Skip to content

GitLab

Commit a163706b authored by Peter Eastman's avatar Peter Eastman
Browse files

cudaThreadExit() was not being called, which caused failures on some platforms

parent 030ef272
Hide whitespace changes
Inline Side-by-side
Showing with 76 additions and 62 deletions
platforms/cuda/src/kernels/gpu.cpp
View file @ a163706b
...@@ -1692,11 +1692,8 @@ void* gpuInit(int numAtoms, unsigned int device, bool useBlockingSync) ...@@ -1692,11 +1692,8 @@ void* gpuInit(int numAtoms, unsigned int device, bool useBlockingSync)
int SMMinor = 0; int SMMinor = 0;
// Select which device to use // Select which device to use
int currentDevice; cudaError_t status = cudaSetDevice(device);
cudaError_t status = cudaGetDevice(&currentDevice); RTERROR(status, "Error setting CUDA device")
RTERROR(status, "Error getting CUDA device")
if (device != currentDevice)
cudaSetDevice(device); // Ignore errors
status = cudaGetDevice(&gpu->device); status = cudaGetDevice(&gpu->device);
RTERROR(status, "Error getting CUDA device") RTERROR(status, "Error getting CUDA device")
status = cudaSetDeviceFlags(useBlockingSync ? cudaDeviceBlockingSync : cudaDeviceScheduleAuto); status = cudaSetDeviceFlags(useBlockingSync ? cudaDeviceBlockingSync : cudaDeviceScheduleAuto);
...@@ -1705,7 +1702,7 @@ void* gpuInit(int numAtoms, unsigned int device, bool useBlockingSync) ...@@ -1705,7 +1702,7 @@ void* gpuInit(int numAtoms, unsigned int device, bool useBlockingSync)
// Determine kernel call configuration // Determine kernel call configuration
cudaDeviceProp deviceProp; cudaDeviceProp deviceProp;
cudaGetDeviceProperties(&deviceProp, currentDevice); cudaGetDeviceProperties(&deviceProp, gpu->device);
// Determine SM version // Determine SM version
if (deviceProp.major == 1) if (deviceProp.major == 1)
...@@ -2089,6 +2086,7 @@ void gpuShutDown(gpuContext gpu) ...@@ -2089,6 +2086,7 @@ void gpuShutDown(gpuContext gpu)
// Wrap up // Wrap up
delete gpu; delete gpu;
cudaThreadExit();
return; return;
} }
......
platforms/cuda/tests/TestCudaCustomNonbondedForce.cpp
View file @ a163706b
...@@ -294,17 +294,27 @@ void testCoulombLennardJones() { ...@@ -294,17 +294,27 @@ void testCoulombLennardJones() {
customSystem.addForce(customNonbonded); customSystem.addForce(customNonbonded);
VerletIntegrator integrator1(0.01); VerletIntegrator integrator1(0.01);
VerletIntegrator integrator2(0.01); VerletIntegrator integrator2(0.01);
Context context1(standardSystem, integrator1, platform); double energy1, energy2;
context1.setPositions(positions); vector<Vec3> forces1, forces2;
context1.setVelocities(velocities); {
State state1 = context1.getState(State::Forces | State::Energy); Context context(standardSystem, integrator1, platform);
Context context2(customSystem, integrator2, platform); context.setPositions(positions);
context2.setPositions(positions); context.setVelocities(velocities);
context2.setVelocities(velocities); State state = context.getState(State::Forces | State::Energy);
State state2 = context2.getState(State::Forces | State::Energy); energy1 = state.getPotentialEnergy();
ASSERT_EQUAL_TOL(state1.getPotentialEnergy(), state2.getPotentialEnergy(), 1e-4); forces1 = state.getForces();
}
{
Context context(customSystem, integrator2, platform);
context.setPositions(positions);
context.setVelocities(velocities);
State state = context.getState(State::Forces | State::Energy);
energy2 = state.getPotentialEnergy();
forces2 = state.getForces();
}
ASSERT_EQUAL_TOL(energy1, energy2, 1e-4);
for (int i = 0; i < numParticles; i++) { for (int i = 0; i < numParticles; i++) {
ASSERT_EQUAL_VEC(state1.getForces()[i], state2.getForces()[i], 1e-4); ASSERT_EQUAL_VEC(forces1[i], forces2[i], 1e-4);
} }
} }
......
platforms/cuda/tests/TestCudaEwald.cpp
View file @ a163706b
...@@ -116,11 +116,11 @@ void testEwaldPME() { ...@@ -116,11 +116,11 @@ void testEwaldPME() {
Vec3 f = cudaState.getForces()[i]; Vec3 f = cudaState.getForces()[i];
positions[i] = Vec3(p[0]-f[0]*step, p[1]-f[1]*step, p[2]-f[2]*step); positions[i] = Vec3(p[0]-f[0]*step, p[1]-f[1]*step, p[2]-f[2]*step);
} }
Context cudaContext2(system, integrator, cuda); cudaContext.reinitialize();
cudaContext2.setPositions(positions); cudaContext.setPositions(positions);
tol = 1e-3; tol = 1e-3;
State cudaState2 = cudaContext2.getState(State::Energy); State cudaState2 = cudaContext.getState(State::Energy);
ASSERT_EQUAL_TOL(norm, (cudaState2.getPotentialEnergy()-cudaState.getPotentialEnergy())/delta, tol) ASSERT_EQUAL_TOL(norm, (cudaState2.getPotentialEnergy()-cudaState.getPotentialEnergy())/delta, tol)
// (3) Check whether the Reference and Cuda platforms agree when using PME // (3) Check whether the Reference and Cuda platforms agree when using PME
...@@ -154,11 +154,11 @@ void testEwaldPME() { ...@@ -154,11 +154,11 @@ void testEwaldPME() {
Vec3 f = cudaState.getForces()[i]; Vec3 f = cudaState.getForces()[i];
positions[i] = Vec3(p[0]-f[0]*step, p[1]-f[1]*step, p[2]-f[2]*step); positions[i] = Vec3(p[0]-f[0]*step, p[1]-f[1]*step, p[2]-f[2]*step);
} }
Context cudaContext3(system, integrator, cuda); cudaContext.reinitialize();
cudaContext3.setPositions(positions); cudaContext.setPositions(positions);
tol = 1e-3; tol = 1e-3;
State cudaState3 = cudaContext3.getState(State::Energy); State cudaState3 = cudaContext.getState(State::Energy);
ASSERT_EQUAL_TOL(norm, (cudaState3.getPotentialEnergy()-cudaState.getPotentialEnergy())/delta, tol) ASSERT_EQUAL_TOL(norm, (cudaState3.getPotentialEnergy()-cudaState.getPotentialEnergy())/delta, tol)
} }
......
platforms/cuda/tests/TestCudaNonbondedForce.cpp
View file @ a163706b
...@@ -131,7 +131,7 @@ void testExclusionsAnd14() { ...@@ -131,7 +131,7 @@ void testExclusionsAnd14() {
for (int i = 1; i < 5; ++i) { for (int i = 1; i < 5; ++i) {
// Test LJ forces // Test LJ forces
vector<Vec3> positions(5); vector<Vec3> positions(5);
const double r = 1.0; const double r = 1.0;
for (int j = 0; j < 5; ++j) { for (int j = 0; j < 5; ++j) {
...@@ -143,49 +143,55 @@ void testExclusionsAnd14() { ...@@ -143,49 +143,55 @@ void testExclusionsAnd14() {
nonbonded->setExceptionParameters(first14, 0, 3, 0, 1.5, i == 3 ? 0.5 : 0.0); nonbonded->setExceptionParameters(first14, 0, 3, 0, 1.5, i == 3 ? 0.5 : 0.0);
nonbonded->setExceptionParameters(second14, 1, 4, 0, 1.5, 0.0); nonbonded->setExceptionParameters(second14, 1, 4, 0, 1.5, 0.0);
positions[i] = Vec3(r, 0, 0); positions[i] = Vec3(r, 0, 0);
Context context(system, integrator, platform); // The following is in its own block, because CUDA can't deal with multiple Contexts
context.setPositions(positions); // existing on the same thread at the same time.
State state = context.getState(State::Forces | State::Energy); {
const vector<Vec3>& forces = state.getForces(); Context context(system, integrator, platform);
double x = 1.5/r; context.setPositions(positions);
double eps = 1.0; State state = context.getState(State::Forces | State::Energy);
double force = 4.0*eps*(12*std::pow(x, 12.0)-6*std::pow(x, 6.0))/r; const vector<Vec3>& forces = state.getForces();
double energy = 4.0*eps*(std::pow(x, 12.0)-std::pow(x, 6.0)); double x = 1.5/r;
if (i == 3) { double eps = 1.0;
force *= 0.5; double force = 4.0*eps*(12*std::pow(x, 12.0)-6*std::pow(x, 6.0))/r;
energy *= 0.5; double energy = 4.0*eps*(std::pow(x, 12.0)-std::pow(x, 6.0));
} if (i == 3) {
if (i < 3) { force *= 0.5;
force = 0; energy *= 0.5;
energy = 0; }
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);
} }
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 // Test Coulomb forces
nonbonded->setParticleParameters(0, 2, 1.5, 0); {
nonbonded->setParticleParameters(i, 2, 1.5, 0); nonbonded->setParticleParameters(0, 2, 1.5, 0);
nonbonded->setExceptionParameters(first14, 0, 3, i == 3 ? 4/1.2 : 0, 1.5, 0); nonbonded->setParticleParameters(i, 2, 1.5, 0);
nonbonded->setExceptionParameters(second14, 1, 4, 0, 1.5, 0); nonbonded->setExceptionParameters(first14, 0, 3, i == 3 ? 4/1.2 : 0, 1.5, 0);
Context context2(system, integrator, platform); nonbonded->setExceptionParameters(second14, 1, 4, 0, 1.5, 0);
context2.setPositions(positions); Context context(system, integrator, platform);
state = context2.getState(State::Forces | State::Energy); context.setPositions(positions);
const vector<Vec3>& forces2 = state.getForces(); State state = context.getState(State::Forces | State::Energy);
force = ONE_4PI_EPS0*4/(r*r); const vector<Vec3>& forces2 = state.getForces();
energy = ONE_4PI_EPS0*4/r; double force = ONE_4PI_EPS0*4/(r*r);
if (i == 3) { double energy = ONE_4PI_EPS0*4/r;
force /= 1.2; if (i == 3) {
energy /= 1.2; force /= 1.2;
} energy /= 1.2;
if (i < 3) { }
force = 0; if (i < 3) {
energy = 0; 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);
} }
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);
} }
} }
......
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