OpenCLSort::OpenCLSort(OpenCLContext& context, SortTrait* trait, unsigned int length) : context(context), trait(trait),
dataRange(NULL), bucketOfElement(NULL), offsetInBucket(NULL), bucketOffset(NULL), buckets(NULL), dataLength(length) {
// Create kernels.
std::map<std::string, std::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();
replacements["VALUE_IS_INT2"] = (trait->getDataType() == std::string("int2") ? "1" : "0");
cl::Program program = context.createProgram(context.replaceStrings(OpenCLKernelSources::sort, replacements));
shortListKernel = cl::Kernel(program, "sortShortList");
computeRangeKernel = cl::Kernel(program, "computeRange");
assignElementsKernel = cl::Kernel(program, "assignElementsToBuckets");
computeBucketPositionsKernel = cl::Kernel(program, "computeBucketPositions");
copyToBucketsKernel = cl::Kernel(program, "copyDataToBuckets");
sortBucketsKernel = cl::Kernel(program, "sortBuckets");
// Work out the work group sizes for various kernels.
unsigned int maxGroupSize = std::min(256, (int) context.getDevice().getInfo<CL_DEVICE_MAX_WORK_GROUP_SIZE>());
int maxSharedMem = context.getDevice().getInfo<CL_DEVICE_LOCAL_MEM_SIZE>();
unsigned int maxLocalBuffer = (unsigned int) ((maxSharedMem/trait->getDataSize())/2);
unsigned int maxRangeSize = std::min(maxGroupSize, (unsigned int) computeRangeKernel.getWorkGroupInfo<CL_KERNEL_WORK_GROUP_SIZE>(context.getDevice()));
unsigned int maxPositionsSize = std::min(maxGroupSize, (unsigned int) computeBucketPositionsKernel.getWorkGroupInfo<CL_KERNEL_WORK_GROUP_SIZE>(context.getDevice()));
unsigned int maxShortListSize = shortListKernel.getWorkGroupInfo<CL_KERNEL_WORK_GROUP_SIZE>(context.getDevice());
// On Qualcomm's OpenCL, it's essential to check against maxShortListSize. Otherwise you get a crash.
// But AMD's OpenCL returns an inappropriately small value for it that is much shorter than the actual
// maximum, so including the check hurts performance. For the moment I'm going to just comment it out.
// If we officially support Qualcomm in the future, we'll need to do something better.
isShortList = (length <= maxLocalBuffer/* && length < maxShortListSize*/);
for (rangeKernelSize = 1; rangeKernelSize*2 <= maxRangeSize; rangeKernelSize *= 2)
;
positionsKernelSize = std::min(rangeKernelSize, maxPositionsSize);
sortKernelSize = (isShortList ? rangeKernelSize : rangeKernelSize/2);
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 = new OpenCLArray(context, 2, trait->getKeySize(), "sortDataRange");
bucketOffset = OpenCLArray::create<cl_uint>(context, numBuckets, "bucketOffset");
bucketOfElement = OpenCLArray::create<cl_uint>(context, length, "bucketOfElement");
offsetInBucket = OpenCLArray::create<cl_uint>(context, length, "offsetInBucket");
buckets = new OpenCLArray(context, length, trait->getDataSize(), "buckets");
}
}
OpenCLSort::OpenCLSort(OpenCLContext& context, SortTrait* trait, unsigned int length) : context(context), trait(trait),
dataRange(NULL), bucketOfElement(NULL), offsetInBucket(NULL), bucketOffset(NULL), buckets(NULL), dataLength(length) {
// Create kernels.
std::map<std::string, std::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();
replacements["VALUE_IS_INT2"] = (trait->getDataType() == std::string("int2") ? "1" : "0");
cl::Program program = context.createProgram(context.replaceStrings(OpenCLKernelSources::sort, replacements));
shortListKernel = cl::Kernel(program, "sortShortList");
computeRangeKernel = cl::Kernel(program, "computeRange");
assignElementsKernel = cl::Kernel(program, "assignElementsToBuckets");
computeBucketPositionsKernel = cl::Kernel(program, "computeBucketPositions");
copyToBucketsKernel = cl::Kernel(program, "copyDataToBuckets");
sortBucketsKernel = cl::Kernel(program, "sortBuckets");
// Work out the work group sizes for various kernels.
unsigned int maxGroupSize = std::min(256, (int) context.getDevice().getInfo<CL_DEVICE_MAX_WORK_GROUP_SIZE>());
int maxSharedMem = context.getDevice().getInfo<CL_DEVICE_LOCAL_MEM_SIZE>();
unsigned int maxLocalBuffer = (unsigned int) ((maxSharedMem/trait->getDataSize())/2);
isShortList = (length <= maxLocalBuffer);
for (rangeKernelSize = 1; rangeKernelSize*2 <= maxGroupSize; rangeKernelSize *= 2)
;
positionsKernelSize = rangeKernelSize;
sortKernelSize = (isShortList ? rangeKernelSize : rangeKernelSize/2);
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 = new OpenCLArray(context, 2, trait->getKeySize(), "sortDataRange");
bucketOffset = OpenCLArray::create<cl_uint>(context, numBuckets, "bucketOffset");
bucketOfElement = OpenCLArray::create<cl_uint>(context, length, "bucketOfElement");
offsetInBucket = OpenCLArray::create<cl_uint>(context, length, "offsetInBucket");
buckets = new OpenCLArray(context, length, trait->getDataSize(), "buckets");
}
}
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;
}
}
}
