Optimizations to CustomGBForce

platforms/opencl/src/OpenCLKernels.h

@@ -687,6 +687,7 @@ private:
     std::vector<std::string> globalParamNames;
     std::vector<cl_float> globalParamValues;
     std::vector<OpenCLArray<mm_float4>*> tabulatedFunctions;
+    std::vector<bool> pairValueUsesParam, pairEnergyUsesParam, pairEnergyUsesValue;
     System& system;
     cl::Kernel pairValueKernel, perParticleValueKernel, pairEnergyKernel, perParticleEnergyKernel, gradientChainRuleKernel;
 };

platforms/opencl/src/OpenCLParameterSet.cpp

@@ -32,11 +32,12 @@
 using namespace OpenMM;
 using namespace std;
 
-OpenCLParameterSet::OpenCLParameterSet(OpenCLContext& context, int numParameters, int numObjects, const string& name) :
+OpenCLParameterSet::OpenCLParameterSet(OpenCLContext& context, int numParameters, int numObjects, const string& name, bool bufferPerParameter) :
             context(context), numParameters(numParameters), numObjects(numObjects), name(name) {
     int params = numParameters;
     int bufferCount = 0;
     try {
+        if (!bufferPerParameter) {
             while (params > 2) {
                 cl::Buffer* buf = new cl::Buffer(context.getContext(), CL_MEM_READ_WRITE, numObjects*sizeof(mm_float4));
                 std::stringstream name;
@@ -51,11 +52,13 @@ OpenCLParameterSet::OpenCLParameterSet(OpenCLContext& context, int numParameters
                 buffers.push_back(OpenCLNonbondedUtilities::ParameterInfo(name.str(), "float", 2, sizeof(mm_float2), *buf));
                 params -= 2;
             }
-        if (params > 0) {
+        }
+        while (params > 0) {
             cl::Buffer* buf = new cl::Buffer(context.getContext(), CL_MEM_READ_WRITE, numObjects*sizeof(cl_float));
             std::stringstream name;
             name << "param" << (++bufferCount);
             buffers.push_back(OpenCLNonbondedUtilities::ParameterInfo(name.str(), "float", 1, sizeof(cl_float), *buf));
+            params--;
         }
     }
     catch (cl::Error err) {
@@ -106,6 +109,7 @@ void OpenCLParameterSet::getParameterValues(vector<vector<cl_float> >& values) c
                 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];
+                base++;
             }
             else
                 throw OpenMMException("Internal error: Unknown buffer type in OpenCLParameterSet");
@@ -151,6 +155,7 @@ void OpenCLParameterSet::setParameterValues(const vector<vector<cl_float> >& val
                 for (int j = 0; j < numObjects; j++)
                     data[j] = values[j][base];
                 context.getQueue().enqueueWriteBuffer(reinterpret_cast<cl::Buffer&>(buffers[i].getMemory()), CL_TRUE, 0, numObjects*buffers[i].getSize(), &data[0]);
+                base++;
             }
             else
                 throw OpenMMException("Internal error: Unknown buffer type in OpenCLParameterSet");

platforms/opencl/src/OpenCLParameterSet.h

@@ -49,8 +49,10 @@ public:
      * @param numParameters    the number of parameters for each object
      * @param numObjects       the number of objects to store parameter values for
      * @param name             the name of the parameter set
+     * @param bufferPerParameter  if true, a separate cl::Buffer is created for each parameter.  If false,
+     *                            multiple parameters may be combined into a single buffer.
      */
-    OpenCLParameterSet(OpenCLContext& context, int numParameters, int numObjects, const std::string& name);
+    OpenCLParameterSet(OpenCLContext& context, int numParameters, int numObjects, const std::string& name, bool bufferPerParameter=false);
     ~OpenCLParameterSet();
     /**
      * Get the number of parameters.

platforms/opencl/src/kernels/customGBEnergyPerParticle.cl

-#define REDUCE_VALUE(NAME, TYPE) \
+#define REDUCE_VALUE(NAME, TYPE) {\
     TYPE sum = NAME[index]; \
     for (int i = index+bufferSize; i < totalSize; i += bufferSize) \
         sum += NAME[i]; \
-    NAME[index] = sum;
+    NAME[index] = sum; \
+}
 
 /**
  * Reduce the derivatives computed in the N^2 energy kernel, and compute all per-particle energy terms.