From 38a4dbfcd7d3e9539e334cee02fecab2c7901793 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Marek=20Ne=C4=8Dada?= <marek@necada.org>
Date: Fri, 3 Jul 2020 15:10:15 +0300
Subject: [PATCH] Set compiler/linker options to make cython parallelisation
 actually work.

---
 qpms/qpms_c.pyx | 31 ++++++++++++++++---------------
 setup.py        |  2 ++
 2 files changed, 18 insertions(+), 15 deletions(-)

diff --git a/qpms/qpms_c.pyx b/qpms/qpms_c.pyx
index fbeafb3..ab76c97 100644
--- a/qpms/qpms_c.pyx
+++ b/qpms/qpms_c.pyx
@@ -17,8 +17,8 @@ from .cymaterials cimport EpsMuGenerator, EpsMu
 from libc.stdlib cimport malloc, free, calloc
 import warnings
 
-from cython.parallel import prange, parallel
-from cython import boundscheck, wraparound
+from cython.parallel cimport prange, parallel
+from cython cimport boundscheck, wraparound
 
 # Set custom GSL error handler. N.B. this is obviously not thread-safe.
 cdef char *pgsl_err_reason
@@ -916,7 +916,7 @@ cdef class ScatteringSystem:
         cdef ccart3_t res
         cdef cart3_t pos
         cdef Py_ssize_t i
-        with nogil, parallel():
+        with nogil, parallel(), boundscheck(False), wraparound(False):
             for i in prange(evalpos_a.shape[0]):
                 pos.x = evalpos_a[i,0]
                 pos.y = evalpos_a[i,1]
@@ -957,7 +957,6 @@ def empty_lattice_modes_xy(EpsMu epsmu, reciprocal_basis, wavevector, double max
     free(omegas_c)
     return omegas
 
-
 cdef class _ScatteringSystemAtOmegaK:
     '''
     Wrapper over the C qpms_scatsys_at_omega_k_t structure
@@ -1121,6 +1120,7 @@ cdef class _ScatteringSystemAtOmega:
     def translation_matrix_packed(self, iri, J = QPMS_HANKEL_PLUS):
         return self.ss_pyref.translation_matrix_packed(wavenumber=self.wavenumber, iri=iri, J=J)
 
+    @boundscheck(False)
     def scattered_E(self, scatcoeffvector_full, evalpos, bint alt=False, btyp=QPMS_HANKEL_PLUS):
         cdef qpms_bessel_t btyp_c = BesselType(btyp)
         evalpos = np.array(evalpos, dtype=float, copy=False)
@@ -1133,17 +1133,18 @@ cdef class _ScatteringSystemAtOmega:
         cdef ccart3_t res
         cdef cart3_t pos
         cdef Py_ssize_t i
-        for i in range(evalpos_a.shape[0]):
-            pos.x = evalpos_a[i,0]
-            pos.y = evalpos_a[i,1]
-            pos.z = evalpos_a[i,2]
-            if alt:
-                res = qpms_scatsysw_scattered_E__alt(self.ssw, btyp_c, &scv_view[0], pos)
-            else:
-                res = qpms_scatsysw_scattered_E(self.ssw, btyp_c, &scv_view[0], pos)
-            results[i,0] = res.x
-            results[i,1] = res.y
-            results[i,2] = res.z
+        with wraparound(False), nogil, parallel():
+            for i in prange(evalpos_a.shape[0]):
+                pos.x = evalpos_a[i,0]
+                pos.y = evalpos_a[i,1]
+                pos.z = evalpos_a[i,2]
+                if alt:
+                    res = qpms_scatsysw_scattered_E__alt(self.ssw, btyp_c, &scv_view[0], pos)
+                else:
+                    res = qpms_scatsysw_scattered_E(self.ssw, btyp_c, &scv_view[0], pos)
+                results[i,0] = res.x
+                results[i,1] = res.y
+                results[i,2] = res.z
         return results.reshape(evalpos.shape)
 
 
diff --git a/setup.py b/setup.py
index 13ee224..29fd31f 100755
--- a/setup.py
+++ b/setup.py
@@ -134,6 +134,8 @@ qpms_c = Extension('qpms.qpms_c',
             #('amos', dict(sources=amos_sources) ),
 	],
         include_dirs=['amos', 'qpms'],
+        extra_link_args=['-fopenmp'],
+        extra_compile_args=['-fopenmp'],
         #extra_link_args=[ 'qpms/libqpms.a','amos/libamos.a', ],
         #runtime_library_dirs=os.environ['LD_LIBRARY_PATH'].split(':') if 'LD_LIBRARY_PATH' in os.environ else [],
         #extra_objects = ['amos/libamos.a'], # FIXME apparently, I would like to eliminate the need to cmake/make first