CudaSort::CudaSort(CudaContext& context, SortTrait* trait, unsigned int length) : context(context), trait(trait), dataLength(length) {
// Create kernels.
map<string, string> replacements;
replacements["DATA_TYPE"] = trait->getDataType();
replacements["KEY_TYPE"] = trait->getKeyType();
replacements["SORT_KEY"] = trait->getSortKey();
replacements["MIN_KEY"] = trait->getMinKey();
replacements["MAX_KEY"] = trait->getMaxKey();
replacements["MAX_VALUE"] = trait->getMaxValue();
CUmodule module = context.createModule(context.replaceStrings(CudaKernelSources::sort, replacements));
shortListKernel = context.getKernel(module, "sortShortList");
shortList2Kernel = context.getKernel(module, "sortShortList2");
computeRangeKernel = context.getKernel(module, "computeRange");
assignElementsKernel = context.getKernel(module, "assignElementsToBuckets");
computeBucketPositionsKernel = context.getKernel(module, "computeBucketPositions");
copyToBucketsKernel = context.getKernel(module, "copyDataToBuckets");
sortBucketsKernel = context.getKernel(module, "sortBuckets");
// Work out the work group sizes for various kernels.
int maxBlockSize;
cuDeviceGetAttribute(&maxBlockSize, CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_X, context.getDevice());
int maxSharedMem;
cuDeviceGetAttribute(&maxSharedMem, CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK, context.getDevice());
int maxLocalBuffer = (maxSharedMem/trait->getDataSize())/2;
int maxShortList = min(8192, max(maxLocalBuffer, CudaContext::ThreadBlockSize*context.getNumThreadBlocks()));
isShortList = (length <= maxShortList);
for (rangeKernelSize = 1; rangeKernelSize*2 <= maxBlockSize; rangeKernelSize *= 2)
;
positionsKernelSize = rangeKernelSize;
sortKernelSize = (isShortList ? rangeKernelSize/2 : rangeKernelSize/4);
if (rangeKernelSize > length)
rangeKernelSize = length;
if (sortKernelSize > maxLocalBuffer)
sortKernelSize = maxLocalBuffer;
unsigned int targetBucketSize = sortKernelSize/2;
unsigned int numBuckets = length/targetBucketSize;
if (numBuckets < 1)
numBuckets = 1;
if (positionsKernelSize > numBuckets)
positionsKernelSize = numBuckets;
// Create workspace arrays.
if (!isShortList) {
dataRange.initialize(context, 2, trait->getKeySize(), "sortDataRange");
bucketOffset.initialize<uint1>(context, numBuckets, "bucketOffset");
bucketOfElement.initialize<uint1>(context, length, "bucketOfElement");
offsetInBucket.initialize<uint1>(context, length, "offsetInBucket");
}
buckets.initialize(context, length, trait->getDataSize(), "buckets");
}
CudaNonbondedUtilities::CudaNonbondedUtilities(CudaContext& context) : context(context), cutoff(-1.0), useCutoff(false), anyExclusions(false), usePadding(true),
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 to use.
string errorMessage = "Error initializing nonbonded utilities";
int multiprocessors;
CHECK_RESULT(cuDeviceGetAttribute(&multiprocessors, CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT, context.getDevice()));
numForceThreadBlocks = 4*multiprocessors;
forceThreadBlockSize = (context.getComputeCapability() < 2.0 ? 128 : 256);
}
