Parallel mode problem matrix creation.

Former-commit-id: a81323f6ba53fc5e790e489314bca98f6ae1ca0a
This commit is contained in:
Marek Nečada 2019-03-18 01:06:04 +00:00
parent 86b87c955c
commit 98da54946c
6 changed files with 191 additions and 4 deletions

View File

@ -1468,13 +1468,16 @@ cdef class ScatteringSystem:
qpms_scatsys_build_modeproblem_matrix_full(&target_view[0][0], self.s, k)
return target
def modeproblem_matrix_packed(self, double k, qpms_iri_t iri, version=None):
def modeproblem_matrix_packed(self, double k, qpms_iri_t iri, version='pR'):
cdef size_t rlen = self.saecv_sizes[iri]
cdef np.ndarray[np.complex_t, ndim=2] target = np.empty(
(rlen,rlen),dtype=complex, order='C')
cdef cdouble[:,::1] target_view = target
if (version == 'R'):
qpms_scatsys_build_modeproblem_matrix_irrep_packed_orbitorderR(&target_view[0][0], self.s, iri, k)
elif (version == 'pR'):
with nogil:
qpms_scatsys_build_modeproblem_matrix_irrep_packed_parallelR(&target_view[0][0], self.s, iri, k)
else:
qpms_scatsys_build_modeproblem_matrix_irrep_packed(&target_view[0][0], self.s, iri, k)
return target

View File

@ -293,6 +293,8 @@ cdef extern from "scatsystem.h":
const qpms_scatsys_t *ss, qpms_iri_t iri, double k)
cdouble *qpms_scatsys_build_modeproblem_matrix_irrep_packed_orbitorderR(
cdouble *target, const qpms_scatsys_t *ss, qpms_iri_t iri, double k)
cdouble *qpms_scatsys_build_modeproblem_matrix_irrep_packed_parallelR(
cdouble *target, const qpms_scatsys_t *ss, qpms_iri_t iri, double k) nogil

View File

@ -1561,3 +1561,177 @@ complex double *qpms_scatsys_build_modeproblem_matrix_irrep_packed_orbitorderR(
free(TSblock);
return target_packed;
}
struct qpms_scatsys_build_modeproblem_matrix_irrep_packed_parallelR_thread_arg{
const qpms_scatsys_t *ss;
qpms_ss_pi_t *opistartR_ptr;
pthread_mutex_t *opistartR_mutex;
qpms_iri_t iri;
complex double *target_packed;
double k;
};
static void *qpms_scatsys_build_modeproblem_matrix_irrep_packed_parallelR_thread(void *arg)
{
const struct qpms_scatsys_build_modeproblem_matrix_irrep_packed_parallelR_thread_arg
*a = arg;
const qpms_scatsys_t *ss = a->ss;
const qpms_iri_t iri = a->iri;
const size_t packedlen = ss->saecv_sizes[iri];
// some of the following workspaces are probably redundant; TODO optimize later.
// workspaces for the uncompressed particle<-particle tranlation matrix block
// and the result of multiplying with a T-matrix
complex double *Sblock, *TSblock;
QPMS_CRASHING_MALLOC(Sblock, sizeof(complex double)*SQ(ss->max_bspecn));
QPMS_CRASHING_MALLOC(TSblock, sizeof(complex double)*SQ(ss->max_bspecn));
// Workspace for the intermediate particle-orbit matrix result
complex double *tmp;
QPMS_CRASHING_MALLOC(tmp, sizeof(complex double) * SQ(ss->max_bspecn) * ss->sym->order);
const complex double one = 1, zero = 0;
while(1) {
// In the beginning, pick a target (row) orbit for this thread
QPMS_ENSURE_SUCCESS(pthread_mutex_lock(a->opistartR_mutex));
if(*(a->opistartR_ptr) >= ss->p_count) {// Everything is already done, end
QPMS_ENSURE_SUCCESS(pthread_mutex_unlock(a->opistartR_mutex));
break;
}
const qpms_ss_pi_t opistartR = *(a->opistartR_ptr);
// Now increment it for another thread:
*(a->opistartR_ptr) += ss->orbit_types[ss->p_orbitinfo[ss->p_by_orbit[opistartR]].t].size;
QPMS_ENSURE_SUCCESS(pthread_mutex_unlock(a->opistartR_mutex));
// Orbit picked (defined by opistartR), do the work.
const qpms_ss_pi_t orbitstartpiR = ss->p_by_orbit[opistartR];
const qpms_ss_oti_t otiR = ss->p_orbitinfo[orbitstartpiR].t;
const qpms_ss_osn_t osnR = ss->p_orbitinfo[orbitstartpiR].osn;
const qpms_ss_orbit_type_t *const otR = ss->orbit_types + otiR;
const qpms_ss_orbit_pi_t orbit_p_countR = otR->size;
const size_t orbit_packedsizeR = otR->irbase_sizes[iri];
if(orbit_packedsizeR) { // avoid zgemm crash on empty irrep
const size_t particle_fullsizeR = otR->bspecn; // == bspecR->n
const qpms_vswf_set_spec_t *bspecR = ss->tm[ss->p[orbitstartpiR].tmatrix_id]->spec;
// This is the orbit-level matrix projecting the whole orbit onto the irrep.
const complex double *omR = otR->irbases + otR->irbase_offsets[iri];
// Orbit coeff vector's full size:
const size_t orbit_fullsizeR = otR->size * otR->bspecn;
// This is where the orbit starts in the "packed" vector:
const size_t packed_orbit_offsetR =
ss->saecv_ot_offsets[iri*ss->orbit_type_count + otiR]
+ osnR * otR->irbase_sizes[iri];
for(qpms_ss_orbit_pi_t opiR = 0; opiR < orbit_p_countR; ++opiR) {
qpms_ss_pi_t piR = ss->p_by_orbit[opistartR + opiR];
assert(opiR == ss->p_orbitinfo[piR].p);
assert(otiR == ss->p_orbitinfo[piR].t);
assert(ss->p_orbitinfo[piR].osn == osnR);
const cart3_t posR = ss->p[piR].pos;
// dest particle T-matrix
const complex double *tmmR = ss->tm[ss->p[piR].tmatrix_id]->m;
for(qpms_ss_pi_t piC = 0; piC < ss->p_count; ++piC) { //Column loop
const qpms_ss_oti_t otiC = ss->p_orbitinfo[piC].t;
const qpms_ss_orbit_type_t *const otC = ss->orbit_types + otiC;
const qpms_ss_osn_t osnC = ss->p_orbitinfo[piC].osn;
const qpms_ss_orbit_pi_t opiC = ss->p_orbitinfo[piC].p;
// This is where the particle's orbit starts in the "packed" vector:
const size_t packed_orbit_offsetC =
ss->saecv_ot_offsets[iri*ss->orbit_type_count + otiC]
+ osnC * otC->irbase_sizes[iri];
const qpms_vswf_set_spec_t *bspecC = ss->tm[ss->p[piC].tmatrix_id]->spec;
// Orbit coeff vector's full size:
const size_t orbit_fullsizeC = otC->size * otC->bspecn;
const size_t particle_fullsizeC = otC->bspecn; // == bspecC->n
const size_t orbit_packedsizeC = otC->irbase_sizes[iri];
// This is the orbit-level matrix projecting the whole orbit onto the irrep.
const complex double *omC = otC->irbases + otC->irbase_offsets[iri];
if(orbit_packedsizeC) { // avoid zgemm crash on empty irrep
if(piC != piR) { // non-diagonal, calculate TS
const cart3_t posC = ss->p[piC].pos;
QPMS_ENSURE_SUCCESS(qpms_trans_calculator_get_trans_array_lc3p(ss->c,
Sblock, // Sblock is S(piR->piC)
bspecR, bspecC->n, bspecC, 1,
a->k, posR, posC));
cblas_zgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans,
bspecR->n /*m*/, bspecC->n /*n*/, bspecR->n /*k*/,
&one/*alpha*/, tmmR/*a*/, bspecR->n/*lda*/,
Sblock/*b*/, bspecC->n/*ldb*/, &zero/*beta*/,
TSblock /*c*/, bspecC->n /*ldc*/);
} else { // diagonal, fill with diagonal -1
for (size_t row = 0; row < bspecR->n; ++row)
for (size_t col = 0; col < bspecC->n; ++col)
TSblock[row * bspecC->n + col] = (row == col)? -1 : 0;
}
// tmp[oiR|piR,piC] = ∑_K M[piR,K] U*[K,piC]
cblas_zgemm(CblasRowMajor, CblasNoTrans, CblasConjTrans,
particle_fullsizeR /*M*/, orbit_packedsizeC /*N*/, particle_fullsizeC /*K*/,
&one /*alpha*/, TSblock/*A*/, particle_fullsizeC/*ldA*/,
omC + opiC*particle_fullsizeC /*B*/,
orbit_fullsizeC/*ldB*/, &zero /*beta*/,
tmp /*C*/, orbit_packedsizeC /*LDC*/);
// target[oiR|piR,oiC|piC] += U[...] tmp[...]
cblas_zgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans,
orbit_packedsizeR /*M*/, orbit_packedsizeC /*N*/, particle_fullsizeR /*K*/,
&one /*alpha*/, omR + opiR*particle_fullsizeR/*A*/, orbit_fullsizeR/*ldA*/,
tmp /*B*/, orbit_packedsizeC /*ldB*/, &one /*beta*/,
a->target_packed + packedlen*packed_orbit_offsetR + packed_orbit_offsetC /*C*/,
packedlen /*ldC*/);
}
}
}
}
}
free(tmp);
free(Sblock);
free(TSblock);
return NULL;
}
complex double *qpms_scatsys_build_modeproblem_matrix_irrep_packed_parallelR(
/// Target memory with capacity for ss->saecv_sizes[iri]**2 elements. If NULL, new will be allocated.
complex double *target_packed,
const qpms_scatsys_t *ss, qpms_iri_t iri,
double k ///< Wave number to use in the translation matrix.
)
{
const size_t packedlen = ss->saecv_sizes[iri];
if (!packedlen) // THIS IS A BIT PROBLEMATIC, TODO how to deal with empty irreps?
return target_packed;
if (target_packed == NULL)
target_packed = malloc(SQ(packedlen)*sizeof(complex double));
if (target_packed == NULL) abort();
memset(target_packed, 0, SQ(packedlen)*sizeof(complex double));
qpms_ss_pi_t opistartR = 0;
pthread_mutex_t opistartR_mutex;
QPMS_ENSURE_SUCCESS(pthread_mutex_init(&opistartR_mutex, NULL));
const struct qpms_scatsys_build_modeproblem_matrix_irrep_packed_parallelR_thread_arg
arg = {ss, &opistartR, &opistartR_mutex, iri, target_packed, k};
// FIXME THIS IS NOT PORTABLE:
long nthreads = sysconf(_SC_NPROCESSORS_ONLN);
if (nthreads < 1) nthreads = 1; // If something goes wrong...
pthread_t thread_ids[nthreads];
for(long thi = 0; thi < nthreads; ++thi)
QPMS_ENSURE_SUCCESS(pthread_create(thread_ids + thi, NULL,
qpms_scatsys_build_modeproblem_matrix_irrep_packed_parallelR_thread,
&arg));
for(long thi = 0; thi < nthreads; ++thi) {
void *retval;
QPMS_ENSURE_SUCCESS(pthread_join(thread_ids[thi], &retval));
}
QPMS_ENSURE_SUCCESS(pthread_mutex_destroy(&opistartR_mutex));
return target_packed;
}

View File

@ -500,6 +500,13 @@ complex double *qpms_scatsys_build_modeproblem_matrix_irrep_packed_orbitorderR(
const qpms_scatsys_t *ss, qpms_iri_t iri,
double k ///< Wave number to use in the translation matrix.
);
/// Alternative implementation of qpms_scatsys_build_modeproblem_matrix_irrep_packed().
complex double *qpms_scatsys_build_modeproblem_matrix_irrep_packed_orbitorder_parallelR(
/// Target memory with capacity for ss->fecv_size**2 elements. If NULL, new will be allocated.
complex double *target,
const qpms_scatsys_t *ss, qpms_iri_t iri,
double k ///< Wave number to use in the translation matrix.
);
/// NOT IMPLEMENTED Dumps a qpms_scatsys_t structure to a file.
qpms_errno_t qpms_scatsys_dump(qpms_scatsys_t *ss, char *path);

View File

@ -35,7 +35,7 @@ qpms_c = Extension('qpms_c',
'-DDISABLE_NDEBUG', # uncomment to enable assertions in the modules
#'-fopenmp',
],
libraries=['gsl', 'lapacke', 'blas', 'gslcblas', #'omp'
libraries=['gsl', 'lapacke', 'blas', 'gslcblas', 'pthread', #'omp'
# TODO resolve the problem with openblas (missing gotoblas symbol) and preferable use other blas library
],
runtime_library_dirs=os.environ['LD_LIBRARY_PATH'].split(':') if 'LD_LIBRARY_PATH' in os.environ else []

View File

@ -38,7 +38,8 @@ ss = ScatteringSystem(particles, sym)
k = n * omega / c
for iri in range(ss.nirreps):
mm_iri_orig = ss.modeproblem_matrix_packed(k, iri)
mm_iri_orig = ss.modeproblem_matrix_packed(k, iri, version = None)
mm_iri_alt = ss.modeproblem_matrix_packed(k, iri, version='R')
print(np.amax(abs(mm_iri_orig-mm_iri_alt)))
mm_iri_paral = ss.modeproblem_matrix_packed(k, iri, version='pR')
print(np.amax(abs(mm_iri_orig-mm_iri_alt)), np.amax(abs(mm_iri_orig-mm_iri_paral)))