OpenCLNonbondedUtilities::OpenCLNonbondedUtilities(OpenCLContext& context) : context(context), cutoff(-1.0), useCutoff(false), anyExclusions(false), usePadding(true),
numForceBuffers(0), exclusionIndices(NULL), exclusionRowIndices(NULL), exclusionTiles(NULL), exclusions(NULL), interactingTiles(NULL), interactingAtoms(NULL),
interactionCount(NULL), blockCenter(NULL), blockBoundingBox(NULL), sortedBlocks(NULL), sortedBlockCenter(NULL), sortedBlockBoundingBox(NULL),
oldPositions(NULL), rebuildNeighborList(NULL), blockSorter(NULL), nonbondedForceGroup(0) {
// Decide how many thread blocks and force buffers to use.
deviceIsCpu = (context.getDevice().getInfo<CL_DEVICE_TYPE>() == CL_DEVICE_TYPE_CPU);
if (deviceIsCpu) {
numForceThreadBlocks = context.getNumThreadBlocks();
forceThreadBlockSize = 1;
numForceBuffers = numForceThreadBlocks;
}
else if (context.getSIMDWidth() == 32) {
if (context.getSupports64BitGlobalAtomics()) {
numForceThreadBlocks = 4*context.getDevice().getInfo<CL_DEVICE_MAX_COMPUTE_UNITS>();
forceThreadBlockSize = 256;
// Even though using longForceBuffer, still need a single forceBuffer for the reduceForces kernel to convert the long results into float4 which will be used by later kernels.
numForceBuffers = 1;
}
else {
numForceThreadBlocks = 3*context.getDevice().getInfo<CL_DEVICE_MAX_COMPUTE_UNITS>();
forceThreadBlockSize = 256;
numForceBuffers = numForceThreadBlocks*forceThreadBlockSize/OpenCLContext::TileSize;
}
}
else {
numForceThreadBlocks = context.getNumThreadBlocks();
forceThreadBlockSize = (context.getSIMDWidth() >= 32 ? OpenCLContext::ThreadBlockSize : 32);
if (context.getSupports64BitGlobalAtomics()) {
// Even though using longForceBuffer, still need a single forceBuffer for the reduceForces kernel to convert the long results into float4 which will be used by later kernels.
numForceBuffers = 1;
}
else {
numForceBuffers = numForceThreadBlocks*forceThreadBlockSize/OpenCLContext::TileSize;
}
}
}
OpenCLCompact::OpenCLCompact(OpenCLContext& context) : context(context), dgBlockCounts(NULL) {
dgBlockCounts = OpenCLArray::create<cl_uint>(context, context.getNumThreadBlocks(), "dgBlockCounts");
cl::Program program = context.createProgram(OpenCLKernelSources::compact);
countKernel = cl::Kernel(program, "countElts");
moveValidKernel = cl::Kernel(program, "moveValidElementsStaged");
}
