Merge branch 'master' into gayberne

platforms/cuda/src/kernels/customGBValueN2.cu

@@ -73,6 +73,7 @@ extern "C" __global__ void computeN2Value(const real4* __restrict__ posq, const
                         COMPUTE_VALUE
                     }
                     value += tempValue1;
+                    ADD_TEMP_DERIVS1
 #ifdef USE_CUTOFF
                 }
 #endif
@@ -121,6 +122,8 @@ extern "C" __global__ void computeN2Value(const real4* __restrict__ posq, const
                     }
                     value += tempValue1;
                     localData[tbx+tj].value += tempValue2;
+                    ADD_TEMP_DERIVS1
+                    ADD_TEMP_DERIVS2
 #ifdef USE_CUTOFF
                 }
 #endif
@@ -133,11 +136,13 @@ extern "C" __global__ void computeN2Value(const real4* __restrict__ posq, const
 
         // Write results.
 
-        unsigned int offset = x*TILE_SIZE + tgx;
-        atomicAdd(&global_value[offset], static_cast<unsigned long long>((long long) (value*0x100000000)));
+        unsigned int offset1 = x*TILE_SIZE + tgx;
+        atomicAdd(&global_value[offset1], static_cast<unsigned long long>((long long) (value*0x100000000)));
+        STORE_PARAM_DERIVS1
         if (x != y) {
-            offset = y*TILE_SIZE + tgx;
-            atomicAdd(&global_value[offset], static_cast<unsigned long long>((long long) (localData[threadIdx.x].value*0x100000000)));
+            unsigned int offset2 = y*TILE_SIZE + tgx;
+            atomicAdd(&global_value[offset2], static_cast<unsigned long long>((long long) (localData[threadIdx.x].value*0x100000000)));
+            STORE_PARAM_DERIVS2
         }
     }
 
@@ -146,6 +151,8 @@ extern "C" __global__ void computeN2Value(const real4* __restrict__ posq, const
 
 #ifdef USE_CUTOFF
     unsigned int numTiles = interactionCount[0];
+    if (numTiles > maxTiles)
+        return; // There wasn't enough memory for the neighbor list.
     int pos = (int) (warp*(numTiles > maxTiles ? NUM_BLOCKS*((long long)NUM_BLOCKS+1)/2 : (long)numTiles)/totalWarps);
     int end = (int) ((warp+1)*(numTiles > maxTiles ? NUM_BLOCKS*((long long)NUM_BLOCKS+1)/2 : (long)numTiles)/totalWarps);
 #else
@@ -167,16 +174,12 @@ extern "C" __global__ void computeN2Value(const real4* __restrict__ posq, const
         int x, y;
         bool singlePeriodicCopy = false;
 #ifdef USE_CUTOFF
-        if (numTiles <= maxTiles) {
         x = tiles[pos];
         real4 blockSizeX = blockSize[x];
         singlePeriodicCopy = (0.5f*periodicBoxSize.x-blockSizeX.x >= CUTOFF &&
                               0.5f*periodicBoxSize.y-blockSizeX.y >= CUTOFF &&
                               0.5f*periodicBoxSize.z-blockSizeX.z >= CUTOFF);
-        }
-        else
-#endif
-        {
+#else
         y = (int) floor(NUM_BLOCKS+0.5f-SQRT((NUM_BLOCKS+0.5f)*(NUM_BLOCKS+0.5f)-2*pos));
         x = (pos-y*NUM_BLOCKS+y*(y+1)/2);
         if (x < y || x >= NUM_BLOCKS) { // Occasionally happens due to roundoff error.
@@ -199,7 +202,7 @@ extern "C" __global__ void computeN2Value(const real4* __restrict__ posq, const
         while (skipTiles[currentSkipIndex] < pos)
             currentSkipIndex++;
         includeTile = (skipTiles[currentSkipIndex] != pos);
-        }
+#endif
         if (includeTile) {
             unsigned int atom1 = x*TILE_SIZE + tgx;
 
@@ -246,6 +249,8 @@ extern "C" __global__ void computeN2Value(const real4* __restrict__ posq, const
                         }
                         value += tempValue1;
                         localData[tbx+tj].value += tempValue2;
+                        ADD_TEMP_DERIVS1
+                        ADD_TEMP_DERIVS2
                     }
                     tj = (tj + 1) & (TILE_SIZE - 1);
                 }
@@ -278,6 +283,8 @@ extern "C" __global__ void computeN2Value(const real4* __restrict__ posq, const
                         }
                         value += tempValue1;
                         localData[tbx+tj].value += tempValue2;
+                        ADD_TEMP_DERIVS1
+                        ADD_TEMP_DERIVS2
 #ifdef USE_CUTOFF
                     }
 #endif
@@ -287,14 +294,19 @@ extern "C" __global__ void computeN2Value(const real4* __restrict__ posq, const
         
             // Write results.
 
-            atomicAdd(&global_value[atom1], static_cast<unsigned long long>((long long) (value*0x100000000)));
+            unsigned int offset1 = atom1;
+            atomicAdd(&global_value[offset1], static_cast<unsigned long long>((long long) (value*0x100000000)));
+            STORE_PARAM_DERIVS1
 #ifdef USE_CUTOFF
             unsigned int atom2 = atomIndices[threadIdx.x];
 #else
             unsigned int atom2 = y*TILE_SIZE + tgx;
 #endif
-            if (atom2 < PADDED_NUM_ATOMS)
-                atomicAdd(&global_value[atom2], static_cast<unsigned long long>((long long) (localData[threadIdx.x].value*0x100000000)));
+            if (atom2 < PADDED_NUM_ATOMS) {
+                unsigned int offset2 = atom2;
+                atomicAdd(&global_value[offset2], static_cast<unsigned long long>((long long) (localData[threadIdx.x].value*0x100000000)));
+                STORE_PARAM_DERIVS2
+            }
         }
         pos++;
     }

platforms/cuda/src/kernels/customGBValuePerParticle.cu

@@ -8,6 +8,7 @@ extern "C" __global__ void computePerParticleValues(real4* posq, long long* valu
         // Load the pairwise value
 
         real sum = valueBuffers[index]/(real) 0x100000000;
+        REDUCE_PARAM0_DERIV
         
         // Now calculate other values
 

platforms/cuda/src/kernels/customIntegratorPerDof.cu

@@ -33,7 +33,8 @@ inline __device__ mixed4 convertFromDouble4(double4 a) {
 
 extern "C" __global__ void computePerDof(real4* __restrict__ posq, real4* __restrict__ posqCorrection, mixed4* __restrict__ posDelta,
         mixed4* __restrict__ velm, const long long* __restrict__ force, const mixed2* __restrict__ dt, const mixed* __restrict__ globals,
-        mixed* __restrict__ sum, const float4* __restrict__ gaussianValues, unsigned int gaussianBaseIndex, const float4* __restrict__ uniformValues, const real energy
+        mixed* __restrict__ sum, const float4* __restrict__ gaussianValues, unsigned int gaussianBaseIndex, const float4* __restrict__ uniformValues,
+        const real energy, mixed* __restrict__ energyParamDerivs
         PARAMETER_ARGUMENTS) {
     mixed stepSize = dt[0].y;
     int index = blockIdx.x*blockDim.x+threadIdx.x;

platforms/cuda/src/kernels/customNonbonded.cu

@@ -4,15 +4,18 @@ if (!isExcluded && r2 < CUTOFF_SQUARED) {
 if (!isExcluded) {
 #endif
     real tempForce = 0;
-    COMPUTE_FORCE
+    real switchValue = 1, switchDeriv = 0;
 #if USE_SWITCH
     if (r > SWITCH_CUTOFF) {
         real x = r-SWITCH_CUTOFF;
-        real switchValue = 1+x*x*x*(SWITCH_C3+x*(SWITCH_C4+x*SWITCH_C5));
-        real switchDeriv = x*x*(3*SWITCH_C3+x*(4*SWITCH_C4+x*5*SWITCH_C5));
+        switchValue = 1+x*x*x*(SWITCH_C3+x*(SWITCH_C4+x*SWITCH_C5));
+        switchDeriv = x*x*(3*SWITCH_C3+x*(4*SWITCH_C4+x*5*SWITCH_C5));
+    }
+#endif
+    COMPUTE_FORCE
+#if USE_SWITCH
     tempForce = tempForce*switchValue - tempEnergy*switchDeriv;
     tempEnergy *= switchValue;
-    }
 #endif
     dEdR += tempForce*invR;
 }

platforms/cuda/src/kernels/gbsaObc1.cu

@@ -204,6 +204,8 @@ extern "C" __global__ void computeBornSum(unsigned long long* __restrict__ globa
 
 #ifdef USE_CUTOFF
     unsigned int numTiles = interactionCount[0];
+    if (numTiles > maxTiles)
+        return; // There wasn't enough memory for the neighbor list.
     int pos = (int) (warp*(numTiles > maxTiles ? NUM_BLOCKS*((long long)NUM_BLOCKS+1)/2 : (long)numTiles)/totalWarps);
     int end = (int) ((warp+1)*(numTiles > maxTiles ? NUM_BLOCKS*((long long)NUM_BLOCKS+1)/2 : (long)numTiles)/totalWarps);
 #else
@@ -225,16 +227,12 @@ extern "C" __global__ void computeBornSum(unsigned long long* __restrict__ globa
         int x, y;
         bool singlePeriodicCopy = false;
 #ifdef USE_CUTOFF
-        if (numTiles <= maxTiles) {
             x = tiles[pos];
             real4 blockSizeX = blockSize[x];
             singlePeriodicCopy = (0.5f*periodicBoxSize.x-blockSizeX.x >= CUTOFF &&
                                   0.5f*periodicBoxSize.y-blockSizeX.y >= CUTOFF &&
                                   0.5f*periodicBoxSize.z-blockSizeX.z >= CUTOFF);
-        }
-        else
-#endif
-        {
+#else
         y = (int) floor(NUM_BLOCKS+0.5f-SQRT((NUM_BLOCKS+0.5f)*(NUM_BLOCKS+0.5f)-2*pos));
         x = (pos-y*NUM_BLOCKS+y*(y+1)/2);
         if (x < y || x >= NUM_BLOCKS) { // Occasionally happens due to roundoff error.
@@ -257,7 +255,7 @@ extern "C" __global__ void computeBornSum(unsigned long long* __restrict__ globa
         while (skipTiles[currentSkipIndex] < pos)
             currentSkipIndex++;
         includeTile = (skipTiles[currentSkipIndex] != pos);
-        }
+#endif
         if (includeTile) {
             unsigned int atom1 = x*TILE_SIZE + tgx;
 
@@ -559,6 +557,8 @@ extern "C" __global__ void computeGBSAForce1(unsigned long long* __restrict__ fo
 
 #ifdef USE_CUTOFF
     unsigned int numTiles = interactionCount[0];
+    if (numTiles > maxTiles)
+        return; // There wasn't enough memory for the neighbor list.
     int pos = (int) (warp*(numTiles > maxTiles ? NUM_BLOCKS*((long long)NUM_BLOCKS+1)/2 : (long)numTiles)/totalWarps);
     int end = (int) ((warp+1)*(numTiles > maxTiles ? NUM_BLOCKS*((long long)NUM_BLOCKS+1)/2 : (long)numTiles)/totalWarps);
 #else
@@ -580,16 +580,12 @@ extern "C" __global__ void computeGBSAForce1(unsigned long long* __restrict__ fo
         int x, y;
         bool singlePeriodicCopy = false;
 #ifdef USE_CUTOFF
-        if (numTiles <= maxTiles) {
         x = tiles[pos];
         real4 blockSizeX = blockSize[x];
         singlePeriodicCopy = (0.5f*periodicBoxSize.x-blockSizeX.x >= CUTOFF &&
                               0.5f*periodicBoxSize.y-blockSizeX.y >= CUTOFF &&
                               0.5f*periodicBoxSize.z-blockSizeX.z >= CUTOFF);
-        }
-        else
-#endif
-        {
+#else
         y = (int) floor(NUM_BLOCKS+0.5f-SQRT((NUM_BLOCKS+0.5f)*(NUM_BLOCKS+0.5f)-2*pos));
         x = (pos-y*NUM_BLOCKS+y*(y+1)/2);
         if (x < y || x >= NUM_BLOCKS) { // Occasionally happens due to roundoff error.
@@ -612,7 +608,7 @@ extern "C" __global__ void computeGBSAForce1(unsigned long long* __restrict__ fo
         while (skipTiles[currentSkipIndex] < pos)
             currentSkipIndex++;
         includeTile = (skipTiles[currentSkipIndex] != pos);
-        }
+#endif
         if (includeTile) {
             unsigned int atom1 = x*TILE_SIZE + tgx;
 

platforms/cuda/src/kernels/nonbonded.cu

@@ -113,11 +113,14 @@ extern "C" __global__ void computeNonbonded(
     const unsigned int tgx = threadIdx.x & (TILE_SIZE-1); // index within the warp
     const unsigned int tbx = threadIdx.x - tgx;           // block warpIndex
     mixed energy = 0;
+    INIT_DERIVATIVES
     // used shared memory if the device cannot shuffle
 #ifndef ENABLE_SHUFFLE
     __shared__ AtomData localData[THREAD_BLOCK_SIZE];
 #endif
+
     // First loop: process tiles that contain exclusions.
+
     const unsigned int firstExclusionTile = FIRST_EXCLUSION_TILE+warp*(LAST_EXCLUSION_TILE-FIRST_EXCLUSION_TILE)/totalWarps;
     const unsigned int lastExclusionTile = FIRST_EXCLUSION_TILE+(warp+1)*(LAST_EXCLUSION_TILE-FIRST_EXCLUSION_TILE)/totalWarps;
     for (int pos = firstExclusionTile; pos < lastExclusionTile; pos++) {
@@ -173,6 +176,7 @@ extern "C" __global__ void computeNonbonded(
                 bool isExcluded = (atom1 >= NUM_ATOMS || atom2 >= NUM_ATOMS || !(excl & 0x1));
 #endif
                 real tempEnergy = 0.0f;
+                const real interactionScale = 0.5f;
                 COMPUTE_INTERACTION
                 energy += 0.5f*tempEnergy;
 #ifdef INCLUDE_FORCES
@@ -241,6 +245,7 @@ extern "C" __global__ void computeNonbonded(
                 bool isExcluded = (atom1 >= NUM_ATOMS || atom2 >= NUM_ATOMS || !(excl & 0x1));
 #endif
                 real tempEnergy = 0.0f;
+                const real interactionScale = 1.0f;
                 COMPUTE_INTERACTION
                 energy += tempEnergy;
 #ifdef INCLUDE_FORCES
@@ -309,8 +314,11 @@ extern "C" __global__ void computeNonbonded(
 
     // Second loop: tiles without exclusions, either from the neighbor list (with cutoff) or just enumerating all
     // of them (no cutoff).
+
 #ifdef USE_CUTOFF
     const unsigned int numTiles = interactionCount[0];
+    if (numTiles > maxTiles)
+        return; // There wasn't enough memory for the neighbor list.
     int pos = (int) (numTiles > maxTiles ? startTileIndex+warp*(long long)numTileIndices/totalWarps : warp*(long long)numTiles/totalWarps);
     int end = (int) (numTiles > maxTiles ? startTileIndex+(warp+1)*(long long)numTileIndices/totalWarps : (warp+1)*(long long)numTiles/totalWarps);
 #else
@@ -335,16 +343,12 @@ extern "C" __global__ void computeNonbonded(
         int x, y;
         bool singlePeriodicCopy = false;
 #ifdef USE_CUTOFF
-        if (numTiles <= maxTiles) {
         x = tiles[pos];
         real4 blockSizeX = blockSize[x];
         singlePeriodicCopy = (0.5f*periodicBoxSize.x-blockSizeX.x >= MAX_CUTOFF &&
                               0.5f*periodicBoxSize.y-blockSizeX.y >= MAX_CUTOFF &&
                               0.5f*periodicBoxSize.z-blockSizeX.z >= MAX_CUTOFF);
-        }
-        else
-#endif
-        {
+#else
         y = (int) floor(NUM_BLOCKS+0.5f-SQRT((NUM_BLOCKS+0.5f)*(NUM_BLOCKS+0.5f)-2*pos));
         x = (pos-y*NUM_BLOCKS+y*(y+1)/2);
         if (x < y || x >= NUM_BLOCKS) { // Occasionally happens due to roundoff error.
@@ -367,7 +371,7 @@ extern "C" __global__ void computeNonbonded(
         while (skipTiles[currentSkipIndex] < pos)
             currentSkipIndex++;
         includeTile = (skipTiles[currentSkipIndex] != pos);
-        }
+#endif
         if (includeTile) {
             unsigned int atom1 = x*TILE_SIZE + tgx;
             // Load atom data for this tile.
@@ -447,6 +451,7 @@ extern "C" __global__ void computeNonbonded(
                     bool isExcluded = (atom1 >= NUM_ATOMS || atom2 >= NUM_ATOMS);
 #endif
                     real tempEnergy = 0.0f;
+                    const real interactionScale = 1.0f;
                     COMPUTE_INTERACTION
                     energy += tempEnergy;
 #ifdef INCLUDE_FORCES
@@ -517,6 +522,7 @@ extern "C" __global__ void computeNonbonded(
                     bool isExcluded = (atom1 >= NUM_ATOMS || atom2 >= NUM_ATOMS);
 #endif
                     real tempEnergy = 0.0f;
+                    const real interactionScale = 1.0f;
                     COMPUTE_INTERACTION
                     energy += tempEnergy;
 #ifdef INCLUDE_FORCES
@@ -585,4 +591,5 @@ extern "C" __global__ void computeNonbonded(
 #ifdef INCLUDE_ENERGY
     energyBuffer[blockIdx.x*blockDim.x+threadIdx.x] += energy;
 #endif
+    SAVE_DERIVATIVES
 }
\ No newline at end of file

platforms/opencl/include/OpenCLBondedUtilities.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) 2011-2015 Stanford University and the Authors.      *
+ * Portions copyright (c) 2011-2016 Stanford University and the Authors.      *
  * Authors: Peter Eastman                                                     *
  * Contributors:                                                              *
  *                                                                            *
@@ -100,6 +100,15 @@ public:
      * refer to it by this name.
      */
     std::string addArgument(cl::Memory& data, const std::string& type);
+    /**
+     * Register that the interaction kernel will be computing the derivative of the potential energy
+     * with respect to a parameter.
+     * 
+     * @param param   the name of the parameter
+     * @return the variable that will be used to accumulate the derivative.  Any code you pass to addInteraction() should
+     * add its contributions to this variable.
+     */
+    std::string addEnergyParameterDerivative(const std::string& param);
     /**
      * Add some OpenCL code that should be included in the program, before the start of the kernel.
      * This can be used, for example, to define functions that will be called by the kernel.
@@ -137,6 +146,7 @@ private:
     std::vector<OpenCLArray*> atomIndices;
     std::vector<OpenCLArray*> bufferIndices;
     std::vector<std::string> prefixCode;
+    std::vector<std::string> energyParameterDerivatives;
     int numForceBuffers, maxBonds, allGroups;
     bool hasInitializedKernels;
 };

platforms/opencl/include/OpenCLContext.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) 2009-2015 Stanford University and the Authors.      *
+ * Portions copyright (c) 2009-2016 Stanford University and the Authors.      *
  * Authors: Peter Eastman                                                     *
  * Contributors:                                                              *
  *                                                                            *
@@ -264,6 +264,12 @@ public:
     OpenCLArray& getEnergyBuffer() {
         return *energyBuffer;
     }
+    /**
+     * Get the array which contains the buffer in which derivatives of the energy with respect to parameters are computed.
+     */
+    OpenCLArray& getEnergyParamDerivBuffer() {
+        return *energyParamDerivBuffer;
+    }
     /**
      * Get a pointer to a block of pinned memory that can be used for efficient transfers between host and device.
      * This is guaranteed to be at least as large as any of the arrays returned by methods of this class.
@@ -408,6 +414,18 @@ public:
     void setStepsSinceReorder(int steps) {
         stepsSinceReorder = steps;
     }
+    /**
+     * Get the flag that marks whether the current force evaluation is valid.
+     */
+    bool getForcesValid() const {
+        return forcesValid;
+    }
+    /**
+     * Get the flag that marks whether the current force evaluation is valid.
+     */
+    void setForcesValid(bool valid) {
+        forcesValid = valid;
+    }
     /**
      * Get the number of atoms.
      */
@@ -647,6 +665,27 @@ public:
     std::vector<ForcePostComputation*>& getPostComputations() {
         return postComputations;
     }
+    /**
+     * Get the names of all parameters with respect to which energy derivatives are computed.
+     */
+    const std::vector<std::string>& getEnergyParamDerivNames() const {
+        return energyParamDerivNames;
+    }
+    /**
+     * Get a workspace data structure used for accumulating the values of derivatives of the energy
+     * with respect to parameters.
+     */
+    std::map<std::string, double>& getEnergyParamDerivWorkspace() {
+        return energyParamDerivWorkspace;
+    }
+    /**
+     * Register that the derivative of potential energy with respect to a context parameter
+     * will need to be calculated.  If this is called multiple times for a single parameter,
+     * it is only added to the list once.
+     * 
+     * @param param    the name of the parameter to add
+     */
+    void addEnergyParameterDerivative(const std::string& param);
     /**
      * 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
@@ -684,7 +723,7 @@ private:
     int numThreadBlocks;
     int numForceBuffers;
     int simdWidth;
-    bool supports64BitGlobalAtomics, supportsDoublePrecision, useDoublePrecision, useMixedPrecision, atomsWereReordered, boxIsTriclinic;
+    bool supports64BitGlobalAtomics, supportsDoublePrecision, useDoublePrecision, useMixedPrecision, atomsWereReordered, boxIsTriclinic, forcesValid;
     mm_float4 periodicBoxSize, invPeriodicBoxSize, periodicBoxVecX, periodicBoxVecY, periodicBoxVecZ;
     mm_double4 periodicBoxSizeDouble, invPeriodicBoxSizeDouble, periodicBoxVecXDouble, periodicBoxVecYDouble, periodicBoxVecZDouble;
     std::string defaultOptimizationOptions;
@@ -713,7 +752,10 @@ private:
     OpenCLArray* forceBuffers;
     OpenCLArray* longForceBuffer;
     OpenCLArray* energyBuffer;
+    OpenCLArray* energyParamDerivBuffer;
     OpenCLArray* atomIndexDevice;
+    std::vector<std::string> energyParamDerivNames;
+    std::map<std::string, double> energyParamDerivWorkspace;
     std::vector<int> atomIndex;
     std::vector<cl::Memory*> autoclearBuffers;
     std::vector<int> autoclearBufferSizes;

platforms/opencl/include/OpenCLKernels.h

@@ -141,6 +141,12 @@ public:
      * @param forces  on exit, this contains the forces
      */
     void getForces(ContextImpl& context, std::vector<Vec3>& forces);
+    /**
+     * Get the current derivatives of the energy with respect to context parameters.
+     *
+     * @param derivs  on exit, this contains the derivatives
+     */
+    void getEnergyParameterDerivatives(ContextImpl& context, std::map<std::string, double>& derivs);
     /**
      * Get the current periodic box vectors.
      *
@@ -709,6 +715,7 @@ private:
     std::vector<cl_float> globalParamValues;
     std::vector<OpenCLArray*> tabulatedFunctions;
     double longRangeCoefficient;
+    std::vector<double> longRangeCoefficientDerivs;
     bool hasInitializedLongRangeCorrection, hasInitializedKernel;
     int numGroupThreadBlocks;
     CustomNonbondedForce* forceCopy;
@@ -801,13 +808,15 @@ public:
     void copyParametersToContext(ContextImpl& context, const CustomGBForce& force);
 private:
     double cutoff;
-    bool hasInitializedKernels, needParameterGradient;
+    bool hasInitializedKernels, needParameterGradient, needEnergyParamDerivs;
     int maxTiles, numComputedValues;
     OpenCLContext& cl;
     OpenCLParameterSet* params;
     OpenCLParameterSet* computedValues;
     OpenCLParameterSet* energyDerivs;
     OpenCLParameterSet* energyDerivChain;
+    std::vector<OpenCLParameterSet*> dValuedParam;
+    std::vector<OpenCLArray*> dValue0dParam;
     OpenCLArray* longEnergyDerivs;
     OpenCLArray* globals;
     OpenCLArray* valueBuffers;
@@ -953,6 +962,7 @@ public:
 
 private:
     int numGroups, numBonds;
+    bool needEnergyParamDerivs;
     OpenCLContext& cl;
     OpenCLParameterSet* params;
     OpenCLArray* globals;
@@ -1339,7 +1349,7 @@ public:
     enum GlobalTargetType {DT, VARIABLE, PARAMETER};
     OpenCLIntegrateCustomStepKernel(std::string name, const Platform& platform, OpenCLContext& cl) : IntegrateCustomStepKernel(name, platform), cl(cl),
             hasInitializedKernels(false), localValuesAreCurrent(false), globalValues(NULL), sumBuffer(NULL), summedValue(NULL), uniformRandoms(NULL),
-            randomSeed(NULL), perDofValues(NULL) {
+            randomSeed(NULL), perDofEnergyParamDerivs(NULL), perDofValues(NULL), needsEnergyParamDerivs(false) {
     }
     ~OpenCLIntegrateCustomStepKernel();
     /**
@@ -1404,8 +1414,11 @@ public:
 private:
     class ReorderListener;
     class GlobalTarget;
+    class DerivFunction;
     std::string createPerDofComputation(const std::string& variable, const Lepton::ParsedExpression& expr, int component, CustomIntegrator& integrator, const std::string& forceName, const std::string& energyName);
     void prepareForComputation(ContextImpl& context, CustomIntegrator& integrator, bool& forcesAreValid);
+    Lepton::ExpressionTreeNode replaceDerivFunctions(const Lepton::ExpressionTreeNode& node, OpenMM::ContextImpl& context);
+    void findExpressionsForDerivs(const Lepton::ExpressionTreeNode& node, std::vector<std::pair<Lepton::ExpressionTreeNode, std::string> >& variableNodes);
     void recordGlobalValue(double value, GlobalTarget target);
     void recordChangedParameters(ContextImpl& context);
     bool evaluateCondition(int step);
@@ -1413,18 +1426,23 @@ private:
     double energy;
     float energyFloat;
     int numGlobalVariables;
-    bool hasInitializedKernels, deviceValuesAreCurrent, deviceGlobalsAreCurrent, modifiesParameters, keNeedsForce, hasAnyConstraints;
+    bool hasInitializedKernels, deviceValuesAreCurrent, deviceGlobalsAreCurrent, modifiesParameters, keNeedsForce, hasAnyConstraints, needsEnergyParamDerivs;
     mutable bool localValuesAreCurrent;
     OpenCLArray* globalValues;
     OpenCLArray* sumBuffer;
     OpenCLArray* summedValue;
     OpenCLArray* uniformRandoms;
     OpenCLArray* randomSeed;
+    OpenCLArray* perDofEnergyParamDerivs;
     std::map<int, OpenCLArray*> savedForces;
     std::set<int> validSavedForces;
     OpenCLParameterSet* perDofValues;
     mutable std::vector<std::vector<cl_float> > localPerDofValuesFloat;
     mutable std::vector<std::vector<cl_double> > localPerDofValuesDouble;
+    std::map<std::string, double> energyParamDerivs;
+    std::vector<std::string> perDofEnergyParamDerivNames;
+    std::vector<cl_float> localPerDofEnergyParamDerivsFloat;
+    std::vector<cl_double> localPerDofEnergyParamDerivsDouble;
     std::vector<float> globalValuesFloat;
     std::vector<double> globalValuesDouble;
     std::vector<double> initialGlobalVariables;

platforms/opencl/include/OpenCLNonbondedUtilities.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) 2009-2013 Stanford University and the Authors.      *
+ * Portions copyright (c) 2009-2016 Stanford University and the Authors.      *
  * Authors: Peter Eastman                                                     *
  * Contributors:                                                              *
  *                                                                            *
@@ -88,6 +88,15 @@ public:
      * Add an array (other than a per-atom parameter) that should be passed as an argument to the default interaction kernel.
      */
     void addArgument(const ParameterInfo& parameter);
+    /**
+     * Register that the interaction kernel will be computing the derivative of the potential energy
+     * with respect to a parameter.
+     * 
+     * @param param   the name of the parameter
+     * @return the variable that will be used to accumulate the derivative.  Any code you pass to addInteraction() should
+     * add its contributions to this variable.
+     */
+    std::string addEnergyParameterDerivative(const std::string& param);
     /**
      * Specify the list of exclusions that an interaction outside the default kernel will depend on.
      * 
@@ -281,9 +290,13 @@ private:
     OpenCLArray* oldPositions;
     OpenCLArray* rebuildNeighborList;
     OpenCLSort* blockSorter;
+    cl::Event downloadCountEvent;
+    cl::Buffer* pinnedCountBuffer;
+    int* pinnedCountMemory;
     std::vector<std::vector<int> > atomExclusions;
     std::vector<ParameterInfo> parameters;
     std::vector<ParameterInfo> arguments;
+    std::vector<std::string> energyParameterDerivatives;
     std::map<int, double> groupCutoff;
     std::map<int, std::string> groupKernelSource;
     double lastCutoff;

platforms/opencl/src/OpenCLBondedUtilities.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) 2011-2015 Stanford University and the Authors.      *
+ * Portions copyright (c) 2011-2016 Stanford University and the Authors.      *
  * Authors: Peter Eastman                                                     *
  * Contributors:                                                              *
  *                                                                            *
@@ -56,12 +56,25 @@ void OpenCLBondedUtilities::addInteraction(const vector<vector<int> >& atoms, co
     }
 }
 
-std::string OpenCLBondedUtilities::addArgument(cl::Memory& data, const string& type) {
+string OpenCLBondedUtilities::addArgument(cl::Memory& data, const string& type) {
     arguments.push_back(&data);
     argTypes.push_back(type);
     return "customArg"+context.intToString(arguments.size());
 }
 
+string OpenCLBondedUtilities::addEnergyParameterDerivative(const string& param) {
+    // See if the parameter has already been added.
+    
+    int index;
+    for (index = 0; index < energyParameterDerivatives.size(); index++)
+        if (param == energyParameterDerivatives[index])
+            break;
+    if (index == energyParameterDerivatives.size())
+        energyParameterDerivatives.push_back(param);
+    context.addEnergyParameterDerivative(param);
+    return string("energyParamDeriv")+context.intToString(index);
+}
+
 void OpenCLBondedUtilities::addPrefixCode(const string& source) {
     for (int i = 0; i < (int) prefixCode.size(); i++)
         if (prefixCode[i] == source)
@@ -190,13 +203,23 @@ void OpenCLBondedUtilities::initialize(const System& system) {
         }
         for (int i = 0; i < (int) arguments.size(); i++)
             s<<", __global "<<argTypes[i]<<"* customArg"<<(i+1);
+        if (energyParameterDerivatives.size() > 0)
+            s<<", __global mixed* restrict energyParamDerivs";
         s<<") {\n";
         s<<"mixed energy = 0;\n";
+        for (int i = 0; i < energyParameterDerivatives.size(); i++)
+            s<<"mixed energyParamDeriv"<<i<<" = 0;\n";
         for (int i = 0; i < setSize; i++) {
             int force = set[i];
             s<<createForceSource(i, forceAtoms[force].size(), forceAtoms[force][0].size(), forceGroup[force], forceSource[force]);
         }
         s<<"energyBuffer[get_global_id(0)] += energy;\n";
+        const vector<string>& allParamDerivNames = context.getEnergyParamDerivNames();
+        int numDerivs = allParamDerivNames.size();
+        for (int i = 0; i < energyParameterDerivatives.size(); i++)
+            for (int index = 0; index < numDerivs; index++)
+                if (allParamDerivNames[index] == energyParameterDerivatives[i])
+                    s<<"energyParamDerivs[get_global_id(0)*"<<numDerivs<<"+"<<index<<"] += energyParamDeriv"<<i<<";\n";
         s<<"}\n";
         map<string, string> defines;
         defines["PADDED_NUM_ATOMS"] = context.intToString(context.getPaddedNumAtoms());
@@ -274,6 +297,8 @@ void OpenCLBondedUtilities::computeInteractions(int groups) {
             }
             for (int j = 0; j < (int) arguments.size(); j++)
                 kernel.setArg<cl::Memory>(index++, *arguments[j]);
+            if (energyParameterDerivatives.size() > 0)
+                kernel.setArg<cl::Memory>(index++, context.getEnergyParamDerivBuffer().getDeviceBuffer());
         }
     }
     for (int i = 0; i < (int) kernels.size(); i++) {

platforms/opencl/src/OpenCLContext.cpp

@@ -69,7 +69,7 @@ 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), atomIndexDevice(NULL), integration(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) {
     if (precision == "single") {
         useDoublePrecision = false;
@@ -179,10 +179,8 @@ OpenCLContext::OpenCLContext(const System& system, int platformIndex, int device
         compilationDefines["WORK_GROUP_SIZE"] = intToString(ThreadBlockSize);
         if (platformVendor.size() >= 5 && platformVendor.substr(0, 5) == "Intel")
             defaultOptimizationOptions = "";
-        else if (platformVendor == "Apple")
-            defaultOptimizationOptions = "-cl-mad-enable -cl-no-signed-zeros";
         else
-            defaultOptimizationOptions = "-cl-fast-relaxed-math";
+            defaultOptimizationOptions = "-cl-mad-enable -cl-no-signed-zeros";
         supports64BitGlobalAtomics = (device.getInfo<CL_DEVICE_EXTENSIONS>().find("cl_khr_int64_base_atomics") != string::npos);
         supportsDoublePrecision = (device.getInfo<CL_DEVICE_EXTENSIONS>().find("cl_khr_fp64") != string::npos);
         if ((useDoublePrecision || useMixedPrecision) && !supportsDoublePrecision)
@@ -437,6 +435,8 @@ OpenCLContext::~OpenCLContext() {
         delete longForceBuffer;
     if (energyBuffer != NULL)
         delete energyBuffer;
+    if (energyParamDerivBuffer != NULL)
+        delete energyParamDerivBuffer;
     if (atomIndexDevice != NULL)
         delete atomIndexDevice;
     if (integration != NULL)
@@ -457,15 +457,16 @@ void OpenCLContext::initialize() {
     numForceBuffers = std::max(numForceBuffers, bonded->getNumForceBuffers());
     for (int i = 0; i < (int) forces.size(); i++)
         numForceBuffers = std::max(numForceBuffers, forces[i]->getRequiredForceBuffers());
+    int energyBufferSize = max(numThreadBlocks*ThreadBlockSize, nonbonded->getNumEnergyBuffers());
     if (useDoublePrecision) {
         forceBuffers = OpenCLArray::create<mm_double4>(*this, paddedNumAtoms*numForceBuffers, "forceBuffers");
         force = OpenCLArray::create<mm_double4>(*this, &forceBuffers->getDeviceBuffer(), paddedNumAtoms, "force");
-        energyBuffer = OpenCLArray::create<cl_double>(*this, max(numThreadBlocks*ThreadBlockSize, nonbonded->getNumEnergyBuffers()), "energyBuffer");
+        energyBuffer = OpenCLArray::create<cl_double>(*this, energyBufferSize, "energyBuffer");
     }
     else {
         forceBuffers = OpenCLArray::create<mm_float4>(*this, paddedNumAtoms*numForceBuffers, "forceBuffers");
         force = OpenCLArray::create<mm_float4>(*this, &forceBuffers->getDeviceBuffer(), paddedNumAtoms, "force");
-        energyBuffer = OpenCLArray::create<cl_double>(*this, max(numThreadBlocks*ThreadBlockSize, nonbonded->getNumEnergyBuffers()), "energyBuffer");
+        energyBuffer = OpenCLArray::create<cl_double>(*this, energyBufferSize, "energyBuffer");
     }
     if (supports64BitGlobalAtomics) {
         longForceBuffer = OpenCLArray::create<cl_long>(*this, 3*paddedNumAtoms, "longForceBuffer");
@@ -477,7 +478,15 @@ void OpenCLContext::initialize() {
     }
     addAutoclearBuffer(*forceBuffers);
     addAutoclearBuffer(*energyBuffer);
-    int bufferBytes = max(velm->getSize()*velm->getElementSize(), energyBuffer->getSize()*energyBuffer->getElementSize());
+    int numEnergyParamDerivs = energyParamDerivNames.size();
+    if (numEnergyParamDerivs > 0) {
+        if (useDoublePrecision || useMixedPrecision)
+            energyParamDerivBuffer = OpenCLArray::create<cl_double>(*this, numEnergyParamDerivs*energyBufferSize, "energyParamDerivBuffer");
+        else
+            energyParamDerivBuffer = OpenCLArray::create<cl_float>(*this, numEnergyParamDerivs*energyBufferSize, "energyParamDerivBuffer");
+        addAutoclearBuffer(*energyParamDerivBuffer);
+    }
+    int bufferBytes = max(velm->getSize()*velm->getElementSize(), energyBufferSize*energyBuffer->getElementSize());
     pinnedBuffer = new cl::Buffer(context, CL_MEM_ALLOC_HOST_PTR, bufferBytes);
     pinnedMemory = currentQueue.enqueueMapBuffer(*pinnedBuffer, CL_TRUE, CL_MAP_READ | CL_MAP_WRITE, 0, bufferBytes);
     for (int i = 0; i < numAtoms; i++) {
@@ -1052,7 +1061,6 @@ void OpenCLContext::reorderAtoms() {
         reorderAtomsImpl<cl_float, mm_float4, cl_double, mm_double4>();
     else
         reorderAtomsImpl<cl_float, mm_float4, cl_float, mm_float4>();
-    nonbonded->updateNeighborListSize();
 }
 
 template <class Real, class Real4, class Mixed, class Mixed4>
@@ -1232,6 +1240,15 @@ void OpenCLContext::addPostComputation(ForcePostComputation* computation) {
     postComputations.push_back(computation);
 }
 
+void OpenCLContext::addEnergyParameterDerivative(const string& param) {
+    // See if this parameter has already been registered.
+    
+    for (int i = 0; i < energyParamDerivNames.size(); i++)
+        if (param == energyParamDerivNames[i])
+            return;
+    energyParamDerivNames.push_back(param);
+}
+
 struct OpenCLContext::WorkThread::ThreadData {
     ThreadData(std::queue<OpenCLContext::WorkTask*>& tasks, bool& waiting,  bool& finished,
             pthread_mutex_t& queueLock, pthread_cond_t& waitForTaskCondition, pthread_cond_t& queueEmptyCondition) :

platforms/opencl/src/OpenCLExpressionUtilities.cpp

@@ -174,8 +174,8 @@ void OpenCLExpressionUtilities::processExpression(stringstream& out, const Expre
                     out << "if (x >= " << paramsFloat[2] << " && x <= " << paramsFloat[3] << " && y >= " << paramsFloat[4] << " && y <= " << paramsFloat[5] << ") {\n";
                     out << "x = (x - " << paramsFloat[2] << ")*" << paramsFloat[6] << ";\n";
                     out << "y = (y - " << paramsFloat[4] << ")*" << paramsFloat[7] << ";\n";
-                    out << "int s = min((int) floor(x), " << paramsInt[0] << ");\n";
-                    out << "int t = min((int) floor(y), " << paramsInt[1] << ");\n";
+                    out << "int s = min((int) floor(x), " << paramsInt[0] << "-1);\n";
+                    out << "int t = min((int) floor(y), " << paramsInt[1] << "-1);\n";
                     out << "int coeffIndex = 4*(s+" << paramsInt[0] << "*t);\n";
                     out << "float4 c[4];\n";
                     for (int j = 0; j < 4; j++)
@@ -217,9 +217,9 @@ void OpenCLExpressionUtilities::processExpression(stringstream& out, const Expre
                     out << "x = (x - " << paramsFloat[3] << ")*" << paramsFloat[9] << ";\n";
                     out << "y = (y - " << paramsFloat[5] << ")*" << paramsFloat[10] << ";\n";
                     out << "z = (z - " << paramsFloat[7] << ")*" << paramsFloat[11] << ";\n";
-                    out << "int s = min((int) floor(x), " << paramsInt[0] << ");\n";
-                    out << "int t = min((int) floor(y), " << paramsInt[1] << ");\n";
-                    out << "int u = min((int) floor(z), " << paramsInt[2] << ");\n";
+                    out << "int s = min((int) floor(x), " << paramsInt[0] << "-1);\n";
+                    out << "int t = min((int) floor(y), " << paramsInt[1] << "-1);\n";
+                    out << "int u = min((int) floor(z), " << paramsInt[2] << "-1);\n";
                     out << "int coeffIndex = 16*(s+" << paramsInt[0] << "*(t+" << paramsInt[1] << "*u));\n";
                     out << "float4 c[16];\n";
                     for (int j = 0; j < 16; j++)

platforms/opencl/src/OpenCLNonbondedUtilities.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) 2009-2015 Stanford University and the Authors.      *
+ * Portions copyright (c) 2009-2016 Stanford University and the Authors.      *
  * Authors: Peter Eastman                                                     *
  * Contributors:                                                              *
  *                                                                            *
@@ -57,7 +57,7 @@ private:
 OpenCLNonbondedUtilities::OpenCLNonbondedUtilities(OpenCLContext& context) : context(context), useCutoff(false), usePeriodic(false), anyExclusions(false), usePadding(true),
         numForceBuffers(0), exclusionIndices(NULL), exclusionRowIndices(NULL), exclusionTiles(NULL), exclusions(NULL), interactingTiles(NULL), interactingAtoms(NULL),
         interactionCount(NULL), blockCenter(NULL), blockBoundingBox(NULL), sortedBlocks(NULL), sortedBlockCenter(NULL), sortedBlockBoundingBox(NULL),
-        oldPositions(NULL), rebuildNeighborList(NULL), blockSorter(NULL), forceRebuildNeighborList(true), lastCutoff(0.0), groupFlags(0) {
+        oldPositions(NULL), rebuildNeighborList(NULL), blockSorter(NULL), pinnedCountBuffer(NULL), pinnedCountMemory(NULL), forceRebuildNeighborList(true), lastCutoff(0.0), groupFlags(0) {
     // Decide how many thread blocks and force buffers to use.
 
     deviceIsCpu = (context.getDevice().getInfo<CL_DEVICE_TYPE>() == CL_DEVICE_TYPE_CPU);
@@ -90,6 +90,8 @@ OpenCLNonbondedUtilities::OpenCLNonbondedUtilities(OpenCLContext& context) : con
             numForceBuffers = numForceThreadBlocks*forceThreadBlockSize/OpenCLContext::TileSize;
         }
     }
+    pinnedCountBuffer = new cl::Buffer(context.getContext(), CL_MEM_ALLOC_HOST_PTR, sizeof(int));
+    pinnedCountMemory = (int*) context.getQueue().enqueueMapBuffer(*pinnedCountBuffer, CL_TRUE, CL_MAP_READ, 0, sizeof(int));
 }
 
 OpenCLNonbondedUtilities::~OpenCLNonbondedUtilities() {
@@ -123,6 +125,8 @@ OpenCLNonbondedUtilities::~OpenCLNonbondedUtilities() {
         delete rebuildNeighborList;
     if (blockSorter != NULL)
         delete blockSorter;
+    if (pinnedCountBuffer != NULL)
+        delete pinnedCountBuffer;
 }
 
 void OpenCLNonbondedUtilities::addInteraction(bool usesCutoff, bool usesPeriodic, bool usesExclusions, double cutoffDistance, const vector<vector<int> >& exclusionList, const string& kernel, int forceGroup) {
@@ -158,6 +162,19 @@ void OpenCLNonbondedUtilities::addArgument(const ParameterInfo& parameter) {
     arguments.push_back(parameter);
 }
 
+string OpenCLNonbondedUtilities::addEnergyParameterDerivative(const string& param) {
+    // See if the parameter has already been added.
+    
+    int index;
+    for (index = 0; index < energyParameterDerivatives.size(); index++)
+        if (param == energyParameterDerivatives[index])
+            break;
+    if (index == energyParameterDerivatives.size())
+        energyParameterDerivatives.push_back(param);
+    context.addEnergyParameterDerivative(param);
+    return string("energyParamDeriv")+context.intToString(index);
+}
+
 void OpenCLNonbondedUtilities::requestExclusions(const vector<vector<int> >& exclusionList) {
     if (anyExclusions) {
         bool sameExclusions = (exclusionList.size() == atomExclusions.size());
@@ -357,8 +374,6 @@ void OpenCLNonbondedUtilities::prepareInteractions(int forceGroups) {
 
     if (lastCutoff != kernels.cutoffDistance)
         forceRebuildNeighborList = true;
-    bool rebuild = false;
-    do {
     setPeriodicBoxArgs(context, kernels.findBlockBoundsKernel, 1);
     context.executeKernel(kernels.findBlockBoundsKernel, context.getNumAtoms());
     blockSorter->sort(*sortedBlocks);
@@ -367,10 +382,8 @@ void OpenCLNonbondedUtilities::prepareInteractions(int forceGroups) {
     setPeriodicBoxArgs(context, kernels.findInteractingBlocksKernel, 0);
     context.executeKernel(kernels.findInteractingBlocksKernel, context.getNumAtoms(), interactingBlocksThreadBlockSize);
     forceRebuildNeighborList = false;
-        if (context.getComputeForceCount() == 1)
-            rebuild = updateNeighborListSize(); // This is the first time step, so check whether our initial guess was large enough.
-    } while (rebuild);
     lastCutoff = kernels.cutoffDistance;
+    context.getQueue().enqueueReadBuffer(interactionCount->getDeviceBuffer(), CL_FALSE, 0, sizeof(int), pinnedCountMemory, NULL, &downloadCountEvent); 
 }
 
 void OpenCLNonbondedUtilities::computeInteractions(int forceGroups, bool includeForces, bool includeEnergy) {
@@ -385,20 +398,22 @@ void OpenCLNonbondedUtilities::computeInteractions(int forceGroups, bool include
             setPeriodicBoxArgs(context, kernel, 9);
         context.executeKernel(kernel, numForceThreadBlocks*forceThreadBlockSize, forceThreadBlockSize);
     }
+    if (useCutoff && numTiles > 0) {
+        downloadCountEvent.wait();
+        updateNeighborListSize();
+    }
 }
 
 bool OpenCLNonbondedUtilities::updateNeighborListSize() {
     if (!useCutoff)
         return false;
-    unsigned int* pinnedInteractionCount = (unsigned int*) context.getPinnedBuffer();
-    interactionCount->download(pinnedInteractionCount);
-    if (pinnedInteractionCount[0] <= (unsigned int) interactingTiles->getSize())
+    if (pinnedCountMemory[0] <= (unsigned int) interactingTiles->getSize())
         return false;
 
     // The most recent timestep had too many interactions to fit in the arrays.  Make the arrays bigger to prevent
     // this from happening in the future.
 
-    int maxTiles = (int) (1.2*pinnedInteractionCount[0]);
+    int maxTiles = (int) (1.2*pinnedCountMemory[0]);
     int totalTiles = context.getNumAtomBlocks()*(context.getNumAtomBlocks()+1)/2;
     if (maxTiles > totalTiles)
         maxTiles = totalTiles;
@@ -430,6 +445,7 @@ bool OpenCLNonbondedUtilities::updateNeighborListSize() {
         kernels.findInteractingBlocksKernel.setArg<cl_uint>(9, maxTiles);
     }
     forceRebuildNeighborList = true;
+    context.setForcesValid(false);
     return true;
 }
 
@@ -588,6 +604,8 @@ cl::Kernel OpenCLNonbondedUtilities::createInteractionKernel(const string& sourc
             args << arguments[i].getName();
         }
     }
+    if (energyParameterDerivatives.size() > 0)
+        args << ", __global mixed* restrict energyParamDerivs";
     replacements["PARAMETER_ARGUMENTS"] = args.str();
     stringstream loadLocal1;
     for (int i = 0; i < (int) params.size(); i++) {
@@ -638,6 +656,18 @@ cl::Kernel OpenCLNonbondedUtilities::createInteractionKernel(const string& sourc
         }
     }
     replacements["LOAD_ATOM2_PARAMETERS"] = load2j.str();
+    stringstream initDerivs;
+    for (int i = 0; i < energyParameterDerivatives.size(); i++)
+        initDerivs<<"mixed energyParamDeriv"<<i<<" = 0;\n";
+    replacements["INIT_DERIVATIVES"] = initDerivs.str();
+    stringstream saveDerivs;
+    const vector<string>& allParamDerivNames = context.getEnergyParamDerivNames();
+    int numDerivs = allParamDerivNames.size();
+    for (int i = 0; i < energyParameterDerivatives.size(); i++)
+        for (int index = 0; index < numDerivs; index++)
+            if (allParamDerivNames[index] == energyParameterDerivatives[i])
+                saveDerivs<<"energyParamDerivs[get_global_id(0)*"<<numDerivs<<"+"<<index<<"] += energyParamDeriv"<<i<<";\n";
+    replacements["SAVE_DERIVATIVES"] = saveDerivs.str();
     map<string, string> defines;
     if (useCutoff)
         defines["USE_CUTOFF"] = "1";
@@ -713,5 +743,7 @@ cl::Kernel OpenCLNonbondedUtilities::createInteractionKernel(const string& sourc
     for (int i = 0; i < (int) arguments.size(); i++) {
         kernel.setArg<cl::Memory>(index++, arguments[i].getMemory());
     }
+    if (energyParameterDerivatives.size() > 0)
+        kernel.setArg<cl::Memory>(index++, context.getEnergyParamDerivBuffer().getDeviceBuffer());
     return kernel;
 }

platforms/opencl/src/kernels/coulombLennardJones.cl

@@ -30,6 +30,10 @@
             tempForce = -prefactor*(erfAlphaR-alphaR*expAlphaRSqr*TWO_OVER_SQRT_PI);
             tempEnergy += -prefactor*erfAlphaR;
         }
+        else {
+            includeInteraction = false;
+            tempEnergy -= TWO_OVER_SQRT_PI*EWALD_ALPHA*138.935456f*posq1.w*posq2.w;
+        }
     }
     else {
 #if HAS_LENNARD_JONES

platforms/opencl/src/kernels/customCentroidBond.cl

@@ -116,10 +116,12 @@ __kernel void computeGroupForces(__global long* restrict groupForce, __global mi
         __global const int* restrict bondGroups, real4 periodicBoxSize, real4 invPeriodicBoxSize, real4 periodicBoxVecX, real4 periodicBoxVecY, real4 periodicBoxVecZ
         EXTRA_ARGS) {
     mixed energy = 0;
+    INIT_PARAM_DERIVS
     for (int index = get_global_id(0); index < NUM_BONDS; index += get_global_size(0)) {
         COMPUTE_FORCE
     }
     energyBuffer[get_global_id(0)] += energy;
+    SAVE_PARAM_DERIVS
 }
 
 /**

platforms/opencl/src/kernels/customGBEnergyPerParticle.cl

@@ -12,6 +12,7 @@
 __kernel void computePerParticleEnergy(int bufferSize, int numBuffers, __global real4* restrict forceBuffers, __global mixed* restrict energyBuffer, __global const real4* restrict posq
         PARAMETER_ARGUMENTS) {
     mixed energy = 0;
+    INIT_PARAM_DERIVS
     unsigned int index = get_global_id(0);
     while (index < NUM_ATOMS) {
         // Reduce the derivatives
@@ -27,4 +28,5 @@ __kernel void computePerParticleEnergy(int bufferSize, int numBuffers, __global
         index += get_global_size(0);
     }
     energyBuffer[get_global_id(0)] += energy;
+    SAVE_PARAM_DERIVS
 }