Merge github.com:leeping/openmm

platforms/opencl/CMakeLists.txt

@@ -16,7 +16,6 @@
 
 IF (APPLE)
     SET (CMAKE_OSX_DEPLOYMENT_TARGET "10.6")
-    SET (CMAKE_OSX_SYSROOT "/Developer/SDKs/MacOSX10.6.sdk")
 ENDIF (APPLE)
 
 set(OPENMM_BUILD_OPENCL_TESTS TRUE CACHE BOOL "Whether to build OpenCL test cases")
@@ -103,4 +102,9 @@ SET(CL_KERNELS_H ${CMAKE_CURRENT_BINARY_DIR}/src/${CL_SOURCE_CLASS}.h)
 SET(SOURCE_FILES ${SOURCE_FILES} ${CL_KERNELS_CPP} ${CL_KERNELS_H})
 INCLUDE_DIRECTORIES(BEFORE ${CMAKE_CURRENT_BINARY_DIR}/src)
 
+# Install headers
+
+FILE(GLOB CORE_HEADERS include/*.h)
+INSTALL_FILES(/include/openmm/opencl FILES ${CORE_HEADERS})
+
 SUBDIRS (sharedTarget)

platforms/opencl/src/OpenCLContext.h → platforms/opencl/include/OpenCLContext.h

@@ -158,6 +158,8 @@ public:
     class WorkTask;
     class WorkThread;
     class ReorderListener;
+    class ForcePreComputation;
+    class ForcePostComputation;
     static const int ThreadBlockSize;
     static const int TileSize;
     OpenCLContext(const System& system, int platformIndex, int deviceIndex, const std::string& precision, OpenCLPlatform::PlatformData& platformData);
@@ -554,6 +556,28 @@ public:
     std::vector<ReorderListener*>& getReorderListeners() {
         return reorderListeners;
     }
+    /**
+     * Add a pre-computation that should be called at the very start of force and energy evaluations.
+     * The OpenCLContext assumes ownership of the object, and deletes it when the context itself is deleted.
+     */
+    void addPreComputation(ForcePreComputation* computation);
+    /**
+     * Get the list of ForcePreComputations.
+     */
+    std::vector<ForcePreComputation*>& getPreComputations() {
+        return preComputations;
+    }
+    /**
+     * Add a post-computation that should be called at the very end of force and energy evaluations.
+     * The OpenCLContext assumes ownership of the object, and deletes it when the context itself is deleted.
+     */
+    void addPostComputation(ForcePostComputation* computation);
+    /**
+     * Get the list of ForcePostComputations.
+     */
+    std::vector<ForcePostComputation*>& getPostComputations() {
+        return postComputations;
+    }
     /**
      * 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
@@ -625,6 +649,8 @@ private:
     std::vector<cl::Memory*> autoclearBuffers;
     std::vector<int> autoclearBufferSizes;
     std::vector<ReorderListener*> reorderListeners;
+    std::vector<ForcePreComputation*> preComputations;
+    std::vector<ForcePostComputation*> postComputations;
     OpenCLIntegrationUtilities* integration;
     OpenCLExpressionUtilities* expression;
     OpenCLBondedUtilities* bonded;
@@ -686,7 +712,7 @@ private:
 
 /**
  * This abstract class defines a function to be executed whenever atoms get reordered.
- * Objects that need to know when reordering happens should create a reorderListener
+ * Objects that need to know when reordering happens should create a ReorderListener
  * and register it by calling addReorderListener().
  */
 class OpenCLContext::ReorderListener {
@@ -696,6 +722,39 @@ public:
     }
 };
 
+/**
+ * This abstract class defines a function to be executed at the very beginning of force and
+ * energy evaluation, before any other calculation has been done.  It is useful for operations
+ * that need to be performed at a nonstandard point in the process.  After creating a
+ * ForcePreComputation, register it by calling addForcePreComputation().
+ */
+class OpenCLContext::ForcePreComputation {
+public:
+    /**
+     * @param includeForce  true if forces should be computed
+     * @param includeEnergy true if potential energy should be computed
+     * @param groups        a set of bit flags for which force groups to include
+     */
+    virtual void computeForceAndEnergy(bool includeForces, bool includeEnergy, int groups) = 0;
+};
+
+/**
+ * This abstract class defines a function to be executed at the very end of force and
+ * energy evaluation, after all other calculations have been done.  It is useful for operations
+ * that need to be performed at a nonstandard point in the process.  After creating a
+ * ForcePostComputation, register it by calling addForcePostComputation().
+ */
+class OpenCLContext::ForcePostComputation {
+public:
+    /**
+     * @param includeForce  true if forces should be computed
+     * @param includeEnergy true if potential energy should be computed
+     * @param groups        a set of bit flags for which force groups to include
+     * @return an optional contribution to add to the potential energy.
+     */
+    virtual double computeForceAndEnergy(bool includeForces, bool includeEnergy, int groups) = 0;
+};
+
 } // namespace OpenMM
 
 #endif /*OPENMM_OPENCLCONTEXT_H_*/

platforms/opencl/src/OpenCLKernels.h → platforms/opencl/include/OpenCLKernels.h

@@ -557,7 +557,7 @@ public:
     OpenCLCalcNonbondedForceKernel(std::string name, const Platform& platform, OpenCLContext& cl, const System& system) : CalcNonbondedForceKernel(name, platform),
             hasInitializedKernel(false), cl(cl), sigmaEpsilon(NULL), exceptionParams(NULL), cosSinSums(NULL), pmeGrid(NULL),
             pmeGrid2(NULL), pmeBsplineModuliX(NULL), pmeBsplineModuliY(NULL), pmeBsplineModuliZ(NULL), pmeBsplineTheta(NULL),
-            pmeAtomRange(NULL), pmeAtomGridIndex(NULL), sort(NULL), fft(NULL) {
+            pmeAtomRange(NULL), pmeAtomGridIndex(NULL), sort(NULL), fft(NULL), pmeio(NULL) {
     }
     ~OpenCLCalcNonbondedForceKernel();
     /**
@@ -596,6 +596,9 @@ private:
         const char* getMaxValue() const {return "(int2) (INT_MAX, INT_MAX)";}
         const char* getSortKey() const {return "value.y";}
     };
+    class PmeIO;
+    class PmePreComputation;
+    class PmePostComputation;
     OpenCLContext& cl;
     bool hasInitializedKernel;
     OpenCLArray* sigmaEpsilon;
@@ -611,6 +614,8 @@ private:
     OpenCLArray* pmeAtomGridIndex;
     OpenCLSort* sort;
     OpenCLFFT3D* fft;
+    Kernel cpuPme;
+    PmeIO* pmeio;
     cl::Kernel ewaldSumsKernel;
     cl::Kernel ewaldForcesKernel;
     cl::Kernel pmeGridIndexKernel;

platforms/opencl/include/OpenCLPlatform.h

@@ -87,17 +87,25 @@ public:
         static const std::string key = "OpenCLPrecision";
         return key;
     }
+    /**
+     * This is the name of the parameter for selecting whether to use the CPU based PME calculation.
+     */
+    static const std::string& OpenCLUseCpuPme() {
+        static const std::string key = "OpenCLUseCpuPme";
+        return key;
+    }
 };
 
 class OPENMM_EXPORT_OPENCL OpenCLPlatform::PlatformData {
 public:
-    PlatformData(const System& system, const std::string& platformPropValue, const std::string& deviceIndexProperty, const std::string& precisionProperty);
+    PlatformData(const System& system, const std::string& platformPropValue, const std::string& deviceIndexProperty, const std::string& precisionProperty, const std::string& cpuPmeProperty);
     ~PlatformData();
     void initializeContexts(const System& system);
     void syncContexts();
+    ContextImpl* context;
     std::vector<OpenCLContext*> contexts;
     std::vector<double> contextEnergy;
-    bool removeCM;
+    bool removeCM, useCpuPme;
     int cmMotionFrequency;
     int stepCount, computeForceCount;
     double time;

platforms/opencl/src/OpenCLContext.cpp

@@ -334,6 +334,10 @@ OpenCLContext::~OpenCLContext() {
         delete forces[i];
     for (int i = 0; i < (int) reorderListeners.size(); i++)
         delete reorderListeners[i];
+    for (int i = 0; i < (int) preComputations.size(); i++)
+        delete preComputations[i];
+    for (int i = 0; i < (int) postComputations.size(); i++)
+        delete postComputations[i];
     if (pinnedBuffer != NULL)
         delete pinnedBuffer;
     if (posq != NULL)
@@ -1106,6 +1110,14 @@ void OpenCLContext::addReorderListener(ReorderListener* listener) {
     reorderListeners.push_back(listener);
 }
 
+void OpenCLContext::addPreComputation(ForcePreComputation* computation) {
+    preComputations.push_back(computation);
+}
+
+void OpenCLContext::addPostComputation(ForcePostComputation* computation) {
+    postComputations.push_back(computation);
+}
+
 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/OpenCLFFT3D.cpp

@@ -27,7 +27,7 @@
 #include "OpenCLFFT3D.h"
 #include "OpenCLExpressionUtilities.h"
 #include "OpenCLKernelSources.h"
-#include "../src/SimTKUtilities/SimTKOpenMMRealType.h"
+#include "SimTKOpenMMRealType.h"
 #include <map>
 #include <sstream>
 #include <string>

platforms/opencl/src/OpenCLKernels.cpp

@@ -44,8 +44,8 @@
 #include "lepton/Operation.h"
 #include "lepton/Parser.h"
 #include "lepton/ParsedExpression.h"
-#include "../src/SimTKUtilities/SimTKOpenMMRealType.h"
-#include "../src/SimTKUtilities/SimTKOpenMMUtilities.h"
+#include "SimTKOpenMMRealType.h"
+#include "SimTKOpenMMUtilities.h"
 #include <cmath>
 #include <set>
 
@@ -104,10 +104,12 @@ void OpenCLCalcForcesAndEnergyKernel::initialize(const System& system) {
 }
 
 void OpenCLCalcForcesAndEnergyKernel::beginComputation(ContextImpl& context, bool includeForces, bool includeEnergy, int groups) {
+    cl.clearAutoclearBuffers();
+    for (vector<OpenCLContext::ForcePreComputation*>::iterator iter = cl.getPreComputations().begin(); iter != cl.getPreComputations().end(); ++iter)
+        (*iter)->computeForceAndEnergy(includeForces, includeEnergy, groups);
     OpenCLNonbondedUtilities& nb = cl.getNonbondedUtilities();
     bool includeNonbonded = ((groups&(1<<nb.getForceGroup())) != 0);
     cl.setComputeForceCount(cl.getComputeForceCount()+1);
-    cl.clearAutoclearBuffers();
     if (includeNonbonded)
         nb.prepareInteractions();
 }
@@ -117,8 +119,10 @@ double OpenCLCalcForcesAndEnergyKernel::finishComputation(ContextImpl& context,
     if ((groups&(1<<cl.getNonbondedUtilities().getForceGroup())) != 0)
         cl.getNonbondedUtilities().computeInteractions();
     cl.reduceForces();
+    double sum = 0.0;
+    for (vector<OpenCLContext::ForcePostComputation*>::iterator iter = cl.getPostComputations().begin(); iter != cl.getPostComputations().end(); ++iter)
+        sum += (*iter)->computeForceAndEnergy(includeForces, includeEnergy, groups);
     cl.getIntegrationUtilities().distributeForcesFromVirtualSites();
-    double sum = 0.0f;
     if (includeEnergy) {
         OpenCLArray& energyArray = cl.getEnergyBuffer();
         if (cl.getUseDoublePrecision()) {
@@ -1323,6 +1327,58 @@ private:
     const NonbondedForce& force;
 };
 
+class OpenCLCalcNonbondedForceKernel::PmeIO : public CalcPmeReciprocalForceKernel::IO {
+public:
+    PmeIO(OpenCLContext& cl, cl::Kernel addForcesKernel) : cl(cl), addForcesKernel(addForcesKernel), forceTemp(NULL) {
+        forceTemp = OpenCLArray::create<mm_float4>(cl, cl.getNumAtoms(), "PmeForce");
+        addForcesKernel.setArg<cl::Buffer>(0, forceTemp->getDeviceBuffer());
+    }
+    ~PmeIO() {
+        if (forceTemp != NULL)
+            delete forceTemp;
+    }
+    float* getPosq() {
+        cl.getPosq().download(posq);
+        return (float*) &posq[0];
+    }
+    void setForce(float* force) {
+        forceTemp->upload(force);
+        addForcesKernel.setArg<cl::Buffer>(1, cl.getForce().getDeviceBuffer());
+        cl.executeKernel(addForcesKernel, cl.getNumAtoms());
+    }
+private:
+    OpenCLContext& cl;
+    vector<mm_float4> posq;
+    OpenCLArray* forceTemp;
+    cl::Kernel addForcesKernel;
+};
+
+class OpenCLCalcNonbondedForceKernel::PmePreComputation : public OpenCLContext::ForcePreComputation {
+public:
+    PmePreComputation(OpenCLContext& cl, Kernel& pme, CalcPmeReciprocalForceKernel::IO& io) : cl(cl), pme(pme), io(io) {
+    }
+    void computeForceAndEnergy(bool includeForces, bool includeEnergy, int groups) {
+        Vec3 boxSize(cl.getPeriodicBoxSize().x, cl.getPeriodicBoxSize().y, cl.getPeriodicBoxSize().z);
+        pme.getAs<CalcPmeReciprocalForceKernel>().beginComputation(io, boxSize, includeEnergy);
+    }
+private:
+    OpenCLContext& cl;
+    Kernel pme;
+    CalcPmeReciprocalForceKernel::IO& io;
+};
+
+class OpenCLCalcNonbondedForceKernel::PmePostComputation : public OpenCLContext::ForcePostComputation {
+public:
+    PmePostComputation(Kernel& pme, CalcPmeReciprocalForceKernel::IO& io) : pme(pme), io(io) {
+    }
+    double computeForceAndEnergy(bool includeForces, bool includeEnergy, int groups) {
+        return pme.getAs<CalcPmeReciprocalForceKernel>().finishComputation(io);
+    }
+private:
+    Kernel pme;
+    CalcPmeReciprocalForceKernel::IO& io;
+};
+
 OpenCLCalcNonbondedForceKernel::~OpenCLCalcNonbondedForceKernel() {
     if (sigmaEpsilon != NULL)
         delete sigmaEpsilon;
@@ -1350,6 +1406,8 @@ OpenCLCalcNonbondedForceKernel::~OpenCLCalcNonbondedForceKernel() {
         delete sort;
     if (fft != NULL)
         delete fft;
+    if (pmeio != NULL)
+        delete pmeio;
 }
 
 void OpenCLCalcNonbondedForceKernel::initialize(const System& system, const NonbondedForce& force) {
@@ -1430,7 +1488,7 @@ void OpenCLCalcNonbondedForceKernel::initialize(const System& system, const Nonb
     else
         dispersionCoefficient = 0.0;
     alpha = 0;
-    if (force.getNonbondedMethod() == NonbondedForce::Ewald) {
+    if (force.getNonbondedMethod() == NonbondedForce::Ewald && cl.getContextIndex() == 0) {
         // Compute the Ewald parameters.
 
         int kmaxx, kmaxy, kmaxz;
@@ -1438,7 +1496,7 @@ void OpenCLCalcNonbondedForceKernel::initialize(const System& system, const Nonb
         defines["EWALD_ALPHA"] = cl.doubleToString(alpha);
         defines["TWO_OVER_SQRT_PI"] = cl.doubleToString(2.0/sqrt(M_PI));
         defines["USE_EWALD"] = "1";
-        ewaldSelfEnergy = (cl.getContextIndex() == 0 ? -ONE_4PI_EPS0*alpha*sumSquaredCharges/sqrt(M_PI) : 0.0);
+        ewaldSelfEnergy = -ONE_4PI_EPS0*alpha*sumSquaredCharges/sqrt(M_PI);
 
         // Create the reciprocal space kernels.
 
@@ -1454,7 +1512,7 @@ void OpenCLCalcNonbondedForceKernel::initialize(const System& system, const Nonb
         int elementSize = (cl.getUseDoublePrecision() ? sizeof(mm_double2) : sizeof(mm_float2));
         cosSinSums = new OpenCLArray(cl, (2*kmaxx-1)*(2*kmaxy-1)*(2*kmaxz-1), elementSize, "cosSinSums");
     }
-    else if (force.getNonbondedMethod() == NonbondedForce::PME) {
+    else if (force.getNonbondedMethod() == NonbondedForce::PME && cl.getContextIndex() == 0) {
         // Compute the PME parameters.
 
         int gridSizeX, gridSizeY, gridSizeZ;
@@ -1465,7 +1523,7 @@ void OpenCLCalcNonbondedForceKernel::initialize(const System& system, const Nonb
         defines["EWALD_ALPHA"] = cl.doubleToString(alpha);
         defines["TWO_OVER_SQRT_PI"] = cl.doubleToString(2.0/sqrt(M_PI));
         defines["USE_EWALD"] = "1";
-        ewaldSelfEnergy = (cl.getContextIndex() == 0 ? -ONE_4PI_EPS0*alpha*sumSquaredCharges/sqrt(M_PI) : 0.0);
+        ewaldSelfEnergy = -ONE_4PI_EPS0*alpha*sumSquaredCharges/sqrt(M_PI);
         pmeDefines["PME_ORDER"] = cl.intToString(PmeOrder);
         pmeDefines["NUM_ATOMS"] = cl.intToString(numParticles);
         pmeDefines["RECIP_EXP_FACTOR"] = cl.doubleToString(M_PI*M_PI/(alpha*alpha));
@@ -1476,7 +1534,23 @@ void OpenCLCalcNonbondedForceKernel::initialize(const System& system, const Nonb
         bool deviceIsCpu = (cl.getDevice().getInfo<CL_DEVICE_TYPE>() == CL_DEVICE_TYPE_CPU);
         if (deviceIsCpu)
             pmeDefines["DEVICE_IS_CPU"] = "1";
+        if (cl.getPlatformData().useCpuPme) {
+            // Create the CPU PME kernel.
             
+            try {
+                cpuPme = getPlatform().createKernel(CalcPmeReciprocalForceKernel::Name(), *cl.getPlatformData().context);
+                cpuPme.getAs<CalcPmeReciprocalForceKernel>().initialize(gridSizeX, gridSizeY, gridSizeZ, numParticles, alpha);
+                cl::Program program = cl.createProgram(OpenCLKernelSources::pme, pmeDefines);
+                cl::Kernel addForcesKernel = cl::Kernel(program, "addForces");
+                pmeio = new PmeIO(cl, addForcesKernel);
+                cl.addPreComputation(new PmePreComputation(cl, cpuPme, *pmeio));
+                cl.addPostComputation(new PmePostComputation(cpuPme, *pmeio));
+            }
+            catch (OpenMMException& ex) {
+                // The CPU PME plugin isn't available.
+            }
+        }
+        if (pmeio == NULL) {
             // Create required data structures.
 
             int elementSize = (cl.getUseDoublePrecision() ? sizeof(double) : sizeof(float));
@@ -1565,6 +1639,7 @@ void OpenCLCalcNonbondedForceKernel::initialize(const System& system, const Nonb
                 }
             }
         }
+    }
     else
         ewaldSelfEnergy = 0.0;
 
@@ -1650,7 +1725,7 @@ double OpenCLCalcNonbondedForceKernel::execute(ContextImpl& context, bool includ
             }
        }
     }
-    if (cosSinSums != NULL && cl.getContextIndex() == 0 && includeReciprocal) {
+    if (cosSinSums != NULL && includeReciprocal) {
         mm_double4 boxSize = cl.getPeriodicBoxSizeDouble();
         mm_double4 recipBoxSize = mm_double4(2*M_PI/boxSize.x, 2*M_PI/boxSize.y, 2*M_PI/boxSize.z, 0.0);
         double recipCoefficient = ONE_4PI_EPS0*4*M_PI/(boxSize.x*boxSize.y*boxSize.z);
@@ -1669,7 +1744,7 @@ double OpenCLCalcNonbondedForceKernel::execute(ContextImpl& context, bool includ
         cl.executeKernel(ewaldSumsKernel, cosSinSums->getSize());
         cl.executeKernel(ewaldForcesKernel, cl.getNumAtoms());
     }
-    if (pmeGrid != NULL && cl.getContextIndex() == 0 && includeReciprocal) {
+    if (pmeGrid != NULL && includeReciprocal) {
         setPeriodicBoxSizeArg(cl, pmeUpdateBsplinesKernel, 4);
         setInvPeriodicBoxSizeArg(cl, pmeUpdateBsplinesKernel, 5);
         cl.executeKernel(pmeUpdateBsplinesKernel, cl.getNumAtoms());
@@ -4905,24 +4980,6 @@ void OpenCLIntegrateCustomStepKernel::prepareForComputation(ContextImpl& context
         defines["NUM_ATOMS"] = cl.intToString(cl.getNumAtoms());
         defines["WORK_GROUP_SIZE"] = cl.intToString(OpenCLContext::ThreadBlockSize);
         
-        // Initialize the random number generator.
-        
-        uniformRandoms = OpenCLArray::create<mm_float4>(cl, cl.getNumAtoms(), "uniformRandoms");
-        randomSeed = OpenCLArray::create<mm_int4>(cl, cl.getNumThreadBlocks()*OpenCLContext::ThreadBlockSize, "randomSeed");
-        vector<mm_int4> seed(randomSeed->getSize());
-        unsigned int r = integrator.getRandomNumberSeed()+1;
-        for (int i = 0; i < randomSeed->getSize(); i++) {
-            seed[i].x = r = (1664525*r + 1013904223) & 0xFFFFFFFF;
-            seed[i].y = r = (1664525*r + 1013904223) & 0xFFFFFFFF;
-            seed[i].z = r = (1664525*r + 1013904223) & 0xFFFFFFFF;
-            seed[i].w = r = (1664525*r + 1013904223) & 0xFFFFFFFF;
-        }
-        randomSeed->upload(seed);
-        cl::Program randomProgram = cl.createProgram(OpenCLKernelSources::customIntegrator, defines);
-        randomKernel = cl::Kernel(randomProgram, "generateRandomNumbers");
-        randomKernel.setArg<cl::Buffer>(0, uniformRandoms->getDeviceBuffer());
-        randomKernel.setArg<cl::Buffer>(1, randomSeed->getDeviceBuffer());
-        
         // Build a list of all variables that affect the forces, so we can tell which
         // steps invalidate them.
         
@@ -5013,10 +5070,10 @@ void OpenCLIntegrateCustomStepKernel::prepareForComputation(ContextImpl& context
         for (int step = 1; step < numSteps; step++) {
             if (needsForces[step] || needsEnergy[step])
                 continue;
-            if (stepType[step-1] == CustomIntegrator::ComputeGlobal && stepType[step] == CustomIntegrator::ComputeGlobal)
+            if (stepType[step-1] == CustomIntegrator::ComputeGlobal && stepType[step] == CustomIntegrator::ComputeGlobal &&
+                    !usesVariable(expression[step], "uniform") && !usesVariable(expression[step], "gaussian"))
                 merged[step] = true;
-            if (stepType[step-1] == CustomIntegrator::ComputePerDof && stepType[step] == CustomIntegrator::ComputePerDof &&
-                    !usesVariable(expression[step], "uniform"))
+            if (stepType[step-1] == CustomIntegrator::ComputePerDof && stepType[step] == CustomIntegrator::ComputePerDof)
                 merged[step] = true;
         }
         
@@ -5035,7 +5092,13 @@ void OpenCLIntegrateCustomStepKernel::prepareForComputation(ContextImpl& context
                 }
                 int numGaussian = 0, numUniform = 0;
                 for (int j = step; j < numSteps && (j == step || merged[j]); j++) {
+                    numGaussian += numAtoms*usesVariable(expression[j], "gaussian");
+                    numUniform += numAtoms*usesVariable(expression[j], "uniform");
                     compute << "{\n";
+                    if (numGaussian > 0)
+                        compute << "float4 gaussian = gaussianValues[gaussianIndex+index];\n";
+                    if (numUniform > 0)
+                        compute << "float4 uniform = uniformValues[uniformIndex+index];\n";
                     for (int i = 0; i < 3; i++)
                         compute << createPerDofComputation(stepType[j] == CustomIntegrator::ComputePerDof ? variable[j] : "", expression[j], i, integrator, forceName[j], energyName[j]);
                     if (variable[j] == "x") {
@@ -5058,9 +5121,11 @@ void OpenCLIntegrateCustomStepKernel::prepareForComputation(ContextImpl& context
                             compute << "perDofValues"<<cl.intToString(i+1)<<"[3*index+2] = perDofz"<<cl.intToString(i+1)<<";\n";
                         }
                     }
+                    if (numGaussian > 0)
+                        compute << "gaussianIndex += NUM_ATOMS;\n";
+                    if (numUniform > 0)
+                        compute << "uniformIndex += NUM_ATOMS;\n";
                     compute << "}\n";
-                    numGaussian += numAtoms*usesVariable(expression[j], "gaussian");
-                    numUniform += numAtoms*usesVariable(expression[j], "uniform");
                 }
                 map<string, string> replacements;
                 replacements["COMPUTE_STEP"] = compute.str();
@@ -5090,9 +5155,7 @@ void OpenCLIntegrateCustomStepKernel::prepareForComputation(ContextImpl& context
                 kernel.setArg<cl::Buffer>(index++, globalValues->getDeviceBuffer());
                 kernel.setArg<cl::Buffer>(index++, contextParameterValues->getDeviceBuffer());
                 kernel.setArg<cl::Buffer>(index++, sumBuffer->getDeviceBuffer());
-                kernel.setArg<cl::Buffer>(index++, integration.getRandom().getDeviceBuffer());
-                index++;
-                kernel.setArg<cl::Buffer>(index++, uniformRandoms->getDeviceBuffer());
+                index += 3;
                 kernel.setArg<cl::Buffer>(index++, potentialEnergy->getDeviceBuffer());
                 for (int i = 0; i < (int) perDofValues->getBuffers().size(); i++)
                     kernel.setArg<cl::Memory>(index++, perDofValues->getBuffers()[i].getMemory());
@@ -5154,6 +5217,28 @@ void OpenCLIntegrateCustomStepKernel::prepareForComputation(ContextImpl& context
             }
         }
         
+        // Initialize the random number generator.
+        
+        int maxUniformRandoms = 1;
+        for (int i = 0; i < (int) requiredUniform.size(); i++)
+            maxUniformRandoms = max(maxUniformRandoms, requiredUniform[i]);
+        uniformRandoms = OpenCLArray::create<mm_float4>(cl, maxUniformRandoms, "uniformRandoms");
+        randomSeed = OpenCLArray::create<mm_int4>(cl, cl.getNumThreadBlocks()*OpenCLContext::ThreadBlockSize, "randomSeed");
+        vector<mm_int4> seed(randomSeed->getSize());
+        unsigned int r = integrator.getRandomNumberSeed()+1;
+        for (int i = 0; i < randomSeed->getSize(); i++) {
+            seed[i].x = r = (1664525*r + 1013904223) & 0xFFFFFFFF;
+            seed[i].y = r = (1664525*r + 1013904223) & 0xFFFFFFFF;
+            seed[i].z = r = (1664525*r + 1013904223) & 0xFFFFFFFF;
+            seed[i].w = r = (1664525*r + 1013904223) & 0xFFFFFFFF;
+        }
+        randomSeed->upload(seed);
+        cl::Program randomProgram = cl.createProgram(OpenCLKernelSources::customIntegrator, defines);
+        randomKernel = cl::Kernel(randomProgram, "generateRandomNumbers");
+        randomKernel.setArg<cl_int>(0, maxUniformRandoms);
+        randomKernel.setArg<cl::Buffer>(1, uniformRandoms->getDeviceBuffer());
+        randomKernel.setArg<cl::Buffer>(2, randomSeed->getDeviceBuffer());
+        
         // Create the kernel for summing the potential energy.
 
         cl::Program program = cl.createProgram(OpenCLKernelSources::customIntegrator, defines);
@@ -5199,8 +5284,7 @@ void OpenCLIntegrateCustomStepKernel::prepareForComputation(ContextImpl& context
         kineticEnergyKernel.setArg<cl::Buffer>(index++, globalValues->getDeviceBuffer());
         kineticEnergyKernel.setArg<cl::Buffer>(index++, contextParameterValues->getDeviceBuffer());
         kineticEnergyKernel.setArg<cl::Buffer>(index++, sumBuffer->getDeviceBuffer());
-        kineticEnergyKernel.setArg<cl::Buffer>(index++, integration.getRandom().getDeviceBuffer());
-        kineticEnergyKernel.setArg<cl_uint>(index++, 0);
+        index += 2;
         kineticEnergyKernel.setArg<cl::Buffer>(index++, uniformRandoms->getDeviceBuffer());
         kineticEnergyKernel.setArg<cl::Buffer>(index++, potentialEnergy->getDeviceBuffer());
         for (int i = 0; i < (int) perDofValues->getBuffers().size(); i++)
@@ -5317,6 +5401,8 @@ void OpenCLIntegrateCustomStepKernel::execute(ContextImpl& context, CustomIntegr
         }
         if (stepType[i] == CustomIntegrator::ComputePerDof && !merged[i]) {
             kernels[i][0].setArg<cl_uint>(10, integration.prepareRandomNumbers(requiredGaussian[i]));
+            kernels[i][0].setArg<cl::Buffer>(9, integration.getRandom().getDeviceBuffer());
+            kernels[i][0].setArg<cl::Buffer>(11, uniformRandoms->getDeviceBuffer());
             if (requiredUniform[i] > 0)
                 cl.executeKernel(randomKernel, numAtoms);
             cl.executeKernel(kernels[i][0], numAtoms);
@@ -5328,6 +5414,8 @@ void OpenCLIntegrateCustomStepKernel::execute(ContextImpl& context, CustomIntegr
         }
         else if (stepType[i] == CustomIntegrator::ComputeSum) {
             kernels[i][0].setArg<cl_uint>(10, integration.prepareRandomNumbers(requiredGaussian[i]));
+            kernels[i][0].setArg<cl::Buffer>(9, integration.getRandom().getDeviceBuffer());
+            kernels[i][0].setArg<cl::Buffer>(11, uniformRandoms->getDeviceBuffer());
             if (requiredUniform[i] > 0)
                 cl.executeKernel(randomKernel, numAtoms);
             cl.clearBuffer(*sumBuffer);
@@ -5379,6 +5467,8 @@ double OpenCLIntegrateCustomStepKernel::computeKineticEnergy(ContextImpl& contex
         forcesAreValid = true;
     }
     cl.clearBuffer(*sumBuffer);
+    kineticEnergyKernel.setArg<cl::Buffer>(9, cl.getIntegrationUtilities().getRandom().getDeviceBuffer());
+    kineticEnergyKernel.setArg<cl_uint>(10, 0);
     cl.executeKernel(kineticEnergyKernel, cl.getNumAtoms());
     cl.executeKernel(sumKineticEnergyKernel, OpenCLContext::ThreadBlockSize, OpenCLContext::ThreadBlockSize);
     if (cl.getUseDoublePrecision() || cl.getUseMixedPrecision()) {

platforms/opencl/src/OpenCLPlatform.cpp

@@ -31,6 +31,7 @@
 #include "openmm/internal/ContextImpl.h"
 #include "openmm/Context.h"
 #include "openmm/System.h"
+#include <algorithm>
 #include <sstream>
 
 using namespace OpenMM;
@@ -78,11 +79,13 @@ OpenCLPlatform::OpenCLPlatform() {
     platformProperties.push_back(OpenCLPlatformIndex());
     platformProperties.push_back(OpenCLPlatformName());
     platformProperties.push_back(OpenCLPrecision());
+    platformProperties.push_back(OpenCLUseCpuPme());
     setPropertyDefaultValue(OpenCLDeviceIndex(), "");
     setPropertyDefaultValue(OpenCLDeviceName(), "");
     setPropertyDefaultValue(OpenCLPlatformIndex(), "");
     setPropertyDefaultValue(OpenCLPlatformName(), "");
     setPropertyDefaultValue(OpenCLPrecision(), "single");
+    setPropertyDefaultValue(OpenCLUseCpuPme(), "false");
 }
 
 double OpenCLPlatform::getSpeed() const {
@@ -112,7 +115,15 @@ void OpenCLPlatform::contextCreated(ContextImpl& context, const map<string, stri
             getPropertyDefaultValue(OpenCLDeviceIndex()) : properties.find(OpenCLDeviceIndex())->second);
     string precisionPropValue = (properties.find(OpenCLPrecision()) == properties.end() ?
             getPropertyDefaultValue(OpenCLPrecision()) : properties.find(OpenCLPrecision())->second);
-    context.setPlatformData(new PlatformData(context.getSystem(), platformPropValue, devicePropValue, precisionPropValue));
+    string cpuPmePropValue = (properties.find(OpenCLUseCpuPme()) == properties.end() ?
+            getPropertyDefaultValue(OpenCLUseCpuPme()) : properties.find(OpenCLUseCpuPme())->second);
+    transform(precisionPropValue.begin(), precisionPropValue.end(), precisionPropValue.begin(), ::tolower);
+    transform(cpuPmePropValue.begin(), cpuPmePropValue.end(), cpuPmePropValue.begin(), ::tolower);
+    vector<string> pmeKernelName;
+    pmeKernelName.push_back(CalcPmeReciprocalForceKernel::Name());
+    if (!supportsKernels(pmeKernelName))
+        cpuPmePropValue = "false";
+    context.setPlatformData(new PlatformData(context.getSystem(), platformPropValue, devicePropValue, precisionPropValue, cpuPmePropValue));
 }
 
 void OpenCLPlatform::contextDestroyed(ContextImpl& context) const {
@@ -121,7 +132,7 @@ void OpenCLPlatform::contextDestroyed(ContextImpl& context) const {
 }
 
 OpenCLPlatform::PlatformData::PlatformData(const System& system, const string& platformPropValue, const string& deviceIndexProperty,
-        const string& precisionProperty) : removeCM(false), stepCount(0), computeForceCount(0), time(0.0)  {
+        const string& precisionProperty, const string& cpuPmeProperty) : removeCM(false), stepCount(0), computeForceCount(0), time(0.0)  {
     int platformIndex = 0;
     if (platformPropValue.length() > 0)
         stringstream(platformPropValue) >> platformIndex;
@@ -150,6 +161,7 @@ OpenCLPlatform::PlatformData::PlatformData(const System& system, const string& p
         deviceIndex << contexts[i]->getDeviceIndex();
         deviceName << contexts[i]->getDevice().getInfo<CL_DEVICE_NAME>();
     }
+    useCpuPme = (cpuPmeProperty == "true" && !contexts[0]->getUseDoublePrecision());
     propertyValues[OpenCLPlatform::OpenCLDeviceIndex()] = deviceIndex.str();
     propertyValues[OpenCLPlatform::OpenCLDeviceName()] = deviceName.str();
     propertyValues[OpenCLPlatform::OpenCLPlatformIndex()] = contexts[0]->intToString(platformIndex);
@@ -157,6 +169,7 @@ OpenCLPlatform::PlatformData::PlatformData(const System& system, const string& p
     cl::Platform::get(&platforms);
     propertyValues[OpenCLPlatform::OpenCLPlatformName()] = platforms[platformIndex].getInfo<CL_PLATFORM_NAME>();
     propertyValues[OpenCLPlatform::OpenCLPrecision()] = precisionProperty;
+    propertyValues[OpenCLPlatform::OpenCLUseCpuPme()] = useCpuPme ? "true" : "false";
     contextEnergy.resize(contexts.size());
 }
 

platforms/opencl/src/kernels/customIntegrator.cl

@@ -52,10 +52,10 @@ __kernel void applyPositionDeltas(__global real4* restrict posq, __global real4*
     }
 }
 
-__kernel void generateRandomNumbers(__global float4* restrict random, __global uint4* restrict seed) {
+__kernel void generateRandomNumbers(int numValues, __global float4* restrict random, __global uint4* restrict seed) {
     uint4 state = seed[get_global_id(0)];
     unsigned int carry = 0;
-    for (int index = get_global_id(0); index < NUM_ATOMS; index += get_global_size(0)) {
+    for (int index = get_global_id(0); index < numValues; index += get_global_size(0)) {
         // Generate three uniform random numbers.
 
         state.x = state.x * 69069 + 1;