[Bf-blender-cvs] [5a02d0da7a8] master: Fluid: Parallelizations for Mantaflow functions (D12002)
Erik Abrahamsson
noreply at git.blender.org
Mon Sep 6 17:30:45 CEST 2021
Commit: 5a02d0da7a88e31422b61d6e4e69c9f464290241
Author: Erik Abrahamsson
Date: Mon Sep 6 17:26:55 2021 +0200
Branches: master
https://developer.blender.org/rB5a02d0da7a88e31422b61d6e4e69c9f464290241
Fluid: Parallelizations for Mantaflow functions (D12002)
This update includes part of a performance boost from D12002.
Contributed by @erik85
===================================================================
M extern/mantaflow/preprocessed/fastmarch.cpp
===================================================================
diff --git a/extern/mantaflow/preprocessed/fastmarch.cpp b/extern/mantaflow/preprocessed/fastmarch.cpp
index 956725e523c..31e43483b49 100644
--- a/extern/mantaflow/preprocessed/fastmarch.cpp
+++ b/extern/mantaflow/preprocessed/fastmarch.cpp
@@ -874,6 +874,136 @@ static const Vec3i nb[6] = {Vec3i(1, 0, 0),
Vec3i(0, 0, 1),
Vec3i(0, 0, -1)};
+struct knMarkSkipCells : public KernelBase {
+ knMarkSkipCells(Grid<Real> &phi, Grid<int> &tmp, bool inside)
+ : KernelBase(&phi, 1), phi(phi), tmp(tmp), inside(inside)
+ {
+ runMessage();
+ run();
+ }
+ inline void op(int i, int j, int k, Grid<Real> &phi, Grid<int> &tmp, bool inside) const
+ {
+ if (!inside && phi(i, j, k) < 0.) {
+ tmp(i, j, k) = 1;
+ }
+ if (inside && phi(i, j, k) > 0.) {
+ tmp(i, j, k) = 1;
+ }
+ }
+ inline Grid<Real> &getArg0()
+ {
+ return phi;
+ }
+ typedef Grid<Real> type0;
+ inline Grid<int> &getArg1()
+ {
+ return tmp;
+ }
+ typedef Grid<int> type1;
+ inline bool &getArg2()
+ {
+ return inside;
+ }
+ typedef bool type2;
+ void runMessage()
+ {
+ debMsg("Executing kernel knMarkSkipCells ", 3);
+ debMsg("Kernel range"
+ << " x " << maxX << " y " << maxY << " z " << minZ << " - " << maxZ << " ",
+ 4);
+ };
+ void operator()(const tbb::blocked_range<IndexInt> &__r) const
+ {
+ const int _maxX = maxX;
+ const int _maxY = maxY;
+ if (maxZ > 1) {
+ for (int k = __r.begin(); k != (int)__r.end(); k++)
+ for (int j = 1; j < _maxY; j++)
+ for (int i = 1; i < _maxX; i++)
+ op(i, j, k, phi, tmp, inside);
+ }
+ else {
+ const int k = 0;
+ for (int j = __r.begin(); j != (int)__r.end(); j++)
+ for (int i = 1; i < _maxX; i++)
+ op(i, j, k, phi, tmp, inside);
+ }
+ }
+ void run()
+ {
+ if (maxZ > 1)
+ tbb::parallel_for(tbb::blocked_range<IndexInt>(minZ, maxZ), *this);
+ else
+ tbb::parallel_for(tbb::blocked_range<IndexInt>(1, maxY), *this);
+ }
+ Grid<Real> φ
+ Grid<int> &tmp;
+ bool inside;
+};
+
+struct knSetFirstLayer : public KernelBase {
+ knSetFirstLayer(Grid<int> &tmp, int dim) : KernelBase(&tmp, 1), tmp(tmp), dim(dim)
+ {
+ runMessage();
+ run();
+ }
+ inline void op(int i, int j, int k, Grid<int> &tmp, int dim) const
+ {
+ Vec3i p(i, j, k);
+ if (tmp(p))
+ return;
+ for (int n = 0; n < 2 * dim; ++n) {
+ if (tmp(p + nb[n]) == 1) {
+ tmp(i, j, k) = 2;
+ break;
+ }
+ }
+ }
+ inline Grid<int> &getArg0()
+ {
+ return tmp;
+ }
+ typedef Grid<int> type0;
+ inline int &getArg1()
+ {
+ return dim;
+ }
+ typedef int type1;
+ void runMessage()
+ {
+ debMsg("Executing kernel knSetFirstLayer ", 3);
+ debMsg("Kernel range"
+ << " x " << maxX << " y " << maxY << " z " << minZ << " - " << maxZ << " ",
+ 4);
+ };
+ void operator()(const tbb::blocked_range<IndexInt> &__r) const
+ {
+ const int _maxX = maxX;
+ const int _maxY = maxY;
+ if (maxZ > 1) {
+ for (int k = __r.begin(); k != (int)__r.end(); k++)
+ for (int j = 1; j < _maxY; j++)
+ for (int i = 1; i < _maxX; i++)
+ op(i, j, k, tmp, dim);
+ }
+ else {
+ const int k = 0;
+ for (int j = __r.begin(); j != (int)__r.end(); j++)
+ for (int i = 1; i < _maxX; i++)
+ op(i, j, k, tmp, dim);
+ }
+ }
+ void run()
+ {
+ if (maxZ > 1)
+ tbb::parallel_for(tbb::blocked_range<IndexInt>(minZ, maxZ), *this);
+ else
+ tbb::parallel_for(tbb::blocked_range<IndexInt>(1, maxY), *this);
+ }
+ Grid<int> &tmp;
+ int dim;
+};
+
template<class S> struct knExtrapolateLsSimple : public KernelBase {
knExtrapolateLsSimple(Grid<S> &val, int distance, Grid<int> &tmp, const int d, S direction)
: KernelBase(&val, 1), val(val), distance(distance), tmp(tmp), d(d), direction(direction)
@@ -1043,39 +1173,12 @@ void extrapolateLsSimple(Grid<Real> &phi, int distance = 4, bool inside = false)
tmp.clear();
const int dim = (phi.is3D() ? 3 : 2);
- // by default, march outside
- Real direction = 1.;
- if (!inside) {
- // mark all inside
- FOR_IJK_BND(phi, 1)
- {
- if (phi(i, j, k) < 0.) {
- tmp(i, j, k) = 1;
- }
- }
- }
- else {
- direction = -1.;
- FOR_IJK_BND(phi, 1)
- {
- if (phi(i, j, k) > 0.) {
- tmp(i, j, k) = 1;
- }
- }
- }
+ // by default, march outside (ie mark all inside to be skipped)
+ Real direction = (inside) ? -1. : 1.;
+ knMarkSkipCells(phi, tmp, inside);
+
// + first layer around
- FOR_IJK_BND(phi, 1)
- {
- Vec3i p(i, j, k);
- if (tmp(p))
- continue;
- for (int n = 0; n < 2 * dim; ++n) {
- if (tmp(p + nb[n]) == 1) {
- tmp(i, j, k) = 2;
- n = 2 * dim;
- }
- }
- }
+ knSetFirstLayer(tmp, dim);
// extrapolate for distance
for (int d = 2; d < 1 + distance; ++d) {
@@ -1126,37 +1229,12 @@ void extrapolateVec3Simple(Grid<Vec3> &vel, Grid<Real> &phi, int distance = 4, b
tmp.clear();
const int dim = (vel.is3D() ? 3 : 2);
- // mark initial cells, by default, march outside
- if (!inside) {
- // mark all inside
- FOR_IJK_BND(phi, 1)
- {
- if (phi(i, j, k) < 0.) {
- tmp(i, j, k) = 1;
- }
- }
- }
- else {
- FOR_IJK_BND(phi, 1)
- {
- if (phi(i, j, k) > 0.) {
- tmp(i, j, k) = 1;
- }
- }
- }
+ // mark initial cells, by default, march outside (ie mark all inside to be skipped)
+ Real direction = (inside) ? -1. : 1.;
+ knMarkSkipCells(phi, tmp, inside);
+
// + first layer next to initial cells
- FOR_IJK_BND(vel, 1)
- {
- Vec3i p(i, j, k);
- if (tmp(p))
- continue;
- for (int n = 0; n < 2 * dim; ++n) {
- if (tmp(p + nb[n]) == 1) {
- tmp(i, j, k) = 2;
- n = 2 * dim;
- }
- }
- }
+ knSetFirstLayer(tmp, dim);
for (int d = 2; d < 1 + distance; ++d) {
knExtrapolateLsSimple<Vec3>(vel, distance, tmp, d, Vec3(0.));
More information about the Bf-blender-cvs
mailing list