Optimizations to updateParametersInContext() for OpenCL

platforms/opencl/include/OpenCLContext.h

@@ -609,6 +609,10 @@ public:
     OpenCLNonbondedUtilities& getNonbondedUtilities() {
         return *nonbonded;
     }
+    /**
+     * Set the particle charges.  These are packed into the fourth element of the posq array.
+     */
+    void setCharges(const std::vector<double>& charges);
     /**
      * Get the thread used by this context for executing parallel computations.
      */
@@ -692,6 +696,12 @@ public:
      * and order to be revalidated.
      */
     void invalidateMolecules();
+    /**
+     * Mark that the current molecule definitions from one particular force (and hence the atom order)
+     * may be invalid.  This should be called whenever force field parameters change.  It will cause the
+     * definitions and order to be revalidated.
+     */
+    bool invalidateMolecules(OpenCLForceInfo* force);
 private:
     struct Molecule;
     struct MoleculeGroup;
@@ -739,6 +749,7 @@ private:
     cl::Kernel clearSixBuffersKernel;
     cl::Kernel reduceReal4Kernel;
     cl::Kernel reduceForcesKernel;
+    cl::Kernel setChargesKernel;
     std::vector<OpenCLForceInfo*> forces;
     std::vector<Molecule> molecules;
     std::vector<MoleculeGroup> moleculeGroups;
@@ -754,6 +765,7 @@ private:
     OpenCLArray* energyBuffer;
     OpenCLArray* energyParamDerivBuffer;
     OpenCLArray* atomIndexDevice;
+    OpenCLArray* chargeBuffer;
     std::vector<std::string> energyParamDerivNames;
     std::map<std::string, double> energyParamDerivWorkspace;
     std::vector<int> atomIndex;

platforms/opencl/include/OpenCLKernels.h

@@ -9,7 +9,7 @@
  * Biological Structures at Stanford, funded under the NIH Roadmap for        *
  * Medical Research, grant U54 GM072970. See https://simtk.org.               *
  *                                                                            *
- * Portions copyright (c) 2008-2016 Stanford University and the Authors.      *
+ * Portions copyright (c) 2008-2017 Stanford University and the Authors.      *
  * Authors: Peter Eastman                                                     *
  * Contributors:                                                              *
  *                                                                            *
@@ -270,9 +270,11 @@ public:
      */
     void copyParametersToContext(ContextImpl& context, const HarmonicBondForce& force);
 private:
+    class ForceInfo;
     int numBonds;
     bool hasInitializedKernel;
     OpenCLContext& cl;
+    ForceInfo* info;
     const System& system;
     OpenCLArray* params;
 };
@@ -310,9 +312,11 @@ public:
      */
     void copyParametersToContext(ContextImpl& context, const CustomBondForce& force);
 private:
+    class ForceInfo;
     int numBonds;
     bool hasInitializedKernel;
     OpenCLContext& cl;
+    ForceInfo* info;
     const System& system;
     OpenCLParameterSet* params;
     OpenCLArray* globals;
@@ -353,9 +357,11 @@ public:
      */
     void copyParametersToContext(ContextImpl& context, const HarmonicAngleForce& force);
 private:
+    class ForceInfo;
     int numAngles;
     bool hasInitializedKernel;
     OpenCLContext& cl;
+    ForceInfo* info;
     const System& system;
     OpenCLArray* params;
 };
@@ -393,9 +399,11 @@ public:
      */
     void copyParametersToContext(ContextImpl& context, const CustomAngleForce& force);
 private:
+    class ForceInfo;
     int numAngles;
     bool hasInitializedKernel;
     OpenCLContext& cl;
+    ForceInfo* info;
     const System& system;
     OpenCLParameterSet* params;
     OpenCLArray* globals;
@@ -436,9 +444,11 @@ public:
      */
     void copyParametersToContext(ContextImpl& context, const PeriodicTorsionForce& force);
 private:
+    class ForceInfo;
     int numTorsions;
     bool hasInitializedKernel;
     OpenCLContext& cl;
+    ForceInfo* info;
     const System& system;
     OpenCLArray* params;
 };
@@ -476,9 +486,11 @@ public:
      */
     void copyParametersToContext(ContextImpl& context, const RBTorsionForce& force);
 private:
+    class ForceInfo;
     int numTorsions;
     bool hasInitializedKernel;
     OpenCLContext& cl;
+    ForceInfo* info;
     const System& system;
     OpenCLArray* params;
 };
@@ -516,9 +528,11 @@ public:
      */
     void copyParametersToContext(ContextImpl& context, const CMAPTorsionForce& force);
 private:
+    class ForceInfo;
     int numTorsions;
     bool hasInitializedKernel;
     OpenCLContext& cl;
+    ForceInfo* info;
     const System& system;
     std::vector<mm_int2> mapPositionsVec;
     OpenCLArray* coefficients;
@@ -559,9 +573,11 @@ public:
      */
     void copyParametersToContext(ContextImpl& context, const CustomTorsionForce& force);
 private:
+    class ForceInfo;
     int numTorsions;
     bool hasInitializedKernel;
     OpenCLContext& cl;
+    ForceInfo* info;
     const System& system;
     OpenCLParameterSet* params;
     OpenCLArray* globals;
@@ -625,12 +641,14 @@ private:
         const char* getMaxValue() const {return "(int2) (INT_MAX, INT_MAX)";}
         const char* getSortKey() const {return "value.y";}
     };
+    class ForceInfo;
     class PmeIO;
     class PmePreComputation;
     class PmePostComputation;
     class SyncQueuePreComputation;
     class SyncQueuePostComputation;
     OpenCLContext& cl;
+    ForceInfo* info;
     bool hasInitializedKernel;
     OpenCLArray* sigmaEpsilon;
     OpenCLArray* exceptionParams;
@@ -704,8 +722,10 @@ public:
      */
     void copyParametersToContext(ContextImpl& context, const CustomNonbondedForce& force);
 private:
+    class ForceInfo;
     void initInteractionGroups(const CustomNonbondedForce& force, const std::string& interactionSource, const std::vector<std::string>& tableTypes);
     OpenCLContext& cl;
+    ForceInfo* info;
     OpenCLParameterSet* params;
     OpenCLArray* globals;
     OpenCLArray* interactionGroupData;
@@ -756,10 +776,12 @@ public:
      */
     void copyParametersToContext(ContextImpl& context, const GBSAOBCForce& force);
 private:
+    class ForceInfo;
     double prefactor, surfaceAreaFactor, cutoff;
     bool hasCreatedKernels;
     int maxTiles;
     OpenCLContext& cl;
+    ForceInfo* info;
     OpenCLArray* params;
     OpenCLArray* bornSum;
     OpenCLArray* longBornSum;
@@ -807,10 +829,12 @@ public:
      */
     void copyParametersToContext(ContextImpl& context, const CustomGBForce& force);
 private:
+    class ForceInfo;
     double cutoff;
     bool hasInitializedKernels, needParameterGradient, needEnergyParamDerivs;
     int maxTiles, numComputedValues;
     OpenCLContext& cl;
+    ForceInfo* info;
     OpenCLParameterSet* params;
     OpenCLParameterSet* computedValues;
     OpenCLParameterSet* energyDerivs;
@@ -864,9 +888,11 @@ public:
      */
     void copyParametersToContext(ContextImpl& context, const CustomExternalForce& force);
 private:
+    class ForceInfo;
     int numParticles;
     bool hasInitializedKernel;
     OpenCLContext& cl;
+    ForceInfo* info;
     const System& system;
     OpenCLParameterSet* params;
     OpenCLArray* globals;
@@ -908,9 +934,11 @@ public:
      */
     void copyParametersToContext(ContextImpl& context, const CustomHbondForce& force);
 private:
+    class ForceInfo;
     int numDonors, numAcceptors;
     bool hasInitializedKernel;
     OpenCLContext& cl;
+    ForceInfo* info;
     OpenCLParameterSet* donorParams;
     OpenCLParameterSet* acceptorParams;
     OpenCLArray* globals;
@@ -961,9 +989,11 @@ public:
     void copyParametersToContext(ContextImpl& context, const CustomCentroidBondForce& force);
 
 private:
+    class ForceInfo;
     int numGroups, numBonds;
     bool needEnergyParamDerivs;
     OpenCLContext& cl;
+    ForceInfo* info;
     OpenCLParameterSet* params;
     OpenCLArray* globals;
     OpenCLArray* groupParticles;
@@ -1013,8 +1043,10 @@ public:
     void copyParametersToContext(ContextImpl& context, const CustomCompoundBondForce& force);
 
 private:
+    class ForceInfo;
     int numBonds;
     OpenCLContext& cl;
+    ForceInfo* info;
     OpenCLParameterSet* params;
     OpenCLArray* globals;
     std::vector<std::string> globalParamNames;
@@ -1059,7 +1091,9 @@ public:
     void copyParametersToContext(ContextImpl& context, const CustomManyParticleForce& force);
 
 private:
+    class ForceInfo;
     OpenCLContext& cl;
+    ForceInfo* info;
     bool hasInitializedKernel;
     NonbondedMethod nonbondedMethod;
     int maxNeighborPairs, forceWorkgroupSize, findNeighborsWorkgroupSize;
@@ -1119,9 +1153,11 @@ public:
      */
     void copyParametersToContext(ContextImpl& context, const GayBerneForce& force);
 private:
+    class ForceInfo;
     class ReorderListener;
     void sortAtoms();
     OpenCLContext& cl;
+    ForceInfo* info;
     bool hasInitializedKernels;
     int numRealParticles, maxNeighborBlocks;
     GayBerneForce::NonbondedMethod nonbondedMethod;

platforms/opencl/src/OpenCLContext.cpp

@@ -69,8 +69,8 @@ static void CL_CALLBACK errorCallback(const char* errinfo, const void* private_i
 
 OpenCLContext::OpenCLContext(const System& system, int platformIndex, int deviceIndex, const string& precision, OpenCLPlatform::PlatformData& platformData) :
         system(system), time(0.0), platformData(platformData), stepCount(0), computeForceCount(0), stepsSinceReorder(99999), atomsWereReordered(false), posq(NULL),
-        posqCorrection(NULL), velm(NULL), forceBuffers(NULL), longForceBuffer(NULL), energyBuffer(NULL), energyParamDerivBuffer(NULL), atomIndexDevice(NULL), integration(NULL),
-        expression(NULL), bonded(NULL), nonbonded(NULL), thread(NULL) {
+        posqCorrection(NULL), velm(NULL), forceBuffers(NULL), longForceBuffer(NULL), energyBuffer(NULL), energyParamDerivBuffer(NULL), atomIndexDevice(NULL),
+        chargeBuffer(NULL), integration(NULL), expression(NULL), bonded(NULL), nonbonded(NULL), thread(NULL) {
     if (precision == "single") {
         useDoublePrecision = false;
         useMixedPrecision = false;
@@ -309,6 +309,7 @@ OpenCLContext::OpenCLContext(const System& system, int platformIndex, int device
     reduceReal4Kernel = cl::Kernel(utilities, "reduceReal4Buffer");
     if (supports64BitGlobalAtomics)
         reduceForcesKernel = cl::Kernel(utilities, "reduceForces");
+    setChargesKernel = cl::Kernel(utilities, "setCharges");
 
     // Decide whether native_sqrt(), native_rsqrt(), and native_recip() are sufficiently accurate to use.
 
@@ -439,6 +440,8 @@ OpenCLContext::~OpenCLContext() {
         delete energyParamDerivBuffer;
     if (atomIndexDevice != NULL)
         delete atomIndexDevice;
+    if (chargeBuffer != NULL)
+        delete chargeBuffer;
     if (integration != NULL)
         delete integration;
     if (expression != NULL)
@@ -747,6 +750,28 @@ void OpenCLContext::reduceBuffer(OpenCLArray& array, int numBuffers) {
     executeKernel(reduceReal4Kernel, bufferSize, 128);
 }
 
+void OpenCLContext::setCharges(const vector<double>& charges) {
+    if (chargeBuffer == NULL)
+        chargeBuffer = new OpenCLArray(*this, numAtoms, useDoublePrecision ? sizeof(double) : sizeof(float), "chargeBuffer");
+    if (getUseDoublePrecision()) {
+        double* c = (double*) getPinnedBuffer();
+        for (int i = 0; i < charges.size(); i++)
+            c[i] = charges[i];
+        chargeBuffer->upload(c);
+    }
+    else {
+        float* c = (float*) getPinnedBuffer();
+        for (int i = 0; i < charges.size(); i++)
+            c[i] = (float) charges[i];
+        chargeBuffer->upload(c);
+    }
+    setChargesKernel.setArg<cl::Buffer>(0, chargeBuffer->getDeviceBuffer());
+    setChargesKernel.setArg<cl::Buffer>(1, posq->getDeviceBuffer());
+    setChargesKernel.setArg<cl::Buffer>(2, atomIndexDevice->getDeviceBuffer());
+    setChargesKernel.setArg<cl_int>(3, numAtoms);
+    executeKernel(setChargesKernel, numAtoms);
+}
+
 /**
  * This class ensures that atom reordering doesn't break virtual sites.
  */
@@ -945,9 +970,19 @@ void OpenCLContext::findMoleculeGroups() {
 }
 
 void OpenCLContext::invalidateMolecules() {
+    for (int i = 0; i < forces.size(); i++)
+        if (invalidateMolecules(forces[i]))
+            return;
+}
+
+bool OpenCLContext::invalidateMolecules(OpenCLForceInfo* force) {
     if (numAtoms == 0 || nonbonded == NULL || !nonbonded->getUseCutoff())
-        return;
+        return false;
     bool valid = true;
+    int forceIndex = -1;
+    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;
@@ -962,22 +997,21 @@ void OpenCLContext::invalidateMolecules() {
             Molecule& m2 = molecules[instances[j]];
             int offset2 = offsets[j];
             for (int i = 0; i < (int) atoms.size() && valid; i++) {
-                for (int k = 0; k < (int) forces.size(); k++)
-                    if (!forces[k]->areParticlesIdentical(atoms[i]+offset1, atoms[i]+offset2))
-                        valid = false;
+                if (!force->areParticlesIdentical(atoms[i]+offset1, atoms[i]+offset2))
+                    valid = false;
             }
 
             // See if the force groups are identical.
 
-            for (int i = 0; i < (int) forces.size() && valid; i++) {
-                for (int k = 0; k < (int) m1.groups[i].size() && valid; k++)
-                    if (!forces[i]->areGroupsIdentical(m1.groups[i][k], m2.groups[i][k]))
+            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)
-        return;
+        return false;
 
     // The list of which molecules are identical is no longer valid.  We need to restore the
     // atoms to their original order, rebuild the list of identical molecules, and sort them
@@ -1045,6 +1079,7 @@ void OpenCLContext::invalidateMolecules() {
     for (int i = 0; i < (int) reorderListeners.size(); i++)
         reorderListeners[i]->execute();
     reorderAtoms();
+    return true;
 }
 
 void OpenCLContext::reorderAtoms() {

platforms/opencl/src/kernels/utilities.cl

@@ -107,3 +107,11 @@ __kernel void determineNativeAccuracy(__global float8* restrict values, int numV
         values[i] = (float8) (v, native_sqrt(v), native_rsqrt(v), native_recip(v), native_exp(v), native_log(v), 0.0f, 0.0f);
     }
 }
+
+/**
+ * Record the atomic charges into the posq array.
+ */
+__kernel void setCharges(__global real* restrict charges, __global real4* restrict posq, __global int* restrict atomOrder, int numAtoms) {
+    for (int i = get_global_id(0); i < numAtoms; i += get_global_size(0))
+        posq[i].w = charges[atomOrder[i]];
+}
\ No newline at end of file