Skip to content

GitLab

Commit fd473eea authored by Peter Eastman's avatar Peter Eastman
Browse files

Merge branch 'master' into nucleic

parents 0a751b5b 6a985cfd
Hide whitespace changes
Inline Side-by-side
Showing with 521 additions and 1961 deletions
platforms/cpu/tests/TestCpuEwald.cpp
View file @ fd473eea
......@@ -6,7 +6,7 @@
* Biological Structures at Stanford, funded under the NIH Roadmap for *
* Medical Research, grant U54 GM072970. See https://simtk.org. *
* *
* Portions copyright (c) 2008-2013 Stanford University and the Authors. *
* Portions copyright (c) 2015 Stanford University and the Authors. *
* Authors: Peter Eastman *
* Contributors: *
* *
......@@ -29,297 +29,8 @@
* USE OR OTHER DEALINGS IN THE SOFTWARE. *
* -------------------------------------------------------------------------- */
/**
* This tests the Ewald summation method CPU implementation of NonbondedForce.
*/
#include "CpuTests.h"
#include "TestEwald.h"
#include "openmm/internal/AssertionUtilities.h"
#include "openmm/Context.h"
#include "CpuPlatform.h"
#include "ReferencePlatform.h"
#include "openmm/NonbondedForce.h"
#include "openmm/System.h"
#include "openmm/LangevinIntegrator.h"
#include "openmm/VerletIntegrator.h"
#include "openmm/internal/ContextImpl.h"
#include "SimTKOpenMMRealType.h"
#include "sfmt/SFMT.h"
#include <iostream>
#include <vector>
using namespace OpenMM;
using namespace std;
CpuPlatform platform;
const double TOL = 1e-5;
void testEwaldPME(bool includeExceptions) {
// Use amorphous NaCl system for the tests
const int numParticles = 894;
const double cutoff = 1.2;
const double boxSize = 3.00646;
double tol = 1e-5;
ReferencePlatform reference;
System system;
NonbondedForce* nonbonded = new NonbondedForce();
nonbonded->setNonbondedMethod(NonbondedForce::Ewald);
nonbonded->setCutoffDistance(cutoff);
nonbonded->setEwaldErrorTolerance(tol);
for (int i = 0; i < numParticles/2; i++)
system.addParticle(22.99);
for (int i = 0; i < numParticles/2; i++)
system.addParticle(35.45);
for (int i = 0; i < numParticles/2; i++)
nonbonded->addParticle(1.0, 1.0,0.0);
for (int i = 0; i < numParticles/2; i++)
nonbonded->addParticle(-1.0, 1.0,0.0);
system.setDefaultPeriodicBoxVectors(Vec3(boxSize, 0, 0), Vec3(0, boxSize, 0), Vec3(0, 0, boxSize));
system.addForce(nonbonded);
vector<Vec3> positions(numParticles);
#include "nacl_amorph.dat"
if (includeExceptions) {
// Add some exclusions.
for (int i = 0; i < numParticles-1; i++) {
Vec3 delta = positions[i]-positions[i+1];
if (sqrt(delta.dot(delta)) < 0.5*cutoff)
nonbonded->addException(i, i+1, i%2 == 0 ? 0.0 : 0.5, 1.0, 0.0);
}
}
// (1) Check whether the Reference and CPU platforms agree when using Ewald Method
VerletIntegrator integrator1(0.01);
VerletIntegrator integrator2(0.01);
Context cpuContext(system, integrator1, platform);
Context referenceContext(system, integrator2, reference);
cpuContext.setPositions(positions);
referenceContext.setPositions(positions);
State cpuState = cpuContext.getState(State::Forces | State::Energy);
State referenceState = referenceContext.getState(State::Forces | State::Energy);
tol = 1e-2;
for (int i = 0; i < numParticles; i++) {
ASSERT_EQUAL_VEC(referenceState.getForces()[i], cpuState.getForces()[i], tol);
}
tol = 1e-5;
ASSERT_EQUAL_TOL(referenceState.getPotentialEnergy(), cpuState.getPotentialEnergy(), tol);
// (2) Check whether Ewald method in CPU is self-consistent
double norm = 0.0;
for (int i = 0; i < numParticles; ++i) {
Vec3 f = cpuState.getForces()[i];
norm += f[0]*f[0] + f[1]*f[1] + f[2]*f[2];
}
norm = std::sqrt(norm);
const double delta = 5e-3;
double step = delta/norm;
for (int i = 0; i < numParticles; ++i) {
Vec3 p = positions[i];
Vec3 f = cpuState.getForces()[i];
positions[i] = Vec3(p[0]-f[0]*step, p[1]-f[1]*step, p[2]-f[2]*step);
}
VerletIntegrator integrator3(0.01);
Context cpuContext2(system, integrator3, platform);
cpuContext2.setPositions(positions);
tol = 1e-2;
State cpuState2 = cpuContext2.getState(State::Energy);
ASSERT_EQUAL_TOL(norm, (cpuState2.getPotentialEnergy()-cpuState.getPotentialEnergy())/delta, tol)
// (3) Check whether the Reference and CPU platforms agree when using PME
nonbonded->setNonbondedMethod(NonbondedForce::PME);
cpuContext.reinitialize();
referenceContext.reinitialize();
cpuContext.setPositions(positions);
referenceContext.setPositions(positions);
cpuState = cpuContext.getState(State::Forces | State::Energy);
referenceState = referenceContext.getState(State::Forces | State::Energy);
tol = 1e-2;
for (int i = 0; i < numParticles; i++) {
ASSERT_EQUAL_VEC(referenceState.getForces()[i], cpuState.getForces()[i], tol);
}
tol = 1e-5;
ASSERT_EQUAL_TOL(referenceState.getPotentialEnergy(), cpuState.getPotentialEnergy(), tol);
// (4) Check whether PME method in CPU is self-consistent
norm = 0.0;
for (int i = 0; i < numParticles; ++i) {
Vec3 f = cpuState.getForces()[i];
norm += f[0]*f[0] + f[1]*f[1] + f[2]*f[2];
}
norm = std::sqrt(norm);
step = delta/norm;
for (int i = 0; i < numParticles; ++i) {
Vec3 p = positions[i];
Vec3 f = cpuState.getForces()[i];
positions[i] = Vec3(p[0]-f[0]*step, p[1]-f[1]*step, p[2]-f[2]*step);
}
VerletIntegrator integrator4(0.01);
Context cpuContext3(system, integrator4, platform);
cpuContext3.setPositions(positions);
tol = 1e-2;
State cpuState3 = cpuContext3.getState(State::Energy);
ASSERT_EQUAL_TOL(norm, (cpuState3.getPotentialEnergy()-cpuState.getPotentialEnergy())/delta, tol)
}
void testEwald2Ions() {
System system;
system.addParticle(1.0);
system.addParticle(1.0);
VerletIntegrator integrator(0.01);
NonbondedForce* nonbonded = new NonbondedForce();
nonbonded->addParticle(1.0, 1, 0);
nonbonded->addParticle(-1.0, 1, 0);
nonbonded->setNonbondedMethod(NonbondedForce::Ewald);
const double cutoff = 2.0;
nonbonded->setCutoffDistance(cutoff);
nonbonded->setEwaldErrorTolerance(TOL);
system.setDefaultPeriodicBoxVectors(Vec3(6, 0, 0), Vec3(0, 6, 0), Vec3(0, 0, 6));
system.addForce(nonbonded);
Context context(system, integrator, platform);
vector<Vec3> positions(2);
positions[0] = Vec3(3.048000,2.764000,3.156000);
positions[1] = Vec3(2.809000,2.888000,2.571000);
context.setPositions(positions);
State state = context.getState(State::Forces | State::Energy);
const vector<Vec3>& forces = state.getForces();
ASSERT_EQUAL_VEC(Vec3(-123.711, 64.1877, -302.716), forces[0], 10*TOL);
ASSERT_EQUAL_VEC(Vec3( 123.711, -64.1877, 302.716), forces[1], 10*TOL);
ASSERT_EQUAL_TOL(-217.276, state.getPotentialEnergy(), 0.01/*10*TOL*/);
}
void testTriclinic() {
// Create a triclinic box containing eight particles.
System system;
system.setDefaultPeriodicBoxVectors(Vec3(2.5, 0, 0), Vec3(0.5, 3.0, 0), Vec3(0.7, 0.9, 3.5));
for (int i = 0; i < 8; i++)
system.addParticle(1.0);
NonbondedForce* force = new NonbondedForce();
system.addForce(force);
force->setNonbondedMethod(NonbondedForce::PME);
force->setCutoffDistance(1.0);
force->setPMEParameters(3.45891, 32, 40, 48);
for (int i = 0; i < 4; i++)
force->addParticle(-1, 0.440104, 0.4184); // Cl parameters
for (int i = 0; i < 4; i++)
force->addParticle(1, 0.332840, 0.0115897); // Na parameters
vector<Vec3> positions(8);
positions[0] = Vec3(1.744, 2.788, 3.162);
positions[1] = Vec3(1.048, 0.762, 2.340);
positions[2] = Vec3(2.489, 1.570, 2.817);
positions[3] = Vec3(1.027, 1.893, 3.271);
positions[4] = Vec3(0.937, 0.825, 0.009);
positions[5] = Vec3(2.290, 1.887, 3.352);
positions[6] = Vec3(1.266, 1.111, 2.894);
positions[7] = Vec3(0.933, 1.862, 3.490);
// Compute the forces and energy.
VerletIntegrator integ(0.001);
Context context(system, integ, platform);
context.setPositions(positions);
State state = context.getState(State::Forces | State::Energy);
// Compare them to values computed by Gromacs.
double expectedEnergy = -963.370;
vector<Vec3> expectedForce(8);
expectedForce[0] = Vec3(4.25253e+01, -1.23503e+02, 1.22139e+02);
expectedForce[1] = Vec3(9.74752e+01, 1.68213e+02, 1.93169e+02);
expectedForce[2] = Vec3(-1.50348e+02, 1.29165e+02, 3.70435e+02);
expectedForce[3] = Vec3(9.18644e+02, -3.52571e+00, -1.34772e+03);
expectedForce[4] = Vec3(-1.61193e+02, 9.01528e+01, -7.12904e+01);
expectedForce[5] = Vec3(2.82630e+02, 2.78029e+01, -3.72864e+02);
expectedForce[6] = Vec3(-1.47454e+02, -2.14448e+02, -3.55789e+02);
expectedForce[7] = Vec3(-8.82195e+02, -7.39132e+01, 1.46202e+03);
for (int i = 0; i < 8; i++) {
ASSERT_EQUAL_VEC(expectedForce[i], state.getForces()[i], 1e-4);
}
ASSERT_EQUAL_TOL(expectedEnergy, state.getPotentialEnergy(), 1e-4);
}
void testErrorTolerance(NonbondedForce::NonbondedMethod method) {
// Create a cloud of random point charges.
const int numParticles = 51;
const double boxWidth = 5.0;
System system;
system.setDefaultPeriodicBoxVectors(Vec3(boxWidth, 0, 0), Vec3(0, boxWidth, 0), Vec3(0, 0, boxWidth));
NonbondedForce* force = new NonbondedForce();
system.addForce(force);
vector<Vec3> positions(numParticles);
OpenMM_SFMT::SFMT sfmt;
init_gen_rand(0, sfmt);
for (int i = 0; i < numParticles; i++) {
system.addParticle(1.0);
force->addParticle(-1.0+i*2.0/(numParticles-1), 1.0, 0.0);
positions[i] = Vec3(boxWidth*genrand_real2(sfmt), boxWidth*genrand_real2(sfmt), boxWidth*genrand_real2(sfmt));
}
force->setNonbondedMethod(method);
// For various values of the cutoff and error tolerance, see if the actual error is reasonable.
for (double cutoff = 1.0; cutoff < boxWidth/2; cutoff *= 1.2) {
force->setCutoffDistance(cutoff);
vector<Vec3> refForces;
double norm = 0.0;
for (double tol = 5e-5; tol < 1e-3; tol *= 2.0) {
force->setEwaldErrorTolerance(tol);
VerletIntegrator integrator(0.01);
Context context(system, integrator, platform);
context.setPositions(positions);
State state = context.getState(State::Forces);
if (refForces.size() == 0) {
refForces = state.getForces();
for (int i = 0; i < numParticles; i++)
norm += refForces[i].dot(refForces[i]);
norm = sqrt(norm);
}
else {
double diff = 0.0;
for (int i = 0; i < numParticles; i++) {
Vec3 delta = refForces[i]-state.getForces()[i];
diff += delta.dot(delta);
}
diff = sqrt(diff)/norm;
ASSERT(diff < 2*tol);
}
}
}
}
int main(int argc, char* argv[]) {
try {
if (!CpuPlatform::isProcessorSupported()) {
cout << "CPU is not supported. Exiting." << endl;
return 0;
}
testEwaldPME(false);
testEwaldPME(true);
// testEwald2Ions();
testTriclinic();
testErrorTolerance(NonbondedForce::Ewald);
testErrorTolerance(NonbondedForce::PME);
}
catch(const exception& e) {
cout << "exception: " << e.what() << endl;
return 1;
}
cout << "Done" << endl;
return 0;
void runPlatformTests() {
}
platforms/cpu/tests/TestCpuGBSAOBCForce.cpp
View file @ fd473eea
......@@ -6,7 +6,7 @@
* Biological Structures at Stanford, funded under the NIH Roadmap for *
* Medical Research, grant U54 GM072970. See https://simtk.org. *
* *
* Portions copyright (c) 2008-2014 Stanford University and the Authors. *
* Portions copyright (c) 2015 Stanford University and the Authors. *
* Authors: Peter Eastman *
* Contributors: *
* *
......@@ -29,244 +29,8 @@
* USE OR OTHER DEALINGS IN THE SOFTWARE. *
* -------------------------------------------------------------------------- */
/**
* This tests the CPU implementation of GBSAOBCForce.
*/
#include "CpuTests.h"
#include "TestGBSAOBCForce.h"
#include "openmm/internal/AssertionUtilities.h"
#include "openmm/Context.h"
#include "CpuPlatform.h"
#include "openmm/GBSAOBCForce.h"
#include "openmm/System.h"
#include "openmm/LangevinIntegrator.h"
#include "openmm/NonbondedForce.h"
#include "SimTKOpenMMRealType.h"
#include "sfmt/SFMT.h"
#include <iostream>
#include <vector>
using namespace OpenMM;
using namespace std;
const double TOL = 1e-5;
void testSingleParticle() {
CpuPlatform platform;
System system;
system.addParticle(2.0);
LangevinIntegrator integrator(0, 0.1, 0.01);
GBSAOBCForce* forceField = new GBSAOBCForce();
forceField->addParticle(0.5, 0.15, 1);
system.addForce(forceField);
Context context(system, integrator, platform);
vector<Vec3> positions(1);
positions[0] = Vec3(0, 0, 0);
context.setPositions(positions);
State state = context.getState(State::Energy);
double bornRadius = 0.15-0.009; // dielectric offset
double eps0 = EPSILON0;
double bornEnergy = (-0.5*0.5/(8*PI_M*eps0))*(1.0/forceField->getSoluteDielectric()-1.0/forceField->getSolventDielectric())/bornRadius;
double extendedRadius = 0.15+0.14; // probe radius
double nonpolarEnergy = 4*PI_M*2.25936*extendedRadius*extendedRadius*std::pow(0.15/bornRadius, 6.0);
ASSERT_EQUAL_TOL((bornEnergy+nonpolarEnergy), state.getPotentialEnergy(), 0.01);
// Change the parameters and see if it is still correct.
forceField->setParticleParameters(0, 0.4, 0.25, 1);
forceField->updateParametersInContext(context);
state = context.getState(State::Energy);
bornRadius = 0.25-0.009; // dielectric offset
bornEnergy = (-0.4*0.4/(8*PI_M*eps0))*(1.0/forceField->getSoluteDielectric()-1.0/forceField->getSolventDielectric())/bornRadius;
extendedRadius = 0.25+0.14;
nonpolarEnergy = 4*PI_M*2.25936*extendedRadius*extendedRadius*std::pow(0.25/bornRadius, 6.0);
ASSERT_EQUAL_TOL((bornEnergy+nonpolarEnergy), state.getPotentialEnergy(), 0.01);
}
void testGlobalSettings() {
CpuPlatform platform;
System system;
system.addParticle(2.0);
LangevinIntegrator integrator(0, 0.1, 0.01);
GBSAOBCForce* forceField = new GBSAOBCForce();
forceField->addParticle(0.5, 0.15, 1);
const double soluteDielectric = 2.1;
const double solventDielectric = 35.0;
const double surfaceAreaEnergy = 0.75;
forceField->setSoluteDielectric(soluteDielectric);
forceField->setSolventDielectric(solventDielectric);
forceField->setSurfaceAreaEnergy(surfaceAreaEnergy);
system.addForce(forceField);
Context context(system, integrator, platform);
vector<Vec3> positions(1);
positions[0] = Vec3(0, 0, 0);
context.setPositions(positions);
State state = context.getState(State::Energy);
double bornRadius = 0.15-0.009; // dielectric offset
double eps0 = EPSILON0;
double bornEnergy = (-0.5*0.5/(8*PI_M*eps0))*(1.0/soluteDielectric-1.0/solventDielectric)/bornRadius;
double extendedRadius = 0.15+0.14; // probe radius
double nonpolarEnergy = 4*PI_M*surfaceAreaEnergy*extendedRadius*extendedRadius*std::pow(0.15/bornRadius, 6.0);
ASSERT_EQUAL_TOL((bornEnergy+nonpolarEnergy), state.getPotentialEnergy(), 0.01);
}
void testCutoffAndPeriodic() {
CpuPlatform platform;
System system;
system.addParticle(1.0);
system.addParticle(1.0);
LangevinIntegrator integrator(0, 0.1, 0.01);
GBSAOBCForce* gbsa = new GBSAOBCForce();
NonbondedForce* nonbonded = new NonbondedForce();
gbsa->addParticle(-1, 0.15, 1);
nonbonded->addParticle(-1, 1, 0);
gbsa->addParticle(1, 0.15, 1);
nonbonded->addParticle(1, 1, 0);
const double cutoffDistance = 3.0;
const double boxSize = 10.0;
nonbonded->setCutoffDistance(cutoffDistance);
gbsa->setCutoffDistance(cutoffDistance);
system.setDefaultPeriodicBoxVectors(Vec3(boxSize, 0, 0), Vec3(0, boxSize, 0), Vec3(0, 0, boxSize));
system.addForce(gbsa);
system.addForce(nonbonded);
vector<Vec3> positions(2);
positions[0] = Vec3(0, 0, 0);
positions[1] = Vec3(2, 0, 0);
// Calculate the forces for both cutoff and periodic with two different atom positions.
nonbonded->setNonbondedMethod(NonbondedForce::CutoffNonPeriodic);
gbsa->setNonbondedMethod(GBSAOBCForce::CutoffNonPeriodic);
Context context(system, integrator, platform);
context.setPositions(positions);
State state1 = context.getState(State::Forces);
nonbonded->setNonbondedMethod(NonbondedForce::CutoffPeriodic);
gbsa->setNonbondedMethod(GBSAOBCForce::CutoffPeriodic);
context.reinitialize();
context.setPositions(positions);
State state2 = context.getState(State::Forces);
positions[1][0]+= boxSize;
nonbonded->setNonbondedMethod(NonbondedForce::CutoffNonPeriodic);
gbsa->setNonbondedMethod(GBSAOBCForce::CutoffNonPeriodic);
context.reinitialize();
context.setPositions(positions);
State state3 = context.getState(State::Forces);
nonbonded->setNonbondedMethod(NonbondedForce::CutoffPeriodic);
gbsa->setNonbondedMethod(GBSAOBCForce::CutoffPeriodic);
context.reinitialize();
context.setPositions(positions);
State state4 = context.getState(State::Forces);
// All forces should be identical, exception state3 which should be zero.
ASSERT_EQUAL_VEC(state1.getForces()[0], state2.getForces()[0], 0.01);
ASSERT_EQUAL_VEC(state1.getForces()[1], state2.getForces()[1], 0.01);
ASSERT_EQUAL_VEC(state1.getForces()[0], state4.getForces()[0], 0.01);
ASSERT_EQUAL_VEC(state1.getForces()[1], state4.getForces()[1], 0.01);
ASSERT_EQUAL_VEC(state3.getForces()[0], Vec3(0, 0, 0), 0.01);
ASSERT_EQUAL_VEC(state3.getForces()[1], Vec3(0, 0, 0), 0.01);
}
void testForce(int numParticles, NonbondedForce::NonbondedMethod method, GBSAOBCForce::NonbondedMethod method2) {
CpuPlatform platform;
ReferencePlatform reference;
System system;
GBSAOBCForce* gbsa = new GBSAOBCForce();
NonbondedForce* nonbonded = new NonbondedForce();
for (int i = 0; i < numParticles; ++i) {
system.addParticle(1.0);
double charge = i%2 == 0 ? -1 : 1;
gbsa->addParticle(charge, 0.15, 1);
nonbonded->addParticle(charge, 1, 0);
}
nonbonded->setNonbondedMethod(method);
gbsa->setNonbondedMethod(method2);
nonbonded->setCutoffDistance(3.0);
gbsa->setCutoffDistance(3.0);
int grid = (int) floor(0.5+pow(numParticles, 1.0/3.0));
if (method == NonbondedForce::CutoffPeriodic) {
double boxSize = (grid+1)*1.1;
system.setDefaultPeriodicBoxVectors(Vec3(boxSize, 0, 0), Vec3(0, boxSize, 0), Vec3(0, 0, boxSize));
}
system.addForce(gbsa);
system.addForce(nonbonded);
LangevinIntegrator integrator1(0, 0.1, 0.01);
LangevinIntegrator integrator2(0, 0.1, 0.01);
Context context(system, integrator1, platform);
Context refContext(system, integrator2, reference);
// Set random (but uniformly distributed) positions for all the particles.
vector<Vec3> positions(numParticles);
OpenMM_SFMT::SFMT sfmt;
init_gen_rand(0, sfmt);
for (int i = 0; i < grid; i++)
for (int j = 0; j < grid; j++)
for (int k = 0; k < grid; k++)
positions[i*grid*grid+j*grid+k] = Vec3(i*1.1, j*1.1, k*1.1);
for (int i = 0; i < numParticles; ++i)
positions[i] = positions[i] + Vec3(0.5*genrand_real2(sfmt), 0.5*genrand_real2(sfmt), 0.5*genrand_real2(sfmt));
context.setPositions(positions);
refContext.setPositions(positions);
State state = context.getState(State::Forces | State::Energy);
State refState = refContext.getState(State::Forces | State::Energy);
// Make sure the CPU and Reference platforms agree.
double norm = 0.0;
double diff = 0.0;
for (int i = 0; i < numParticles; ++i) {
Vec3 f = state.getForces()[i];
norm += f[0]*f[0] + f[1]*f[1] + f[2]*f[2];
Vec3 delta = f-refState.getForces()[i];
diff += delta[0]*delta[0] + delta[1]*delta[1] + delta[2]*delta[2];
}
norm = std::sqrt(norm);
diff = std::sqrt(diff);
ASSERT_EQUAL_TOL(0.0, diff, 0.001*norm);
ASSERT_EQUAL_TOL(state.getPotentialEnergy(), refState.getPotentialEnergy(), 1e-3);
// Take a small step in the direction of the energy gradient and see whether the potential energy changes by the expected amount.
// (This doesn't work with cutoffs, since the energy changes discontinuously at the cutoff distance.)
if (method == NonbondedForce::NoCutoff)
{
const double delta = 0.3;
double step = 0.5*delta/norm;
vector<Vec3> positions2(numParticles), positions3(numParticles);
for (int i = 0; i < numParticles; ++i) {
Vec3 p = positions[i];
Vec3 f = state.getForces()[i];
positions2[i] = Vec3(p[0]-f[0]*step, p[1]-f[1]*step, p[2]-f[2]*step);
positions3[i] = Vec3(p[0]+f[0]*step, p[1]+f[1]*step, p[2]+f[2]*step);
}
context.setPositions(positions2);
State state2 = context.getState(State::Energy);
context.setPositions(positions3);
State state3 = context.getState(State::Energy);
ASSERT_EQUAL_TOL(norm, (state2.getPotentialEnergy()-state3.getPotentialEnergy())/delta, 1e-2)
}
}
int main() {
try {
if (!CpuPlatform::isProcessorSupported()) {
cout << "CPU is not supported. Exiting." << endl;
return 0;
}
testSingleParticle();
testGlobalSettings();
testCutoffAndPeriodic();
for (int i = 5; i < 11; i++) {
testForce(i*i*i, NonbondedForce::NoCutoff, GBSAOBCForce::NoCutoff);
testForce(i*i*i, NonbondedForce::CutoffNonPeriodic, GBSAOBCForce::CutoffNonPeriodic);
testForce(i*i*i, NonbondedForce::CutoffPeriodic, GBSAOBCForce::CutoffPeriodic);
}
}
catch(const exception& e) {
cout << "exception: " << e.what() << endl;
return 1;
}
cout << "Done" << endl;
return 0;
void runPlatformTests() {
}
platforms/cpu/tests/TestCpuHarmonicAngleForce.cpp 0 → 100644
View file @ fd473eea
/* -------------------------------------------------------------------------- *
* OpenMM *
* -------------------------------------------------------------------------- *
* This is part of the OpenMM molecular simulation toolkit originating from *
* Simbios, the NIH National Center for Physics-Based Simulation of *
* Biological Structures at Stanford, funded under the NIH Roadmap for *
* Medical Research, grant U54 GM072970. See https://simtk.org. *
* *
* Portions copyright (c) 2008-2015 Stanford University and the Authors. *
* Authors: Peter Eastman *
* Contributors: *
* *
* Permission is hereby granted, free of charge, to any person obtaining a *
* copy of this software and associated documentation files (the "Software"), *
* to deal in the Software without restriction, including without limitation *
* the rights to use, copy, modify, merge, publish, distribute, sublicense, *
* and/or sell copies of the Software, and to permit persons to whom the *
* Software is furnished to do so, subject to the following conditions: *
* *
* The above copyright notice and this permission notice shall be included in *
* all copies or substantial portions of the Software. *
* *
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR *
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, *
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL *
* THE AUTHORS, CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, *
* DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR *
* OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE *
* USE OR OTHER DEALINGS IN THE SOFTWARE. *
* -------------------------------------------------------------------------- */
#include "CpuTests.h"
#include "TestHarmonicAngleForce.h"
void testParallelComputation() {
System system;
const int numParticles = 200;
for (int i = 0; i < numParticles; i++)
system.addParticle(1.0);
HarmonicAngleForce* force = new HarmonicAngleForce();
for (int i = 2; i < numParticles; i++)
force->addAngle(i-2, i-1, i, 1.1, i);
system.addForce(force);
vector<Vec3> positions(numParticles);
for (int i = 0; i < numParticles; i++)
positions[i] = Vec3(i, i%2, 0);
VerletIntegrator integrator1(0.01);
ReferencePlatform reference;
Context context1(system, integrator1, reference);
context1.setPositions(positions);
State state1 = context1.getState(State::Forces | State::Energy);
VerletIntegrator integrator2(0.01);
Context context2(system, integrator2, platform);
context2.setPositions(positions);
State state2 = context2.getState(State::Forces | State::Energy);
ASSERT_EQUAL_TOL(state1.getPotentialEnergy(), state2.getPotentialEnergy(), 1e-5);
for (int i = 0; i < numParticles; i++)
ASSERT_EQUAL_VEC(state1.getForces()[i], state2.getForces()[i], 1e-5);
}
void runPlatformTests() {
testParallelComputation();
}
platforms/cpu/tests/TestCpuLangevinIntegrator.cpp
View file @ fd473eea
......@@ -6,7 +6,7 @@
* Biological Structures at Stanford, funded under the NIH Roadmap for *
* Medical Research, grant U54 GM072970. See https://simtk.org. *
* *
* Portions copyright (c) 2008-2013 Stanford University and the Authors. *
* Portions copyright (c) 2015 Stanford University and the Authors. *
* Authors: Peter Eastman *
* Contributors: *
* *
......@@ -29,255 +29,8 @@
* USE OR OTHER DEALINGS IN THE SOFTWARE. *
* -------------------------------------------------------------------------- */
/**
* This tests the reference implementation of LangevinIntegrator.
*/
#include "CpuTests.h"
#include "TestLangevinIntegrator.h"
#include "openmm/internal/AssertionUtilities.h"
#include "openmm/Context.h"
#include "CpuPlatform.h"
#include "openmm/HarmonicBondForce.h"
#include "openmm/NonbondedForce.h"
#include "openmm/System.h"
#include "openmm/LangevinIntegrator.h"
#include "SimTKOpenMMRealType.h"
#include "sfmt/SFMT.h"
#include <iostream>
#include <vector>
using namespace OpenMM;
using namespace std;
const double TOL = 1e-5;
void testSingleBond() {
CpuPlatform platform;
System system;
system.addParticle(2.0);
system.addParticle(2.0);
LangevinIntegrator integrator(0, 0.1, 0.01);
HarmonicBondForce* forceField = new HarmonicBondForce();
forceField->addBond(0, 1, 1.5, 1);
system.addForce(forceField);
Context context(system, integrator, platform);
vector<Vec3> positions(2);
positions[0] = Vec3(-1, 0, 0);
positions[1] = Vec3(1, 0, 0);
context.setPositions(positions);
// This is simply a damped harmonic oscillator, so compare it to the analytical solution.
double freq = std::sqrt(1-0.05*0.05);
for (int i = 0; i < 1000; ++i) {
State state = context.getState(State::Positions | State::Velocities);
double time = state.getTime();
double expectedDist = 1.5+0.5*std::exp(-0.05*time)*std::cos(freq*time);
ASSERT_EQUAL_VEC(Vec3(-0.5*expectedDist, 0, 0), state.getPositions()[0], 0.02);
ASSERT_EQUAL_VEC(Vec3(0.5*expectedDist, 0, 0), state.getPositions()[1], 0.02);
double expectedSpeed = -0.5*std::exp(-0.05*time)*(0.05*std::cos(freq*time)+freq*std::sin(freq*time));
ASSERT_EQUAL_VEC(Vec3(-0.5*expectedSpeed, 0, 0), state.getVelocities()[0], 0.02);
ASSERT_EQUAL_VEC(Vec3(0.5*expectedSpeed, 0, 0), state.getVelocities()[1], 0.02);
integrator.step(1);
}
// Not set the friction to a tiny value and see if it conserves energy.
integrator.setFriction(5e-5);
context.setPositions(positions);
State state = context.getState(State::Energy);
double initialEnergy = state.getKineticEnergy()+state.getPotentialEnergy();
for (int i = 0; i < 1000; ++i) {
state = context.getState(State::Energy);
double energy = state.getKineticEnergy()+state.getPotentialEnergy();
ASSERT_EQUAL_TOL(initialEnergy, energy, 0.01);
integrator.step(1);
}
}
void testTemperature() {
const int numParticles = 8;
const double temp = 100.0;
CpuPlatform platform;
System system;
LangevinIntegrator integrator(temp, 2.0, 0.01);
NonbondedForce* forceField = new NonbondedForce();
for (int i = 0; i < numParticles; ++i) {
system.addParticle(2.0);
forceField->addParticle((i%2 == 0 ? 1.0 : -1.0), 1.0, 5.0);
}
system.addForce(forceField);
Context context(system, integrator, platform);
vector<Vec3> positions(numParticles);
for (int i = 0; i < numParticles; ++i)
positions[i] = Vec3((i%2 == 0 ? 2 : -2), (i%4 < 2 ? 2 : -2), (i < 4 ? 2 : -2));
context.setPositions(positions);
// Let it equilibrate.
integrator.step(10000);
// Now run it for a while and see if the temperature is correct.
double ke = 0.0;
for (int i = 0; i < 10000; ++i) {
State state = context.getState(State::Energy);
ke += state.getKineticEnergy();
integrator.step(1);
}
ke /= 10000;
double expected = 0.5*numParticles*3*BOLTZ*temp;
ASSERT_USUALLY_EQUAL_TOL(expected, ke, 6/std::sqrt(10000.0));
}
void testConstraints() {
const int numParticles = 8;
const double temp = 100.0;
CpuPlatform platform;
System system;
LangevinIntegrator integrator(temp, 2.0, 0.01);
integrator.setConstraintTolerance(1e-5);
NonbondedForce* forceField = new NonbondedForce();
for (int i = 0; i < numParticles; ++i) {
system.addParticle(10.0);
forceField->addParticle((i%2 == 0 ? 0.2 : -0.2), 0.5, 5.0);
}
for (int i = 0; i < numParticles-1; ++i)
system.addConstraint(i, i+1, 1.0);
system.addForce(forceField);
Context context(system, integrator, platform);
vector<Vec3> positions(numParticles);
vector<Vec3> velocities(numParticles);
OpenMM_SFMT::SFMT sfmt;
init_gen_rand(0, sfmt);
for (int i = 0; i < numParticles; ++i) {
positions[i] = Vec3(i/2, (i+1)/2, 0);
velocities[i] = Vec3(genrand_real2(sfmt)-0.5, genrand_real2(sfmt)-0.5, genrand_real2(sfmt)-0.5);
}
context.setPositions(positions);
context.setVelocities(velocities);
// Simulate it and see whether the constraints remain satisfied.
for (int i = 0; i < 1000; ++i) {
State state = context.getState(State::Positions);
for (int j = 0; j < numParticles-1; ++j) {
Vec3 p1 = state.getPositions()[j];
Vec3 p2 = state.getPositions()[j+1];
double dist = std::sqrt((p1[0]-p2[0])*(p1[0]-p2[0])+(p1[1]-p2[1])*(p1[1]-p2[1])+(p1[2]-p2[2])*(p1[2]-p2[2]));
ASSERT_EQUAL_TOL(1.0, dist, 2e-5);
}
integrator.step(1);
}
}
void testConstrainedMasslessParticles() {
CpuPlatform platform;
System system;
system.addParticle(0.0);
system.addParticle(1.0);
system.addConstraint(0, 1, 1.5);
vector<Vec3> positions(2);
positions[0] = Vec3(-1, 0, 0);
positions[1] = Vec3(1, 0, 0);
LangevinIntegrator integrator(300.0, 2.0, 0.01);
bool failed = false;
try {
// This should throw an exception.
Context context(system, integrator, platform);
}
catch (exception& ex) {
failed = true;
}
ASSERT(failed);
// Now make both particles massless, which should work.
system.setParticleMass(1, 0.0);
Context context(system, integrator, platform);
context.setPositions(positions);
context.setVelocitiesToTemperature(300.0);
integrator.step(1);
State state = context.getState(State::Velocities | State::Positions);
ASSERT_EQUAL(0.0, state.getVelocities()[0][0]);
}
void testRandomSeed() {
const int numParticles = 8;
const double temp = 100.0;
const double collisionFreq = 10.0;
CpuPlatform platform;
System system;
LangevinIntegrator integrator(temp, 2.0, 0.01);
NonbondedForce* forceField = new NonbondedForce();
for (int i = 0; i < numParticles; ++i) {
system.addParticle(2.0);
forceField->addParticle((i%2 == 0 ? 1.0 : -1.0), 1.0, 5.0);
}
system.addForce(forceField);
vector<Vec3> positions(numParticles);
vector<Vec3> velocities(numParticles);
for (int i = 0; i < numParticles; ++i) {
positions[i] = Vec3((i%2 == 0 ? 2 : -2), (i%4 < 2 ? 2 : -2), (i < 4 ? 2 : -2));
velocities[i] = Vec3(0, 0, 0);
}
// Try twice with the same random seed.
integrator.setRandomNumberSeed(5);
Context context(system, integrator, platform);
context.setPositions(positions);
context.setVelocities(velocities);
integrator.step(10);
State state1 = context.getState(State::Positions);
context.reinitialize();
context.setPositions(positions);
context.setVelocities(velocities);
integrator.step(10);
State state2 = context.getState(State::Positions);
// Try twice with a different random seed.
integrator.setRandomNumberSeed(10);
context.reinitialize();
context.setPositions(positions);
context.setVelocities(velocities);
integrator.step(10);
State state3 = context.getState(State::Positions);
context.reinitialize();
context.setPositions(positions);
context.setVelocities(velocities);
integrator.step(10);
State state4 = context.getState(State::Positions);
// Compare the results.
for (int i = 0; i < numParticles; i++) {
for (int j = 0; j < 3; j++) {
ASSERT_EQUAL_TOL(state1.getPositions()[i][j], state2.getPositions()[i][j], 1e-5);
ASSERT_EQUAL_TOL(state3.getPositions()[i][j], state4.getPositions()[i][j], 1e-5);
ASSERT(state1.getPositions()[i][j] != state3.getPositions()[i][j]);
}
}
}
int main() {
try {
if (!CpuPlatform::isProcessorSupported()) {
cout << "CPU is not supported. Exiting." << endl;
return 0;
}
testSingleBond();
testTemperature();
testConstraints();
testConstrainedMasslessParticles();
testRandomSeed();
}
catch(const exception& e) {
cout << "exception: " << e.what() << endl;
return 1;
}
cout << "Done" << endl;
return 0;
void runPlatformTests() {
}
platforms/cpu/tests/TestCpuNonbondedForce.cpp
View file @ fd473eea
......@@ -6,7 +6,7 @@
* Biological Structures at Stanford, funded under the NIH Roadmap for *
* Medical Research, grant U54 GM072970. See https://simtk.org. *
* *
* Portions copyright (c) 2008-2015 Stanford University and the Authors. *
* Portions copyright (c) 2015 Stanford University and the Authors. *
* Authors: Peter Eastman *
* Contributors: *
* *
......@@ -29,684 +29,8 @@
* USE OR OTHER DEALINGS IN THE SOFTWARE. *
* -------------------------------------------------------------------------- */
/**
* This tests all the different force terms in the CUDA implementation of NonbondedForce.
*/
#include "CpuTests.h"
#include "TestNonbondedForce.h"
#include "openmm/internal/AssertionUtilities.h"
#include "openmm/Context.h"
#include "CpuPlatform.h"
#include "ReferencePlatform.h"
#include "openmm/HarmonicBondForce.h"
#include "openmm/NonbondedForce.h"
#include "openmm/System.h"
#include "openmm/LangevinIntegrator.h"
#include "openmm/VerletIntegrator.h"
#include "openmm/internal/ContextImpl.h"
#include "SimTKOpenMMRealType.h"
#include "sfmt/SFMT.h"
#include <iostream>
#include <vector>
using namespace OpenMM;
using namespace std;
CpuPlatform platform;
const double TOL = 1e-5;
void testCoulomb() {
System system;
system.addParticle(1.0);
system.addParticle(1.0);
VerletIntegrator integrator(0.01);
NonbondedForce* forceField = new NonbondedForce();
forceField->addParticle(0.5, 1, 0);
forceField->addParticle(-1.5, 1, 0);
system.addForce(forceField);
Context context(system, integrator, platform);
vector<Vec3> positions(2);
positions[0] = Vec3(0, 0, 0);
positions[1] = Vec3(2, 0, 0);
context.setPositions(positions);
State state = context.getState(State::Forces | State::Energy);
const vector<Vec3>& forces = state.getForces();
double force = ONE_4PI_EPS0*(-0.75)/4.0;
ASSERT_EQUAL_VEC(Vec3(-force, 0, 0), forces[0], TOL);
ASSERT_EQUAL_VEC(Vec3(force, 0, 0), forces[1], TOL);
ASSERT_EQUAL_TOL(ONE_4PI_EPS0*(-0.75)/2.0, state.getPotentialEnergy(), TOL);
}
void testLJ() {
System system;
system.addParticle(1.0);
system.addParticle(1.0);
VerletIntegrator integrator(0.01);
NonbondedForce* forceField = new NonbondedForce();
forceField->addParticle(0, 1.2, 1);
forceField->addParticle(0, 1.4, 2);
system.addForce(forceField);
Context context(system, integrator, platform);
vector<Vec3> positions(2);
positions[0] = Vec3(0, 0, 0);
positions[1] = Vec3(2, 0, 0);
context.setPositions(positions);
State state = context.getState(State::Forces | State::Energy);
const vector<Vec3>& forces = state.getForces();
double x = 1.3/2.0;
double eps = SQRT_TWO;
double force = 4.0*eps*(12*std::pow(x, 12.0)-6*std::pow(x, 6.0))/2.0;
ASSERT_EQUAL_VEC(Vec3(-force, 0, 0), forces[0], TOL);
ASSERT_EQUAL_VEC(Vec3(force, 0, 0), forces[1], TOL);
ASSERT_EQUAL_TOL(4.0*eps*(std::pow(x, 12.0)-std::pow(x, 6.0)), state.getPotentialEnergy(), TOL);
}
void testExclusionsAnd14() {
System system;
NonbondedForce* nonbonded = new NonbondedForce();
for (int i = 0; i < 5; ++i) {
system.addParticle(1.0);
nonbonded->addParticle(0, 1.5, 0);
}
vector<pair<int, int> > bonds;
bonds.push_back(pair<int, int>(0, 1));
bonds.push_back(pair<int, int>(1, 2));
bonds.push_back(pair<int, int>(2, 3));
bonds.push_back(pair<int, int>(3, 4));
nonbonded->createExceptionsFromBonds(bonds, 0.0, 0.0);
int first14, second14;
for (int i = 0; i < nonbonded->getNumExceptions(); i++) {
int particle1, particle2;
double chargeProd, sigma, epsilon;
nonbonded->getExceptionParameters(i, particle1, particle2, chargeProd, sigma, epsilon);
if ((particle1 == 0 && particle2 == 3) || (particle1 == 3 && particle2 == 0))
first14 = i;
if ((particle1 == 1 && particle2 == 4) || (particle1 == 4 && particle2 == 1))
second14 = i;
}
system.addForce(nonbonded);
VerletIntegrator integrator(0.01);
Context context(system, integrator, platform);
for (int i = 1; i < 5; ++i) {
// Test LJ forces
vector<Vec3> positions(5);
const double r = 1.0;
for (int j = 0; j < 5; ++j) {
nonbonded->setParticleParameters(j, 0, 1.5, 0);
positions[j] = Vec3(0, j, 0);
}
nonbonded->setParticleParameters(0, 0, 1.5, 1);
nonbonded->setParticleParameters(i, 0, 1.5, 1);
nonbonded->setExceptionParameters(first14, 0, 3, 0, 1.5, i == 3 ? 0.5 : 0.0);
nonbonded->setExceptionParameters(second14, 1, 4, 0, 1.5, 0.0);
positions[i] = Vec3(r, 0, 0);
context.reinitialize();
context.setPositions(positions);
State state = context.getState(State::Forces | State::Energy);
const vector<Vec3>& forces = state.getForces();
double x = 1.5/r;
double eps = 1.0;
double force = 4.0*eps*(12*std::pow(x, 12.0)-6*std::pow(x, 6.0))/r;
double energy = 4.0*eps*(std::pow(x, 12.0)-std::pow(x, 6.0));
if (i == 3) {
force *= 0.5;
energy *= 0.5;
}
if (i < 3) {
force = 0;
energy = 0;
}
ASSERT_EQUAL_VEC(Vec3(-force, 0, 0), forces[0], TOL);
ASSERT_EQUAL_VEC(Vec3(force, 0, 0), forces[i], TOL);
ASSERT_EQUAL_TOL(energy, state.getPotentialEnergy(), TOL);
// Test Coulomb forces
nonbonded->setParticleParameters(0, 2, 1.5, 0);
nonbonded->setParticleParameters(i, 2, 1.5, 0);
nonbonded->setExceptionParameters(first14, 0, 3, i == 3 ? 4/1.2 : 0, 1.5, 0);
nonbonded->setExceptionParameters(second14, 1, 4, 0, 1.5, 0);
context.reinitialize();
context.setPositions(positions);
state = context.getState(State::Forces | State::Energy);
const vector<Vec3>& forces2 = state.getForces();
force = ONE_4PI_EPS0*4/(r*r);
energy = ONE_4PI_EPS0*4/r;
if (i == 3) {
force /= 1.2;
energy /= 1.2;
}
if (i < 3) {
force = 0;
energy = 0;
}
ASSERT_EQUAL_VEC(Vec3(-force, 0, 0), forces2[0], TOL);
ASSERT_EQUAL_VEC(Vec3(force, 0, 0), forces2[i], TOL);
ASSERT_EQUAL_TOL(energy, state.getPotentialEnergy(), TOL);
}
}
void testCutoff() {
System system;
system.addParticle(1.0);
system.addParticle(1.0);
system.addParticle(1.0);
VerletIntegrator integrator(0.01);
NonbondedForce* forceField = new NonbondedForce();
forceField->addParticle(1.0, 1, 0);
forceField->addParticle(1.0, 1, 0);
forceField->addParticle(1.0, 1, 0);
forceField->setNonbondedMethod(NonbondedForce::CutoffNonPeriodic);
const double cutoff = 2.9;
forceField->setCutoffDistance(cutoff);
const double eps = 50.0;
forceField->setReactionFieldDielectric(eps);
system.addForce(forceField);
Context context(system, integrator, platform);
vector<Vec3> positions(3);
positions[0] = Vec3(0, 0, 0);
positions[1] = Vec3(0, 2, 0);
positions[2] = Vec3(0, 3, 0);
context.setPositions(positions);
State state = context.getState(State::Forces | State::Energy);
const vector<Vec3>& forces = state.getForces();
const double krf = (1.0/(cutoff*cutoff*cutoff))*(eps-1.0)/(2.0*eps+1.0);
const double crf = (1.0/cutoff)*(3.0*eps)/(2.0*eps+1.0);
const double force1 = ONE_4PI_EPS0*(1.0)*(0.25-2.0*krf*2.0);
const double force2 = ONE_4PI_EPS0*(1.0)*(1.0-2.0*krf*1.0);
ASSERT_EQUAL_VEC(Vec3(0, -force1, 0), forces[0], TOL);
ASSERT_EQUAL_VEC(Vec3(0, force1-force2, 0), forces[1], TOL);
ASSERT_EQUAL_VEC(Vec3(0, force2, 0), forces[2], TOL);
const double energy1 = ONE_4PI_EPS0*(1.0)*(0.5+krf*4.0-crf);
const double energy2 = ONE_4PI_EPS0*(1.0)*(1.0+krf*1.0-crf);
ASSERT_EQUAL_TOL(energy1+energy2, state.getPotentialEnergy(), TOL);
}
void testCutoff14() {
System system;
VerletIntegrator integrator(0.01);
NonbondedForce* nonbonded = new NonbondedForce();
nonbonded->setNonbondedMethod(NonbondedForce::CutoffNonPeriodic);
for (int i = 0; i < 5; ++i) {
system.addParticle(1.0);
nonbonded->addParticle(0, 1.5, 0);
}
const double cutoff = 3.5;
nonbonded->setCutoffDistance(cutoff);
const double eps = 30.0;
nonbonded->setReactionFieldDielectric(eps);
vector<pair<int, int> > bonds;
bonds.push_back(pair<int, int>(0, 1));
bonds.push_back(pair<int, int>(1, 2));
bonds.push_back(pair<int, int>(2, 3));
bonds.push_back(pair<int, int>(3, 4));
nonbonded->createExceptionsFromBonds(bonds, 0.0, 0.0);
int first14, second14;
for (int i = 0; i < nonbonded->getNumExceptions(); i++) {
int particle1, particle2;
double chargeProd, sigma, epsilon;
nonbonded->getExceptionParameters(i, particle1, particle2, chargeProd, sigma, epsilon);
if ((particle1 == 0 && particle2 == 3) || (particle1 == 3 && particle2 == 0))
first14 = i;
if ((particle1 == 1 && particle2 == 4) || (particle1 == 4 && particle2 == 1))
second14 = i;
}
system.addForce(nonbonded);
Context context(system, integrator, platform);
vector<Vec3> positions(5);
positions[0] = Vec3(0, 0, 0);
positions[1] = Vec3(1, 0, 0);
positions[2] = Vec3(2, 0, 0);
positions[3] = Vec3(3, 0, 0);
positions[4] = Vec3(4, 0, 0);
for (int i = 1; i < 5; ++i) {
// Test LJ forces
nonbonded->setParticleParameters(0, 0, 1.5, 1);
for (int j = 1; j < 5; ++j)
nonbonded->setParticleParameters(j, 0, 1.5, 0);
nonbonded->setParticleParameters(i, 0, 1.5, 1);
nonbonded->setExceptionParameters(first14, 0, 3, 0, 1.5, i == 3 ? 0.5 : 0.0);
nonbonded->setExceptionParameters(second14, 1, 4, 0, 1.5, 0.0);
context.reinitialize();
context.setPositions(positions);
State state = context.getState(State::Forces | State::Energy);
const vector<Vec3>& forces = state.getForces();
double r = positions[i][0];
double x = 1.5/r;
double e = 1.0;
double force = 4.0*e*(12*std::pow(x, 12.0)-6*std::pow(x, 6.0))/r;
double energy = 4.0*e*(std::pow(x, 12.0)-std::pow(x, 6.0));
if (i == 3) {
force *= 0.5;
energy *= 0.5;
}
if (i < 3 || r > cutoff) {
force = 0;
energy = 0;
}
ASSERT_EQUAL_VEC(Vec3(-force, 0, 0), forces[0], TOL);
ASSERT_EQUAL_VEC(Vec3(force, 0, 0), forces[i], TOL);
ASSERT_EQUAL_TOL(energy, state.getPotentialEnergy(), TOL);
// Test Coulomb forces
const double q = 0.7;
nonbonded->setParticleParameters(0, q, 1.5, 0);
nonbonded->setParticleParameters(i, q, 1.5, 0);
nonbonded->setExceptionParameters(first14, 0, 3, i == 3 ? q*q/1.2 : 0, 1.5, 0);
nonbonded->setExceptionParameters(second14, 1, 4, 0, 1.5, 0);
context.reinitialize();
context.setPositions(positions);
state = context.getState(State::Forces | State::Energy);
const vector<Vec3>& forces2 = state.getForces();
force = ONE_4PI_EPS0*q*q/(r*r);
energy = ONE_4PI_EPS0*q*q/r;
if (i == 3) {
force /= 1.2;
energy /= 1.2;
}
if (i < 3 || r > cutoff) {
force = 0;
energy = 0;
}
ASSERT_EQUAL_VEC(Vec3(-force, 0, 0), forces2[0], TOL);
ASSERT_EQUAL_VEC(Vec3(force, 0, 0), forces2[i], TOL);
ASSERT_EQUAL_TOL(energy, state.getPotentialEnergy(), TOL);
}
}
void testPeriodic() {
System system;
system.addParticle(1.0);
system.addParticle(1.0);
system.addParticle(1.0);
VerletIntegrator integrator(0.01);
NonbondedForce* nonbonded = new NonbondedForce();
nonbonded->addParticle(1.0, 1, 0);
nonbonded->addParticle(1.0, 1, 0);
nonbonded->addParticle(1.0, 1, 0);
nonbonded->addException(0, 1, 0.0, 1.0, 0.0);
nonbonded->setNonbondedMethod(NonbondedForce::CutoffPeriodic);
const double cutoff = 2.0;
nonbonded->setCutoffDistance(cutoff);
system.setDefaultPeriodicBoxVectors(Vec3(4, 0, 0), Vec3(0, 4, 0), Vec3(0, 0, 4));
system.addForce(nonbonded);
Context context(system, integrator, platform);
vector<Vec3> positions(3);
positions[0] = Vec3(0, 0, 0);
positions[1] = Vec3(2, 0, 0);
positions[2] = Vec3(3, 0, 0);
context.setPositions(positions);
State state = context.getState(State::Forces | State::Energy);
const vector<Vec3>& forces = state.getForces();
const double eps = 78.3;
const double krf = (1.0/(cutoff*cutoff*cutoff))*(eps-1.0)/(2.0*eps+1.0);
const double crf = (1.0/cutoff)*(3.0*eps)/(2.0*eps+1.0);
const double force = ONE_4PI_EPS0*(1.0)*(1.0-2.0*krf*1.0);
ASSERT_EQUAL_VEC(Vec3(force, 0, 0), forces[0], TOL);
ASSERT_EQUAL_VEC(Vec3(-force, 0, 0), forces[1], TOL);
ASSERT_EQUAL_VEC(Vec3(0, 0, 0), forces[2], TOL);
ASSERT_EQUAL_TOL(2*ONE_4PI_EPS0*(1.0)*(1.0+krf*1.0-crf), state.getPotentialEnergy(), TOL);
}
void testTriclinic() {
System system;
system.addParticle(1.0);
system.addParticle(1.0);
Vec3 a(3.1, 0, 0);
Vec3 b(0.4, 3.5, 0);
Vec3 c(-0.1, -0.5, 4.0);
system.setDefaultPeriodicBoxVectors(a, b, c);
VerletIntegrator integrator(0.01);
NonbondedForce* nonbonded = new NonbondedForce();
nonbonded->addParticle(1.0, 1, 0);
nonbonded->addParticle(1.0, 1, 0);
nonbonded->setNonbondedMethod(NonbondedForce::CutoffPeriodic);
const double cutoff = 1.5;
nonbonded->setCutoffDistance(cutoff);
system.addForce(nonbonded);
Context context(system, integrator, platform);
vector<Vec3> positions(2);
OpenMM_SFMT::SFMT sfmt;
init_gen_rand(0, sfmt);
const double eps = 78.3;
const double krf = (1.0/(cutoff*cutoff*cutoff))*(eps-1.0)/(2.0*eps+1.0);
const double crf = (1.0/cutoff)*(3.0*eps)/(2.0*eps+1.0);
for (int iteration = 0; iteration < 50; iteration++) {
// Generate random positions for the two particles.
positions[0] = a*genrand_real2(sfmt) + b*genrand_real2(sfmt) + c*genrand_real2(sfmt);
positions[1] = a*genrand_real2(sfmt) + b*genrand_real2(sfmt) + c*genrand_real2(sfmt);
context.setPositions(positions);
// Loop over all possible periodic copies and find the nearest one.
Vec3 delta;
double distance2 = 100.0;
for (int i = -1; i < 2; i++)
for (int j = -1; j < 2; j++)
for (int k = -1; k < 2; k++) {
Vec3 d = positions[1]-positions[0]+a*i+b*j+c*k;
if (d.dot(d) < distance2) {
delta = d;
distance2 = d.dot(d);
}
}
double distance = sqrt(distance2);
// See if the force and energy are correct.
State state = context.getState(State::Forces | State::Energy);
if (distance >= cutoff) {
ASSERT_EQUAL(0.0, state.getPotentialEnergy());
ASSERT_EQUAL_VEC(Vec3(0, 0, 0), state.getForces()[0], 0);
ASSERT_EQUAL_VEC(Vec3(0, 0, 0), state.getForces()[1], 0);
}
else {
const Vec3 force = delta*ONE_4PI_EPS0*(-1.0/(distance*distance*distance)+2.0*krf);
ASSERT_EQUAL_TOL(ONE_4PI_EPS0*(1.0/distance+krf*distance*distance-crf), state.getPotentialEnergy(), 1e-4);
ASSERT_EQUAL_VEC(force, state.getForces()[0], 2e-5);
ASSERT_EQUAL_VEC(-force, state.getForces()[1], 2e-5);
}
}
}
void testLargeSystem() {
const int numMolecules = 600;
const int numParticles = numMolecules*2;
const double cutoff = 2.0;
const double boxSize = 20.0;
const double tol = 2e-3;
ReferencePlatform reference;
System system;
for (int i = 0; i < numParticles; i++)
system.addParticle(1.0);
NonbondedForce* nonbonded = new NonbondedForce();
HarmonicBondForce* bonds = new HarmonicBondForce();
vector<Vec3> positions(numParticles);
vector<Vec3> velocities(numParticles);
OpenMM_SFMT::SFMT sfmt;
init_gen_rand(0, sfmt);
for (int i = 0; i < numMolecules; i++) {
if (i < numMolecules/2) {
nonbonded->addParticle(-1.0, 0.2, 0.1);
nonbonded->addParticle(1.0, 0.1, 0.1);
}
else {
nonbonded->addParticle(-1.0, 0.2, 0.2);
nonbonded->addParticle(1.0, 0.1, 0.2);
}
positions[2*i] = Vec3(boxSize*genrand_real2(sfmt), boxSize*genrand_real2(sfmt), boxSize*genrand_real2(sfmt));
positions[2*i+1] = Vec3(positions[2*i][0]+1.0, positions[2*i][1], positions[2*i][2]);
velocities[2*i] = Vec3(genrand_real2(sfmt), genrand_real2(sfmt), genrand_real2(sfmt));
velocities[2*i+1] = Vec3(genrand_real2(sfmt), genrand_real2(sfmt), genrand_real2(sfmt));
bonds->addBond(2*i, 2*i+1, 1.0, 0.1);
nonbonded->addException(2*i, 2*i+1, 0.0, 0.15, 0.0);
}
// Try with cutoffs but not periodic boundary conditions, and make sure the cl and Reference
// platforms agree.
nonbonded->setNonbondedMethod(NonbondedForce::CutoffNonPeriodic);
nonbonded->setCutoffDistance(cutoff);
system.addForce(nonbonded);
system.addForce(bonds);
VerletIntegrator integrator1(0.01);
VerletIntegrator integrator2(0.01);
Context cpuContext(system, integrator1, platform);
Context referenceContext(system, integrator2, reference);
cpuContext.setPositions(positions);
cpuContext.setVelocities(velocities);
referenceContext.setPositions(positions);
referenceContext.setVelocities(velocities);
State cpuState = cpuContext.getState(State::Positions | State::Velocities | State::Forces | State::Energy);
State referenceState = referenceContext.getState(State::Positions | State::Velocities | State::Forces | State::Energy);
for (int i = 0; i < numParticles; i++) {
ASSERT_EQUAL_VEC(cpuState.getPositions()[i], referenceState.getPositions()[i], tol);
ASSERT_EQUAL_VEC(cpuState.getVelocities()[i], referenceState.getVelocities()[i], tol);
ASSERT_EQUAL_VEC(cpuState.getForces()[i], referenceState.getForces()[i], tol);
}
ASSERT_EQUAL_TOL(cpuState.getPotentialEnergy(), referenceState.getPotentialEnergy(), tol);
// Now do the same thing with periodic boundary conditions.
nonbonded->setNonbondedMethod(NonbondedForce::CutoffPeriodic);
system.setDefaultPeriodicBoxVectors(Vec3(boxSize, 0, 0), Vec3(0, boxSize, 0), Vec3(0, 0, boxSize));
cpuContext.reinitialize();
referenceContext.reinitialize();
cpuContext.setPositions(positions);
cpuContext.setVelocities(velocities);
referenceContext.setPositions(positions);
referenceContext.setVelocities(velocities);
cpuState = cpuContext.getState(State::Positions | State::Velocities | State::Forces | State::Energy);
referenceState = referenceContext.getState(State::Positions | State::Velocities | State::Forces | State::Energy);
for (int i = 0; i < numParticles; i++) {
double dx = cpuState.getPositions()[i][0]-referenceState.getPositions()[i][0];
double dy = cpuState.getPositions()[i][1]-referenceState.getPositions()[i][1];
double dz = cpuState.getPositions()[i][2]-referenceState.getPositions()[i][2];
ASSERT_EQUAL_TOL(fmod(cpuState.getPositions()[i][0]-referenceState.getPositions()[i][0], boxSize), 0, tol);
ASSERT_EQUAL_TOL(fmod(cpuState.getPositions()[i][1]-referenceState.getPositions()[i][1], boxSize), 0, tol);
ASSERT_EQUAL_TOL(fmod(cpuState.getPositions()[i][2]-referenceState.getPositions()[i][2], boxSize), 0, tol);
ASSERT_EQUAL_VEC(cpuState.getVelocities()[i], referenceState.getVelocities()[i], tol);
ASSERT_EQUAL_VEC(cpuState.getForces()[i], referenceState.getForces()[i], tol);
}
ASSERT_EQUAL_TOL(cpuState.getPotentialEnergy(), referenceState.getPotentialEnergy(), tol);
}
void testDispersionCorrection() {
// Create a box full of identical particles.
int gridSize = 5;
int numParticles = gridSize*gridSize*gridSize;
double boxSize = gridSize*0.7;
double cutoff = boxSize/3;
System system;
VerletIntegrator integrator(0.01);
NonbondedForce* nonbonded = new NonbondedForce();
vector<Vec3> positions(numParticles);
int index = 0;
for (int i = 0; i < gridSize; i++)
for (int j = 0; j < gridSize; j++)
for (int k = 0; k < gridSize; k++) {
system.addParticle(1.0);
nonbonded->addParticle(0, 1.1, 0.5);
positions[index] = Vec3(i*boxSize/gridSize, j*boxSize/gridSize, k*boxSize/gridSize);
index++;
}
nonbonded->setNonbondedMethod(NonbondedForce::CutoffPeriodic);
nonbonded->setCutoffDistance(cutoff);
system.setDefaultPeriodicBoxVectors(Vec3(boxSize, 0, 0), Vec3(0, boxSize, 0), Vec3(0, 0, boxSize));
system.addForce(nonbonded);
// See if the correction has the correct value.
Context context(system, integrator, platform);
context.setPositions(positions);
double energy1 = context.getState(State::Energy).getPotentialEnergy();
nonbonded->setUseDispersionCorrection(false);
context.reinitialize();
context.setPositions(positions);
double energy2 = context.getState(State::Energy).getPotentialEnergy();
double term1 = (0.5*pow(1.1, 12)/pow(cutoff, 9))/9;
double term2 = (0.5*pow(1.1, 6)/pow(cutoff, 3))/3;
double expected = 8*M_PI*numParticles*numParticles*(term1-term2)/(boxSize*boxSize*boxSize);
ASSERT_EQUAL_TOL(expected, energy1-energy2, 1e-4);
// Now modify half the particles to be different, and see if it is still correct.
int numType2 = 0;
for (int i = 0; i < numParticles; i += 2) {
nonbonded->setParticleParameters(i, 0, 1, 1);
numType2++;
}
int numType1 = numParticles-numType2;
nonbonded->updateParametersInContext(context);
energy2 = context.getState(State::Energy).getPotentialEnergy();
nonbonded->setUseDispersionCorrection(true);
context.reinitialize();
context.setPositions(positions);
energy1 = context.getState(State::Energy).getPotentialEnergy();
term1 = ((numType1*(numType1+1))/2)*(0.5*pow(1.1, 12)/pow(cutoff, 9))/9;
term2 = ((numType1*(numType1+1))/2)*(0.5*pow(1.1, 6)/pow(cutoff, 3))/3;
term1 += ((numType2*(numType2+1))/2)*(1*pow(1.0, 12)/pow(cutoff, 9))/9;
term2 += ((numType2*(numType2+1))/2)*(1*pow(1.0, 6)/pow(cutoff, 3))/3;
double combinedSigma = 0.5*(1+1.1);
double combinedEpsilon = sqrt(1*0.5);
term1 += (numType1*numType2)*(combinedEpsilon*pow(combinedSigma, 12)/pow(cutoff, 9))/9;
term2 += (numType1*numType2)*(combinedEpsilon*pow(combinedSigma, 6)/pow(cutoff, 3))/3;
term1 /= (numParticles*(numParticles+1))/2;
term2 /= (numParticles*(numParticles+1))/2;
expected = 8*M_PI*numParticles*numParticles*(term1-term2)/(boxSize*boxSize*boxSize);
ASSERT_EQUAL_TOL(expected, energy1-energy2, 1e-4);
}
void testChangingParameters() {
const int numMolecules = 600;
const int numParticles = numMolecules*2;
const double cutoff = 2.0;
const double boxSize = 20.0;
const double tol = 2e-3;
ReferencePlatform reference;
System system;
for (int i = 0; i < numParticles; i++)
system.addParticle(1.0);
NonbondedForce* nonbonded = new NonbondedForce();
vector<Vec3> positions(numParticles);
OpenMM_SFMT::SFMT sfmt;
init_gen_rand(0, sfmt);
for (int i = 0; i < numMolecules; i++) {
if (i < numMolecules/2) {
nonbonded->addParticle(-1.0, 0.2, 0.1);
nonbonded->addParticle(1.0, 0.1, 0.1);
}
else {
nonbonded->addParticle(-1.0, 0.2, 0.2);
nonbonded->addParticle(1.0, 0.1, 0.2);
}
positions[2*i] = Vec3(boxSize*genrand_real2(sfmt), boxSize*genrand_real2(sfmt), boxSize*genrand_real2(sfmt));
positions[2*i+1] = Vec3(positions[2*i][0]+1.0, positions[2*i][1], positions[2*i][2]);
system.addConstraint(2*i, 2*i+1, 1.0);
nonbonded->addException(2*i, 2*i+1, 0.0, 0.15, 0.0);
}
nonbonded->setNonbondedMethod(NonbondedForce::PME);
nonbonded->setCutoffDistance(cutoff);
system.addForce(nonbonded);
system.setDefaultPeriodicBoxVectors(Vec3(boxSize, 0, 0), Vec3(0, boxSize, 0), Vec3(0, 0, boxSize));
// See if Reference and CPU give the same forces and energies.
VerletIntegrator integrator1(0.01);
VerletIntegrator integrator2(0.01);
Context cpuContext(system, integrator1, platform);
Context referenceContext(system, integrator2, reference);
cpuContext.setPositions(positions);
referenceContext.setPositions(positions);
State cpuState = cpuContext.getState(State::Forces | State::Energy);
State referenceState = referenceContext.getState(State::Forces | State::Energy);
for (int i = 0; i < numParticles; i++)
ASSERT_EQUAL_VEC(cpuState.getForces()[i], referenceState.getForces()[i], tol);
ASSERT_EQUAL_TOL(cpuState.getPotentialEnergy(), referenceState.getPotentialEnergy(), tol);
// Now modify parameters and see if they still agree.
for (int i = 0; i < numParticles; i += 5) {
double charge, sigma, epsilon;
nonbonded->getParticleParameters(i, charge, sigma, epsilon);
nonbonded->setParticleParameters(i, 1.5*charge, 1.1*sigma, 1.7*epsilon);
}
nonbonded->updateParametersInContext(cpuContext);
nonbonded->updateParametersInContext(referenceContext);
cpuState = cpuContext.getState(State::Forces | State::Energy);
referenceState = referenceContext.getState(State::Forces | State::Energy);
for (int i = 0; i < numParticles; i++)
ASSERT_EQUAL_VEC(cpuState.getForces()[i], referenceState.getForces()[i], tol);
ASSERT_EQUAL_TOL(cpuState.getPotentialEnergy(), referenceState.getPotentialEnergy(), tol);
}
void testSwitchingFunction(NonbondedForce::NonbondedMethod method) {
System system;
system.setDefaultPeriodicBoxVectors(Vec3(6, 0, 0), Vec3(0, 6, 0), Vec3(0, 0, 6));
system.addParticle(1.0);
system.addParticle(1.0);
VerletIntegrator integrator(0.01);
NonbondedForce* nonbonded = new NonbondedForce();
nonbonded->addParticle(0, 1.2, 1);
nonbonded->addParticle(0, 1.4, 2);
nonbonded->setNonbondedMethod(method);
nonbonded->setCutoffDistance(2.0);
nonbonded->setUseSwitchingFunction(true);
nonbonded->setSwitchingDistance(1.5);
nonbonded->setUseDispersionCorrection(false);
system.addForce(nonbonded);
Context context(system, integrator, platform);
vector<Vec3> positions(2);
positions[0] = Vec3(0, 0, 0);
double eps = SQRT_TWO;
// Compute the interaction at various distances.
for (double r = 1.0; r < 2.5; r += 0.1) {
positions[1] = Vec3(r, 0, 0);
context.setPositions(positions);
State state = context.getState(State::Forces | State::Energy);
// See if the energy is correct.
double x = 1.3/r;
double expectedEnergy = 4.0*eps*(std::pow(x, 12.0)-std::pow(x, 6.0));
double switchValue;
if (r <= 1.5)
switchValue = 1;
else if (r >= 2.0)
switchValue = 0;
else {
double t = (r-1.5)/0.5;
switchValue = 1+t*t*t*(-10+t*(15-t*6));
}
ASSERT_EQUAL_TOL(switchValue*expectedEnergy, state.getPotentialEnergy(), TOL);
// See if the force is the gradient of the energy.
double delta = 1e-3;
positions[1] = Vec3(r-delta, 0, 0);
context.setPositions(positions);
double e1 = context.getState(State::Energy).getPotentialEnergy();
positions[1] = Vec3(r+delta, 0, 0);
context.setPositions(positions);
double e2 = context.getState(State::Energy).getPotentialEnergy();
ASSERT_EQUAL_TOL((e2-e1)/(2*delta), state.getForces()[0][0], 1e-3);
}
}
int main(int argc, char* argv[]) {
try {
if (!CpuPlatform::isProcessorSupported()) {
cout << "CPU is not supported. Exiting." << endl;
return 0;
}
testCoulomb();
testLJ();
testExclusionsAnd14();
testCutoff();
testCutoff14();
testPeriodic();
testTriclinic();
testLargeSystem();
testDispersionCorrection();
testChangingParameters();
testSwitchingFunction(NonbondedForce::CutoffNonPeriodic);
testSwitchingFunction(NonbondedForce::PME);
}
catch(const exception& e) {
cout << "exception: " << e.what() << endl;
return 1;
}
cout << "Done" << endl;
return 0;
void runPlatformTests() {
}
platforms/cpu/tests/TestCpuPeriodicTorsionForce.cpp
View file @ fd473eea
......@@ -6,7 +6,7 @@
* Biological Structures at Stanford, funded under the NIH Roadmap for *
* Medical Research, grant U54 GM072970. See https://simtk.org. *
* *
* Portions copyright (c) 2008-2014 Stanford University and the Authors. *
* Portions copyright (c) 2008-2015 Stanford University and the Authors. *
* Authors: Peter Eastman *
* Contributors: *
* *
......@@ -29,69 +29,8 @@
* USE OR OTHER DEALINGS IN THE SOFTWARE. *
* -------------------------------------------------------------------------- */
/**
* This tests the CPU implementation of PeriodicTorsionForce.
*/
#include "openmm/internal/AssertionUtilities.h"
#include "openmm/Context.h"
#include "CpuPlatform.h"
#include "openmm/PeriodicTorsionForce.h"
#include "openmm/System.h"
#include "openmm/VerletIntegrator.h"
#include "SimTKOpenMMRealType.h"
#include <iostream>
#include <vector>
using namespace OpenMM;
using namespace std;
CpuPlatform platform;
const double TOL = 1e-5;
void testPeriodicTorsions() {
System system;
system.addParticle(1.0);
system.addParticle(1.0);
system.addParticle(1.0);
system.addParticle(1.0);
VerletIntegrator integrator(0.01);
PeriodicTorsionForce* forceField = new PeriodicTorsionForce();
forceField->addTorsion(0, 1, 2, 3, 2, PI_M/3, 1.1);
system.addForce(forceField);
Context context(system, integrator, platform);
vector<Vec3> positions(4);
positions[0] = Vec3(0, 1, 0);
positions[1] = Vec3(0, 0, 0);
positions[2] = Vec3(1, 0, 0);
positions[3] = Vec3(1, 0, 2);
context.setPositions(positions);
State state = context.getState(State::Forces | State::Energy);
{
const vector<Vec3>& forces = state.getForces();
double torque = -2*1.1*std::sin(2*PI_M/3);
ASSERT_EQUAL_VEC(Vec3(0, 0, torque), forces[0], TOL);
ASSERT_EQUAL_VEC(Vec3(0, 0.5*torque, 0), forces[3], TOL);
ASSERT_EQUAL_VEC(Vec3(forces[0][0]+forces[1][0]+forces[2][0]+forces[3][0], forces[0][1]+forces[1][1]+forces[2][1]+forces[3][1], forces[0][2]+forces[1][2]+forces[2][2]+forces[3][2]), Vec3(0, 0, 0), TOL);
ASSERT_EQUAL_TOL(1.1*(1+std::cos(2*PI_M/3)), state.getPotentialEnergy(), TOL);
}
// Try changing the torsion parameters and make sure it's still correct.
forceField->setTorsionParameters(0, 0, 1, 2, 3, 3, PI_M/3.2, 1.3);
forceField->updateParametersInContext(context);
state = context.getState(State::Forces | State::Energy);
{
const vector<Vec3>& forces = state.getForces();
double dtheta = (3*PI_M/2)-(PI_M/3.2);
double torque = -3*1.3*std::sin(dtheta);
ASSERT_EQUAL_VEC(Vec3(0, 0, torque), forces[0], TOL);
ASSERT_EQUAL_VEC(Vec3(0, 0.5*torque, 0), forces[3], TOL);
ASSERT_EQUAL_VEC(Vec3(forces[0][0]+forces[1][0]+forces[2][0]+forces[3][0], forces[0][1]+forces[1][1]+forces[2][1]+forces[3][1], forces[0][2]+forces[1][2]+forces[2][2]+forces[3][2]), Vec3(0, 0, 0), TOL);
ASSERT_EQUAL_TOL(1.3*(1+std::cos(dtheta)), state.getPotentialEnergy(), TOL);
}
}
#include "CpuTests.h"
#include "TestPeriodicTorsionForce.h"
void testParallelComputation() {
System system;
......@@ -119,15 +58,6 @@ void testParallelComputation() {
ASSERT_EQUAL_VEC(state1.getForces()[i], state2.getForces()[i], 1e-5);
}
int main(int argc, char* argv[]) {
try {
testPeriodicTorsions();
testParallelComputation();
}
catch(const exception& e) {
cout << "exception: " << e.what() << endl;
return 1;
}
cout << "Done" << endl;
return 0;
void runPlatformTests() {
testParallelComputation();
}
platforms/cpu/tests/TestCpuRBTorsionForce.cpp
View file @ fd473eea
......@@ -6,7 +6,7 @@
* Biological Structures at Stanford, funded under the NIH Roadmap for *
* Medical Research, grant U54 GM072970. See https://simtk.org. *
* *
* Portions copyright (c) 2008-2014 Stanford University and the Authors. *
* Portions copyright (c) 2008-2015 Stanford University and the Authors. *
* Authors: Peter Eastman *
* Contributors: *
* *
......@@ -29,88 +29,8 @@
* USE OR OTHER DEALINGS IN THE SOFTWARE. *
* -------------------------------------------------------------------------- */
/**
* This tests the CUDA implementation of RBTorsionForce.
*/
#include "openmm/internal/AssertionUtilities.h"
#include "openmm/Context.h"
#include "CpuPlatform.h"
#include "openmm/RBTorsionForce.h"
#include "openmm/System.h"
#include "openmm/VerletIntegrator.h"
#include "SimTKOpenMMRealType.h"
#include <iostream>
#include <vector>
using namespace OpenMM;
using namespace std;
CpuPlatform platform;
const double TOL = 1e-5;
void testRBTorsions() {
System system;
system.addParticle(1.0);
system.addParticle(1.0);
system.addParticle(1.0);
system.addParticle(1.0);
VerletIntegrator integrator(0.01);
RBTorsionForce* forceField = new RBTorsionForce();
forceField->addTorsion(0, 1, 2, 3, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6);
system.addForce(forceField);
Context context(system, integrator, platform);
vector<Vec3> positions(4);
positions[0] = Vec3(0, 1, 0);
positions[1] = Vec3(0, 0, 0);
positions[2] = Vec3(1, 0, 0);
positions[3] = Vec3(1, 1, 1);
context.setPositions(positions);
State state = context.getState(State::Forces | State::Energy);
{
const vector<Vec3>& forces = state.getForces();
double psi = 0.25*PI_M - PI_M;
double torque = 0.0;
for (int i = 1; i < 6; ++i) {
double c = 0.1*(i+1);
torque += -c*i*std::pow(std::cos(psi), i-1)*std::sin(psi);
}
ASSERT_EQUAL_VEC(Vec3(0, 0, torque), forces[0], TOL);
ASSERT_EQUAL_VEC(Vec3(0, 0.5*torque, -0.5*torque), forces[3], TOL);
ASSERT_EQUAL_VEC(Vec3(forces[0][0]+forces[1][0]+forces[2][0]+forces[3][0], forces[0][1]+forces[1][1]+forces[2][1]+forces[3][1], forces[0][2]+forces[1][2]+forces[2][2]+forces[3][2]), Vec3(0, 0, 0), TOL);
double energy = 0.0;
for (int i = 0; i < 6; ++i) {
double c = 0.1*(i+1);
energy += c*std::pow(std::cos(psi), i);
}
ASSERT_EQUAL_TOL(energy, state.getPotentialEnergy(), TOL);
}
// Try changing the torsion parameters and make sure it's still correct.
forceField->setTorsionParameters(0, 0, 1, 2, 3, 0.11, 0.22, 0.33, 0.44, 0.55, 0.66);
forceField->updateParametersInContext(context);
state = context.getState(State::Forces | State::Energy);
{
const vector<Vec3>& forces = state.getForces();
double psi = 0.25*PI_M - PI_M;
double torque = 0.0;
for (int i = 1; i < 6; ++i) {
double c = 0.11*(i+1);
torque += -c*i*std::pow(std::cos(psi), i-1)*std::sin(psi);
}
ASSERT_EQUAL_VEC(Vec3(0, 0, torque), forces[0], TOL);
ASSERT_EQUAL_VEC(Vec3(0, 0.5*torque, -0.5*torque), forces[3], TOL);
ASSERT_EQUAL_VEC(Vec3(forces[0][0]+forces[1][0]+forces[2][0]+forces[3][0], forces[0][1]+forces[1][1]+forces[2][1]+forces[3][1], forces[0][2]+forces[1][2]+forces[2][2]+forces[3][2]), Vec3(0, 0, 0), TOL);
double energy = 0.0;
for (int i = 0; i < 6; ++i) {
double c = 0.11*(i+1);
energy += c*std::pow(std::cos(psi), i);
}
ASSERT_EQUAL_TOL(energy, state.getPotentialEnergy(), TOL);
}
}
#include "CpuTests.h"
#include "TestRBTorsionForce.h"
void testParallelComputation() {
System system;
......@@ -138,15 +58,6 @@ void testParallelComputation() {
ASSERT_EQUAL_VEC(state1.getForces()[i], state2.getForces()[i], 1e-5);
}
int main(int argc, char* argv[]) {
try {
testRBTorsions();
testParallelComputation();
}
catch(const exception& e) {
cout << "exception: " << e.what() << endl;
return 1;
}
cout << "Done" << endl;
return 0;
void runPlatformTests() {
testParallelComputation();
}
platforms/cpu/tests/TestCpuSettle.cpp
View file @ fd473eea
/* -------------------------------------------------------------------------- *
* OpenMM *
* -------------------------------------------------------------------------- *
......@@ -7,7 +6,7 @@
* Biological Structures at Stanford, funded under the NIH Roadmap for *
* Medical Research, grant U54 GM072970. See https://simtk.org. *
* *
* Portions copyright (c) 2008-2013 Stanford University and the Authors. *
* Portions copyright (c) 2015 Stanford University and the Authors. *
* Authors: Peter Eastman *
* Contributors: *
* *
......@@ -30,91 +29,8 @@
* USE OR OTHER DEALINGS IN THE SOFTWARE. *
* -------------------------------------------------------------------------- */
/**
* This tests the CPU implementation of the SETTLE algorithm.
*/
#include "openmm/internal/AssertionUtilities.h"
#include "openmm/Context.h"
#include "CpuPlatform.h"
#include "openmm/NonbondedForce.h"
#include "openmm/System.h"
#include "openmm/LangevinIntegrator.h"
#include "sfmt/SFMT.h"
#include <iostream>
#include <vector>
using namespace OpenMM;
using namespace std;
void testConstraints() {
const int numMolecules = 10;
const int numParticles = numMolecules*3;
const int numConstraints = numMolecules*3;
const double temp = 100.0;
CpuPlatform platform;
System system;
LangevinIntegrator integrator(temp, 2.0, 0.001);
integrator.setConstraintTolerance(1e-5);
NonbondedForce* forceField = new NonbondedForce();
for (int i = 0; i < numMolecules; ++i) {
system.addParticle(16.0);
system.addParticle(1.0);
system.addParticle(1.0);
forceField->addParticle(-0.82, 0.317, 0.65);
forceField->addParticle(0.41, 1.0, 0.0);
forceField->addParticle(0.41, 1.0, 0.0);
system.addConstraint(i*3, i*3+1, 0.1);
system.addConstraint(i*3, i*3+2, 0.1);
system.addConstraint(i*3+1, i*3+2, 0.163);
}
system.addForce(forceField);
Context context(system, integrator, platform);
vector<Vec3> positions(numParticles);
vector<Vec3> velocities(numParticles);
OpenMM_SFMT::SFMT sfmt;
init_gen_rand(0, sfmt);
for (int i = 0; i < numMolecules; ++i) {
positions[i*3] = Vec3((i%4)*0.4, (i/4)*0.4, 0);
positions[i*3+1] = positions[i*3]+Vec3(0.1, 0, 0);
positions[i*3+2] = positions[i*3]+Vec3(-0.03333, 0.09428, 0);
velocities[i*3] = Vec3(genrand_real2(sfmt)-0.5, genrand_real2(sfmt)-0.5, genrand_real2(sfmt)-0.5);
velocities[i*3+1] = Vec3(genrand_real2(sfmt)-0.5, genrand_real2(sfmt)-0.5, genrand_real2(sfmt)-0.5);
velocities[i*3+2] = Vec3(genrand_real2(sfmt)-0.5, genrand_real2(sfmt)-0.5, genrand_real2(sfmt)-0.5);
}
context.setPositions(positions);
context.setVelocities(velocities);
// Simulate it and see whether the constraints remain satisfied.
for (int i = 0; i < 1000; ++i) {
integrator.step(1);
State state = context.getState(State::Positions | State::Forces);
for (int j = 0; j < numConstraints; ++j) {
int particle1, particle2;
double distance;
system.getConstraintParameters(j, particle1, particle2, distance);
Vec3 p1 = state.getPositions()[particle1];
Vec3 p2 = state.getPositions()[particle2];
double dist = std::sqrt((p1[0]-p2[0])*(p1[0]-p2[0])+(p1[1]-p2[1])*(p1[1]-p2[1])+(p1[2]-p2[2])*(p1[2]-p2[2]));
ASSERT_EQUAL_TOL(distance, dist, 1e-5);
}
}
}
#include "CpuTests.h"
#include "TestSettle.h"
int main(int argc, char* argv[]) {
try {
if (!CpuPlatform::isProcessorSupported()) {
cout << "CPU is not supported. Exiting." << endl;
return 0;
}
testConstraints();
}
catch(const exception& e) {
cout << "exception: " << e.what() << endl;
return 1;
}
cout << "Done" << endl;
return 0;
void runPlatformTests() {
}
platforms/cuda/CMakeLists.txt
View file @ fd473eea
......@@ -13,9 +13,9 @@
#----------------------------------------------------
set(OPENMM_BUILD_CUDA_TESTS TRUE CACHE BOOL "Whether to build CUDA test cases")
if(OPENMM_BUILD_CUDA_TESTS)
if(BUILD_TESTING AND OPENMM_BUILD_CUDA_TESTS)
SUBDIRS (tests)
endif(OPENMM_BUILD_CUDA_TESTS)
endif(BUILD_TESTING AND OPENMM_BUILD_CUDA_TESTS)
# The source is organized into subdirectories, but we handle them all from
# this CMakeLists file rather than letting CMake visit them as SUBDIRS.
......
platforms/cuda/include/CudaContext.h
View file @ fd473eea
......@@ -30,6 +30,7 @@
#include <map>
#include <queue>
#include <string>
#include <utility>
#define __CL_ENABLE_EXCEPTIONS
#ifdef _MSC_VER
// Prevent Windows from defining macros that interfere with other code.
......@@ -538,6 +539,11 @@ public:
*/
void invalidateMolecules();
private:
/**
* Compute a sorted list of device indices in decreasing order of desirability
*/
std::vector<int> getDevicePrecedence();
struct Molecule;
struct MoleculeGroup;
class VirtualSiteInfo;
......
platforms/cuda/include/CudaExpressionUtilities.h
View file @ fd473eea
......@@ -9,7 +9,7 @@
* Biological Structures at Stanford, funded under the NIH Roadmap for *
* Medical Research, grant U54 GM072970. See https://simtk.org. *
* *
* Portions copyright (c) 2009-2014 Stanford University and the Authors. *
* Portions copyright (c) 2009-2015 Stanford University and the Authors. *
* Authors: Peter Eastman *
* Contributors: *
* *
......@@ -89,6 +89,10 @@ public:
* @param function the function for which to get a placeholder
*/
Lepton::CustomFunction* getFunctionPlaceholder(const TabulatedFunction& function);
/**
* Get a Lepton::CustomFunction that can be used to represent the periodicdistance() function when parsing expressions.
*/
Lepton::CustomFunction* getPeriodicDistancePlaceholder();
private:
class FunctionPlaceholder : public Lepton::CustomFunction {
public:
......@@ -114,13 +118,13 @@ private:
const std::vector<const TabulatedFunction*>& functions, const std::vector<std::pair<std::string, std::string> >& functionNames,
const std::string& prefix, const std::vector<std::vector<double> >& functionParams, const std::vector<Lepton::ParsedExpression>& allExpressions, const std::string& tempType);
std::string getTempName(const Lepton::ExpressionTreeNode& node, const std::vector<std::pair<Lepton::ExpressionTreeNode, std::string> >& temps);
void findRelatedTabulatedFunctions(const Lepton::ExpressionTreeNode& node, const Lepton::ExpressionTreeNode& searchNode,
void findRelatedCustomFunctions(const Lepton::ExpressionTreeNode& node, const Lepton::ExpressionTreeNode& searchNode,
std::vector<const Lepton::ExpressionTreeNode*>& nodes);
void findRelatedPowers(const Lepton::ExpressionTreeNode& node, const Lepton::ExpressionTreeNode& searchNode,
std::map<int, const Lepton::ExpressionTreeNode*>& powers);
std::vector<std::vector<double> > computeFunctionParameters(const std::vector<const TabulatedFunction*>& functions);
CudaContext& context;
FunctionPlaceholder fp1, fp2, fp3;
FunctionPlaceholder fp1, fp2, fp3, periodicDistance;
};
} // namespace OpenMM
......
platforms/cuda/include/CudaKernels.h
View file @ fd473eea
......@@ -620,6 +620,15 @@ public:
* @param force the NonbondedForce to copy the parameters from
*/
void copyParametersToContext(ContextImpl& context, const NonbondedForce& force);
/**
* Get the parameters being used for PME.
*
* @param alpha the separation parameter
* @param nx the number of grid points along the X axis
* @param ny the number of grid points along the Y axis
* @param nz the number of grid points along the Z axis
*/
void getPMEParameters(double& alpha, int& nx, int& ny, int& nz) const;
private:
class SortTrait : public CudaSort::SortTrait {
int getDataSize() const {return 8;}
......@@ -668,7 +677,9 @@ private:
std::vector<std::pair<int, int> > exceptionAtoms;
double ewaldSelfEnergy, dispersionCoefficient, alpha;
int interpolateForceThreads;
int gridSizeX, gridSizeY, gridSizeZ;
bool hasCoulomb, hasLJ, usePmeStream, useCudaFFT;
NonbondedMethod nonbondedMethod;
static const int PmeOrder = 5;
};
......@@ -922,6 +933,58 @@ private:
CUfunction donorKernel, acceptorKernel;
};
/**
* This kernel is invoked by CustomCentroidBondForce to calculate the forces acting on the system.
*/
class CudaCalcCustomCentroidBondForceKernel : public CalcCustomCentroidBondForceKernel {
public:
CudaCalcCustomCentroidBondForceKernel(std::string name, const Platform& platform, CudaContext& cu, const System& system) : CalcCustomCentroidBondForceKernel(name, platform),
cu(cu), params(NULL), globals(NULL), groupParticles(NULL), groupWeights(NULL), groupOffsets(NULL), groupForces(NULL), bondGroups(NULL), centerPositions(NULL), system(system) {
}
~CudaCalcCustomCentroidBondForceKernel();
/**
* Initialize the kernel.
*
* @param system the System this kernel will be applied to
* @param force the CustomCentroidBondForce this kernel will be used for
*/
void initialize(const System& system, const CustomCentroidBondForce& force);
/**
* Execute the kernel to calculate the forces and/or energy.
*
* @param context the context in which to execute this kernel
* @param includeForces true if forces should be calculated
* @param includeEnergy true if the energy should be calculated
* @return the potential energy due to the force
*/
double execute(ContextImpl& context, bool includeForces, bool includeEnergy);
/**
* Copy changed parameters over to a context.
*
* @param context the context to copy parameters to
* @param force the CustomCentroidBondForce to copy the parameters from
*/
void copyParametersToContext(ContextImpl& context, const CustomCentroidBondForce& force);
private:
int numGroups, numBonds;
CudaContext& cu;
CudaParameterSet* params;
CudaArray* globals;
CudaArray* groupParticles;
CudaArray* groupWeights;
CudaArray* groupOffsets;
CudaArray* groupForces;
CudaArray* bondGroups;
CudaArray* centerPositions;
std::vector<std::string> globalParamNames;
std::vector<float> globalParamValues;
std::vector<CudaArray*> tabulatedFunctions;
std::vector<void*> groupForcesArgs;
CUfunction computeCentersKernel, groupForcesKernel, applyForcesKernel;
const System& system;
};
/**
* This kernel is invoked by CustomCompoundBondForce to calculate the forces acting on the system.
*/
......
platforms/cuda/include/CudaNonbondedUtilities.h
View file @ fd473eea
......@@ -138,12 +138,18 @@ public:
void prepareInteractions(int forceGroups);
/**
* Compute the nonbonded interactions.
*
* @param forceGroups the flags specifying which force groups to include
* @param includeForces whether to compute forces
* @param includeEnergy whether to compute the potential energy
*/
void computeInteractions(int forceGroups);
void computeInteractions(int forceGroups, bool includeForces, bool includeEnergy);
/**
* Check to see if the neighbor list arrays are large enough, and make them bigger if necessary.
*
* @return true if the neighbor list needed to be enlarged.
*/
void updateNeighborListSize();
bool updateNeighborListSize();
/**
* Get the array containing the center of each atom block.
*/
......@@ -233,8 +239,10 @@ public:
* @param useExclusions specifies whether exclusions are applied to this interaction
* @param isSymmetric specifies whether the interaction is symmetric
* @param groups the set of force groups this kernel is for
* @param includeForces whether this kernel should compute forces
* @param includeEnergy whether this kernel should compute potential energy
*/
CUfunction createInteractionKernel(const std::string& source, std::vector<ParameterInfo>& params, std::vector<ParameterInfo>& arguments, bool useExclusions, bool isSymmetric, int groups);
CUfunction createInteractionKernel(const std::string& source, std::vector<ParameterInfo>& params, std::vector<ParameterInfo>& arguments, bool useExclusions, bool isSymmetric, int groups, bool includeForces, bool includeEnergy);
/**
* Create the set of kernels that will be needed for a particular combination of force groups.
*
......@@ -280,7 +288,8 @@ class CudaNonbondedUtilities::KernelSet {
public:
bool hasForces;
double cutoffDistance;
CUfunction forceKernel;
std::string source;
CUfunction forceKernel, energyKernel, forceEnergyKernel;
CUfunction findBlockBoundsKernel;
CUfunction sortBoxDataKernel;
CUfunction findInteractingBlocksKernel;
......
platforms/cuda/include/CudaParallelKernels.h
View file @ fd473eea
......@@ -430,6 +430,15 @@ public:
* @param force the NonbondedForce to copy the parameters from
*/
void copyParametersToContext(ContextImpl& context, const NonbondedForce& force);
/**
* Get the parameters being used for PME.
*
* @param alpha the separation parameter
* @param nx the number of grid points along the X axis
* @param ny the number of grid points along the Y axis
* @param nz the number of grid points along the Z axis
*/
void getPMEParameters(double& alpha, int& nx, int& ny, int& nz) const;
private:
class Task;
CudaPlatform::PlatformData& data;
......
platforms/cuda/sharedTarget/CMakeLists.txt
View file @ fd473eea
......@@ -18,7 +18,7 @@ SET_TARGET_PROPERTIES(${SHARED_TARGET} PROPERTIES COMPILE_FLAGS "${EXTRA_COMPILE
IF (APPLE)
SET_TARGET_PROPERTIES(${SHARED_TARGET} PROPERTIES LINK_FLAGS "${EXTRA_COMPILE_FLAGS} -F/Library/Frameworks -framework CUDA")
ELSE (APPLE)
SET_TARGET_PROPERTIES(${SHARED_TARGET} PROPERTIES LINK_FLAGS "${EXTRA_COMPILE_FLAGS}")
SET_TARGET_PROPERTIES(${SHARED_TARGET} PROPERTIES LINK_FLAGS "${EXTRA_LINK_FLAGS}")
ENDIF (APPLE)
INSTALL_TARGETS(/lib/plugins RUNTIME_DIRECTORY /lib/plugins ${SHARED_TARGET})
platforms/cuda/src/CudaBondedUtilities.cpp
View file @ fd473eea
......@@ -99,7 +99,7 @@ void CudaBondedUtilities::initialize(const System& system) {
s<<CudaKernelSources::vectorOps;
for (int i = 0; i < (int) prefixCode.size(); i++)
s<<prefixCode[i];
s<<"extern \"C\" __global__ void computeBondedForces(unsigned long long* __restrict__ forceBuffer, real* __restrict__ energyBuffer, const real4* __restrict__ posq, int groups";
s<<"extern \"C\" __global__ void computeBondedForces(unsigned long long* __restrict__ forceBuffer, mixed* __restrict__ energyBuffer, const real4* __restrict__ posq, int groups, real4 periodicBoxSize, real4 invPeriodicBoxSize, real4 periodicBoxVecX, real4 periodicBoxVecY, real4 periodicBoxVecZ";
for (int force = 0; force < numForces; force++) {
for (int i = 0; i < (int) atomIndices[force].size(); i++) {
int indexWidth = atomIndices[force][i]->getElementSize()/4;
......@@ -110,7 +110,7 @@ void CudaBondedUtilities::initialize(const System& system) {
for (int i = 0; i < (int) arguments.size(); i++)
s<<", "<<argTypes[i]<<"* customArg"<<(i+1);
s<<") {\n";
s<<"real energy = 0;\n";
s<<"mixed energy = 0;\n";
for (int force = 0; force < numForces; force++)
s<<createForceSource(force, forceAtoms[force].size(), forceAtoms[force][0].size(), forceGroup[force], forceSource[force]);
s<<"energyBuffer[blockIdx.x*blockDim.x+threadIdx.x] += energy;\n";
......@@ -161,6 +161,11 @@ void CudaBondedUtilities::computeInteractions(int groups) {
kernelArgs.push_back(&context.getEnergyBuffer().getDevicePointer());
kernelArgs.push_back(&context.getPosq().getDevicePointer());
kernelArgs.push_back(NULL);
kernelArgs.push_back(context.getPeriodicBoxSizePointer());
kernelArgs.push_back(context.getInvPeriodicBoxSizePointer());
kernelArgs.push_back(context.getPeriodicBoxVecXPointer());
kernelArgs.push_back(context.getPeriodicBoxVecYPointer());
kernelArgs.push_back(context.getPeriodicBoxVecZPointer());
for (int i = 0; i < (int) atomIndices.size(); i++)
for (int j = 0; j < (int) atomIndices[i].size(); j++)
kernelArgs.push_back(&atomIndices[i][j]->getDevicePointer());
......
platforms/cuda/src/CudaContext.cpp
View file @ fd473eea
......@@ -120,49 +120,50 @@ CudaContext::CudaContext(const System& system, int deviceIndex, bool useBlocking
int numDevices;
string errorMessage = "Error initializing Context";
CHECK_RESULT(cuDeviceGetCount(&numDevices));
if (deviceIndex < 0 || deviceIndex >= numDevices) {
// Try to figure out which device is the fastest.
int bestSpeed = -1;
int bestCompute = -1;
for (int i = 0; i < numDevices; i++) {
CHECK_RESULT(cuDeviceGet(&device, i));
int major, minor, clock, multiprocessors;
CHECK_RESULT(cuDeviceComputeCapability(&major, &minor, device));
if (major == 1 && minor < 2)
continue; // 1.0 and 1.1 are not supported
CHECK_RESULT(cuDeviceGetAttribute(&clock, CU_DEVICE_ATTRIBUTE_CLOCK_RATE, device));
CHECK_RESULT(cuDeviceGetAttribute(&multiprocessors, CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT, device));
int speed = clock*multiprocessors;
if (major > bestCompute || (major == bestCompute && speed > bestSpeed)) {
deviceIndex = i;
bestSpeed = speed;
bestCompute = major;
}
if (deviceIndex < -1 || deviceIndex >= numDevices)
throw OpenMMException("Illegal value for CudaDeviceIndex: "+intToString(deviceIndex));
vector<int> devicePrecedence;
if (deviceIndex == -1) {
devicePrecedence = getDevicePrecedence();
} else {
devicePrecedence.push_back(deviceIndex);
}
this->deviceIndex = -1;
for (int i = 0; i < static_cast<int>(devicePrecedence.size()); i++) {
int trialDeviceIndex = devicePrecedence[i];
CHECK_RESULT(cuDeviceGet(&device, trialDeviceIndex));
defaultOptimizationOptions = "--use_fast_math";
unsigned int flags = CU_CTX_MAP_HOST;
if (useBlockingSync)
flags += CU_CTX_SCHED_BLOCKING_SYNC;
else
flags += CU_CTX_SCHED_SPIN;
if (cuCtxCreate(&context, flags, device) == CUDA_SUCCESS) {
this->deviceIndex = trialDeviceIndex;
break;
}
}
if (deviceIndex == -1)
throw OpenMMException("No compatible CUDA device is available");
CHECK_RESULT(cuDeviceGet(&device, deviceIndex));
this->deviceIndex = deviceIndex;
if (this->deviceIndex == -1)
if (deviceIndex != -1)
throw OpenMMException("The requested CUDA device could not be loaded");
else
throw OpenMMException("No compatible CUDA device is available");
int major, minor;
CHECK_RESULT(cuDeviceComputeCapability(&major, &minor, device));
// This is a workaround to support GTX 980 with CUDA 6.5. It reports its compute capability
// as 5.2, but the compiler doesn't support anything beyond 5.0. We can remove this once
// CUDA 7.0 is released.
if (major == 5)
minor = 0;
#if __CUDA_API_VERSION < 7000
// This is a workaround to support GTX 980 with CUDA 6.5. It reports
// its compute capability as 5.2, but the compiler doesn't support
// anything beyond 5.0.
if (major == 5)
minor = 0;
#endif
gpuArchitecture = intToString(major)+intToString(minor);
computeCapability = major+0.1*minor;
if ((useDoublePrecision || useMixedPrecision) && computeCapability < 1.3)
throw OpenMMException("This device does not support double precision");
defaultOptimizationOptions = "--use_fast_math";
unsigned int flags = CU_CTX_MAP_HOST;
if (useBlockingSync)
flags += CU_CTX_SCHED_BLOCKING_SYNC;
else
flags += CU_CTX_SCHED_SPIN;
CHECK_RESULT(cuCtxCreate(&context, flags, device));
contextIsValid = true;
CHECK_RESULT(cuCtxSetCacheConfig(CU_FUNC_CACHE_PREFER_SHARED));
if (contextIndex > 0) {
......@@ -243,9 +244,9 @@ CudaContext::CudaContext(const System& system, int deviceIndex, bool useBlocking
compilationDefines["ATAN"] = useDoublePrecision ? "atan" : "atanf";
compilationDefines["ERF"] = useDoublePrecision ? "erf" : "erff";
compilationDefines["ERFC"] = useDoublePrecision ? "erfc" : "erfcf";
// Set defines for applying periodic boundary conditions.
Vec3 boxVectors[3];
system.getDefaultPeriodicBoxVectors(boxVectors[0], boxVectors[1], boxVectors[2]);
boxIsTriclinic = (boxVectors[0][1] != 0.0 || boxVectors[0][2] != 0.0 ||
......@@ -305,11 +306,11 @@ CudaContext::CudaContext(const System& system, int deviceIndex, bool useBlocking
}
// Create the work thread used for parallelization when running on multiple devices.
thread = new WorkThread();
// Create utilities objects.
bonded = new CudaBondedUtilities(*this);
nonbonded = new CudaNonbondedUtilities(*this);
integration = new CudaIntegrationUtilities(*this, system);
......@@ -366,7 +367,7 @@ void CudaContext::initialize() {
CHECK_RESULT(cuMemHostAlloc(&pinnedBuffer, pinnedBufferSize*sizeof(double), 0));
}
else if (useMixedPrecision) {
energyBuffer = CudaArray::create<float>(*this, numEnergyBuffers, "energyBuffer");
energyBuffer = CudaArray::create<double>(*this, numEnergyBuffers, "energyBuffer");
int pinnedBufferSize = max(paddedNumAtoms*4, numEnergyBuffers);
CHECK_RESULT(cuMemHostAlloc(&pinnedBuffer, pinnedBufferSize*sizeof(double), 0));
}
......@@ -425,7 +426,7 @@ string CudaContext::replaceStrings(const string& input, const std::map<std::stri
if (index != result.npos) {
if ((index == 0 || symbolChars.find(result[index-1]) == symbolChars.end()) && (index == result.size()-size || symbolChars.find(result[index+size]) == symbolChars.end())) {
// We have found a complete symbol, not part of a longer symbol.
result.replace(index, size, iter->second);
index += iter->second.size();
}
......@@ -460,11 +461,11 @@ static bool compileInWindows(const string &command) {
return -1;
}
WaitForSingleObject(pi.hProcess, INFINITE);
DWORD exitCode = -1;
DWORD exitCode = -1;
if(!GetExitCodeProcess(pi.hProcess, &exitCode)) {
throw(OpenMMException("Could not get nvcc.exe's exit code\n"));
} else {
if(exitCode == 0)
if(exitCode == 0)
return 0;
else
return -1;
......@@ -520,9 +521,9 @@ CUmodule CudaContext::createModule(const string source, const map<string, string
if (!defines.empty())
src << endl;
src << source << endl;
// See whether we already have PTX for this kernel cached.
CSHA1 sha1;
sha1.Update((const UINT_8*) src.str().c_str(), src.str().size());
sha1.Final();
......@@ -537,9 +538,9 @@ CUmodule CudaContext::createModule(const string source, const map<string, string
CUmodule module;
if (cuModuleLoad(&module, cacheFile.str().c_str()) == CUDA_SUCCESS)
return module;
// Select names for the various temporary files.
stringstream tempFileName;
tempFileName << "openmmTempKernel" << this; // Include a pointer to this context as part of the filename to avoid collisions.
#ifdef WIN32
......@@ -553,12 +554,12 @@ CUmodule CudaContext::createModule(const string source, const map<string, string
int res = 0;
// If the runtime compiler plugin is available, use it.
if (hasCompilerKernel) {
string ptx = compilerKernel.getAs<CudaCompilerKernel>().createModule(src.str(), "-arch=compute_"+gpuArchitecture+" "+options, *this);
// If possible, write the PTX out to a temporary file so we can cache it for later use.
bool wroteCache = false;
try {
ofstream out(outputFile.c_str());
......@@ -572,7 +573,7 @@ CUmodule CudaContext::createModule(const string source, const map<string, string
}
if (!wroteCache) {
// An error occurred. Possibly we don't have permission to write to the temp directory. Just try to load the module directly.
CHECK_RESULT2(cuModuleLoadDataEx(&module, &ptx[0], 0, NULL, NULL), "Error loading CUDA module");
return module;
}
......@@ -881,7 +882,7 @@ private:
void CudaContext::findMoleculeGroups() {
// The first time this is called, we need to identify all the molecules in the system.
if (moleculeGroups.size() == 0) {
// Add a ForceInfo that makes sure reordering doesn't break virtual sites.
......@@ -964,7 +965,7 @@ void CudaContext::findMoleculeGroups() {
if (!forces[k]->areParticlesIdentical(mol.atoms[i], mol2.atoms[i]))
identical = false;
}
// See if the constraints are identical.
for (int i = 0; i < (int) mol.constraints.size() && identical; i++) {
......@@ -1045,11 +1046,11 @@ void CudaContext::invalidateMolecules() {
}
if (valid)
return;
// The list of which molecules are identical is no longer valid. We need to restore the
// atoms to their original order, rebuild the list of identical molecules, and sort them
// again.
vector<int4> newCellOffsets(numAtoms);
if (useDoublePrecision) {
vector<double4> oldPosq(paddedNumAtoms);
......@@ -1194,6 +1195,8 @@ void CudaContext::reorderAtomsImpl() {
molPos[i].x *= invNumAtoms;
molPos[i].y *= invNumAtoms;
molPos[i].z *= invNumAtoms;
if (molPos[i].x != molPos[i].x)
throw OpenMMException("Particle coordinate is nan");
}
if (nonbonded->getUsePeriodic()) {
// Move each molecule position into the same box.
......@@ -1389,3 +1392,41 @@ void CudaContext::WorkThread::flush() {
pthread_cond_wait(&queueEmptyCondition, &queueLock);
pthread_mutex_unlock(&queueLock);
}
vector<int> CudaContext::getDevicePrecedence() {
int numDevices;
CUdevice thisDevice;
string errorMessage = "Error initializing Context";
vector<pair<pair<int, int>, int> > devices;
CHECK_RESULT(cuDeviceGetCount(&numDevices));
for (int i = 0; i < numDevices; i++) {
CHECK_RESULT(cuDeviceGet(&thisDevice, i));
int major, minor, clock, multiprocessors, speed;
CHECK_RESULT(cuDeviceComputeCapability(&major, &minor, thisDevice));
if (major == 1 && minor < 2)
continue;
if ((useDoublePrecision || useMixedPrecision) && (major+0.1*minor < 1.3))
continue;
CHECK_RESULT(cuDeviceGetAttribute(&clock, CU_DEVICE_ATTRIBUTE_CLOCK_RATE, thisDevice));
CHECK_RESULT(cuDeviceGetAttribute(&multiprocessors, CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT, thisDevice));
speed = clock*multiprocessors;
pair<int, int> deviceProperties = std::make_pair(major, speed);
devices.push_back(std::make_pair(deviceProperties, -i));
}
// sort first by compute capability (higher is better), then speed
// (higher is better), and finally device index (lower is better)
std::sort(devices.begin(), devices.end());
std::reverse(devices.begin(), devices.end());
vector<int> precedence;
for (int i = 0; i < static_cast<int>(devices.size()); i++) {
precedence.push_back(-devices[i].second);
}
return precedence;
}
platforms/cuda/src/CudaExpressionUtilities.cpp
View file @ fd473eea
......@@ -6,7 +6,7 @@
* Biological Structures at Stanford, funded under the NIH Roadmap for *
* Medical Research, grant U54 GM072970. See https://simtk.org. *
* *
* Portions copyright (c) 2009-2014 Stanford University and the Authors. *
* Portions copyright (c) 2009-2015 Stanford University and the Authors. *
* Authors: Peter Eastman *
* Contributors: *
* *
......@@ -33,7 +33,7 @@ using namespace OpenMM;
using namespace Lepton;
using namespace std;
CudaExpressionUtilities::CudaExpressionUtilities(CudaContext& context) : context(context), fp1(1), fp2(2), fp3(3) {
CudaExpressionUtilities::CudaExpressionUtilities(CudaContext& context) : context(context), fp1(1), fp2(2), fp3(3), periodicDistance(6) {
}
string CudaExpressionUtilities::createExpressions(const map<string, ParsedExpression>& expressions, const map<string, string>& variables,
......@@ -79,11 +79,6 @@ void CudaExpressionUtilities::processExpression(stringstream& out, const Express
throw OpenMMException("Unknown variable in expression: "+node.getOperation().getName());
case Operation::CUSTOM:
{
int i;
for (i = 0; i < (int) functionNames.size() && functionNames[i].first != node.getOperation().getName(); i++)
;
if (i == functionNames.size())
throw OpenMMException("Unknown function in expression: "+node.getOperation().getName());
out << "0.0f;\n";
temps.push_back(make_pair(node, name));
hasRecordedNode = true;
......@@ -93,7 +88,7 @@ void CudaExpressionUtilities::processExpression(stringstream& out, const Express
vector<const ExpressionTreeNode*> nodes;
for (int j = 0; j < (int) allExpressions.size(); j++)
findRelatedTabulatedFunctions(node, allExpressions[j].getRootNode(), nodes);
findRelatedCustomFunctions(node, allExpressions[j].getRootNode(), nodes);
vector<string> nodeNames;
nodeNames.push_back(name);
for (int j = 1; j < (int) nodes.size(); j++) {
......@@ -103,175 +98,223 @@ void CudaExpressionUtilities::processExpression(stringstream& out, const Express
temps.push_back(make_pair(*nodes[j], name2));
}
out << "{\n";
vector<string> paramsFloat, paramsInt;
for (int j = 0; j < (int) functionParams[i].size(); j++) {
paramsFloat.push_back(context.doubleToString(functionParams[i][j]));
paramsInt.push_back(context.intToString((int) functionParams[i][j]));
}
if (dynamic_cast<const Continuous1DFunction*>(functions[i]) != NULL) {
out << "real x = " << getTempName(node.getChildren()[0], temps) << ";\n";
out << "if (x >= " << paramsFloat[0] << " && x <= " << paramsFloat[1] << ") {\n";
out << "x = (x - " << paramsFloat[0] << ")*" << paramsFloat[2] << ";\n";
out << "int index = (int) (floor(x));\n";
out << "index = min(index, (int) " << paramsInt[3] << ");\n";
out << "float4 coeff = " << functionNames[i].second << "[index];\n";
out << "real b = x-index;\n";
out << "real a = 1.0f-b;\n";
for (int j = 0; j < nodes.size(); j++) {
const vector<int>& derivOrder = dynamic_cast<const Operation::Custom*>(&nodes[j]->getOperation())->getDerivOrder();
if (derivOrder[0] == 0)
out << nodeNames[j] << " = a*coeff.x+b*coeff.y+((a*a*a-a)*coeff.z+(b*b*b-b)*coeff.w)/(" << paramsFloat[2] << "*" << paramsFloat[2] << ");\n";
else
out << nodeNames[j] << " = (coeff.y-coeff.x)*" << paramsFloat[2] << "+((1.0f-3.0f*a*a)*coeff.z+(3.0f*b*b-1.0f)*coeff.w)/" << paramsFloat[2] << ";\n";
if (node.getOperation().getName() == "periodicdistance") {
// This is the periodicdistance() function.
out << tempType << "3 periodicDistance_delta = make_real3(";
for (int i = 0; i < 3; i++) {
if (i > 0)
out << ", ";
out << getTempName(node.getChildren()[i], temps) << "-" << getTempName(node.getChildren()[i+3], temps);
}
out << "}\n";
}
else if (dynamic_cast<const Continuous2DFunction*>(functions[i]) != NULL) {
out << "real x = " << getTempName(node.getChildren()[0], temps) << ";\n";
out << "real y = " << getTempName(node.getChildren()[1], temps) << ";\n";
out << "if (x >= " << paramsFloat[2] << " && x <= " << paramsFloat[3] << " && y >= " << paramsFloat[4] << " && y <= " << paramsFloat[5] << ") {\n";
out << "x = (x - " << paramsFloat[2] << ")*" << paramsFloat[6] << ";\n";
out << "y = (y - " << paramsFloat[4] << ")*" << paramsFloat[7] << ";\n";
out << "int s = min((int) floor(x), " << paramsInt[0] << ");\n";
out << "int t = min((int) floor(y), " << paramsInt[1] << ");\n";
out << "int coeffIndex = 4*(s+" << paramsInt[0] << "*t);\n";
out << "float4 c[4];\n";
for (int j = 0; j < 4; j++)
out << "c[" << j << "] = " << functionNames[i].second << "[coeffIndex+" << j << "];\n";
out << "real da = x-s;\n";
out << "real db = y-t;\n";
out << ");\n";
out << "APPLY_PERIODIC_TO_DELTA(periodicDistance_delta)\n";
out << tempType << " periodicDistance_r2 = periodicDistance_delta.x*periodicDistance_delta.x + periodicDistance_delta.y*periodicDistance_delta.y + periodicDistance_delta.z*periodicDistance_delta.z;\n";
out << tempType << " periodicDistance_rinv = RSQRT(periodicDistance_r2);\n";
for (int j = 0; j < nodes.size(); j++) {
const vector<int>& derivOrder = dynamic_cast<const Operation::Custom*>(&nodes[j]->getOperation())->getDerivOrder();
if (derivOrder[0] == 0 && derivOrder[1] == 0) {
out << nodeNames[j] << " = da*" << nodeNames[j] << " + ((c[3].w*db + c[3].z)*db + c[3].y)*db + c[3].x;\n";
out << nodeNames[j] << " = da*" << nodeNames[j] << " + ((c[2].w*db + c[2].z)*db + c[2].y)*db + c[2].x;\n";
out << nodeNames[j] << " = da*" << nodeNames[j] << " + ((c[1].w*db + c[1].z)*db + c[1].y)*db + c[1].x;\n";
out << nodeNames[j] << " = da*" << nodeNames[j] << " + ((c[0].w*db + c[0].z)*db + c[0].y)*db + c[0].x;\n";
}
else if (derivOrder[0] == 1 && derivOrder[1] == 0) {
out << nodeNames[j] << " = db*" << nodeNames[j] << " + (3.0f*c[3].w*da + 2.0f*c[2].w)*da + c[1].w;\n";
out << nodeNames[j] << " = db*" << nodeNames[j] << " + (3.0f*c[3].z*da + 2.0f*c[2].z)*da + c[1].z;\n";
out << nodeNames[j] << " = db*" << nodeNames[j] << " + (3.0f*c[3].y*da + 2.0f*c[2].y)*da + c[1].y;\n";
out << nodeNames[j] << " = db*" << nodeNames[j] << " + (3.0f*c[3].x*da + 2.0f*c[2].x)*da + c[1].x;\n";
out << nodeNames[j] << " *= " << paramsFloat[6] << ";\n";
}
else if (derivOrder[0] == 0 && derivOrder[1] == 1) {
out << nodeNames[j] << " = da*" << nodeNames[j] << " + (3.0f*c[3].w*db + 2.0f*c[3].z)*db + c[3].y;\n";
out << nodeNames[j] << " = da*" << nodeNames[j] << " + (3.0f*c[2].w*db + 2.0f*c[2].z)*db + c[2].y;\n";
out << nodeNames[j] << " = da*" << nodeNames[j] << " + (3.0f*c[1].w*db + 2.0f*c[1].z)*db + c[1].y;\n";
out << nodeNames[j] << " = da*" << nodeNames[j] << " + (3.0f*c[0].w*db + 2.0f*c[0].z)*db + c[0].y;\n";
out << nodeNames[j] << " *= " << paramsFloat[7] << ";\n";
int argIndex = -1;
for (int k = 0; k < 6; k++) {
if (derivOrder[k] > 0) {
if (derivOrder[k] > 1 || argIndex != -1)
throw OpenMMException("Unsupported derivative of periodicdistance"); // Should be impossible for this to happen.
argIndex = k;
}
}
else
throw OpenMMException("Unsupported derivative order for Continuous2DFunction");
if (argIndex == -1)
out << nodeNames[j] << " = RECIP(periodicDistance_rinv);\n";
else if (argIndex == 0)
out << nodeNames[j] << " = (periodicDistance_r2 > 0 ? periodicDistance_delta.x*periodicDistance_rinv : 0);\n";
else if (argIndex == 1)
out << nodeNames[j] << " = (periodicDistance_r2 > 0 ? periodicDistance_delta.y*periodicDistance_rinv : 0);\n";
else if (argIndex == 2)
out << nodeNames[j] << " = (periodicDistance_r2 > 0 ? periodicDistance_delta.z*periodicDistance_rinv : 0);\n";
else if (argIndex == 3)
out << nodeNames[j] << " = (periodicDistance_r2 > 0 ? -periodicDistance_delta.x*periodicDistance_rinv : 0);\n";
else if (argIndex == 4)
out << nodeNames[j] << " = (periodicDistance_r2 > 0 ? -periodicDistance_delta.y*periodicDistance_rinv : 0);\n";
else if (argIndex == 5)
out << nodeNames[j] << " = (periodicDistance_r2 > 0 ? -periodicDistance_delta.z*periodicDistance_rinv : 0);\n";
}
out << "}\n";
}
else if (dynamic_cast<const Continuous3DFunction*>(functions[i]) != NULL) {
out << "real x = " << getTempName(node.getChildren()[0], temps) << ";\n";
out << "real y = " << getTempName(node.getChildren()[1], temps) << ";\n";
out << "real z = " << getTempName(node.getChildren()[2], temps) << ";\n";
out << "if (x >= " << paramsFloat[3] << " && x <= " << paramsFloat[4] << " && y >= " << paramsFloat[5] << " && y <= " << paramsFloat[6] << " && z >= " << paramsFloat[7] << " && z <= " << paramsFloat[8] << ") {\n";
out << "x = (x - " << paramsFloat[3] << ")*" << paramsFloat[9] << ";\n";
out << "y = (y - " << paramsFloat[5] << ")*" << paramsFloat[10] << ";\n";
out << "z = (z - " << paramsFloat[7] << ")*" << paramsFloat[11] << ";\n";
out << "int s = min((int) floor(x), " << paramsInt[0] << ");\n";
out << "int t = min((int) floor(y), " << paramsInt[1] << ");\n";
out << "int u = min((int) floor(z), " << paramsInt[2] << ");\n";
out << "int coeffIndex = 16*(s+" << paramsInt[0] << "*(t+" << paramsInt[1] << "*u));\n";
out << "float4 c[16];\n";
for (int j = 0; j < 16; j++)
out << "c[" << j << "] = " << functionNames[i].second << "[coeffIndex+" << j << "];\n";
out << "real da = x-s;\n";
out << "real db = y-t;\n";
out << "real dc = z-u;\n";
for (int j = 0; j < nodes.size(); j++) {
const vector<int>& derivOrder = dynamic_cast<const Operation::Custom*>(&nodes[j]->getOperation())->getDerivOrder();
if (derivOrder[0] == 0 && derivOrder[1] == 0 && derivOrder[2] == 0) {
out << "real value[4] = {0, 0, 0, 0};\n";
for (int k = 3; k >= 0; k--)
for (int m = 0; m < 4; m++) {
int base = k + 4*m;
out << "value[" << m << "] = db*value[" << m << "] + ((c[" << base << "].w*da + c[" << base << "].z)*da + c[" << base << "].y)*da + c[" << base << "].x;\n";
}
out << nodeNames[j] << " = value[0] + dc*(value[1] + dc*(value[2] + dc*value[3]));\n";
}
else if (derivOrder[0] == 1 && derivOrder[1] == 0 && derivOrder[2] == 0) {
out << "real derivx[4] = {0, 0, 0, 0};\n";
for (int k = 3; k >= 0; k--)
for (int m = 0; m < 4; m++) {
int base = k + 4*m;
out << "derivx[" << m << "] = db*derivx[" << m << "] + (3*c[" << base << "].w*da + 2*c[" << base << "].z)*da + c[" << base << "].y;\n";
}
out << nodeNames[j] << " = derivx[0] + dc*(derivx[1] + dc*(derivx[2] + dc*derivx[3]));\n";
out << nodeNames[j] << " *= " << paramsFloat[9] << ";\n";
else {
// This is a tabulated function.
int i;
for (i = 0; i < (int) functionNames.size() && functionNames[i].first != node.getOperation().getName(); i++)
;
if (i == functionNames.size())
throw OpenMMException("Unknown function in expression: "+node.getOperation().getName());
vector<string> paramsFloat, paramsInt;
for (int j = 0; j < (int) functionParams[i].size(); j++) {
paramsFloat.push_back(context.doubleToString(functionParams[i][j]));
paramsInt.push_back(context.intToString((int) functionParams[i][j]));
}
if (dynamic_cast<const Continuous1DFunction*>(functions[i]) != NULL) {
out << "real x = " << getTempName(node.getChildren()[0], temps) << ";\n";
out << "if (x >= " << paramsFloat[0] << " && x <= " << paramsFloat[1] << ") {\n";
out << "x = (x - " << paramsFloat[0] << ")*" << paramsFloat[2] << ";\n";
out << "int index = (int) (floor(x));\n";
out << "index = min(index, (int) " << paramsInt[3] << ");\n";
out << "float4 coeff = " << functionNames[i].second << "[index];\n";
out << "real b = x-index;\n";
out << "real a = 1.0f-b;\n";
for (int j = 0; j < nodes.size(); j++) {
const vector<int>& derivOrder = dynamic_cast<const Operation::Custom*>(&nodes[j]->getOperation())->getDerivOrder();
if (derivOrder[0] == 0)
out << nodeNames[j] << " = a*coeff.x+b*coeff.y+((a*a*a-a)*coeff.z+(b*b*b-b)*coeff.w)/(" << paramsFloat[2] << "*" << paramsFloat[2] << ");\n";
else
out << nodeNames[j] << " = (coeff.y-coeff.x)*" << paramsFloat[2] << "+((1.0f-3.0f*a*a)*coeff.z+(3.0f*b*b-1.0f)*coeff.w)/" << paramsFloat[2] << ";\n";
}
else if (derivOrder[0] == 0 && derivOrder[1] == 1 && derivOrder[2] == 0) {
const string suffixes[] = {".x", ".y", ".z", ".w"};
out << "real derivy[4] = {0, 0, 0, 0};\n";
for (int k = 3; k >= 0; k--)
for (int m = 0; m < 4; m++) {
int base = 4*m;
string suffix = suffixes[m];
out << "derivy[" << m << "] = da*derivy[" << m << "] + (3*c[" << (base+3) << "]" << suffix << "*db + 2*c[" << (base+2) << "]" << suffix << ")*db + c[" << (base+1) << "]" << suffix << ";\n";
}
out << nodeNames[j] << " = derivy[0] + dc*(derivy[1] + dc*(derivy[2] + dc*derivy[3]));\n";
out << nodeNames[j] << " *= " << paramsFloat[10] << ";\n";
out << "}\n";
}
else if (dynamic_cast<const Continuous2DFunction*>(functions[i]) != NULL) {
out << "real x = " << getTempName(node.getChildren()[0], temps) << ";\n";
out << "real y = " << getTempName(node.getChildren()[1], temps) << ";\n";
out << "if (x >= " << paramsFloat[2] << " && x <= " << paramsFloat[3] << " && y >= " << paramsFloat[4] << " && y <= " << paramsFloat[5] << ") {\n";
out << "x = (x - " << paramsFloat[2] << ")*" << paramsFloat[6] << ";\n";
out << "y = (y - " << paramsFloat[4] << ")*" << paramsFloat[7] << ";\n";
out << "int s = min((int) floor(x), " << paramsInt[0] << ");\n";
out << "int t = min((int) floor(y), " << paramsInt[1] << ");\n";
out << "int coeffIndex = 4*(s+" << paramsInt[0] << "*t);\n";
out << "float4 c[4];\n";
for (int j = 0; j < 4; j++)
out << "c[" << j << "] = " << functionNames[i].second << "[coeffIndex+" << j << "];\n";
out << "real da = x-s;\n";
out << "real db = y-t;\n";
for (int j = 0; j < nodes.size(); j++) {
const vector<int>& derivOrder = dynamic_cast<const Operation::Custom*>(&nodes[j]->getOperation())->getDerivOrder();
if (derivOrder[0] == 0 && derivOrder[1] == 0) {
out << nodeNames[j] << " = da*" << nodeNames[j] << " + ((c[3].w*db + c[3].z)*db + c[3].y)*db + c[3].x;\n";
out << nodeNames[j] << " = da*" << nodeNames[j] << " + ((c[2].w*db + c[2].z)*db + c[2].y)*db + c[2].x;\n";
out << nodeNames[j] << " = da*" << nodeNames[j] << " + ((c[1].w*db + c[1].z)*db + c[1].y)*db + c[1].x;\n";
out << nodeNames[j] << " = da*" << nodeNames[j] << " + ((c[0].w*db + c[0].z)*db + c[0].y)*db + c[0].x;\n";
}
else if (derivOrder[0] == 1 && derivOrder[1] == 0) {
out << nodeNames[j] << " = db*" << nodeNames[j] << " + (3.0f*c[3].w*da + 2.0f*c[2].w)*da + c[1].w;\n";
out << nodeNames[j] << " = db*" << nodeNames[j] << " + (3.0f*c[3].z*da + 2.0f*c[2].z)*da + c[1].z;\n";
out << nodeNames[j] << " = db*" << nodeNames[j] << " + (3.0f*c[3].y*da + 2.0f*c[2].y)*da + c[1].y;\n";
out << nodeNames[j] << " = db*" << nodeNames[j] << " + (3.0f*c[3].x*da + 2.0f*c[2].x)*da + c[1].x;\n";
out << nodeNames[j] << " *= " << paramsFloat[6] << ";\n";
}
else if (derivOrder[0] == 0 && derivOrder[1] == 1) {
out << nodeNames[j] << " = da*" << nodeNames[j] << " + (3.0f*c[3].w*db + 2.0f*c[3].z)*db + c[3].y;\n";
out << nodeNames[j] << " = da*" << nodeNames[j] << " + (3.0f*c[2].w*db + 2.0f*c[2].z)*db + c[2].y;\n";
out << nodeNames[j] << " = da*" << nodeNames[j] << " + (3.0f*c[1].w*db + 2.0f*c[1].z)*db + c[1].y;\n";
out << nodeNames[j] << " = da*" << nodeNames[j] << " + (3.0f*c[0].w*db + 2.0f*c[0].z)*db + c[0].y;\n";
out << nodeNames[j] << " *= " << paramsFloat[7] << ";\n";
}
else
throw OpenMMException("Unsupported derivative order for Continuous2DFunction");
}
else if (derivOrder[0] == 0 && derivOrder[1] == 0 && derivOrder[2] == 1) {
out << "real derivz[4] = {0, 0, 0, 0};\n";
for (int k = 3; k >= 0; k--)
for (int m = 0; m < 4; m++) {
int base = k + 4*m;
out << "derivz[" << m << "] = db*derivz[" << m << "] + ((c[" << base << "].w*da + c[" << base << "].z)*da + c[" << base << "].y)*da + c[" << base << "].x;\n";
}
out << nodeNames[j] << " = derivz[1] + dc*(2*derivz[2] + dc*3*derivz[3]);\n";
out << nodeNames[j] << " *= " << paramsFloat[11] << ";\n";
out << "}\n";
}
else if (dynamic_cast<const Continuous3DFunction*>(functions[i]) != NULL) {
out << "real x = " << getTempName(node.getChildren()[0], temps) << ";\n";
out << "real y = " << getTempName(node.getChildren()[1], temps) << ";\n";
out << "real z = " << getTempName(node.getChildren()[2], temps) << ";\n";
out << "if (x >= " << paramsFloat[3] << " && x <= " << paramsFloat[4] << " && y >= " << paramsFloat[5] << " && y <= " << paramsFloat[6] << " && z >= " << paramsFloat[7] << " && z <= " << paramsFloat[8] << ") {\n";
out << "x = (x - " << paramsFloat[3] << ")*" << paramsFloat[9] << ";\n";
out << "y = (y - " << paramsFloat[5] << ")*" << paramsFloat[10] << ";\n";
out << "z = (z - " << paramsFloat[7] << ")*" << paramsFloat[11] << ";\n";
out << "int s = min((int) floor(x), " << paramsInt[0] << ");\n";
out << "int t = min((int) floor(y), " << paramsInt[1] << ");\n";
out << "int u = min((int) floor(z), " << paramsInt[2] << ");\n";
out << "int coeffIndex = 16*(s+" << paramsInt[0] << "*(t+" << paramsInt[1] << "*u));\n";
out << "float4 c[16];\n";
for (int j = 0; j < 16; j++)
out << "c[" << j << "] = " << functionNames[i].second << "[coeffIndex+" << j << "];\n";
out << "real da = x-s;\n";
out << "real db = y-t;\n";
out << "real dc = z-u;\n";
for (int j = 0; j < nodes.size(); j++) {
const vector<int>& derivOrder = dynamic_cast<const Operation::Custom*>(&nodes[j]->getOperation())->getDerivOrder();
if (derivOrder[0] == 0 && derivOrder[1] == 0 && derivOrder[2] == 0) {
out << "real value[4] = {0, 0, 0, 0};\n";
for (int k = 3; k >= 0; k--)
for (int m = 0; m < 4; m++) {
int base = k + 4*m;
out << "value[" << m << "] = db*value[" << m << "] + ((c[" << base << "].w*da + c[" << base << "].z)*da + c[" << base << "].y)*da + c[" << base << "].x;\n";
}
out << nodeNames[j] << " = value[0] + dc*(value[1] + dc*(value[2] + dc*value[3]));\n";
}
else if (derivOrder[0] == 1 && derivOrder[1] == 0 && derivOrder[2] == 0) {
out << "real derivx[4] = {0, 0, 0, 0};\n";
for (int k = 3; k >= 0; k--)
for (int m = 0; m < 4; m++) {
int base = k + 4*m;
out << "derivx[" << m << "] = db*derivx[" << m << "] + (3*c[" << base << "].w*da + 2*c[" << base << "].z)*da + c[" << base << "].y;\n";
}
out << nodeNames[j] << " = derivx[0] + dc*(derivx[1] + dc*(derivx[2] + dc*derivx[3]));\n";
out << nodeNames[j] << " *= " << paramsFloat[9] << ";\n";
}
else if (derivOrder[0] == 0 && derivOrder[1] == 1 && derivOrder[2] == 0) {
const string suffixes[] = {".x", ".y", ".z", ".w"};
out << "real derivy[4] = {0, 0, 0, 0};\n";
for (int k = 3; k >= 0; k--)
for (int m = 0; m < 4; m++) {
int base = 4*m;
string suffix = suffixes[m];
out << "derivy[" << m << "] = da*derivy[" << m << "] + (3*c[" << (base+3) << "]" << suffix << "*db + 2*c[" << (base+2) << "]" << suffix << ")*db + c[" << (base+1) << "]" << suffix << ";\n";
}
out << nodeNames[j] << " = derivy[0] + dc*(derivy[1] + dc*(derivy[2] + dc*derivy[3]));\n";
out << nodeNames[j] << " *= " << paramsFloat[10] << ";\n";
}
else if (derivOrder[0] == 0 && derivOrder[1] == 0 && derivOrder[2] == 1) {
out << "real derivz[4] = {0, 0, 0, 0};\n";
for (int k = 3; k >= 0; k--)
for (int m = 0; m < 4; m++) {
int base = k + 4*m;
out << "derivz[" << m << "] = db*derivz[" << m << "] + ((c[" << base << "].w*da + c[" << base << "].z)*da + c[" << base << "].y)*da + c[" << base << "].x;\n";
}
out << nodeNames[j] << " = derivz[1] + dc*(2*derivz[2] + dc*3*derivz[3]);\n";
out << nodeNames[j] << " *= " << paramsFloat[11] << ";\n";
}
else
throw OpenMMException("Unsupported derivative order for Continuous2DFunction");
}
else
throw OpenMMException("Unsupported derivative order for Continuous2DFunction");
out << "}\n";
}
out << "}\n";
}
else if (dynamic_cast<const Discrete1DFunction*>(functions[i]) != NULL) {
for (int j = 0; j < nodes.size(); j++) {
const vector<int>& derivOrder = dynamic_cast<const Operation::Custom*>(&nodes[j]->getOperation())->getDerivOrder();
if (derivOrder[0] == 0) {
out << "real x = " << getTempName(node.getChildren()[0], temps) << ";\n";
out << "if (x >= 0 && x < " << paramsInt[0] << ") {\n";
out << "int index = (int) floor(x+0.5f);\n";
out << nodeNames[j] << " = " << functionNames[i].second << "[index];\n";
out << "}\n";
else if (dynamic_cast<const Discrete1DFunction*>(functions[i]) != NULL) {
for (int j = 0; j < nodes.size(); j++) {
const vector<int>& derivOrder = dynamic_cast<const Operation::Custom*>(&nodes[j]->getOperation())->getDerivOrder();
if (derivOrder[0] == 0) {
out << "real x = " << getTempName(node.getChildren()[0], temps) << ";\n";
out << "if (x >= 0 && x < " << paramsInt[0] << ") {\n";
out << "int index = (int) floor(x+0.5f);\n";
out << nodeNames[j] << " = " << functionNames[i].second << "[index];\n";
out << "}\n";
}
}
}
}
else if (dynamic_cast<const Discrete2DFunction*>(functions[i]) != NULL) {
for (int j = 0; j < nodes.size(); j++) {
const vector<int>& derivOrder = dynamic_cast<const Operation::Custom*>(&nodes[j]->getOperation())->getDerivOrder();
if (derivOrder[0] == 0 && derivOrder[1] == 0) {
out << "int x = (int) floor(" << getTempName(node.getChildren()[0], temps) << "+0.5f);\n";
out << "int y = (int) floor(" << getTempName(node.getChildren()[1], temps) << "+0.5f);\n";
out << "int xsize = (int) " << paramsInt[0] << ";\n";
out << "int ysize = (int) " << paramsInt[1] << ";\n";
out << "int index = x+y*xsize;\n";
out << "if (index >= 0 && index < xsize*ysize)\n";
out << nodeNames[j] << " = " << functionNames[i].second << "[index];\n";
else if (dynamic_cast<const Discrete2DFunction*>(functions[i]) != NULL) {
for (int j = 0; j < nodes.size(); j++) {
const vector<int>& derivOrder = dynamic_cast<const Operation::Custom*>(&nodes[j]->getOperation())->getDerivOrder();
if (derivOrder[0] == 0 && derivOrder[1] == 0) {
out << "int x = (int) floor(" << getTempName(node.getChildren()[0], temps) << "+0.5f);\n";
out << "int y = (int) floor(" << getTempName(node.getChildren()[1], temps) << "+0.5f);\n";
out << "int xsize = (int) " << paramsInt[0] << ";\n";
out << "int ysize = (int) " << paramsInt[1] << ";\n";
out << "int index = x+y*xsize;\n";
out << "if (index >= 0 && index < xsize*ysize)\n";
out << nodeNames[j] << " = " << functionNames[i].second << "[index];\n";
}
}
}
}
else if (dynamic_cast<const Discrete3DFunction*>(functions[i]) != NULL) {
for (int j = 0; j < nodes.size(); j++) {
const vector<int>& derivOrder = dynamic_cast<const Operation::Custom*>(&nodes[j]->getOperation())->getDerivOrder();
if (derivOrder[0] == 0 && derivOrder[1] == 0 && derivOrder[2] == 0) {
out << "int x = (int) floor(" << getTempName(node.getChildren()[0], temps) << "+0.5f);\n";
out << "int y = (int) floor(" << getTempName(node.getChildren()[1], temps) << "+0.5f);\n";
out << "int z = (int) floor(" << getTempName(node.getChildren()[2], temps) << "+0.5f);\n";
out << "int xsize = (int) " << paramsInt[0] << ";\n";
out << "int ysize = (int) " << paramsInt[1] << ";\n";
out << "int zsize = (int) " << paramsInt[2] << ";\n";
out << "int index = x+(y+z*ysize)*xsize;\n";
out << "if (index >= 0 && index < xsize*ysize*zsize)\n";
out << nodeNames[j] << " = " << functionNames[i].second << "[index];\n";
else if (dynamic_cast<const Discrete3DFunction*>(functions[i]) != NULL) {
for (int j = 0; j < nodes.size(); j++) {
const vector<int>& derivOrder = dynamic_cast<const Operation::Custom*>(&nodes[j]->getOperation())->getDerivOrder();
if (derivOrder[0] == 0 && derivOrder[1] == 0 && derivOrder[2] == 0) {
out << "int x = (int) floor(" << getTempName(node.getChildren()[0], temps) << "+0.5f);\n";
out << "int y = (int) floor(" << getTempName(node.getChildren()[1], temps) << "+0.5f);\n";
out << "int z = (int) floor(" << getTempName(node.getChildren()[2], temps) << "+0.5f);\n";
out << "int xsize = (int) " << paramsInt[0] << ";\n";
out << "int ysize = (int) " << paramsInt[1] << ";\n";
out << "int zsize = (int) " << paramsInt[2] << ";\n";
out << "int index = x+(y+z*ysize)*xsize;\n";
out << "if (index >= 0 && index < xsize*ysize*zsize)\n";
out << nodeNames[j] << " = " << functionNames[i].second << "[index];\n";
}
}
}
}
......@@ -483,7 +526,7 @@ string CudaExpressionUtilities::getTempName(const ExpressionTreeNode& node, cons
throw OpenMMException(out.str());
}
void CudaExpressionUtilities::findRelatedTabulatedFunctions(const ExpressionTreeNode& node, const ExpressionTreeNode& searchNode,
void CudaExpressionUtilities::findRelatedCustomFunctions(const ExpressionTreeNode& node, const ExpressionTreeNode& searchNode,
vector<const Lepton::ExpressionTreeNode*>& nodes) {
if (searchNode.getOperation().getId() == Operation::CUSTOM && node.getOperation().getName() == searchNode.getOperation().getName()) {
// Make sure the arguments are identical.
......@@ -504,7 +547,7 @@ void CudaExpressionUtilities::findRelatedTabulatedFunctions(const ExpressionTree
}
else
for (int i = 0; i < (int) searchNode.getChildren().size(); i++)
findRelatedTabulatedFunctions(node, searchNode.getChildren()[i], nodes);
findRelatedCustomFunctions(node, searchNode.getChildren()[i], nodes);
}
void CudaExpressionUtilities::findRelatedPowers(const ExpressionTreeNode& node, const ExpressionTreeNode& searchNode, map<int, const ExpressionTreeNode*>& powers) {
......@@ -730,3 +773,7 @@ Lepton::CustomFunction* CudaExpressionUtilities::getFunctionPlaceholder(const Ta
return &fp3;
throw OpenMMException("getFunctionPlaceholder: Unknown function type");
}
Lepton::CustomFunction* CudaExpressionUtilities::getPeriodicDistancePlaceholder() {
return &periodicDistance;
}
platforms/cuda/src/CudaIntegrationUtilities.cpp
View file @ fd473eea
......@@ -201,15 +201,17 @@ CudaIntegrationUtilities::CudaIntegrationUtilities(CudaContext& context, const S
params.push_back(make_float2(dist13, dist12));
}
else
throw OpenMMException("Two of the three distances constrained with SETTLE must be the same.");
continue; // We can't handle this with SETTLE
isShakeAtom[atom1] = true;
isShakeAtom[atom2] = true;
isShakeAtom[atom3] = true;
}
settleAtoms = CudaArray::create<int4>(context, atoms.size(), "settleAtoms");
settleParams = CudaArray::create<float2>(context, params.size(), "settleParams");
settleAtoms->upload(atoms);
settleParams->upload(params);
if (atoms.size() > 0) {
settleAtoms = CudaArray::create<int4>(context, atoms.size(), "settleAtoms");
settleParams = CudaArray::create<float2>(context, params.size(), "settleParams");
settleAtoms->upload(atoms);
settleParams->upload(params);
}
}
// Find clusters consisting of a central atom with up to three peripheral atoms.
......
platforms/cuda/src/CudaKernelFactory.cpp
View file @ fd473eea
......@@ -104,6 +104,8 @@ KernelImpl* CudaKernelFactory::createKernelImpl(std::string name, const Platform
return new CudaCalcCustomExternalForceKernel(name, platform, cu, context.getSystem());
if (name == CalcCustomHbondForceKernel::Name())
return new CudaCalcCustomHbondForceKernel(name, platform, cu, context.getSystem());
if (name == CalcCustomCentroidBondForceKernel::Name())
return new CudaCalcCustomCentroidBondForceKernel(name, platform, cu, context.getSystem());
if (name == CalcCustomCompoundBondForceKernel::Name())
return new CudaCalcCustomCompoundBondForceKernel(name, platform, cu, context.getSystem());
if (name == CalcCustomManyParticleForceKernel::Name())
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment