More bug fixes to multi-GPU

platforms/cuda/src/CudaNonbondedUtilities.cpp

@@ -264,14 +264,6 @@ void CudaNonbondedUtilities::initialize(const System& system) {
         sortedBlockCenter = new CudaArray(context, numAtomBlocks+1, 4*elementSize, "sortedBlockCenter");
         sortedBlockBoundingBox = new CudaArray(context, numAtomBlocks+1, 4*elementSize, "sortedBlockBoundingBox");
         oldPositions = new CudaArray(context, numAtoms, 4*elementSize, "oldPositions");
-        if (context.getUseDoublePrecision()) {
-            vector<double4> oldPositionsVec(numAtoms, make_double4(1e30, 1e30, 1e30, 0));
-            oldPositions->upload(oldPositionsVec);
-        }
-        else {
-            vector<float4> oldPositionsVec(numAtoms, make_float4(1e30f, 1e30f, 1e30f, 0));
-            oldPositions->upload(oldPositionsVec);
-        }
         rebuildNeighborList = CudaArray::create<int>(context, 1, "rebuildNeighborList");
         blockSorter = new CudaSort(context, new BlockSortTrait(context.getUseDoublePrecision()), numAtomBlocks);
         vector<unsigned int> count(1, 0);
@@ -402,14 +394,7 @@ void CudaNonbondedUtilities::updateNeighborListSize() {
     if (forceArgs.size() > 0)
         forceArgs[17] = &interactingAtoms->getDevicePointer();
     findInteractingBlocksArgs[7] = &interactingAtoms->getDevicePointer();
-    if (context.getUseDoublePrecision()) {
-        vector<double4> oldPositionsVec(numAtoms, make_double4(1e30, 1e30, 1e30, 0));
-        oldPositions->upload(oldPositionsVec);
-    }
-    else {
-        vector<float4> oldPositionsVec(numAtoms, make_float4(1e30f, 1e30f, 1e30f, 0));
-        oldPositions->upload(oldPositionsVec);
-    }
+    forceRebuildNeighborList = true;
 }
 
 void CudaNonbondedUtilities::setUsePadding(bool padding) {

platforms/cuda/src/CudaParallelKernels.cpp

@@ -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) 2011-2013 Stanford University and the Authors.      *
+ * Portions copyright (c) 2011-2015 Stanford University and the Authors.      *
  * Authors: Peter Eastman                                                     *
  * Contributors:                                                              *
  *                                                                            *
@@ -93,9 +93,9 @@ private:
 class CudaParallelCalcForcesAndEnergyKernel::FinishComputationTask : public CudaContext::WorkTask {
 public:
     FinishComputationTask(ContextImpl& context, CudaContext& cu, CudaCalcForcesAndEnergyKernel& kernel,
-            bool includeForce, bool includeEnergy, int groups, double& energy, long long& completionTime, long long* pinnedMemory, CudaArray& contextForces, bool& valid) :
+            bool includeForce, bool includeEnergy, int groups, double& energy, long long& completionTime, long long* pinnedMemory, CudaArray& contextForces, bool& valid, int& numTiles) :
             context(context), cu(cu), kernel(kernel), includeForce(includeForce), includeEnergy(includeEnergy), groups(groups), energy(energy),
-            completionTime(completionTime), pinnedMemory(pinnedMemory), contextForces(contextForces), valid(valid) {
+            completionTime(completionTime), pinnedMemory(pinnedMemory), contextForces(contextForces), valid(valid), numTiles(numTiles) {
     }
     void execute() {
         // Execute the kernel, then download forces.
@@ -120,6 +120,10 @@ public:
                     cu.getForce().download(&pinnedMemory[(cu.getContextIndex()-1)*numAtoms*3]);
             }
         }
+        if (cu.getNonbondedUtilities().getUsePeriodic() && numTiles > cu.getNonbondedUtilities().getInteractingTiles().getSize()) {
+            valid = false;
+            cu.getNonbondedUtilities().updateNeighborListSize();
+        }
     }
 private:
     ContextImpl& context;
@@ -132,6 +136,7 @@ private:
     long long* pinnedMemory;
     CudaArray& contextForces;
     bool& valid;
+    int& numTiles;
 };
 
 CudaParallelCalcForcesAndEnergyKernel::CudaParallelCalcForcesAndEnergyKernel(string name, const Platform& platform, CudaPlatform::PlatformData& data) :
@@ -201,16 +206,9 @@ double CudaParallelCalcForcesAndEnergyKernel::finishComputation(ContextImpl& con
     for (int i = 0; i < (int) data.contexts.size(); i++) {
         CudaContext& cu = *data.contexts[i];
         CudaContext::WorkThread& thread = cu.getWorkThread();
-        thread.addTask(new FinishComputationTask(context, cu, getKernel(i), includeForce, includeEnergy, groups, data.contextEnergy[i], completionTimes[i], pinnedForceBuffer, *contextForces, valid));
+        thread.addTask(new FinishComputationTask(context, cu, getKernel(i), includeForce, includeEnergy, groups, data.contextEnergy[i], completionTimes[i], pinnedForceBuffer, *contextForces, valid, tileCounts[i]));
     }
     data.syncContexts();
-    if (data.contexts[0]->getNonbondedUtilities().getUsePeriodic()) {
-        for (int i = 0; i < (int) tileCounts.size(); i++)
-            if (tileCounts[i] > data.contexts[i]->getNonbondedUtilities().getInteractingTiles().getSize()) {
-                valid = false;
-                data.contexts[i]->getNonbondedUtilities().updateNeighborListSize();
-            }
-    }
     double energy = 0.0;
     for (int i = 0; i < (int) data.contextEnergy.size(); i++)
         energy += data.contextEnergy[i];

platforms/opencl/src/OpenCLNonbondedUtilities.cpp

@@ -282,14 +282,6 @@ void OpenCLNonbondedUtilities::initialize(const System& system) {
         sortedBlockCenter = new OpenCLArray(context, numAtomBlocks+1, 4*elementSize, "sortedBlockCenter");
         sortedBlockBoundingBox = new OpenCLArray(context, numAtomBlocks+1, 4*elementSize, "sortedBlockBoundingBox");
         oldPositions = new OpenCLArray(context, numAtoms, 4*elementSize, "oldPositions");
-        if (context.getUseDoublePrecision()) {
-            vector<mm_double4> oldPositionsVec(numAtoms, mm_double4(1e30, 1e30, 1e30, 0));
-            oldPositions->upload(oldPositionsVec);
-        }
-        else {
-            vector<mm_float4> oldPositionsVec(numAtoms, mm_float4(1e30f, 1e30f, 1e30f, 0));
-            oldPositions->upload(oldPositionsVec);
-        }
         rebuildNeighborList = OpenCLArray::create<int>(context, 1, "rebuildNeighborList");
         blockSorter = new OpenCLSort(context, new BlockSortTrait(context.getUseDoublePrecision()), numAtomBlocks);
         vector<cl_uint> count(1, 0);
@@ -447,15 +439,7 @@ void OpenCLNonbondedUtilities::updateNeighborListSize() {
     findInteractingBlocksKernel.setArg<cl::Buffer>(6, interactingTiles->getDeviceBuffer());
     findInteractingBlocksKernel.setArg<cl::Buffer>(7, interactingAtoms->getDeviceBuffer());
     findInteractingBlocksKernel.setArg<cl_uint>(9, maxTiles);
-    int numAtoms = context.getNumAtoms();
-    if (context.getUseDoublePrecision()) {
-        vector<mm_double4> oldPositionsVec(numAtoms, mm_double4(1e30, 1e30, 1e30, 0));
-        oldPositions->upload(oldPositionsVec);
-    }
-    else {
-        vector<mm_float4> oldPositionsVec(numAtoms, mm_float4(1e30f, 1e30f, 1e30f, 0));
-        oldPositions->upload(oldPositionsVec);
-    }
+    sortBoxDataKernel.setArg<cl_int>(9, true);
 }
 
 void OpenCLNonbondedUtilities::setUsePadding(bool padding) {

platforms/opencl/src/OpenCLParallelKernels.cpp

@@ -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) 2011-2013 Stanford University and the Authors.      *
+ * Portions copyright (c) 2011-2015 Stanford University and the Authors.      *
  * Authors: Peter Eastman                                                     *
  * Contributors:                                                              *
  *                                                                            *
@@ -79,9 +79,9 @@ private:
 class OpenCLParallelCalcForcesAndEnergyKernel::FinishComputationTask : public OpenCLContext::WorkTask {
 public:
     FinishComputationTask(ContextImpl& context, OpenCLContext& cl, OpenCLCalcForcesAndEnergyKernel& kernel,
-            bool includeForce, bool includeEnergy, int groups, double& energy, long long& completionTime, void* pinnedMemory, bool& valid) :
+            bool includeForce, bool includeEnergy, int groups, double& energy, long long& completionTime, void* pinnedMemory, bool& valid, int& numTiles) :
             context(context), cl(cl), kernel(kernel), includeForce(includeForce), includeEnergy(includeEnergy), groups(groups), energy(energy),
-            completionTime(completionTime), pinnedMemory(pinnedMemory), valid(valid) {
+            completionTime(completionTime), pinnedMemory(pinnedMemory), valid(valid), numTiles(numTiles) {
     }
     void execute() {
         // Execute the kernel, then download forces.
@@ -98,6 +98,10 @@ public:
                 cl.getQueue().finish();
         }
         completionTime = getTime();
+        if (cl.getNonbondedUtilities().getUsePeriodic() && numTiles > cl.getNonbondedUtilities().getInteractingTiles().getSize()) {
+            valid = false;
+            cl.getNonbondedUtilities().updateNeighborListSize();
+        }
     }
 private:
     ContextImpl& context;
@@ -109,6 +113,7 @@ private:
     long long& completionTime;
     void* pinnedMemory;
     bool& valid;
+    int& numTiles;
 };
 
 OpenCLParallelCalcForcesAndEnergyKernel::OpenCLParallelCalcForcesAndEnergyKernel(string name, const Platform& platform, OpenCLPlatform::PlatformData& data) :
@@ -162,16 +167,9 @@ double OpenCLParallelCalcForcesAndEnergyKernel::finishComputation(ContextImpl& c
     for (int i = 0; i < (int) data.contexts.size(); i++) {
         OpenCLContext& cl = *data.contexts[i];
         OpenCLContext::WorkThread& thread = cl.getWorkThread();
-        thread.addTask(new FinishComputationTask(context, cl, getKernel(i), includeForce, includeEnergy, groups, data.contextEnergy[i], completionTimes[i], pinnedForceMemory, valid));
+        thread.addTask(new FinishComputationTask(context, cl, getKernel(i), includeForce, includeEnergy, groups, data.contextEnergy[i], completionTimes[i], pinnedForceMemory, valid, tileCounts[i]));
     }
     data.syncContexts();
-    if (data.contexts[0]->getNonbondedUtilities().getUsePeriodic()) {
-        for (int i = 0; i < (int) tileCounts.size(); i++)
-            if (tileCounts[i] > data.contexts[i]->getNonbondedUtilities().getInteractingTiles().getSize()) {
-                valid = false;
-                data.contexts[i]->getNonbondedUtilities().updateNeighborListSize();
-            }
-    }
     double energy = 0.0;
     for (int i = 0; i < (int) data.contextEnergy.size(); i++)
         energy += data.contextEnergy[i];