Skip to content

GitLab

Commit 68f8aed8 authored by Peter's avatar Peter
Browse files

Merge branch 'pme' of https://github.com/peastman/openmm into pme

parents 84acc13d 5e0b0e3a
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 1310 additions and 120 deletions
cmake_modules/FindFFTW.cmake 0 → 100644
View file @ 68f8aed8
# - Find FFTW
# Find the native FFTW includes and library
#
# FFTW_INCLUDES - where to find fftw3.h
# FFTW_LIBRARY - the main FFTW library.
# FFTW_THREADS_LIBRARY - the FFTW multithreading support library.
# FFTW_FOUND - True if FFTW found.
if (FFTW_INCLUDES)
# Already in cache, be silent
set (FFTW_FIND_QUIETLY TRUE)
endif (FFTW_INCLUDES)
find_path (FFTW_INCLUDES fftw3.h)
find_library (FFTW_LIBRARY NAMES fftw3f)
find_library (FFTW_THREADS_LIBRARY NAMES fftw3f_threads)
# handle the QUIETLY and REQUIRED arguments and set FFTW_FOUND to TRUE if
# all listed variables are TRUE
include (FindPackageHandleStandardArgs)
find_package_handle_standard_args (FFTW DEFAULT_MSG FFTW_LIBRARY FFTW_INCLUDES)
find_package_handle_standard_args (FFTW_THREADS DEFAULT_MSG FFTW_THREADS_LIBRARY FFTW_INCLUDES)
mark_as_advanced (FFTW_LIBRARY FFTW_THREADS_LIBRARY FFTW_INCLUDES)
olla/include/openmm/kernels.h
View file @ 68f8aed8
......@@ -1152,6 +1152,70 @@ public:
virtual void execute(ContextImpl& context) = 0;
};
/**
* This kernel performs the reciprocal space calculation for PME. In most cases, this
* calculation is done directly by CalcNonbondedForceKernel so this kernel is unneeded.
* In some cases it may want to outsource the work to a different kernel. In particular,
* GPU based platforms sometimes use a CPU based implementation provided by a separate
* plugin.
*/
class CalcPmeReciprocalForceKernel : public KernelImpl {
public:
class IO;
static std::string Name() {
return "CalcPmeReciprocalForce";
}
CalcPmeReciprocalForceKernel(std::string name, const Platform& platform) : KernelImpl(name, platform) {
}
/**
* Initialize the kernel.
*
* @param gridx the x size of the PME grid
* @param gridy the y size of the PME grid
* @param gridz the z size of the PME grid
* @param numParticles the number of particles in the system
* @param alpha the Ewald blending parameter
*/
virtual void initialize(int gridx, int gridy, int gridz, int numParticles, double alpha) = 0;
/**
* Begin computing the force and energy.
*
* @param io an object that coordinates data transfer
* @param periodicBoxSize the size of the periodic box (measured in nm)
* @param includeEnergy true if potential energy should be computed
*/
virtual void beginComputation(IO& io, Vec3 periodicBoxSize, bool includeEnergy) = 0;
/**
* Finish computing the force and energy.
*
* @param io an object that coordinates data transfer
* @return the potential energy due to the PME reciprocal space interactions
*/
virtual double finishComputation(IO& io) = 0;
};
/**
* Any class that uses CalcPmeReciprocalForceKernel should create an implementation of this
* class, then pass it to the kernel to manage communication with it.
*/
class CalcPmeReciprocalForceKernel::IO {
public:
/**
* Get a pointer to the atom charges and positions. This array should contain four
* elements for each atom: x, y, z, and q in that order.
*/
virtual float* getPosq() = 0;
/**
* Record the forces calculated by the kernel.
*
* @param force an array containing four elements for each atom. The first three
* are the x, y, and z components of the force, while the fourth element
* should be ignored.
*/
virtual void setForce(float* force) = 0;
};
} // namespace OpenMM
#endif /*OPENMM_KERNELS_H_*/
platforms/cuda/include/CudaPlatform.h
View file @ 68f8aed8
......@@ -99,11 +99,12 @@ public:
class OPENMM_EXPORT_CUDA CudaPlatform::PlatformData {
public:
PlatformData(const System& system, const std::string& deviceIndexProperty, const std::string& blockingProperty, const std::string& precisionProperty,
PlatformData(ContextImpl* context, const System& system, const std::string& deviceIndexProperty, const std::string& blockingProperty, const std::string& precisionProperty,
const std::string& compilerProperty, const std::string& tempProperty);
~PlatformData();
void initializeContexts(const System& system);
void syncContexts();
ContextImpl* context;
std::vector<CudaContext*> contexts;
std::vector<double> contextEnergy;
bool removeCM, peerAccessSupported;
......
platforms/cuda/src/CudaContext.cpp
View file @ 68f8aed8
......@@ -231,6 +231,10 @@ CudaContext::~CudaContext() {
delete forces[i];
for (int i = 0; i < (int) reorderListeners.size(); i++)
delete reorderListeners[i];
for (int i = 0; i < (int) preComputations.size(); i++)
delete preComputations[i];
for (int i = 0; i < (int) postComputations.size(); i++)
delete postComputations[i];
if (pinnedBuffer != NULL)
cuMemFreeHost(pinnedBuffer);
if (posq != NULL)
......@@ -1102,6 +1106,14 @@ void CudaContext::addReorderListener(ReorderListener* listener) {
reorderListeners.push_back(listener);
}
void CudaContext::addPreComputation(ForcePreComputation* computation) {
preComputations.push_back(computation);
}
void CudaContext::addPostComputation(ForcePostComputation* computation) {
postComputations.push_back(computation);
}
struct CudaContext::WorkThread::ThreadData {
ThreadData(std::queue<CudaContext::WorkTask*>& tasks, bool& waiting, bool& finished,
pthread_mutex_t& queueLock, pthread_cond_t& waitForTaskCondition, pthread_cond_t& queueEmptyCondition) :
......
platforms/cuda/src/CudaContext.h
View file @ 68f8aed8
......@@ -70,6 +70,8 @@ public:
class WorkTask;
class WorkThread;
class ReorderListener;
class ForcePreComputation;
class ForcePostComputation;
static const int ThreadBlockSize;
static const int TileSize;
CudaContext(const System& system, int deviceIndex, bool useBlockingSync, const std::string& precision,
......@@ -454,6 +456,28 @@ public:
std::vector<ReorderListener*>& getReorderListeners() {
return reorderListeners;
}
/**
* Add a pre-computation that should be called at the very start of force and energy evalutations.
* The CudaContext assumes ownership of the object, and deletes it when the context itself is deleted.
*/
void addPreComputation(ForcePreComputation* computation);
/**
* Get the list of ForcePreComputations.
*/
std::vector<ForcePreComputation*>& getPreComputations() {
return preComputations;
}
/**
* Add a post-computation that should be called at the very end of force and energy evalutations.
* The CudaContext assumes ownership of the object, and deletes it when the context itself is deleted.
*/
void addPostComputation(ForcePostComputation* computation);
/**
* Get the list of ForcePostComputations.
*/
std::vector<ForcePostComputation*>& getPostComputations() {
return postComputations;
}
/**
* Mark that the current molecule definitions (and hence the atom order) may be invalid.
* This should be called whenever force field parameters change. It will cause the definitions
......@@ -519,6 +543,8 @@ private:
std::vector<CUdeviceptr> autoclearBuffers;
std::vector<int> autoclearBufferSizes;
std::vector<ReorderListener*> reorderListeners;
std::vector<ForcePreComputation*> preComputations;
std::vector<ForcePostComputation*> postComputations;
CudaIntegrationUtilities* integration;
CudaExpressionUtilities* expression;
CudaBondedUtilities* bonded;
......@@ -580,7 +606,7 @@ private:
/**
* This abstract class defines a function to be executed whenever atoms get reordered.
* Objects that need to know when reordering happens should create a reorderListener
* Objects that need to know when reordering happens should create a ReorderListener
* and register it by calling addReorderListener().
*/
class CudaContext::ReorderListener {
......@@ -590,6 +616,38 @@ public:
}
};
/**
* This abstract class defines a function to be executed at the very beginning of force and
* energy evaluation, before any other calculation has been done. It is useful for operations
* that need to be performed at a nonstandard point in the process. After creating a
* ForcePreComputation, register it by calling addForcePreComputation().
*/
class CudaContext::ForcePreComputation {
public:
/**
* @param includeForce true if forces should be computed
* @param includeEnergy true if potential energy should be computed
* @param groups a set of bit flags for which force groups to include
*/
virtual void computeForceAndEnergy(bool includeForces, bool includeEnergy, int groups) = 0;
};
/**
* This abstract class defines a function to be executed at the very end of force and
* energy evaluation, after all other calculations have been done. It is useful for operations
* that need to be performed at a nonstandard point in the process. After creating a
* ForcePostComputation, register it by calling addForcePostComputation().
*/
class CudaContext::ForcePostComputation {
public:
/**
* @param includeForce true if forces should be computed
* @param includeEnergy true if potential energy should be computed
* @param groups a set of bit flags for which force groups to include
* @return an optional contribution to add to the potential energy. */
virtual double computeForceAndEnergy(bool includeForces, bool includeEnergy, int groups) = 0;
};
} // namespace OpenMM
#endif /*OPENMM_CUDACONTEXT_H_*/
platforms/cuda/src/CudaKernels.cpp
View file @ 68f8aed8
......@@ -84,10 +84,12 @@ void CudaCalcForcesAndEnergyKernel::initialize(const System& system) {
void CudaCalcForcesAndEnergyKernel::beginComputation(ContextImpl& context, bool includeForces, bool includeEnergy, int groups) {
cu.setAsCurrent();
cu.clearAutoclearBuffers();
for (vector<CudaContext::ForcePreComputation*>::iterator iter = cu.getPreComputations().begin(); iter != cu.getPreComputations().end(); ++iter)
(*iter)->computeForceAndEnergy(includeForces, includeEnergy, groups);
CudaNonbondedUtilities& nb = cu.getNonbondedUtilities();
bool includeNonbonded = ((groups&(1<<nb.getForceGroup())) != 0);
cu.setComputeForceCount(cu.getComputeForceCount()+1);
cu.clearAutoclearBuffers();
if (includeNonbonded)
nb.prepareInteractions();
}
......@@ -96,8 +98,10 @@ double CudaCalcForcesAndEnergyKernel::finishComputation(ContextImpl& context, bo
cu.getBondedUtilities().computeInteractions(groups);
if ((groups&(1<<cu.getNonbondedUtilities().getForceGroup())) != 0)
cu.getNonbondedUtilities().computeInteractions();
cu.getIntegrationUtilities().distributeForcesFromVirtualSites();
double sum = 0.0;
for (vector<CudaContext::ForcePostComputation*>::iterator iter = cu.getPostComputations().begin(); iter != cu.getPostComputations().end(); ++iter)
sum += (*iter)->computeForceAndEnergy(includeForces, includeEnergy, groups);
cu.getIntegrationUtilities().distributeForcesFromVirtualSites();
if (includeEnergy) {
CudaArray& energyArray = cu.getEnergyBuffer();
if (cu.getUseDoublePrecision()) {
......@@ -1330,6 +1334,59 @@ private:
const NonbondedForce& force;
};
class CudaCalcNonbondedForceKernel::PmeIO : public CalcPmeReciprocalForceKernel::IO {
public:
PmeIO(CudaContext& cu, CUfunction addForcesKernel) : cu(cu), addForcesKernel(addForcesKernel), forceTemp(NULL) {
int elementSize = (cu.getUseDoublePrecision() ? sizeof(double4) : sizeof(float4));
forceTemp = new CudaArray(cu, cu.getNumAtoms(), elementSize, "PmeForce");
}
~PmeIO() {
if (forceTemp != NULL)
delete forceTemp;
}
float* getPosq() {
cu.setAsCurrent();
cu.getPosq().download(posq);
return (float*) &posq[0];
}
void setForce(float* force) {
forceTemp->upload(force);
void* args[] = {&forceTemp->getDevicePointer(), &cu.getForce().getDevicePointer()};
cu.executeKernel(addForcesKernel, args, cu.getNumAtoms());
}
private:
CudaContext& cu;
vector<float4> posq;
CudaArray* forceTemp;
CUfunction addForcesKernel;
};
class CudaCalcNonbondedForceKernel::PmePreComputation : public CudaContext::ForcePreComputation {
public:
PmePreComputation(CudaContext& cu, Kernel& pme, CalcPmeReciprocalForceKernel::IO& io) : cu(cu), pme(pme), io(io) {
}
void computeForceAndEnergy(bool includeForces, bool includeEnergy, int groups) {
Vec3 boxSize(cu.getPeriodicBoxSize().x, cu.getPeriodicBoxSize().y, cu.getPeriodicBoxSize().z);
pme.getAs<CalcPmeReciprocalForceKernel>().beginComputation(io, boxSize, includeEnergy);
}
private:
CudaContext& cu;
Kernel pme;
CalcPmeReciprocalForceKernel::IO& io;
};
class CudaCalcNonbondedForceKernel::PmePostComputation : public CudaContext::ForcePostComputation {
public:
PmePostComputation(Kernel& pme, CalcPmeReciprocalForceKernel::IO& io) : pme(pme), io(io) {
}
double computeForceAndEnergy(bool includeForces, bool includeEnergy, int groups) {
return pme.getAs<CalcPmeReciprocalForceKernel>().finishComputation(io);
}
private:
Kernel pme;
CalcPmeReciprocalForceKernel::IO& io;
};
CudaCalcNonbondedForceKernel::~CudaCalcNonbondedForceKernel() {
cu.setAsCurrent();
if (sigmaEpsilon != NULL)
......@@ -1354,6 +1411,8 @@ CudaCalcNonbondedForceKernel::~CudaCalcNonbondedForceKernel() {
delete pmeAtomGridIndex;
if (sort != NULL)
delete sort;
if (pmeio != NULL)
delete pmeio;
if (hasInitializedFFT) {
cufftDestroy(fftForward);
cufftDestroy(fftBackward);
......@@ -1457,7 +1516,7 @@ void CudaCalcNonbondedForceKernel::initialize(const System& system, const Nonbon
else
dispersionCoefficient = 0.0;
alpha = 0;
if (force.getNonbondedMethod() == NonbondedForce::Ewald) {
if (force.getNonbondedMethod() == NonbondedForce::Ewald && cu.getContextIndex() == 0) {
// Compute the Ewald parameters.
int kmaxx, kmaxy, kmaxz;
......@@ -1465,7 +1524,7 @@ void CudaCalcNonbondedForceKernel::initialize(const System& system, const Nonbon
defines["EWALD_ALPHA"] = cu.doubleToString(alpha);
defines["TWO_OVER_SQRT_PI"] = cu.doubleToString(2.0/sqrt(M_PI));
defines["USE_EWALD"] = "1";
ewaldSelfEnergy = (cu.getContextIndex() == 0 ? -ONE_4PI_EPS0*alpha*sumSquaredCharges/sqrt(M_PI) : 0.0);
ewaldSelfEnergy = -ONE_4PI_EPS0*alpha*sumSquaredCharges/sqrt(M_PI);
// Create the reciprocal space kernels.
......@@ -1484,7 +1543,7 @@ void CudaCalcNonbondedForceKernel::initialize(const System& system, const Nonbon
int elementSize = (cu.getUseDoublePrecision() ? sizeof(double2) : sizeof(float2));
cosSinSums = new CudaArray(cu, (2*kmaxx-1)*(2*kmaxy-1)*(2*kmaxz-1), elementSize, "cosSinSums");
}
else if (force.getNonbondedMethod() == NonbondedForce::PME) {
else if (force.getNonbondedMethod() == NonbondedForce::PME && cu.getContextIndex() == 0) {
// Compute the PME parameters.
int gridSizeX, gridSizeY, gridSizeZ;
......@@ -1497,7 +1556,7 @@ void CudaCalcNonbondedForceKernel::initialize(const System& system, const Nonbon
defines["EWALD_ALPHA"] = cu.doubleToString(alpha);
defines["TWO_OVER_SQRT_PI"] = cu.doubleToString(2.0/sqrt(M_PI));
defines["USE_EWALD"] = "1";
ewaldSelfEnergy = (cu.getContextIndex() == 0 ? -ONE_4PI_EPS0*alpha*sumSquaredCharges/sqrt(M_PI) : 0.0);
ewaldSelfEnergy = -ONE_4PI_EPS0*alpha*sumSquaredCharges/sqrt(M_PI);
pmeDefines["PME_ORDER"] = cu.intToString(PmeOrder);
pmeDefines["NUM_ATOMS"] = cu.intToString(numParticles);
pmeDefines["PADDED_NUM_ATOMS"] = cu.intToString(cu.getPaddedNumAtoms());
......@@ -1510,111 +1569,127 @@ void CudaCalcNonbondedForceKernel::initialize(const System& system, const Nonbon
if (cu.getUseDoublePrecision())
pmeDefines["USE_DOUBLE_PRECISION"] = "1";
CUmodule module = cu.createModule(CudaKernelSources::vectorOps+CudaKernelSources::pme, pmeDefines);
pmeGridIndexKernel = cu.getKernel(module, "findAtomGridIndex");
pmeSpreadChargeKernel = cu.getKernel(module, "gridSpreadCharge");
pmeConvolutionKernel = cu.getKernel(module, "reciprocalConvolution");
pmeInterpolateForceKernel = cu.getKernel(module, "gridInterpolateForce");
pmeEvalEnergyKernel = cu.getKernel(module, "gridEvaluateEnergy");
pmeFinishSpreadChargeKernel = cu.getKernel(module, "finishSpreadCharge");
cuFuncSetCacheConfig(pmeSpreadChargeKernel, CU_FUNC_CACHE_PREFER_L1);
cuFuncSetCacheConfig(pmeInterpolateForceKernel, CU_FUNC_CACHE_PREFER_L1);
// Create required data structures.
int elementSize = (cu.getUseDoublePrecision() ? sizeof(double) : sizeof(float));
directPmeGrid = new CudaArray(cu, gridSizeX*gridSizeY*gridSizeZ, cu.getComputeCapability() >= 2.0 ? elementSize : sizeof(long long), "originalPmeGrid");
reciprocalPmeGrid = new CudaArray(cu, gridSizeX*gridSizeY*(gridSizeZ/2+1), 2*elementSize, "reciprocalPmeGrid");
cu.addAutoclearBuffer(*directPmeGrid);
pmeBsplineModuliX = new CudaArray(cu, gridSizeX, elementSize, "pmeBsplineModuliX");
pmeBsplineModuliY = new CudaArray(cu, gridSizeY, elementSize, "pmeBsplineModuliY");
pmeBsplineModuliZ = new CudaArray(cu, gridSizeZ, elementSize, "pmeBsplineModuliZ");
pmeAtomRange = CudaArray::create<int>(cu, gridSizeX*gridSizeY*gridSizeZ+1, "pmeAtomRange");
pmeAtomGridIndex = CudaArray::create<int2>(cu, numParticles, "pmeAtomGridIndex");
sort = new CudaSort(cu, new SortTrait(), cu.getNumAtoms());
cufftResult result = cufftPlan3d(&fftForward, gridSizeX, gridSizeY, gridSizeZ, cu.getUseDoublePrecision() ? CUFFT_D2Z : CUFFT_R2C);
if (result != CUFFT_SUCCESS)
throw OpenMMException("Error initializing FFT: "+cu.intToString(result));
result = cufftPlan3d(&fftBackward, gridSizeX, gridSizeY, gridSizeZ, cu.getUseDoublePrecision() ? CUFFT_Z2D : CUFFT_C2R);
if (result != CUFFT_SUCCESS)
throw OpenMMException("Error initializing FFT: "+cu.intToString(result));
cufftSetCompatibilityMode(fftForward, CUFFT_COMPATIBILITY_NATIVE);
cufftSetCompatibilityMode(fftBackward, CUFFT_COMPATIBILITY_NATIVE);
bool useCpuPme = true;
if (useCpuPme) {
try {
cpuPme = getPlatform().createKernel(CalcPmeReciprocalForceKernel::Name(), *cu.getPlatformData().context);
cpuPme.getAs<CalcPmeReciprocalForceKernel>().initialize(gridSizeX, gridSizeY, gridSizeZ, numParticles, alpha);
CUfunction addForcesKernel = cu.getKernel(module, "addForces");
pmeio = new PmeIO(cu, addForcesKernel);
cu.addPreComputation(new PmePreComputation(cu, cpuPme, *pmeio));
cu.addPostComputation(new PmePostComputation(cpuPme, *pmeio));
}
catch (OpenMMException& ex) {
// The CPU PME plugin isn't available.
}
}
if (pmeio == NULL) {
pmeGridIndexKernel = cu.getKernel(module, "findAtomGridIndex");
pmeSpreadChargeKernel = cu.getKernel(module, "gridSpreadCharge");
pmeConvolutionKernel = cu.getKernel(module, "reciprocalConvolution");
pmeInterpolateForceKernel = cu.getKernel(module, "gridInterpolateForce");
pmeEvalEnergyKernel = cu.getKernel(module, "gridEvaluateEnergy");
pmeFinishSpreadChargeKernel = cu.getKernel(module, "finishSpreadCharge");
cuFuncSetCacheConfig(pmeSpreadChargeKernel, CU_FUNC_CACHE_PREFER_L1);
cuFuncSetCacheConfig(pmeInterpolateForceKernel, CU_FUNC_CACHE_PREFER_L1);
// Create required data structures.
int elementSize = (cu.getUseDoublePrecision() ? sizeof(double) : sizeof(float));
directPmeGrid = new CudaArray(cu, gridSizeX*gridSizeY*gridSizeZ, cu.getComputeCapability() >= 2.0 ? elementSize : sizeof(long long), "originalPmeGrid");
reciprocalPmeGrid = new CudaArray(cu, gridSizeX*gridSizeY*(gridSizeZ/2+1), 2*elementSize, "reciprocalPmeGrid");
cu.addAutoclearBuffer(*directPmeGrid);
pmeBsplineModuliX = new CudaArray(cu, gridSizeX, elementSize, "pmeBsplineModuliX");
pmeBsplineModuliY = new CudaArray(cu, gridSizeY, elementSize, "pmeBsplineModuliY");
pmeBsplineModuliZ = new CudaArray(cu, gridSizeZ, elementSize, "pmeBsplineModuliZ");
pmeAtomRange = CudaArray::create<int>(cu, gridSizeX*gridSizeY*gridSizeZ+1, "pmeAtomRange");
pmeAtomGridIndex = CudaArray::create<int2>(cu, numParticles, "pmeAtomGridIndex");
sort = new CudaSort(cu, new SortTrait(), cu.getNumAtoms());
cufftResult result = cufftPlan3d(&fftForward, gridSizeX, gridSizeY, gridSizeZ, cu.getUseDoublePrecision() ? CUFFT_D2Z : CUFFT_R2C);
if (result != CUFFT_SUCCESS)
throw OpenMMException("Error initializing FFT: "+cu.intToString(result));
result = cufftPlan3d(&fftBackward, gridSizeX, gridSizeY, gridSizeZ, cu.getUseDoublePrecision() ? CUFFT_Z2D : CUFFT_C2R);
if (result != CUFFT_SUCCESS)
throw OpenMMException("Error initializing FFT: "+cu.intToString(result));
cufftSetCompatibilityMode(fftForward, CUFFT_COMPATIBILITY_NATIVE);
cufftSetCompatibilityMode(fftBackward, CUFFT_COMPATIBILITY_NATIVE);
hasInitializedFFT = true;
// Initialize the b-spline moduli.
int maxSize = max(max(gridSizeX, gridSizeY), gridSizeZ);
vector<double> data(PmeOrder);
vector<double> ddata(PmeOrder);
vector<double> bsplines_data(maxSize);
data[PmeOrder-1] = 0.0;
data[1] = 0.0;
data[0] = 1.0;
for (int i = 3; i < PmeOrder; i++) {
double div = 1.0/(i-1.0);
data[i-1] = 0.0;
for (int j = 1; j < (i-1); j++)
data[i-j-1] = div*(j*data[i-j-2]+(i-j)*data[i-j-1]);
data[0] = div*data[0];
}
hasInitializedFFT = true;
// Differentiate.
// Initialize the b-spline moduli.
int maxSize = max(max(gridSizeX, gridSizeY), gridSizeZ);
vector<double> data(PmeOrder);
vector<double> ddata(PmeOrder);
vector<double> bsplines_data(maxSize);
data[PmeOrder-1] = 0.0;
data[1] = 0.0;
data[0] = 1.0;
for (int i = 3; i < PmeOrder; i++) {
double div = 1.0/(i-1.0);
data[i-1] = 0.0;
for (int j = 1; j < (i-1); j++)
data[i-j-1] = div*(j*data[i-j-2]+(i-j)*data[i-j-1]);
ddata[0] = -data[0];
for (int i = 1; i < PmeOrder; i++)
ddata[i] = data[i-1]-data[i];
double div = 1.0/(PmeOrder-1);
data[PmeOrder-1] = 0.0;
for (int i = 1; i < (PmeOrder-1); i++)
data[PmeOrder-i-1] = div*(i*data[PmeOrder-i-2]+(PmeOrder-i)*data[PmeOrder-i-1]);
data[0] = div*data[0];
}
// Differentiate.
ddata[0] = -data[0];
for (int i = 1; i < PmeOrder; i++)
ddata[i] = data[i-1]-data[i];
double div = 1.0/(PmeOrder-1);
data[PmeOrder-1] = 0.0;
for (int i = 1; i < (PmeOrder-1); i++)
data[PmeOrder-i-1] = div*(i*data[PmeOrder-i-2]+(PmeOrder-i)*data[PmeOrder-i-1]);
data[0] = div*data[0];
for (int i = 0; i < maxSize; i++)
bsplines_data[i] = 0.0;
for (int i = 1; i <= PmeOrder; i++)
bsplines_data[i] = data[i-1];
// Evaluate the actual bspline moduli for X/Y/Z.
for(int dim = 0; dim < 3; dim++) {
int ndata = (dim == 0 ? gridSizeX : dim == 1 ? gridSizeY : gridSizeZ);
vector<double> moduli(ndata);
for (int i = 0; i < ndata; i++) {
double sc = 0.0;
double ss = 0.0;
for (int j = 0; j < ndata; j++) {
double arg = (2.0*M_PI*i*j)/ndata;
sc += bsplines_data[j]*cos(arg);
ss += bsplines_data[j]*sin(arg);
for (int i = 0; i < maxSize; i++)
bsplines_data[i] = 0.0;
for (int i = 1; i <= PmeOrder; i++)
bsplines_data[i] = data[i-1];
// Evaluate the actual bspline moduli for X/Y/Z.
for(int dim = 0; dim < 3; dim++) {
int ndata = (dim == 0 ? gridSizeX : dim == 1 ? gridSizeY : gridSizeZ);
vector<double> moduli(ndata);
for (int i = 0; i < ndata; i++) {
double sc = 0.0;
double ss = 0.0;
for (int j = 0; j < ndata; j++) {
double arg = (2.0*M_PI*i*j)/ndata;
sc += bsplines_data[j]*cos(arg);
ss += bsplines_data[j]*sin(arg);
}
moduli[i] = sc*sc+ss*ss;
}
moduli[i] = sc*sc+ss*ss;
}
for (int i = 0; i < ndata; i++)
if (moduli[i] < 1.0e-7)
moduli[i] = (moduli[i-1]+moduli[i+1])*0.5;
if (cu.getUseDoublePrecision()) {
if (dim == 0)
pmeBsplineModuliX->upload(moduli);
else if (dim == 1)
pmeBsplineModuliY->upload(moduli);
else
pmeBsplineModuliZ->upload(moduli);
}
else {
vector<float> modulif(ndata);
for (int i = 0; i < ndata; i++)
modulif[i] = (float) moduli[i];
if (dim == 0)
pmeBsplineModuliX->upload(modulif);
else if (dim == 1)
pmeBsplineModuliY->upload(modulif);
else
pmeBsplineModuliZ->upload(modulif);
if (moduli[i] < 1.0e-7)
moduli[i] = (moduli[i-1]+moduli[i+1])*0.5;
if (cu.getUseDoublePrecision()) {
if (dim == 0)
pmeBsplineModuliX->upload(moduli);
else if (dim == 1)
pmeBsplineModuliY->upload(moduli);
else
pmeBsplineModuliZ->upload(moduli);
}
else {
vector<float> modulif(ndata);
for (int i = 0; i < ndata; i++)
modulif[i] = (float) moduli[i];
if (dim == 0)
pmeBsplineModuliX->upload(modulif);
else if (dim == 1)
pmeBsplineModuliY->upload(modulif);
else
pmeBsplineModuliZ->upload(modulif);
}
}
}
}
......@@ -1654,13 +1729,14 @@ void CudaCalcNonbondedForceKernel::initialize(const System& system, const Nonbon
}
double CudaCalcNonbondedForceKernel::execute(ContextImpl& context, bool includeForces, bool includeEnergy, bool includeDirect, bool includeReciprocal) {
if (cosSinSums != NULL && cu.getContextIndex() == 0 && includeReciprocal) {
double energy = (includeReciprocal ? ewaldSelfEnergy : 0.0);
if (cosSinSums != NULL && includeReciprocal) {
void* sumsArgs[] = {&cu.getEnergyBuffer().getDevicePointer(), &cu.getPosq().getDevicePointer(), &cosSinSums->getDevicePointer(), cu.getPeriodicBoxSizePointer()};
cu.executeKernel(ewaldSumsKernel, sumsArgs, cosSinSums->getSize());
void* forcesArgs[] = {&cu.getForce().getDevicePointer(), &cu.getPosq().getDevicePointer(), &cosSinSums->getDevicePointer(), cu.getPeriodicBoxSizePointer()};
cu.executeKernel(ewaldForcesKernel, forcesArgs, cu.getNumAtoms());
}
if (directPmeGrid != NULL && cu.getContextIndex() == 0 && includeReciprocal) {
if (directPmeGrid != NULL && includeReciprocal) {
void* gridIndexArgs[] = {&cu.getPosq().getDevicePointer(), &pmeAtomGridIndex->getDevicePointer(), cu.getPeriodicBoxSizePointer(), cu.getInvPeriodicBoxSizePointer()};
cu.executeKernel(pmeGridIndexKernel, gridIndexArgs, cu.getNumAtoms());
......@@ -1699,7 +1775,6 @@ double CudaCalcNonbondedForceKernel::execute(ContextImpl& context, bool includeF
cu.executeKernel(pmeInterpolateForceKernel, interpolateArgs, cu.getNumAtoms(), 128);
}
double energy = (includeReciprocal ? ewaldSelfEnergy : 0.0);
if (dispersionCoefficient != 0.0 && includeDirect) {
double4 boxSize = cu.getPeriodicBoxSize();
energy += dispersionCoefficient/(boxSize.x*boxSize.y*boxSize.z);
......
platforms/cuda/src/CudaKernels.h
View file @ 68f8aed8
......@@ -557,7 +557,7 @@ class CudaCalcNonbondedForceKernel : public CalcNonbondedForceKernel {
public:
CudaCalcNonbondedForceKernel(std::string name, const Platform& platform, CudaContext& cu, const System& system) : CalcNonbondedForceKernel(name, platform),
cu(cu), hasInitializedFFT(false), sigmaEpsilon(NULL), exceptionParams(NULL), cosSinSums(NULL), directPmeGrid(NULL), reciprocalPmeGrid(NULL),
pmeBsplineModuliX(NULL), pmeBsplineModuliY(NULL), pmeBsplineModuliZ(NULL), pmeAtomRange(NULL), pmeAtomGridIndex(NULL), sort(NULL) {
pmeBsplineModuliX(NULL), pmeBsplineModuliY(NULL), pmeBsplineModuliZ(NULL), pmeAtomRange(NULL), pmeAtomGridIndex(NULL), sort(NULL), pmeio(NULL) {
}
~CudaCalcNonbondedForceKernel();
/**
......@@ -596,6 +596,9 @@ private:
const char* getMaxValue() const {return "make_int2(INT_MAX, INT_MAX)";}
const char* getSortKey() const {return "value.y";}
};
class PmeIO;
class PmePreComputation;
class PmePostComputation;
CudaContext& cu;
bool hasInitializedFFT;
CudaArray* sigmaEpsilon;
......@@ -609,6 +612,8 @@ private:
CudaArray* pmeAtomRange;
CudaArray* pmeAtomGridIndex;
CudaSort* sort;
Kernel cpuPme;
PmeIO* pmeio;
cufftHandle fftForward;
cufftHandle fftBackward;
CUfunction ewaldSumsKernel;
......
platforms/cuda/src/CudaPlatform.cpp
View file @ 68f8aed8
......@@ -147,7 +147,7 @@ void CudaPlatform::contextCreated(ContextImpl& context, const map<string, string
getPropertyDefaultValue(CudaTempDirectory()) : properties.find(CudaTempDirectory())->second);
transform(blockingPropValue.begin(), blockingPropValue.end(), blockingPropValue.begin(), ::tolower);
transform(precisionPropValue.begin(), precisionPropValue.end(), precisionPropValue.begin(), ::tolower);
context.setPlatformData(new PlatformData(context.getSystem(), devicePropValue, blockingPropValue, precisionPropValue, compilerPropValue, tempPropValue));
context.setPlatformData(new PlatformData(&context, context.getSystem(), devicePropValue, blockingPropValue, precisionPropValue, compilerPropValue, tempPropValue));
}
void CudaPlatform::contextDestroyed(ContextImpl& context) const {
......@@ -155,8 +155,8 @@ void CudaPlatform::contextDestroyed(ContextImpl& context) const {
delete data;
}
CudaPlatform::PlatformData::PlatformData(const System& system, const string& deviceIndexProperty, const string& blockingProperty, const string& precisionProperty,
const string& compilerProperty, const string& tempProperty) : removeCM(false), stepCount(0), computeForceCount(0), time(0.0) {
CudaPlatform::PlatformData::PlatformData(ContextImpl* context, const System& system, const string& deviceIndexProperty, const string& blockingProperty, const string& precisionProperty,
const string& compilerProperty, const string& tempProperty) : context(context), removeCM(false), stepCount(0), computeForceCount(0), time(0.0) {
bool blocking = (blockingProperty == "true");
vector<string> devices;
size_t searchPos = 0, nextPos;
......
platforms/cuda/src/kernels/pme.cu
View file @ 68f8aed8
......@@ -266,8 +266,18 @@ void gridInterpolateForce(const real4* __restrict__ posq, unsigned long long* __
}
}
real q = pos.w*EPSILON_FACTOR;
forceBuffers[atom] += static_cast<unsigned long long>((long long) (-q*force.x*GRID_SIZE_X*invPeriodicBoxSize.x*0x100000000));
forceBuffers[atom+PADDED_NUM_ATOMS] += static_cast<unsigned long long>((long long) (-q*force.y*GRID_SIZE_Y*invPeriodicBoxSize.y*0x100000000));
forceBuffers[atom+2*PADDED_NUM_ATOMS] += static_cast<unsigned long long>((long long) (-q*force.z*GRID_SIZE_Z*invPeriodicBoxSize.z*0x100000000));
forceBuffers[atom] += static_cast<unsigned long long>((long long) (-q*force.x*GRID_SIZE_X*invPeriodicBoxSize.x*0x100000000));
forceBuffers[atom+PADDED_NUM_ATOMS] += static_cast<unsigned long long>((long long) (-q*force.y*GRID_SIZE_Y*invPeriodicBoxSize.y*0x100000000));
forceBuffers[atom+2*PADDED_NUM_ATOMS] += static_cast<unsigned long long>((long long) (-q*force.z*GRID_SIZE_Z*invPeriodicBoxSize.z*0x100000000));
}
}
extern "C" __global__
void addForces(const real4* __restrict__ forces, unsigned long long* __restrict__ forceBuffers) {
for (int atom = blockIdx.x*blockDim.x+threadIdx.x; atom < NUM_ATOMS; atom += blockDim.x*gridDim.x) {
real4 f = forces[atom];
forceBuffers[atom] += static_cast<unsigned long long>((long long) (f.x*0x100000000));
forceBuffers[atom+PADDED_NUM_ATOMS] += static_cast<unsigned long long>((long long) (f.y*0x100000000));
forceBuffers[atom+2*PADDED_NUM_ATOMS] += static_cast<unsigned long long>((long long) (f.z*0x100000000));
}
}
\ No newline at end of file
platforms/cuda/tests/TestCudaRandom.cpp
View file @ 68f8aed8
......@@ -54,7 +54,7 @@ void testGaussian() {
System system;
for (int i = 0; i < numAtoms; i++)
system.addParticle(1.0);
CudaPlatform::PlatformData platformData(system, "", "true", platform.getPropertyDefaultValue("CudaPrecision"),
CudaPlatform::PlatformData platformData(NULL, system, "", "true", platform.getPropertyDefaultValue("CudaPrecision"),
platform.getPropertyDefaultValue(CudaPlatform::CudaCompiler()), platform.getPropertyDefaultValue(CudaPlatform::CudaTempDirectory()));
CudaContext& context = *platformData.contexts[0];
context.initialize();
......
platforms/cuda/tests/TestCudaSort.cpp
View file @ 68f8aed8
......@@ -64,7 +64,7 @@ void verifySorting(vector<float> array) {
System system;
system.addParticle(0.0);
CudaPlatform::PlatformData platformData(system, "", "true", platform.getPropertyDefaultValue("CudaPrecision"),
CudaPlatform::PlatformData platformData(NULL, system, "", "true", platform.getPropertyDefaultValue("CudaPrecision"),
platform.getPropertyDefaultValue(CudaPlatform::CudaCompiler()), platform.getPropertyDefaultValue(CudaPlatform::CudaTempDirectory()));
CudaContext& context = *platformData.contexts[0];
context.initialize();
......
plugins/cpupme/CMakeLists.txt 0 → 100644
View file @ 68f8aed8
#---------------------------------------------------
# OpenMM CPU PME Plugin
#
# Creates plugin library, base name=OpenMMPME.
# Default libraries are shared & optimized.
#
# Windows:
# OpenMMPME[_d].dll
# OpenMMPME[_d].lib
# Unix:
# libOpenMMPME[_d].so
#----------------------------------------------------
IF (APPLE)
SET (CMAKE_OSX_DEPLOYMENT_TARGET "10.6")
ENDIF (APPLE)
# The source is organized into subdirectories, but we handle them all from
# this CMakeLists file rather than letting CMake visit them as SUBDIRS.
SET(OPENMM_SOURCE_SUBDIRS .)
# Collect up information about the version of the OpenMM library we're building
# and make it available to the code so it can be built into the binaries.
SET(OPENMMPME_LIBRARY_NAME OpenMMPME)
SET(SHARED_TARGET ${OPENMMPME_LIBRARY_NAME})
# Ensure that debug libraries have "_d" appended to their names.
# CMake gets this right on Windows automatically with this definition.
IF (${CMAKE_GENERATOR} MATCHES "Visual Studio")
SET(CMAKE_DEBUG_POSTFIX "_d" CACHE INTERNAL "" FORCE)
ENDIF (${CMAKE_GENERATOR} MATCHES "Visual Studio")
# But on Unix or Cygwin we have to add the suffix manually
IF (UNIX AND CMAKE_BUILD_TYPE MATCHES Debug)
SET(SHARED_TARGET ${SHARED_TARGET}_d)
SET(STATIC_TARGET ${STATIC_TARGET}_d)
ENDIF (UNIX AND CMAKE_BUILD_TYPE MATCHES Debug)
# These are all the places to search for header files which are
# to be part of the API.
SET(API_INCLUDE_DIRS) # start empty
FOREACH(subdir ${OPENMM_SOURCE_SUBDIRS})
# append
SET(API_INCLUDE_DIRS ${API_INCLUDE_DIRS}
${CMAKE_CURRENT_SOURCE_DIR}/${subdir}/include
${CMAKE_CURRENT_SOURCE_DIR}/${subdir}/include/internal)
ENDFOREACH(subdir)
# Find the include files.
SET(API_INCLUDE_FILES)
FOREACH(dir ${API_INCLUDE_DIRS})
FILE(GLOB fullpaths ${dir}/*.h) # returns full pathnames
SET(API_INCLUDE_FILES ${API_INCLUDE_FILES} ${fullpaths})
ENDFOREACH(dir)
# collect up source files
SET(SOURCE_FILES) # empty
SET(SOURCE_INCLUDE_FILES)
FOREACH(subdir ${OPENMM_SOURCE_SUBDIRS})
FILE(GLOB_RECURSE src_files ${CMAKE_CURRENT_SOURCE_DIR}/${subdir}/src/*.cpp ${CMAKE_CURRENT_SOURCE_DIR}/${subdir}/src/*.c)
FILE(GLOB incl_files ${CMAKE_CURRENT_SOURCE_DIR}/${subdir}/src/*.h)
SET(SOURCE_FILES ${SOURCE_FILES} ${src_files}) #append
SET(SOURCE_INCLUDE_FILES ${SOURCE_INCLUDE_FILES} ${incl_files})
INCLUDE_DIRECTORIES(BEFORE ${CMAKE_CURRENT_SOURCE_DIR}/${subdir}/include)
ENDFOREACH(subdir)
INCLUDE_DIRECTORIES(BEFORE ${CMAKE_CURRENT_SOURCE_DIR}/src)
# Include FFTW related files.
INCLUDE(FindFFTW)
INCLUDE_DIRECTORIES(${FFTW_INCLUDES})
# Build the plugin library.
ADD_LIBRARY(${SHARED_TARGET} SHARED ${SOURCE_FILES} ${SOURCE_INCLUDE_FILES} ${API_INCLUDE_FILES})
IF (UNIX AND CMAKE_BUILD_TYPE MATCHES Debug)
SET(MAIN_OPENMM_LIB ${OPENMM_LIBRARY_NAME}_d)
ELSE (UNIX AND CMAKE_BUILD_TYPE MATCHES Debug)
SET(MAIN_OPENMM_LIB ${OPENMM_LIBRARY_NAME})
ENDIF (UNIX AND CMAKE_BUILD_TYPE MATCHES Debug)
TARGET_LINK_LIBRARIES(${SHARED_TARGET} ${MAIN_OPENMM_LIB} ${PTHREADS_LIB} ${FFTW_LIBRARY} ${FFTW_THREADS_LIBRARY})
SET_TARGET_PROPERTIES(${SHARED_TARGET} PROPERTIES COMPILE_FLAGS "-DOPENMM_PME_BUILDING_SHARED_LIBRARY")
INSTALL_TARGETS(/lib/plugins RUNTIME_DIRECTORY /lib/plugins ${SHARED_TARGET})
plugins/cpupme/include/internal/windowsExportPme.h 0 → 100644
View file @ 68f8aed8
#ifndef OPENMM_WINDOWSEXPORTPME_H_
#define OPENMM_WINDOWSEXPORTPME_H_
/*
* Shared libraries are messy in Visual Studio. We have to distinguish three
* cases:
* (1) this header is being used to build the OpenMM shared library
* (dllexport)
* (2) this header is being used by a *client* of the OpenMM shared
* library (dllimport)
* (3) we are building the OpenMM static library, or the client is
* being compiled with the expectation of linking with the
* OpenMM static library (nothing special needed)
* In the CMake script for building this library, we define one of the symbols
* OPENMM_PME_BUILDING_{SHARED|STATIC}_LIBRARY
* Client code normally has no special symbol defined, in which case we'll
* assume it wants to use the shared library. However, if the client defines
* the symbol OPENMM_USE_STATIC_LIBRARIES we'll suppress the dllimport so
* that the client code can be linked with static libraries. Note that
* the client symbol is not library dependent, while the library symbols
* affect only the OpenMM library, meaning that other libraries can
* be clients of this one. However, we are assuming all-static or all-shared.
*/
#ifdef _MSC_VER
// We don't want to hear about how sprintf is "unsafe".
#pragma warning(disable:4996)
// Keep MS VC++ quiet about lack of dll export of private members.
#pragma warning(disable:4251)
#if defined(OPENMM_PME_BUILDING_SHARED_LIBRARY)
#define OPENMM_EXPORT_PME __declspec(dllexport)
#elif defined(OPENMM_PME_BUILDING_STATIC_LIBRARY) || defined(OPENMM_PME_USE_STATIC_LIBRARIES)
#define OPENMM_EXPORT_PME
#else
#define OPENMM_EXPORT_PME __declspec(dllimport) // i.e., a client of a shared library
#endif
#else
#define OPENMM_EXPORT_PME // Linux, Mac
#endif
#endif // OPENMM_WINDOWSEXPORTPME_H_
plugins/cpupme/src/CpuPmeKernelFactory.cpp 0 → 100644
View file @ 68f8aed8
/* -------------------------------------------------------------------------- *
* OpenMMAmoeba *
* -------------------------------------------------------------------------- *
* 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) 2013 Stanford University and the Authors. *
* Authors: Peter Eastman *
* Contributors: *
* *
* This program is free software: you can redistribute it and/or modify *
* it under the terms of the GNU Lesser General Public License as published *
* by the Free Software Foundation, either version 3 of the License, or *
* (at your option) any later version. *
* *
* This program is distributed in the hope that it will be useful, *
* but WITHOUT ANY WARRANTY; without even the implied warranty of *
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the *
* GNU Lesser General Public License for more details. *
* *
* You should have received a copy of the GNU Lesser General Public License *
* along with this program. If not, see <http://www.gnu.org/licenses/>. *
* -------------------------------------------------------------------------- */
#include "CpuPmeKernelFactory.h"
#include "CpuPmeKernels.h"
#include "internal/windowsExportPme.h"
#include "openmm/internal/ContextImpl.h"
#include "openmm/OpenMMException.h"
using namespace OpenMM;
extern "C" void registerPlatforms() {
}
extern "C" void registerKernelFactories() {
if (CpuCalcPmeReciprocalForceKernel::isProcessorSupported()) {
CpuPmeKernelFactory* factory = new CpuPmeKernelFactory();
for (int i = 0; i < Platform::getNumPlatforms(); i++)
Platform::getPlatform(i).registerKernelFactory(CalcPmeReciprocalForceKernel::Name(), factory);
}
}
extern "C" OPENMM_EXPORT_PME void registerCpuPmeKernelFactories() {
registerKernelFactories();
}
KernelImpl* CpuPmeKernelFactory::createKernelImpl(std::string name, const Platform& platform, ContextImpl& context) const {
if (name == CalcPmeReciprocalForceKernel::Name())
return new CpuCalcPmeReciprocalForceKernel(name, platform);
throw OpenMMException((std::string("Tried to create kernel with illegal kernel name '")+name+"'").c_str());
}
plugins/cpupme/src/CpuPmeKernelFactory.h 0 → 100644
View file @ 68f8aed8
#ifndef OPENMM_CPUPMEKERNELFACTORY_H_
#define OPENMM_CPUPMEKERNELFACTORY_H_
/* -------------------------------------------------------------------------- *
* 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) 2013 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 "openmm/KernelFactory.h"
namespace OpenMM {
/**
* This KernelFactory creates kernels for the CPU implementation of PME.
*/
class CpuPmeKernelFactory : public KernelFactory {
public:
KernelImpl* createKernelImpl(std::string name, const Platform& platform, ContextImpl& context) const;
};
} // namespace OpenMM
#endif /*OPENMM_CPUPMEKERNELFACTORY_H_*/
plugins/cpupme/src/CpuPmeKernels.cpp 0 → 100644
View file @ 68f8aed8
This diff is collapsed.
plugins/cpupme/src/CpuPmeKernels.h 0 → 100644
View file @ 68f8aed8
#ifndef OPENMM_CPU_PME_KERNELS_H_
#define OPENMM_CPU_PME_KERNELS_H_
/* -------------------------------------------------------------------------- *
* 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) 2013 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 "internal/windowsExportPme.h"
#include "openmm/kernels.h"
#include "openmm/Vec3.h"
#include <fftw3.h>
#include <pthread.h>
#include <vector>
namespace OpenMM {
/**
*/
class OPENMM_EXPORT_PME CpuCalcPmeReciprocalForceKernel : public CalcPmeReciprocalForceKernel {
public:
class ThreadData;
CpuCalcPmeReciprocalForceKernel(std::string name, const Platform& platform) : CalcPmeReciprocalForceKernel(name, platform),
hasCreatedPlan(false), isDeleted(false), realGrid(NULL), complexGrid(NULL) {
}
void initialize(int gridx, int gridy, int gridz, int numParticles, double alpha);
~CpuCalcPmeReciprocalForceKernel();
void beginComputation(IO& io, Vec3 periodicBoxSize, bool includeEnergy);
double finishComputation(IO& io);
void runThread(int index);
static bool isProcessorSupported();
private:
void threadWait();
void advanceThreads();
static bool hasInitializedThreads;
static int numThreads;
int gridx, gridy, gridz, numParticles;
double alpha;
bool hasCreatedPlan, isFinished, isDeleted;
std::vector<float> force;
std::vector<float> bsplineModuli[3];
float* realGrid;
fftwf_complex* complexGrid;
fftwf_plan forwardFFT, backwardFFT;
int waitCount;
pthread_cond_t startCondition, endCondition;
pthread_cond_t mainThreadStartCondition, mainThreadEndCondition;
pthread_mutex_t lock;
pthread_t mainThread;
std::vector<pthread_t> thread;
std::vector<ThreadData*> threadData;
// The following variables are used to store information about the calculation currently being performed.
IO* io;
float energy;
float* posq;
Vec3 periodicBoxSize;
bool includeEnergy;
};
} // namespace OpenMM
#endif /*OPENMM_CPU_PME_KERNELS_H_*/
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