Esempio n. 1
0
def tran_rdm2(t1, t2):
    nocc, nvir = t1.shape
    tau = numpy.empty((1, nocc, nvir, nvir))
    rdm2 = numpy.empty((nocc, nocc, nvir, nvir))
    for p0 in range(nocc):
        p1 = p0 + 1
        from pyscf.cc import _ccsd
        _ccsd.make_tau(t2[p0:p1], t1[p0:p1], t1, 1, out=tau)
        theta = tau * 2 - tau.transpose(0, 1, 3, 2)
        rdm2[p0, :, :, :] = theta[0, :, :, :]
    rdm2 = numpy.einsum('ijab->iajb', rdm2)  # rdm2 as iajb
    return rdm2 * 2.0
Esempio n. 2
0
def update_amps(mycc, t1, t2, l1, l2, eris=None, saved=None):
    if saved is None:
        saved = make_intermediates(mycc, t1, t2, eris)
    time1 = time0 = time.clock(), time.time()
    log = logger.Logger(mycc.stdout, mycc.verbose)
    nocc, nvir = t1.shape
    nov = nocc * nvir
    foo = eris.fock[:nocc, :nocc]
    fov = eris.fock[:nocc, nocc:]
    fvv = eris.fock[:nocc, :nocc]

    #:mba = numpy.einsum('klca,klcb->ba', l2, t2*2-t2.transpose(0,1,3,2))
    #:mij = numpy.einsum('ikcd,jkcd->ij', l2, t2*2-t2.transpose(0,1,3,2))
    #:theta = t2*2 - t2.transpose(0,1,3,2)
    theta = _ccsd.make_0132(t2, t2, 2, -1)
    mba = lib.dot(theta.reshape(-1, nvir).T, l2.reshape(-1, nvir))
    mij = lib.dot(l2.reshape(nocc, -1), theta.reshape(nocc, -1).T)
    theta = None
    mba1 = numpy.einsum("jc,jb->bc", l1, t1) + mba
    mij1 = numpy.einsum("kb,jb->kj", l1, t1) + mij
    mia1 = (
        t1
        + numpy.einsum("kc,jkbc->jb", l1, t2) * 2
        - numpy.einsum("kc,jkcb->jb", l1, t2)
        - reduce(numpy.dot, (t1, l1.T, t1))
        - numpy.einsum("bd,jd->jb", mba, t1)
        - numpy.einsum("lj,lb->jb", mij, t1)
    )

    tmp = mycc.add_wvvVV(numpy.zeros_like(l1), l2, eris)
    l2new = numpy.empty((nocc, nocc, nvir, nvir))
    ij = 0
    for i in range(nocc):
        for j in range(i):
            tmp1 = tmp[ij] * 0.5  # *.5 because of l2+l2.transpose(1,0,3,2) later
            l2new[i, j] = tmp1
            l2new[j, i] = tmp1.T
            ij += 1
        l2new[i, i] = tmp[ij] * 0.5
        ij += 1
    l1new = numpy.einsum("ijab,jb->ia", l2new, t1) * 4 - numpy.einsum("jiab,jb->ia", l2new, t1) * 2
    tmp = tmp1 = None

    l1new += fov
    l1new += numpy.einsum("ib,ba->ia", l1, saved.w1)
    l1new -= numpy.einsum("ja,ij->ia", l1, saved.w2)
    l1new -= numpy.einsum("ik,ka->ia", mij, saved.w4)
    l1new -= numpy.einsum("ca,ic->ia", mba, saved.w4)
    l1new += numpy.einsum("ijab,bj->ia", l2, saved.w3) * 2
    l1new -= numpy.einsum("ijba,bj->ia", l2, saved.w3)

    l2new += numpy.einsum("ia,jb->ijab", l1, saved.w4)
    #:l2new += numpy.einsum('jibc,ca->jiba', l2, saved.w1)
    #:l2new -= numpy.einsum('kiba,jk->jiba', l2, saved.w2)
    lib.dot(l2.reshape(-1, nvir), saved.w1, 1, l2new.reshape(-1, nvir), 1)
    lib.dot(saved.w2, l2.reshape(nocc, -1), -1, l2new.reshape(nocc, -1), 1)

    eris_ooov = _cp(eris.ooov)
    l1new -= numpy.einsum("jkia,kj->ia", eris_ooov, mij1) * 2
    l1new += numpy.einsum("ikja,kj->ia", eris_ooov, mij1)
    #:l2new -= numpy.einsum('ka,kijb->jiba', l1, eris_ooov)
    lib.dot(_cp(eris_ooov.transpose(0, 2, 1, 3).reshape(nocc, -1)).T, l1, -1, l2new.reshape(-1, nvir), 1)
    eris_ooov = None

    tau = _ccsd.make_tau(t2, t1, t1)
    #:l2tau = numpy.einsum('ijcd,klcd->ijkl', l2, tau)
    l2tau = lib.dot(l2.reshape(nocc ** 2, -1), tau.reshape(nocc ** 2, -1).T).reshape((nocc,) * 4)
    tau = None
    #:l2t1 = numpy.einsum('jidc,kc->ijkd', l2, t1)
    l2t1 = lib.dot(l2.reshape(-1, nvir), t1.T).reshape(nocc, nocc, nvir, nocc)
    l2t1 = _cp(l2t1.transpose(1, 0, 3, 2))

    max_memory = mycc.max_memory - lib.current_memory()[0]
    unit = max(nvir ** 3 * 2 + nocc * nvir ** 2, nocc * nvir ** 2 * 5)
    blksize = min(nocc, max(ccsd.BLKMIN, int(max_memory * 0.95e6 / 8 / unit)))
    log.debug1(
        "block size = %d, nocc = %d is divided into %d blocks", blksize, nocc, int((nocc + blksize - 1) / blksize)
    )
    for p0, p1 in prange(0, nocc, blksize):
        eris_ovvv = _cp(eris.ovvv[p0:p1])
        eris_ovvv = lib.unpack_tril(eris_ovvv.reshape((p1 - p0) * nvir, -1))
        eris_ovvv = eris_ovvv.reshape(p1 - p0, nvir, nvir, nvir)

        l1new[p0:p1] += numpy.einsum("iabc,bc->ia", eris_ovvv, mba1) * 2
        l1new[p0:p1] -= numpy.einsum("ibca,bc->ia", eris_ovvv, mba1)
        #:l2new[p0:p1] += numpy.einsum('ic,jbac->jiba', l1, eris_ovvv)
        tmp = lib.dot(l1, eris_ovvv.reshape(-1, nvir).T)
        l2new[p0:p1] += tmp.reshape(nocc, -1, nvir, nvir).transpose(1, 0, 2, 3)
        tmp = None
        m4buf = numpy.empty((blksize, nocc, nvir, nvir))
        eris_ovvv = _cp(eris_ovvv.transpose(0, 2, 1, 3).reshape(-1, nvir ** 2))
        for j0, j1 in prange(0, nocc, blksize):
            #:m4 = numpy.einsum('ijkd,kadb->ijab', l2t1[j0:j1,:,p0:p1], eris_ovvv)
            m4 = m4buf[: j1 - j0]
            lib.dot(_cp(l2t1[j0:j1, :, p0:p1].reshape((j1 - j0) * nocc, -1)), eris_ovvv, 1, m4.reshape(-1, nvir ** 2))
            l2new[j0:j1] -= m4
            l1new[j0:j1] -= numpy.einsum("ijab,jb->ia", m4, t1) * 2
            l1new -= numpy.einsum("ijab,ia->jb", m4, t1[j0:j1]) * 2
            l1new += numpy.einsum("jiab,jb->ia", m4, t1[j0:j1])
            l1new[j0:j1] += numpy.einsum("jiab,ia->jb", m4, t1)
        eris_ovvv = m4buf = m4 = None
        # ==== mem usage nvir**3*2 + nocc*nvir**2

        eris_ovov = _cp(eris.ovov[p0:p1])
        l1new[p0:p1] += numpy.einsum("jb,iajb->ia", l1, eris_ovov) * 2
        for i in range(p1 - p0):
            l2new[p0 + i] += eris_ovov[i].transpose(1, 0, 2) * 0.5
        #:l2new[p0:p1] -= numpy.einsum('icjb,ca->ijab', eris_ovov, mba1)
        #:l2new[p0:p1] -= numpy.einsum('jbka,ik->jiba', eris_ovov, mij1)
        tmp = numpy.empty((nocc, nvir, nvir))
        for j in range(p0, p1):
            lib.dot(eris_ovov[j - p0].reshape(nvir, -1).T, mba1, 1, tmp.reshape(-1, nvir))
            l2new[j] -= tmp.transpose(0, 2, 1)
            lib.dot(
                mij1, _cp(eris_ovov[j - p0].transpose(1, 0, 2).reshape(nocc, -1)), -1, l2new[j].reshape(nocc, -1), 1
            )
        tmp = None
        l1new[p0:p1] += numpy.einsum("iajb,jb->ia", eris_ovov, mia1) * 2
        l1new[p0:p1] -= numpy.einsum("ibja,jb->ia", eris_ovov, mia1)
        m4buf = numpy.empty((blksize, nocc, nvir, nvir))
        for j0, j1 in prange(0, nocc, blksize):
            #:m4 = numpy.einsum('kalb,ijkl->ijab', eris_ovov, l2tau[j0:j1,:,p0:p1])
            m4 = m4buf[: j1 - j0]
            lib.dot(
                l2tau[j0:j1, :, p0:p1].reshape((j1 - j0) * nocc, -1).copy(),
                _cp(eris_ovov.transpose(0, 2, 1, 3).reshape(-1, nvir ** 2)),
                0.5,
                m4.reshape(-1, nvir ** 2),
            )
            l2new[j0:j1] += m4
            l1new[j0:j1] += numpy.einsum("ijab,jb->ia", m4, t1) * 4
            l1new[j0:j1] -= numpy.einsum("ijba,jb->ia", m4, t1) * 2
        eris_ovov = m4buf = m4 = None
        # ==== mem usage nocc*nvir**2 * 3

        eris_oovv = _cp(eris.oovv[p0:p1])
        l1new[p0:p1] -= numpy.einsum("jb,ijba->ia", l1, eris_oovv)
        eris_oovv = None

        saved_wooov = _cp(saved.wooov[p0:p1])
        #:l1new[p0:p1] -= numpy.einsum('jkca,ijkc->ia', l2, saved_wooov)
        l1new[p0:p1] -= lib.dot(saved_wooov.reshape(p1 - p0, -1), l2.reshape(-1, nvir))
        saved_wovvv = _cp(saved.wovvv[p0:p1])
        #:l1new += numpy.einsum('kibc,kabc->ia', l2[p0:p1], saved_wovvv)
        for j in range(p1 - p0):
            lib.dot(l2[p0 + j].reshape(nocc, -1), saved_wovvv[j].reshape(nvir, -1).T, 1, l1new, 1)
        saved_wooov = saved_wovvv = None
        # ==== mem usage nvir**3 + nocc**2*nvir

        saved_wOvOv = _cp(saved.wOvOv[p0:p1])
        tmp_ovov = _cp(saved.wOVov[p0:p1]) * 2
        tmp_ovov += saved_wOvOv
        tmp_ovov = lib.transpose(tmp_ovov.reshape(-1, nov)).reshape(nocc, nvir, -1, nvir)
        tmp1 = numpy.empty((p1 - p0, nvir, nocc, nvir))
        tmp = numpy.empty((blksize, nvir, nocc, nvir))
        for j0, j1 in prange(0, nocc, blksize):
            #:tmp = l2[j0:j1].transpose(0,2,1,3) - l2[j0:j1].transpose(0,3,1,2)*.5
            #:l2new[p0:p1] += numpy.einsum('kcia,kcjb->jiba', tmp, tmp_ovov[j0:j1])
            for i in range(j1 - j0):
                tmp[i] = -0.5 * l2[j0 + i].transpose(2, 0, 1)
                tmp[i] += l2[j0 + i].transpose(1, 0, 2)
            lib.dot(
                tmp_ovov[j0:j1].reshape((j1 - j0) * nvir, -1).T,
                tmp[: j1 - j0].reshape((j1 - j0) * nvir, -1),
                1,
                tmp1.reshape(-1, nov),
            )
            l2new[p0:p1] += tmp1.transpose(0, 2, 1, 3)
        tmp = tmp1 = tmp_ovov = None
        # ==== mem usage nocc*nvir**2 * 5

        #:tmp = numpy.einsum('jkca,ibkc->ijab', l2, saved_wOvOv)
        tmp = numpy.empty((p1 - p0, nvir, nvir))
        for j in range(nocc):
            lib.dot(saved_wOvOv.reshape(-1, nov), l2[j].reshape(nov, -1), 1, tmp.reshape(-1, nvir))
            l2new[p0:p1, j] += tmp.transpose(0, 2, 1)
            l2new[p0:p1, j] += tmp * 0.5
        saved_wOvOv = tmp = None

        saved_woooo = _cp(saved.woooo[p0:p1])
        #:m3 = numpy.einsum('klab,ijkl->ijab', l2, saved_woooo)
        m3 = lib.dot(saved_woooo.reshape(-1, nocc ** 2), l2.reshape(nocc ** 2, -1), 0.5).reshape(-1, nocc, nvir, nvir)
        l2new[p0:p1] += m3
        l1new[p0:p1] += numpy.einsum("ijab,jb->ia", m3, t1) * 4
        l1new[p0:p1] -= numpy.einsum("ijba,jb->ia", m3, t1) * 2
        saved_woooo = m3 = None
        time1 = log.timer_debug1("lambda pass [%d:%d]" % (p0, p1), *time1)

    mo_e = eris.fock.diagonal()
    eia = lib.direct_sum("i-a->ia", mo_e[:nocc], mo_e[nocc:])
    l1new /= eia
    l1new += l1

    #    l2new = l2new + l2new.transpose(1,0,3,2)
    #    l2new /= lib.direct_sum('ia+jb->ijab', eia, eia)
    #    l2new += l2
    ij = 0
    for i in range(nocc):
        for j in range(i):
            dab = lib.direct_sum("a+b->ab", eia[i], eia[j])
            tmp = (l2new[i, j] + l2new[j, i].T) / dab + l2[i, j]
            l2new[i, j] = tmp
            l2new[j, i] = tmp.T
            ij += 1
        dab = lib.direct_sum("a+b->ab", eia[i], eia[i])
        l2new[i, i] = (l2new[i, i] + l2new[i, i].T) / dab + l2[i, i]
        ij += 1

    time0 = log.timer_debug1("update l1 l2", *time0)
    return l1new, l2new
Esempio n. 3
0
def make_intermediates(mycc, t1, t2, eris, max_memory=2000):
    log = logger.Logger(mycc.stdout, mycc.verbose)
    nocc, nvir = t1.shape
    nov = nocc * nvir
    foo = eris.fock[:nocc, :nocc]
    fov = eris.fock[:nocc, nocc:]
    fvv = eris.fock[nocc:, nocc:]

    class _Saved(object):
        def __init__(self):
            self._tmpfile = tempfile.NamedTemporaryFile()
            self.ftmp = h5py.File(self._tmpfile.name)

        def __del__(self):
            if hasattr(self, 'ftmp'):
                self.ftmp.close()
                self._tmpfile = None

    saved = _Saved()
    saved.woooo = saved.ftmp.create_dataset('woooo', (nocc, nocc, nocc, nocc),
                                            'f8')
    saved.wooov = saved.ftmp.create_dataset('wooov', (nocc, nocc, nocc, nvir),
                                            'f8')
    saved.wOVov = saved.ftmp.create_dataset('wOVov', (nocc, nvir, nocc, nvir),
                                            'f8')
    saved.wOvOv = saved.ftmp.create_dataset('wOvOv', (nocc, nvir, nocc, nvir),
                                            'f8')
    saved.wovvv = saved.ftmp.create_dataset('wovvv', (nocc, nvir, nvir, nvir),
                                            'f8')

    # As we don't have l2 in memory, hold tau temporarily in memory
    w1 = fvv - numpy.einsum('ja,jb->ba', fov, t1)
    w2 = foo + numpy.einsum('ib,jb->ij', fov, t1)
    w3 = numpy.einsum('kc,jkbc->bj', fov, t2) * 2 + fov.T
    w3 -= numpy.einsum('kc,kjbc->bj', fov, t2)
    w3 += reduce(numpy.dot, (t1.T, fov, t1.T))
    w4 = fov.copy()

    _tmpfile = tempfile.NamedTemporaryFile()
    fswap = h5py.File(_tmpfile.name)

    time1 = time.clock(), time.time()
    max_memory = max_memory - lib.current_memory()[0]
    unit = max(nocc * nvir**2 * 4 + nvir**3 * 2, nvir**3 * 3 + nocc * nvir**2,
               nocc * nvir**2 * 6 + nocc**2 * nvir + nocc**3 + nocc**2 * nvir)
    blksize = max(ccsd.BLKMIN, int(max_memory * .95e6 / 8 / unit))
    log.debug1(
        'ccsd lambda make_intermediates: block size = %d, nocc = %d in %d blocks',
        blksize, nocc, int((nocc + blksize - 1) // blksize))
    for istep, (p0, p1) in enumerate(prange(0, nocc, blksize)):
        eris_ovvv = _cp(eris.ovvv[p0:p1])
        eris_ovvv = _ccsd.unpack_tril(eris_ovvv.reshape((p1 - p0) * nvir, -1))
        eris_ovvv = eris_ovvv.reshape(p1 - p0, nvir, nvir, nvir)
        w1 += numpy.einsum('jcba,jc->ba', eris_ovvv, t1[p0:p1] * 2)
        w1 -= numpy.einsum('jabc,jc->ba', eris_ovvv, t1[p0:p1])
        #:w3 += numpy.einsum('kdcb,kjdc->bj', eris_ovvv, theta)
        for i in range(p1 - p0):
            theta = t2[p0 + i] * 2
            theta -= t2[p0 + i].transpose(0, 2, 1)
            w3 += lib.dot(eris_ovvv[i].reshape(-1, nvir).T,
                          _cp(theta.reshape(nocc, -1)).T)
        theta = None
        #:wOVov = numpy.einsum('jbcd,kd->jbkc', eris_ovvv, t1)
        #:wOvOv = numpy.einsum('jdcb,kd->jbkc', eris_ovvv, -t1)
        wOVov = lib.dot(eris_ovvv.reshape(-1, nvir),
                        t1.T).reshape(-1, nvir, nvir,
                                      nocc).transpose(0, 1, 3, 2).copy()
        g2ovvv = _cp(eris_ovvv.transpose(0, 2, 3, 1))
        wOvOv = lib.dot(g2ovvv.reshape(-1, nvir),
                        -t1.T).reshape(-1, nvir, nvir,
                                       nocc).transpose(0, 1, 3, 2).copy()
        for i in range(p1 - p0):
            g2ovvv[i] *= 2
            g2ovvv[i] -= eris_ovvv[i].transpose(1, 0, 2)
        wooov = numpy.empty((p1 - p0, nocc, nocc, nvir))
        woooo = numpy.empty((p1 - p0, nocc, nocc, nocc))
        eris_ovov = _cp(_cp(eris.ovov[p0:p1]).transpose(0, 2, 1, 3))
        for j0, j1 in prange(0, nocc, blksize):
            tau = _ccsd.make_tau(t2[j0:j1], t1[j0:j1], t1)
            #:wooov[:,j0:j1] = numpy.einsum('icbd,jkbd->ijkc', g2ovvv, tau)
            #:woooo[:,:,j0:j1] = numpy.einsum('icjd,klcd->ijkl', eris_ovov, tau)
            tmp = lib.dot(g2ovvv.reshape(-1, nvir**2),
                          tau.reshape(-1, nvir**2).T)
            wooov[:, j0:j1] = tmp.reshape(-1, nvir, j1 - j0,
                                          nocc).transpose(0, 2, 3, 1)
            woooo[:, :, j0:j1] = lib.dot(eris_ovov.reshape(-1, nvir**2),
                                         tau.reshape(-1, nvir**2).T).reshape(
                                             -1, nocc, j1 - j0, nocc)
        eris_ovov = eris_ovvv = g2ovvv = tau = tmp = None
        #==== mem usage nocc*nvir**2*2 + nocc**2*nvir + nocc**3 + nvir**3*2 + nocc*nvir**2*2

        eris_ooov = _cp(eris.ooov[p0:p1])
        w2[p0:p1] += numpy.einsum('ijkb,kb->ij', eris_ooov, t1) * 2
        w2 -= numpy.einsum('kjib,kb->ij', eris_ooov, t1[p0:p1])
        #:w3 -= numpy.einsum('kjlc,klbc->bj', eris_ooov, theta)
        for i in range(p1 - p0):
            theta = t2[p0 + i].transpose(0, 2, 1) * 2
            theta -= t2[p0 + i]
            w3 -= lib.dot(
                theta.reshape(-1, nvir).T, eris_ooov[i].reshape(nocc, -1).T)
        theta = None
        #:woooo += numpy.einsum('ikjc,lc->ijkl', eris_ooov, t1)
        #:wOvOv += numpy.einsum('jklb,lc->jbkc', eris_ooov, t1)
        woooo += lib.dot(eris_ooov.reshape(-1, nvir), t1.T).reshape(
            (-1, nocc, nocc, nocc)).transpose(0, 2, 1, 3)
        for i in range(p1 - p0):
            lib.dot(
                _cp(eris_ooov[i].transpose(2, 0, 1)).reshape(-1, nocc), t1, 1,
                wOvOv[i].reshape(-1, nvir), 1)
            wooov[i] += eris_ooov[i].transpose(1, 0, 2) * 2
            wooov[i] -= eris_ooov[i]

        eris_ovoo = _cp(eris.ovoo[p0:p1])
        #:woooo += numpy.einsum('icjl,kc->ijkl', eris_ovoo, t1)
        #:wOVov += numpy.einsum('jbkl,lc->jbkc', eris_ovoo, -t1)
        for i in range(p1 - p0):
            woooo[i] += lib.dot(t1, eris_ovoo[i].reshape(nvir, -1)).reshape(
                (nocc, ) * 3).transpose(1, 0, 2)
        lib.dot(eris_ovoo.reshape(-1, nocc), t1, -1, wOVov.reshape(-1, nvir),
                1)
        #:wooov -= numpy.einsum('iblj,klbc->ijkc', eris_ovoo*1.5, t2)
        tmp_ovoo = _cp(-eris_ovoo.transpose(0, 2, 3, 1)).reshape(-1, nov)
        for j in range(nocc):
            wooov[:, :, j] += lib.dot(tmp_ovoo, t2[j].reshape(-1, nvir),
                                      1.5).reshape(-1, nocc, nvir)
        #:g2ooov = eris_ooov * 2 - eris_ovoo.transpose(0,3,2,1)
        g2ooov, tmp_ovoo = tmp_ovoo.reshape(p1 - p0, nocc, nocc, nvir), None
        g2ooov += eris_ooov * 2
        thetabuf = numpy.empty((blksize, nvir, nocc, nvir))
        vikjc = numpy.empty((p1 - p0, nocc, blksize, nvir))
        for j0, j1 in prange(0, nocc, blksize):
            theta = thetabuf[:j1 - j0]
            for i in range(j1 - j0):
                theta[i] = t2[j0 + i].transpose(
                    1, 0, 2) * 2 - t2[j0 + i].transpose(2, 0, 1)
            #:vikjc = numpy.einsum('iklb,jlcb->ikjc', g2ooov, theta)
            if j1 == j0 + blksize:
                lib.dot(g2ooov.reshape(-1, nov),
                        _cp(theta.reshape(-1, nov)).T, 1,
                        vikjc.reshape((p1 - p0) * nocc, -1), 0)
            else:
                vikjc = lib.dot(g2ooov.reshape(-1, nov),
                                _cp(theta.reshape(-1, nov)).T)
                vikjc = vikjc.reshape(p1 - p0, nocc, j1 - j0, nvir)
            wooov[:, j0:j1, :] += vikjc.transpose(0, 2, 1, 3)
            wooov[:, :, j0:j1] -= vikjc * .5
        eris_ooov = eris_ovoo = g2ooov = vikjc = theta = thetabuf = None
        #==== mem usage nocc*nvir**2*3 + nocc**2*nvir + nocc**3 + nocc*nvir**2 + nocc**2*nvir*3

        eris_ovov = _cp(eris.ovov[p0:p1])
        g2ovov = eris_ovov * 2
        g2ovov -= eris_ovov.transpose(0, 3, 2, 1)
        tmpw4 = numpy.einsum('kcld,ld->kc', g2ovov, t1)
        #:w1 -= numpy.einsum('kcja,kjcb->ba', g2ovov, t2[p0:p1])
        w1 -= lib.dot(t2[p0:p1].reshape(-1, nvir).T,
                      _cp(g2ovov.transpose(0, 2, 1, 3).reshape(-1, nvir)))
        w1 -= numpy.einsum('ja,jb->ba', tmpw4, t1[p0:p1])
        #:w2[p0:p1] += numpy.einsum('ibkc,jkbc->ij', g2ovov, t2)
        w2[p0:p1] += lib.dot(
            _cp(g2ovov.transpose(0, 2, 1, 3)).reshape(p1 - p0, -1),
            t2.reshape(nocc, -1).T)
        w2[p0:p1] += numpy.einsum('ib,jb->ij', tmpw4, t1)
        w3 += reduce(numpy.dot, (t1[p0:p1].T, tmpw4, t1.T))
        w4[p0:p1] += tmpw4
        vOVov = numpy.empty((nocc, nvir, p1 - p0, nvir))
        #:vOVov += numpy.einsum('jbld,klcd->kcjb', g2ovov, t2)
        #:vOVov -= numpy.einsum('jbld,kldc->kcjb', eris_ovov, t2)
        for j in range(nocc):
            lib.dot(
                _cp(t2[j].transpose(1, 0, 2)).reshape(-1, nov),
                g2ovov.reshape(-1, nov).T, 1, vOVov[j].reshape(nvir, -1))
            lib.dot(t2[j].reshape(nov, -1).T,
                    eris_ovov.reshape(-1, nov).T, -1,
                    vOVov[j].reshape(nvir, -1), 1)
        vOVov = lib.transpose(vOVov.reshape(nov, -1)).reshape(
            p1 - p0, nvir, nocc, nvir)
        vOVov += eris_ovov
        g2ovov = tmp = tmpw4 = None
        #==== mem usage nocc*nvir**2*4 + nocc**2*nvir + nocc**3 + nocc*nvir**2

        #:tmp = numpy.einsum('jbld,kd->jlbk', eris_ovov, t1)
        #:wOVov -= numpy.einsum('jlbk,lc->jbkc', tmp, t1)
        #:tmp = numpy.einsum('jdlb,kd->jlbk', eris_ovov, t1)
        #:wOvOv += numpy.einsum('jlbk,lc->jbkc', tmp, t1)
        tmp = numpy.empty((nocc, nvir, nocc))
        for j in range(p1 - p0):
            lib.dot(
                _cp(eris_ovov[j].transpose(1, 0, 2)).reshape(-1, nvir), t1.T,
                1, tmp.reshape(-1, nocc))
            lib.dot(
                tmp.reshape(nocc, -1).T, t1, -1, wOVov[j].reshape(-1, nvir), 1)
            lib.dot(eris_ovov[j].reshape(nvir, -1).T, t1.T, 1,
                    tmp.reshape(-1, nocc))
            lib.dot(
                tmp.reshape(nocc, -1).T, t1, 1, wOvOv[j].reshape(-1, nvir), 1)
        tmp = None

        #:vOvOv = numpy.einsum('jdlb,kldc->kcjb', eris_ovov, t2)
        ovovtmp = _cp(eris_ovov.transpose(0, 3, 2, 1)).reshape(-1, nov)
        vOvOv = numpy.empty((nocc, nvir, p1 - p0, nvir))
        for j in range(nocc):
            lib.dot(t2[j].reshape(-1, nvir).T, ovovtmp.T, 1,
                    vOvOv[j].reshape(nvir, -1))
        ovovtmp = eris_ovov = None
        vOvOv = lib.transpose(vOvOv.reshape(nov, -1)).reshape(
            p1 - p0, nvir, nocc, nvir)
        vOvOv -= _cp(eris.oovv[p0:p1]).transpose(0, 3, 1, 2)
        wOVov += vOVov
        wOvOv += vOvOv
        saved.wOVov[p0:p1] = wOVov
        saved.wOvOv[p0:p1] = wOvOv
        wOVov = wOvOv = None
        #==== mem usage nocc*nvir**2*6 + nocc**2*nvir + nocc**3 + nocc**2*nvir

        ov1 = vOvOv * 2 + vOVov
        #:wooov -= numpy.einsum('ibkc,jb->ijkc', ov1, t1)
        for i in range(p1 - p0):
            lib.dot(t1, ov1[i].reshape(nvir, -1), -1,
                    wooov[i].reshape(nocc, -1), 1)
        ov1 = lib.transpose(ov1.reshape(-1, nov))
        fswap['2vOvOv/%d' % istep] = ov1.reshape(nocc, nvir, -1, nvir)
        ov1 = None
        ov2 = vOVov * 2 + vOvOv
        w3 += numpy.einsum('kcjb,kc->bj', ov2, t1[p0:p1])
        #:wooov += numpy.einsum('ibjc,kb->ijkc', ov2, t1)
        for i in range(p1 - p0):
            wooov[i] += lib.dot(t1, ov2[i].reshape(nvir, -1)).reshape(
                nocc, nocc, nvir).transpose(1, 0, 2)
        ov2 = lib.transpose(ov2.reshape(-1, nov))
        fswap['2vovOV/%d' % istep] = ov2.reshape(nocc, nvir, -1, nvir)
        vOVov = vOvOv = None
        ov2 = None
        #==== mem usage nocc*nvir**2*5 + nocc**2*nvir + nocc**3

        woooo += _cp(eris.oooo[p0:p1]).transpose(0, 2, 1, 3)
        saved.woooo[p0:p1] = woooo
        saved.wooov[p0:p1] = wooov
        woooo = wooov = None
        time1 = log.timer_debug1('pass1 [%d:%d]' % (p0, p1), *time1)

    w3 += numpy.einsum('bc,jc->bj', w1, t1)
    w3 -= numpy.einsum('kj,kb->bj', w2, t1)

    for p0, p1 in prange(0, nocc, blksize):
        eris_ooov = _cp(eris.ooov[p0:p1])
        g2ooov = eris_ooov * 2
        g2ooov -= eris_ooov.transpose(0, 2, 1, 3)
        #:tmp = numpy.einsum('kjla,jb->kabl', g2ooov, t1)
        #:wovvv = numpy.einsum('kabl,lc->kabc', tmp, t1)
        #:wovvv += numpy.einsum('kjla,jlbc->kabc', g2ooov, t2)
        tmp = lib.dot(
            _cp(g2ooov.transpose(1, 0, 2, 3).reshape(nocc, -1)).T,
            t1).reshape(-1, nocc, nvir, nvir).transpose(0, 2, 3, 1)
        wovvv = lib.dot(_cp(tmp.reshape(-1, nocc)),
                        t1).reshape(-1, nvir, nvir, nvir)
        wovvv += lib.dot(
            _cp(g2ooov.transpose(0, 3, 1, 2).reshape(-1, nocc**2)),
            t2.reshape(nocc**2, -1)).reshape(-1, nvir, nvir, nvir)
        tmp = g2ooov = None
        ov1 = numpy.empty((p1 - p0, nvir, nocc, nvir))
        ov2 = numpy.empty((p1 - p0, nvir, nocc, nvir))
        for istep, (j0, j1) in enumerate(prange(0, nocc, blksize)):
            ov1[:, :, j0:j1] = fswap['2vOvOv/%d' % istep][p0:p1]
            ov2[:, :, j0:j1] = fswap['2vovOV/%d' % istep][p0:p1]
        #:wovvv += numpy.einsum('kcja,jb->kabc', ov1, t1)
        #:wovvv -= numpy.einsum('kbja,jc->kabc', ov2, t1)
        wovvv += lib.dot(_cp(ov1.transpose(0, 1, 3, 2).reshape(-1, nocc)),
                         t1).reshape(-1, nvir, nvir,
                                     nvir).transpose(0, 2, 3, 1)
        wovvv -= lib.dot(_cp(ov2.transpose(0, 1, 3, 2).reshape(-1, nocc)),
                         t1).reshape(-1, nvir, nvir,
                                     nvir).transpose(0, 2, 1, 3)
        #==== mem usage nvir**3 + nocc*nvir**2*2
        eris_ooov = ov1 = ov2 = None

        for j0, j1 in prange(0, nocc, blksize):
            eris_ovvv = _cp(eris.ovvv[j0:j1])
            eris_ovvv = _ccsd.unpack_tril(
                eris_ovvv.reshape((j1 - j0) * nvir, -1))
            eris_ovvv = eris_ovvv.reshape(j1 - j0, nvir, nvir, nvir)
            #:wovvv += numpy.einsum('jabd,kjdc->kabc', eris_ovvv, t2[p0:p1,j0:j1]) * -1.5
            tmp_ovvv = numpy.empty((j1 - j0, nvir, nvir, nvir))
            for i in range(j1 - j0):
                tmp_ovvv[i] = eris_ovvv[i].transpose(1, 0, 2) * 2
            tmp = lib.dot(
                _cp(t2[p0:p1, j0:j1].transpose(0, 3, 1, 2).reshape(
                    (p1 - p0) * nvir, -1)), tmp_ovvv.reshape(-1, nvir**2),
                -1.5 / 2).reshape(-1, nvir, nvir, nvir)
            wovvv += tmp.transpose(0, 2, 3, 1)
            if p0 == j0:
                for i in range(p1 - p0):
                    tmp_ovvv[i] -= eris_ovvv[i].transpose(1, 2, 0)
                    wovvv[i] += tmp_ovvv[i]
            tmp = tmp_ovvv = None
            g2ovvv = numpy.empty((j1 - j0, nvir, nvir, nvir))
            for i in range(j1 - j0):
                g2ovvv[i] = eris_ovvv[i] * 2
                g2ovvv[i] -= eris_ovvv[i].transpose(1, 2, 0)
#==== mem usage nvir**3*3
            eris_ovvv = None
            theta = _cp(t2[p0:p1, j0:j1].transpose(0, 2, 1, 3) * 2)
            for i in range(p1 - p0):
                theta[i] -= t2[p0 + i, j0:j1].transpose(2, 0, 1)
            #:vkbca = numpy.einsum('jdca,kbjd->kbca', g2ovvv, theta)
            vkbca = lib.dot(theta.reshape((p1 - p0) * nvir, -1),
                            g2ovvv.reshape(-1, nvir * nvir)).reshape(
                                -1, nvir, nvir, nvir)
            wovvv += vkbca.transpose(0, 3, 1, 2)
            wovvv -= vkbca.transpose(0, 3, 2, 1) * .5
            #==== mem usage nvir**3*3 + nocc*nvir**2
            g2ovvv = theta = vkabc = None
        saved.wovvv[p0:p1] = wovvv
        time1 = log.timer_debug1('pass2 [%d:%d]' % (p0, p1), *time1)

    fswap.close()

    saved.w1 = w1
    saved.w2 = w2
    saved.w3 = w3
    saved.w4 = w4
    saved.ftmp.flush()
    return saved
Esempio n. 4
0
def update_amps(mycc, t1, t2, l1, l2, eris=None, saved=None, max_memory=2000):
    if saved is None:
        saved = make_intermediates(mycc, t1, t2, eris)
    time1 = time0 = time.clock(), time.time()
    log = logger.Logger(mycc.stdout, mycc.verbose)
    nocc, nvir = t1.shape
    nov = nocc * nvir
    foo = eris.fock[:nocc, :nocc]
    fov = eris.fock[:nocc, nocc:]
    fvv = eris.fock[:nocc, :nocc]

    #:mba = numpy.einsum('klca,klcb->ba', l2, t2*2-t2.transpose(0,1,3,2))
    #:mij = numpy.einsum('ikcd,jkcd->ij', l2, t2*2-t2.transpose(0,1,3,2))
    #:theta = t2*2 - t2.transpose(0,1,3,2)
    theta = _ccsd.make_0132(t2, t2, 2, -1)
    mba = lib.dot(theta.reshape(-1, nvir).T, l2.reshape(-1, nvir))
    mij = lib.dot(l2.reshape(nocc, -1), theta.reshape(nocc, -1).T)
    theta = None
    mba1 = numpy.einsum('jc,jb->bc', l1, t1) + mba
    mij1 = numpy.einsum('kb,jb->kj', l1, t1) + mij
    mia1 = (t1 + numpy.einsum('kc,jkbc->jb', l1, t2) * 2 -
            numpy.einsum('kc,jkcb->jb', l1, t2) -
            reduce(numpy.dot,
                   (t1, l1.T, t1)) - numpy.einsum('bd,jd->jb', mba, t1) -
            numpy.einsum('lj,lb->jb', mij, t1))

    tmp = mycc.add_wvvVV(numpy.zeros_like(l1), l2, eris, max_memory)
    l2new = numpy.empty((nocc, nocc, nvir, nvir))
    ij = 0
    for i in range(nocc):
        for j in range(i):
            tmp1 = tmp[ij] * .5  # *.5 because of l2+l2.transpose(1,0,3,2) later
            l2new[i, j] = tmp1
            l2new[j, i] = tmp1.T
            ij += 1
        l2new[i, i] = tmp[ij] * .5
        ij += 1
    l1new = (numpy.einsum('ijab,jb->ia', l2new, t1) * 4 -
             numpy.einsum('jiab,jb->ia', l2new, t1) * 2)
    tmp = tmp1 = None

    l1new += fov
    l1new += numpy.einsum('ib,ba->ia', l1, saved.w1)
    l1new -= numpy.einsum('ja,ij->ia', l1, saved.w2)
    l1new -= numpy.einsum('ik,ka->ia', mij, saved.w4)
    l1new -= numpy.einsum('ca,ic->ia', mba, saved.w4)
    l1new += numpy.einsum('ijab,bj->ia', l2, saved.w3) * 2
    l1new -= numpy.einsum('ijba,bj->ia', l2, saved.w3)

    l2new += numpy.einsum('ia,jb->ijab', l1, saved.w4)
    #:l2new += numpy.einsum('jibc,ca->jiba', l2, saved.w1)
    #:l2new -= numpy.einsum('kiba,jk->jiba', l2, saved.w2)
    lib.dot(l2.reshape(-1, nvir), saved.w1, 1, l2new.reshape(-1, nvir), 1)
    lib.dot(saved.w2, l2.reshape(nocc, -1), -1, l2new.reshape(nocc, -1), 1)

    eris_ooov = _cp(eris.ooov)
    l1new -= numpy.einsum('jkia,kj->ia', eris_ooov, mij1) * 2
    l1new += numpy.einsum('ikja,kj->ia', eris_ooov, mij1)
    #:l2new -= numpy.einsum('ka,kijb->jiba', l1, eris_ooov)
    lib.dot(
        _cp(eris_ooov.transpose(0, 2, 1, 3).reshape(nocc, -1)).T, l1, -1,
        l2new.reshape(-1, nvir), 1)
    eris_ooov = None

    tau = _ccsd.make_tau(t2, t1, t1)
    #:l2tau = numpy.einsum('ijcd,klcd->ijkl', l2, tau)
    l2tau = lib.dot(l2.reshape(nocc**2, -1),
                    tau.reshape(nocc**2, -1).T).reshape((nocc, ) * 4)
    tau = None
    #:l2t1 = numpy.einsum('jidc,kc->ijkd', l2, t1)
    l2t1 = lib.dot(l2.reshape(-1, nvir), t1.T).reshape(nocc, nocc, nvir, nocc)
    l2t1 = _cp(l2t1.transpose(1, 0, 3, 2))

    max_memory = max_memory - lib.current_memory()[0]
    unit = max(nvir**3 * 2 + nocc * nvir**2, nocc * nvir**2 * 5)
    blksize = min(nocc, max(ccsd.BLKMIN, int(max_memory * .95e6 / 8 / unit)))
    log.debug1('block size = %d, nocc = %d is divided into %d blocks', blksize,
               nocc, int((nocc + blksize - 1) / blksize))
    for p0, p1 in prange(0, nocc, blksize):
        eris_ovvv = _cp(eris.ovvv[p0:p1])
        eris_ovvv = _ccsd.unpack_tril(eris_ovvv.reshape((p1 - p0) * nvir, -1))
        eris_ovvv = eris_ovvv.reshape(p1 - p0, nvir, nvir, nvir)

        l1new[p0:p1] += numpy.einsum('iabc,bc->ia', eris_ovvv, mba1) * 2
        l1new[p0:p1] -= numpy.einsum('ibca,bc->ia', eris_ovvv, mba1)
        #:l2new[p0:p1] += numpy.einsum('ic,jbac->jiba', l1, eris_ovvv)
        tmp = lib.dot(l1, eris_ovvv.reshape(-1, nvir).T)
        l2new[p0:p1] += tmp.reshape(nocc, -1, nvir, nvir).transpose(1, 0, 2, 3)
        tmp = None
        m4buf = numpy.empty((blksize, nocc, nvir, nvir))
        eris_ovvv = _cp(eris_ovvv.transpose(0, 2, 1, 3).reshape(-1, nvir**2))
        for j0, j1 in prange(0, nocc, blksize):
            #:m4 = numpy.einsum('ijkd,kadb->ijab', l2t1[j0:j1,:,p0:p1], eris_ovvv)
            m4 = m4buf[:j1 - j0]
            lib.dot(_cp(l2t1[j0:j1, :, p0:p1].reshape((j1 - j0) * nocc, -1)),
                    eris_ovvv, 1, m4.reshape(-1, nvir**2))
            l2new[j0:j1] -= m4
            l1new[j0:j1] -= numpy.einsum('ijab,jb->ia', m4, t1) * 2
            l1new -= numpy.einsum('ijab,ia->jb', m4, t1[j0:j1]) * 2
            l1new += numpy.einsum('jiab,jb->ia', m4, t1[j0:j1])
            l1new[j0:j1] += numpy.einsum('jiab,ia->jb', m4, t1)
        eris_ovvv = m4buf = m4 = None
        #==== mem usage nvir**3*2 + nocc*nvir**2

        eris_ovov = _cp(eris.ovov[p0:p1])
        l1new[p0:p1] += numpy.einsum('jb,iajb->ia', l1, eris_ovov) * 2
        for i in range(p1 - p0):
            l2new[p0 + i] += eris_ovov[i].transpose(1, 0, 2) * .5
        #:l2new[p0:p1] -= numpy.einsum('icjb,ca->ijab', eris_ovov, mba1)
        #:l2new[p0:p1] -= numpy.einsum('jbka,ik->jiba', eris_ovov, mij1)
        tmp = numpy.empty((nocc, nvir, nvir))
        for j in range(p0, p1):
            lib.dot(eris_ovov[j - p0].reshape(nvir, -1).T, mba1, 1,
                    tmp.reshape(-1, nvir))
            l2new[j] -= tmp.transpose(0, 2, 1)
            lib.dot(
                mij1,
                _cp(eris_ovov[j - p0].transpose(1, 0, 2).reshape(nocc, -1)),
                -1, l2new[j].reshape(nocc, -1), 1)
        tmp = None
        l1new[p0:p1] += numpy.einsum('iajb,jb->ia', eris_ovov, mia1) * 2
        l1new[p0:p1] -= numpy.einsum('ibja,jb->ia', eris_ovov, mia1)
        m4buf = numpy.empty((blksize, nocc, nvir, nvir))
        for j0, j1 in prange(0, nocc, blksize):
            #:m4 = numpy.einsum('kalb,ijkl->ijab', eris_ovov, l2tau[j0:j1,:,p0:p1])
            m4 = m4buf[:j1 - j0]
            lib.dot(
                l2tau[j0:j1, :, p0:p1].reshape((j1 - j0) * nocc, -1).copy(),
                _cp(eris_ovov.transpose(0, 2, 1, 3).reshape(-1, nvir**2)), .5,
                m4.reshape(-1, nvir**2))
            l2new[j0:j1] += m4
            l1new[j0:j1] += numpy.einsum('ijab,jb->ia', m4, t1) * 4
            l1new[j0:j1] -= numpy.einsum('ijba,jb->ia', m4, t1) * 2
        eris_ovov = m4buf = m4 = None
        #==== mem usage nocc*nvir**2 * 3

        eris_oovv = _cp(eris.oovv[p0:p1])
        l1new[p0:p1] -= numpy.einsum('jb,ijba->ia', l1, eris_oovv)
        eris_oovv = None

        saved_wooov = _cp(saved.wooov[p0:p1])
        #:l1new[p0:p1] -= numpy.einsum('jkca,ijkc->ia', l2, saved_wooov)
        l1new[p0:p1] -= lib.dot(saved_wooov.reshape(p1 - p0, -1),
                                l2.reshape(-1, nvir))
        saved_wovvv = _cp(saved.wovvv[p0:p1])
        #:l1new += numpy.einsum('kibc,kabc->ia', l2[p0:p1], saved_wovvv)
        for j in range(p1 - p0):
            lib.dot(l2[p0 + j].reshape(nocc, -1),
                    saved_wovvv[j].reshape(nvir, -1).T, 1, l1new, 1)
        saved_wooov = saved_wovvv = None
        #==== mem usage nvir**3 + nocc**2*nvir

        saved_wOvOv = _cp(saved.wOvOv[p0:p1])
        tmp_ovov = _cp(saved.wOVov[p0:p1]) * 2
        tmp_ovov += saved_wOvOv
        tmp_ovov = lib.transpose(tmp_ovov.reshape(-1, nov)).reshape(
            nocc, nvir, -1, nvir)
        tmp1 = numpy.empty((p1 - p0, nvir, nocc, nvir))
        tmp = numpy.empty((blksize, nvir, nocc, nvir))
        for j0, j1 in prange(0, nocc, blksize):
            #:tmp = l2[j0:j1].transpose(0,2,1,3) - l2[j0:j1].transpose(0,3,1,2)*.5
            #:l2new[p0:p1] += numpy.einsum('kcia,kcjb->jiba', tmp, tmp_ovov[j0:j1])
            for i in range(j1 - j0):
                tmp[i] = -.5 * l2[j0 + i].transpose(2, 0, 1)
                tmp[i] += l2[j0 + i].transpose(1, 0, 2)
            lib.dot(tmp_ovov[j0:j1].reshape((j1 - j0) * nvir, -1).T,
                    tmp[:j1 - j0].reshape((j1 - j0) * nvir, -1), 1,
                    tmp1.reshape(-1, nov))
            l2new[p0:p1] += tmp1.transpose(0, 2, 1, 3)
        tmp = tmp1 = tmp_ovov = None
        #==== mem usage nocc*nvir**2 * 5

        #:tmp = numpy.einsum('jkca,ibkc->ijab', l2, saved_wOvOv)
        tmp = numpy.empty((p1 - p0, nvir, nvir))
        for j in range(nocc):
            lib.dot(saved_wOvOv.reshape(-1, nov), l2[j].reshape(nov, -1), 1,
                    tmp.reshape(-1, nvir))
            l2new[p0:p1, j] += tmp.transpose(0, 2, 1)
            l2new[p0:p1, j] += tmp * .5
        saved_wOvOv = tmp = None

        saved_woooo = _cp(saved.woooo[p0:p1])
        #:m3 = numpy.einsum('klab,ijkl->ijab', l2, saved_woooo)
        m3 = lib.dot(saved_woooo.reshape(-1, nocc**2), l2.reshape(nocc**2, -1),
                     .5).reshape(-1, nocc, nvir, nvir)
        l2new[p0:p1] += m3
        l1new[p0:p1] += numpy.einsum('ijab,jb->ia', m3, t1) * 4
        l1new[p0:p1] -= numpy.einsum('ijba,jb->ia', m3, t1) * 2
        saved_woooo = m3 = None
        time1 = log.timer_debug1('lambda pass [%d:%d]' % (p0, p1), *time1)

    mo_e = eris.fock.diagonal()
    eia = lib.direct_sum('i-a->ia', mo_e[:nocc], mo_e[nocc:])
    l1new /= eia
    l1new += l1

    #    l2new = l2new + l2new.transpose(1,0,3,2)
    #    l2new /= lib.direct_sum('ia+jb->ijab', eia, eia)
    #    l2new += l2
    ij = 0
    for i in range(nocc):
        for j in range(i):
            dab = lib.direct_sum('a+b->ab', eia[i], eia[j])
            tmp = (l2new[i, j] + l2new[j, i].T) / dab + l2[i, j]
            l2new[i, j] = tmp
            l2new[j, i] = tmp.T
            ij += 1
        dab = lib.direct_sum('a+b->ab', eia[i], eia[i])
        l2new[i, i] = (l2new[i, i] + l2new[i, i].T) / dab + l2[i, i]
        ij += 1

    time0 = log.timer_debug1('update l1 l2', *time0)
    return l1new, l2new
Esempio n. 5
0
def gamma2_outcore(mycc, t1, t2, l1, l2, h5fobj):
    log = logger.Logger(mycc.stdout, mycc.verbose)
    nocc, nvir = t1.shape
    nov = nocc * nvir
    nvir_pair = nvir * (nvir+1) //2
    dovov = h5fobj.create_dataset('dovov', (nocc,nvir,nocc,nvir), 'f8')
    dvvvv = h5fobj.create_dataset('dvvvv', (nvir_pair,nvir_pair), 'f8')
    doooo = h5fobj.create_dataset('doooo', (nocc,nocc,nocc,nocc), 'f8')
    doovv = h5fobj.create_dataset('doovv', (nocc,nocc,nvir,nvir), 'f8')
    dovvo = h5fobj.create_dataset('dovvo', (nocc,nvir,nvir,nocc), 'f8')
    dooov = h5fobj.create_dataset('dooov', (nocc,nocc,nocc,nvir), 'f8')

    _tmpfile = tempfile.NamedTemporaryFile(dir=lib.param.TMPDIR)
    fswap = h5py.File(_tmpfile.name)
    mOvOv = fswap.create_dataset('mOvOv', (nocc,nvir,nocc,nvir), 'f8')
    mOVov = fswap.create_dataset('mOVov', (nocc,nvir,nocc,nvir), 'f8')

    moo = numpy.empty((nocc,nocc))
    mvv = numpy.zeros((nvir,nvir))

    max_memory = mycc.max_memory - lib.current_memory()[0]
    unit = nocc*nvir**2 * 5
    blksize = max(ccsd.BLKMIN, int(max_memory*.95e6/8/unit))
    log.debug1('rdm intermediates pass 1: block size = %d, nocc = %d in %d blocks',
               blksize, nocc, int((nocc+blksize-1)/blksize))
    time1 = time.clock(), time.time()
    for istep, (p0, p1) in enumerate(prange(0, nocc, blksize)):
        #:theta = make_theta(t2[p0:p1])
        #:pOvOv = numpy.einsum('ikca,jkcb->jbia', l2, t2[p0:p1])
        #:pOVov = -numpy.einsum('ikca,jkbc->jbia', l2, t2[p0:p1])
        #:pOVov += numpy.einsum('ikac,jkbc->jbia', l2, theta)
        pOvOv = numpy.empty((nocc,p1-p0,nvir,nvir))
        pOVov = numpy.empty((nocc,p1-p0,nvir,nvir))
        t2a = numpy.empty((p1-p0,nvir,nocc,nvir))
        t2b = numpy.empty((p1-p0,nvir,nocc,nvir))
        theta = make_theta(t2[p0:p1])
        tmp = numpy.empty_like(t2a)
        for i in range(p1-p0):
            t2a[i] = t2[p0+i].transpose(2,0,1)
            t2b[i] = t2[p0+i].transpose(1,0,2)
            tmp[i] = theta[i].transpose(1,0,2)
        t2a = t2a.reshape(-1,nov)
        t2b = t2b.reshape(-1,nov)
        theta, tmp = tmp.reshape(-1,nov), None
        for i in range(nocc):
            pOvOv[i] = lib.dot(t2a, l2[i].reshape(nov,-1)).reshape(-1,nvir,nvir)
            pOVov[i] = lib.dot(t2b, l2[i].reshape(nov,-1), -1).reshape(-1,nvir,nvir)
            pOVov[i] += lib.dot(theta, _cp(l2[i].transpose(0,2,1).reshape(nov,-1))).reshape(-1,nvir,nvir)
        theta = t2a = t2b = None
        mOvOv[p0:p1] = pOvOv.transpose(1,2,0,3)
        mOVov[p0:p1] = pOVov.transpose(1,2,0,3)
        fswap['mvOvO/%d'%istep] = pOvOv.transpose(3,1,2,0)
        fswap['mvOVo/%d'%istep] = pOVov.transpose(3,1,2,0)
        moo[p0:p1] =(numpy.einsum('ljdd->jl', pOvOv) * 2
                   + numpy.einsum('ljdd->jl', pOVov))
        mvv +=(numpy.einsum('llbd->bd', pOvOv[p0:p1]) * 2
             + numpy.einsum('llbd->bd', pOVov[p0:p1]))
        pOvOv = pOVov = None
        time1 = log.timer_debug1('rdm intermediates pass1 [%d:%d]'%(p0, p1), *time1)
    mia =(numpy.einsum('kc,ikac->ia', l1, t2) * 2
        - numpy.einsum('kc,ikca->ia', l1, t2))
    mab = numpy.einsum('kc,kb->cb', l1, t1)
    mij = numpy.einsum('kc,jc->jk', l1, t1) + moo*.5

    gooov = numpy.einsum('ji,ka->jkia', moo*-.5, t1)
    max_memory = mycc.max_memory - lib.current_memory()[0]
    unit = nocc**3 + nocc**2*nvir + nocc*nvir**2*6
    blksize = max(ccsd.BLKMIN, int(max_memory*.95e6/8/unit))
    log.debug1('rdm intermediates pass 2: block size = %d, nocc = %d in %d blocks',
               blksize, nocc, int((nocc+blksize-1)/blksize))
    for p0, p1 in prange(0, nocc, blksize):
        tau = _ccsd.make_tau(t2[p0:p1], t1[p0:p1], t1)
        #:goooo = numpy.einsum('ijab,klab->klij', l2, tau)*.5
        goooo = lib.dot(tau.reshape(-1,nvir**2), l2.reshape(-1,nvir**2).T, .5)
        goooo = goooo.reshape(-1,nocc,nocc,nocc)
        h5fobj['doooo'][p0:p1] = make_theta(goooo).transpose(0,2,1,3)

        #:gooov[p0:p1] -= numpy.einsum('ib,jkba->jkia', l1, tau)
        #:gooov[p0:p1] -= numpy.einsum('jkba,ib->jkia', l2[p0:p1], t1)
        #:gooov[p0:p1] += numpy.einsum('jkil,la->jkia', goooo, t1*2)
        for i in range(p0,p1):
            gooov[i] -= lib.dot(_cp(tau[i-p0].transpose(0,2,1).reshape(-1,nvir)),
                                l1.T).reshape(nocc,nvir,nocc).transpose(0,2,1)
            gooov[i] -= lib.dot(_cp(l2[i].transpose(0,2,1).reshape(-1,nvir)),
                                t1.T).reshape(nocc,nvir,nocc).transpose(0,2,1)
        lib.dot(goooo.reshape(-1,nocc), t1, 2, gooov[p0:p1].reshape(-1,nvir), 1)

        #:goovv -= numpy.einsum('jk,ikab->ijab', mij, tau)
        goovv = numpy.einsum('ia,jb->ijab', mia[p0:p1], t1)
        for i in range(p1-p0):
            lib.dot(mij, tau[i].reshape(nocc,-1), -1, goovv[i].reshape(nocc,-1), 1)
            goovv[i] += .5 * l2[p0+i]
            goovv[i] += .5 * tau[i]
        #:goovv -= numpy.einsum('cb,ijac->ijab', mab, t2[p0:p1])
        #:goovv -= numpy.einsum('bd,ijad->ijab', mvv*.5, tau)
        lib.dot(t2[p0:p1].reshape(-1,nvir), mab, -1, goovv.reshape(-1,nvir), 1)
        lib.dot(tau.reshape(-1,nvir), mvv.T, -.5, goovv.reshape(-1,nvir), 1)
        tau = None
#==== mem usage nocc**3 + nocc*nvir**2

        pOvOv = _cp(mOvOv[p0:p1])
        pOVov = _cp(mOVov[p0:p1])
        #:gooov[p0:p1,:] += numpy.einsum('jaic,kc->jkia', pOvOv, t1)
        #:gooov[:,p0:p1] -= numpy.einsum('kaic,jc->jkia', pOVov, t1)
        tmp = lib.dot(pOvOv.reshape(-1,nvir), t1.T).reshape(p1-p0,-1,nocc,nocc)
        gooov[p0:p1,:] += tmp.transpose(0,3,2,1)
        lib.dot(t1, pOVov.reshape(-1,nvir).T, 1, tmp.reshape(nocc,-1), 0)
        gooov[:,p0:p1] -= tmp.reshape(nocc,p1-p0,nvir,nocc).transpose(0,1,3,2)
        #:tmp = numpy.einsum('ikac,jc->jika', l2, t1[p0:p1])
        #:gOvVo -= numpy.einsum('jika,kb->jabi', tmp, t1)
        #:gOvvO = numpy.einsum('jkia,kb->jabi', tmp, t1) + pOvOv.transpose(0,3,1,2)
        tmp = tmp.reshape(-1,nocc,nocc,nvir)
        lib.dot(t1[p0:p1], l2.reshape(-1,nvir).T, 1, tmp.reshape(p1-p0,-1))
        gOvVo = numpy.einsum('ia,jb->jabi', l1, t1[p0:p1])
        gOvvO = numpy.empty((p1-p0,nvir,nvir,nocc))
        for i in range(p1-p0):
            gOvVo[i] -= lib.dot(_cp(tmp[i].transpose(0,2,1).reshape(-1,nocc)),
                                t1).reshape(nocc,nvir,-1).transpose(1,2,0)
            gOvVo[i] += pOVov[i].transpose(2,0,1)
            gOvvO[i] = lib.dot(tmp[i].reshape(nocc,-1).T,
                               t1).reshape(nocc,nvir,-1).transpose(1,2,0)
            gOvvO[i] += pOvOv[i].transpose(2,0,1)
        tmp = None
#==== mem usage nocc**3 + nocc*nvir**6
        dovvo[p0:p1] = (gOvVo*2 + gOvvO).transpose(0,2,1,3)
        gOvvO *= -2
        gOvvO -= gOvVo
        doovv[p0:p1] = gOvvO.transpose(0,3,1,2)
        gOvvO = gOvVo = None

        for j0, j1 in prange(0, nocc, blksize):
            tau2 = _ccsd.make_tau(t2[j0:j1], t1[j0:j1], t1)
            #:goovv += numpy.einsum('ijkl,klab->ijab', goooo[:,:,j0:j1], tau2)
            lib.dot(goooo[:,:,j0:j1].copy().reshape((p1-p0)*nocc,-1),
                    tau2.reshape(-1,nvir**2), 1, goovv.reshape(-1,nvir**2), 1)
            tau2 += numpy.einsum('ia,jb->ijab', t1[j0:j1], t1)
            tau2 = _cp(tau2.transpose(0,3,1,2).reshape(-1,nov))
            #:goovv[:,j0:j1] += numpy.einsum('ibld,jlda->ijab', pOvOv, tau2) * .5
            #:goovv[:,j0:j1] -= numpy.einsum('iald,jldb->ijab', pOVov, tau2) * .5
            goovv[:,j0:j1] += lib.dot(pOvOv.reshape(-1,nov), tau2.T,
                                      .5).reshape(p1-p0,nvir,-1,nvir).transpose(0,2,3,1)
            goovv[:,j0:j1] += lib.dot(pOVov.reshape(-1,nov), tau2.T,
                                      -.5).reshape(p1-p0,nvir,-1,nvir).transpose(0,2,1,3)
            tau2 = None
#==== mem usage nocc**3 + nocc*nvir**2*7
        #:goovv += numpy.einsum('iald,jlbd->ijab', pOVov*2+pOvOv, t2) * .5
        pOVov *= 2
        pOVov += pOvOv
        for j in range(nocc):
            tmp = lib.dot(pOVov.reshape(-1,nov),
                          _cp(t2[j].transpose(0,2,1).reshape(-1,nvir)), .5)
            goovv[:,j] += tmp.reshape(-1,nvir,nvir)
            tmp = None
        dovov[p0:p1] = make_theta(goovv).transpose(0,2,1,3)
        goooo = goovv = pOvOv = pOVov = None
        time1 = log.timer_debug1('rdm intermediates pass2 [%d:%d]'%(p0, p1), *time1)

    h5fobj['dooov'][:] = gooov.transpose(0,2,1,3)*2 - gooov.transpose(1,2,0,3)
    gooov = None

    max_memory = mycc.max_memory - lib.current_memory()[0]
    unit = max(nocc**2*nvir*2+nocc*nvir**2*2, nvir**3*2+nocc*nvir**2)
    blksize = max(ccsd.BLKMIN, int(max_memory*.95e6/8/unit))
    iobuflen = int(256e6/8/blksize)
    log.debug1('rdm intermediates pass 3: block size = %d, nvir = %d in %d blocks',
               blksize, nocc, int((nvir+blksize-1)/blksize))
    h5fobj.create_group('dovvv')
    for istep, (p0, p1) in enumerate(prange(0, nvir, blksize)):
        pvOvO = numpy.empty((p1-p0,nocc,nvir,nocc))
        pvOVo = numpy.empty((p1-p0,nocc,nvir,nocc))
        ao2mo.outcore._load_from_h5g(fswap['mvOvO'], p0, p1, pvOvO)
        ao2mo.outcore._load_from_h5g(fswap['mvOVo'], p0, p1, pvOVo)
        #:gvovv -= numpy.einsum('aibk,kc->aibc', pvOvO, t1)
        #:gvovv += numpy.einsum('aick,kb->aibc', pvOVo, t1)
        gvovv = lib.dot(pvOVo.reshape(-1,nocc), t1).reshape(-1,nocc,nvir,nvir)
        for i in range(p1-p0):
            gvovv[i] = gvovv[i].transpose(0,2,1)
        lib.dot(pvOvO.reshape(-1,nocc), t1, -1, gvovv.reshape(-1,nvir), 1)
        pvOvO = pvOVo = None
#==== mem usage nocc**2*nvir*2 + nocc*nvir**2*2

        l2tmp = l2[:,:,p0:p1] * .5
        #:gvvvv = numpy.einsum('ijab,ijcd->abcd', l2tmp, t2)
        #:jabc = numpy.einsum('ijab,ic->jabc', l2tmp, t1)
        #:gvvvv += numpy.einsum('jabc,jd->abcd', jabc, t1)
        gvvvv = lib.dot(l2tmp.reshape(nocc**2,-1).T, t2.reshape(nocc**2,-1))
        jabc = lib.dot(l2tmp.reshape(nocc,-1).T, t1)
        lib.dot(jabc.reshape(nocc,-1).T, t1, 1, gvvvv.reshape(-1,nvir), 1)
        gvvvv = gvvvv.reshape(-1,nvir,nvir,nvir)
        l2tmp = jabc = None

        #:gvovv = numpy.einsum('ja,jibc->aibc', l1[:,p0:p1], t2)
        #:gvovv += numpy.einsum('jibc,ja->aibc', l2, t1[:,p0:p1])
        lib.dot(l1[:,p0:p1].copy().T, t2.reshape(nocc,-1), 1, gvovv.reshape(p1-p0,-1), 1)
        lib.dot(t1[:,p0:p1].copy().T, l2.reshape(nocc,-1), 1, gvovv.reshape(p1-p0,-1), 1)
        tmp = numpy.einsum('ja,jb->ab', l1[:,p0:p1], t1)
        gvovv += numpy.einsum('ab,ic->aibc', tmp, t1)
        gvovv += numpy.einsum('ba,ic->aibc', mvv[:,p0:p1]*.5, t1)
        #:gvovv -= numpy.einsum('adbc,id->aibc', gvvvv, t1*2)
        for j in range(p1-p0):
            lib.dot(t1, gvvvv[j].reshape(nvir,-1), -2,
                    gvovv[j].reshape(nocc,-1), 1)

# symmetrize dvvvv because it is symmetrized in ccsd_grad and make_rdm2 anyway
#:dvvvv = .5*(gvvvv+gvvvv.transpose(0,1,3,2))
#:dvvvv = .5*(dvvvv+dvvvv.transpose(1,0,3,2))
# now dvvvv == dvvvv.transpose(2,3,0,1) == dvvvv.transpose(0,1,3,2) == dvvvv.transpose(1,0,3,2)
        tmp = numpy.empty((nvir,nvir,nvir))
        tmp1 = numpy.empty((nvir,nvir,nvir))
        tmpvvvv = numpy.empty((p1-p0,nvir,nvir_pair))
        for i in range(p1-p0):
            make_theta(gvvvv[i:i+1], out=tmp)
            tmp1[:] = tmp.transpose(1,0,2)
            _ccsd.precontract(tmp1, diag_fac=2, out=tmpvvvv[i])
        # tril of (dvvvv[p0:p1,p0:p1]+dvvvv[p0:p1,p0:p1].T)
        for i in range(p0, p1):
            for j in range(p0, i):
                tmpvvvv[i-p0,j] += tmpvvvv[j-p0,i]
            tmpvvvv[i-p0,i] *= 2
        for i in range(p0, p1):
            off = i * (i+1) // 2
            if p0 > 0:
                tmpvvvv[i-p0,:p0] += dvvvv[off:off+p0]
            dvvvv[off:off+i+1] = tmpvvvv[i-p0,:i+1] * .25
        for i in range(p1, nvir):
            off = i * (i+1) // 2
            dvvvv[off+p0:off+p1] = tmpvvvv[:,i]
        tmp = tmp1 = tmpvvvv = None
#==== mem usage nvir**3 + nocc*nvir**2
        gvvov = make_theta(gvovv).transpose(0,2,1,3)
        ao2mo.outcore._transpose_to_h5g(h5fobj, 'dovvv/%d'%istep,
                                        gvvov.reshape(-1,nov), iobuflen)
        gvvvv = None
        gvovv = None
        time1 = log.timer_debug1('rdm intermediates pass3 [%d:%d]'%(p0, p1), *time1)

    del(fswap['mOvOv'])
    del(fswap['mOVov'])
    del(fswap['mvOvO'])
    del(fswap['mvOVo'])
    fswap.close()
    _tmpfile = None
    return (h5fobj['dovov'], h5fobj['dvvvv'], h5fobj['doooo'], h5fobj['doovv'],
            h5fobj['dovvo'], None, h5fobj['dovvv'], h5fobj['dooov'])
Esempio n. 6
0
def make_intermediates(mycc, t1, t2, eris):
    log = logger.Logger(mycc.stdout, mycc.verbose)
    nocc, nvir = t1.shape
    foo = eris.fock[:nocc,:nocc]
    fov = eris.fock[:nocc,nocc:]
    fvo = eris.fock[nocc:,:nocc]
    fvv = eris.fock[nocc:,nocc:]

    tau = _ccsd.make_tau(t2, t1, t1)
    ovov = np.asarray(eris.ovov)
    ovoo = np.asarray(eris.ovoo)
    ovov1 = ovov * 2 - ovov.transpose(0,3,2,1)
    ovoo1 = ovoo * 2 - ovoo.transpose(2,1,0,3)

    v1  = fvv - lib.einsum('ja,jb->ba', fov, t1)
    v1 -= lib.einsum('jakc,jkbc->ba', ovov1, tau)
    v2  = foo + lib.einsum('ib,jb->ij', fov, t1)
    v2 += lib.einsum('ibkc,jkbc->ij', ovov1, tau)
    v2 += np.einsum('kbij,kb->ij', ovoo1, t1)
    v4 = fov + np.einsum('jbkc,kc->jb', ovov1, t1)

    v5  = np.einsum('kc,jkbc->bj', fov, t2) * 2
    v5 -= np.einsum('kc,jkcb->bj', fov, t2)
    v5 += fvo
    v5 += lib.einsum('kc,kb,jc->bj', v4, t1, t1)
    v5 -= lib.einsum('lckj,klbc->bj', ovoo1, t2)

    oooo = np.asarray(eris.oooo)
    woooo  = lib.einsum('icjl,kc->ikjl', ovoo, t1)
    woooo += lib.einsum('jcil,kc->iljk', ovoo, t1)
    woooo += oooo.copy()
    woooo += lib.einsum('icjd,klcd->ikjl', ovov, tau)

    theta = t2*2 - t2.transpose(0,1,3,2)
    v4OVvo  = lib.einsum('ldjb,klcd->jbck', ovov1, t2)
    v4OVvo -= lib.einsum('ldjb,kldc->jbck', ovov, t2)
    v4OVvo += np.asarray(eris.ovvo)

    v4oVVo  = lib.einsum('jdlb,kldc->jbck', ovov, t2)
    v4oVVo -= np.asarray(eris.oovv).transpose(0,3,2,1)

    v4ovvo = v4OVvo*2 + v4oVVo
    w3 = np.einsum('jbck,jb->ck', v4ovvo, t1)

    woovo  = lib.einsum('ibck,jb->ijck', v4ovvo, t1)
    woovo = woovo - woovo.transpose(0,3,2,1)
    woovo += lib.einsum('ibck,jb->ikcj', v4OVvo-v4oVVo, t1)
    woovo += ovoo1.conj().transpose(3,2,1,0)

    woovo += lib.einsum('lcik,jlbc->ikbj', ovoo1, theta)
    woovo -= lib.einsum('lcik,jlbc->ijbk', ovoo1, t2)
    woovo -= lib.einsum('iclk,ljbc->ijbk', ovoo1, t2)

    wvvvo  = lib.einsum('jack,jb->back', v4ovvo, t1)
    wvvvo = wvvvo - wvvvo.transpose(2,1,0,3)
    wvvvo += lib.einsum('jack,jb->cabk', v4OVvo-v4oVVo, t1)
    wvvvo -= lib.einsum('lajk,jlbc->cabk', ovoo1, tau)

    wOVvo  = v4OVvo
    woVVo  = v4oVVo
    wOVvo -= np.einsum('jbld,kd,lc->jbck', ovov, t1, t1)
    woVVo += np.einsum('jdlb,kd,lc->jbck', ovov, t1, t1)
    wOVvo -= lib.einsum('jblk,lc->jbck', ovoo, t1)
    woVVo += lib.einsum('lbjk,lc->jbck', ovoo, t1)
    v4ovvo = v4OVvo = v4oVVo = None

    ovvv = np.asarray(eris.get_ovvv())
    wvvvo += lib.einsum('kacd,kjbd->bacj', ovvv, t2) * 1.5

    wOVvo += lib.einsum('jbcd,kd->jbck', ovvv, t1)
    woVVo -= lib.einsum('jdcb,kd->jbck', ovvv, t1)

    ovvv = ovvv*2 - ovvv.transpose(0,3,2,1)
    v1 += np.einsum('jcba,jc->ba', ovvv, t1)
    v5 += lib.einsum('kdbc,jkcd->bj', ovvv, t2)
    woovo += lib.einsum('idcb,jkdb->ijck', ovvv, tau)

    tmp = lib.einsum('kdca,jkbd->cabj', ovvv, theta)
    wvvvo -= tmp
    wvvvo += tmp.transpose(2,1,0,3) * .5
    wvvvo -= ovvv.conj().transpose(3,2,1,0)
    ovvv = tmp = None

    w3 += v5
    w3 += np.einsum('cb,jb->cj', v1, t1)
    w3 -= np.einsum('jk,jb->bk', v2, t1)

    class _IMDS: pass
    imds = _IMDS()
    imds.ftmp = lib.H5TmpFile()
    dtype = np.result_type(t2, eris.vvvv).char
    imds.woooo = imds.ftmp.create_dataset('woooo', (nocc,nocc,nocc,nocc), dtype)
    imds.wovvo = imds.ftmp.create_dataset('wovvo', (nocc,nvir,nvir,nocc), dtype)
    imds.woVVo = imds.ftmp.create_dataset('woVVo', (nocc,nvir,nvir,nocc), dtype)
    imds.woovo = imds.ftmp.create_dataset('woovo', (nocc,nocc,nvir,nocc), dtype)
    imds.wvvvo = imds.ftmp.create_dataset('wvvvo', (nvir,nvir,nvir,nocc), dtype)

    imds.woooo[:] = woooo
    imds.wovvo[:] = wOVvo*2 + woVVo
    imds.woVVo[:] = woVVo
    imds.woovo[:] = woovo
    imds.wvvvo[:] = wvvvo
    imds.v1 = v1
    imds.v2 = v2
    imds.w3 = w3
    imds.v4 = v4
    imds.ftmp.flush()
    return imds
Esempio n. 7
0
def make_tau(t2, t1a, t1b, fac=1, out=None):
    return _ccsd.make_tau(t2, t1a, t1b, fac, out)
Esempio n. 8
0
def gamma2_incore(mycc, t1, t2, l1, l2):
    log = logger.Logger(mycc.stdout, mycc.verbose)
    nocc, nvir = t1.shape
    nov = nocc * nvir

    time1 = time.clock(), time.time()
    #:theta = make_theta(t2)
    #:mOvOv = numpy.einsum('ikca,jkcb->jbia', l2, t2)
    #:mOVov = -numpy.einsum('ikca,jkbc->jbia', l2, t2)
    #:mOVov += numpy.einsum('ikac,jkbc->jbia', l2, theta)
    l2a = numpy.empty((nocc,nvir,nocc,nvir))
    t2a = numpy.empty((nocc,nvir,nocc,nvir))
    for i in range(nocc):
        l2a[i] = l2[i].transpose(2,0,1)
        t2a[i] = t2[i].transpose(2,0,1)
    mOvOv = lib.dot(t2a.reshape(-1,nov), l2a.reshape(-1,nov).T).reshape(nocc,nvir,nocc,nvir)
    for i in range(nocc):
        t2a[i] = t2[i].transpose(1,0,2)
    mOVov = lib.dot(t2a.reshape(-1,nov), l2a.reshape(-1,nov).T, -1).reshape(nocc,nvir,nocc,nvir)
    theta = t2a
    for i in range(nocc):
        l2a[i] = l2[i].transpose(1,0,2)
        theta[i] *= 2
        theta[i] -= t2[i].transpose(2,0,1)
    lib.dot(theta.reshape(-1,nov), l2a.reshape(nov,-1).T, 1, mOVov.reshape(nov,-1), 1)
    theta = l2a = t2a = None
    moo =(numpy.einsum('jdld->jl', mOvOv) * 2
        + numpy.einsum('jdld->jl', mOVov))
    mvv =(numpy.einsum('lbld->bd', mOvOv) * 2
        + numpy.einsum('lbld->bd', mOVov))
    mia =(numpy.einsum('kc,ikac->ia', l1, t2) * 2
        - numpy.einsum('kc,ikca->ia', l1, t2))
    mab = numpy.einsum('kc,kb->cb', l1, t1)
    mij = numpy.einsum('kc,jc->jk', l1, t1) + moo*.5

    gooov = numpy.zeros((nocc,nocc,nocc,nvir))
    tau = _ccsd.make_tau(t2, t1, t1)
    #:goooo = numpy.einsum('ijab,klab->klij', l2, tau)*.5
    goooo = lib.dot(tau.reshape(-1,nvir**2), l2.reshape(-1,nvir**2).T, .5)
    goooo = goooo.reshape(-1,nocc,nocc,nocc)
    doooo = _cp(make_theta(goooo).transpose(0,2,1,3))

    #:gooov -= numpy.einsum('ib,kjab->jkia', l1, tau)
    #:gooov -= numpy.einsum('kjab,ib->jkia', l2, t1)
    #:gooov += numpy.einsum('jkil,la->jkia', goooo, t1*2)
    gooov = lib.dot(_cp(tau.reshape(-1,nvir)), l1.T, -1)
    lib.dot(_cp(l2.reshape(-1,nvir)), t1.T, -1, gooov, 1)
    gooov = gooov.reshape(nocc,nocc,nvir,nocc)
    tmp = numpy.einsum('ji,ka->jkia', moo*-.5, t1)
    tmp += gooov.transpose(1,0,3,2)
    gooov, tmp = tmp, None
    lib.dot(goooo.reshape(-1,nocc), t1, 2, gooov.reshape(-1,nvir), 1)

    goovv = numpy.einsum('ia,jb->ijab', mia, t1)
    for i in range(nocc):
        goovv[i] += .5 * l2 [i]
        goovv[i] += .5 * tau[i]
    #:goovv -= numpy.einsum('jk,kiba->jiba', mij, tau)
    lib.dot(mij, tau.reshape(nocc,-1), -1, goovv.reshape(nocc,-1), 1)
    #:goovv -= numpy.einsum('cb,ijac->ijab', mab, t2)
    #:goovv -= numpy.einsum('bd,ijad->ijab', mvv*.5, tau)
    lib.dot(t2.reshape(-1,nvir), mab, -1, goovv.reshape(-1,nvir), 1)
    lib.dot(tau.reshape(-1,nvir), mvv.T, -.5, goovv.reshape(-1,nvir), 1)
    tau = None

    #:gooov += numpy.einsum('jaic,kc->jkia', mOvOv, t1)
    #:gooov -= numpy.einsum('kaic,jc->jkia', mOVov, t1)
    tmp = lib.dot(mOvOv.reshape(-1,nvir), t1.T).reshape(nocc,-1,nocc,nocc)
    gooov += tmp.transpose(0,3,2,1)
    lib.dot(t1, mOVov.reshape(-1,nvir).T, 1, tmp.reshape(nocc,-1), 0)
    gooov -= tmp.reshape(nocc,nocc,nvir,nocc).transpose(0,1,3,2)
    dooov = gooov.transpose(0,2,1,3)*2 - gooov.transpose(1,2,0,3)
    gooov = None
    #:tmp = numpy.einsum('ikac,jc->jaik', l2, t1)
    #:gOvVo -= numpy.einsum('jaik,kb->jabi', tmp, t1)
    #:gOvvO = numpy.einsum('jaki,kb->jabi', tmp, t1) + mOvOv.transpose(0,3,1,2)
    tmp = tmp.reshape(nocc,nocc,nocc,nvir)
    lib.dot(t1, l2.reshape(-1,nvir).T, 1, tmp.reshape(nocc,-1))
    gOvVo = numpy.einsum('ia,jb->jabi', l1, t1)
    gOvvO = numpy.empty((nocc,nvir,nvir,nocc))
    for i in range(nocc):
        gOvVo[i] -= lib.dot(_cp(tmp[i].transpose(0,2,1).reshape(-1,nocc)),
                            t1).reshape(nocc,nvir,-1).transpose(1,2,0)
        gOvVo[i] += mOVov[i].transpose(2,0,1)
        gOvvO[i] = lib.dot(tmp[i].reshape(nocc,-1).T,
                           t1).reshape(nocc,nvir,-1).transpose(1,2,0)
        gOvvO[i] += mOvOv[i].transpose(2,0,1)
    tmp = None

    dovvo = numpy.empty((nocc,nvir,nvir,nocc))
    doovv = numpy.empty((nocc,nocc,nvir,nvir))
    for i in range(nocc):
        tmp = gOvVo[i] * 2 + gOvvO[i]
        dovvo[i] = tmp.transpose(1,0,2)
        tmp = gOvvO[i] * -2 - gOvVo[i]
        doovv[i] = tmp.transpose(2,0,1)
    gOvvO = gOvVo = None

    tau2 = _ccsd.make_tau(t2, t1, t1)
    #:goovv += numpy.einsum('ijkl,klab->ijab', goooo[:,:,j0:j1], tau2)
    lib.dot(goooo.reshape(nocc*nocc,-1),
            tau2.reshape(-1,nvir**2), 1, goovv.reshape(-1,nvir**2), 1)
    tau2 += numpy.einsum('ia,jb->ijab', t1, t1)
    tau2p = tau2.reshape(nocc,nvir,nocc,nvir)
    for i in range(nocc):
        tau2p[i] = tau2[i].transpose(2,0,1)
    tau2, tau2p = tau2p.reshape(nov,-1), None
    #:goovv += numpy.einsum('ibld,jlda->ijab', mOvOv, tau2) * .5
    #:goovv -= numpy.einsum('iald,jldb->ijab', mOVov, tau2) * .5
    tmp = lib.dot(mOvOv.reshape(-1,nov), tau2.T, .5).reshape(nocc,nvir,-1,nvir)
    for i in range(nocc):
        tmp[i] = goovv[i].transpose(1,0,2) + tmp[i].transpose(2,1,0)
    goovv, tmp = tmp, None
    lib.dot(mOVov.reshape(-1,nov), tau2.T, -.5, goovv.reshape(nov,-1), 1)

    #:goovv += numpy.einsum('iald,jlbd->ijab', mOVov*2+mOvOv, t2) * .5
    t2a, tau2 = tau2.reshape(nocc,nvir,nocc,nvir), None
    for i in range(nocc):
        t2a[i] = t2[i].transpose(1,0,2)
    tmp = mOVov*2
    tmp += mOvOv
    lib.dot(tmp.reshape(-1,nov), t2a.reshape(nov,-1), .5, goovv.reshape(nov,-1), 1)
    t2a = tmp = None
    for i in range(nocc):
        goovv[i] = goovv[i] * 2 - goovv[i].transpose(2,1,0)
    dovov = goovv
    goooo = goovv = None

    #:gvovv += numpy.einsum('aick,kb->aibc', pvOVo, t1)
    mOVov = lib.transpose(mOVov.reshape(nov,-1))
    gvovv = lib.dot(mOVov.reshape(nocc,-1).T, t1).reshape(nvir,nocc,nvir,nvir)
    mOVov = None
    tmp = numpy.einsum('ja,jb->ab', l1, t1)
    #:gvovv += numpy.einsum('ab,ic->aibc', tmp, t1)
    #:gvovv += numpy.einsum('ba,ic->aibc', mvv, t1*.5)
    for i in range(nvir):
        gvovv[i] += numpy.einsum('b,ic->icb', tmp[i], t1)
        gvovv[i] += numpy.einsum('b,ic->icb', mvv[:,i]*.5, t1)
        gvovv[i] = gvovv[i].transpose(0,2,1)

    #:gvovv += numpy.einsum('ja,jibc->aibc', l1, t2)
    #:gvovv += numpy.einsum('jibc,ja->aibc', l2, t1)
    #:gvovv -= numpy.einsum('aibk,kc->aibc', pvOvO, t1)
    mOvOv = lib.transpose(mOvOv.reshape(nov,-1))
    lib.dot(mOvOv.reshape(nocc,-1).T, t1, -1, gvovv.reshape(-1,nvir), 1)
    mOvOv = None
    lib.dot(l1.T, t2.reshape(nocc,-1), 1, gvovv.reshape(nvir,-1), 1)
    lib.dot(t1.T, l2.reshape(nocc,-1), 1, gvovv.reshape(nvir,-1), 1)

    tmp = numpy.empty((nocc,nvir,nvir))
    for i in range(nvir):
        #:gvovv*2 - gvovv.transpose(0,1,3,2)
        gvovv[i] = _ccsd.make_021(gvovv[i], gvovv[i], 2, -1, out=tmp)

    #:gvvvv = numpy.einsum('ijab,ijcd->abcd', l2, t2)*.5
    #:jabc = numpy.einsum('ijab,ic->jabc', l2, t1) * .5
    #:gvvvv += numpy.einsum('jabc,jd->abcd', jabc, t1)
    #:gvovv -= numpy.einsum('adbc,id->aibc', gvvvv, t1*2)
    tau = _ccsd.make_tau(t2, t1, t1)
    theta = make_theta(tau)
    tau = None
    l2tmp = lib.pack_tril(l2.reshape(-1,nvir,nvir))
    gtmp = lib.dot(l2tmp.T, theta.reshape(nocc**2,-1), .5).reshape(-1,nvir,nvir)
    l2tmp = theta = None
    nvir_pair = nvir * (nvir+1) //2
    tmp = numpy.empty((nvir,nvir,nvir))
    tmp1 = numpy.empty((nvir,nvir,nvir))
    tmptril = numpy.empty((nvir,nvir_pair))
    diag_idx = numpy.arange(nvir)
    diag_idx = diag_idx*(diag_idx+1)//2 + diag_idx

    dvvvv = numpy.empty((nvir_pair,nvir_pair))
    dovvv = numpy.empty((nocc,nvir,nvir,nvir))
# dvvov = (gvovv*2 - gvovv.transpose(0,1,3,2)).transpose(0,2,1,3)
# dovvv = dvvov.transpose(2,3,0,1)
    p0 = 0
    for i in range(nvir):
        tmp[:i+1] = gtmp[p0:p0+i+1]
        for j in range(i+1, nvir):
            tmp[j] = gtmp[j*(j+1)//2+i].T
        lib.dot(t1, tmp.reshape(nvir,-1), -2, gvovv[i].reshape(nocc,-1), 1)
        dovvv[:,:,i] = gvovv[i].transpose(0,2,1)

        #:gvvvv[i] = (tmp*2-tmp.transpose(0,2,1)).transpose(1,0,2)
        #:gvvvv = .5*(gvvvv+gvvvv.transpose(0,1,3,2))
        #:dvvvv = .5*(gvvvv+gvvvv.transpose(1,0,3,2))
        tmp1[:] = tmp.transpose(1,0,2)
        _ccsd.precontract(tmp1, diag_fac=2, out=tmptril)
        dvvvv[p0:p0+i] += tmptril[:i]
        dvvvv[p0:p0+i] *= .25
        dvvvv[i*(i+1)//2+i] = tmptril[i] * .5
        for j in range(i+1, nvir):
            dvvvv[j*(j+1)//2+i] = tmptril[j]
        p0 += i + 1
    gtmp = tmp = tmp1 = tmptril = gvovv = None
    dvvov = dovvv.transpose(2,3,0,1)
    return (dovov, dvvvv, doooo, doovv, dovvo, dvvov, dovvv, dooov)
Esempio n. 9
0
def update_amps(mycc, t1, t2, l1, l2, eris=None, saved=None):
    if saved is None:
        saved = make_intermediates(mycc, t1, t2, eris)
    time1 = time0 = time.clock(), time.time()
    log = logger.Logger(mycc.stdout, mycc.verbose)
    nocc, nvir = t1.shape
    nov = nocc * nvir
    foo = eris.fock[:nocc,:nocc]
    fov = eris.fock[:nocc,nocc:]
    fvv = eris.fock[:nocc,:nocc]

    #:mba = numpy.einsum('klca,klcb->ba', l2, t2*2-t2.transpose(0,1,3,2))
    #:mij = numpy.einsum('ikcd,jkcd->ij', l2, t2*2-t2.transpose(0,1,3,2))
    #:theta = t2*2 - t2.transpose(0,1,3,2)
    theta = _ccsd.make_0132(t2, t2, 2, -1)
    mba = lib.dot(theta.reshape(-1,nvir).T, l2.reshape(-1,nvir))
    mij = lib.dot(l2.reshape(nocc,-1), theta.reshape(nocc,-1).T)
    theta = None
    mba1 = numpy.einsum('jc,jb->bc', l1, t1) + mba
    mij1 = numpy.einsum('kb,jb->kj', l1, t1) + mij
    mia1 =(t1 + numpy.einsum('kc,jkbc->jb', l1, t2) * 2
         - numpy.einsum('kc,jkcb->jb', l1, t2)
         - reduce(numpy.dot, (t1, l1.T, t1))
         - numpy.einsum('bd,jd->jb', mba, t1)
         - numpy.einsum('lj,lb->jb', mij, t1))

    tmp = mycc.add_wvvVV(numpy.zeros_like(l1), l2, eris)
    l2new = numpy.empty((nocc,nocc,nvir,nvir))
    ij = 0
    for i in range(nocc):
        for j in range(i):
            tmp1 = tmp[ij] * .5  # *.5 because of l2+l2.transpose(1,0,3,2) later
            l2new[i,j] = tmp1
            l2new[j,i] = tmp1.T
            ij += 1
        l2new[i,i] = tmp[ij] * .5
        ij += 1
    l1new =(numpy.einsum('ijab,jb->ia', l2new, t1) * 4
          - numpy.einsum('jiab,jb->ia', l2new, t1) * 2)
    tmp = tmp1 = None

    l1new += eris.fock[:nocc,nocc:]
    l1new += numpy.einsum('ib,ba->ia', l1, saved.w1)
    l1new -= numpy.einsum('ja,ij->ia', l1, saved.w2)
    l1new -= numpy.einsum('ik,ka->ia', mij, saved.w4)
    l1new -= numpy.einsum('ca,ic->ia', mba, saved.w4)
    l1new += numpy.einsum('ijab,bj->ia', l2, saved.w3) * 2
    l1new -= numpy.einsum('ijba,bj->ia', l2, saved.w3)

    l2new += numpy.einsum('ia,jb->ijab', l1, saved.w4)
    #:l2new += numpy.einsum('jibc,ca->jiba', l2, saved.w1)
    #:l2new -= numpy.einsum('kiba,jk->jiba', l2, saved.w2)
    lib.dot(l2.reshape(-1,nvir), saved.w1, 1, l2new.reshape(-1,nvir), 1)
    lib.dot(saved.w2, l2.reshape(nocc,-1),-1, l2new.reshape(nocc,-1), 1)

    eris_ooov = _cp(eris.ooov)
    l1new -= numpy.einsum('jkia,kj->ia', eris_ooov, mij1) * 2
    l1new += numpy.einsum('ikja,kj->ia', eris_ooov, mij1)
    #:l2new -= numpy.einsum('ka,kijb->jiba', l1, eris_ooov)
    lib.dot(_cp(eris_ooov.transpose(0,2,1,3).reshape(nocc,-1)).T,
            l1, -1, l2new.reshape(-1,nvir), 1)
    eris_ooov = None

    tau = _ccsd.make_tau(t2, t1, t1)
    #:l2tau = numpy.einsum('ijcd,klcd->ijkl', l2, tau)
    l2tau = lib.dot(l2.reshape(nocc**2,-1),
                    tau.reshape(nocc**2,-1).T).reshape((nocc,)*4)
    tau = None
    l2t1 = numpy.einsum('ijcd,kc->ijkd', l2, t1)

    eris_ovvv = _cp(eris.ovvv)
    eris_ovvv = _ccsd.unpack_tril(eris_ovvv.reshape(nov,-1))
    eris_ovvv = eris_ovvv.reshape(nocc,nvir,nvir,nvir)

    l1new += numpy.einsum('iabc,bc->ia', eris_ovvv, mba1) * 2
    l1new -= numpy.einsum('ibca,bc->ia', eris_ovvv, mba1)
    #:l2new += numpy.einsum('ic,jbac->jiba', l1, eris_ovvv)
    tmp = lib.dot(l1, eris_ovvv.reshape(-1,nvir).T).reshape(nocc,-1,nvir,nvir)
    for i in range(nocc):
        l2new[i] += tmp[i].transpose(0,2,1)
    #:m4 = numpy.einsum('ijkd,kadb->ijab', l2t1, eris_ovvv)
    m4 = tmp
    lib.dot(_cp(l2t1.reshape(nocc*nocc,-1)),
            _cp(eris_ovvv.transpose(0,2,1,3).reshape(-1,nvir**2)),
            1, m4.reshape(nocc*nocc,-1))
    l2new -= m4
    l1new -= numpy.einsum('ijab,jb->ia', m4, t1) * 2
    l1new -= numpy.einsum('ijab,ia->jb', m4, t1) * 2
    l1new += numpy.einsum('jiab,jb->ia', m4, t1)
    l1new += numpy.einsum('jiab,ia->jb', m4, t1)
    eris_ovvv = tmp = None

    eris_ovov = _cp(eris.ovov)
    l1new += numpy.einsum('jb,iajb->ia', l1, eris_ovov) * 2
    #:l2new -= numpy.einsum('jbic,ca->jiba', eris_ovov, mba1)
    #:l2new -= numpy.einsum('kajb,ik->ijab', eris_ovov, mij1)
    tmp = lib.dot(eris_ovov.reshape(-1,nvir), mba1).reshape(nocc,nvir,nocc,nvir)
    lib.dot(mij1, eris_ovov.reshape(nocc,-1), 1, tmp.reshape(nocc,-1), 1)
    tmp_oovv = numpy.empty((nocc,nocc,nvir,nvir))
    for i in range(nocc):
        tmp_oovv[i] = eris_ovov[i].transpose(1,0,2) * .5
        l2new[i] += tmp_oovv[i]
        l2new[i] -= tmp[i].transpose(1,0,2)
    tmp = None
    l1new += numpy.einsum('iajb,jb->ia', eris_ovov, mia1) * 2
    l1new -= numpy.einsum('ibja,jb->ia', eris_ovov, mia1)
    #:m4 = numpy.einsum('kalb,ijkl->ijab', eris_ovov, l2tau)
    lib.dot(l2tau.reshape(nocc*nocc,-1), tmp_oovv.reshape(-1,nvir**2),
            1, m4.reshape(nocc**2,-1))
    l2new += m4
    l1new += numpy.einsum('ijab,jb->ia', m4, t1) * 4
    l1new -= numpy.einsum('ijba,jb->ia', m4, t1) * 2
    eris_ovov = m4 = tmp_oovv = None

    eris_oovv = _cp(eris.oovv)
    l1new -= numpy.einsum('jb,ijba->ia', l1, eris_oovv)
    eris_oovv = None

    saved_wooov = _cp(saved.wooov)
    #:l1new -= numpy.einsum('jkca,ijkc->ia', l2, saved_wooov)
    l1new -= lib.dot(saved_wooov.reshape(nocc,-1), l2.reshape(-1,nvir))
    saved_wovvv = _cp(saved.wovvv)
    #:l1new += numpy.einsum('kibc,kabc->ia', l2, saved_wovvv)
    for j in range(nocc):
        l1new += lib.dot(l2[j].reshape(nocc,-1),
                         saved_wovvv[j].reshape(nvir,-1).T)
    saved_wooov = saved_wovvv = None

    saved_wOvOv = _cp(saved.wOvOv)
    tmp_ovov = _cp(saved.wOVov) * 2
    tmp_ovov += saved_wOvOv
    #:tmp = l2.transpose(0,2,1,3) - l2.transpose(0,3,1,2)*.5
    #:l2new += numpy.einsum('kcia,kcjb->jiba', tmp, tmp_ovov)
    tmp = numpy.empty((nocc,nvir,nocc,nvir))
    for i in range(nocc):
        tmp[i] = l2[i].transpose(2,0,1)*-.5
        tmp[i] += l2[i].transpose(1,0,2)
    tmp = lib.dot(tmp_ovov.reshape(-1,nov),
                  tmp.reshape(nov,-1)).reshape(-1,nvir,nocc,nvir)
    #:tmp = numpy.einsum('jkca,ibkc->ijab', l2, saved_wOvOv)
    for i in range(nocc):
        l2new[i] += tmp[i].transpose(1,0,2)
        tmp_ovov[i] = l2[i].transpose(2,0,1)
    lib.dot(saved_wOvOv.reshape(-1,nov), tmp_ovov.reshape(nov,-1),
            1, tmp.reshape(nov,-1))
    for i in range(nocc):
        l2new[i] += tmp[i].transpose(1,2,0)
        l2new[i] += tmp[i].transpose(1,0,2) * .5
    saved_wOvOv = tmp = tmp_ovov = None

    saved_woooo = _cp(saved.woooo)
    #:m3 = numpy.einsum('klab,ijkl->ijab', l2, saved_woooo)
    m3 = lib.dot(saved_woooo.reshape(-1,nocc**2),
                 l2.reshape(nocc**2,-1), .5).reshape(-1,nocc,nvir,nvir)
    l2new += m3
    l1new += numpy.einsum('ijab,jb->ia', m3, t1) * 4
    l1new -= numpy.einsum('ijba,jb->ia', m3, t1) * 2
    saved_woooo = m3 = None

    mo_e = eris.fock.diagonal()
    eia = lib.direct_sum('i-j->ij', mo_e[:nocc], mo_e[nocc:])
    l1new /= eia
    l1new += l1

#    l2new = l2new + l2new.transpose(1,0,3,2)
#    l2new /= lib.direct_sum('ia+jb->ijab', eia, eia)
#    l2new += l2
    ij = 0
    for i in range(nocc):
        for j in range(i):
            dab = lib.direct_sum('a+b->ab', eia[i], eia[j])
            tmp = (l2new[i,j]+l2new[j,i].T) / dab + l2[i,j]
            l2new[i,j] = tmp
            l2new[j,i] = tmp.T
            ij += 1
        dab = lib.direct_sum('a+b->ab', eia[i], eia[i])
        l2new[i,i] = (l2new[i,i]+l2new[i,i].T)/dab + l2[i,i]
        ij += 1

    time0 = log.timer_debug1('update l1 l2', *time0)
    return l1new, l2new
Esempio n. 10
0
def make_intermediates(mycc, t1, t2, eris):
    log = logger.Logger(mycc.stdout, mycc.verbose)
    nocc, nvir = t1.shape
    nov = nocc * nvir
    foo = eris.fock[:nocc,:nocc]
    fov = eris.fock[:nocc,nocc:]
    fvv = eris.fock[nocc:,nocc:]

    class _Saved(object):
        pass
    saved = _Saved()

# As we don't have l2 in memory, hold tau temporarily in memory
    w1 = fvv - numpy.einsum('ja,jb->ba', fov, t1)
    w2 = foo + numpy.einsum('ib,jb->ij', fov, t1)
    w3 = _cp(numpy.einsum('kc,jkbc->bj', fov, t2) * 2 + fov.T)
    w3 -= numpy.einsum('kc,kjbc->bj', fov, t2)
    w3 += reduce(numpy.dot, (t1.T, fov, t1.T))
    w4 = fov.copy()

    eris_ovvv = _cp(eris.ovvv)
    eris_ovvv = _ccsd.unpack_tril(eris_ovvv.reshape(nov,-1))
    eris_ovvv = eris_ovvv.reshape(nocc,nvir,nvir,nvir)

    wovvv = numpy.empty((nocc,nvir,nvir,nvir))
    t2tmp = numpy.empty((nocc,nvir,nocc,nvir))
    for i in range(nocc):
        wovvv[i] = eris_ovvv[i].transpose(1,0,2) * 2
        t2tmp[i] = t2[i].transpose(2,0,1)
    #:wovvv += numpy.einsum('jabd,kjdc->kabc', eris_ovvv, t2) * -1.5
    tmp = lib.dot(t2tmp.reshape(nov,-1), wovvv.reshape(-1,nvir**2),
                  -1.5/2).reshape(-1,nvir,nvir,nvir)
    g2ovvv = tmp
    for i in range(nocc):
        wovvv[i] -= eris_ovvv[i].transpose(1,2,0)
        wovvv[i] += tmp[i].transpose(1,2,0)
        g2ovvv[i] = eris_ovvv[i]*2
        g2ovvv[i] -= eris_ovvv[i].transpose(1,2,0)
    tmp = t2tmp = None

    w1 += numpy.einsum('jcba,jc->ba', eris_ovvv, t1*2)
    w1 -= numpy.einsum('jabc,jc->ba', eris_ovvv, t1)
    #:w3 += numpy.einsum('kdcb,kjdc->bj', eris_ovvv, theta)
    theta = numpy.empty(t2.shape)
    for i in range(nocc):
        theta[i] = t2[i] * 2
        theta[i] -= t2[i].transpose(0,2,1)
        lib.dot(eris_ovvv[i].reshape(-1,nvir).T,
                _cp(theta[i].reshape(nocc,-1)).T, 1, w3, 1)

    theta = _cp(theta.transpose(0,2,1,3))
    #:vkbca = numpy.einsum('jdca,kbjd->kbca', g2ovvv, theta)
    vkbca = lib.dot(_cp(theta.reshape(nov,-1)),
                    g2ovvv.reshape(-1,nvir*nvir)).reshape(-1,nvir,nvir,nvir)
    for i in range(nocc):
        wovvv[i] += vkbca[i].transpose(2,0,1)
        wovvv[i] -= vkbca[i].transpose(2,1,0) * .5
    vkabc = None

    #:wOVov = numpy.einsum('jbcd,kd->jbkc', eris_ovvv, t1)
    #:wOvOv = numpy.einsum('jdcb,kd->jbkc', eris_ovvv, -t1)
    wOVov = lib.dot(eris_ovvv.reshape(-1,nvir),
                    t1.T).reshape(-1,nvir,nvir,nocc).transpose(0,1,3,2).copy()
    for i in range(nocc):
        g2ovvv[i] = eris_ovvv[i].transpose(1,2,0) * 2
    wOvOv = lib.dot(g2ovvv.reshape(-1,nvir),
                    -t1.T, .5).reshape(-1,nvir,nvir,nocc).transpose(0,1,3,2).copy()
    for i in range(nocc):
        g2ovvv[i] -= eris_ovvv[i].transpose(1,0,2)
    eris_ovov = _cp(_cp(eris.ovov).transpose(0,2,1,3))
    tau = _ccsd.make_tau(t2, t1, t1)
    #:wooov[:,j0:j1] = numpy.einsum('icbd,jkbd->ijkc', g2ovvv, tau)
    #:woooo[:,:,j0:j1] = numpy.einsum('icjd,klcd->ijkl', eris_ovov, tau)
    tmp = lib.dot(g2ovvv.reshape(-1,nvir**2), tau.reshape(-1,nvir**2).T)
    wooov = _cp(tmp.reshape(-1,nvir,nocc,nocc).transpose(0,2,3,1))
    woooo = lib.dot(eris_ovov.reshape(-1,nvir**2),
                    tau.reshape(-1,nvir**2).T).reshape(-1,nocc,nocc,nocc)
    eris_ovov = eris_ovvv = g2ovvv = tau = tmp = None

    eris_ooov = _cp(eris.ooov)
    eris_ovoo = _cp(eris.ovoo)
    #:woooo += numpy.einsum('icjl,kc->ijkl', eris_ovoo, t1)
    #:wOVov += numpy.einsum('jblk,lc->jbkc', eris_ovoo, -t1)
    for i in range(nocc):
        woooo[i] += lib.dot(t1, eris_ovoo[i].reshape(nvir,-1)).reshape((nocc,)*3).transpose(1,0,2)
    lib.dot(eris_ovoo.reshape(-1,nocc), t1, -1, wOVov.reshape(-1,nvir), 1)
    #:wooov -= numpy.einsum('ibjl,lkcb->ijkc', eris_ovoo*1.5, t2)
    t2tmp = numpy.empty((nocc,nvir,nocc,nvir))
    for i in range(nocc):
        t2tmp[i] = t2[i].transpose(2,0,1)
    tmp_ooov = _cp(-eris_ooov.transpose(2,0,1,3)).reshape(-1,nov)
    lib.dot(tmp_ooov, t2tmp.reshape(nov,-1), 1.5, wooov.reshape(-1,nov), 1)
    t2tmp = None

    g2ooov, tmp_ooov = tmp_ooov.reshape(nocc,nocc,nocc,nvir), None
    g2ooov += eris_ooov * 2
    #:vikjc = numpy.einsum('iklb,jlcb->ikjc', g2ooov, theta)
    vikjc = lib.dot(g2ooov.reshape(-1,nov), theta.reshape(-1,nov).T)
    vikjc = vikjc.reshape(nocc,nocc,nocc,nvir)
    wooov += vikjc.transpose(0,2,1,3)
    wooov -= vikjc*.5
    g2ooov = vikjc = eris_ovoo = None

    w2 += numpy.einsum('ijkb,kb->ij', eris_ooov, t1) * 2
    w2 -= numpy.einsum('kjib,kb->ij', eris_ooov, t1)
    #:w3 -= numpy.einsum('kjlc,klbc->bj', eris_ooov, theta)
    for i in range(nocc):
        lib.dot(_cp(theta[i].transpose(1,2,0)).reshape(-1,nvir).T,
                eris_ooov[i].reshape(nocc,-1).T, -1, w3, 1)
    #:woooo += numpy.einsum('ikjc,lc->ijkl', eris_ooov, t1)
    #:wOvOv += numpy.einsum('jklb,lc->jbkc', eris_ooov, t1)
    woooo += lib.dot(eris_ooov.reshape(-1,nvir),
                     t1.T).reshape((-1,nocc,nocc,nocc)).transpose(0,2,1,3)
    for i in range(nocc):
        lib.dot(_cp(eris_ooov[i].transpose(2,0,1)).reshape(-1,nocc),
                t1, 1, wOvOv[i].reshape(-1,nvir), 1)
        wooov[i] += eris_ooov[i].transpose(1,0,2)*2
        wooov[i] -= eris_ooov[i]
    eris_ooov = theta = None

    eris_ovov = _cp(eris.ovov)
    g2ovov = numpy.empty((nocc,nocc,nvir,nvir))
    for i in range(nocc):
        g2ovov[i] = eris_ovov[i].transpose(1,0,2)*2
        g2ovov[i] -= eris_ovov[i].transpose(1,2,0)
    tmpw4 = numpy.einsum('klcd,ld->kc', g2ovov, t1)
    #:w1 -= numpy.einsum('kcja,kjcb->ba', g2ovov, t2)
    w1 -= lib.dot(t2.reshape(-1,nvir).T, g2ovov.reshape(-1,nvir))
    w1 -= numpy.einsum('ja,jb->ba', tmpw4, t1)
    #:w2 += numpy.einsum('ibkc,jkbc->ij', g2ovov, t2)
    w2 += lib.dot(g2ovov.reshape(nocc,-1), t2.reshape(nocc,-1).T)
    w2 += numpy.einsum('ib,jb->ij', tmpw4, t1)
    w3 += reduce(numpy.dot, (t1.T, tmpw4, t1.T))
    w4 += tmpw4
    vOVov  = eris_ovov.copy()
    #:vOVov += numpy.einsum('jbld,klcd->jbkc', g2ovov, t2)
    #:vOVov -= numpy.einsum('jbld,kldc->jbkc', eris_ovov, t2)
    lib.dot(_cp(g2ovov.transpose(0,2,1,3)).reshape(-1,nov),
            _cp(t2.transpose(0,2,1,3).reshape(nov,-1).T), 1,
            vOVov.reshape(nov,-1), 1)
    lib.dot(eris_ovov.reshape(-1,nov),
            _cp(t2.transpose(0,3,1,2).reshape(nov,-1).T), -1,
            vOVov.reshape(nov,-1), 1)
    g2ovov = None

    #:tmp = numpy.einsum('jbld,kd->ljbk', eris_ovov, t1)
    #:wOVov -= numpy.einsum('ljbk,lc->jbkc', tmp, t1)
    #:tmp = numpy.einsum('jdlb,kd->ljbk', eris_ovov, t1)
    #:wOvOv += numpy.einsum('ljbk,lc->jbkc', tmp, t1)
    tmp = numpy.empty((nocc,nvir,nocc))
    for j in range(nocc):
        lib.dot(_cp(eris_ovov[j].transpose(1,0,2)).reshape(-1,nvir),
                t1.T, 1, tmp.reshape(-1,nocc))
        lib.dot(tmp.reshape(nocc,-1).T, t1, -1, wOVov[j].reshape(-1,nvir), 1)
        lib.dot(eris_ovov[j].reshape(nvir,-1).T, t1.T, 1,
                tmp.reshape(-1,nocc))
        lib.dot(tmp.reshape(nocc,-1).T, t1, 1, wOvOv[j].reshape(-1,nvir), 1)
    tmp = None

    #:vOvOv = numpy.einsum('jdlb,kldc->jbkc', eris_ovov, t2)
    ovovtmp = _cp(eris_ovov.transpose(0,3,2,1).reshape(-1,nov))
    vOvOv = numpy.empty((nocc,nvir,nocc,nvir))
    for j in range(nocc):
        lib.dot(t2[j].reshape(-1,nvir).T, ovovtmp.T, 1,
                vOvOv[j].reshape(nvir,-1))
        vOvOv[j] -= eris.oovv[j].transpose(2,0,1)
    ovovtmp = eris_ovov = None
    vOvOv = lib.transpose(vOvOv.reshape(nov,-1)).reshape(nocc,nvir,nocc,nvir)
    wOVov += vOVov
    wOvOv += vOvOv
    saved.wOVov = wOVov
    saved.wOvOv = wOvOv
    ovovtmp = wOVov = wOvOv = eris_ovov = None

    ov2 = vOVov*2 + vOvOv
    w3 += numpy.einsum('kcjb,kc->bj', ov2, t1)
    #:wooov += numpy.einsum('ibjc,kb->ijkc', ov2, t1)
    #:wovvv -= numpy.einsum('jakb,jc->kabc', ov2, t1)
    for i in range(nocc):
        wooov[i] += lib.dot(t1, ov2[i].reshape(nvir,-1)).reshape(nocc,nocc,nvir).transpose(1,0,2)
    lib.dot(_cp(ov2.transpose(0,2,1,3).reshape(nocc,-1)).T,
            t1, -1, wovvv.reshape(-1,nvir), 1)
    ov2 = None
    ov1 = vOvOv*2 + vOVov
    #:wooov -= numpy.einsum('ibkc,jb->ijkc', ov1, t1)
    #:wovvv += numpy.einsum('jakc,jb->kabc', ov1, t1)
    for i in range(nocc):
        lib.dot(t1, ov1[i].reshape(nvir,-1), -1, wooov[i].reshape(nocc,-1), 1)
    wovvv += lib.dot(_cp(ov1.reshape(nocc,-1)).T,
                     t1).reshape(nvir,-1,nvir,nvir).transpose(1,0,3,2)
    ov1 = None

    woooo += _cp(eris.oooo).transpose(0,2,1,3)
    saved.woooo = woooo
    saved.wooov = wooov
    woooo = wooov = None

    w3 += numpy.einsum('bc,jc->bj', w1, t1)
    w3 -= numpy.einsum('kj,kb->bj', w2, t1)

    eris_ooov = _cp(eris.ooov)
    g2ooov = eris_ooov * 2
    g2ooov -= eris_ooov.transpose(2,0,1,3)
    #:tmp = numpy.einsum('kjla,jb->kabl', g2ooov, t1)
    #:wovvv = numpy.einsum('kabl,lc->kabc', tmp, t1)
    #:wovvv += numpy.einsum('kjla,jlbc->kabc', g2ooov, t2)
    tmp = lib.dot(g2ooov.reshape(nocc,-1).T, t1).reshape(-1,nocc,nvir,nvir).transpose(0,2,3,1)
    lib.dot(_cp(tmp.reshape(-1,nocc)), t1, 1, wovvv.reshape(-1,nvir), 1)
    tmp = None
    lib.dot(_cp(g2ooov.transpose(0,2,1,3).reshape(nocc**2,-1)).T,
            t2.reshape(nocc**2,-1), 1, wovvv.reshape(nov,-1), 1)
    g2ooov = eris_ooov = vOVov = vOvOv = None

    saved.wovvv = wovvv
    saved.w1 = w1
    saved.w2 = w2
    saved.w3 = w3
    saved.w4 = w4
    return saved
Esempio n. 11
0
def make_intermediates(mycc, t1, t2, eris):
    log = logger.Logger(mycc.stdout, mycc.verbose)
    nocc, nvir = t1.shape
    nov = nocc * nvir
    foo = eris.fock[:nocc, :nocc]
    fov = eris.fock[:nocc, nocc:]
    fvv = eris.fock[nocc:, nocc:]

    class _Saved(object):
        pass

    saved = _Saved()

    # As we don't have l2 in memory, hold tau temporarily in memory
    w1 = fvv - numpy.einsum('ja,jb->ba', fov, t1)
    w2 = foo + numpy.einsum('ib,jb->ij', fov, t1)
    w3 = _cp(numpy.einsum('kc,jkbc->bj', fov, t2) * 2 + fov.T)
    w3 -= numpy.einsum('kc,kjbc->bj', fov, t2)
    w3 += reduce(numpy.dot, (t1.T, fov, t1.T))
    w4 = fov.copy()

    eris_ovvv = _cp(eris.ovvv)
    eris_ovvv = lib.unpack_tril(eris_ovvv.reshape(nov, -1))
    eris_ovvv = eris_ovvv.reshape(nocc, nvir, nvir, nvir)

    wovvv = numpy.empty((nocc, nvir, nvir, nvir))
    t2tmp = numpy.empty((nocc, nvir, nocc, nvir))
    for i in range(nocc):
        wovvv[i] = eris_ovvv[i].transpose(1, 0, 2) * 2
        t2tmp[i] = t2[i].transpose(2, 0, 1)
    #:wovvv += numpy.einsum('jabd,kjdc->kabc', eris_ovvv, t2) * -1.5
    tmp = lib.dot(t2tmp.reshape(nov, -1), wovvv.reshape(-1, nvir**2),
                  -1.5 / 2).reshape(-1, nvir, nvir, nvir)
    g2ovvv = tmp
    for i in range(nocc):
        wovvv[i] -= eris_ovvv[i].transpose(1, 2, 0)
        wovvv[i] += tmp[i].transpose(1, 2, 0)
        g2ovvv[i] = eris_ovvv[i] * 2
        g2ovvv[i] -= eris_ovvv[i].transpose(1, 2, 0)
    tmp = t2tmp = None

    w1 += numpy.einsum('jcba,jc->ba', eris_ovvv, t1 * 2)
    w1 -= numpy.einsum('jabc,jc->ba', eris_ovvv, t1)
    #:w3 += numpy.einsum('kdcb,kjdc->bj', eris_ovvv, theta)
    theta = numpy.empty(t2.shape)
    for i in range(nocc):
        theta[i] = t2[i] * 2
        theta[i] -= t2[i].transpose(0, 2, 1)
        lib.dot(eris_ovvv[i].reshape(-1, nvir).T,
                _cp(theta[i].reshape(nocc, -1)).T, 1, w3, 1)

    theta = _cp(theta.transpose(0, 2, 1, 3))
    #:vkbca = numpy.einsum('jdca,kbjd->kbca', g2ovvv, theta)
    vkbca = lib.dot(_cp(theta.reshape(nov, -1)),
                    g2ovvv.reshape(-1,
                                   nvir * nvir)).reshape(-1, nvir, nvir, nvir)
    for i in range(nocc):
        wovvv[i] += vkbca[i].transpose(2, 0, 1)
        wovvv[i] -= vkbca[i].transpose(2, 1, 0) * .5
    vkabc = None

    #:wOVov = numpy.einsum('jbcd,kd->jbkc', eris_ovvv, t1)
    #:wOvOv = numpy.einsum('jdcb,kd->jbkc', eris_ovvv, -t1)
    wOVov = lib.dot(eris_ovvv.reshape(-1, nvir),
                    t1.T).reshape(-1, nvir, nvir, nocc).transpose(0, 1, 3,
                                                                  2).copy()
    for i in range(nocc):
        g2ovvv[i] = eris_ovvv[i].transpose(1, 2, 0) * 2
    wOvOv = lib.dot(g2ovvv.reshape(-1, nvir), -t1.T,
                    .5).reshape(-1, nvir, nvir, nocc).transpose(0, 1, 3,
                                                                2).copy()
    for i in range(nocc):
        g2ovvv[i] -= eris_ovvv[i].transpose(1, 0, 2)
    eris_ovov = _cp(_cp(eris.ovov).transpose(0, 2, 1, 3))
    tau = _ccsd.make_tau(t2, t1, t1)
    #:wooov[:,j0:j1] = numpy.einsum('icbd,jkbd->ijkc', g2ovvv, tau)
    #:woooo[:,:,j0:j1] = numpy.einsum('icjd,klcd->ijkl', eris_ovov, tau)
    tmp = lib.dot(g2ovvv.reshape(-1, nvir**2), tau.reshape(-1, nvir**2).T)
    wooov = _cp(tmp.reshape(-1, nvir, nocc, nocc).transpose(0, 2, 3, 1))
    woooo = lib.dot(eris_ovov.reshape(-1, nvir**2),
                    tau.reshape(-1, nvir**2).T).reshape(-1, nocc, nocc, nocc)
    eris_ovov = eris_ovvv = g2ovvv = tau = tmp = None

    eris_ooov = _cp(eris.ooov)
    eris_ovoo = _cp(eris.ovoo)
    #:woooo += numpy.einsum('icjl,kc->ijkl', eris_ovoo, t1)
    #:wOVov += numpy.einsum('jblk,lc->jbkc', eris_ovoo, -t1)
    for i in range(nocc):
        woooo[i] += lib.dot(t1, eris_ovoo[i].reshape(nvir, -1)).reshape(
            (nocc, ) * 3).transpose(1, 0, 2)
    lib.dot(eris_ovoo.reshape(-1, nocc), t1, -1, wOVov.reshape(-1, nvir), 1)
    #:wooov -= numpy.einsum('ibjl,lkcb->ijkc', eris_ovoo*1.5, t2)
    t2tmp = numpy.empty((nocc, nvir, nocc, nvir))
    for i in range(nocc):
        t2tmp[i] = t2[i].transpose(2, 0, 1)
    tmp_ooov = _cp(-eris_ooov.transpose(2, 0, 1, 3)).reshape(-1, nov)
    lib.dot(tmp_ooov, t2tmp.reshape(nov, -1), 1.5, wooov.reshape(-1, nov), 1)
    t2tmp = None

    g2ooov, tmp_ooov = tmp_ooov.reshape(nocc, nocc, nocc, nvir), None
    g2ooov += eris_ooov * 2
    #:vikjc = numpy.einsum('iklb,jlcb->ikjc', g2ooov, theta)
    vikjc = lib.dot(g2ooov.reshape(-1, nov), theta.reshape(-1, nov).T)
    vikjc = vikjc.reshape(nocc, nocc, nocc, nvir)
    wooov += vikjc.transpose(0, 2, 1, 3)
    wooov -= vikjc * .5
    g2ooov = vikjc = eris_ovoo = None

    w2 += numpy.einsum('ijkb,kb->ij', eris_ooov, t1) * 2
    w2 -= numpy.einsum('kjib,kb->ij', eris_ooov, t1)
    #:w3 -= numpy.einsum('kjlc,klbc->bj', eris_ooov, theta)
    for i in range(nocc):
        lib.dot(
            _cp(theta[i].transpose(1, 2, 0)).reshape(-1, nvir).T,
            eris_ooov[i].reshape(nocc, -1).T, -1, w3, 1)
    #:woooo += numpy.einsum('ikjc,lc->ijkl', eris_ooov, t1)
    #:wOvOv += numpy.einsum('jklb,lc->jbkc', eris_ooov, t1)
    woooo += lib.dot(eris_ooov.reshape(-1, nvir), t1.T).reshape(
        (-1, nocc, nocc, nocc)).transpose(0, 2, 1, 3)
    for i in range(nocc):
        lib.dot(
            _cp(eris_ooov[i].transpose(2, 0, 1)).reshape(-1, nocc), t1, 1,
            wOvOv[i].reshape(-1, nvir), 1)
        wooov[i] += eris_ooov[i].transpose(1, 0, 2) * 2
        wooov[i] -= eris_ooov[i]
    eris_ooov = theta = None

    eris_ovov = _cp(eris.ovov)
    g2ovov = numpy.empty((nocc, nocc, nvir, nvir))
    for i in range(nocc):
        g2ovov[i] = eris_ovov[i].transpose(1, 0, 2) * 2
        g2ovov[i] -= eris_ovov[i].transpose(1, 2, 0)
    tmpw4 = numpy.einsum('klcd,ld->kc', g2ovov, t1)
    #:w1 -= numpy.einsum('kcja,kjcb->ba', g2ovov, t2)
    w1 -= lib.dot(t2.reshape(-1, nvir).T, g2ovov.reshape(-1, nvir))
    w1 -= numpy.einsum('ja,jb->ba', tmpw4, t1)
    #:w2 += numpy.einsum('ibkc,jkbc->ij', g2ovov, t2)
    w2 += lib.dot(g2ovov.reshape(nocc, -1), t2.reshape(nocc, -1).T)
    w2 += numpy.einsum('ib,jb->ij', tmpw4, t1)
    w3 += reduce(numpy.dot, (t1.T, tmpw4, t1.T))
    w4 += tmpw4
    vOVov = eris_ovov.copy()
    #:vOVov += numpy.einsum('jbld,klcd->jbkc', g2ovov, t2)
    #:vOVov -= numpy.einsum('jbld,kldc->jbkc', eris_ovov, t2)
    lib.dot(
        _cp(g2ovov.transpose(0, 2, 1, 3)).reshape(-1, nov),
        _cp(t2.transpose(0, 2, 1, 3).reshape(nov, -1).T), 1,
        vOVov.reshape(nov, -1), 1)
    lib.dot(eris_ovov.reshape(-1, nov),
            _cp(t2.transpose(0, 3, 1, 2).reshape(nov, -1).T), -1,
            vOVov.reshape(nov, -1), 1)
    g2ovov = None

    #:tmp = numpy.einsum('jbld,kd->ljbk', eris_ovov, t1)
    #:wOVov -= numpy.einsum('ljbk,lc->jbkc', tmp, t1)
    #:tmp = numpy.einsum('jdlb,kd->ljbk', eris_ovov, t1)
    #:wOvOv += numpy.einsum('ljbk,lc->jbkc', tmp, t1)
    tmp = numpy.empty((nocc, nvir, nocc))
    for j in range(nocc):
        lib.dot(
            _cp(eris_ovov[j].transpose(1, 0, 2)).reshape(-1, nvir), t1.T, 1,
            tmp.reshape(-1, nocc))
        lib.dot(tmp.reshape(nocc, -1).T, t1, -1, wOVov[j].reshape(-1, nvir), 1)
        lib.dot(eris_ovov[j].reshape(nvir, -1).T, t1.T, 1,
                tmp.reshape(-1, nocc))
        lib.dot(tmp.reshape(nocc, -1).T, t1, 1, wOvOv[j].reshape(-1, nvir), 1)
    tmp = None

    #:vOvOv = numpy.einsum('jdlb,kldc->jbkc', eris_ovov, t2)
    ovovtmp = _cp(eris_ovov.transpose(0, 3, 2, 1).reshape(-1, nov))
    vOvOv = numpy.empty((nocc, nvir, nocc, nvir))
    for j in range(nocc):
        lib.dot(t2[j].reshape(-1, nvir).T, ovovtmp.T, 1,
                vOvOv[j].reshape(nvir, -1))
        vOvOv[j] -= eris.oovv[j].transpose(2, 0, 1)
    ovovtmp = eris_ovov = None
    vOvOv = lib.transpose(vOvOv.reshape(nov,
                                        -1)).reshape(nocc, nvir, nocc, nvir)
    wOVov += vOVov
    wOvOv += vOvOv
    saved.wOVov = wOVov
    saved.wOvOv = wOvOv
    ovovtmp = wOVov = wOvOv = eris_ovov = None

    ov2 = vOVov * 2 + vOvOv
    w3 += numpy.einsum('kcjb,kc->bj', ov2, t1)
    #:wooov += numpy.einsum('ibjc,kb->ijkc', ov2, t1)
    #:wovvv -= numpy.einsum('jakb,jc->kabc', ov2, t1)
    for i in range(nocc):
        wooov[i] += lib.dot(t1, ov2[i].reshape(nvir, -1)).reshape(
            nocc, nocc, nvir).transpose(1, 0, 2)
    lib.dot(
        _cp(ov2.transpose(0, 2, 1, 3).reshape(nocc, -1)).T, t1, -1,
        wovvv.reshape(-1, nvir), 1)
    ov2 = None
    ov1 = vOvOv * 2 + vOVov
    #:wooov -= numpy.einsum('ibkc,jb->ijkc', ov1, t1)
    #:wovvv += numpy.einsum('jakc,jb->kabc', ov1, t1)
    for i in range(nocc):
        lib.dot(t1, ov1[i].reshape(nvir, -1), -1, wooov[i].reshape(nocc, -1),
                1)
    wovvv += lib.dot(_cp(ov1.reshape(nocc, -1)).T,
                     t1).reshape(nvir, -1, nvir, nvir).transpose(1, 0, 3, 2)
    ov1 = None

    woooo += _cp(eris.oooo).transpose(0, 2, 1, 3)
    saved.woooo = woooo
    saved.wooov = wooov
    woooo = wooov = None

    w3 += numpy.einsum('bc,jc->bj', w1, t1)
    w3 -= numpy.einsum('kj,kb->bj', w2, t1)

    eris_ooov = _cp(eris.ooov)
    g2ooov = eris_ooov * 2
    g2ooov -= eris_ooov.transpose(2, 0, 1, 3)
    #:tmp = numpy.einsum('kjla,jb->kabl', g2ooov, t1)
    #:wovvv = numpy.einsum('kabl,lc->kabc', tmp, t1)
    #:wovvv += numpy.einsum('kjla,jlbc->kabc', g2ooov, t2)
    tmp = lib.dot(g2ooov.reshape(nocc, -1).T,
                  t1).reshape(-1, nocc, nvir, nvir).transpose(0, 2, 3, 1)
    lib.dot(_cp(tmp.reshape(-1, nocc)), t1, 1, wovvv.reshape(-1, nvir), 1)
    tmp = None
    lib.dot(
        _cp(g2ooov.transpose(0, 2, 1, 3).reshape(nocc**2, -1)).T,
        t2.reshape(nocc**2, -1), 1, wovvv.reshape(nov, -1), 1)
    g2ooov = eris_ooov = vOVov = vOvOv = None

    saved.wovvv = wovvv
    saved.w1 = w1
    saved.w2 = w2
    saved.w3 = w3
    saved.w4 = w4
    return saved
Esempio n. 12
0
def update_amps(mycc, t1, t2, l1, l2, eris=None, saved=None):
    if saved is None:
        saved = make_intermediates(mycc, t1, t2, eris)
    time1 = time0 = time.clock(), time.time()
    log = logger.Logger(mycc.stdout, mycc.verbose)
    nocc, nvir = t1.shape
    nov = nocc * nvir
    foo = eris.fock[:nocc, :nocc]
    fov = eris.fock[:nocc, nocc:]
    fvv = eris.fock[:nocc, :nocc]

    #:mba = numpy.einsum('klca,klcb->ba', l2, t2*2-t2.transpose(0,1,3,2))
    #:mij = numpy.einsum('ikcd,jkcd->ij', l2, t2*2-t2.transpose(0,1,3,2))
    #:theta = t2*2 - t2.transpose(0,1,3,2)
    theta = _ccsd.make_0132(t2, t2, 2, -1)
    mba = lib.dot(theta.reshape(-1, nvir).T, l2.reshape(-1, nvir))
    mij = lib.dot(l2.reshape(nocc, -1), theta.reshape(nocc, -1).T)
    theta = None
    mba1 = numpy.einsum('jc,jb->bc', l1, t1) + mba
    mij1 = numpy.einsum('kb,jb->kj', l1, t1) + mij
    mia1 = (t1 + numpy.einsum('kc,jkbc->jb', l1, t2) * 2 -
            numpy.einsum('kc,jkcb->jb', l1, t2) -
            reduce(numpy.dot,
                   (t1, l1.T, t1)) - numpy.einsum('bd,jd->jb', mba, t1) -
            numpy.einsum('lj,lb->jb', mij, t1))

    tmp = mycc.add_wvvVV(numpy.zeros_like(l1), l2, eris)
    l2new = numpy.empty((nocc, nocc, nvir, nvir))
    ij = 0
    for i in range(nocc):
        for j in range(i):
            tmp1 = tmp[ij] * .5  # *.5 because of l2+l2.transpose(1,0,3,2) later
            l2new[i, j] = tmp1
            l2new[j, i] = tmp1.T
            ij += 1
        l2new[i, i] = tmp[ij] * .5
        ij += 1
    l1new = (numpy.einsum('ijab,jb->ia', l2new, t1) * 4 -
             numpy.einsum('jiab,jb->ia', l2new, t1) * 2)
    tmp = tmp1 = None

    l1new += eris.fock[:nocc, nocc:]
    l1new += numpy.einsum('ib,ba->ia', l1, saved.w1)
    l1new -= numpy.einsum('ja,ij->ia', l1, saved.w2)
    l1new -= numpy.einsum('ik,ka->ia', mij, saved.w4)
    l1new -= numpy.einsum('ca,ic->ia', mba, saved.w4)
    l1new += numpy.einsum('ijab,bj->ia', l2, saved.w3) * 2
    l1new -= numpy.einsum('ijba,bj->ia', l2, saved.w3)

    l2new += numpy.einsum('ia,jb->ijab', l1, saved.w4)
    #:l2new += numpy.einsum('jibc,ca->jiba', l2, saved.w1)
    #:l2new -= numpy.einsum('kiba,jk->jiba', l2, saved.w2)
    lib.dot(l2.reshape(-1, nvir), saved.w1, 1, l2new.reshape(-1, nvir), 1)
    lib.dot(saved.w2, l2.reshape(nocc, -1), -1, l2new.reshape(nocc, -1), 1)

    eris_ooov = _cp(eris.ooov)
    l1new -= numpy.einsum('jkia,kj->ia', eris_ooov, mij1) * 2
    l1new += numpy.einsum('ikja,kj->ia', eris_ooov, mij1)
    #:l2new -= numpy.einsum('ka,kijb->jiba', l1, eris_ooov)
    lib.dot(
        _cp(eris_ooov.transpose(0, 2, 1, 3).reshape(nocc, -1)).T, l1, -1,
        l2new.reshape(-1, nvir), 1)
    eris_ooov = None

    tau = _ccsd.make_tau(t2, t1, t1)
    #:l2tau = numpy.einsum('ijcd,klcd->ijkl', l2, tau)
    l2tau = lib.dot(l2.reshape(nocc**2, -1),
                    tau.reshape(nocc**2, -1).T).reshape((nocc, ) * 4)
    tau = None
    l2t1 = numpy.einsum('ijcd,kc->ijkd', l2, t1)

    eris_ovvv = _cp(eris.ovvv)
    eris_ovvv = lib.unpack_tril(eris_ovvv.reshape(nov, -1))
    eris_ovvv = eris_ovvv.reshape(nocc, nvir, nvir, nvir)

    l1new += numpy.einsum('iabc,bc->ia', eris_ovvv, mba1) * 2
    l1new -= numpy.einsum('ibca,bc->ia', eris_ovvv, mba1)
    #:l2new += numpy.einsum('ic,jbac->jiba', l1, eris_ovvv)
    tmp = lib.dot(l1,
                  eris_ovvv.reshape(-1, nvir).T).reshape(nocc, -1, nvir, nvir)
    for i in range(nocc):
        l2new[i] += tmp[i].transpose(0, 2, 1)
    #:m4 = numpy.einsum('ijkd,kadb->ijab', l2t1, eris_ovvv)
    m4 = tmp
    lib.dot(_cp(l2t1.reshape(nocc * nocc, -1)),
            _cp(eris_ovvv.transpose(0, 2, 1, 3).reshape(-1, nvir**2)), 1,
            m4.reshape(nocc * nocc, -1))
    l2new -= m4
    l1new -= numpy.einsum('ijab,jb->ia', m4, t1) * 2
    l1new -= numpy.einsum('ijab,ia->jb', m4, t1) * 2
    l1new += numpy.einsum('jiab,jb->ia', m4, t1)
    l1new += numpy.einsum('jiab,ia->jb', m4, t1)
    eris_ovvv = tmp = None

    eris_ovov = _cp(eris.ovov)
    l1new += numpy.einsum('jb,iajb->ia', l1, eris_ovov) * 2
    #:l2new -= numpy.einsum('jbic,ca->jiba', eris_ovov, mba1)
    #:l2new -= numpy.einsum('kajb,ik->ijab', eris_ovov, mij1)
    tmp = lib.dot(eris_ovov.reshape(-1, nvir),
                  mba1).reshape(nocc, nvir, nocc, nvir)
    lib.dot(mij1, eris_ovov.reshape(nocc, -1), 1, tmp.reshape(nocc, -1), 1)
    tmp_oovv = numpy.empty((nocc, nocc, nvir, nvir))
    for i in range(nocc):
        tmp_oovv[i] = eris_ovov[i].transpose(1, 0, 2) * .5
        l2new[i] += tmp_oovv[i]
        l2new[i] -= tmp[i].transpose(1, 0, 2)
    tmp = None
    l1new += numpy.einsum('iajb,jb->ia', eris_ovov, mia1) * 2
    l1new -= numpy.einsum('ibja,jb->ia', eris_ovov, mia1)
    #:m4 = numpy.einsum('kalb,ijkl->ijab', eris_ovov, l2tau)
    lib.dot(l2tau.reshape(nocc * nocc, -1), tmp_oovv.reshape(-1, nvir**2), 1,
            m4.reshape(nocc**2, -1))
    l2new += m4
    l1new += numpy.einsum('ijab,jb->ia', m4, t1) * 4
    l1new -= numpy.einsum('ijba,jb->ia', m4, t1) * 2
    eris_ovov = m4 = tmp_oovv = None

    eris_oovv = _cp(eris.oovv)
    l1new -= numpy.einsum('jb,ijba->ia', l1, eris_oovv)
    eris_oovv = None

    saved_wooov = _cp(saved.wooov)
    #:l1new -= numpy.einsum('jkca,ijkc->ia', l2, saved_wooov)
    l1new -= lib.dot(saved_wooov.reshape(nocc, -1), l2.reshape(-1, nvir))
    saved_wovvv = _cp(saved.wovvv)
    #:l1new += numpy.einsum('kibc,kabc->ia', l2, saved_wovvv)
    for j in range(nocc):
        l1new += lib.dot(l2[j].reshape(nocc, -1),
                         saved_wovvv[j].reshape(nvir, -1).T)
    saved_wooov = saved_wovvv = None

    saved_wOvOv = _cp(saved.wOvOv)
    tmp_ovov = _cp(saved.wOVov) * 2
    tmp_ovov += saved_wOvOv
    #:tmp = l2.transpose(0,2,1,3) - l2.transpose(0,3,1,2)*.5
    #:l2new += numpy.einsum('kcia,kcjb->jiba', tmp, tmp_ovov)
    tmp = numpy.empty((nocc, nvir, nocc, nvir))
    for i in range(nocc):
        tmp[i] = l2[i].transpose(2, 0, 1) * -.5
        tmp[i] += l2[i].transpose(1, 0, 2)
    tmp = lib.dot(tmp_ovov.reshape(-1, nov),
                  tmp.reshape(nov, -1)).reshape(-1, nvir, nocc, nvir)
    #:tmp = numpy.einsum('jkca,ibkc->ijab', l2, saved_wOvOv)
    for i in range(nocc):
        l2new[i] += tmp[i].transpose(1, 0, 2)
        tmp_ovov[i] = l2[i].transpose(2, 0, 1)
    lib.dot(saved_wOvOv.reshape(-1, nov), tmp_ovov.reshape(nov, -1), 1,
            tmp.reshape(nov, -1))
    for i in range(nocc):
        l2new[i] += tmp[i].transpose(1, 2, 0)
        l2new[i] += tmp[i].transpose(1, 0, 2) * .5
    saved_wOvOv = tmp = tmp_ovov = None

    saved_woooo = _cp(saved.woooo)
    #:m3 = numpy.einsum('klab,ijkl->ijab', l2, saved_woooo)
    m3 = lib.dot(saved_woooo.reshape(-1, nocc**2), l2.reshape(nocc**2, -1),
                 .5).reshape(-1, nocc, nvir, nvir)
    l2new += m3
    l1new += numpy.einsum('ijab,jb->ia', m3, t1) * 4
    l1new -= numpy.einsum('ijba,jb->ia', m3, t1) * 2
    saved_woooo = m3 = None

    mo_e = eris.fock.diagonal()
    eia = lib.direct_sum('i-j->ij', mo_e[:nocc], mo_e[nocc:])
    l1new /= eia
    l1new += l1

    #    l2new = l2new + l2new.transpose(1,0,3,2)
    #    l2new /= lib.direct_sum('ia+jb->ijab', eia, eia)
    #    l2new += l2
    ij = 0
    for i in range(nocc):
        for j in range(i):
            dab = lib.direct_sum('a+b->ab', eia[i], eia[j])
            tmp = (l2new[i, j] + l2new[j, i].T) / dab + l2[i, j]
            l2new[i, j] = tmp
            l2new[j, i] = tmp.T
            ij += 1
        dab = lib.direct_sum('a+b->ab', eia[i], eia[i])
        l2new[i, i] = (l2new[i, i] + l2new[i, i].T) / dab + l2[i, i]
        ij += 1

    time0 = log.timer_debug1('update l1 l2', *time0)
    return l1new, l2new
Esempio n. 13
0
def update_lambda(mycc, t1, t2, l1, l2, eris=None, imds=None):
    if imds is None: imds = make_intermediates(mycc, t1, t2, eris)
    time1 = time0 = time.clock(), time.time()
    log = logger.Logger(mycc.stdout, mycc.verbose)
    nocc, nvir = t1.shape
    nov = nocc * nvir
    fov = eris.fock[:nocc,nocc:]
    mo_e_o = eris.mo_energy[:nocc]
    mo_e_v = eris.mo_energy[nocc:] + mycc.level_shift

    theta = t2*2 - t2.transpose(0,1,3,2)
    mba = lib.einsum('klca,klcb->ba', l2, theta)
    mij = lib.einsum('ikcd,jkcd->ij', l2, theta)
    theta = None
    mba1 = numpy.einsum('jc,jb->bc', l1, t1) + mba
    mij1 = numpy.einsum('kb,jb->kj', l1, t1) + mij
    mia1 = t1 + numpy.einsum('kc,jkbc->jb', l1, t2) * 2
    mia1 -= numpy.einsum('kc,jkcb->jb', l1, t2)
    mia1 -= reduce(numpy.dot, (t1, l1.T, t1))
    mia1 -= numpy.einsum('bd,jd->jb', mba, t1)
    mia1 -= numpy.einsum('lj,lb->jb', mij, t1)

    l2new = mycc._add_vvvv(None, l2, eris, with_ovvv=False, t2sym='jiba')
    l1new  = numpy.einsum('ijab,jb->ia', l2new, t1) * 2
    l1new -= numpy.einsum('jiab,jb->ia', l2new, t1)
    l2new *= .5  # *.5 because of l2+l2.transpose(1,0,3,2) in the end
    tmp = tmp1 = None

    w1 = imds.w1 - numpy.diag(mo_e_v)
    w2 = imds.w2 - numpy.diag(mo_e_o)

    l1new += fov
    l1new += numpy.einsum('ib,ba->ia', l1, w1)
    l1new -= numpy.einsum('ja,ij->ia', l1, w2)
    l1new -= numpy.einsum('ik,ka->ia', mij, imds.w4)
    l1new -= numpy.einsum('ca,ic->ia', mba, imds.w4)
    l1new += numpy.einsum('ijab,bj->ia', l2, imds.w3) * 2
    l1new -= numpy.einsum('ijba,bj->ia', l2, imds.w3)

    l2new += numpy.einsum('ia,jb->ijab', l1, imds.w4)
    l2new += lib.einsum('jibc,ca->jiba', l2, w1)
    l2new -= lib.einsum('jk,kiba->jiba', w2, l2)

    eris_ovoo = _cp(eris.ovoo)
    l1new -= numpy.einsum('iajk,kj->ia', eris_ovoo, mij1) * 2
    l1new += numpy.einsum('jaik,kj->ia', eris_ovoo, mij1)
    l2new -= lib.einsum('jbki,ka->jiba', eris_ovoo, l1)
    eris_ovoo = None

    tau = _ccsd.make_tau(t2, t1, t1)
    l2tau = lib.einsum('ijcd,klcd->ijkl', l2, tau)
    tau = None
    l2t1 = lib.einsum('jidc,kc->ijkd', l2, t1)

    max_memory = max(0, mycc.max_memory - lib.current_memory()[0])
    unit = nocc*nvir**2*5
    blksize = min(nocc, max(ccsd.BLKMIN, int(max_memory*.95e6/8/unit)))
    log.debug1('block size = %d, nocc = %d is divided into %d blocks',
               blksize, nocc, int((nocc+blksize-1)/blksize))

    l1new -= numpy.einsum('jb,jiab->ia', l1, _cp(eris.oovv))
    for p0, p1 in lib.prange(0, nvir, blksize):
        eris_ovvv = eris.get_ovvv(slice(None), slice(p0,p1))
        l1new[:,p0:p1] += numpy.einsum('iabc,bc->ia', eris_ovvv, mba1) * 2
        l1new -= numpy.einsum('ibca,bc->ia', eris_ovvv, mba1[p0:p1])
        l2new[:,:,p0:p1] += lib.einsum('jbac,ic->jiba', eris_ovvv, l1)
        m4 = lib.einsum('ijkd,kadb->ijab', l2t1, eris_ovvv)
        l2new[:,:,p0:p1] -= m4
        l1new[:,p0:p1] -= numpy.einsum('ijab,jb->ia', m4, t1) * 2
        l1new -= numpy.einsum('ijab,ia->jb', m4, t1[:,p0:p1]) * 2
        l1new[:,p0:p1] += numpy.einsum('jiab,jb->ia', m4, t1)
        l1new += numpy.einsum('jiab,ia->jb', m4, t1[:,p0:p1])
        eris_ovvv = m4buf = m4 = None

        eris_voov = _cp(eris.ovvo[:,p0:p1].transpose(1,0,3,2))
        l1new[:,p0:p1] += numpy.einsum('jb,aijb->ia', l1, eris_voov) * 2
        l2new[:,:,p0:p1] += eris_voov.transpose(1,2,0,3) * .5
        l2new[:,:,p0:p1] -= lib.einsum('bjic,ca->jiba', eris_voov, mba1)
        l2new[:,:,p0:p1] -= lib.einsum('bjka,ik->jiba', eris_voov, mij1)
        l1new[:,p0:p1] += numpy.einsum('aijb,jb->ia', eris_voov, mia1) * 2
        l1new -= numpy.einsum('bija,jb->ia', eris_voov, mia1[:,p0:p1])
        m4 = lib.einsum('ijkl,aklb->ijab', l2tau, eris_voov)
        l2new[:,:,p0:p1] += m4 * .5
        l1new[:,p0:p1] += numpy.einsum('ijab,jb->ia', m4, t1) * 2
        l1new -= numpy.einsum('ijba,jb->ia', m4, t1[:,p0:p1])

        saved_wvooo = _cp(imds.wvooo[p0:p1])
        l1new -= lib.einsum('ckij,jkca->ia', saved_wvooo, l2[:,:,p0:p1])
        saved_wvovv = _cp(imds.wvvov[p0:p1])
        # Watch out memory usage here, due to the l2 transpose
        l1new[:,p0:p1] += lib.einsum('abkc,kibc->ia', saved_wvovv, l2)
        saved_wvooo = saved_wvovv = None

        saved_wvOOv = _cp(imds.wvOOv[p0:p1])
        tmp_voov = _cp(imds.wVOov[p0:p1]) * 2
        tmp_voov += saved_wvOOv
        tmp = l2.transpose(0,2,1,3) - l2.transpose(0,3,1,2)*.5
        l2new[:,:,p0:p1] += lib.einsum('iakc,bjkc->jiba', tmp, tmp_voov)
        tmp = tmp1 = tmp_ovov = None

        tmp = lib.einsum('jkca,bikc->jiba', l2, saved_wvOOv)
        l2new[:,:,p0:p1] += tmp
        l2new[:,:,p0:p1] += tmp.transpose(1,0,2,3) * .5
        saved_wvOOv = tmp = None

    saved_woooo = _cp(imds.woooo)
    m3 = lib.einsum('ijkl,klab->ijab', saved_woooo, l2)
    l2new += m3 * .5
    l1new += numpy.einsum('ijab,jb->ia', m3, t1) * 2
    l1new -= numpy.einsum('ijba,jb->ia', m3, t1)
    saved_woooo = m3 = None
    #time1 = log.timer_debug1('lambda pass [%d:%d]'%(p0, p1), *time1)

    eia = lib.direct_sum('i-a->ia', mo_e_o, mo_e_v)
    l1new /= eia

#    l2new = l2new + l2new.transpose(1,0,3,2)
#    l2new /= lib.direct_sum('ia+jb->ijab', eia, eia)
#    l2new += l2
    ij = 0
    for i in range(nocc):
        if i > 0:
            l2new[i,:i] += l2new[:i,i].transpose(0,2,1)
            l2new[i,:i] /= lib.direct_sum('a,jb->jab', eia[i], eia[:i])
            l2new[:i,i] = l2new[i,:i].transpose(0,2,1)
        l2new[i,i] = l2new[i,i] + l2new[i,i].T
        l2new[i,i] /= lib.direct_sum('a,b->ab', eia[i], eia[i])

    time0 = log.timer_debug1('update l1 l2', *time0)
    return l1new, l2new
Esempio n. 14
0
def make_intermediates(mycc, t1, t2, eris):
    log = logger.Logger(mycc.stdout, mycc.verbose)
    nocc, nvir = t1.shape
    nov = nocc * nvir
    foo = eris.fock[:nocc, :nocc]
    fov = eris.fock[:nocc, nocc:]
    fvv = eris.fock[nocc:, nocc:]

    class _Saved:
        pass

    saved = _Saved()
    saved._tmpfile = tempfile.NamedTemporaryFile(dir=lib.param.TMPDIR)
    saved.ftmp = ftmp = h5py.File(saved._tmpfile.name)

    def __del__():
        ftmp.close()

    saved.ftmp.__del__ = __del__
    saved.woooo = saved.ftmp.create_dataset("woooo", (nocc, nocc, nocc, nocc), "f8")
    saved.wooov = saved.ftmp.create_dataset("wooov", (nocc, nocc, nocc, nvir), "f8")
    saved.wOVov = saved.ftmp.create_dataset("wOVov", (nocc, nvir, nocc, nvir), "f8")
    saved.wOvOv = saved.ftmp.create_dataset("wOvOv", (nocc, nvir, nocc, nvir), "f8")
    saved.wovvv = saved.ftmp.create_dataset("wovvv", (nocc, nvir, nvir, nvir), "f8")

    # As we don't have l2 in memory, hold tau temporarily in memory
    w1 = fvv - numpy.einsum("ja,jb->ba", fov, t1)
    w2 = foo + numpy.einsum("ib,jb->ij", fov, t1)
    w3 = numpy.einsum("kc,jkbc->bj", fov, t2) * 2 + fov.T
    w3 -= numpy.einsum("kc,kjbc->bj", fov, t2)
    w3 += reduce(numpy.dot, (t1.T, fov, t1.T))
    w4 = fov.copy()

    _tmpfile = tempfile.NamedTemporaryFile(dir=lib.param.TMPDIR)
    fswap = h5py.File(_tmpfile.name)

    time1 = time.clock(), time.time()
    unit = max(
        nocc * nvir ** 2 * 4 + nvir ** 3 * 2,
        nvir ** 3 * 3 + nocc * nvir ** 2,
        nocc * nvir ** 2 * 6 + nocc ** 2 * nvir + nocc ** 3 + nocc ** 2 * nvir,
    )
    max_memory = mycc.max_memory - lib.current_memory()[0]
    blksize = max(ccsd.BLKMIN, int(max_memory * 0.95e6 / 8 / unit))
    log.debug1(
        "ccsd lambda make_intermediates: block size = %d, nocc = %d in %d blocks",
        blksize,
        nocc,
        int((nocc + blksize - 1) // blksize),
    )
    for istep, (p0, p1) in enumerate(prange(0, nocc, blksize)):
        eris_ovvv = _cp(eris.ovvv[p0:p1])
        eris_ovvv = lib.unpack_tril(eris_ovvv.reshape((p1 - p0) * nvir, -1))
        eris_ovvv = eris_ovvv.reshape(p1 - p0, nvir, nvir, nvir)
        w1 += numpy.einsum("jcba,jc->ba", eris_ovvv, t1[p0:p1] * 2)
        w1 -= numpy.einsum("jabc,jc->ba", eris_ovvv, t1[p0:p1])
        #:w3 += numpy.einsum('kdcb,kjdc->bj', eris_ovvv, theta)
        for i in range(p1 - p0):
            theta = t2[p0 + i] * 2
            theta -= t2[p0 + i].transpose(0, 2, 1)
            w3 += lib.dot(eris_ovvv[i].reshape(-1, nvir).T, _cp(theta.reshape(nocc, -1)).T)
        theta = None
        #:wOVov = numpy.einsum('jbcd,kd->jbkc', eris_ovvv, t1)
        #:wOvOv = numpy.einsum('jdcb,kd->jbkc', eris_ovvv, -t1)
        wOVov = lib.dot(eris_ovvv.reshape(-1, nvir), t1.T).reshape(-1, nvir, nvir, nocc).transpose(0, 1, 3, 2).copy()
        g2ovvv = _cp(eris_ovvv.transpose(0, 2, 3, 1))
        wOvOv = lib.dot(g2ovvv.reshape(-1, nvir), -t1.T).reshape(-1, nvir, nvir, nocc).transpose(0, 1, 3, 2).copy()
        for i in range(p1 - p0):
            g2ovvv[i] *= 2
            g2ovvv[i] -= eris_ovvv[i].transpose(1, 0, 2)
        wooov = numpy.empty((p1 - p0, nocc, nocc, nvir))
        woooo = numpy.empty((p1 - p0, nocc, nocc, nocc))
        eris_ovov = _cp(_cp(eris.ovov[p0:p1]).transpose(0, 2, 1, 3))
        for j0, j1 in prange(0, nocc, blksize):
            tau = _ccsd.make_tau(t2[j0:j1], t1[j0:j1], t1)
            #:wooov[:,j0:j1] = numpy.einsum('icbd,jkbd->ijkc', g2ovvv, tau)
            #:woooo[:,:,j0:j1] = numpy.einsum('icjd,klcd->ijkl', eris_ovov, tau)
            tmp = lib.dot(g2ovvv.reshape(-1, nvir ** 2), tau.reshape(-1, nvir ** 2).T)
            wooov[:, j0:j1] = tmp.reshape(-1, nvir, j1 - j0, nocc).transpose(0, 2, 3, 1)
            woooo[:, :, j0:j1] = lib.dot(eris_ovov.reshape(-1, nvir ** 2), tau.reshape(-1, nvir ** 2).T).reshape(
                -1, nocc, j1 - j0, nocc
            )
        eris_ovov = eris_ovvv = g2ovvv = tau = tmp = None
        # ==== mem usage nocc*nvir**2*2 + nocc**2*nvir + nocc**3 + nvir**3*2 + nocc*nvir**2*2

        eris_ooov = _cp(eris.ooov[p0:p1])
        w2[p0:p1] += numpy.einsum("ijkb,kb->ij", eris_ooov, t1) * 2
        w2 -= numpy.einsum("kjib,kb->ij", eris_ooov, t1[p0:p1])
        #:w3 -= numpy.einsum('kjlc,klbc->bj', eris_ooov, theta)
        for i in range(p1 - p0):
            theta = t2[p0 + i].transpose(0, 2, 1) * 2
            theta -= t2[p0 + i]
            w3 -= lib.dot(theta.reshape(-1, nvir).T, eris_ooov[i].reshape(nocc, -1).T)
        theta = None
        #:woooo += numpy.einsum('ikjc,lc->ijkl', eris_ooov, t1)
        #:wOvOv += numpy.einsum('jklb,lc->jbkc', eris_ooov, t1)
        woooo += lib.dot(eris_ooov.reshape(-1, nvir), t1.T).reshape((-1, nocc, nocc, nocc)).transpose(0, 2, 1, 3)
        for i in range(p1 - p0):
            lib.dot(_cp(eris_ooov[i].transpose(2, 0, 1)).reshape(-1, nocc), t1, 1, wOvOv[i].reshape(-1, nvir), 1)
            wooov[i] += eris_ooov[i].transpose(1, 0, 2) * 2
            wooov[i] -= eris_ooov[i]

        eris_ovoo = _cp(eris.ovoo[p0:p1])
        #:woooo += numpy.einsum('icjl,kc->ijkl', eris_ovoo, t1)
        #:wOVov += numpy.einsum('jbkl,lc->jbkc', eris_ovoo, -t1)
        for i in range(p1 - p0):
            woooo[i] += lib.dot(t1, eris_ovoo[i].reshape(nvir, -1)).reshape((nocc,) * 3).transpose(1, 0, 2)
        lib.dot(eris_ovoo.reshape(-1, nocc), t1, -1, wOVov.reshape(-1, nvir), 1)
        #:wooov -= numpy.einsum('iblj,klbc->ijkc', eris_ovoo*1.5, t2)
        tmp_ovoo = _cp(-eris_ovoo.transpose(0, 2, 3, 1)).reshape(-1, nov)
        for j in range(nocc):
            wooov[:, :, j] += lib.dot(tmp_ovoo, t2[j].reshape(-1, nvir), 1.5).reshape(-1, nocc, nvir)
        #:g2ooov = eris_ooov * 2 - eris_ovoo.transpose(0,3,2,1)
        g2ooov, tmp_ovoo = tmp_ovoo.reshape(p1 - p0, nocc, nocc, nvir), None
        g2ooov += eris_ooov * 2
        thetabuf = numpy.empty((blksize, nvir, nocc, nvir))
        vikjc = numpy.empty((p1 - p0, nocc, blksize, nvir))
        for j0, j1 in prange(0, nocc, blksize):
            theta = thetabuf[: j1 - j0]
            for i in range(j1 - j0):
                theta[i] = t2[j0 + i].transpose(1, 0, 2) * 2 - t2[j0 + i].transpose(2, 0, 1)
            #:vikjc = numpy.einsum('iklb,jlcb->ikjc', g2ooov, theta)
            if j1 == j0 + blksize:
                lib.dot(
                    g2ooov.reshape(-1, nov), _cp(theta.reshape(-1, nov)).T, 1, vikjc.reshape((p1 - p0) * nocc, -1), 0
                )
            else:
                vikjc = lib.dot(g2ooov.reshape(-1, nov), _cp(theta.reshape(-1, nov)).T)
                vikjc = vikjc.reshape(p1 - p0, nocc, j1 - j0, nvir)
            wooov[:, j0:j1, :] += vikjc.transpose(0, 2, 1, 3)
            wooov[:, :, j0:j1] -= vikjc * 0.5
        eris_ooov = eris_ovoo = g2ooov = vikjc = theta = thetabuf = None
        # ==== mem usage nocc*nvir**2*3 + nocc**2*nvir + nocc**3 + nocc*nvir**2 + nocc**2*nvir*3

        eris_ovov = _cp(eris.ovov[p0:p1])
        g2ovov = eris_ovov * 2
        g2ovov -= eris_ovov.transpose(0, 3, 2, 1)
        tmpw4 = numpy.einsum("kcld,ld->kc", g2ovov, t1)
        #:w1 -= numpy.einsum('kcja,kjcb->ba', g2ovov, t2[p0:p1])
        w1 -= lib.dot(t2[p0:p1].reshape(-1, nvir).T, _cp(g2ovov.transpose(0, 2, 1, 3).reshape(-1, nvir)))
        w1 -= numpy.einsum("ja,jb->ba", tmpw4, t1[p0:p1])
        #:w2[p0:p1] += numpy.einsum('ibkc,jkbc->ij', g2ovov, t2)
        w2[p0:p1] += lib.dot(_cp(g2ovov.transpose(0, 2, 1, 3)).reshape(p1 - p0, -1), t2.reshape(nocc, -1).T)
        w2[p0:p1] += numpy.einsum("ib,jb->ij", tmpw4, t1)
        w3 += reduce(numpy.dot, (t1[p0:p1].T, tmpw4, t1.T))
        w4[p0:p1] += tmpw4
        vOVov = numpy.empty((nocc, nvir, p1 - p0, nvir))
        #:vOVov += numpy.einsum('jbld,klcd->kcjb', g2ovov, t2)
        #:vOVov -= numpy.einsum('jbld,kldc->kcjb', eris_ovov, t2)
        for j in range(nocc):
            lib.dot(
                _cp(t2[j].transpose(1, 0, 2)).reshape(-1, nov), g2ovov.reshape(-1, nov).T, 1, vOVov[j].reshape(nvir, -1)
            )
            lib.dot(t2[j].reshape(nov, -1).T, eris_ovov.reshape(-1, nov).T, -1, vOVov[j].reshape(nvir, -1), 1)
        vOVov = lib.transpose(vOVov.reshape(nov, -1)).reshape(p1 - p0, nvir, nocc, nvir)
        vOVov += eris_ovov
        g2ovov = tmp = tmpw4 = None
        # ==== mem usage nocc*nvir**2*4 + nocc**2*nvir + nocc**3 + nocc*nvir**2

        #:tmp = numpy.einsum('jbld,kd->jlbk', eris_ovov, t1)
        #:wOVov -= numpy.einsum('jlbk,lc->jbkc', tmp, t1)
        #:tmp = numpy.einsum('jdlb,kd->jlbk', eris_ovov, t1)
        #:wOvOv += numpy.einsum('jlbk,lc->jbkc', tmp, t1)
        tmp = numpy.empty((nocc, nvir, nocc))
        for j in range(p1 - p0):
            lib.dot(_cp(eris_ovov[j].transpose(1, 0, 2)).reshape(-1, nvir), t1.T, 1, tmp.reshape(-1, nocc))
            lib.dot(tmp.reshape(nocc, -1).T, t1, -1, wOVov[j].reshape(-1, nvir), 1)
            lib.dot(eris_ovov[j].reshape(nvir, -1).T, t1.T, 1, tmp.reshape(-1, nocc))
            lib.dot(tmp.reshape(nocc, -1).T, t1, 1, wOvOv[j].reshape(-1, nvir), 1)
        tmp = None

        #:vOvOv = numpy.einsum('jdlb,kldc->kcjb', eris_ovov, t2)
        ovovtmp = _cp(eris_ovov.transpose(0, 3, 2, 1)).reshape(-1, nov)
        vOvOv = numpy.empty((nocc, nvir, p1 - p0, nvir))
        for j in range(nocc):
            lib.dot(t2[j].reshape(-1, nvir).T, ovovtmp.T, 1, vOvOv[j].reshape(nvir, -1))
        ovovtmp = eris_ovov = None
        vOvOv = lib.transpose(vOvOv.reshape(nov, -1)).reshape(p1 - p0, nvir, nocc, nvir)
        vOvOv -= _cp(eris.oovv[p0:p1]).transpose(0, 3, 1, 2)
        wOVov += vOVov
        wOvOv += vOvOv
        saved.wOVov[p0:p1] = wOVov
        saved.wOvOv[p0:p1] = wOvOv
        wOVov = wOvOv = None
        # ==== mem usage nocc*nvir**2*6 + nocc**2*nvir + nocc**3 + nocc**2*nvir

        ov1 = vOvOv * 2 + vOVov
        #:wooov -= numpy.einsum('ibkc,jb->ijkc', ov1, t1)
        for i in range(p1 - p0):
            lib.dot(t1, ov1[i].reshape(nvir, -1), -1, wooov[i].reshape(nocc, -1), 1)
        ov1 = lib.transpose(ov1.reshape(-1, nov))
        fswap["2vOvOv/%d" % istep] = ov1.reshape(nocc, nvir, -1, nvir)
        ov1 = None
        ov2 = vOVov * 2 + vOvOv
        w3 += numpy.einsum("kcjb,kc->bj", ov2, t1[p0:p1])
        #:wooov += numpy.einsum('ibjc,kb->ijkc', ov2, t1)
        for i in range(p1 - p0):
            wooov[i] += lib.dot(t1, ov2[i].reshape(nvir, -1)).reshape(nocc, nocc, nvir).transpose(1, 0, 2)
        ov2 = lib.transpose(ov2.reshape(-1, nov))
        fswap["2vovOV/%d" % istep] = ov2.reshape(nocc, nvir, -1, nvir)
        vOVov = vOvOv = None
        ov2 = None
        # ==== mem usage nocc*nvir**2*5 + nocc**2*nvir + nocc**3

        woooo += _cp(eris.oooo[p0:p1]).transpose(0, 2, 1, 3)
        saved.woooo[p0:p1] = woooo
        saved.wooov[p0:p1] = wooov
        woooo = wooov = None
        time1 = log.timer_debug1("pass1 [%d:%d]" % (p0, p1), *time1)

    w3 += numpy.einsum("bc,jc->bj", w1, t1)
    w3 -= numpy.einsum("kj,kb->bj", w2, t1)

    for p0, p1 in prange(0, nocc, blksize):
        eris_ooov = _cp(eris.ooov[p0:p1])
        g2ooov = eris_ooov * 2
        g2ooov -= eris_ooov.transpose(0, 2, 1, 3)
        #:tmp = numpy.einsum('kjla,jb->kabl', g2ooov, t1)
        #:wovvv = numpy.einsum('kabl,lc->kabc', tmp, t1)
        #:wovvv += numpy.einsum('kjla,jlbc->kabc', g2ooov, t2)
        tmp = (
            lib.dot(_cp(g2ooov.transpose(1, 0, 2, 3).reshape(nocc, -1)).T, t1)
            .reshape(-1, nocc, nvir, nvir)
            .transpose(0, 2, 3, 1)
        )
        wovvv = lib.dot(_cp(tmp.reshape(-1, nocc)), t1).reshape(-1, nvir, nvir, nvir)
        wovvv += lib.dot(_cp(g2ooov.transpose(0, 3, 1, 2).reshape(-1, nocc ** 2)), t2.reshape(nocc ** 2, -1)).reshape(
            -1, nvir, nvir, nvir
        )
        tmp = g2ooov = None
        ov1 = numpy.empty((p1 - p0, nvir, nocc, nvir))
        ov2 = numpy.empty((p1 - p0, nvir, nocc, nvir))
        for istep, (j0, j1) in enumerate(prange(0, nocc, blksize)):
            ov1[:, :, j0:j1] = fswap["2vOvOv/%d" % istep][p0:p1]
            ov2[:, :, j0:j1] = fswap["2vovOV/%d" % istep][p0:p1]
        #:wovvv += numpy.einsum('kcja,jb->kabc', ov1, t1)
        #:wovvv -= numpy.einsum('kbja,jc->kabc', ov2, t1)
        wovvv += (
            lib.dot(_cp(ov1.transpose(0, 1, 3, 2).reshape(-1, nocc)), t1)
            .reshape(-1, nvir, nvir, nvir)
            .transpose(0, 2, 3, 1)
        )
        wovvv -= (
            lib.dot(_cp(ov2.transpose(0, 1, 3, 2).reshape(-1, nocc)), t1)
            .reshape(-1, nvir, nvir, nvir)
            .transpose(0, 2, 1, 3)
        )
        # ==== mem usage nvir**3 + nocc*nvir**2*2
        eris_ooov = ov1 = ov2 = None

        for j0, j1 in prange(0, nocc, blksize):
            eris_ovvv = _cp(eris.ovvv[j0:j1])
            eris_ovvv = lib.unpack_tril(eris_ovvv.reshape((j1 - j0) * nvir, -1))
            eris_ovvv = eris_ovvv.reshape(j1 - j0, nvir, nvir, nvir)
            #:wovvv += numpy.einsum('jabd,kjdc->kabc', eris_ovvv, t2[p0:p1,j0:j1]) * -1.5
            tmp_ovvv = numpy.empty((j1 - j0, nvir, nvir, nvir))
            for i in range(j1 - j0):
                tmp_ovvv[i] = eris_ovvv[i].transpose(1, 0, 2) * 2
            tmp = lib.dot(
                _cp(t2[p0:p1, j0:j1].transpose(0, 3, 1, 2).reshape((p1 - p0) * nvir, -1)),
                tmp_ovvv.reshape(-1, nvir ** 2),
                -1.5 / 2,
            ).reshape(-1, nvir, nvir, nvir)
            wovvv += tmp.transpose(0, 2, 3, 1)
            if p0 == j0:
                for i in range(p1 - p0):
                    tmp_ovvv[i] -= eris_ovvv[i].transpose(1, 2, 0)
                    wovvv[i] += tmp_ovvv[i]
            tmp = tmp_ovvv = None
            g2ovvv = numpy.empty((j1 - j0, nvir, nvir, nvir))
            for i in range(j1 - j0):
                g2ovvv[i] = eris_ovvv[i] * 2
                g2ovvv[i] -= eris_ovvv[i].transpose(1, 2, 0)
            # ==== mem usage nvir**3*3
            eris_ovvv = None
            theta = _cp(t2[p0:p1, j0:j1].transpose(0, 2, 1, 3) * 2)
            for i in range(p1 - p0):
                theta[i] -= t2[p0 + i, j0:j1].transpose(2, 0, 1)
            #:vkbca = numpy.einsum('jdca,kbjd->kbca', g2ovvv, theta)
            vkbca = lib.dot(theta.reshape((p1 - p0) * nvir, -1), g2ovvv.reshape(-1, nvir * nvir)).reshape(
                -1, nvir, nvir, nvir
            )
            wovvv += vkbca.transpose(0, 3, 1, 2)
            wovvv -= vkbca.transpose(0, 3, 2, 1) * 0.5
            # ==== mem usage nvir**3*3 + nocc*nvir**2
            g2ovvv = theta = vkabc = None
        saved.wovvv[p0:p1] = wovvv
        time1 = log.timer_debug1("pass2 [%d:%d]" % (p0, p1), *time1)

    fswap.close()

    saved.w1 = w1
    saved.w2 = w2
    saved.w3 = w3
    saved.w4 = w4
    saved.ftmp.flush()
    return saved
Esempio n. 15
0
def update_lambda(mycc, t1, t2, l1, l2, eris, imds):
    time0 = time.clock(), time.time()
    log = logger.Logger(mycc.stdout, mycc.verbose)
    nocc, nvir = t1.shape
    l1new = np.zeros_like(l1)

    foo = eris.fock[:nocc,:nocc]
    fov = eris.fock[:nocc,nocc:]
    fvv = eris.fock[nocc:,nocc:]

    tau = _ccsd.make_tau(t2, t1, t1)

    theta = t2*2 - t2.transpose(0,1,3,2)
    mvv = lib.einsum('klca,klcb->ba', l2, theta)
    moo = lib.einsum('kicd,kjcd->ij', l2, theta)
    mvv1 = lib.einsum('jc,jb->bc', l1, t1) + mvv
    moo1 = lib.einsum('ic,kc->ik', l1, t1) + moo

    # m3 = einsum('ijab,acbd->ijcd', l2, vvvv)
    #    = einsum('ijab,cadb->ijcd', l2.conj(), vvvv).conj()
    m3 = mycc._add_vvvv(None, l2.conj(), eris, with_ovvv=False).conj()
    m3 += lib.einsum('klab,ikjl->ijab', l2, imds.woooo)
    m3 *= .5

    ovov = np.asarray(eris.ovov)
    l2tau = np.einsum('ijcd,klcd->ijkl', l2, tau)
    m3 += np.einsum('kalb,ijkl->ijab', ovov, l2tau) * .5
    l2tau = None

    l2new = ovov.transpose(0,2,1,3) * .5
    l2new += lib.einsum('ijac,cb->ijab', l2, imds.v1)
    l2new -= lib.einsum('ikab,jk->ijab', l2, imds.v2)
    l2new -= lib.einsum('ca,icjb->ijab', mvv1, ovov)
    l2new -= lib.einsum('ik,kajb->ijab', moo1, ovov)

    ovov = ovov * 2 - ovov.transpose(0,3,2,1)
    l1new -= np.einsum('ik,ka->ia', moo, imds.v4)
    l1new -= np.einsum('ca,ic->ia', mvv, imds.v4)
    l2new += np.einsum('ia,jb->ijab', l1, imds.v4)

    tmp  = t1 + np.einsum('kc,kjcb->jb', l1, theta)
    tmp -= lib.einsum('bd,jd->jb', mvv1, t1)
    tmp -= lib.einsum('lj,lb->jb', moo, t1)
    l1new += np.einsum('jbia,jb->ia', ovov, tmp)
    ovov = tmp = None

    ovvv = np.asarray(eris.get_ovvv())
    l1new += np.einsum('iacb,bc->ia', ovvv, mvv1) * 2
    l1new -= np.einsum('ibca,bc->ia', ovvv, mvv1)
    l2new += lib.einsum('ic,jbca->jiba', l1, ovvv)
    l2t1 = np.einsum('ijcd,kd->ijck', l2, t1)
    m3 -= np.einsum('kbca,ijck->ijab', ovvv, l2t1)
    l2t1 = ovvv = None

    l2new += m3
    l1new += np.einsum('ijab,jb->ia', m3, t1) * 2
    l1new += np.einsum('jiba,jb->ia', m3, t1) * 2
    l1new -= np.einsum('ijba,jb->ia', m3, t1)
    l1new -= np.einsum('jiab,jb->ia', m3, t1)

    ovoo = np.asarray(eris.ovoo)
    l1new -= np.einsum('iajk,kj->ia', ovoo, moo1) * 2
    l1new += np.einsum('jaik,kj->ia', ovoo, moo1)
    l2new -= lib.einsum('ka,jbik->ijab', l1, ovoo)
    ovoo = None

    l2theta = l2*2 - l2.transpose(0,1,3,2)
    l2new += lib.einsum('ikac,jbck->ijab', l2theta, imds.wovvo) * .5
    tmp = lib.einsum('ikca,jbck->ijab', l2, imds.woVVo)
    l2new += tmp * .5
    l2new += tmp.transpose(1,0,2,3)
    l2theta = None

    l1new += fov
    l1new += lib.einsum('ib,ba->ia', l1, imds.v1)
    l1new -= lib.einsum('ja,ij->ia', l1, imds.v2)

    l1new += np.einsum('jb,iabj->ia', l1, eris.ovvo) * 2
    l1new -= np.einsum('jb,ijba->ia', l1, eris.oovv)

    l1new -= lib.einsum('ijbc,bacj->ia', l2, imds.wvvvo)
    l1new -= lib.einsum('kjca,ijck->ia', l2, imds.woovo)

    l1new += np.einsum('ijab,bj->ia', l2, imds.w3) * 2
    l1new -= np.einsum('ijba,bj->ia', l2, imds.w3)

    eia = lib.direct_sum('i-j->ij', foo.diagonal(), fvv.diagonal() + mycc.level_shift)
    l1new /= eia
    l1new += l1

    l2new = l2new + l2new.transpose(1,0,3,2)
    l2new /= lib.direct_sum('ia+jb->ijab', eia, eia)
    l2new += l2

    time0 = log.timer_debug1('update l1 l2', *time0)
    return l1new, l2new
Esempio n. 16
0
def gamma2_outcore(mycc, t1, t2, l1, l2, h5fobj):
    log = logger.Logger(mycc.stdout, mycc.verbose)
    nocc, nvir = t1.shape
    nov = nocc * nvir
    nvir_pair = nvir * (nvir + 1) // 2
    dovov = h5fobj.create_dataset('dovov', (nocc, nvir, nocc, nvir), 'f8')
    dvvvv = h5fobj.create_dataset('dvvvv', (nvir_pair, nvir_pair), 'f8')
    doooo = h5fobj.create_dataset('doooo', (nocc, nocc, nocc, nocc), 'f8')
    doovv = h5fobj.create_dataset('doovv', (nocc, nocc, nvir, nvir), 'f8')
    dovvo = h5fobj.create_dataset('dovvo', (nocc, nvir, nvir, nocc), 'f8')
    dooov = h5fobj.create_dataset('dooov', (nocc, nocc, nocc, nvir), 'f8')

    _tmpfile = tempfile.NamedTemporaryFile(dir=lib.param.TMPDIR)
    fswap = h5py.File(_tmpfile.name)
    mOvOv = fswap.create_dataset('mOvOv', (nocc, nvir, nocc, nvir), 'f8')
    mOVov = fswap.create_dataset('mOVov', (nocc, nvir, nocc, nvir), 'f8')

    moo = numpy.empty((nocc, nocc))
    mvv = numpy.zeros((nvir, nvir))

    max_memory = mycc.max_memory - lib.current_memory()[0]
    unit = nocc * nvir**2 * 5
    blksize = max(ccsd.BLKMIN, int(max_memory * .95e6 / 8 / unit))
    log.debug1(
        'rdm intermediates pass 1: block size = %d, nocc = %d in %d blocks',
        blksize, nocc, int((nocc + blksize - 1) / blksize))
    time1 = time.clock(), time.time()
    for istep, (p0, p1) in enumerate(prange(0, nocc, blksize)):
        #:theta = make_theta(t2[p0:p1])
        #:pOvOv = numpy.einsum('ikca,jkcb->jbia', l2, t2[p0:p1])
        #:pOVov = -numpy.einsum('ikca,jkbc->jbia', l2, t2[p0:p1])
        #:pOVov += numpy.einsum('ikac,jkbc->jbia', l2, theta)
        pOvOv = numpy.empty((nocc, p1 - p0, nvir, nvir))
        pOVov = numpy.empty((nocc, p1 - p0, nvir, nvir))
        t2a = numpy.empty((p1 - p0, nvir, nocc, nvir))
        t2b = numpy.empty((p1 - p0, nvir, nocc, nvir))
        theta = make_theta(t2[p0:p1])
        tmp = numpy.empty_like(t2a)
        for i in range(p1 - p0):
            t2a[i] = t2[p0 + i].transpose(2, 0, 1)
            t2b[i] = t2[p0 + i].transpose(1, 0, 2)
            tmp[i] = theta[i].transpose(1, 0, 2)
        t2a = t2a.reshape(-1, nov)
        t2b = t2b.reshape(-1, nov)
        theta, tmp = tmp.reshape(-1, nov), None
        for i in range(nocc):
            pOvOv[i] = lib.dot(t2a, l2[i].reshape(nov,
                                                  -1)).reshape(-1, nvir, nvir)
            pOVov[i] = lib.dot(t2b, l2[i].reshape(nov, -1),
                               -1).reshape(-1, nvir, nvir)
            pOVov[i] += lib.dot(theta,
                                _cp(l2[i].transpose(0, 2, 1).reshape(
                                    nov, -1))).reshape(-1, nvir, nvir)
        theta = t2a = t2b = None
        mOvOv[p0:p1] = pOvOv.transpose(1, 2, 0, 3)
        mOVov[p0:p1] = pOVov.transpose(1, 2, 0, 3)
        fswap['mvOvO/%d' % istep] = pOvOv.transpose(3, 1, 2, 0)
        fswap['mvOVo/%d' % istep] = pOVov.transpose(3, 1, 2, 0)
        moo[p0:p1] = (numpy.einsum('ljdd->jl', pOvOv) * 2 +
                      numpy.einsum('ljdd->jl', pOVov))
        mvv += (numpy.einsum('llbd->bd', pOvOv[p0:p1]) * 2 +
                numpy.einsum('llbd->bd', pOVov[p0:p1]))
        pOvOv = pOVov = None
        time1 = log.timer_debug1('rdm intermediates pass1 [%d:%d]' % (p0, p1),
                                 *time1)
    mia = (numpy.einsum('kc,ikac->ia', l1, t2) * 2 -
           numpy.einsum('kc,ikca->ia', l1, t2))
    mab = numpy.einsum('kc,kb->cb', l1, t1)
    mij = numpy.einsum('kc,jc->jk', l1, t1) + moo * .5

    gooov = numpy.einsum('ji,ka->jkia', moo * -.5, t1)
    max_memory = mycc.max_memory - lib.current_memory()[0]
    unit = nocc**3 + nocc**2 * nvir + nocc * nvir**2 * 6
    blksize = max(ccsd.BLKMIN, int(max_memory * .95e6 / 8 / unit))
    log.debug1(
        'rdm intermediates pass 2: block size = %d, nocc = %d in %d blocks',
        blksize, nocc, int((nocc + blksize - 1) / blksize))
    for p0, p1 in prange(0, nocc, blksize):
        tau = _ccsd.make_tau(t2[p0:p1], t1[p0:p1], t1)
        #:goooo = numpy.einsum('ijab,klab->klij', l2, tau)*.5
        goooo = lib.dot(tau.reshape(-1, nvir**2),
                        l2.reshape(-1, nvir**2).T, .5)
        goooo = goooo.reshape(-1, nocc, nocc, nocc)
        h5fobj['doooo'][p0:p1] = make_theta(goooo).transpose(0, 2, 1, 3)

        #:gooov[p0:p1] -= numpy.einsum('ib,jkba->jkia', l1, tau)
        #:gooov[p0:p1] -= numpy.einsum('jkba,ib->jkia', l2[p0:p1], t1)
        #:gooov[p0:p1] += numpy.einsum('jkil,la->jkia', goooo, t1*2)
        for i in range(p0, p1):
            gooov[i] -= lib.dot(
                _cp(tau[i - p0].transpose(0, 2, 1).reshape(-1, nvir)),
                l1.T).reshape(nocc, nvir, nocc).transpose(0, 2, 1)
            gooov[i] -= lib.dot(
                _cp(l2[i].transpose(0, 2, 1).reshape(-1, nvir)),
                t1.T).reshape(nocc, nvir, nocc).transpose(0, 2, 1)
        lib.dot(goooo.reshape(-1, nocc), t1, 2, gooov[p0:p1].reshape(-1, nvir),
                1)

        #:goovv -= numpy.einsum('jk,ikab->ijab', mij, tau)
        goovv = numpy.einsum('ia,jb->ijab', mia[p0:p1], t1)
        for i in range(p1 - p0):
            lib.dot(mij, tau[i].reshape(nocc, -1), -1,
                    goovv[i].reshape(nocc, -1), 1)
            goovv[i] += .5 * l2[p0 + i]
            goovv[i] += .5 * tau[i]
        #:goovv -= numpy.einsum('cb,ijac->ijab', mab, t2[p0:p1])
        #:goovv -= numpy.einsum('bd,ijad->ijab', mvv*.5, tau)
        lib.dot(t2[p0:p1].reshape(-1, nvir), mab, -1, goovv.reshape(-1, nvir),
                1)
        lib.dot(tau.reshape(-1, nvir), mvv.T, -.5, goovv.reshape(-1, nvir), 1)
        tau = None
        #==== mem usage nocc**3 + nocc*nvir**2

        pOvOv = _cp(mOvOv[p0:p1])
        pOVov = _cp(mOVov[p0:p1])
        #:gooov[p0:p1,:] += numpy.einsum('jaic,kc->jkia', pOvOv, t1)
        #:gooov[:,p0:p1] -= numpy.einsum('kaic,jc->jkia', pOVov, t1)
        tmp = lib.dot(pOvOv.reshape(-1, nvir),
                      t1.T).reshape(p1 - p0, -1, nocc, nocc)
        gooov[p0:p1, :] += tmp.transpose(0, 3, 2, 1)
        lib.dot(t1, pOVov.reshape(-1, nvir).T, 1, tmp.reshape(nocc, -1), 0)
        gooov[:, p0:p1] -= tmp.reshape(nocc, p1 - p0, nvir,
                                       nocc).transpose(0, 1, 3, 2)
        #:tmp = numpy.einsum('ikac,jc->jika', l2, t1[p0:p1])
        #:gOvVo -= numpy.einsum('jika,kb->jabi', tmp, t1)
        #:gOvvO = numpy.einsum('jkia,kb->jabi', tmp, t1) + pOvOv.transpose(0,3,1,2)
        tmp = tmp.reshape(-1, nocc, nocc, nvir)
        lib.dot(t1[p0:p1], l2.reshape(-1, nvir).T, 1, tmp.reshape(p1 - p0, -1))
        gOvVo = numpy.einsum('ia,jb->jabi', l1, t1[p0:p1])
        gOvvO = numpy.empty((p1 - p0, nvir, nvir, nocc))
        for i in range(p1 - p0):
            gOvVo[i] -= lib.dot(
                _cp(tmp[i].transpose(0, 2, 1).reshape(-1, nocc)),
                t1).reshape(nocc, nvir, -1).transpose(1, 2, 0)
            gOvVo[i] += pOVov[i].transpose(2, 0, 1)
            gOvvO[i] = lib.dot(tmp[i].reshape(nocc, -1).T,
                               t1).reshape(nocc, nvir, -1).transpose(1, 2, 0)
            gOvvO[i] += pOvOv[i].transpose(2, 0, 1)
        tmp = None
        #==== mem usage nocc**3 + nocc*nvir**6
        dovvo[p0:p1] = (gOvVo * 2 + gOvvO).transpose(0, 2, 1, 3)
        gOvvO *= -2
        gOvvO -= gOvVo
        doovv[p0:p1] = gOvvO.transpose(0, 3, 1, 2)
        gOvvO = gOvVo = None

        for j0, j1 in prange(0, nocc, blksize):
            tau2 = _ccsd.make_tau(t2[j0:j1], t1[j0:j1], t1)
            #:goovv += numpy.einsum('ijkl,klab->ijab', goooo[:,:,j0:j1], tau2)
            lib.dot(goooo[:, :, j0:j1].copy().reshape((p1 - p0) * nocc, -1),
                    tau2.reshape(-1, nvir**2), 1, goovv.reshape(-1, nvir**2),
                    1)
            tau2 += numpy.einsum('ia,jb->ijab', t1[j0:j1], t1)
            tau2 = _cp(tau2.transpose(0, 3, 1, 2).reshape(-1, nov))
            #:goovv[:,j0:j1] += numpy.einsum('ibld,jlda->ijab', pOvOv, tau2) * .5
            #:goovv[:,j0:j1] -= numpy.einsum('iald,jldb->ijab', pOVov, tau2) * .5
            goovv[:, j0:j1] += lib.dot(pOvOv.reshape(-1, nov), tau2.T,
                                       .5).reshape(p1 - p0, nvir, -1,
                                                   nvir).transpose(0, 2, 3, 1)
            goovv[:,
                  j0:j1] += lib.dot(pOVov.reshape(-1, nov), tau2.T,
                                    -.5).reshape(p1 - p0, nvir, -1,
                                                 nvir).transpose(0, 2, 1, 3)
            tau2 = None
#==== mem usage nocc**3 + nocc*nvir**2*7
#:goovv += numpy.einsum('iald,jlbd->ijab', pOVov*2+pOvOv, t2) * .5
        pOVov *= 2
        pOVov += pOvOv
        for j in range(nocc):
            tmp = lib.dot(pOVov.reshape(-1, nov),
                          _cp(t2[j].transpose(0, 2, 1).reshape(-1, nvir)), .5)
            goovv[:, j] += tmp.reshape(-1, nvir, nvir)
            tmp = None
        dovov[p0:p1] = make_theta(goovv).transpose(0, 2, 1, 3)
        goooo = goovv = pOvOv = pOVov = None
        time1 = log.timer_debug1('rdm intermediates pass2 [%d:%d]' % (p0, p1),
                                 *time1)

    h5fobj['dooov'][:] = gooov.transpose(0, 2, 1, 3) * 2 - gooov.transpose(
        1, 2, 0, 3)
    gooov = None

    max_memory = mycc.max_memory - lib.current_memory()[0]
    unit = max(nocc**2 * nvir * 2 + nocc * nvir**2 * 2,
               nvir**3 * 2 + nocc * nvir**2)
    blksize = min(nvir, max(ccsd.BLKMIN, int(max_memory * .95e6 / 8 / unit)))
    iobuflen = int(256e6 / 8 / blksize)
    log.debug1(
        'rdm intermediates pass 3: block size = %d, nvir = %d in %d blocks',
        blksize, nocc, int((nvir + blksize - 1) / blksize))
    h5fobj.create_group('dovvv')
    for istep, (p0, p1) in enumerate(prange(0, nvir, blksize)):
        pvOvO = numpy.empty((p1 - p0, nocc, nvir, nocc))
        pvOVo = numpy.empty((p1 - p0, nocc, nvir, nocc))
        ao2mo.outcore._load_from_h5g(fswap['mvOvO'], p0, p1, pvOvO)
        ao2mo.outcore._load_from_h5g(fswap['mvOVo'], p0, p1, pvOVo)
        #:gvovv -= numpy.einsum('aibk,kc->aibc', pvOvO, t1)
        #:gvovv += numpy.einsum('aick,kb->aibc', pvOVo, t1)
        gvovv = lib.dot(pvOVo.reshape(-1, nocc),
                        t1).reshape(-1, nocc, nvir, nvir)
        for i in range(p1 - p0):
            gvovv[i] = gvovv[i].transpose(0, 2, 1)
        lib.dot(pvOvO.reshape(-1, nocc), t1, -1, gvovv.reshape(-1, nvir), 1)
        pvOvO = pvOVo = None
        #==== mem usage nocc**2*nvir*2 + nocc*nvir**2*2

        l2tmp = l2[:, :, p0:p1] * .5
        #:gvvvv = numpy.einsum('ijab,ijcd->abcd', l2tmp, t2)
        #:jabc = numpy.einsum('ijab,ic->jabc', l2tmp, t1)
        #:gvvvv += numpy.einsum('jabc,jd->abcd', jabc, t1)
        gvvvv = lib.dot(l2tmp.reshape(nocc**2, -1).T, t2.reshape(nocc**2, -1))
        jabc = lib.dot(l2tmp.reshape(nocc, -1).T, t1)
        lib.dot(jabc.reshape(nocc, -1).T, t1, 1, gvvvv.reshape(-1, nvir), 1)
        gvvvv = gvvvv.reshape(-1, nvir, nvir, nvir)
        l2tmp = jabc = None

        #:gvovv = numpy.einsum('ja,jibc->aibc', l1[:,p0:p1], t2)
        #:gvovv += numpy.einsum('jibc,ja->aibc', l2, t1[:,p0:p1])
        lib.dot(l1[:, p0:p1].copy().T, t2.reshape(nocc, -1), 1,
                gvovv.reshape(p1 - p0, -1), 1)
        lib.dot(t1[:, p0:p1].copy().T, l2.reshape(nocc, -1), 1,
                gvovv.reshape(p1 - p0, -1), 1)
        tmp = numpy.einsum('ja,jb->ab', l1[:, p0:p1], t1)
        gvovv += numpy.einsum('ab,ic->aibc', tmp, t1)
        gvovv += numpy.einsum('ba,ic->aibc', mvv[:, p0:p1] * .5, t1)
        #:gvovv -= numpy.einsum('adbc,id->aibc', gvvvv, t1*2)
        for j in range(p1 - p0):
            lib.dot(t1, gvvvv[j].reshape(nvir, -1), -2,
                    gvovv[j].reshape(nocc, -1), 1)


# symmetrize dvvvv because it is symmetrized in ccsd_grad and make_rdm2 anyway
#:dvvvv = .5*(gvvvv+gvvvv.transpose(0,1,3,2))
#:dvvvv = .5*(dvvvv+dvvvv.transpose(1,0,3,2))
# now dvvvv == dvvvv.transpose(2,3,0,1) == dvvvv.transpose(0,1,3,2) == dvvvv.transpose(1,0,3,2)
        tmp = numpy.empty((nvir, nvir, nvir))
        tmp1 = numpy.empty((nvir, nvir, nvir))
        tmpvvvv = numpy.empty((p1 - p0, nvir, nvir_pair))
        for i in range(p1 - p0):
            make_theta(gvvvv[i:i + 1], out=tmp)
            tmp1[:] = tmp.transpose(1, 0, 2)
            _ccsd.precontract(tmp1, diag_fac=2, out=tmpvvvv[i])
        # tril of (dvvvv[p0:p1,p0:p1]+dvvvv[p0:p1,p0:p1].T)
        for i in range(p0, p1):
            for j in range(p0, i):
                tmpvvvv[i - p0, j] += tmpvvvv[j - p0, i]
            tmpvvvv[i - p0, i] *= 2
        for i in range(p0, p1):
            off = i * (i + 1) // 2
            if p0 > 0:
                tmpvvvv[i - p0, :p0] += dvvvv[off:off + p0]
            dvvvv[off:off + i + 1] = tmpvvvv[i - p0, :i + 1] * .25
        for i in range(p1, nvir):
            off = i * (i + 1) // 2
            dvvvv[off + p0:off + p1] = tmpvvvv[:, i]
        tmp = tmp1 = tmpvvvv = None
        #==== mem usage nvir**3 + nocc*nvir**2
        gvvov = make_theta(gvovv).transpose(0, 2, 1, 3)
        ao2mo.outcore._transpose_to_h5g(h5fobj, 'dovvv/%d' % istep,
                                        gvvov.reshape(-1, nov), iobuflen)
        gvvvv = None
        gvovv = None
        time1 = log.timer_debug1('rdm intermediates pass3 [%d:%d]' % (p0, p1),
                                 *time1)

    del (fswap['mOvOv'])
    del (fswap['mOVov'])
    del (fswap['mvOvO'])
    del (fswap['mvOVo'])
    fswap.close()
    _tmpfile = None
    return (h5fobj['dovov'], h5fobj['dvvvv'], h5fobj['doooo'], h5fobj['doovv'],
            h5fobj['dovvo'], None, h5fobj['dovvv'], h5fobj['dooov'])
Esempio n. 17
0
def make_intermediates(mycc, t1, t2, eris):
    nocc, nvir = t1.shape
    foo = eris.fock[:nocc,:nocc]
    fov = eris.fock[:nocc,nocc:]
    fvo = eris.fock[nocc:,:nocc]
    fvv = eris.fock[nocc:,nocc:]

    tau = _ccsd.make_tau(t2, t1, t1)
    ovov = np.asarray(eris.ovov)
    ovoo = np.asarray(eris.ovoo)
    ovov1 = ovov * 2 - ovov.transpose(0,3,2,1)
    ovoo1 = ovoo * 2 - ovoo.transpose(2,1,0,3)

    v1  = fvv - lib.einsum('ja,jb->ba', fov, t1)
    v1 -= lib.einsum('jakc,jkbc->ba', ovov1, tau)
    v2  = foo + lib.einsum('ib,jb->ij', fov, t1)
    v2 += lib.einsum('ibkc,jkbc->ij', ovov1, tau)
    v2 += np.einsum('kbij,kb->ij', ovoo1, t1)
    v4 = fov + np.einsum('jbkc,kc->jb', ovov1, t1)

    v5  = np.einsum('kc,jkbc->bj', fov, t2) * 2
    v5 -= np.einsum('kc,jkcb->bj', fov, t2)
    v5 += fvo
    v5 += lib.einsum('kc,kb,jc->bj', v4, t1, t1)
    v5 -= lib.einsum('lckj,klbc->bj', ovoo1, t2)

    oooo = np.asarray(eris.oooo)
    woooo  = lib.einsum('icjl,kc->ikjl', ovoo, t1)
    woooo += lib.einsum('jcil,kc->iljk', ovoo, t1)
    woooo += oooo.copy()
    woooo += lib.einsum('icjd,klcd->ikjl', ovov, tau)

    theta = t2*2 - t2.transpose(0,1,3,2)
    v4OVvo  = lib.einsum('ldjb,klcd->jbck', ovov1, t2)
    v4OVvo -= lib.einsum('ldjb,kldc->jbck', ovov, t2)
    v4OVvo += np.asarray(eris.ovvo)

    v4oVVo  = lib.einsum('jdlb,kldc->jbck', ovov, t2)
    v4oVVo -= np.asarray(eris.oovv).transpose(0,3,2,1)

    v4ovvo = v4OVvo*2 + v4oVVo
    w3 = np.einsum('jbck,jb->ck', v4ovvo, t1)

    woovo  = lib.einsum('ibck,jb->ijck', v4ovvo, t1)
    woovo = woovo - woovo.transpose(0,3,2,1)
    woovo += lib.einsum('ibck,jb->ikcj', v4OVvo-v4oVVo, t1)
    woovo += ovoo1.conj().transpose(3,2,1,0)

    woovo += lib.einsum('lcik,jlbc->ikbj', ovoo1, theta)
    woovo -= lib.einsum('lcik,jlbc->ijbk', ovoo1, t2)
    woovo -= lib.einsum('iclk,ljbc->ijbk', ovoo1, t2)

    wvvvo  = lib.einsum('jack,jb->back', v4ovvo, t1)
    wvvvo = wvvvo - wvvvo.transpose(2,1,0,3)
    wvvvo += lib.einsum('jack,jb->cabk', v4OVvo-v4oVVo, t1)
    wvvvo -= lib.einsum('lajk,jlbc->cabk', ovoo1, tau)

    wOVvo  = v4OVvo
    woVVo  = v4oVVo
    wOVvo -= np.einsum('jbld,kd,lc->jbck', ovov, t1, t1)
    woVVo += np.einsum('jdlb,kd,lc->jbck', ovov, t1, t1)
    wOVvo -= lib.einsum('jblk,lc->jbck', ovoo, t1)
    woVVo += lib.einsum('lbjk,lc->jbck', ovoo, t1)
    v4ovvo = v4OVvo = v4oVVo = None

    ovvv = np.asarray(eris.get_ovvv())
    wvvvo += lib.einsum('kacd,kjbd->bacj', ovvv, t2) * 1.5

    wOVvo += lib.einsum('jbcd,kd->jbck', ovvv, t1)
    woVVo -= lib.einsum('jdcb,kd->jbck', ovvv, t1)

    ovvv = ovvv*2 - ovvv.transpose(0,3,2,1)
    v1 += np.einsum('jcba,jc->ba', ovvv, t1)
    v5 += lib.einsum('kdbc,jkcd->bj', ovvv, t2)
    woovo += lib.einsum('idcb,jkdb->ijck', ovvv, tau)

    tmp = lib.einsum('kdca,jkbd->cabj', ovvv, theta)
    wvvvo -= tmp
    wvvvo += tmp.transpose(2,1,0,3) * .5
    wvvvo -= ovvv.conj().transpose(3,2,1,0)
    ovvv = tmp = None

    w3 += v5
    w3 += np.einsum('cb,jb->cj', v1, t1)
    w3 -= np.einsum('jk,jb->bk', v2, t1)

    class _IMDS: pass
    imds = _IMDS()
    imds.ftmp = lib.H5TmpFile()
    dtype = np.result_type(t2, eris.vvvv).char
    imds.woooo = imds.ftmp.create_dataset('woooo', (nocc,nocc,nocc,nocc), dtype)
    imds.wovvo = imds.ftmp.create_dataset('wovvo', (nocc,nvir,nvir,nocc), dtype)
    imds.woVVo = imds.ftmp.create_dataset('woVVo', (nocc,nvir,nvir,nocc), dtype)
    imds.woovo = imds.ftmp.create_dataset('woovo', (nocc,nocc,nvir,nocc), dtype)
    imds.wvvvo = imds.ftmp.create_dataset('wvvvo', (nvir,nvir,nvir,nocc), dtype)

    imds.woooo[:] = woooo
    imds.wovvo[:] = wOVvo*2 + woVVo
    imds.woVVo[:] = woVVo
    imds.woovo[:] = woovo
    imds.wvvvo[:] = wvvvo
    imds.v1 = v1
    imds.v2 = v2
    imds.w3 = w3
    imds.v4 = v4
    imds.ftmp.flush()
    return imds
Esempio n. 18
0
def gamma2_incore(mycc, t1, t2, l1, l2):
    log = logger.Logger(mycc.stdout, mycc.verbose)
    nocc, nvir = t1.shape
    nov = nocc * nvir

    time1 = time.clock(), time.time()
    #:theta = make_theta(t2)
    #:mOvOv = numpy.einsum('ikca,jkcb->jbia', l2, t2)
    #:mOVov = -numpy.einsum('ikca,jkbc->jbia', l2, t2)
    #:mOVov += numpy.einsum('ikac,jkbc->jbia', l2, theta)
    l2a = numpy.empty((nocc, nvir, nocc, nvir))
    t2a = numpy.empty((nocc, nvir, nocc, nvir))
    for i in range(nocc):
        l2a[i] = l2[i].transpose(2, 0, 1)
        t2a[i] = t2[i].transpose(2, 0, 1)
    mOvOv = lib.dot(t2a.reshape(-1, nov),
                    l2a.reshape(-1, nov).T).reshape(nocc, nvir, nocc, nvir)
    for i in range(nocc):
        t2a[i] = t2[i].transpose(1, 0, 2)
    mOVov = lib.dot(t2a.reshape(-1, nov),
                    l2a.reshape(-1, nov).T,
                    -1).reshape(nocc, nvir, nocc, nvir)
    theta = t2a
    for i in range(nocc):
        l2a[i] = l2[i].transpose(1, 0, 2)
        theta[i] *= 2
        theta[i] -= t2[i].transpose(2, 0, 1)
    lib.dot(theta.reshape(-1, nov),
            l2a.reshape(nov, -1).T, 1, mOVov.reshape(nov, -1), 1)
    theta = l2a = t2a = None
    moo = (numpy.einsum('jdld->jl', mOvOv) * 2 +
           numpy.einsum('jdld->jl', mOVov))
    mvv = (numpy.einsum('lbld->bd', mOvOv) * 2 +
           numpy.einsum('lbld->bd', mOVov))
    mia = (numpy.einsum('kc,ikac->ia', l1, t2) * 2 -
           numpy.einsum('kc,ikca->ia', l1, t2))
    mab = numpy.einsum('kc,kb->cb', l1, t1)
    mij = numpy.einsum('kc,jc->jk', l1, t1) + moo * .5

    gooov = numpy.zeros((nocc, nocc, nocc, nvir))
    tau = _ccsd.make_tau(t2, t1, t1)
    #:goooo = numpy.einsum('ijab,klab->klij', l2, tau)*.5
    goooo = lib.dot(tau.reshape(-1, nvir**2), l2.reshape(-1, nvir**2).T, .5)
    goooo = goooo.reshape(-1, nocc, nocc, nocc)
    doooo = _cp(make_theta(goooo).transpose(0, 2, 1, 3))

    #:gooov -= numpy.einsum('ib,kjab->jkia', l1, tau)
    #:gooov -= numpy.einsum('kjab,ib->jkia', l2, t1)
    #:gooov += numpy.einsum('jkil,la->jkia', goooo, t1*2)
    gooov = lib.dot(_cp(tau.reshape(-1, nvir)), l1.T, -1)
    lib.dot(_cp(l2.reshape(-1, nvir)), t1.T, -1, gooov, 1)
    gooov = gooov.reshape(nocc, nocc, nvir, nocc)
    tmp = numpy.einsum('ji,ka->jkia', moo * -.5, t1)
    tmp += gooov.transpose(1, 0, 3, 2)
    gooov, tmp = tmp, None
    lib.dot(goooo.reshape(-1, nocc), t1, 2, gooov.reshape(-1, nvir), 1)

    goovv = numpy.einsum('ia,jb->ijab', mia, t1)
    for i in range(nocc):
        goovv[i] += .5 * l2[i]
        goovv[i] += .5 * tau[i]
    #:goovv -= numpy.einsum('jk,kiba->jiba', mij, tau)
    lib.dot(mij, tau.reshape(nocc, -1), -1, goovv.reshape(nocc, -1), 1)
    #:goovv -= numpy.einsum('cb,ijac->ijab', mab, t2)
    #:goovv -= numpy.einsum('bd,ijad->ijab', mvv*.5, tau)
    lib.dot(t2.reshape(-1, nvir), mab, -1, goovv.reshape(-1, nvir), 1)
    lib.dot(tau.reshape(-1, nvir), mvv.T, -.5, goovv.reshape(-1, nvir), 1)
    tau = None

    #:gooov += numpy.einsum('jaic,kc->jkia', mOvOv, t1)
    #:gooov -= numpy.einsum('kaic,jc->jkia', mOVov, t1)
    tmp = lib.dot(mOvOv.reshape(-1, nvir), t1.T).reshape(nocc, -1, nocc, nocc)
    gooov += tmp.transpose(0, 3, 2, 1)
    lib.dot(t1, mOVov.reshape(-1, nvir).T, 1, tmp.reshape(nocc, -1), 0)
    gooov -= tmp.reshape(nocc, nocc, nvir, nocc).transpose(0, 1, 3, 2)
    dooov = gooov.transpose(0, 2, 1, 3) * 2 - gooov.transpose(1, 2, 0, 3)
    gooov = None
    #:tmp = numpy.einsum('ikac,jc->jaik', l2, t1)
    #:gOvVo -= numpy.einsum('jaik,kb->jabi', tmp, t1)
    #:gOvvO = numpy.einsum('jaki,kb->jabi', tmp, t1) + mOvOv.transpose(0,3,1,2)
    tmp = tmp.reshape(nocc, nocc, nocc, nvir)
    lib.dot(t1, l2.reshape(-1, nvir).T, 1, tmp.reshape(nocc, -1))
    gOvVo = numpy.einsum('ia,jb->jabi', l1, t1)
    gOvvO = numpy.empty((nocc, nvir, nvir, nocc))
    for i in range(nocc):
        gOvVo[i] -= lib.dot(_cp(tmp[i].transpose(0, 2, 1).reshape(-1, nocc)),
                            t1).reshape(nocc, nvir, -1).transpose(1, 2, 0)
        gOvVo[i] += mOVov[i].transpose(2, 0, 1)
        gOvvO[i] = lib.dot(tmp[i].reshape(nocc, -1).T,
                           t1).reshape(nocc, nvir, -1).transpose(1, 2, 0)
        gOvvO[i] += mOvOv[i].transpose(2, 0, 1)
    tmp = None

    dovvo = numpy.empty((nocc, nvir, nvir, nocc))
    doovv = numpy.empty((nocc, nocc, nvir, nvir))
    for i in range(nocc):
        tmp = gOvVo[i] * 2 + gOvvO[i]
        dovvo[i] = tmp.transpose(1, 0, 2)
        tmp = gOvvO[i] * -2 - gOvVo[i]
        doovv[i] = tmp.transpose(2, 0, 1)
    gOvvO = gOvVo = None

    tau2 = _ccsd.make_tau(t2, t1, t1)
    #:goovv += numpy.einsum('ijkl,klab->ijab', goooo[:,:,j0:j1], tau2)
    lib.dot(goooo.reshape(nocc * nocc, -1), tau2.reshape(-1, nvir**2), 1,
            goovv.reshape(-1, nvir**2), 1)
    tau2 += numpy.einsum('ia,jb->ijab', t1, t1)
    tau2p = tau2.reshape(nocc, nvir, nocc, nvir)
    for i in range(nocc):
        tau2p[i] = tau2[i].transpose(2, 0, 1)
    tau2, tau2p = tau2p.reshape(nov, -1), None
    #:goovv += numpy.einsum('ibld,jlda->ijab', mOvOv, tau2) * .5
    #:goovv -= numpy.einsum('iald,jldb->ijab', mOVov, tau2) * .5
    tmp = lib.dot(mOvOv.reshape(-1, nov), tau2.T,
                  .5).reshape(nocc, nvir, -1, nvir)
    for i in range(nocc):
        tmp[i] = goovv[i].transpose(1, 0, 2) + tmp[i].transpose(2, 1, 0)
    goovv, tmp = tmp, None
    lib.dot(mOVov.reshape(-1, nov), tau2.T, -.5, goovv.reshape(nov, -1), 1)

    #:goovv += numpy.einsum('iald,jlbd->ijab', mOVov*2+mOvOv, t2) * .5
    t2a, tau2 = tau2.reshape(nocc, nvir, nocc, nvir), None
    for i in range(nocc):
        t2a[i] = t2[i].transpose(1, 0, 2)
    tmp = mOVov * 2
    tmp += mOvOv
    lib.dot(tmp.reshape(-1, nov), t2a.reshape(nov, -1), .5,
            goovv.reshape(nov, -1), 1)
    t2a = tmp = None
    for i in range(nocc):
        goovv[i] = goovv[i] * 2 - goovv[i].transpose(2, 1, 0)
    dovov = goovv
    goooo = goovv = None

    #:gvovv += numpy.einsum('aick,kb->aibc', pvOVo, t1)
    mOVov = lib.transpose(mOVov.reshape(nov, -1))
    gvovv = lib.dot(mOVov.reshape(nocc, -1).T,
                    t1).reshape(nvir, nocc, nvir, nvir)
    mOVov = None
    tmp = numpy.einsum('ja,jb->ab', l1, t1)
    #:gvovv += numpy.einsum('ab,ic->aibc', tmp, t1)
    #:gvovv += numpy.einsum('ba,ic->aibc', mvv, t1*.5)
    for i in range(nvir):
        gvovv[i] += numpy.einsum('b,ic->icb', tmp[i], t1)
        gvovv[i] += numpy.einsum('b,ic->icb', mvv[:, i] * .5, t1)
        gvovv[i] = gvovv[i].transpose(0, 2, 1)

    #:gvovv += numpy.einsum('ja,jibc->aibc', l1, t2)
    #:gvovv += numpy.einsum('jibc,ja->aibc', l2, t1)
    #:gvovv -= numpy.einsum('aibk,kc->aibc', pvOvO, t1)
    mOvOv = lib.transpose(mOvOv.reshape(nov, -1))
    lib.dot(mOvOv.reshape(nocc, -1).T, t1, -1, gvovv.reshape(-1, nvir), 1)
    mOvOv = None
    lib.dot(l1.T, t2.reshape(nocc, -1), 1, gvovv.reshape(nvir, -1), 1)
    lib.dot(t1.T, l2.reshape(nocc, -1), 1, gvovv.reshape(nvir, -1), 1)

    tmp = numpy.empty((nocc, nvir, nvir))
    for i in range(nvir):
        #:gvovv*2 - gvovv.transpose(0,1,3,2)
        gvovv[i] = _ccsd.make_021(gvovv[i], gvovv[i], 2, -1, out=tmp)

    #:gvvvv = numpy.einsum('ijab,ijcd->abcd', l2, t2)*.5
    #:jabc = numpy.einsum('ijab,ic->jabc', l2, t1) * .5
    #:gvvvv += numpy.einsum('jabc,jd->abcd', jabc, t1)
    #:gvovv -= numpy.einsum('adbc,id->aibc', gvvvv, t1*2)
    tau = _ccsd.make_tau(t2, t1, t1)
    theta = make_theta(tau)
    tau = None
    l2tmp = _ccsd.pack_tril(l2.reshape(-1, nvir, nvir))
    gtmp = lib.dot(l2tmp.T, theta.reshape(nocc**2, -1),
                   .5).reshape(-1, nvir, nvir)
    l2tmp = theta = None
    nvir_pair = nvir * (nvir + 1) // 2
    tmp = numpy.empty((nvir, nvir, nvir))
    tmp1 = numpy.empty((nvir, nvir, nvir))
    tmptril = numpy.empty((nvir, nvir_pair))
    diag_idx = numpy.arange(nvir)
    diag_idx = diag_idx * (diag_idx + 1) // 2 + diag_idx

    dvvvv = numpy.empty((nvir_pair, nvir_pair))
    dovvv = numpy.empty((nocc, nvir, nvir, nvir))
    # dvvov = (gvovv*2 - gvovv.transpose(0,1,3,2)).transpose(0,2,1,3)
    # dovvv = dvvov.transpose(2,3,0,1)
    p0 = 0
    for i in range(nvir):
        tmp[:i + 1] = gtmp[p0:p0 + i + 1]
        for j in range(i + 1, nvir):
            tmp[j] = gtmp[j * (j + 1) // 2 + i].T
        lib.dot(t1, tmp.reshape(nvir, -1), -2, gvovv[i].reshape(nocc, -1), 1)
        dovvv[:, :, i] = gvovv[i].transpose(0, 2, 1)

        #:gvvvv[i] = (tmp*2-tmp.transpose(0,2,1)).transpose(1,0,2)
        #:gvvvv = .5*(gvvvv+gvvvv.transpose(0,1,3,2))
        #:dvvvv = .5*(gvvvv+gvvvv.transpose(1,0,3,2))
        tmp1[:] = tmp.transpose(1, 0, 2)
        _ccsd.precontract(tmp1, diag_fac=2, out=tmptril)
        dvvvv[p0:p0 + i] += tmptril[:i]
        dvvvv[p0:p0 + i] *= .25
        dvvvv[i * (i + 1) // 2 + i] = tmptril[i] * .5
        for j in range(i + 1, nvir):
            dvvvv[j * (j + 1) // 2 + i] = tmptril[j]
        p0 += i + 1
    gtmp = tmp = tmp1 = tmptril = gvovv = None
    dvvov = dovvv.transpose(2, 3, 0, 1)
    return (dovov, dvvvv, doooo, doovv, dovvo, dvvov, dovvv, dooov)
Esempio n. 19
0
def update_lambda(mycc, t1, t2, l1, l2, eris=None, imds=None):
    if imds is None: imds = make_intermediates(mycc, t1, t2, eris)
    time0 = logger.process_clock(), logger.perf_counter()
    log = logger.Logger(mycc.stdout, mycc.verbose)
    nocc, nvir = t1.shape
    fov = eris.fock[:nocc, nocc:]
    mo_e_o = eris.mo_energy[:nocc]
    mo_e_v = eris.mo_energy[nocc:] + mycc.level_shift

    theta = t2 * 2 - t2.transpose(0, 1, 3, 2)
    mba = lib.einsum('klca,klcb->ba', l2, theta)
    mij = lib.einsum('ikcd,jkcd->ij', l2, theta)
    theta = None
    mba1 = numpy.einsum('jc,jb->bc', l1, t1) + mba
    mij1 = numpy.einsum('kb,jb->kj', l1, t1) + mij
    mia1 = t1 + numpy.einsum('kc,jkbc->jb', l1, t2) * 2
    mia1 -= numpy.einsum('kc,jkcb->jb', l1, t2)
    mia1 -= reduce(numpy.dot, (t1, l1.T, t1))
    mia1 -= numpy.einsum('bd,jd->jb', mba, t1)
    mia1 -= numpy.einsum('lj,lb->jb', mij, t1)

    l2new = mycc._add_vvvv(None, l2, eris, with_ovvv=False, t2sym='jiba')
    l1new = numpy.einsum('ijab,jb->ia', l2new, t1) * 2
    l1new -= numpy.einsum('jiab,jb->ia', l2new, t1)
    l2new *= .5  # *.5 because of l2+l2.transpose(1,0,3,2) in the end
    tmp = None

    w1 = imds.w1 - numpy.diag(mo_e_v)
    w2 = imds.w2 - numpy.diag(mo_e_o)

    l1new += fov
    l1new += numpy.einsum('ib,ba->ia', l1, w1)
    l1new -= numpy.einsum('ja,ij->ia', l1, w2)
    l1new -= numpy.einsum('ik,ka->ia', mij, imds.w4)
    l1new -= numpy.einsum('ca,ic->ia', mba, imds.w4)
    l1new += numpy.einsum('ijab,bj->ia', l2, imds.w3) * 2
    l1new -= numpy.einsum('ijba,bj->ia', l2, imds.w3)

    l2new += numpy.einsum('ia,jb->ijab', l1, imds.w4)
    l2new += lib.einsum('jibc,ca->jiba', l2, w1)
    l2new -= lib.einsum('jk,kiba->jiba', w2, l2)

    eris_ovoo = _cp(eris.ovoo)
    l1new -= numpy.einsum('iajk,kj->ia', eris_ovoo, mij1) * 2
    l1new += numpy.einsum('jaik,kj->ia', eris_ovoo, mij1)
    l2new -= lib.einsum('jbki,ka->jiba', eris_ovoo, l1)
    eris_ovoo = None

    tau = _ccsd.make_tau(t2, t1, t1)
    l2tau = lib.einsum('ijcd,klcd->ijkl', l2, tau)
    tau = None
    l2t1 = lib.einsum('jidc,kc->ijkd', l2, t1)

    max_memory = max(0, mycc.max_memory - lib.current_memory()[0])
    unit = nocc * nvir**2 * 5
    blksize = min(nocc, max(ccsd.BLKMIN, int(max_memory * .95e6 / 8 / unit)))
    log.debug1('block size = %d, nocc = %d is divided into %d blocks', blksize,
               nocc, int((nocc + blksize - 1) / blksize))

    l1new -= numpy.einsum('jb,jiab->ia', l1, _cp(eris.oovv))
    for p0, p1 in lib.prange(0, nvir, blksize):
        eris_ovvv = eris.get_ovvv(slice(None), slice(p0, p1))
        l1new[:, p0:p1] += numpy.einsum('iabc,bc->ia', eris_ovvv, mba1) * 2
        l1new -= numpy.einsum('ibca,bc->ia', eris_ovvv, mba1[p0:p1])
        l2new[:, :, p0:p1] += lib.einsum('jbac,ic->jiba', eris_ovvv, l1)
        m4 = lib.einsum('ijkd,kadb->ijab', l2t1, eris_ovvv)
        l2new[:, :, p0:p1] -= m4
        l1new[:, p0:p1] -= numpy.einsum('ijab,jb->ia', m4, t1) * 2
        l1new -= numpy.einsum('ijab,ia->jb', m4, t1[:, p0:p1]) * 2
        l1new[:, p0:p1] += numpy.einsum('jiab,jb->ia', m4, t1)
        l1new += numpy.einsum('jiab,ia->jb', m4, t1[:, p0:p1])
        eris_ovvv = m4 = None

        eris_voov = _cp(eris.ovvo[:, p0:p1].transpose(1, 0, 3, 2))
        l1new[:, p0:p1] += numpy.einsum('jb,aijb->ia', l1, eris_voov) * 2
        l2new[:, :, p0:p1] += eris_voov.transpose(1, 2, 0, 3) * .5
        l2new[:, :, p0:p1] -= lib.einsum('bjic,ca->jiba', eris_voov, mba1)
        l2new[:, :, p0:p1] -= lib.einsum('bjka,ik->jiba', eris_voov, mij1)
        l1new[:, p0:p1] += numpy.einsum('aijb,jb->ia', eris_voov, mia1) * 2
        l1new -= numpy.einsum('bija,jb->ia', eris_voov, mia1[:, p0:p1])
        m4 = lib.einsum('ijkl,aklb->ijab', l2tau, eris_voov)
        l2new[:, :, p0:p1] += m4 * .5
        l1new[:, p0:p1] += numpy.einsum('ijab,jb->ia', m4, t1) * 2
        l1new -= numpy.einsum('ijba,jb->ia', m4, t1[:, p0:p1])

        saved_wvooo = _cp(imds.wvooo[p0:p1])
        l1new -= lib.einsum('ckij,jkca->ia', saved_wvooo, l2[:, :, p0:p1])
        saved_wvovv = _cp(imds.wvvov[p0:p1])
        # Watch out memory usage here, due to the l2 transpose
        l1new[:, p0:p1] += lib.einsum('abkc,kibc->ia', saved_wvovv, l2)
        saved_wvooo = saved_wvovv = None

        saved_wvOOv = _cp(imds.wvOOv[p0:p1])
        tmp_voov = _cp(imds.wVOov[p0:p1]) * 2
        tmp_voov += saved_wvOOv
        tmp = l2.transpose(0, 2, 1, 3) - l2.transpose(0, 3, 1, 2) * .5
        l2new[:, :, p0:p1] += lib.einsum('iakc,bjkc->jiba', tmp, tmp_voov)
        tmp = None

        tmp = lib.einsum('jkca,bikc->jiba', l2, saved_wvOOv)
        l2new[:, :, p0:p1] += tmp
        l2new[:, :, p0:p1] += tmp.transpose(1, 0, 2, 3) * .5
        saved_wvOOv = tmp = None

    saved_woooo = _cp(imds.woooo)
    m3 = lib.einsum('ijkl,klab->ijab', saved_woooo, l2)
    l2new += m3 * .5
    l1new += numpy.einsum('ijab,jb->ia', m3, t1) * 2
    l1new -= numpy.einsum('ijba,jb->ia', m3, t1)
    saved_woooo = m3 = None
    #time1 = log.timer_debug1('lambda pass [%d:%d]'%(p0, p1), *time1)

    eia = lib.direct_sum('i-a->ia', mo_e_o, mo_e_v)
    l1new /= eia

    #    l2new = l2new + l2new.transpose(1,0,3,2)
    #    l2new /= lib.direct_sum('ia+jb->ijab', eia, eia)
    #    l2new += l2
    for i in range(nocc):
        if i > 0:
            l2new[i, :i] += l2new[:i, i].transpose(0, 2, 1)
            l2new[i, :i] /= lib.direct_sum('a,jb->jab', eia[i], eia[:i])
            l2new[:i, i] = l2new[i, :i].transpose(0, 2, 1)
        l2new[i, i] = l2new[i, i] + l2new[i, i].T
        l2new[i, i] /= lib.direct_sum('a,b->ab', eia[i], eia[i])

    time0 = log.timer_debug1('update l1 l2', *time0)
    return l1new, l2new
Esempio n. 20
0
def make_tau(t2, t1a, t1b, fac=1, out=None):
    return _ccsd.make_tau(t2, t1a, t1b, fac, out)
Esempio n. 21
0
def update_lambda(mycc, t1, t2, l1, l2, eris, imds):
    time0 = time.clock(), time.time()
    log = logger.Logger(mycc.stdout, mycc.verbose)
    nocc, nvir = t1.shape
    l1new = np.zeros_like(l1)

    foo = eris.fock[:nocc,:nocc]
    fov = eris.fock[:nocc,nocc:]
    fvv = eris.fock[nocc:,nocc:]

    tau = _ccsd.make_tau(t2, t1, t1)

    theta = t2*2 - t2.transpose(0,1,3,2)
    mvv = lib.einsum('klca,klcb->ba', l2, theta)
    moo = lib.einsum('kicd,kjcd->ij', l2, theta)
    mvv1 = lib.einsum('jc,jb->bc', l1, t1) + mvv
    moo1 = lib.einsum('ic,kc->ik', l1, t1) + moo

    # m3 = einsum('ijab,acbd->ijcd', l2, vvvv)
    #    = einsum('ijab,cadb->ijcd', l2.conj(), vvvv).conj()
    m3 = mycc._add_vvvv(None, l2.conj(), eris, with_ovvv=False).conj()
    m3 += lib.einsum('klab,ikjl->ijab', l2, imds.woooo)
    m3 *= .5

    ovov = np.asarray(eris.ovov)
    l2tau = np.einsum('ijcd,klcd->ijkl', l2, tau)
    m3 += np.einsum('kalb,ijkl->ijab', ovov, l2tau) * .5
    l2tau = None

    l2new = ovov.transpose(0,2,1,3) * .5
    l2new += lib.einsum('ijac,cb->ijab', l2, imds.v1)
    l2new -= lib.einsum('ikab,jk->ijab', l2, imds.v2)
    l2new -= lib.einsum('ca,icjb->ijab', mvv1, ovov)
    l2new -= lib.einsum('ik,kajb->ijab', moo1, ovov)

    ovov = ovov * 2 - ovov.transpose(0,3,2,1)
    l1new -= np.einsum('ik,ka->ia', moo, imds.v4)
    l1new -= np.einsum('ca,ic->ia', mvv, imds.v4)
    l2new += np.einsum('ia,jb->ijab', l1, imds.v4)

    tmp  = t1 + np.einsum('kc,kjcb->jb', l1, theta)
    tmp -= lib.einsum('bd,jd->jb', mvv1, t1)
    tmp -= lib.einsum('lj,lb->jb', moo, t1)
    l1new += np.einsum('jbia,jb->ia', ovov, tmp)
    ovov = tmp = None

    ovvv = np.asarray(eris.get_ovvv())
    l1new += np.einsum('iacb,bc->ia', ovvv, mvv1) * 2
    l1new -= np.einsum('ibca,bc->ia', ovvv, mvv1)
    l2new += lib.einsum('ic,jbca->jiba', l1, ovvv)
    l2t1 = np.einsum('ijcd,kd->ijck', l2, t1)
    m3 -= np.einsum('kbca,ijck->ijab', ovvv, l2t1)
    l2t1 = ovvv = None

    l2new += m3
    l1new += np.einsum('ijab,jb->ia', m3, t1) * 2
    l1new += np.einsum('jiba,jb->ia', m3, t1) * 2
    l1new -= np.einsum('ijba,jb->ia', m3, t1)
    l1new -= np.einsum('jiab,jb->ia', m3, t1)

    ovoo = np.asarray(eris.ovoo)
    l1new -= np.einsum('iajk,kj->ia', ovoo, moo1) * 2
    l1new += np.einsum('jaik,kj->ia', ovoo, moo1)
    l2new -= lib.einsum('ka,jbik->ijab', l1, ovoo)
    ovoo = None

    l2theta = l2*2 - l2.transpose(0,1,3,2)
    l2new += lib.einsum('ikac,jbck->ijab', l2theta, imds.wovvo) * .5
    tmp = lib.einsum('ikca,jbck->ijab', l2, imds.woVVo)
    l2new += tmp * .5
    l2new += tmp.transpose(1,0,2,3)
    l2theta = None

    l1new += fov
    l1new += lib.einsum('ib,ba->ia', l1, imds.v1)
    l1new -= lib.einsum('ja,ij->ia', l1, imds.v2)

    l1new += np.einsum('jb,iabj->ia', l1, eris.ovvo) * 2
    l1new -= np.einsum('jb,ijba->ia', l1, eris.oovv)

    l1new -= lib.einsum('ijbc,bacj->ia', l2, imds.wvvvo)
    l1new -= lib.einsum('kjca,ijck->ia', l2, imds.woovo)

    l1new += np.einsum('ijab,bj->ia', l2, imds.w3) * 2
    l1new -= np.einsum('ijba,bj->ia', l2, imds.w3)

    eia = lib.direct_sum('i-j->ij', foo.diagonal(), fvv.diagonal())
    l1new /= eia
    l1new += l1

    l2new = l2new + l2new.transpose(1,0,3,2)
    l2new /= lib.direct_sum('ia+jb->ijab', eia, eia)
    l2new += l2

    time0 = log.timer_debug1('update l1 l2', *time0)
    return l1new, l2new