Optimization to OpenCL platform: use tabulated values for erfc()

platforms/opencl/src/OpenCLContext.cpp

@@ -214,8 +214,8 @@ void OpenCLContext::clearBuffer(OpenCLArray<mm_float4>& array) {
     clearBuffer(array.getDeviceBuffer(), array.getSize()*4);
 }
 
-void OpenCLContext::clearBuffer(cl::Buffer& buffer, int size) {
-    clearBufferKernel.setArg<cl::Buffer>(0, buffer);
+void OpenCLContext::clearBuffer(cl::Memory& memory, int size) {
+    clearBufferKernel.setArg<cl::Memory>(0, memory);
     clearBufferKernel.setArg<cl_int>(1, size);
     executeKernel(clearBufferKernel, size);
 }

platforms/opencl/src/OpenCLContext.h

@@ -230,10 +230,10 @@ public:
     /**
      * Set all elements of an array to 0.
      *
-     * @param buffer     the Buffer to clear
+     * @param memory     the Memory to clear
      * @param size       the number of float elements in the buffer
      */
-    void clearBuffer(cl::Buffer& buffer, int size);
+    void clearBuffer(cl::Memory& memory, int size);
     /**
      * Given a collection of buffers packed into an array, sum them and store
      * the sum in the first buffer.

platforms/opencl/src/OpenCLKernels.cpp

@@ -35,10 +35,11 @@
 #include "OpenCLIntegrationUtilities.h"
 #include "OpenCLNonbondedUtilities.h"
 #include "OpenCLKernelSources.h"
+#include "lepton/Operation.h"
 #include "lepton/Parser.h"
 #include "lepton/ParsedExpression.h"
 #include "../src/SimTKUtilities/SimTKOpenMMRealType.h"
-#include "lepton/Operation.h"
+#include "openmm/internal/MSVC_erfc.h"
 #include <cmath>
 #include <set>
 
@@ -390,7 +391,7 @@ void OpenCLCalcCustomBondForceKernel::executeForces(ContextImpl& context) {
             kernel.setArg<cl::Buffer>(nextIndex++, globals->getDeviceBuffer());
         for (int i = 0; i < (int) params->getBuffers().size(); i++) {
             const OpenCLNonbondedUtilities::ParameterInfo& buffer = params->getBuffers()[i];
-            kernel.setArg<cl::Buffer>(nextIndex++, buffer.getBuffer());
+            kernel.setArg<cl::Memory>(nextIndex++, buffer.getMemory());
         }
     }
     cl.executeKernel(kernel, numBonds);
@@ -625,7 +626,7 @@ void OpenCLCalcCustomAngleForceKernel::executeForces(ContextImpl& context) {
             kernel.setArg<cl::Buffer>(nextIndex++, globals->getDeviceBuffer());
         for (int i = 0; i < (int) params->getBuffers().size(); i++) {
             const OpenCLNonbondedUtilities::ParameterInfo& buffer = params->getBuffers()[i];
-            kernel.setArg<cl::Buffer>(nextIndex++, buffer.getBuffer());
+            kernel.setArg<cl::Memory>(nextIndex++, buffer.getMemory());
         }
     }
     cl.executeKernel(kernel, numAngles);
@@ -947,7 +948,7 @@ void OpenCLCalcCustomTorsionForceKernel::executeForces(ContextImpl& context) {
             kernel.setArg<cl::Buffer>(nextIndex++, globals->getDeviceBuffer());
         for (int i = 0; i < (int) params->getBuffers().size(); i++) {
             const OpenCLNonbondedUtilities::ParameterInfo& buffer = params->getBuffers()[i];
-            kernel.setArg<cl::Buffer>(nextIndex++, buffer.getBuffer());
+            kernel.setArg<cl::Memory>(nextIndex++, buffer.getMemory());
         }
     }
     cl.executeKernel(kernel, numTorsions);
@@ -1015,6 +1016,8 @@ OpenCLCalcNonbondedForceKernel::~OpenCLCalcNonbondedForceKernel() {
         delete pmeAtomRange;
     if (pmeAtomGridIndex != NULL)
         delete pmeAtomGridIndex;
+    if (erfcTable != NULL)
+        delete erfcTable;
     if (sort != NULL)
         delete sort;
     if (fft != NULL)
@@ -1079,16 +1082,15 @@ void OpenCLCalcNonbondedForceKernel::initialize(const System& system, const Nonb
         defines["REACTION_FIELD_K"] = doubleToString(reactionFieldK);
         defines["REACTION_FIELD_C"] = doubleToString(reactionFieldC);
     }
+    double alpha = 0;
     if (force.getNonbondedMethod() == NonbondedForce::Ewald) {
         // Compute the Ewald parameters.
 
-        double alpha;
         int kmaxx, kmaxy, kmaxz;
         NonbondedForceImpl::calcEwaldParameters(system, force, alpha, kmaxx, kmaxy, kmaxz);
         defines["EWALD_ALPHA"] = doubleToString(alpha);
         defines["TWO_OVER_SQRT_PI"] = doubleToString(2.0/sqrt(M_PI));
         defines["USE_EWALD"] = "1";
-        double selfEnergyScale = ONE_4PI_EPS0*alpha/std::sqrt(M_PI);
         ewaldSelfEnergy = -ONE_4PI_EPS0*alpha*sumSquaredCharges/std::sqrt(M_PI);
 
         // Create the reciprocal space kernels.
@@ -1111,7 +1113,6 @@ void OpenCLCalcNonbondedForceKernel::initialize(const System& system, const Nonb
     else if (force.getNonbondedMethod() == NonbondedForce::PME) {
         // Compute the PME parameters.
 
-        double alpha;
         int gridSizeX, gridSizeY, gridSizeZ;
         NonbondedForceImpl::calcPMEParameters(system, force, alpha, gridSizeX, gridSizeY, gridSizeZ);
         gridSizeX = OpenCLFFT3D::findLegalDimension(gridSizeX);
@@ -1120,7 +1121,6 @@ void OpenCLCalcNonbondedForceKernel::initialize(const System& system, const Nonb
         defines["EWALD_ALPHA"] = doubleToString(alpha);
         defines["TWO_OVER_SQRT_PI"] = doubleToString(2.0/sqrt(M_PI));
         defines["USE_EWALD"] = "1";
-        double selfEnergyScale = ONE_4PI_EPS0*alpha/std::sqrt(M_PI);
         ewaldSelfEnergy = -ONE_4PI_EPS0*alpha*sumSquaredCharges/std::sqrt(M_PI);
         pmeDefines["PME_ORDER"] = intToString(PmeOrder);
         pmeDefines["NUM_ATOMS"] = intToString(numParticles);
@@ -1205,6 +1205,37 @@ void OpenCLCalcNonbondedForceKernel::initialize(const System& system, const Nonb
     }
     else
         ewaldSelfEnergy = 0.0;
+    
+    // Tabulate values of erfc().
+
+    if (force.getNonbondedMethod() == NonbondedForce::Ewald || force.getNonbondedMethod() == NonbondedForce::PME) {
+        if (cl.getDevice().getInfo<CL_DEVICE_IMAGE_SUPPORT>()) {
+            try
+            {
+                const int tableSize = 2048;
+                defines["USE_TABULATED_ERFC"] = "1";
+                defines["ERFC_TABLE_SCALE"] = doubleToString((tableSize-1)/(alpha*force.getCutoffDistance()));
+                erfcTable = new cl::Image2D(cl.getContext(), CL_MEM_READ_ONLY, cl::ImageFormat(CL_INTENSITY, CL_FLOAT), tableSize, 1, 0);
+                vector<cl_float> erfcVector(tableSize);
+                for (int i = 0; i < tableSize; ++i)
+                    erfcVector[i] = (float) erfc(i*(alpha*force.getCutoffDistance())/(tableSize-1));
+                cl::size_t<3> origin, region;
+                origin.push_back(0);
+                origin.push_back(0);
+                origin.push_back(0);
+                region.push_back(tableSize);
+                region.push_back(1);
+                region.push_back(1);
+                cl.getQueue().enqueueWriteImage(*erfcTable, CL_TRUE, origin, region, 0, 0, &erfcVector[0]);
+                cl.getNonbondedUtilities().addArgument(OpenCLNonbondedUtilities::ParameterInfo("erfcTable", "image2d_t", sizeof(cl_float), *erfcTable));
+            }
+            catch (cl::Error err) {
+                std::stringstream str;
+                str<<"Error creating erfc() image: "<<err.what()<<" ("<<err.err()<<")";
+                throw OpenMMException(str.str());
+            }
+        }
+    }
 
     // Add the interaction to the default nonbonded kernel.
     
@@ -1466,7 +1497,7 @@ void OpenCLCalcCustomNonbondedForceKernel::initialize(const System& system, cons
     cl.getNonbondedUtilities().addInteraction(useCutoff, usePeriodic, true, force.getCutoffDistance(), exclusionList, source);
     for (int i = 0; i < (int) params->getBuffers().size(); i++) {
         const OpenCLNonbondedUtilities::ParameterInfo& buffer = params->getBuffers()[i];
-        cl.getNonbondedUtilities().addParameter(OpenCLNonbondedUtilities::ParameterInfo(prefix+"params"+intToString(i+1), buffer.getType(), buffer.getSize(), buffer.getBuffer()));
+        cl.getNonbondedUtilities().addParameter(OpenCLNonbondedUtilities::ParameterInfo(prefix+"params"+intToString(i+1), buffer.getType(), buffer.getSize(), buffer.getMemory()));
     }
     if (globals != NULL) {
         globals->upload(globalParamValues);
@@ -2149,17 +2180,17 @@ void OpenCLCalcCustomGBForceKernel::initialize(const System& system, const Custo
         for (int i = 0; i < (int) params->getBuffers().size(); i++) {
             const OpenCLNonbondedUtilities::ParameterInfo& buffer = params->getBuffers()[i];
             string paramName = prefix+"params"+intToString(i+1);
-            cl.getNonbondedUtilities().addParameter(OpenCLNonbondedUtilities::ParameterInfo(paramName, buffer.getType(), buffer.getSize(), buffer.getBuffer()));
+            cl.getNonbondedUtilities().addParameter(OpenCLNonbondedUtilities::ParameterInfo(paramName, buffer.getType(), buffer.getSize(), buffer.getMemory()));
         }
         for (int i = 0; i < (int) computedValues->getBuffers().size(); i++) {
             const OpenCLNonbondedUtilities::ParameterInfo& buffer = computedValues->getBuffers()[i];
             string paramName = prefix+"values"+intToString(i+1);
-            cl.getNonbondedUtilities().addParameter(OpenCLNonbondedUtilities::ParameterInfo(paramName, buffer.getType(), buffer.getSize(), buffer.getBuffer()));
+            cl.getNonbondedUtilities().addParameter(OpenCLNonbondedUtilities::ParameterInfo(paramName, buffer.getType(), buffer.getSize(), buffer.getMemory()));
         }
         for (int i = 0; i < (int) energyDerivs->getBuffers().size(); i++) {
             const OpenCLNonbondedUtilities::ParameterInfo& buffer = energyDerivs->getBuffers()[i];
             string paramName = prefix+"dEdV"+intToString(i+1);
-            cl.getNonbondedUtilities().addParameter(OpenCLNonbondedUtilities::ParameterInfo(paramName, buffer.getType(), buffer.getSize(), buffer.getBuffer()));
+            cl.getNonbondedUtilities().addParameter(OpenCLNonbondedUtilities::ParameterInfo(paramName, buffer.getType(), buffer.getSize(), buffer.getMemory()));
         }
         if (globals != NULL) {
             globals->upload(globalParamValues);
@@ -2195,7 +2226,7 @@ void OpenCLCalcCustomGBForceKernel::executeForces(ContextImpl& context) {
             pairValueKernel.setArg<cl::Buffer>(index++, globals->getDeviceBuffer());
         for (int i = 0; i < (int) params->getBuffers().size(); i++) {
             const OpenCLNonbondedUtilities::ParameterInfo& buffer = params->getBuffers()[i];
-            pairValueKernel.setArg<cl::Buffer>(index++, buffer.getBuffer());
+            pairValueKernel.setArg<cl::Memory>(index++, buffer.getMemory());
             pairValueKernel.setArg(index++, OpenCLContext::ThreadBlockSize*buffer.getSize(), NULL);
         }
         if (tabulatedFunctionParams != NULL) {
@@ -2210,9 +2241,9 @@ void OpenCLCalcCustomGBForceKernel::executeForces(ContextImpl& context) {
         if (globals != NULL)
             perParticleValueKernel.setArg<cl::Buffer>(index++, globals->getDeviceBuffer());
         for (int i = 0; i < (int) params->getBuffers().size(); i++)
-            perParticleValueKernel.setArg<cl::Buffer>(index++, params->getBuffers()[i].getBuffer());
+            perParticleValueKernel.setArg<cl::Memory>(index++, params->getBuffers()[i].getMemory());
         for (int i = 0; i < (int) computedValues->getBuffers().size(); i++)
-            perParticleValueKernel.setArg<cl::Buffer>(index++, computedValues->getBuffers()[i].getBuffer());
+            perParticleValueKernel.setArg<cl::Memory>(index++, computedValues->getBuffers()[i].getMemory());
         if (tabulatedFunctionParams != NULL) {
             for (int i = 0; i < (int) tabulatedFunctions.size(); i++)
                 perParticleValueKernel.setArg<cl::Buffer>(index++, tabulatedFunctions[i]->getDeviceBuffer());
@@ -2240,17 +2271,17 @@ void OpenCLCalcCustomGBForceKernel::executeForces(ContextImpl& context) {
             pairEnergyKernel.setArg<cl::Buffer>(index++, globals->getDeviceBuffer());
         for (int i = 0; i < (int) params->getBuffers().size(); i++) {
             const OpenCLNonbondedUtilities::ParameterInfo& buffer = params->getBuffers()[i];
-            pairEnergyKernel.setArg<cl::Buffer>(index++, buffer.getBuffer());
+            pairEnergyKernel.setArg<cl::Memory>(index++, buffer.getMemory());
             pairEnergyKernel.setArg(index++, OpenCLContext::ThreadBlockSize*buffer.getSize(), NULL);
         }
         for (int i = 0; i < (int) computedValues->getBuffers().size(); i++) {
             const OpenCLNonbondedUtilities::ParameterInfo& buffer = computedValues->getBuffers()[i];
-            pairEnergyKernel.setArg<cl::Buffer>(index++, buffer.getBuffer());
+            pairEnergyKernel.setArg<cl::Memory>(index++, buffer.getMemory());
             pairEnergyKernel.setArg(index++, OpenCLContext::ThreadBlockSize*buffer.getSize(), NULL);
         }
         for (int i = 0; i < (int) energyDerivs->getBuffers().size(); i++) {
             const OpenCLNonbondedUtilities::ParameterInfo& buffer = energyDerivs->getBuffers()[i];
-            pairEnergyKernel.setArg<cl::Buffer>(index++, buffer.getBuffer());
+            pairEnergyKernel.setArg<cl::Memory>(index++, buffer.getMemory());
             pairEnergyKernel.setArg(index++, OpenCLContext::ThreadBlockSize*buffer.getSize(), NULL);
         }
         if (tabulatedFunctionParams != NULL) {
@@ -2265,11 +2296,11 @@ void OpenCLCalcCustomGBForceKernel::executeForces(ContextImpl& context) {
         if (globals != NULL)
             perParticleEnergyKernel.setArg<cl::Buffer>(index++, globals->getDeviceBuffer());
         for (int i = 0; i < (int) params->getBuffers().size(); i++)
-            perParticleEnergyKernel.setArg<cl::Buffer>(index++, params->getBuffers()[i].getBuffer());
+            perParticleEnergyKernel.setArg<cl::Memory>(index++, params->getBuffers()[i].getMemory());
         for (int i = 0; i < (int) computedValues->getBuffers().size(); i++)
-            perParticleEnergyKernel.setArg<cl::Buffer>(index++, computedValues->getBuffers()[i].getBuffer());
+            perParticleEnergyKernel.setArg<cl::Memory>(index++, computedValues->getBuffers()[i].getMemory());
         for (int i = 0; i < (int) energyDerivs->getBuffers().size(); i++)
-            perParticleEnergyKernel.setArg<cl::Buffer>(index++, energyDerivs->getBuffers()[i].getBuffer());
+            perParticleEnergyKernel.setArg<cl::Memory>(index++, energyDerivs->getBuffers()[i].getMemory());
         if (tabulatedFunctionParams != NULL) {
             for (int i = 0; i < (int) tabulatedFunctions.size(); i++)
                 perParticleEnergyKernel.setArg<cl::Buffer>(index++, tabulatedFunctions[i]->getDeviceBuffer());
@@ -2290,7 +2321,7 @@ void OpenCLCalcCustomGBForceKernel::executeForces(ContextImpl& context) {
     cl.clearBuffer(*valueBuffers);
     for (int i = 0; i < (int) energyDerivs->getBuffers().size(); i++) {
         const OpenCLNonbondedUtilities::ParameterInfo& buffer = energyDerivs->getBuffers()[i];
-        cl.clearBuffer(buffer.getBuffer(), buffer.getSize()*energyDerivs->getNumObjects()/sizeof(cl_float));
+        cl.clearBuffer(buffer.getMemory(), buffer.getSize()*energyDerivs->getNumObjects()/sizeof(cl_float));
     }
     cl.executeKernel(pairValueKernel, nb.getTiles().getSize()*OpenCLContext::TileSize);
     cl.executeKernel(perParticleValueKernel, cl.getPaddedNumAtoms());
@@ -2440,7 +2471,7 @@ void OpenCLCalcCustomExternalForceKernel::executeForces(ContextImpl& context) {
             kernel.setArg<cl::Buffer>(nextIndex++, globals->getDeviceBuffer());
         for (int i = 0; i < (int) params->getBuffers().size(); i++) {
             const OpenCLNonbondedUtilities::ParameterInfo& buffer = params->getBuffers()[i];
-            kernel.setArg<cl::Buffer>(nextIndex++, buffer.getBuffer());
+            kernel.setArg<cl::Memory>(nextIndex++, buffer.getMemory());
         }
     }
     cl.executeKernel(kernel, numParticles);
@@ -2905,11 +2936,11 @@ void OpenCLCalcCustomHbondForceKernel::executeForces(ContextImpl& context) {
             donorKernel.setArg<cl::Buffer>(index++, globals->getDeviceBuffer());
         for (int i = 0; i < (int) donorParams->getBuffers().size(); i++) {
             const OpenCLNonbondedUtilities::ParameterInfo& buffer = donorParams->getBuffers()[i];
-            donorKernel.setArg<cl::Buffer>(index++, buffer.getBuffer());
+            donorKernel.setArg<cl::Memory>(index++, buffer.getMemory());
         }
         for (int i = 0; i < (int) acceptorParams->getBuffers().size(); i++) {
             const OpenCLNonbondedUtilities::ParameterInfo& buffer = acceptorParams->getBuffers()[i];
-            donorKernel.setArg<cl::Buffer>(index++, buffer.getBuffer());
+            donorKernel.setArg<cl::Memory>(index++, buffer.getMemory());
         }
         if (tabulatedFunctionParams != NULL) {
             for (int i = 0; i < (int) tabulatedFunctions.size(); i++)
@@ -2929,11 +2960,11 @@ void OpenCLCalcCustomHbondForceKernel::executeForces(ContextImpl& context) {
             acceptorKernel.setArg<cl::Buffer>(index++, globals->getDeviceBuffer());
         for (int i = 0; i < (int) donorParams->getBuffers().size(); i++) {
             const OpenCLNonbondedUtilities::ParameterInfo& buffer = donorParams->getBuffers()[i];
-            acceptorKernel.setArg<cl::Buffer>(index++, buffer.getBuffer());
+            acceptorKernel.setArg<cl::Memory>(index++, buffer.getMemory());
         }
         for (int i = 0; i < (int) acceptorParams->getBuffers().size(); i++) {
             const OpenCLNonbondedUtilities::ParameterInfo& buffer = acceptorParams->getBuffers()[i];
-            acceptorKernel.setArg<cl::Buffer>(index++, buffer.getBuffer());
+            acceptorKernel.setArg<cl::Memory>(index++, buffer.getMemory());
         }
         if (tabulatedFunctionParams != NULL) {
             for (int i = 0; i < (int) tabulatedFunctions.size(); i++)

platforms/opencl/src/OpenCLKernels.h

@@ -438,7 +438,7 @@ public:
     OpenCLCalcNonbondedForceKernel(std::string name, const Platform& platform, OpenCLContext& cl, System& system) : CalcNonbondedForceKernel(name, platform),
             hasInitializedKernel(false), cl(cl), sigmaEpsilon(NULL), exceptionParams(NULL), exceptionIndices(NULL), cosSinSums(NULL), pmeGrid(NULL),
             pmeBsplineModuliX(NULL), pmeBsplineModuliY(NULL), pmeBsplineModuliZ(NULL), pmeBsplineTheta(NULL), pmeBsplineDtheta(NULL), pmeAtomRange(NULL),
-            pmeAtomGridIndex(NULL), sort(NULL), fft(NULL) {
+            pmeAtomGridIndex(NULL), erfcTable(NULL), sort(NULL), fft(NULL) {
     }
     ~OpenCLCalcNonbondedForceKernel();
     /**
@@ -478,6 +478,7 @@ private:
     OpenCLArray<mm_float2>* pmeAtomGridIndex;
     OpenCLSort<mm_float2>* sort;
     OpenCLFFT3D* fft;
+    cl::Image2D* erfcTable;
     cl::Kernel exceptionsKernel;
     cl::Kernel ewaldSumsKernel;
     cl::Kernel ewaldForcesKernel;

platforms/opencl/src/OpenCLNonbondedUtilities.cpp

@@ -294,13 +294,19 @@ cl::Kernel OpenCLNonbondedUtilities::createInteractionKernel(const string& sourc
         args << params[i].getName();
     }
     for (int i = 0; i < (int) arguments.size(); i++) {
-        if ((arguments[i].getBuffer().getInfo<CL_MEM_FLAGS>() & CL_MEM_READ_ONLY) == 0)
-            args << ", __global ";
-        else
-            args << ", __constant ";
-        args << arguments[i].getType();
-        args << "* ";
-        args << arguments[i].getName();
+        if (arguments[i].getMemory().getInfo<CL_MEM_TYPE>() == CL_MEM_OBJECT_IMAGE2D) {
+            args << ", __read_only image2d_t ";
+            args << arguments[i].getName();
+        }
+        else {
+            if ((arguments[i].getMemory().getInfo<CL_MEM_FLAGS>() & CL_MEM_READ_ONLY) == 0)
+                args << ", __global ";
+            else
+                args << ", __constant ";
+            args << arguments[i].getType();
+            args << "* ";
+            args << arguments[i].getName();
+        }
     }
     replacements["PARAMETER_ARGUMENTS"] = args.str();
     stringstream loadLocal1;
@@ -388,12 +394,12 @@ cl::Kernel OpenCLNonbondedUtilities::createInteractionKernel(const string& sourc
         kernel.setArg<cl_uint>(9, tiles->getSize());
     }
     for (int i = 0; i < (int) params.size(); i++) {
-        kernel.setArg<cl::Buffer>(i*2+paramBase, params[i].getBuffer());
+        kernel.setArg<cl::Memory>(i*2+paramBase, params[i].getMemory());
         kernel.setArg(i*2+paramBase+1, OpenCLContext::ThreadBlockSize*params[i].getSize(), NULL);
     }
     paramBase += 2*params.size();
     for (int i = 0; i < (int) arguments.size(); i++) {
-        kernel.setArg<cl::Buffer>(i+paramBase, arguments[i].getBuffer());
+        kernel.setArg<cl::Memory>(i+paramBase, arguments[i].getMemory());
     }
     return kernel;
 }

platforms/opencl/src/OpenCLNonbondedUtilities.h

@@ -231,10 +231,10 @@ public:
      * @param name      the name of the parameter
      * @param type      the data type of the parameter
      * @param size      the size of the parameter in bytes
-     * @param buffer    the buffer containing the parameter values
+     * @param memory    the memory containing the parameter values
      */
-    ParameterInfo(const std::string& name, const std::string& type, int size, cl::Buffer& buffer) :
-            name(name), type(type), size(size), buffer(&buffer) {
+    ParameterInfo(const std::string& name, const std::string& type, int size, cl::Memory& memory) :
+            name(name), type(type), size(size), memory(&memory) {
     }
     const std::string& getName() const {
         return name;
@@ -245,14 +245,14 @@ public:
     int getSize() const {
         return size;
     }
-    cl::Buffer& getBuffer() const {
-        return *buffer;
+    cl::Memory& getMemory() const {
+        return *memory;
     }
 private:
     std::string name;
     std::string type;
     int size;
-    cl::Buffer* buffer;
+    cl::Memory* memory;
 };
 
 } // namespace OpenMM

platforms/opencl/src/OpenCLParameterSet.cpp

@@ -67,7 +67,7 @@ OpenCLParameterSet::OpenCLParameterSet(OpenCLContext& context, int numParameters
 
 OpenCLParameterSet::~OpenCLParameterSet() {
     for (int i = 0; i < (int) buffers.size(); i++)
-        delete &buffers[i].getBuffer();
+        delete &buffers[i].getMemory();
 }
 
 void OpenCLParameterSet::getParameterValues(vector<vector<cl_float> >& values) const {
@@ -79,7 +79,7 @@ void OpenCLParameterSet::getParameterValues(vector<vector<cl_float> >& values) c
         for (int i = 0; i < (int) buffers.size(); i++) {
             if (buffers[i].getType() == "float4") {
                 vector<mm_float4> data(numObjects);
-                context.getQueue().enqueueReadBuffer(buffers[i].getBuffer(), CL_TRUE, 0, numObjects*buffers[i].getSize(), &data[0]);
+                context.getQueue().enqueueReadBuffer(reinterpret_cast<cl::Buffer&>(buffers[i].getMemory()), CL_TRUE, 0, numObjects*buffers[i].getSize(), &data[0]);
                 for (int j = 0; j < numObjects; j++) {
                     values[j][base] = data[j].x;
                     if (base+1 < numParameters)
@@ -93,7 +93,7 @@ void OpenCLParameterSet::getParameterValues(vector<vector<cl_float> >& values) c
             }
             else if (buffers[i].getType() == "float2") {
                 vector<mm_float2> data(numObjects);
-                context.getQueue().enqueueReadBuffer(buffers[i].getBuffer(), CL_TRUE, 0, numObjects*buffers[i].getSize(), &data[0]);
+                context.getQueue().enqueueReadBuffer(reinterpret_cast<cl::Buffer&>(buffers[i].getMemory()), CL_TRUE, 0, numObjects*buffers[i].getSize(), &data[0]);
                 for (int j = 0; j < numObjects; j++) {
                     values[j][base] = data[j].x;
                     if (base+1 < numParameters)
@@ -103,7 +103,7 @@ void OpenCLParameterSet::getParameterValues(vector<vector<cl_float> >& values) c
             }
             else if (buffers[i].getType() == "float") {
                 vector<cl_float> data(numObjects);
-                context.getQueue().enqueueReadBuffer(buffers[i].getBuffer(), CL_TRUE, 0, numObjects*buffers[i].getSize(), &data[0]);
+                context.getQueue().enqueueReadBuffer(reinterpret_cast<cl::Buffer&>(buffers[i].getMemory()), CL_TRUE, 0, numObjects*buffers[i].getSize(), &data[0]);
                 for (int j = 0; j < numObjects; j++)
                     values[j][base] = data[j];
             }
@@ -133,7 +133,7 @@ void OpenCLParameterSet::setParameterValues(const vector<vector<cl_float> >& val
                     if (base+3 < numParameters)
                         data[j].w = values[j][base+3];
                 }
-                context.getQueue().enqueueWriteBuffer(buffers[i].getBuffer(), CL_TRUE, 0, numObjects*buffers[i].getSize(), &data[0]);
+                context.getQueue().enqueueWriteBuffer(reinterpret_cast<cl::Buffer&>(buffers[i].getMemory()), CL_TRUE, 0, numObjects*buffers[i].getSize(), &data[0]);
                 base += 4;
             }
             else if (buffers[i].getType() == "float2") {
@@ -143,14 +143,14 @@ void OpenCLParameterSet::setParameterValues(const vector<vector<cl_float> >& val
                     if (base+1 < numParameters)
                         data[j].y = values[j][base+1];
                 }
-                context.getQueue().enqueueWriteBuffer(buffers[i].getBuffer(), CL_TRUE, 0, numObjects*buffers[i].getSize(), &data[0]);
+                context.getQueue().enqueueWriteBuffer(reinterpret_cast<cl::Buffer&>(buffers[i].getMemory()), CL_TRUE, 0, numObjects*buffers[i].getSize(), &data[0]);
                 base += 2;
             }
             else if (buffers[i].getType() == "float") {
                 vector<cl_float> data(numObjects);
                 for (int j = 0; j < numObjects; j++)
                     data[j] = values[j][base];
-                context.getQueue().enqueueWriteBuffer(buffers[i].getBuffer(), CL_TRUE, 0, numObjects*buffers[i].getSize(), &data[0]);
+                context.getQueue().enqueueWriteBuffer(reinterpret_cast<cl::Buffer&>(buffers[i].getMemory()), CL_TRUE, 0, numObjects*buffers[i].getSize(), &data[0]);
             }
             else
                 throw OpenMMException("Internal error: Unknown buffer type in OpenCLParameterSet");

platforms/opencl/src/kernels/coulombLennardJones.cl

@@ -3,7 +3,15 @@ bool needCorrection = isExcluded && atom1 != atom2 && atom1 < NUM_ATOMS && atom2
 if (!isExcluded || needCorrection) {
     const float prefactor = 138.935456f*posq1.w*posq2.w*invR;
     float alphaR = EWALD_ALPHA*r;
+#ifdef USE_TABULATED_ERFC
+    float normalized = ERFC_TABLE_SCALE*alphaR;
+    int tableIndex = (int) normalized;
+    float fract2 = normalized-tableIndex;
+    float fract1 = 1.0f-fract2;
+    float erfcAlphaR = fract1*read_imagef(erfcTable, sampler, (int2)(tableIndex, 0)).x + fract2*read_imagef(erfcTable, sampler, (int2)(tableIndex+1, 0)).x;
+#else
     float erfcAlphaR = erfc(alphaR);
+#endif
     float tempForce = 0.0f;
     if (needCorrection) {
         // Subtract off the part of this interaction that was included in the reciprocal space contribution.

platforms/opencl/src/kernels/nonbonded_default.cl

 #define TILE_SIZE 32
+const sampler_t sampler = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP_TO_EDGE | CLK_FILTER_NEAREST;
 
 /**
  * Compute nonbonded interactions.

platforms/opencl/src/kernels/nonbonded_nvidia.cl

 #define TILE_SIZE 32
+const sampler_t sampler = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP_TO_EDGE | CLK_FILTER_NEAREST;
 
 /**
  * Compute nonbonded interactions.