Bug fix on AMD GPUs.

platforms/opencl/src/OpenCLNonbondedUtilities.cpp

@@ -461,13 +461,12 @@ cl::Kernel OpenCLNonbondedUtilities::createInteractionKernel(const string& sourc
     kernel.setArg<cl::Buffer>(index++, exclusionIndices->getDeviceBuffer());
     kernel.setArg<cl::Buffer>(index++, exclusionRowIndices->getDeviceBuffer());
     kernel.setArg(index++, (deviceIsCpu ? OpenCLContext::TileSize*localDataSize : OpenCLContext::ThreadBlockSize*localDataSize), NULL);
-    kernel.setArg(index++, 3*OpenCLContext::ThreadBlockSize*sizeof(cl_float), NULL);
+    kernel.setArg(index++, 4*OpenCLContext::ThreadBlockSize*sizeof(cl_float), NULL);
     if (useCutoff) {
         kernel.setArg<cl::Buffer>(index++, interactingTiles->getDeviceBuffer());
         kernel.setArg<cl::Buffer>(index++, interactionCount->getDeviceBuffer());
         index += 2; // The periodic box size arguments are set when the kernel is executed.
         kernel.setArg<cl_uint>(index++, interactingTiles->getSize());
-        if (context.getSIMDWidth() == 32 || deviceIsCpu)
         kernel.setArg<cl::Buffer>(index++, interactionFlags->getDeviceBuffer());
     }
     else {

platforms/opencl/src/kernels/nonbonded_default.cl

@@ -15,7 +15,7 @@ __kernel __attribute__((reqd_work_group_size(WORK_GROUP_SIZE, 1, 1)))
 void computeNonbonded(__global float4* forceBuffers, __global float* energyBuffer, __global float4* posq, __global unsigned int* exclusions,
         __global unsigned int* exclusionIndices, __global unsigned int* exclusionRowIndices, __local AtomData* localData, __local float4* tempBuffer,
 #ifdef USE_CUTOFF
-        __global ushort2* tiles, __global unsigned int* interactionCount, float4 periodicBoxSize, float4 invPeriodicBoxSize, unsigned int maxTiles
+        __global ushort2* tiles, __global unsigned int* interactionCount, float4 periodicBoxSize, float4 invPeriodicBoxSize, unsigned int maxTiles, __global unsigned int* interactionFlags
 #else
         unsigned int numTiles
 #endif