b0), ctypes.c_int(nao),
ctypes.c_int(0), ctypes.c_int(nmo),
ctypes.c_int(0), ctypes.c_int(nmo),
ctypes.c_void_p(0), ctypes.c_int(0))
if log.verbose >= pyscf.lib.logger.DEBUG1:
t1 = log.timer('transform [%d:%d]'%(b0,b1), *t1)
bufd = numpy.einsum('kii->ki', buf).copy()
#:eris.j_cp += numpy.einsum('ki,kj->ij', bufd[:,:ncore], bufd)
pyscf.lib.dot(bufd[:,:ncore].T.copy(), bufd, 1, eris.j_cp, 1)
eris.k_cp += numpy.einsum('kij,kij->ij', buf[:,:ncore], buf[:,:ncore])
if log.verbose >= pyscf.lib.logger.DEBUG1:
t1 = log.timer('j_cp and k_cp', *t1)
return eris
def ao2mo_aaaa(casscf, mo):
dfhf.get_jk_(casscf, casscf.mol, [])
nao, nmo = mo.shape
buf = numpy.empty((casscf._naoaux,nmo*(nmo+1)//2))
mo = numpy.asarray(mo, order='F')
fmmm = _ao2mo._fpointer('AO2MOmmm_nr_s2_s2')
fdrv = _ao2mo.libao2mo.AO2MOnr_e2_drv
ftrans = _ao2mo._fpointer('AO2MOtranse2_nr_s2')
with df.load(casscf._cderi) as feri:
for b0, b1 in dfhf.prange(0, casscf._naoaux, dfhf.BLOCKDIM):
eri1 = numpy.asarray(feri[b0:b1], order='C')
fdrv(ftrans, fmmm,
buf[b0:b1].ctypes.data_as(ctypes.c_void_p),
eri1.ctypes.data_as(ctypes.c_void_p),
mo.ctypes.data_as(ctypes.c_void_p),
ctypes.c_int(b1-b0), ctypes.c_int(nao),
ctypes.c_int(0), ctypes.c_int(nmo),
ctypes.c_int(0), ctypes.c_int(nmo),
ctypes.c_void_p(0), ctypes.c_int(0))
eri1 = None
eri = pyscf.lib.dot(buf.T, buf)
return eri
def __init__(self, casscf, mo):
assert(casscf._scf._tag_df)
import gc
gc.collect()
log = logger.Logger(casscf.stdout, casscf.verbose)
mol = casscf.mol
nao, nmo = mo.shape
ncore = casscf.ncore
ncas = casscf.ncas
nocc = ncore + ncas
naoaux = casscf._naoaux
mem_incore, mem_outcore, mem_basic = _mem_usage(ncore, ncas, nmo)
mem_now = pyscf.lib.current_memory()[0]
max_memory = max(3000, casscf.max_memory*.9-mem_now)
if max_memory < mem_basic:
log.warn('Calculation needs %d MB memory, over CASSCF.max_memory (%d MB) limit',
(mem_basic+mem_now)/.9, casscf.max_memory)
t0 = (time.clock(), time.time())
self._tmpfile = tempfile.NamedTemporaryFile()
self.feri = h5py.File(self._tmpfile.name, 'w')
self.ppaa = self.feri.create_dataset('ppaa', (nmo,nmo,ncas,ncas), 'f8')
self.papa = self.feri.create_dataset('papa', (nmo,ncas,nmo,ncas), 'f8')
self.j_pc = numpy.zeros((nmo,ncore))
k_cp = numpy.zeros((ncore,nmo))
mo = numpy.asarray(mo, order='F')
_tmpfile1 = tempfile.NamedTemporaryFile()
fxpp = h5py.File(_tmpfile1.name)
bufpa = numpy.empty((naoaux,nmo,ncas))
bufs1 = numpy.empty((dfhf.BLOCKDIM,nmo,nmo))
fmmm = _ao2mo._fpointer('AO2MOmmm_nr_s2_iltj')
fdrv = _ao2mo.libao2mo.AO2MOnr_e2_drv
ftrans = _ao2mo._fpointer('AO2MOtranse2_nr_s2')
t2 = t1 = t0
fxpp_keys = []
with df.load(casscf._cderi) as feri:
for b0, b1 in dfhf.prange(0, naoaux, dfhf.BLOCKDIM):
eri1 = numpy.asarray(feri[b0:b1], order='C')
if log.verbose >= logger.DEBUG1:
t2 = log.timer('load buf %d:%d'%(b0,b1), *t2)
bufpp = bufs1[:b1-b0]
fdrv(ftrans, fmmm,
bufpp.ctypes.data_as(ctypes.c_void_p),
eri1.ctypes.data_as(ctypes.c_void_p),
mo.ctypes.data_as(ctypes.c_void_p),
ctypes.c_int(b1-b0), ctypes.c_int(nao),
ctypes.c_int(0), ctypes.c_int(nmo),
ctypes.c_int(0), ctypes.c_int(nmo),
ctypes.c_void_p(0), ctypes.c_int(0))
fxpp_keys.append([str(b0), b0, b1])
fxpp[str(b0)] = bufpp.transpose(1,2,0)
bufpa[b0:b1] = bufpp[:,:,ncore:nocc]
bufd = numpy.einsum('kii->ki', bufpp)
self.j_pc += numpy.einsum('ki,kj->ij', bufd, bufd[:,:ncore])
k_cp += numpy.einsum('kij,kij->ij', bufpp[:,:ncore], bufpp[:,:ncore])
if log.verbose >= logger.DEBUG1:
t1 = log.timer('j_pc and k_pc', *t1)
eri1 = None
self.k_pc = k_cp.T.copy()
bufs1 = bufpp = None
t1 = log.timer('density fitting ao2mo pass1', *t0)
mem_now = pyscf.lib.current_memory()[0]
nblk = int(max(8, min(nmo, ((max_memory-mem_now)*1e6/8-bufpa.size)/(ncas**2*nmo))))
bufs1 = numpy.empty((nblk,ncas,nmo,ncas))
dgemm = pyscf.lib.numpy_helper._dgemm
for p0, p1 in prange(0, nmo, nblk):
#tmp = numpy.dot(bufpa[:,p0:p1].reshape(naoaux,-1).T,
# bufpa.reshape(naoaux,-1))
tmp = bufs1[:p1-p0]
dgemm('T', 'N', (p1-p0)*ncas, nmo*ncas, naoaux,
bufpa.reshape(naoaux,-1), bufpa.reshape(naoaux,-1),
tmp.reshape(-1,nmo*ncas), 1, 0, p0*ncas, 0, 0)
self.papa[p0:p1] = tmp.reshape(p1-p0,ncas,nmo,ncas)
bufaa = bufpa[:,ncore:nocc,:].copy().reshape(-1,ncas**2)
bufs1 = bufpa = None
t1 = log.timer('density fitting papa pass2', *t1)
mem_now = pyscf.lib.current_memory()[0]
nblk = int(max(8, min(nmo, (max_memory-mem_now)*1e6/8/(nmo*naoaux+ncas**2*nmo))))
bufs1 = numpy.empty((nblk,nmo,naoaux))
bufs2 = numpy.empty((nblk,nmo,ncas,ncas))
for p0, p1 in prange(0, nmo, nblk):
nrow = p1 - p0
buf = bufs1[:nrow]
tmp = bufs2[:nrow].reshape(-1,ncas**2)
col0 = 0
for key, col0, col1 in fxpp_keys:
buf[:nrow,:,col0:col1] = fxpp[key][p0:p1]
pyscf.lib.dot(buf.reshape(-1,naoaux), bufaa, 1, tmp)
self.ppaa[p0:p1] = tmp.reshape(p1-p0,nmo,ncas,ncas)
bufs1 = bufs2 = buf = None
t1 = log.timer('density fitting ppaa pass2', *t1)
fxpp.close()
self.feri.flush()
dm_core = numpy.dot(mo[:,:ncore], mo[:,:ncore].T)
vj, vk = casscf.get_jk(mol, dm_core)
self.vhf_c = reduce(numpy.dot, (mo.T, vj*2-vk, mo))
t0 = log.timer('density fitting ao2mo', *t0)
