def _ao2mo_ovov(mp, orbs, feri, max_memory=2000, verbose=None):
time0 = (time.clock(), time.time())
log = logger.new_logger(mp, verbose)
orboa = numpy.asarray(orbs[0], order='F')
orbva = numpy.asarray(orbs[1], order='F')
orbob = numpy.asarray(orbs[2], order='F')
orbvb = numpy.asarray(orbs[3], order='F')
nao, nocca = orboa.shape
noccb = orbob.shape[1]
nvira = orbva.shape[1]
nvirb = orbvb.shape[1]
mol = mp.mol
int2e = mol._add_suffix('int2e')
ao2mopt = _ao2mo.AO2MOpt(mol, int2e, 'CVHFnr_schwarz_cond',
'CVHFsetnr_direct_scf')
nbas = mol.nbas
assert (nvira <= nao)
assert (nvirb <= nao)
ao_loc = mol.ao_loc_nr()
dmax = max(
4, min(nao / 3, numpy.sqrt(max_memory * .95e6 / 8 / (nao + nocca)**2)))
sh_ranges = ao2mo.outcore.balance_partition(ao_loc, dmax)
dmax = max(x[2] for x in sh_ranges)
eribuf = numpy.empty((nao, dmax, dmax, nao))
ftmp = lib.H5TmpFile()
disk = (nocca**2 * (nao * (nao + dmax) / 2 + nvira**2) + noccb**2 *
(nao * (nao + dmax) / 2 + nvirb**2) + nocca * noccb *
(nao**2 + nvira * nvirb))
log.debug('max_memory %s MB (dmax = %s) required disk space %g MB',
max_memory, dmax, disk * 8 / 1e6)
fint = gto.moleintor.getints4c
aa_blk_slices = []
ab_blk_slices = []
count_ab = 0
count_aa = 0
time1 = time0
with lib.call_in_background(ftmp.__setitem__) as save:
for ish0, ish1, ni in sh_ranges:
for jsh0, jsh1, nj in sh_ranges:
i0, i1 = ao_loc[ish0], ao_loc[ish1]
j0, j1 = ao_loc[jsh0], ao_loc[jsh1]
eri = fint(int2e,
mol._atm,
mol._bas,
mol._env,
shls_slice=(0, nbas, ish0, ish1, jsh0, jsh1, 0,
nbas),
aosym='s1',
ao_loc=ao_loc,
cintopt=ao2mopt._cintopt,
out=eribuf)
tmp_i = lib.ddot(orboa.T,
eri.reshape(nao, (i1 - i0) * (j1 - j0) * nao))
tmp_li = lib.ddot(
orbob.T,
tmp_i.reshape(nocca * (i1 - i0) * (j1 - j0), nao).T)
tmp_li = tmp_li.reshape(noccb, nocca, (i1 - i0), (j1 - j0))
save('ab/%d' % count_ab, tmp_li.transpose(1, 0, 2, 3))
ab_blk_slices.append((i0, i1, j0, j1))
count_ab += 1
if ish0 >= jsh0:
tmp_li = lib.ddot(
orboa.T,
tmp_i.reshape(nocca * (i1 - i0) * (j1 - j0), nao).T)
tmp_li = tmp_li.reshape(nocca, nocca, (i1 - i0), (j1 - j0))
save('aa/%d' % count_aa, tmp_li.transpose(1, 0, 2, 3))
tmp_i = lib.ddot(
orbob.T, eri.reshape(nao, (i1 - i0) * (j1 - j0) * nao))
tmp_li = lib.ddot(
orbob.T,
tmp_i.reshape(noccb * (i1 - i0) * (j1 - j0), nao).T)
tmp_li = tmp_li.reshape(noccb, noccb, (i1 - i0), (j1 - j0))
save('bb/%d' % count_aa, tmp_li.transpose(1, 0, 2, 3))
aa_blk_slices.append((i0, i1, j0, j1))
count_aa += 1
time1 = log.timer_debug1(
'partial ao2mo [%d:%d,%d:%d]' % (ish0, ish1, jsh0, jsh1),
*time1)
time1 = time0 = log.timer('mp2 ao2mo_ovov pass1', *time0)
eri = eribuf = tmp_i = tmp_li = None
fovov = feri.create_dataset('ovov', (nocca * nvira, nocca * nvira),
'f8',
chunks=(nvira, nvira))
fovOV = feri.create_dataset('ovOV', (nocca * nvira, noccb * nvirb),
'f8',
chunks=(nvira, nvirb))
fOVOV = feri.create_dataset('OVOV', (noccb * nvirb, noccb * nvirb),
'f8',
chunks=(nvirb, nvirb))
occblk = int(
min(max(nocca, noccb),
max(4, 250 / nocca, max_memory * .9e6 / 8 / (nao**2 * nocca) / 5)))
def load_aa(h5g, nocc, i0, eri):
if i0 < nocc:
i1 = min(i0 + occblk, nocc)
for k, (p0, p1, q0, q1) in enumerate(aa_blk_slices):
eri[:i1 - i0, :, p0:p1, q0:q1] = h5g[str(k)][i0:i1]
if p0 != q0:
dat = numpy.asarray(h5g[str(k)][:, i0:i1])
eri[:i1 - i0, :, q0:q1, p0:p1] = dat.transpose(1, 0, 3, 2)
def load_ab(h5g, nocca, i0, eri):
if i0 < nocca:
i1 = min(i0 + occblk, nocca)
for k, (p0, p1, q0, q1) in enumerate(ab_blk_slices):
eri[:i1 - i0, :, p0:p1, q0:q1] = h5g[str(k)][i0:i1]
def save(h5dat, nvir, i0, i1, dat):
for i in range(i0, i1):
h5dat[i * nvir:(i + 1) * nvir] = dat[i - i0].reshape(nvir, -1)
with lib.call_in_background(save) as bsave:
with lib.call_in_background(load_aa) as prefetch:
buf_prefecth = numpy.empty((occblk, nocca, nao, nao))
buf = numpy.empty_like(buf_prefecth)
load_aa(ftmp['aa'], nocca, 0, buf_prefecth)
for i0, i1 in lib.prange(0, nocca, occblk):
buf, buf_prefecth = buf_prefecth, buf
prefetch(ftmp['aa'], nocca, i1, buf_prefecth)
eri = buf[:i1 - i0].reshape((i1 - i0) * nocca, nao, nao)
dat = _ao2mo.nr_e2(eri, orbva, (0, nvira, 0, nvira), 's1',
's1')
bsave(
fovov, nvira, i0, i1,
dat.reshape(i1 - i0, nocca, nvira,
nvira).transpose(0, 2, 1, 3))
time1 = log.timer_debug1(
'pass2 ao2mo for aa [%d:%d]' % (i0, i1), *time1)
buf_prefecth = numpy.empty((occblk, noccb, nao, nao))
buf = numpy.empty_like(buf_prefecth)
load_aa(ftmp['bb'], noccb, 0, buf_prefecth)
for i0, i1 in lib.prange(0, noccb, occblk):
buf, buf_prefecth = buf_prefecth, buf
prefetch(ftmp['bb'], noccb, i1, buf_prefecth)
eri = buf[:i1 - i0].reshape((i1 - i0) * noccb, nao, nao)
dat = _ao2mo.nr_e2(eri, orbvb, (0, nvirb, 0, nvirb), 's1',
's1')
bsave(
fOVOV, nvirb, i0, i1,
dat.reshape(i1 - i0, noccb, nvirb,
nvirb).transpose(0, 2, 1, 3))
time1 = log.timer_debug1(
'pass2 ao2mo for bb [%d:%d]' % (i0, i1), *time1)
orbvab = numpy.asarray(numpy.hstack((orbva, orbvb)), order='F')
with lib.call_in_background(load_ab) as prefetch:
load_ab(ftmp['ab'], nocca, 0, buf_prefecth)
for i0, i1 in lib.prange(0, nocca, occblk):
buf, buf_prefecth = buf_prefecth, buf
prefetch(ftmp['ab'], nocca, i1, buf_prefecth)
eri = buf[:i1 - i0].reshape((i1 - i0) * noccb, nao, nao)
dat = _ao2mo.nr_e2(eri, orbvab,
(0, nvira, nvira, nvira + nvirb), 's1',
's1')
bsave(
fovOV, nvira, i0, i1,
dat.reshape(i1 - i0, noccb, nvira,
nvirb).transpose(0, 2, 1, 3))
time1 = log.timer_debug1(
'pass2 ao2mo for ab [%d:%d]' % (i0, i1), *time1)
time0 = log.timer('mp2 ao2mo_ovov pass2', *time0)
def _ao2mo_ovov(mp, orbo, orbv, feri, max_memory=2000, verbose=None):
time0 = (time.clock(), time.time())
log = logger.new_logger(mp, verbose)
mol = mp.mol
int2e = mol._add_suffix('int2e')
ao2mopt = _ao2mo.AO2MOpt(mol, int2e, 'CVHFnr_schwarz_cond',
'CVHFsetnr_direct_scf')
nao, nocc = orbo.shape
nvir = orbv.shape[1]
nbas = mol.nbas
assert (nvir <= nao)
ao_loc = mol.ao_loc_nr()
dmax = max(
4, min(nao / 3, numpy.sqrt(max_memory * .95e6 / 8 / (nao + nocc)**2)))
sh_ranges = ao2mo.outcore.balance_partition(ao_loc, dmax)
dmax = max(x[2] for x in sh_ranges)
eribuf = numpy.empty((nao, dmax, dmax, nao))
ftmp = lib.H5TmpFile()
log.debug('max_memory %s MB (dmax = %s) required disk space %g MB',
max_memory, dmax,
nocc**2 * (nao * (nao + dmax) / 2 + nvir**2) * 8 / 1e6)
buf_i = numpy.empty((nocc * dmax**2 * nao))
buf_li = numpy.empty((nocc**2 * dmax**2))
buf1 = numpy.empty_like(buf_li)
fint = gto.moleintor.getints4c
jk_blk_slices = []
count = 0
time1 = time0
with lib.call_in_background(ftmp.__setitem__) as save:
for ip, (ish0, ish1, ni) in enumerate(sh_ranges):
for jsh0, jsh1, nj in sh_ranges[:ip + 1]:
i0, i1 = ao_loc[ish0], ao_loc[ish1]
j0, j1 = ao_loc[jsh0], ao_loc[jsh1]
jk_blk_slices.append((i0, i1, j0, j1))
eri = fint(int2e,
mol._atm,
mol._bas,
mol._env,
shls_slice=(0, nbas, ish0, ish1, jsh0, jsh1, 0,
nbas),
aosym='s1',
ao_loc=ao_loc,
cintopt=ao2mopt._cintopt,
out=eribuf)
tmp_i = numpy.ndarray((nocc, (i1 - i0) * (j1 - j0) * nao),
buffer=buf_i)
tmp_li = numpy.ndarray((nocc, nocc * (i1 - i0) * (j1 - j0)),
buffer=buf_li)
lib.ddot(orbo.T,
eri.reshape(nao, (i1 - i0) * (j1 - j0) * nao),
c=tmp_i)
lib.ddot(orbo.T,
tmp_i.reshape(nocc * (i1 - i0) * (j1 - j0), nao).T,
c=tmp_li)
tmp_li = tmp_li.reshape(nocc, nocc, (i1 - i0), (j1 - j0))
save(str(count), tmp_li.transpose(1, 0, 2, 3))
buf_li, buf1 = buf1, buf_li
count += 1
time1 = log.timer_debug1(
'partial ao2mo [%d:%d,%d:%d]' % (ish0, ish1, jsh0, jsh1),
*time1)
time1 = time0 = log.timer('mp2 ao2mo_ovov pass1', *time0)
eri = eribuf = tmp_i = tmp_li = buf_i = buf_li = buf1 = None
h5dat = feri.create_dataset('ovov', (nocc * nvir, nocc * nvir),
'f8',
chunks=(nvir, nvir))
occblk = int(
min(nocc,
max(4, 250 / nocc, max_memory * .9e6 / 8 / (nao**2 * nocc) / 5)))
def load(i0, eri):
if i0 < nocc:
i1 = min(i0 + occblk, nocc)
for k, (p0, p1, q0, q1) in enumerate(jk_blk_slices):
eri[:i1 - i0, :, p0:p1, q0:q1] = ftmp[str(k)][i0:i1]
if p0 != q0:
dat = numpy.asarray(ftmp[str(k)][:, i0:i1])
eri[:i1 - i0, :, q0:q1, p0:p1] = dat.transpose(1, 0, 3, 2)
def save(i0, i1, dat):
for i in range(i0, i1):
h5dat[i * nvir:(i + 1) * nvir] = dat[i - i0].reshape(
nvir, nocc * nvir)
orbv = numpy.asarray(orbv, order='F')
buf_prefecth = numpy.empty((occblk, nocc, nao, nao))
buf = numpy.empty_like(buf_prefecth)
bufw = numpy.empty((occblk * nocc, nvir**2))
bufw1 = numpy.empty_like(bufw)
with lib.call_in_background(load) as prefetch:
with lib.call_in_background(save) as bsave:
load(0, buf_prefecth)
for i0, i1 in lib.prange(0, nocc, occblk):
buf, buf_prefecth = buf_prefecth, buf
prefetch(i1, buf_prefecth)
eri = buf[:i1 - i0].reshape((i1 - i0) * nocc, nao, nao)
dat = _ao2mo.nr_e2(eri,
orbv, (0, nvir, 0, nvir),
's1',
's1',
out=bufw)
bsave(
i0, i1,
dat.reshape(i1 - i0, nocc, nvir,
nvir).transpose(0, 2, 1, 3))
bufw, bufw1 = bufw1, bufw
time1 = log.timer_debug1('pass2 ao2mo [%d:%d]' % (i0, i1),
*time1)
time0 = log.timer('mp2 ao2mo_ovov pass2', *time0)
return h5dat
def half_e1(mol, mo_coeffs, swapfile,
intor='int2e', aosym='s4', comp=1,
max_memory=MAX_MEMORY, ioblk_size=IOBLK_SIZE, verbose=logger.WARN,
compact=True, ao2mopt=None):
r'''Half transform arbitrary spherical AO integrals to MO integrals
for the given two sets of orbitals
Args:
mol : :class:`Mole` object
AO integrals will be generated in terms of mol._atm, mol._bas, mol._env
mo_coeff : ndarray
Transform (ij|kl) with the same set of orbitals.
swapfile : str or h5py File or h5py Group object
To store the transformed integrals, in HDF5 format. The transformed
integrals are saved in blocks.
Kwargs
intor : str
Name of the 2-electron integral. Ref to :func:`getints_by_shell`
for the complete list of available 2-electron integral names
aosym : int or str
Permutation symmetry for the AO integrals
| 4 or '4' or 's4': 4-fold symmetry (default)
| '2ij' or 's2ij' : symmetry between i, j in (ij|kl)
| '2kl' or 's2kl' : symmetry between k, l in (ij|kl)
| 1 or '1' or 's1': no symmetry
| 'a4ij' : 4-fold symmetry with anti-symmetry between i, j in (ij|kl) (TODO)
| 'a4kl' : 4-fold symmetry with anti-symmetry between k, l in (ij|kl) (TODO)
| 'a2ij' : anti-symmetry between i, j in (ij|kl) (TODO)
| 'a2kl' : anti-symmetry between k, l in (ij|kl) (TODO)
comp : int
Components of the integrals, e.g. int2e_ip_sph has 3 components.
verbose : int
Print level
max_memory : float or int
The maximum size of cache to use (in MB), large cache may **not**
improve performance.
ioblk_size : float or int
The block size for IO, large block size may **not** improve performance
verbose : int
Print level
compact : bool
When compact is True, depending on the four oribital sets, the
returned MO integrals has (up to 4-fold) permutation symmetry.
If it's False, the function will abandon any permutation symmetry,
and return the "plain" MO integrals
ao2mopt : :class:`AO2MOpt` object
Precomputed data to improve perfomance
Returns:
None
'''
if any(c.dtype == numpy.complex128 for c in mo_coeffs):
raise NotImplementedError('Integral transformation for complex orbitals')
intor = mol._add_suffix(intor)
time0 = (logger.process_clock(), logger.perf_counter())
log = logger.new_logger(mol, verbose)
nao = mo_coeffs[0].shape[0]
aosym = _stand_sym_code(aosym)
if aosym in ('s4', 's2ij'):
nao_pair = nao * (nao+1) // 2
else:
nao_pair = nao * nao
ijmosym, nij_pair, moij, ijshape = \
incore._conc_mos(mo_coeffs[0], mo_coeffs[1],
compact and aosym in ('s4', 's2ij'))
e1buflen, mem_words, iobuf_words, ioblk_words = \
guess_e1bufsize(max_memory, ioblk_size, nij_pair, nao_pair, comp)
ioblk_size = ioblk_words * 8/1e6
# The buffer to hold AO integrals in C code, see line (@)
aobuflen = max(int((mem_words - 2*comp*e1buflen*nij_pair) // (nao_pair*comp)),
IOBUF_ROW_MIN)
ao_loc = mol.ao_loc_nr('_cart' in intor)
shranges = guess_shell_ranges(mol, (aosym in ('s4', 's2kl')), e1buflen,
aobuflen, ao_loc)
if ao2mopt is None:
if intor == 'int2e_cart' or intor == 'int2e_sph':
ao2mopt = _ao2mo.AO2MOpt(mol, intor, 'CVHFnr_schwarz_cond',
'CVHFsetnr_direct_scf')
else:
ao2mopt = _ao2mo.AO2MOpt(mol, intor)
if isinstance(swapfile, h5py.Group):
fswap = swapfile
else:
fswap = lib.H5TmpFile(swapfile)
for icomp in range(comp):
fswap.create_group(str(icomp)) # for h5py old version
log.debug('step1: tmpfile %s %.8g MB', fswap.filename, nij_pair*nao_pair*8/1e6)
log.debug('step1: (ij,kl) = (%d,%d), mem cache %.8g MB, iobuf %.8g MB',
nij_pair, nao_pair, mem_words*8/1e6, iobuf_words*8/1e6)
nstep = len(shranges)
e1buflen = max([x[2] for x in shranges])
e2buflen, chunks = guess_e2bufsize(ioblk_size, nij_pair, e1buflen)
def save(istep, iobuf):
for icomp in range(comp):
_transpose_to_h5g(fswap, '%d/%d'%(icomp,istep), iobuf[icomp],
e2buflen, None)
# transform e1
ti0 = log.timer('Initializing ao2mo.outcore.half_e1', *time0)
with lib.call_in_background(save) as async_write:
buf1 = numpy.empty((comp*e1buflen,nao_pair))
buf2 = numpy.empty((comp*e1buflen,nij_pair))
buf_write = numpy.empty_like(buf2)
fill = _ao2mo.nr_e1fill
f_e1 = _ao2mo.nr_e1
for istep,sh_range in enumerate(shranges):
log.debug1('step 1 [%d/%d], AO [%d:%d], len(buf) = %d',
istep+1, nstep, *(sh_range[:3]))
buflen = sh_range[2]
iobuf = numpy.ndarray((comp,buflen,nij_pair), buffer=buf2)
nmic = len(sh_range[3])
p1 = 0
for imic, aoshs in enumerate(sh_range[3]):
log.debug2(' fill iobuf micro [%d/%d], AO [%d:%d], len(aobuf) = %d',
imic+1, nmic, *aoshs)
buf = fill(intor, aoshs, mol._atm, mol._bas, mol._env,
aosym, comp, ao2mopt, out=buf1).reshape(-1,nao_pair)
buf = f_e1(buf, moij, ijshape, aosym, ijmosym)
p0, p1 = p1, p1 + aoshs[2]
iobuf[:,p0:p1] = buf.reshape(comp,aoshs[2],nij_pair)
ti0 = log.timer_debug1('gen AO/transform MO [%d/%d]'%(istep+1,nstep), *ti0)
async_write(istep, iobuf)
buf2, buf_write = buf_write, buf2
fswap = None
return swapfile
def _make_ao_ints(mol, mo_coeff, nocc, dtype):
NS = ctf.SYM.NS
SY = ctf.SYM.SY
ao_loc = mol.ao_loc_nr()
mo = np.asarray(mo_coeff, order='F')
nao, nmo = mo.shape
nvir = nmo - nocc
ppoo = ctf.tensor((nao, nao, nocc, nocc),
sym=[SY, NS, NS, NS],
dtype=dtype)
ppov = ctf.tensor((nao, nao, nocc, nvir),
sym=[SY, NS, NS, NS],
dtype=dtype)
ppvv = ctf.tensor((nao, nao, nvir, nvir),
sym=[SY, NS, SY, NS],
dtype=dtype)
intor = mol._add_suffix('int2e')
ao2mopt = _ao2mo.AO2MOpt(mol, intor, 'CVHFnr_schwarz_cond',
'CVHFsetnr_direct_scf')
blksize = int(max(4, min(nao / 3, nao / size**.5, 2000e6 / 8 / nao**3)))
sh_ranges = ao2mo.outcore.balance_partition(ao_loc, blksize)
tasks = []
for k, (ish0, ish1, di) in enumerate(sh_ranges):
for jsh0, jsh1, dj in sh_ranges[:k + 1]:
tasks.append((ish0, ish1, jsh0, jsh1))
sqidx = np.arange(nao**2).reshape(nao, nao)
trilidx = sqidx[np.tril_indices(nao)]
vsqidx = np.arange(nvir**2).reshape(nvir, nvir)
vtrilidx = vsqidx[np.tril_indices(nvir)]
subtasks = list(static_partition(tasks))
ntasks = max(comm.allgather(len(subtasks)))
for itask in range(ntasks):
if itask >= len(subtasks):
ppoo.write([], [])
ppov.write([], [])
ppvv.write([], [])
continue
shls_slice = subtasks[itask]
ish0, ish1, jsh0, jsh1 = shls_slice
i0, i1 = ao_loc[ish0], ao_loc[ish1]
j0, j1 = ao_loc[jsh0], ao_loc[jsh1]
di = i1 - i0
dj = j1 - j0
if i0 != j0:
eri = gto.moleintor.getints4c(intor,
mol._atm,
mol._bas,
mol._env,
shls_slice=shls_slice,
aosym='s2kl',
ao_loc=ao_loc,
cintopt=ao2mopt._cintopt)
idx = sqidx[i0:i1, j0:j1].ravel()
eri = _ao2mo.nr_e2(eri.reshape(di * dj, -1), mo, (0, nmo, 0, nmo),
's2kl', 's1')
else:
eri = gto.moleintor.getints4c(intor,
mol._atm,
mol._bas,
mol._env,
shls_slice=shls_slice,
aosym='s4',
ao_loc=ao_loc,
cintopt=ao2mopt._cintopt)
eri = _ao2mo.nr_e2(eri, mo, (0, nmo, 0, nmo), 's4', 's1')
idx = sqidx[i0:i1, j0:j1][np.tril_indices(i1 - i0)]
ooidx = idx[:, None] * nocc**2 + np.arange(nocc**2)
ovidx = idx[:, None] * (nocc * nvir) + np.arange(nocc * nvir)
vvidx = idx[:, None] * nvir**2 + vtrilidx
eri = eri.reshape(-1, nmo, nmo)
ppoo.write(ooidx.ravel(), eri[:, :nocc, :nocc].ravel())
ppov.write(ovidx.ravel(), eri[:, :nocc, nocc:].ravel())
ppvv.write(vvidx.ravel(),
pyscflib.pack_tril(eri[:, nocc:, nocc:]).ravel())
idx = eri = None
return ppoo, ppov, ppvv
def trans_e1_outcore(mol,
mo,
ncore,
ncas,
erifile,
max_memory=None,
level=1,
verbose=logger.WARN):
time0 = (time.clock(), time.time())
if isinstance(verbose, logger.Logger):
log = verbose
else:
log = logger.Logger(mol.stdout, verbose)
log.debug1('trans_e1_outcore level %d max_memory %d', level, max_memory)
nao, nmo = mo.shape
nao_pair = nao * (nao + 1) // 2
nocc = ncore + ncas
_tmpfile1 = tempfile.NamedTemporaryFile(dir=lib.param.TMPDIR)
faapp_buf = h5py.File(_tmpfile1.name)
feri = h5py.File(erifile, 'w')
mo_c = numpy.asarray(mo, order='C')
mo = numpy.asarray(mo, order='F')
pashape = (0, nmo, ncore, nocc)
papa_buf = numpy.zeros((nao, ncas, nmo * ncas))
j_pc = numpy.zeros((nmo, ncore))
k_pc = numpy.zeros((nmo, ncore))
mem_words = int(max(2000, max_memory - papa_buf.nbytes / 1e6) * 1e6 / 8)
aobuflen = mem_words // (nao_pair + nocc * nmo) + 1
ao_loc = numpy.array(mol.ao_loc_nr(), dtype=numpy.int32)
shranges = outcore.guess_shell_ranges(mol, True, aobuflen, None, ao_loc)
intor = mol._add_suffix('int2e')
ao2mopt = _ao2mo.AO2MOpt(mol, intor, 'CVHFnr_schwarz_cond',
'CVHFsetnr_direct_scf')
nstep = len(shranges)
paapp = 0
maxbuflen = max([x[2] for x in shranges])
log.debug('mem_words %.8g MB, maxbuflen = %d', mem_words * 8 / 1e6,
maxbuflen)
bufs1 = numpy.empty((maxbuflen, nao_pair))
bufs2 = numpy.empty((maxbuflen, nmo * ncas))
if level == 1:
bufs3 = numpy.empty((maxbuflen, nao * ncore))
log.debug('mem cache %.8g MB',
(bufs1.nbytes + bufs2.nbytes + bufs3.nbytes) / 1e6)
else:
log.debug('mem cache %.8g MB', (bufs1.nbytes + bufs2.nbytes) / 1e6)
ti0 = log.timer('Initializing trans_e1_outcore', *time0)
# fmmm, ftrans, fdrv for level 1
fmmm = libmcscf.AO2MOmmm_ket_nr_s2
ftrans = libmcscf.AO2MOtranse1_nr_s4
fdrv = libmcscf.AO2MOnr_e2_drv
for istep, sh_range in enumerate(shranges):
log.debug('[%d/%d], AO [%d:%d], len(buf) = %d', istep + 1, nstep,
*sh_range)
buf = bufs1[:sh_range[2]]
_ao2mo.nr_e1fill(intor, sh_range, mol._atm, mol._bas, mol._env, 's4',
1, ao2mopt, buf)
if log.verbose >= logger.DEBUG1:
ti1 = log.timer('AO integrals buffer', *ti0)
bufpa = bufs2[:sh_range[2]]
_ao2mo.nr_e1(buf, mo, pashape, 's4', 's1', out=bufpa)
# jc_pp, kc_pp
if level == 1: # ppaa, papa and vhf, jcp, kcp
if log.verbose >= logger.DEBUG1:
ti1 = log.timer('buffer-pa', *ti1)
buf1 = bufs3[:sh_range[2]]
fdrv(ftrans, fmmm, buf1.ctypes.data_as(ctypes.c_void_p),
buf.ctypes.data_as(ctypes.c_void_p),
mo.ctypes.data_as(ctypes.c_void_p), ctypes.c_int(sh_range[2]),
ctypes.c_int(nao), (ctypes.c_int * 4)(0, nao, 0, ncore),
ctypes.POINTER(ctypes.c_void_p)(), ctypes.c_int(0))
p0 = 0
for ij in range(sh_range[0], sh_range[1]):
i, j = _ao2mo._extract_pair(ij)
i0 = ao_loc[i]
j0 = ao_loc[j]
i1 = ao_loc[i + 1]
j1 = ao_loc[j + 1]
di = i1 - i0
dj = j1 - j0
if i == j:
dij = di * (di + 1) // 2
buf = numpy.empty((di, di, nao * ncore))
idx = numpy.tril_indices(di)
buf[idx] = buf1[p0:p0 + dij]
buf[idx[1], idx[0]] = buf1[p0:p0 + dij]
buf = buf.reshape(di, di, nao, ncore)
mo1 = mo_c[i0:i1]
tmp = numpy.einsum('uvpc,pc->uvc', buf, mo[:, :ncore])
tmp = lib.dot(mo1.T, tmp.reshape(di, -1))
j_pc += numpy.einsum('vp,pvc->pc', mo1,
tmp.reshape(nmo, di, ncore))
tmp = numpy.einsum('uvpc,uc->vcp', buf, mo1[:, :ncore])
tmp = lib.dot(tmp.reshape(-1, nmo),
mo).reshape(di, ncore, nmo)
k_pc += numpy.einsum('vp,vcp->pc', mo1, tmp)
else:
dij = di * dj
buf = buf1[p0:p0 + dij].reshape(di, dj, nao, ncore)
mo1 = mo_c[i0:i1]
mo2 = mo_c[j0:j1]
tmp = numpy.einsum('uvpc,pc->uvc', buf, mo[:, :ncore])
tmp = lib.dot(mo1.T, tmp.reshape(di, -1))
j_pc += numpy.einsum('vp,pvc->pc', mo2,
tmp.reshape(nmo, dj, ncore)) * 2
tmp = numpy.einsum('uvpc,uc->vcp', buf, mo1[:, :ncore])
tmp = lib.dot(tmp.reshape(-1, nmo),
mo).reshape(dj, ncore, nmo)
k_pc += numpy.einsum('vp,vcp->pc', mo2, tmp)
tmp = numpy.einsum('uvpc,vc->ucp', buf, mo2[:, :ncore])
tmp = lib.dot(tmp.reshape(-1, nmo),
mo).reshape(di, ncore, nmo)
k_pc += numpy.einsum('up,ucp->pc', mo1, tmp)
p0 += dij
if log.verbose >= logger.DEBUG1:
ti1 = log.timer('j_cp and k_cp', *ti1)
if log.verbose >= logger.DEBUG1:
ti1 = log.timer('half transformation of the buffer', *ti1)
# ppaa, papa
faapp_buf[str(istep)] = \
bufpa.reshape(sh_range[2],nmo,ncas)[:,ncore:nocc].reshape(-1,ncas**2).T
p0 = 0
for ij in range(sh_range[0], sh_range[1]):
i, j = _ao2mo._extract_pair(ij)
i0 = ao_loc[i]
j0 = ao_loc[j]
i1 = ao_loc[i + 1]
j1 = ao_loc[j + 1]
di = i1 - i0
dj = j1 - j0
if i == j:
dij = di * (di + 1) // 2
buf1 = numpy.empty((di, di, nmo * ncas))
idx = numpy.tril_indices(di)
buf1[idx] = bufpa[p0:p0 + dij]
buf1[idx[1], idx[0]] = bufpa[p0:p0 + dij]
else:
dij = di * dj
buf1 = bufpa[p0:p0 + dij].reshape(di, dj, -1)
mo1 = mo[j0:j1, ncore:nocc].copy()
for i in range(di):
lib.dot(mo1.T, buf1[i], 1, papa_buf[i0 + i], 1)
mo1 = mo[i0:i1, ncore:nocc].copy()
buf1 = lib.dot(mo1.T, buf1.reshape(di, -1))
papa_buf[j0:j1] += buf1.reshape(ncas, dj, -1).transpose(1, 0, 2)
p0 += dij
if log.verbose >= logger.DEBUG1:
ti1 = log.timer('ppaa and papa buffer', *ti1)
ti0 = log.timer('gen AO/transform MO [%d/%d]' % (istep + 1, nstep),
*ti0)
buf = buf1 = bufpa = None
bufs1 = bufs2 = bufs3 = None
time1 = log.timer('mc_ao2mo pass 1', *time0)
log.debug1('Half transformation done. Current memory %d',
lib.current_memory()[0])
nblk = int(
max(8,
min(nmo,
(max_memory * 1e6 / 8 - papa_buf.size) / (ncas**2 * nmo))))
log.debug1('nblk for papa = %d', nblk)
dset = feri.create_dataset('papa', (nmo, ncas, nmo, ncas), 'f8')
for i0, i1 in prange(0, nmo, nblk):
tmp = lib.dot(mo[:, i0:i1].T, papa_buf.reshape(nao, -1))
dset[i0:i1] = tmp.reshape(i1 - i0, ncas, nmo, ncas)
papa_buf = tmp = None
time1 = log.timer('papa pass 2', *time1)
tmp = numpy.empty((ncas**2, nao_pair))
p0 = 0
for istep, sh_range in enumerate(shranges):
tmp[:, p0:p0 + sh_range[2]] = faapp_buf[str(istep)]
p0 += sh_range[2]
nblk = int(
max(8, min(nmo,
(max_memory * 1e6 / 8 - tmp.size) / (ncas**2 * nmo) - 1)))
log.debug1('nblk for ppaa = %d', nblk)
dset = feri.create_dataset('ppaa', (nmo, nmo, ncas, ncas), 'f8')
for i0, i1 in prange(0, nmo, nblk):
tmp1 = _ao2mo.nr_e2(tmp,
mo, (i0, i1, 0, nmo),
's4',
's1',
ao_loc=ao_loc)
tmp1 = tmp1.reshape(ncas, ncas, i1 - i0, nmo)
for j in range(i1 - i0):
dset[i0 + j] = tmp1[:, :, j].transpose(2, 0, 1)
tmp = tmp1 = None
time1 = log.timer('ppaa pass 2', *time1)
faapp_buf.close()
feri.close()
_tmpfile1 = None
time0 = log.timer('mc_ao2mo', *time0)
return j_pc, k_pc
def half_e1(mol, mo_coeffs, swapfile,
intor='int2e_spinor', aosym='s4', comp=None,
max_memory=MAX_MEMORY, ioblk_size=IOBLK_SIZE, verbose=logger.WARN,
ao2mopt=None):
time0 = (time.clock(), time.time())
log = logger.new_logger(mol, verbose)
ijsame = iden_coeffs(mo_coeffs[0], mo_coeffs[1])
nmoi = mo_coeffs[0].shape[1]
nmoj = mo_coeffs[1].shape[1]
nao = mo_coeffs[0].shape[0]
aosym = outcore._stand_sym_code(aosym)
if aosym in ('s1', 's2kl', 'a2kl'):
nao_pair = nao * nao
else:
nao_pair = _count_naopair(mol, nao)
nij_pair = nmoi * nmoj
if ijsame and aosym in ('s4', 's2ij', 'a2ij', 'a4ij', 'a4kl', 'a4'):
log.debug('i-mo == j-mo')
moij = numpy.asarray(mo_coeffs[0], order='F')
ijshape = (0, nmoi, 0, nmoi)
else:
moij = numpy.asarray(numpy.hstack((mo_coeffs[0],mo_coeffs[1])), order='F')
ijshape = (0, nmoi, nmoi, nmoi+nmoj)
e1buflen, mem_words, iobuf_words, ioblk_words = \
guess_e1bufsize(max_memory, ioblk_size, nij_pair, nao_pair, comp)
# The buffer to hold AO integrals in C code
aobuflen = int((mem_words - iobuf_words) // (nao*nao*comp))
shranges = outcore.guess_shell_ranges(mol, (aosym not in ('s1', 's2ij', 'a2ij')),
aobuflen, e1buflen, mol.ao_loc_2c(), False)
if ao2mopt is None:
# if intor == 'int2e_spinor':
# ao2mopt = _ao2mo.AO2MOpt(mol, intor, 'CVHFnr_schwarz_cond',
# 'CVHFsetnr_direct_scf')
# elif intor == 'int2e_spsp1_spinor':
# elif intor == 'int2e_spsp1spsp2_spinor':
# else:
# ao2mopt = _ao2mo.AO2MOpt(mol, intor)
ao2mopt = _ao2mo.AO2MOpt(mol, intor)
log.debug('step1: tmpfile %.8g MB', nij_pair*nao_pair*16/1e6)
log.debug('step1: (ij,kl) = (%d,%d), mem cache %.8g MB, iobuf %.8g MB',
nij_pair, nao_pair, mem_words*16/1e6, iobuf_words*16/1e6)
fswap = h5py.File(swapfile, 'w')
for icomp in range(comp):
g = fswap.create_group(str(icomp)) # for h5py old version
tao = numpy.asarray(mol.tmap(), dtype=numpy.int32)
# transform e1
ti0 = log.timer('Initializing ao2mo.outcore.half_e1', *time0)
nstep = len(shranges)
for istep,sh_range in enumerate(shranges):
log.debug('step 1 [%d/%d], AO [%d:%d], len(buf) = %d', \
istep+1, nstep, *(sh_range[:3]))
buflen = sh_range[2]
iobuf = numpy.empty((comp,buflen,nij_pair), dtype=numpy.complex)
nmic = len(sh_range[3])
p0 = 0
for imic, aoshs in enumerate(sh_range[3]):
log.debug1(' fill iobuf micro [%d/%d], AO [%d:%d], len(aobuf) = %d', \
imic+1, nmic, *aoshs)
buf = _ao2mo.r_e1(intor, moij, ijshape, aoshs,
mol._atm, mol._bas, mol._env,
tao, aosym, comp, ao2mopt)
iobuf[:,p0:p0+aoshs[2]] = buf
p0 += aoshs[2]
ti2 = log.timer('gen AO/transform MO [%d/%d]'%(istep+1,nstep), *ti0)
e2buflen, chunks = guess_e2bufsize(ioblk_size, nij_pair, buflen)
for icomp in range(comp):
dset = fswap.create_dataset('%d/%d'%(icomp,istep),
(nij_pair,iobuf.shape[1]), 'c16',
chunks=None)
for col0, col1 in prange(0, nij_pair, e2buflen):
dset[col0:col1] = lib.transpose(iobuf[icomp,:,col0:col1])
ti0 = log.timer('transposing to disk', *ti2)
fswap.close()
return swapfile
def _contract_vvvv_t2(mycc, vvvv, t2T, task_locs, out=None, verbose=None):
'''Ht2 = numpy.einsum('ijcd,acbd->ijab', t2, vvvv)
where vvvv has to be real and has the 4-fold permutation symmetry
Args:
vvvv : None or integral object
if vvvv is None, contract t2 to AO-integrals using AO-direct algorithm
'''
time0 = time.clock(), time.time()
mol = mycc.mol
log = logger.new_logger(mycc, verbose)
if callable(t2T):
t2T = t2T()
assert (t2T.dtype == numpy.double)
nvira, nvirb = t2T.shape[:2]
nvir2 = nvira * nvirb
t2T = t2T.reshape(nvira, nvirb, -1)
nocc2 = t2T.shape[2]
Ht2 = numpy.ndarray(t2T.shape, dtype=t2T.dtype, buffer=out)
Ht2[:] = 0
_dgemm = lib.numpy_helper._dgemm
def contract_blk_(Ht2, t2T, eri, i0, i1, j0, j1):
ic = i1 - i0
jc = j1 - j0
#:Ht2[j0:j1] += numpy.einsum('efx,efab->abx', t2T[i0:i1], eri)
_dgemm('T', 'N', jc * nvirb, nocc2, ic * nvirb,
eri.reshape(ic * nvirb, jc * nvirb), t2T.reshape(-1, nocc2),
Ht2.reshape(nvir2, nocc2), 1, 1, 0, i0 * nvirb * nocc2,
j0 * nvirb * nocc2)
max_memory = max(MEMORYMIN, mycc.max_memory - lib.current_memory()[0])
if vvvv is None: # AO-direct CCSD
ao_loc = mol.ao_loc_nr()
intor = mol._add_suffix('int2e')
ao2mopt = _ao2mo.AO2MOpt(mol, intor, 'CVHFnr_schwarz_cond',
'CVHFsetnr_direct_scf')
blksize = max(BLKMIN, numpy.sqrt(max_memory * .9e6 / 8 / nvirb**2 / 2))
fint = gto.moleintor.getints4c
fload = ccsd._ccsd.libcc.CCload_eri
ntasks = mpi.pool.size
task_sh_locs = task_locs
sh_ranges_tasks = []
for task in range(ntasks):
sh0 = task_sh_locs[task]
sh1 = task_sh_locs[task + 1]
sh_ranges = ao2mo.outcore.balance_partition(
ao_loc, blksize, sh0, sh1)
sh_ranges_tasks.append(sh_ranges)
blksize = max(
max(x[2] for x in sh_ranges) if sh_ranges else 0
for sh_ranges in sh_ranges_tasks)
eribuf = numpy.empty((blksize, blksize, nvirb, nvirb))
loadbuf = numpy.empty((blksize, blksize, nvirb, nvirb))
ao_sh_ranges = sh_ranges_tasks[rank]
ao_sh0 = task_sh_locs[rank]
ao_sh1 = task_sh_locs[rank + 1]
ao_offset = ao_loc[ao_sh0]
assert (nvira == ao_loc[ao_sh1] - ao_loc[ao_sh0])
for task_id, t2T in _rotate_tensor_block(t2T):
sh_ranges = sh_ranges_tasks[task_id]
sh0 = task_sh_locs[task_id]
cur_offset = ao_loc[sh0]
for ish0, ish1, ni in sh_ranges:
for jsh0, jsh1, nj in ao_sh_ranges:
eri = fint(intor,
mol._atm,
mol._bas,
mol._env,
shls_slice=(ish0, ish1, jsh0, jsh1),
aosym='s2kl',
ao_loc=ao_loc,
cintopt=ao2mopt._cintopt,
out=eribuf)
i0, i1 = ao_loc[ish0] - cur_offset, ao_loc[
ish1] - cur_offset
j0, j1 = ao_loc[jsh0] - ao_offset, ao_loc[jsh1] - ao_offset
tmp = numpy.ndarray((i1 - i0, nvirb, j1 - j0, nvirb),
buffer=loadbuf)
fload(tmp.ctypes.data_as(ctypes.c_void_p),
eri.ctypes.data_as(ctypes.c_void_p),
(ctypes.c_int * 4)(i0, i1, j0, j1),
ctypes.c_int(nvirb))
contract_blk_(Ht2, t2T, tmp, i0, i1, j0, j1)
time0 = log.timer_debug1(
'AO-vvvv [%d:%d,%d:%d]' % (ish0, ish1, jsh0, jsh1),
*time0)
else:
raise NotImplementedError
return Ht2
def add_wvvVV_(self, t1, t2, eris, t2new_tril, with_ovvv=True):
time0 = time.clock(), time.time()
nocc, nvir = t1.shape
#: tau = t2 + numpy.einsum('ia,jb->ijab', t1, t1)
#: t2new += numpy.einsum('ijcd,acdb->ijab', tau, vvvv)
def contract_rec_(t2new_tril, tau, eri, i0, i1, j0, j1):
nao = tau.shape[-1]
ic = i1 - i0
jc = j1 - j0
#: t2tril[:,j0:j1] += numpy.einsum('xcd,cdab->xab', tau[:,i0:i1], eri)
_dgemm('N', 'N', nocc*(nocc+1)//2, jc*nao, ic*nao,
tau.reshape(-1,nao*nao), eri.reshape(-1,jc*nao),
t2new_tril.reshape(-1,nao*nao), 1, 1, i0*nao, 0, j0*nao)
#: t2tril[:,i0:i1] += numpy.einsum('xcd,abcd->xab', tau[:,j0:j1], eri)
_dgemm('N', 'T', nocc*(nocc+1)//2, ic*nao, jc*nao,
tau.reshape(-1,nao*nao), eri.reshape(-1,jc*nao),
t2new_tril.reshape(-1,nao*nao), 1, 1, j0*nao, 0, i0*nao)
def contract_tril_(t2new_tril, tau, eri, a0, a):
nvir = tau.shape[-1]
#: t2new[i,:i+1, a] += numpy.einsum('xcd,cdb->xb', tau[:,a0:a+1], eri)
_dgemm('N', 'N', nocc*(nocc+1)//2, nvir, (a+1-a0)*nvir,
tau.reshape(-1,nvir*nvir), eri.reshape(-1,nvir),
t2new_tril.reshape(-1,nvir*nvir), 1, 1, a0*nvir, 0, a*nvir)
#: t2new[i,:i+1,a0:a] += numpy.einsum('xd,abd->xab', tau[:,a], eri[:a])
if a > a0:
_dgemm('N', 'T', nocc*(nocc+1)//2, (a-a0)*nvir, nvir,
tau.reshape(-1,nvir*nvir), eri.reshape(-1,nvir),
t2new_tril.reshape(-1,nvir*nvir), 1, 1, a*nvir, 0, a0*nvir)
if self.direct: # AO-direct CCSD
mol = self.mol
mo = _mo_without_core(self, self.mo_coeff)
nao, nmo = mo.shape
nao_pair = nao * (nao+1) // 2
aos = numpy.asarray(mo[:,nocc:].T, order='F')
nocc2 = nocc*(nocc+1)//2
outbuf = numpy.empty((nocc2,nao,nao))
tau = numpy.ndarray((nocc2,nvir,nvir), buffer=outbuf)
p0 = 0
for i in range(nocc):
tau[p0:p0+i+1] = numpy.einsum('a,jb->jab', t1[i], t1[:i+1])
tau[p0:p0+i+1] += t2[i,:i+1]
p0 += i + 1
tau = _ao2mo.nr_e2(tau.reshape(nocc2,nvir**2), aos, (0,nao,0,nao), 's1', 's1')
tau = tau.reshape(nocc2,nao,nao)
time0 = logger.timer_debug1(self, 'vvvv-tau', *time0)
ao2mopt = _ao2mo.AO2MOpt(mol, 'cint2e_sph', 'CVHFnr_schwarz_cond',
'CVHFsetnr_direct_scf')
outbuf[:] = 0
ao_loc = mol.ao_loc_nr()
max_memory = max(0, self.max_memory - lib.current_memory()[0])
dmax = max(4, int(numpy.sqrt(max_memory*.95e6/8/nao**2/2)))
sh_ranges = ao2mo.outcore.balance_partition(ao_loc, dmax)
dmax = max(x[2] for x in sh_ranges)
eribuf = numpy.empty((dmax,dmax,nao,nao))
loadbuf = numpy.empty((dmax,dmax,nao,nao))
fint = gto.moleintor.getints2e
for ip, (ish0, ish1, ni) in enumerate(sh_ranges):
for jsh0, jsh1, nj in sh_ranges[:ip]:
eri = fint('cint2e_sph', mol._atm, mol._bas, mol._env,
shls_slice=(ish0,ish1,jsh0,jsh1), aosym='s2kl',
ao_loc=ao_loc, cintopt=ao2mopt._cintopt, out=eribuf)
i0, i1 = ao_loc[ish0], ao_loc[ish1]
j0, j1 = ao_loc[jsh0], ao_loc[jsh1]
tmp = numpy.ndarray((i1-i0,nao,j1-j0,nao), buffer=loadbuf)
_ccsd.libcc.CCload_eri(tmp.ctypes.data_as(ctypes.c_void_p),
eri.ctypes.data_as(ctypes.c_void_p),
(ctypes.c_int*4)(i0, i1, j0, j1),
ctypes.c_int(nao))
contract_rec_(outbuf, tau, tmp, i0, i1, j0, j1)
time0 = logger.timer_debug1(self, 'AO-vvvv [%d:%d,%d:%d]' %
(ish0,ish1,jsh0,jsh1), *time0)
eri = fint('cint2e_sph', mol._atm, mol._bas, mol._env,
shls_slice=(ish0,ish1,ish0,ish1), aosym='s4',
ao_loc=ao_loc, cintopt=ao2mopt._cintopt, out=eribuf)
i0, i1 = ao_loc[ish0], ao_loc[ish1]
for i in range(i1-i0):
p0, p1 = i*(i+1)//2, (i+1)*(i+2)//2
tmp = lib.unpack_tril(eri[p0:p1], out=loadbuf)
contract_tril_(outbuf, tau, tmp, i0, i0+i)
time0 = logger.timer_debug1(self, 'AO-vvvv [%d:%d,%d:%d]' %
(ish0,ish1,ish0,ish1), *time0)
eribuf = loadbuf = eri = tmp = None
tmp = _ao2mo.nr_e2(outbuf, mo, (nocc,nmo,nocc,nmo), 's1', 's1', out=tau)
t2new_tril += tmp.reshape(nocc2,nvir,nvir)
if with_ovvv:
#: tmp = numpy.einsum('ijcd,ka,kdcb->ijba', tau, t1, eris.ovvv)
#: t2new -= tmp + tmp.transpose(1,0,3,2)
tmp = _ao2mo.nr_e2(outbuf, mo, (nocc,nmo,0,nocc), 's1', 's1', out=tau)
t2new_tril -= lib.ddot(tmp.reshape(nocc2*nvir,nocc), t1).reshape(nocc2,nvir,nvir)
tmp = _ao2mo.nr_e2(outbuf, mo, (0,nocc,nocc,nmo), 's1', 's1', out=tau)
#: t2new_tril -= numpy.einsum('xkb,ka->xab', tmp.reshape(-1,nocc,nvir), t1)
tmp = lib.transpose(tmp.reshape(nocc2,nocc,nvir), axes=(0,2,1), out=outbuf)
tmp = lib.ddot(tmp.reshape(nocc2*nvir,nocc), t1, 1,
numpy.ndarray((nocc2*nvir,nvir), buffer=tau), 0)
tmp = lib.transpose(tmp.reshape(nocc2,nvir,nvir), axes=(0,2,1), out=outbuf)
t2new_tril -= tmp.reshape(nocc2,nvir,nvir)
else:
#: tau = t2 + numpy.einsum('ia,jb->ijab', t1, t1)
#: t2new += numpy.einsum('ijcd,acdb->ijab', tau, vvvv)
tau = numpy.empty((nocc*(nocc+1)//2,nvir,nvir))
p0 = 0
for i in range(nocc):
tau[p0:p0+i+1] = numpy.einsum('a,jb->jab', t1[i], t1[:i+1])
tau[p0:p0+i+1] += t2[i,:i+1]
p0 += i + 1
time0 = logger.timer_debug1(self, 'vvvv-tau', *time0)
p0 = 0
outbuf = numpy.empty((nvir,nvir,nvir))
outbuf1 = numpy.empty((nvir,nvir,nvir))
handler = None
for a in range(nvir):
buf = lib.unpack_tril(eris.vvvv[p0:p0+a+1], out=outbuf)
outbuf, outbuf1 = outbuf1, outbuf
handler = async_do(handler, contract_tril_, t2new_tril, tau, buf, 0, a)
p0 += a+1
time0 = logger.timer_debug1(self, 'vvvv %d'%a, *time0)
handler.join()
return t2new_tril
def _make_eris(mp, mo_coeff=None, verbose=None):
log = logger.new_logger(mp, verbose)
time0 = (time.clock(), time.time())
log.debug('transform (ia|jb) outcore')
mol = mp.mol
nocc = mp.nocc
nmo = mp.nmo
nvir = nmo - nocc
eris = mp2._ChemistsERIs(mp, mo_coeff)
nao = eris.mo_coeff.shape[0]
assert (nvir <= nao)
orbo = eris.mo_coeff[:, :nocc]
orbv = numpy.asarray(eris.mo_coeff[:, nocc:], order='F')
eris.feri = lib.H5TmpFile()
int2e = mol._add_suffix('int2e')
ao2mopt = _ao2mo.AO2MOpt(mol, int2e, 'CVHFnr_schwarz_cond',
'CVHFsetnr_direct_scf')
fint = gto.moleintor.getints4c
ntasks = mpi.pool.size
olocs = [_task_location(nocc, task_id) for task_id in range(ntasks)]
oloc0, oloc1 = olocs[rank]
nocc_seg = oloc1 - oloc0
log.debug2('olocs %s', olocs)
ao_loc = mol.ao_loc_nr()
task_sh_locs = lib.misc._balanced_partition(ao_loc, ntasks)
log.debug2('task_sh_locs %s', task_sh_locs)
ao_sh0 = task_sh_locs[rank]
ao_sh1 = task_sh_locs[rank + 1]
ao_loc0 = ao_loc[ao_sh0]
ao_loc1 = ao_loc[ao_sh1]
nao_seg = ao_loc1 - ao_loc0
orbo_seg = orbo[ao_loc0:ao_loc1]
mem_now = lib.current_memory()[0]
max_memory = max(0, mp.max_memory - mem_now)
dmax = numpy.sqrt(max_memory * .9e6 / 8 / ((nao + nocc) *
(nao_seg + nocc)))
dmax = min(nao // 4 + 2, max(BLKMIN, min(comm.allgather(dmax))))
sh_ranges = ao2mo.outcore.balance_partition(ao_loc, dmax)
sh_ranges = comm.bcast(sh_ranges)
dmax = max(x[2] for x in sh_ranges)
eribuf = numpy.empty((nao, dmax, dmax, nao_seg))
ftmp = lib.H5TmpFile()
log.debug('max_memory %s MB (dmax = %s) required disk space %g MB',
max_memory, dmax,
nocc * nocc_seg * (nao * (nao + dmax) / 2 + nvir**2) * 8 / 1e6)
def save(count, tmp_xo):
di, dj = tmp_xo.shape[2:4]
tmp_xo = [tmp_xo[p0:p1] for p0, p1 in olocs]
tmp_xo = mpi.alltoall(tmp_xo, split_recvbuf=True)
tmp_xo = sum(tmp_xo).reshape(nocc_seg, nocc, di, dj)
ftmp[str(count) + 'b'] = tmp_xo
tmp_ox = mpi.alltoall([tmp_xo[:, p0:p1] for p0, p1 in olocs],
split_recvbuf=True)
tmp_ox = [
tmp_ox[i].reshape(p1 - p0, nocc_seg, di, dj)
for i, (p0, p1) in enumerate(olocs)
]
ftmp[str(count) + 'a'] = numpy.vstack(tmp_ox)
jk_blk_slices = []
count = 0
time1 = time0
with lib.call_in_background(save) as bg_save:
for ip, (ish0, ish1, ni) in enumerate(sh_ranges):
for jsh0, jsh1, nj in sh_ranges[:ip + 1]:
i0, i1 = ao_loc[ish0], ao_loc[ish1]
j0, j1 = ao_loc[jsh0], ao_loc[jsh1]
jk_blk_slices.append((i0, i1, j0, j1))
shls_slice = (0, mol.nbas, ish0, ish1, jsh0, jsh1, ao_sh0,
ao_sh1)
eri = fint(int2e,
mol._atm,
mol._bas,
mol._env,
shls_slice=shls_slice,
aosym='s1',
ao_loc=ao_loc,
cintopt=ao2mopt._cintopt,
out=eribuf)
tmp_xo = lib.einsum('pi,pqrs->iqrs', orbo, eri)
tmp_xo = lib.einsum('iqrs,sl->ilqr', tmp_xo, orbo_seg)
bg_save(count, tmp_xo)
tmp_xo = None
count += 1
time1 = log.timer_debug1(
'partial ao2mo [%d:%d,%d:%d]' % (ish0, ish1, jsh0, jsh1),
*time1)
eri = eribuf = None
time1 = time0 = log.timer('mp2 ao2mo_ovov pass1', *time0)
eris.ovov = eris.feri.create_dataset('ovov', (nocc, nvir, nocc_seg, nvir),
'f8')
occblk = int(
min(nocc,
max(BLKMIN, max_memory * .9e6 / 8 / (nao**2 * nocc_seg + 1) / 5)))
def load(i0, eri):
if i0 < nocc:
i1 = min(i0 + occblk, nocc)
for k, (p0, p1, q0, q1) in enumerate(jk_blk_slices):
eri[:i1 - i0, :, p0:p1, q0:q1] = ftmp[str(k) + 'a'][i0:i1]
if p0 != q0:
dat = numpy.asarray(ftmp[str(k) + 'b'][:, i0:i1])
eri[:i1 - i0, :, q0:q1, p0:p1] = dat.transpose(1, 0, 3, 2)
def save(i0, i1, dat):
eris.ovov[i0:i1] = dat
buf_prefecth = numpy.empty((occblk, nocc_seg, nao, nao))
buf = numpy.empty_like(buf_prefecth)
bufw = numpy.empty((occblk * nocc_seg, nvir**2))
bufw1 = numpy.empty_like(bufw)
with lib.call_in_background(load) as prefetch:
with lib.call_in_background(save) as bsave:
load(0, buf_prefecth)
for i0, i1 in lib.prange(0, nocc, occblk):
buf, buf_prefecth = buf_prefecth, buf
prefetch(i1, buf_prefecth)
eri = buf[:i1 - i0].reshape((i1 - i0) * nocc_seg, nao, nao)
dat = _ao2mo.nr_e2(eri,
orbv, (0, nvir, 0, nvir),
's1',
's1',
out=bufw)
bsave(
i0, i1,
dat.reshape(i1 - i0, nocc_seg, nvir,
nvir).transpose(0, 2, 1, 3))
bufw, bufw1 = bufw1, bufw
time1 = log.timer_debug1('pass2 ao2mo [%d:%d]' % (i0, i1),
*time1)
time0 = log.timer('mp2 ao2mo_ovov pass2', *time0)
mp._eris = eris
return eris
def half_e1(mol, mo_coeffs, swapfile,
intor='cint2e_sph', aosym='s4', comp=1,
max_memory=2000, ioblk_size=IOBLK_SIZE, verbose=logger.WARN, compact=True,
ao2mopt=None):
r'''Half transform arbitrary spherical AO integrals to MO integrals
for the given two sets of orbitals
Args:
mol : :class:`Mole` object
AO integrals will be generated in terms of mol._atm, mol._bas, mol._env
mo_coeff : ndarray
Transform (ij|kl) with the same set of orbitals.
swapfile : str or h5py File or h5py Group object
To store the transformed integrals, in HDF5 format. The transformed
integrals are saved in blocks.
Kwargs
intor : str
Name of the 2-electron integral. Ref to :func:`getints_by_shell`
for the complete list of available 2-electron integral names
aosym : int or str
Permutation symmetry for the AO integrals
| 4 or '4' or 's4': 4-fold symmetry (default)
| '2ij' or 's2ij' : symmetry between i, j in (ij|kl)
| '2kl' or 's2kl' : symmetry between k, l in (ij|kl)
| 1 or '1' or 's1': no symmetry
| 'a4ij' : 4-fold symmetry with anti-symmetry between i, j in (ij|kl) (TODO)
| 'a4kl' : 4-fold symmetry with anti-symmetry between k, l in (ij|kl) (TODO)
| 'a2ij' : anti-symmetry between i, j in (ij|kl) (TODO)
| 'a2kl' : anti-symmetry between k, l in (ij|kl) (TODO)
comp : int
Components of the integrals, e.g. cint2e_ip_sph has 3 components.
verbose : int
Print level
max_memory : float or int
The maximum size of cache to use (in MB), large cache may **not**
improve performance.
ioblk_size : float or int
The block size for IO, large block size may **not** improve performance
verbose : int
Print level
compact : bool
When compact is True, depending on the four oribital sets, the
returned MO integrals has (up to 4-fold) permutation symmetry.
If it's False, the function will abandon any permutation symmetry,
and return the "plain" MO integrals
ao2mopt : :class:`AO2MOpt` object
Precomputed data to improve perfomance
Returns:
None
'''
time0 = (time.clock(), time.time())
if isinstance(verbose, logger.Logger):
log = verbose
else:
log = logger.Logger(mol.stdout, verbose)
nao = mo_coeffs[0].shape[0]
aosym = _stand_sym_code(aosym)
if aosym in ('s4', 's2ij'):
nao_pair = nao * (nao+1) // 2
else:
nao_pair = nao * nao
ijmosym, nij_pair, moij, ijshape = \
incore._conc_mos(mo_coeffs[0], mo_coeffs[1],
compact and aosym in ('s4', 's2ij'))
e1buflen, mem_words, iobuf_words, ioblk_words = \
guess_e1bufsize(max_memory, ioblk_size, nij_pair, nao_pair, comp)
# The buffer to hold AO integrals in C code, see line (@)
aobuflen = int((mem_words - iobuf_words) // (nao_pair*comp))
shranges = guess_shell_ranges(mol, (aosym in ('s4', 's2kl')), e1buflen, aobuflen)
if ao2mopt is None:
if intor == 'cint2e_sph':
ao2mopt = _ao2mo.AO2MOpt(mol, intor, 'CVHFnr_schwarz_cond',
'CVHFsetnr_direct_scf')
else:
ao2mopt = _ao2mo.AO2MOpt(mol, intor)
if isinstance(swapfile, str):
fswap = h5py.File(swapfile, 'w')
else:
fswap = swapfile
for icomp in range(comp):
g = fswap.create_group(str(icomp)) # for h5py old version
log.debug('step1: tmpfile %s %.8g MB', fswap.filename, nij_pair*nao_pair*8/1e6)
log.debug('step1: (ij,kl) = (%d,%d), mem cache %.8g MB, iobuf %.8g MB',
nij_pair, nao_pair, mem_words*8/1e6, iobuf_words*8/1e6)
# transform e1
ti0 = log.timer('Initializing ao2mo.outcore.half_e1', *time0)
nstep = len(shranges)
maxbuflen = max([x[2] for x in shranges])
bufs1 = numpy.empty((comp*maxbuflen,nao_pair))
bufs2 = numpy.empty((comp*maxbuflen,nij_pair))
for istep,sh_range in enumerate(shranges):
log.debug1('step 1 [%d/%d], AO [%d:%d], len(buf) = %d', \
istep+1, nstep, *(sh_range[:3]))
buflen = sh_range[2]
iobuf = bufs2[:comp*buflen].reshape(comp,buflen,nij_pair)
nmic = len(sh_range[3])
p0 = 0
for imic, aoshs in enumerate(sh_range[3]):
log.debug2(' fill iobuf micro [%d/%d], AO [%d:%d], len(aobuf) = %d', \
imic+1, nmic, *aoshs)
buf = bufs1[:comp*aoshs[2]] # (@)
_ao2mo.nr_e1fill(intor, aoshs, mol._atm, mol._bas, mol._env,
aosym, comp, ao2mopt, out=buf)
buf = _ao2mo.nr_e1(buf, moij, ijshape, aosym, ijmosym)
iobuf[:,p0:p0+aoshs[2]] = buf.reshape(comp,aoshs[2],-1)
p0 += aoshs[2]
ti2 = log.timer_debug1('gen AO/transform MO [%d/%d]'%(istep+1,nstep), *ti0)
e2buflen, chunks = guess_e2bufsize(ioblk_size, nij_pair, buflen)
for icomp in range(comp):
_transpose_to_h5g(fswap, '%d/%d'%(icomp,istep), iobuf[icomp],
e2buflen, None)
ti0 = log.timer_debug1('transposing to disk', *ti2)
bufs1 = bufs2 = None
if isinstance(swapfile, str):
fswap.close()
return swapfile
def _ao2mo_ovov(mp, orbo, orbv, feri, max_memory=2000, verbose=None):
time0 = (time.clock(), time.time())
log = logger.new_logger(mp, verbose)
mol = mp.mol
int2e = mol._add_suffix('int2e')
ao2mopt = _ao2mo.AO2MOpt(mol, int2e, 'CVHFnr_schwarz_cond',
'CVHFsetnr_direct_scf')
nao, nocc = orbo.shape
nvir = orbv.shape[1]
nbas = mol.nbas
assert (nvir <= nao)
ao_loc = mol.ao_loc_nr()
dmax = max(
4, min(nao / 3, numpy.sqrt(max_memory * .95e6 / 8 / (nao + nocc)**2)))
sh_ranges = ao2mo.outcore.balance_partition(ao_loc, dmax)
dmax = max(x[2] for x in sh_ranges)
eribuf = numpy.empty((nao, dmax, dmax, nao))
ftmp = lib.H5TmpFile()
log.debug('max_memory %s MB (dmax = %s) required disk space %g MB',
max_memory, dmax,
nocc**2 * (nao * (nao + dmax) / 2 + nvir**2) * 8 / 1e6)
buf_i = numpy.empty((nocc * dmax**2 * nao))
buf_li = numpy.empty((nocc**2 * dmax**2))
buf1 = numpy.empty_like(buf_li)
fint = gto.moleintor.getints4c
jk_blk_slices = []
count = 0
time1 = time0
with lib.call_in_background(ftmp.__setitem__) as save:
for ip, (ish0, ish1, ni) in enumerate(sh_ranges):
for jsh0, jsh1, nj in sh_ranges[:ip + 1]:
i0, i1 = ao_loc[ish0], ao_loc[ish1]
j0, j1 = ao_loc[jsh0], ao_loc[jsh1]
jk_blk_slices.append((i0, i1, j0, j1))
eri = fint(int2e,
mol._atm,
mol._bas,
mol._env,
shls_slice=(0, nbas, ish0, ish1, jsh0, jsh1, 0,
nbas),
aosym='s1',
ao_loc=ao_loc,
cintopt=ao2mopt._cintopt,
out=eribuf)
tmp_i = numpy.ndarray((nocc, (i1 - i0) * (j1 - j0) * nao),
buffer=buf_i)
tmp_li = numpy.ndarray((nocc, nocc * (i1 - i0) * (j1 - j0)),
buffer=buf_li)
lib.ddot(orbo.T,
eri.reshape(nao, (i1 - i0) * (j1 - j0) * nao),
c=tmp_i)
lib.ddot(orbo.T,
tmp_i.reshape(nocc * (i1 - i0) * (j1 - j0), nao).T,
c=tmp_li)
tmp_li = tmp_li.reshape(nocc, nocc, (i1 - i0), (j1 - j0))
save(str(count), tmp_li.transpose(1, 0, 2, 3))
buf_li, buf1 = buf1, buf_li
count += 1
time1 = log.timer_debug1(
'partial ao2mo [%d:%d,%d:%d]' % (ish0, ish1, jsh0, jsh1),
*time1)
time1 = time0 = log.timer('mp2 ao2mo_ovov pass1', *time0)
eri = eribuf = tmp_i = tmp_li = buf_i = buf_li = buf1 = None
chunks = (nvir, nvir)
h5dat = feri.create_dataset('ovov', (nocc * nvir, nocc * nvir),
'f8',
chunks=chunks)
# jk_where is the sorting indices for the stacked (oO|pP) integrals in pass 2
jk_where = []
aoao_idx = numpy.arange(nao * nao).reshape(nao, nao)
for i0, i1, j0, j1 in jk_blk_slices:
# idx of pP in <oO|pP>
jk_where.append(aoao_idx[i0:i1, j0:j1].ravel())
if i0 != j0:
# idx of pP in (<oO|pP>).transpose(1,0,3,2)
jk_where.append(aoao_idx[j0:j1, i0:i1].ravel())
jk_where = numpy.argsort(numpy.hstack(jk_where)).astype(numpy.int32)
orbv = numpy.asarray(orbv, order='F')
occblk = int(
min(nocc,
max(4, 250 / nocc, max_memory * .9e6 / 8 / (nao**2 * nocc) / 3)))
def load(i0, eri):
if i0 >= nocc:
return
i1 = min(i0 + occblk, nocc)
eri = eri[:(i1 - i0) * nocc]
p1 = 0
for k, jk_slice in enumerate(jk_blk_slices):
dat = numpy.asarray(ftmp[str(k)][i0:i1]).reshape((i1 - i0) * nocc,
-1)
p0, p1 = p1, p1 + dat.shape[1]
eri[:, p0:p1] = dat
if jk_slice[0] != jk_slice[2]:
dat = numpy.asarray(ftmp[str(k)][:, i0:i1])
dat = dat.transpose(1, 0, 3, 2).reshape((i1 - i0) * nocc, -1)
p0, p1 = p1, p1 + dat.shape[1]
eri[:, p0:p1] = dat
def save(i0, i1, dat):
for i in range(i0, i1):
h5dat[i * nvir:(i + 1) * nvir] = dat[i - i0].reshape(
nvir, nocc * nvir)
buf_prefecth = numpy.empty((occblk * nocc, nao**2))
buf = numpy.empty_like(buf_prefecth)
buf1 = numpy.empty_like(buf_prefecth)
with lib.call_in_background(load) as prefetch:
with lib.call_in_background(save) as bsave:
load(0, buf_prefecth)
for i0, i1 in lib.prange(0, nocc, occblk):
buf, buf_prefecth = buf_prefecth, buf
eri = buf[:(i1 - i0) * nocc]
prefetch(i1, buf_prefecth)
idx = numpy.arange(eri.shape[0], dtype=numpy.int32)
dat = lib.take_2d(eri, idx, jk_where, out=buf1)
dat = _ao2mo.nr_e2(dat,
orbv, (0, nvir, 0, nvir),
's1',
's1',
out=eri)
bsave(
i0, i1,
dat.reshape(i1 - i0, nocc, nvir,
nvir).transpose(0, 2, 1, 3))
time1 = log.timer_debug1('pass2 ao2mo [%d:%d]' % (i0, i1),
*time1)
time0 = log.timer('mp2 ao2mo_ovov pass2', *time0)
return h5dat
