Use ParallelFor instead of ParallelForRange

better readability and no performance difference (if using
        NETGEN_LAMBDA_INLINE)
This commit is contained in:
Matthias Hochsteger 2019-10-09 09:57:37 +02:00
parent 95df0ea73e
commit 6e5d806d92

View File

@ -7,11 +7,6 @@
namespace netgen
{
using ngcore::ParallelForRange;
class trionedge
{
public:
@ -191,15 +186,12 @@ namespace netgen
if(faceindex==0)
{
seia.SetSize(mesh.GetNSE());
ParallelForRange( Range(seia), [&] (auto myrange)
ParallelFor( Range(seia), [&] (auto i) NETGEN_LAMBDA_INLINE
{
for (auto i : myrange)
{
SurfaceElementIndex sei(i);
seia[i] = sei;
if (mesh[sei].GetNP() != 3)
mixed = true;
}
SurfaceElementIndex sei(i);
seia[i] = sei;
if (mesh[sei].GetNP() != 3)
mixed = true;
});
}
else
@ -225,103 +217,90 @@ namespace netgen
if(faceindex == 0)
{
ParallelForRange( Range(pangle), [&] (auto myrange)
ParallelFor( Range(pangle), [&] (auto i) NETGEN_LAMBDA_INLINE
{
for (auto i : myrange)
pangle[i] = 0.0;
pangle[i] = 0.0;
});
}
else
{
ParallelForRange( Range(seia), [&] (auto myrange)
ParallelFor( Range(seia), [&] (auto i) NETGEN_LAMBDA_INLINE
{
for (auto i : myrange)
{
const Element2d & sel = mesh[seia[i]];
for (int j = 0; j < 3; j++)
pangle[sel[j]] = 0.0;
}
const Element2d & sel = mesh[seia[i]];
for (int j = 0; j < 3; j++)
pangle[sel[j]] = 0.0;
});
}
ParallelForRange( Range(seia), [&] (auto myrange)
ParallelFor( Range(seia), [&] (auto i) NETGEN_LAMBDA_INLINE
{
for (auto i : myrange)
const Element2d & sel = mesh[seia[i]];
for (int j = 0; j < 3; j++)
{
const Element2d & sel = mesh[seia[i]];
for (int j = 0; j < 3; j++)
POINTTYPE typ = mesh[sel[j]].Type();
if (typ == FIXEDPOINT || typ == EDGEPOINT)
{
POINTTYPE typ = mesh[sel[j]].Type();
if (typ == FIXEDPOINT || typ == EDGEPOINT)
{
AtomicAdd(pangle[sel[j]],
Angle (mesh[sel[(j+1)%3]] - mesh[sel[j]],
mesh[sel[(j+2)%3]] - mesh[sel[j]]));
}
}
}
});
ParallelForRange( Range(seia), [&] (auto myrange)
{
for (auto i : myrange)
{
const Element2d & sel = mesh[seia[i]];
for (int j = 0; j < 3; j++)
{
PointIndex pi = sel[j];
if (mesh[pi].Type() == INNERPOINT || mesh[pi].Type() == SURFACEPOINT)
pdef[pi] = -6;
else
for (int j = 0; j < 8; j++)
if (pangle[pi] >= minangle[j])
pdef[pi] = -1-j;
AtomicAdd(pangle[sel[j]],
Angle (mesh[sel[(j+1)%3]] - mesh[sel[j]],
mesh[sel[(j+2)%3]] - mesh[sel[j]]));
}
}
});
ParallelForRange( Range(seia), [&pdef, &neighbors, &mesh, &seia, &elements_on_node] (auto myrange)
ParallelFor( Range(seia), [&] (auto i) NETGEN_LAMBDA_INLINE
{
for (auto i : myrange)
const Element2d & sel = mesh[seia[i]];
for (int j = 0; j < 3; j++)
{
auto sei = seia[i];
for (PointIndex pi : mesh[sei].template PNums<3>())
AsAtomic(pdef[pi])++;
for (int j = 0; j < 3; j++)
{
neighbors[sei].SetNr (j, -1);
neighbors[sei].SetOrientation (j, 0);
}
PointIndex pi = sel[j];
if (mesh[pi].Type() == INNERPOINT || mesh[pi].Type() == SURFACEPOINT)
pdef[pi] = -6;
else
for (int j = 0; j < 8; j++)
if (pangle[pi] >= minangle[j])
pdef[pi] = -1-j;
}
});
const auto sel = mesh[sei];
for (int j = 0; j < 3; j++)
{
PointIndex pi1 = sel.PNumMod(j+2);
PointIndex pi2 = sel.PNumMod(j+3);
ParallelFor( Range(seia), [&pdef, &neighbors, &mesh, &seia, &elements_on_node] (auto i) NETGEN_LAMBDA_INLINE
{
auto sei = seia[i];
for (PointIndex pi : mesh[sei].template PNums<3>())
AsAtomic(pdef[pi])++;
for (int j = 0; j < 3; j++)
{
neighbors[sei].SetNr (j, -1);
neighbors[sei].SetOrientation (j, 0);
}
for (auto sei_other : elements_on_node[pi1])
const auto sel = mesh[sei];
for (int j = 0; j < 3; j++)
{
PointIndex pi1 = sel.PNumMod(j+2);
PointIndex pi2 = sel.PNumMod(j+3);
for (auto sei_other : elements_on_node[pi1])
{
if(sei_other==sei) continue;
const auto & other = mesh[sei_other];
int pi1_other = -1;
int pi2_other = -1;
bool common_edge = false;
for (int k = 0; k < 3; k++)
{
if(sei_other==sei) continue;
const auto & other = mesh[sei_other];
int pi1_other = -1;
int pi2_other = -1;
bool common_edge = false;
for (int k = 0; k < 3; k++)
if(other[k] == pi1)
pi1_other = k;
if(other[k] == pi2)
{
if(other[k] == pi1)
pi1_other = k;
if(other[k] == pi2)
{
pi2_other = k;
common_edge = true;
}
pi2_other = k;
common_edge = true;
}
}
if(common_edge)
{
neighbors[sei].SetNr (j, sei_other);
neighbors[sei].SetOrientation (j, 3-pi1_other-pi2_other);
}
if(common_edge)
{
neighbors[sei].SetNr (j, sei_other);
neighbors[sei].SetOrientation (j, 3-pi1_other-pi2_other);
}
}
}
@ -343,25 +322,22 @@ namespace netgen
while (!done && t >= 2)
{
cnt = 0;
ParallelForRange( Range(seia), [&] (auto myrange)
ParallelFor( Range(seia), [&] (auto i) NETGEN_LAMBDA_INLINE
{
for (auto i : myrange)
{
SurfaceElementIndex t1 = seia[i];
SurfaceElementIndex t1 = seia[i];
if (mesh[t1].IsDeleted())
continue;
if (mesh[t1].IsDeleted())
return;
if (mesh[t1].GetIndex() != faceindex)
continue;
if (mesh[t1].GetIndex() != faceindex)
return;
if (multithread.terminate)
throw NgException ("Meshing stopped");
if (multithread.terminate)
throw NgException ("Meshing stopped");
for (int o1 = 0; o1 < 3; o1++)
if(EdgeSwapping(mesh, usemetric, neighbors, swapped, t1, o1, t, pdef, true))
improvement_candidates[cnt++]= std::make_pair(t1,o1);
}
for (int o1 = 0; o1 < 3; o1++)
if(EdgeSwapping(mesh, usemetric, neighbors, swapped, t1, o1, t, pdef, true))
improvement_candidates[cnt++]= std::make_pair(t1,o1);
});
auto elements_with_improvement = improvement_candidates.Range(cnt.load());