Bug fixes to OpenCL on CPU

platforms/opencl/src/OpenCLKernels.cpp

@@ -3239,14 +3239,23 @@ void OpenCLCalcCustomGBForceKernel::initialize(const System& system, const Custo
                 load2 << "real temp_" << derivName << "_1 = 0;\n";
                 load2 << "real temp_" << derivName << "_2 = 0;\n";
                 tempDerivs1 << derivName << " += temp_" << derivName << "_1;\n";
-                tempDerivs2 << "local_" << derivName << "[tbx+tj] += temp_" << derivName << "_2;\n";
+                if (deviceIsCpu)
+                    tempDerivs2 << "local_" << derivName << "[j] += temp_" << derivName << "_2;\n";
+                else
+                    tempDerivs2 << "local_" << derivName << "[tbx+tj] += temp_" << derivName << "_2;\n";
                 if (useLong) {
                     storeDeriv1 << "atom_add(&global_" << derivName << "[offset1], (long) (" << derivName << "*0x100000000));\n";
-                    storeDeriv2 << "atom_add(&global_" << derivName << "[offset2], (long) (local_" << derivName << "[get_local_id(0)]*0x100000000));\n";
+                    if (deviceIsCpu)
+                        storeDeriv2 << "atom_add(&global_" << derivName << "[offset2], (long) (local_" << derivName << "[tgx]*0x100000000));\n";
+                    else
+                        storeDeriv2 << "atom_add(&global_" << derivName << "[offset2], (long) (local_" << derivName << "[get_local_id(0)]*0x100000000));\n";
                 }
                 else {
                     storeDeriv1 << "global_" << derivName << "[offset1] += " << derivName << ";\n";
-                    storeDeriv2 << "global_" << derivName << "[offset2] += local_" << derivName << "[get_local_id(0)];\n";
+                    if (deviceIsCpu)
+                        storeDeriv2 << "global_" << derivName << "[offset2] += local_" << derivName << "[tgx];\n";
+                    else
+                        storeDeriv2 << "global_" << derivName << "[offset2] += local_" << derivName << "[get_local_id(0)];\n";
                 }
             }
         }

platforms/opencl/src/kernels/customGBValueN2_cpu.cl

@@ -87,11 +87,13 @@ __kernel void computeN2Value(__global const real4* restrict posq, __local real4*
                 // Write results.
 
 #ifdef SUPPORTS_64_BIT_ATOMICS
-                atom_add(&global_value[atom1], (long) (value*0x100000000));
+                unsigned int offset1 = atom1;
+                atom_add(&global_value[offset1], (long) (value*0x100000000));
 #else
-                unsigned int offset = atom1 + get_group_id(0)*PADDED_NUM_ATOMS;
-                global_value[offset] += value;
+                unsigned int offset1 = atom1 + get_group_id(0)*PADDED_NUM_ATOMS;
+                global_value[offset1] += value;
 #endif
+                STORE_PARAM_DERIVS1
             }
         }
         else {
@@ -274,11 +276,13 @@ __kernel void computeN2Value(__global const real4* restrict posq, __local real4*
                     // Write results for atom1.
 
 #ifdef SUPPORTS_64_BIT_ATOMICS
-                    atom_add(&global_value[atom1], (long) (value*0x100000000));
+                    unsigned int offset1 = atom1;
+                    atom_add(&global_value[offset1], (long) (value*0x100000000));
 #else
-                    unsigned int offset = atom1 + get_group_id(0)*PADDED_NUM_ATOMS;
-                    global_value[offset] += value;
+                    unsigned int offset1 = atom1 + get_group_id(0)*PADDED_NUM_ATOMS;
+                    global_value[offset1] += value;
 #endif
+                    STORE_PARAM_DERIVS1
                 }
             }
             else

tests/TestCustomGBForce.h

@@ -491,7 +491,7 @@ void testIllegalVariable() {
 void testEnergyParameterDerivatives() {
     // Create a box of particles.
     
-    const int numParticles = 30;
+    const int numParticles = 40;
     const int numParameters = 4;
     const double boxSize = 2.0;
     const double delta = 1e-3;