Fixed bug when running on a CPU device

platforms/cuda/src/CudaKernels.cpp

@@ -2271,7 +2271,7 @@ void CudaCalcCustomNonbondedForceKernel::initInteractionGroups(const CustomNonbo
         defines["USE_CUTOFF"] = "1";
     if (force.getNonbondedMethod() == CustomNonbondedForce::CutoffPeriodic)
         defines["USE_PERIODIC"] = "1";
-    defines["THREAD_BLOCK_SIZE"] = cu.intToString(cu.getNonbondedUtilities().getForceThreadBlockSize());
+    defines["LOCAL_MEMORY_SIZE"] = cu.intToString(max(32, cu.getNonbondedUtilities().getForceThreadBlockSize()));
     double cutoff = force.getCutoffDistance();
     defines["CUTOFF_SQUARED"] = cu.doubleToString(cutoff*cutoff);
     defines["PADDED_NUM_ATOMS"] = cu.intToString(cu.getPaddedNumAtoms());

platforms/cuda/src/kernels/customNonbondedGroups.cu

-#define WARPS_PER_GROUP (THREAD_BLOCK_SIZE/TILE_SIZE)
-
 typedef struct {
     real x, y, z;
     real q;
@@ -19,7 +17,7 @@ extern "C" __global__ void computeInteractionGroups(
     const unsigned int tgx = threadIdx.x & (TILE_SIZE-1); // index within the warp
     const unsigned int tbx = threadIdx.x - tgx;           // block warpIndex
     real energy = 0.0f;
-    __shared__ AtomData localData[THREAD_BLOCK_SIZE];
+    __shared__ AtomData localData[LOCAL_MEMORY_SIZE];
 
     const unsigned int startTile = FIRST_TILE+warp*(LAST_TILE-FIRST_TILE)/totalWarps;
     const unsigned int endTile = FIRST_TILE+(warp+1)*(LAST_TILE-FIRST_TILE)/totalWarps;
@@ -86,4 +84,4 @@ extern "C" __global__ void computeInteractionGroups(
         atomicAdd(&forceBuffers[atom2+2*PADDED_NUM_ATOMS], static_cast<unsigned long long>((long long) (localData[threadIdx.x].fz*0x100000000)));
     }
     energyBuffer[blockIdx.x*blockDim.x+threadIdx.x] += energy;
-}
\ No newline at end of file
+}

platforms/opencl/src/OpenCLKernels.cpp

@@ -2289,7 +2289,7 @@ void OpenCLCalcCustomNonbondedForceKernel::initInteractionGroups(const CustomNon
         defines["USE_CUTOFF"] = "1";
     if (force.getNonbondedMethod() == CustomNonbondedForce::CutoffPeriodic)
         defines["USE_PERIODIC"] = "1";
-    defines["THREAD_BLOCK_SIZE"] = cl.intToString(cl.getNonbondedUtilities().getForceThreadBlockSize());
+    defines["LOCAL_MEMORY_SIZE"] = cl.intToString(max(32, cl.getNonbondedUtilities().getForceThreadBlockSize()));
     double cutoff = force.getCutoffDistance();
     defines["CUTOFF_SQUARED"] = cl.doubleToString(cutoff*cutoff);
     defines["PADDED_NUM_ATOMS"] = cl.intToString(cl.getPaddedNumAtoms());

platforms/opencl/src/kernels/customNonbondedGroups.cl

@@ -2,8 +2,6 @@
 #pragma OPENCL EXTENSION cl_khr_int64_base_atomics : enable
 #endif
 
-#define WARPS_PER_GROUP (THREAD_BLOCK_SIZE/TILE_SIZE)
-
 typedef struct {
     real x, y, z;
     real q;
@@ -52,7 +50,7 @@ __kernel void computeInteractionGroups(
     const unsigned int tgx = get_local_id(0) & (TILE_SIZE-1); // index within the warp
     const unsigned int tbx = get_local_id(0) - tgx;           // block warpIndex
     real energy = 0.0f;
-    __local AtomData localData[THREAD_BLOCK_SIZE];
+    __local AtomData localData[LOCAL_MEMORY_SIZE];
 
     const unsigned int startTile = FIRST_TILE+warp*(LAST_TILE-FIRST_TILE)/totalWarps;
     const unsigned int endTile = FIRST_TILE+(warp+1)*(LAST_TILE-FIRST_TILE)/totalWarps;
@@ -127,4 +125,4 @@ __kernel void computeInteractionGroups(
 #endif
     }
     energyBuffer[get_global_id(0)] += energy;
-}
\ No newline at end of file
+}