Skip to content

GitLab

Commit 71d89f7f authored by Peter Eastman's avatar Peter Eastman
Browse files

Eliminated unnecessary copy of the PME grid

parent 5c2569b0
Show whitespace changes
Inline Side-by-side
Showing with 16 additions and 20 deletions
platforms/cuda/src/CudaKernels.cpp
View file @ 71d89f7f
...@@ -1340,12 +1340,10 @@ CudaCalcNonbondedForceKernel::~CudaCalcNonbondedForceKernel() { ...@@ -1340,12 +1340,10 @@ CudaCalcNonbondedForceKernel::~CudaCalcNonbondedForceKernel() {
delete exceptionParams; delete exceptionParams;
if (cosSinSums != NULL) if (cosSinSums != NULL)
delete cosSinSums; delete cosSinSums;
if (originalPmeGrid != NULL) if (directPmeGrid != NULL)
delete originalPmeGrid; delete directPmeGrid;
if (reciprocalPmeGrid != NULL) if (reciprocalPmeGrid != NULL)
delete reciprocalPmeGrid; delete reciprocalPmeGrid;
if (convolvedPmeGrid != NULL)
delete convolvedPmeGrid;
if (pmeBsplineModuliX != NULL) if (pmeBsplineModuliX != NULL)
delete pmeBsplineModuliX; delete pmeBsplineModuliX;
if (pmeBsplineModuliY != NULL) if (pmeBsplineModuliY != NULL)
...@@ -1521,11 +1519,10 @@ void CudaCalcNonbondedForceKernel::initialize(const System& system, const Nonbon ...@@ -1521,11 +1519,10 @@ void CudaCalcNonbondedForceKernel::initialize(const System& system, const Nonbon
int elementSize = (cu.getUseDoublePrecision() ? sizeof(double) : sizeof(float)); int elementSize = (cu.getUseDoublePrecision() ? sizeof(double) : sizeof(float));
originalPmeGrid = new CudaArray(cu, gridSizeX*gridSizeY*gridSizeZ, cu.getComputeCapability() >= 2.0 ? elementSize : sizeof(long long), "originalPmeGrid"); directPmeGrid = new CudaArray(cu, gridSizeX*gridSizeY*gridSizeZ, cu.getComputeCapability() >= 2.0 ? elementSize : sizeof(long long), "originalPmeGrid");
convolvedPmeGrid = new CudaArray(cu, gridSizeX*gridSizeY*gridSizeZ, elementSize, "convolvedPmeGrid");
reciprocalPmeGrid = new CudaArray(cu, gridSizeX*gridSizeY*(gridSizeZ/2+1), 2*elementSize, "reciprocalPmeGrid"); reciprocalPmeGrid = new CudaArray(cu, gridSizeX*gridSizeY*(gridSizeZ/2+1), 2*elementSize, "reciprocalPmeGrid");
cu.addAutoclearBuffer(*originalPmeGrid); cu.addAutoclearBuffer(*directPmeGrid);
pmeBsplineModuliX = new CudaArray(cu, gridSizeX, elementSize, "pmeBsplineModuliX"); pmeBsplineModuliX = new CudaArray(cu, gridSizeX, elementSize, "pmeBsplineModuliX");
pmeBsplineModuliY = new CudaArray(cu, gridSizeY, elementSize, "pmeBsplineModuliY"); pmeBsplineModuliY = new CudaArray(cu, gridSizeY, elementSize, "pmeBsplineModuliY");
...@@ -1660,7 +1657,7 @@ double CudaCalcNonbondedForceKernel::execute(ContextImpl& context, bool includeF ...@@ -1660,7 +1657,7 @@ double CudaCalcNonbondedForceKernel::execute(ContextImpl& context, bool includeF
void* forcesArgs[] = {&cu.getForce().getDevicePointer(), &cu.getPosq().getDevicePointer(), &cosSinSums->getDevicePointer(), cu.getPeriodicBoxSizePointer()}; void* forcesArgs[] = {&cu.getForce().getDevicePointer(), &cu.getPosq().getDevicePointer(), &cosSinSums->getDevicePointer(), cu.getPeriodicBoxSizePointer()};
cu.executeKernel(ewaldForcesKernel, forcesArgs, cu.getNumAtoms()); cu.executeKernel(ewaldForcesKernel, forcesArgs, cu.getNumAtoms());
} }
if (originalPmeGrid != NULL && cu.getContextIndex() == 0 && includeReciprocal) { if (directPmeGrid != NULL && cu.getContextIndex() == 0 && includeReciprocal) {
void* bsplinesArgs[] = {&cu.getPosq().getDevicePointer(), &pmeBsplineTheta->getDevicePointer(), &pmeAtomGridIndex->getDevicePointer(), void* bsplinesArgs[] = {&cu.getPosq().getDevicePointer(), &pmeBsplineTheta->getDevicePointer(), &pmeAtomGridIndex->getDevicePointer(),
cu.getPeriodicBoxSizePointer(), cu.getInvPeriodicBoxSizePointer()}; cu.getPeriodicBoxSizePointer(), cu.getInvPeriodicBoxSizePointer()};
int bsplinesSharedSize = cu.ThreadBlockSize*PmeOrder*(cu.getUseDoublePrecision() ? sizeof(double4) : sizeof(float4)); int bsplinesSharedSize = cu.ThreadBlockSize*PmeOrder*(cu.getUseDoublePrecision() ? sizeof(double4) : sizeof(float4));
...@@ -1672,19 +1669,19 @@ double CudaCalcNonbondedForceKernel::execute(ContextImpl& context, bool includeF ...@@ -1672,19 +1669,19 @@ double CudaCalcNonbondedForceKernel::execute(ContextImpl& context, bool includeF
cu.getPeriodicBoxSizePointer(), cu.getInvPeriodicBoxSizePointer()}; cu.getPeriodicBoxSizePointer(), cu.getInvPeriodicBoxSizePointer()};
cu.executeKernel(pmeAtomRangeKernel, rangeArgs, cu.getNumAtoms()); cu.executeKernel(pmeAtomRangeKernel, rangeArgs, cu.getNumAtoms());
void* spreadArgs[] = {&cu.getPosq().getDevicePointer(), &originalPmeGrid->getDevicePointer(), &pmeBsplineTheta->getDevicePointer(), cu.getPeriodicBoxSizePointer(), cu.getInvPeriodicBoxSizePointer()}; void* spreadArgs[] = {&cu.getPosq().getDevicePointer(), &directPmeGrid->getDevicePointer(), &pmeBsplineTheta->getDevicePointer(), cu.getPeriodicBoxSizePointer(), cu.getInvPeriodicBoxSizePointer()};
cu.executeKernel(pmeSpreadChargeKernel, spreadArgs, cu.getNumAtoms(), PmeOrder*PmeOrder*PmeOrder); cu.executeKernel(pmeSpreadChargeKernel, spreadArgs, cu.getNumAtoms(), PmeOrder*PmeOrder*PmeOrder);
void* finishSpreadArgs[] = {&originalPmeGrid->getDevicePointer()}; void* finishSpreadArgs[] = {&directPmeGrid->getDevicePointer()};
if (cu.getUseDoublePrecision() || cu.getComputeCapability() < 2.0) { if (cu.getUseDoublePrecision() || cu.getComputeCapability() < 2.0) {
void* finishSpreadArgs[] = {&originalPmeGrid->getDevicePointer()}; void* finishSpreadArgs[] = {&directPmeGrid->getDevicePointer()};
cu.executeKernel(pmeFinishSpreadChargeKernel, finishSpreadArgs, originalPmeGrid->getSize()); cu.executeKernel(pmeFinishSpreadChargeKernel, finishSpreadArgs, directPmeGrid->getSize());
} }
if (cu.getUseDoublePrecision()) if (cu.getUseDoublePrecision())
cufftExecD2Z(fftForward, (double*) originalPmeGrid->getDevicePointer(), (double2*) reciprocalPmeGrid->getDevicePointer()); cufftExecD2Z(fftForward, (double*) directPmeGrid->getDevicePointer(), (double2*) reciprocalPmeGrid->getDevicePointer());
else else
cufftExecR2C(fftForward, (float*) originalPmeGrid->getDevicePointer(), (float2*) reciprocalPmeGrid->getDevicePointer()); cufftExecR2C(fftForward, (float*) directPmeGrid->getDevicePointer(), (float2*) reciprocalPmeGrid->getDevicePointer());
if (includeEnergy) { if (includeEnergy) {
void* computeEnergyArgs[] = {&reciprocalPmeGrid->getDevicePointer(), &cu.getEnergyBuffer().getDevicePointer(), &pmeBsplineModuliX->getDevicePointer(), &pmeBsplineModuliY->getDevicePointer(), &pmeBsplineModuliZ->getDevicePointer(), cu.getPeriodicBoxSizePointer(), cu.getInvPeriodicBoxSizePointer()}; void* computeEnergyArgs[] = {&reciprocalPmeGrid->getDevicePointer(), &cu.getEnergyBuffer().getDevicePointer(), &pmeBsplineModuliX->getDevicePointer(), &pmeBsplineModuliY->getDevicePointer(), &pmeBsplineModuliZ->getDevicePointer(), cu.getPeriodicBoxSizePointer(), cu.getInvPeriodicBoxSizePointer()};
...@@ -1695,12 +1692,12 @@ double CudaCalcNonbondedForceKernel::execute(ContextImpl& context, bool includeF ...@@ -1695,12 +1692,12 @@ double CudaCalcNonbondedForceKernel::execute(ContextImpl& context, bool includeF
cu.executeKernel(pmeConvolutionKernel, convolutionArgs, cu.getNumAtoms()); cu.executeKernel(pmeConvolutionKernel, convolutionArgs, cu.getNumAtoms());
if (cu.getUseDoublePrecision()) if (cu.getUseDoublePrecision())
cufftExecZ2D(fftBackward, (double2*) reciprocalPmeGrid->getDevicePointer(), (double*) convolvedPmeGrid->getDevicePointer()); cufftExecZ2D(fftBackward, (double2*) reciprocalPmeGrid->getDevicePointer(), (double*) directPmeGrid->getDevicePointer());
else else
cufftExecC2R(fftBackward, (float2*) reciprocalPmeGrid->getDevicePointer(), (float*) convolvedPmeGrid->getDevicePointer()); cufftExecC2R(fftBackward, (float2*) reciprocalPmeGrid->getDevicePointer(), (float*) directPmeGrid->getDevicePointer());
void* interpolateArgs[] = {&cu.getPosq().getDevicePointer(), &cu.getForce().getDevicePointer(), &convolvedPmeGrid->getDevicePointer(), void* interpolateArgs[] = {&cu.getPosq().getDevicePointer(), &cu.getForce().getDevicePointer(), &directPmeGrid->getDevicePointer(),
cu.getPeriodicBoxSizePointer(), cu.getInvPeriodicBoxSizePointer()}; cu.getPeriodicBoxSizePointer(), cu.getInvPeriodicBoxSizePointer()};
cu.executeKernel(pmeInterpolateForceKernel, interpolateArgs, cu.getNumAtoms()); cu.executeKernel(pmeInterpolateForceKernel, interpolateArgs, cu.getNumAtoms());
......
platforms/cuda/src/CudaKernels.h
View file @ 71d89f7f
...@@ -556,7 +556,7 @@ private: ...@@ -556,7 +556,7 @@ private:
class CudaCalcNonbondedForceKernel : public CalcNonbondedForceKernel { class CudaCalcNonbondedForceKernel : public CalcNonbondedForceKernel {
public: public:
CudaCalcNonbondedForceKernel(std::string name, const Platform& platform, CudaContext& cu, System& system) : CalcNonbondedForceKernel(name, platform), CudaCalcNonbondedForceKernel(std::string name, const Platform& platform, CudaContext& cu, System& system) : CalcNonbondedForceKernel(name, platform),
cu(cu), hasInitializedFFT(false), sigmaEpsilon(NULL), exceptionParams(NULL), cosSinSums(NULL), originalPmeGrid(NULL), reciprocalPmeGrid(NULL), convolvedPmeGrid(NULL), cu(cu), hasInitializedFFT(false), sigmaEpsilon(NULL), exceptionParams(NULL), cosSinSums(NULL), directPmeGrid(NULL), reciprocalPmeGrid(NULL),
pmeBsplineModuliX(NULL), pmeBsplineModuliY(NULL), pmeBsplineModuliZ(NULL), pmeBsplineTheta(NULL), pmeBsplineDTheta(NULL), pmeBsplineModuliX(NULL), pmeBsplineModuliY(NULL), pmeBsplineModuliZ(NULL), pmeBsplineTheta(NULL), pmeBsplineDTheta(NULL),
pmeAtomRange(NULL), pmeAtomGridIndex(NULL), sort(NULL) { pmeAtomRange(NULL), pmeAtomGridIndex(NULL), sort(NULL) {
} }
...@@ -602,9 +602,8 @@ private: ...@@ -602,9 +602,8 @@ private:
CudaArray* sigmaEpsilon; CudaArray* sigmaEpsilon;
CudaArray* exceptionParams; CudaArray* exceptionParams;
CudaArray* cosSinSums; CudaArray* cosSinSums;
CudaArray* originalPmeGrid; CudaArray* directPmeGrid;
CudaArray* reciprocalPmeGrid; CudaArray* reciprocalPmeGrid;
CudaArray* convolvedPmeGrid;
CudaArray* pmeBsplineModuliX; CudaArray* pmeBsplineModuliX;
CudaArray* pmeBsplineModuliY; CudaArray* pmeBsplineModuliY;
CudaArray* pmeBsplineModuliZ; CudaArray* pmeBsplineModuliZ;
......
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