Merge pull request #2024 from peastman/copyparams

Optimizations to updateParametersInContext()

platforms/cuda/src/CudaContext.cpp

@@ -1119,33 +1119,39 @@ bool CudaContext::invalidateMolecules(CudaForceInfo* force) {
     for (int i = 0; i < forces.size(); i++)
         if (forces[i] == force)
             forceIndex = i;
-    for (int group = 0; valid && group < (int) moleculeGroups.size(); group++) {
-        MoleculeGroup& mol = moleculeGroups[group];
-        vector<int>& instances = mol.instances;
-        vector<int>& offsets = mol.offsets;
-        vector<int>& atoms = mol.atoms;
-        int numMolecules = instances.size();
-        Molecule& m1 = molecules[instances[0]];
-        int offset1 = offsets[0];
-        for (int j = 1; valid && j < numMolecules; j++) {
-            // See if the atoms are identical.
-
-            Molecule& m2 = molecules[instances[j]];
-            int offset2 = offsets[j];
-            for (int i = 0; i < (int) atoms.size() && valid; i++) {
-                if (!force->areParticlesIdentical(atoms[i]+offset1, atoms[i]+offset2))
-                    valid = false;
-            }
+    getPlatformData().threads.execute([&] (ThreadPool& threads, int threadIndex) {
+        for (int group = 0; valid && group < (int) moleculeGroups.size(); group++) {
+            MoleculeGroup& mol = moleculeGroups[group];
+            vector<int>& instances = mol.instances;
+            vector<int>& offsets = mol.offsets;
+            vector<int>& atoms = mol.atoms;
+            int numMolecules = instances.size();
+            Molecule& m1 = molecules[instances[0]];
+            int offset1 = offsets[0];
+            int numThreads = threads.getNumThreads();
+            int start = max(1, threadIndex*numMolecules/numThreads);
+            int end = (threadIndex+1)*numMolecules/numThreads;
+            for (int j = start; j < end; j++) {
+                // See if the atoms are identical.
+
+                Molecule& m2 = molecules[instances[j]];
+                int offset2 = offsets[j];
+                for (int i = 0; i < (int) atoms.size() && valid; i++) {
+                    if (!force->areParticlesIdentical(atoms[i]+offset1, atoms[i]+offset2))
+                        valid = false;
+                }
 
-            // See if the force groups are identical.
+                // See if the force groups are identical.
 
-            if (valid && forceIndex > -1) {
-                for (int k = 0; k < (int) m1.groups[forceIndex].size() && valid; k++)
-                    if (!force->areGroupsIdentical(m1.groups[forceIndex][k], m2.groups[forceIndex][k]))
-                        valid = false;
+                if (valid && forceIndex > -1) {
+                    for (int k = 0; k < (int) m1.groups[forceIndex].size() && valid; k++)
+                        if (!force->areGroupsIdentical(m1.groups[forceIndex][k], m2.groups[forceIndex][k]))
+                            valid = false;
+                }
             }
         }
-    }
+    });
+    getPlatformData().threads.waitForThreads();
     if (valid)
         return false;
 

platforms/cuda/src/CudaKernels.cpp

@@ -2165,7 +2165,7 @@ void CudaCalcNonbondedForceKernel::copyParametersToContext(ContextImpl& context,
     // Record the per-particle parameters.
     
     vector<double> chargeVector(cu.getNumAtoms());
-    vector<float2> sigmaEpsilonVector(cu.getPaddedNumAtoms(), make_float2(0, 0));
+    vector<float2> sigmaEpsilonVector(cu.getPaddedNumAtoms());
     double sumSquaredCharges = 0.0;
     double sumSquaredC6 = 0.0;
     const vector<int>& order = cu.getAtomIndex();
@@ -2180,6 +2180,8 @@ void CudaCalcNonbondedForceKernel::copyParametersToContext(ContextImpl& context,
         sumSquaredC6 += C6*C6;
         sumSquaredCharges += charge*charge;
     }
+    for (int i = force.getNumParticles(); i < cu.getPaddedNumAtoms(); i++)
+        sigmaEpsilonVector[i] = make_float2(0,0);
     cu.setCharges(chargeVector);
     sigmaEpsilon.upload(sigmaEpsilonVector);
     
@@ -2977,7 +2979,7 @@ void CudaCalcGBSAOBCForceKernel::copyParametersToContext(ContextImpl& context, c
     // Record the per-particle parameters.
     
     vector<double> chargeVector(cu.getNumAtoms());
-    vector<float2> paramsVector(cu.getPaddedNumAtoms(), make_float2(1, 1));
+    vector<float2> paramsVector(cu.getPaddedNumAtoms());
     const double dielectricOffset = 0.009;
     for (int i = 0; i < numParticles; i++) {
         double charge, radius, scalingFactor;
@@ -2986,6 +2988,8 @@ void CudaCalcGBSAOBCForceKernel::copyParametersToContext(ContextImpl& context, c
         radius -= dielectricOffset;
         paramsVector[i] = make_float2((float) radius, (float) (scalingFactor*radius));
     }
+    for (int i = numParticles; i < cu.getPaddedNumAtoms(); i++)
+        paramsVector[i] = make_float2(1, 1);
     cu.setCharges(chargeVector);
     params.upload(paramsVector);
     

platforms/opencl/src/OpenCLContext.cpp

@@ -1006,33 +1006,39 @@ bool OpenCLContext::invalidateMolecules(OpenCLForceInfo* force) {
     for (int i = 0; i < forces.size(); i++)
         if (forces[i] == force)
             forceIndex = i;
-    for (int group = 0; valid && group < (int) moleculeGroups.size(); group++) {
-        MoleculeGroup& mol = moleculeGroups[group];
-        vector<int>& instances = mol.instances;
-        vector<int>& offsets = mol.offsets;
-        vector<int>& atoms = mol.atoms;
-        int numMolecules = instances.size();
-        Molecule& m1 = molecules[instances[0]];
-        int offset1 = offsets[0];
-        for (int j = 1; valid && j < numMolecules; j++) {
-            // See if the atoms are identical.
-
-            Molecule& m2 = molecules[instances[j]];
-            int offset2 = offsets[j];
-            for (int i = 0; i < (int) atoms.size() && valid; i++) {
-                if (!force->areParticlesIdentical(atoms[i]+offset1, atoms[i]+offset2))
-                    valid = false;
-            }
+    getPlatformData().threads.execute([&] (ThreadPool& threads, int threadIndex) {
+        for (int group = 0; valid && group < (int) moleculeGroups.size(); group++) {
+            MoleculeGroup& mol = moleculeGroups[group];
+            vector<int>& instances = mol.instances;
+            vector<int>& offsets = mol.offsets;
+            vector<int>& atoms = mol.atoms;
+            int numMolecules = instances.size();
+            Molecule& m1 = molecules[instances[0]];
+            int offset1 = offsets[0];
+            int numThreads = threads.getNumThreads();
+            int start = max(1, threadIndex*numMolecules/numThreads);
+            int end = (threadIndex+1)*numMolecules/numThreads;
+            for (int j = start; j < end; j++) {
+                // See if the atoms are identical.
+
+                Molecule& m2 = molecules[instances[j]];
+                int offset2 = offsets[j];
+                for (int i = 0; i < (int) atoms.size(); i++) {
+                    if (!force->areParticlesIdentical(atoms[i]+offset1, atoms[i]+offset2))
+                        valid = false;
+                }
 
-            // See if the force groups are identical.
+                // See if the force groups are identical.
 
-            if (valid && forceIndex > -1) {
-                for (int k = 0; k < (int) m1.groups[forceIndex].size() && valid; k++)
-                    if (!force->areGroupsIdentical(m1.groups[forceIndex][k], m2.groups[forceIndex][k]))
-                        valid = false;
+                if (valid && forceIndex > -1) {
+                    for (int k = 0; k < (int) m1.groups[forceIndex].size(); k++)
+                        if (!force->areGroupsIdentical(m1.groups[forceIndex][k], m2.groups[forceIndex][k]))
+                            valid = false;
+                }
             }
         }
-    }
+    });
+    getPlatformData().threads.waitForThreads();
     if (valid)
         return false;
 

platforms/opencl/src/OpenCLKernels.cpp

@@ -2296,7 +2296,7 @@ void OpenCLCalcNonbondedForceKernel::copyParametersToContext(ContextImpl& contex
     // Record the per-particle parameters.
     
     vector<double> chargeVector(cl.getNumAtoms());
-    vector<mm_float2> sigmaEpsilonVector(cl.getPaddedNumAtoms(), mm_float2(0,0));
+    vector<mm_float2> sigmaEpsilonVector(cl.getPaddedNumAtoms());
     double sumSquaredCharges = 0.0;
     for (int i = 0; i < force.getNumParticles(); i++) {
         double charge, sigma, epsilon;
@@ -2305,6 +2305,8 @@ void OpenCLCalcNonbondedForceKernel::copyParametersToContext(ContextImpl& contex
         sigmaEpsilonVector[i] = mm_float2((float) (0.5*sigma), (float) (2.0*sqrt(epsilon)));
         sumSquaredCharges += charge*charge;
     }
+    for (int i = force.getNumParticles(); i < cl.getPaddedNumAtoms(); i++)
+        sigmaEpsilonVector[i] = mm_float2(0,0);
     cl.setCharges(chargeVector);
     sigmaEpsilon.upload(sigmaEpsilonVector);
     
@@ -3123,7 +3125,7 @@ void OpenCLCalcGBSAOBCForceKernel::copyParametersToContext(ContextImpl& context,
     // Record the per-particle parameters.
     
     vector<double> chargeVector(cl.getNumAtoms());
-    vector<mm_float2> paramsVector(cl.getPaddedNumAtoms(), mm_float2(1,1));
+    vector<mm_float2> paramsVector(cl.getPaddedNumAtoms());
     const double dielectricOffset = 0.009;
     for (int i = 0; i < numParticles; i++) {
         double charge, radius, scalingFactor;
@@ -3132,6 +3134,8 @@ void OpenCLCalcGBSAOBCForceKernel::copyParametersToContext(ContextImpl& context,
         radius -= dielectricOffset;
         paramsVector[i] = mm_float2((float) radius, (float) (scalingFactor*radius));
     }
+    for (int i = numParticles; i < cl.getPaddedNumAtoms(); i++)
+        paramsVector[i] = mm_float2(1,1);
     cl.setCharges(chargeVector);
     params.upload(paramsVector);