void MultiplyMD::execEventScalar(typename MDEventWorkspace<MDE, nd>::sptr ws)
{
// Get the scalar multiplying
float scalar = float(m_rhs_scalar->dataY(0)[0]);
float scalarError = float(m_rhs_scalar->dataE(0)[0]);
float scalarRelativeErrorSquared = (scalarError * scalarError) / (scalar * scalar);
// Get all the MDBoxes contained
MDBoxBase<MDE,nd> * parentBox = ws->getBox();
std::vector<API::IMDNode *> boxes;
parentBox->getBoxes(boxes, 1000, true);
bool fileBackedTarget(false);
Kernel::DiskBuffer *dbuff(NULL);
if(ws->isFileBacked())
{
fileBackedTarget = true;
dbuff = ws->getBoxController()->getFileIO();
}
for (size_t i=0; i<boxes.size(); i++)
{
MDBox<MDE,nd> * box = dynamic_cast<MDBox<MDE,nd> *>(boxes[i]);
if (box)
{
typename std::vector<MDE> & events = box->getEvents();
size_t ic(events.size());
typename std::vector<MDE>::iterator it = events.begin();
typename std::vector<MDE>::iterator it_end = events.end();
for (; it != it_end; it++)
{
// Multiply weight by a scalar, propagating error
float oldSignal = it->getSignal();
float signal = oldSignal * scalar;
float errorSquared = signal * signal * (it->getErrorSquared() / (oldSignal * oldSignal) + scalarRelativeErrorSquared);
it->setSignal(signal);
it->setErrorSquared(errorSquared);
}
box->releaseEvents();
if(fileBackedTarget && ic>0)
{
Kernel::ISaveable *const pSaver(box->getISaveable());
dbuff->toWrite(pSaver);
}
}
}
// Recalculate the totals
ws->refreshCache();
// Mark file-backed workspace as dirty
ws->setFileNeedsUpdating(true);
}
void CreateMDWorkspace::finish(typename MDEventWorkspace<MDE, nd>::sptr ws)
{
// ------------ Set up the box controller ----------------------------------
BoxController_sptr bc = ws->getBoxController();
this->setBoxController(bc);
// Split to level 1
ws->splitBox();
// Do we split more due to MinRecursionDepth?
int minDepth = this->getProperty("MinRecursionDepth");
if (minDepth<0) throw std::invalid_argument("MinRecursionDepth must be >= 0.");
ws->setMinRecursionDepth(size_t(minDepth));
}
void CloneMDWorkspace::doClone(const typename MDEventWorkspace<MDE, nd>::sptr ws)
{
std::string outWSName = getPropertyValue("OutputWorkspace");
Progress prog(this, 0.0, 10.0, 100);
BoxController_sptr bc = ws->getBoxController();
if (!bc) throw std::runtime_error("Error with InputWorkspace: no BoxController!");
if (bc->isFileBacked())
{
// Generate a new filename to copy to
prog.report("Copying File");
std::string originalFile = bc->getFilename();
std::string outFilename = getPropertyValue("Filename");
if (outFilename.empty())
{
// Auto-generated name
Poco::Path path = Poco::Path(originalFile).absolute();
std::string newName = path.getBaseName() + "_clone." + path.getExtension();
path.setFileName(newName);
outFilename = path.toString();
}
// Perform the copying
g_log.notice() << "Cloned workspace file being copied to: " << outFilename << std::endl;
Poco::File(originalFile).copyTo(outFilename);
g_log.information() << "File copied successfully." << std::endl;
// Now load it back
IAlgorithm_sptr alg = createSubAlgorithm("LoadMD", 0.5, 1.0, false);
alg->setPropertyValue("Filename", outFilename);
alg->setPropertyValue("FileBackEnd", "1");
alg->setPropertyValue("Memory", "0"); //TODO: How much memory?
alg->setPropertyValue("OutputWorkspace", outWSName);
alg->executeAsSubAlg();
// Set the output workspace to this
IMDEventWorkspace_sptr outWS = alg->getProperty("OutputWorkspace");
this->setProperty("OutputWorkspace", outWS);
}
else
{
// Perform the clone in memory.
boost::shared_ptr<MDEventWorkspace<MDE,nd> > outWS(new MDEventWorkspace<MDE,nd>(*ws));
this->setProperty("OutputWorkspace", boost::dynamic_pointer_cast<IMDEventWorkspace>(outWS) );
}
}
void SliceMD::slice(typename MDEventWorkspace<MDE, nd>::sptr ws) {
// Create the ouput workspace
typename MDEventWorkspace<OMDE, ond>::sptr outWS(
new MDEventWorkspace<OMDE, ond>());
for (size_t od = 0; od < m_binDimensions.size(); od++) {
outWS->addDimension(m_binDimensions[od]);
}
outWS->setCoordinateSystem(ws->getSpecialCoordinateSystem());
outWS->initialize();
// Copy settings from the original box controller
BoxController_sptr bc = ws->getBoxController();
// store wrute buffer size for the future
// uint64_t writeBufSize =
// bc->getFileIO()getDiskBuffer().getWriteBufferSize();
// and disable write buffer (if any) for input MD Events for this algorithm
// purposes;
// bc->setCacheParameters(1,0);
BoxController_sptr obc = outWS->getBoxController();
// Use the "number of bins" as the "split into" parameter
for (size_t od = 0; od < m_binDimensions.size(); od++)
obc->setSplitInto(od, m_binDimensions[od]->getNBins());
obc->setSplitThreshold(bc->getSplitThreshold());
bool bTakeDepthFromInputWorkspace =
getProperty("TakeMaxRecursionDepthFromInput");
int tempDepth = getProperty("MaxRecursionDepth");
size_t maxDepth =
bTakeDepthFromInputWorkspace ? bc->getMaxDepth() : size_t(tempDepth);
obc->setMaxDepth(maxDepth);
// size_t outputSize = writeBufSize;
// obc->setCacheParameters(sizeof(OMDE),outputSize);
obc->resetNumBoxes();
// Perform the first box splitting
outWS->splitBox();
size_t lastNumBoxes = obc->getTotalNumMDBoxes();
// --- File back end ? ----------------
std::string filename = getProperty("OutputFilename");
if (!filename.empty()) {
// First save to the NXS file
g_log.notice() << "Running SaveMD to create file back-end" << std::endl;
IAlgorithm_sptr alg = createChildAlgorithm("SaveMD");
alg->setPropertyValue("Filename", filename);
alg->setProperty("InputWorkspace", outWS);
alg->setProperty("MakeFileBacked", true);
alg->executeAsChildAlg();
if (!obc->isFileBacked())
throw std::runtime_error("SliceMD with file-backed output: Can not set "
"up file-backed output workspace ");
auto IOptr = obc->getFileIO();
size_t outBufSize = IOptr->getWriteBufferSize();
// the buffer size for resulting workspace; reasonable size is at least 10
// data chunk sizes (nice to verify)
if (outBufSize < 10 * IOptr->getDataChunk()) {
outBufSize = 10 * IOptr->getDataChunk();
IOptr->setWriteBufferSize(outBufSize);
}
}
// Function defining which events (in the input dimensions) to place in the
// output
MDImplicitFunction *function = this->getImplicitFunctionForChunk(NULL, NULL);
std::vector<API::IMDNode *> boxes;
// Leaf-only; no depth limit; with the implicit function passed to it.
ws->getBox()->getBoxes(boxes, 1000, true, function);
// Sort boxes by file position IF file backed. This reduces seeking time,
// hopefully.
bool fileBackedWS = bc->isFileBacked();
if (fileBackedWS)
API::IMDNode::sortObjByID(boxes);
Progress *prog = new Progress(this, 0.0, 1.0, boxes.size());
// The root of the output workspace
MDBoxBase<OMDE, ond> *outRootBox = outWS->getBox();
// if target workspace has events, we should count them as added
uint64_t totalAdded = outWS->getNEvents();
uint64_t numSinceSplit = 0;
// Go through every box for this chunk.
// PARALLEL_FOR_IF( !bc->isFileBacked() )
for (int i = 0; i < int(boxes.size()); i++) {
MDBox<MDE, nd> *box = dynamic_cast<MDBox<MDE, nd> *>(boxes[i]);
// Perform the binning in this separate method.
if (box) {
// An array to hold the rotated/transformed coordinates
coord_t outCenter[ond];
const std::vector<MDE> &events = box->getConstEvents();
typename std::vector<MDE>::const_iterator it = events.begin();
typename std::vector<MDE>::const_iterator it_end = events.end();
for (; it != it_end; it++) {
// Cache the center of the event (again for speed)
const coord_t *inCenter = it->getCenter();
if (function->isPointContained(inCenter)) {
// Now transform to the output dimensions
m_transformFromOriginal->apply(inCenter, outCenter);
// Create the event
OMDE newEvent(it->getSignal(), it->getErrorSquared(), outCenter);
// Copy extra data, if any
copyEvent(*it, newEvent);
// Add it to the workspace
outRootBox->addEvent(newEvent);
numSinceSplit++;
}
}
box->releaseEvents();
// Ask BC if one needs to split boxes
if (obc->shouldSplitBoxes(totalAdded, numSinceSplit, lastNumBoxes))
// if (numSinceSplit > 20000000 || (i == int(boxes.size()-1)))
{
// This splits up all the boxes according to split thresholds and sizes.
Kernel::ThreadScheduler *ts = new ThreadSchedulerFIFO();
ThreadPool tp(ts);
outWS->splitAllIfNeeded(ts);
tp.joinAll();
// Accumulate stats
totalAdded += numSinceSplit;
numSinceSplit = 0;
lastNumBoxes = obc->getTotalNumMDBoxes();
// Progress reporting
if (!fileBackedWS)
prog->report(i);
}
if (fileBackedWS) {
if (!(i % 10))
prog->report(i);
}
} // is box
} // for each box in the vector
prog->report();
outWS->splitAllIfNeeded(NULL);
// Refresh all cache.
outWS->refreshCache();
g_log.notice() << totalAdded << " " << OMDE::getTypeName()
<< "'s added to the output workspace." << std::endl;
if (outWS->isFileBacked()) {
// Update the file-back-end
g_log.notice() << "Running SaveMD" << std::endl;
IAlgorithm_sptr alg = createChildAlgorithm("SaveMD");
alg->setProperty("UpdateFileBackEnd", true);
alg->setProperty("InputWorkspace", outWS);
alg->executeAsChildAlg();
}
// return the size of the input workspace write buffer to its initial value
// bc->setCacheParameters(sizeof(MDE),writeBufSize);
this->setProperty("OutputWorkspace",
boost::dynamic_pointer_cast<IMDEventWorkspace>(outWS));
delete prog;
}
void BinMD::binByIterating(typename MDEventWorkspace<MDE, nd>::sptr ws) {
BoxController_sptr bc = ws->getBoxController();
// store exisiting write buffer size for the future
// uint64_t writeBufSize =bc->getDiskBuffer().getWriteBufferSize();
// and disable write buffer (if any) for input MD Events for this algorithm
// purposes;
// bc->setCacheParameters(1,0);
// Cache some data to speed up accessing them a bit
indexMultiplier = new size_t[m_outD];
for (size_t d = 0; d < m_outD; d++) {
if (d > 0)
indexMultiplier[d] = outWS->getIndexMultiplier()[d - 1];
else
indexMultiplier[d] = 1;
}
signals = outWS->getSignalArray();
errors = outWS->getErrorSquaredArray();
numEvents = outWS->getNumEventsArray();
// Start with signal/error/numEvents at 0.0
outWS->setTo(0.0, 0.0, 0.0);
// The dimension (in the output workspace) along which we chunk for parallel
// processing
// TODO: Find the smartest dimension to chunk against
size_t chunkDimension = 0;
// How many bins (in that dimension) per chunk.
// Try to split it so each core will get 2 tasks:
int chunkNumBins = int(m_binDimensions[chunkDimension]->getNBins() /
(PARALLEL_GET_MAX_THREADS * 2));
if (chunkNumBins < 1)
chunkNumBins = 1;
// Do we actually do it in parallel?
bool doParallel = getProperty("Parallel");
// Not if file-backed!
if (bc->isFileBacked())
doParallel = false;
if (!doParallel)
chunkNumBins = int(m_binDimensions[chunkDimension]->getNBins());
// Total number of steps
size_t progNumSteps = 0;
if (prog)
prog->setNotifyStep(0.1);
if (prog)
prog->resetNumSteps(100, 0.00, 1.0);
// Run the chunks in parallel. There is no overlap in the output workspace so
// it is thread safe to write to it..
// cppcheck-suppress syntaxError
PRAGMA_OMP( parallel for schedule(dynamic,1) if (doParallel) )
for (int chunk = 0;
chunk < int(m_binDimensions[chunkDimension]->getNBins());
chunk += chunkNumBins) {
PARALLEL_START_INTERUPT_REGION
// Region of interest for this chunk.
std::vector<size_t> chunkMin(m_outD);
std::vector<size_t> chunkMax(m_outD);
for (size_t bd = 0; bd < m_outD; bd++) {
// Same limits in the other dimensions
chunkMin[bd] = 0;
chunkMax[bd] = m_binDimensions[bd]->getNBins();
}
// Parcel out a chunk in that single dimension dimension
chunkMin[chunkDimension] = size_t(chunk);
if (size_t(chunk + chunkNumBins) >
m_binDimensions[chunkDimension]->getNBins())
chunkMax[chunkDimension] = m_binDimensions[chunkDimension]->getNBins();
else
chunkMax[chunkDimension] = size_t(chunk + chunkNumBins);
// Build an implicit function (it needs to be in the space of the
// MDEventWorkspace)
MDImplicitFunction *function =
this->getImplicitFunctionForChunk(chunkMin.data(), chunkMax.data());
// Use getBoxes() to get an array with a pointer to each box
std::vector<API::IMDNode *> boxes;
// Leaf-only; no depth limit; with the implicit function passed to it.
ws->getBox()->getBoxes(boxes, 1000, true, function);
// Sort boxes by file position IF file backed. This reduces seeking time,
// hopefully.
if (bc->isFileBacked())
API::IMDNode::sortObjByID(boxes);
// For progress reporting, the # of boxes
if (prog) {
PARALLEL_CRITICAL(BinMD_progress) {
g_log.debug() << "Chunk " << chunk << ": found " << boxes.size()
<< " boxes within the implicit function.\n";
progNumSteps += boxes.size();
prog->setNumSteps(progNumSteps);
}
}
// Go through every box for this chunk.
for (auto &boxe : boxes) {
MDBox<MDE, nd> *box = dynamic_cast<MDBox<MDE, nd> *>(boxe);
// Perform the binning in this separate method.
if (box && !box->getIsMasked())
this->binMDBox(box, chunkMin.data(), chunkMax.data());
// Progress reporting
if (prog)
prog->report();
// For early cancelling of the loop
if (this->m_cancel)
break;
} // for each box in the vector
PARALLEL_END_INTERUPT_REGION
} // for each chunk in parallel
PARALLEL_CHECK_INTERUPT_REGION
// Now the implicit function
if (implicitFunction) {
if (prog)
prog->report("Applying implicit function.");
signal_t nan = std::numeric_limits<signal_t>::quiet_NaN();
outWS->applyImplicitFunction(implicitFunction, nan, nan);
}
// return the size of the input workspace write buffer to its initial value
// bc->setCacheParameters(sizeof(MDE),writeBufSize);
}
void LoadMD::doLoad(typename MDEventWorkspace<MDE, nd>::sptr ws) {
// Are we using the file back end?
bool fileBackEnd = getProperty("FileBackEnd");
if (fileBackEnd && m_BoxStructureAndMethadata)
throw std::invalid_argument("Combination of BoxStructureOnly or "
"MetaDataOnly were set to TRUE with "
"fileBackEnd "
": this is not possible.");
CPUTimer tim;
auto prog = new Progress(this, 0.0, 1.0, 100);
prog->report("Opening file.");
std::string title;
try {
m_file->getAttr("title", title);
} catch (std::exception &) {
// Leave the title blank if error on loading
}
ws->setTitle(title);
// Load the WorkspaceHistory "process"
if (this->getProperty("LoadHistory")) {
ws->history().loadNexus(m_file.get());
}
this->loadAffineMatricies(boost::dynamic_pointer_cast<IMDWorkspace>(ws));
m_file->closeGroup();
m_file->close();
// Add each of the dimension
for (size_t d = 0; d < nd; d++)
ws->addDimension(m_dims[d]);
// Coordinate system
ws->setCoordinateSystem(m_coordSystem);
// ----------------------------------------- Box Structure
// ------------------------------
prog->report("Reading box structure from HDD.");
MDBoxFlatTree FlatBoxTree;
int nDims = static_cast<int>(nd); // should be safe
FlatBoxTree.loadBoxStructure(m_filename, nDims, MDE::getTypeName());
BoxController_sptr bc = ws->getBoxController();
bc->fromXMLString(FlatBoxTree.getBCXMLdescr());
prog->report("Restoring box structure and connectivity");
std::vector<API::IMDNode *> boxTree;
FlatBoxTree.restoreBoxTree(boxTree, bc, fileBackEnd,
m_BoxStructureAndMethadata);
size_t numBoxes = boxTree.size();
// ---------------------------------------- DEAL WITH BOXES
// ------------------------------------
if (fileBackEnd) { // TODO:: call to the file format factory
auto loader = boost::shared_ptr<API::IBoxControllerIO>(
new DataObjects::BoxControllerNeXusIO(bc.get()));
loader->setDataType(sizeof(coord_t), MDE::getTypeName());
bc->setFileBacked(loader, m_filename);
// boxes have been already made file-backed when restoring the boxTree;
// How much memory for the cache?
{
// TODO: Clean up, only a write buffer now
double mb = getProperty("Memory");
// Defaults have changed, default disk buffer size should be 10 data
// chunks TODO: find optimal, 100 may be better.
if (mb <= 0)
mb = double(10 * loader->getDataChunk() * sizeof(MDE)) /
double(1024 * 1024);
// Express the cache memory in units of number of events.
uint64_t cacheMemory =
static_cast<uint64_t>((mb * 1024. * 1024.) / sizeof(MDE)) + 1;
// Set these values in the diskMRU
bc->getFileIO()->setWriteBufferSize(cacheMemory);
g_log.information() << "Setting a DiskBuffer cache size of " << mb
<< " MB, or " << cacheMemory << " events.\n";
}
} // Not file back end
else if (!m_BoxStructureAndMethadata) {
// ---------------------------------------- READ IN THE BOXES
// ------------------------------------
// TODO:: call to the file format factory
auto loader =
file_holder_type(new DataObjects::BoxControllerNeXusIO(bc.get()));
loader->setDataType(sizeof(coord_t), MDE::getTypeName());
loader->openFile(m_filename, "r");
const std::vector<uint64_t> &BoxEventIndex = FlatBoxTree.getEventIndex();
prog->setNumSteps(numBoxes);
for (size_t i = 0; i < numBoxes; i++) {
prog->report();
MDBox<MDE, nd> *box = dynamic_cast<MDBox<MDE, nd> *>(boxTree[i]);
if (!box)
continue;
if (BoxEventIndex[2 * i + 1] >
0) // Load in memory NOT using the file as the back-end,
{
boxTree[i]->reserveMemoryForLoad(BoxEventIndex[2 * i + 1]);
boxTree[i]->loadAndAddFrom(
loader.get(), BoxEventIndex[2 * i],
static_cast<size_t>(BoxEventIndex[2 * i + 1]));
}
}
loader->closeFile();
} else // box structure and metadata only
{
}
g_log.debug() << tim
<< " to create all the boxes and fill them with events.\n";
// Box of ID 0 is the head box.
ws->setBox(boxTree[0]);
// Make sure the max ID is ok for later ID generation
bc->setMaxId(numBoxes);
// end-of bMetaDataOnly
// Refresh cache
// TODO:if(!fileBackEnd)ws->refreshCache();
ws->refreshCache();
g_log.debug() << tim << " to refreshCache(). " << ws->getNPoints()
<< " points after refresh.\n";
g_log.debug() << tim << " to finish up.\n";
delete prog;
}
void SaveMD::doSaveEvents(typename MDEventWorkspace<MDE, nd>::sptr ws) {
std::string filename = getPropertyValue("Filename");
bool update = getProperty("UpdateFileBackEnd");
bool MakeFileBacked = getProperty("MakeFileBacked");
bool wsIsFileBacked = ws->isFileBacked();
if (update && MakeFileBacked)
throw std::invalid_argument(
"Please choose either UpdateFileBackEnd or MakeFileBacked, not both.");
if (MakeFileBacked && wsIsFileBacked)
throw std::invalid_argument(
"You picked MakeFileBacked but the workspace is already file-backed!");
BoxController_sptr bc = ws->getBoxController();
if (!wsIsFileBacked) { // Erase the file if it exists
Poco::File oldFile(filename);
if (oldFile.exists())
oldFile.remove();
}
auto prog = new Progress(this, 0.0, 0.05, 1);
if (update) // workspace has its own file and ignores any changes to the
// algorithm parameters
{
if (!ws->isFileBacked())
throw std::runtime_error(" attempt to update non-file backed workspace");
filename = bc->getFileIO()->getFileName();
}
//-----------------------------------------------------------------------------------------------------
// create or open WS group and put there additional information about WS and
// its dimensions
int nDims = static_cast<int>(nd);
bool data_exist;
auto file = file_holder_type(MDBoxFlatTree::createOrOpenMDWSgroup(
filename, nDims, MDE::getTypeName(), false, data_exist));
// Save each NEW ExperimentInfo to a spot in the file
MDBoxFlatTree::saveExperimentInfos(file.get(), ws);
if (!update || !data_exist) {
MDBoxFlatTree::saveWSGenericInfo(file.get(), ws);
}
file->closeGroup();
file->close();
MDBoxFlatTree BoxFlatStruct;
//-----------------------------------------------------------------------------------------------------
if (update) // the workspace is already file backed;
{
// remove all boxes from the DiskBuffer. DB will calculate boxes positions
// on HDD.
bc->getFileIO()->flushCache();
// flatten the box structure; this will remember boxes file positions in the
// box structure
BoxFlatStruct.initFlatStructure(ws, filename);
} else // not file backed;
{
// the boxes file positions are unknown and we need to calculate it.
BoxFlatStruct.initFlatStructure(ws, filename);
// create saver class
auto Saver = boost::shared_ptr<API::IBoxControllerIO>(
new DataObjects::BoxControllerNeXusIO(bc.get()));
Saver->setDataType(sizeof(coord_t), MDE::getTypeName());
if (MakeFileBacked) {
// store saver with box controller
bc->setFileBacked(Saver, filename);
// get access to boxes array
std::vector<API::IMDNode *> &boxes = BoxFlatStruct.getBoxes();
// calculate the position of the boxes on file, indicating to make them
// saveable and that the boxes were not saved.
BoxFlatStruct.setBoxesFilePositions(true);
prog->resetNumSteps(boxes.size(), 0.06, 0.90);
for (auto &boxe : boxes) {
auto saveableTag = boxe->getISaveable();
if (saveableTag) // only boxes can be saveable
{
// do not spend time on empty boxes
if (boxe->getDataInMemorySize() == 0)
continue;
// save boxes directly using the boxes file postion, precalculated in
// boxFlatStructure.
saveableTag->save();
// remove boxes data from memory. This will actually correctly set the
// tag indicatin that data were not loaded.
saveableTag->clearDataFromMemory();
// put boxes into write buffer wich will save them when necessary
// Saver->toWrite(saveTag);
prog->report("Saving Box");
}
}
// remove everything from diskBuffer; (not sure if it really necessary
// but just in case , should not make any harm)
Saver->flushCache();
// drop NeXus on HDD (not sure if it really necessary but just in case )
Saver->flushData();
} else // just save data, and finish with it
{
Saver->openFile(filename, "w");
BoxFlatStruct.setBoxesFilePositions(false);
std::vector<API::IMDNode *> &boxes = BoxFlatStruct.getBoxes();
std::vector<uint64_t> &eventIndex = BoxFlatStruct.getEventIndex();
prog->resetNumSteps(boxes.size(), 0.06, 0.90);
for (size_t i = 0; i < boxes.size(); i++) {
if (eventIndex[2 * i + 1] == 0)
continue;
boxes[i]->saveAt(Saver.get(), eventIndex[2 * i]);
prog->report("Saving Box");
}
Saver->closeFile();
}
}
// -------------- Save Box Structure -------------------------------------
// OK, we've filled these big arrays of data representing flat box structrre.
// Save them.
progress(0.91, "Writing Box Data");
prog->resetNumSteps(8, 0.92, 1.00);
// Save box structure;
BoxFlatStruct.saveBoxStructure(filename);
delete prog;
ws->setFileNeedsUpdating(false);
}
void BinToMDHistoWorkspace::do_centerpointBin(typename MDEventWorkspace<MDE, nd>::sptr ws)
{
bool DODEBUG = true;
CPUTimer tim;
// Number of output binning dimensions found
size_t outD = binDimensions.size();
//Since the costs are not known ahead of time, use a simple FIFO buffer.
ThreadScheduler * ts = new ThreadSchedulerFIFO();
// Create the threadpool with: all CPUs, a progress reporter
ThreadPool tp(ts, 0, prog);
// Big efficiency gain is obtained by grouping a few bins per task.
size_t binsPerTask = 100;
// For progress reporting, the approx # of tasks
if (prog)
prog->setNumSteps( int(outWS->getNPoints()/100) );
// The root-level box.
IMDBox<MDE,nd> * rootBox = ws->getBox();
// This is the limit to loop over in each dimension
size_t * index_max = new size_t[outD];
for (size_t bd=0; bd<outD; bd++) index_max[bd] = binDimensions[bd]->getNBins();
// Cache a calculation to convert indices x,y,z,t into a linear index.
size_t * index_maker = new size_t[outD];
Utils::NestedForLoop::SetUpIndexMaker(outD, index_maker, index_max);
int numPoints = int(outWS->getNPoints());
// Run in OpenMP with dynamic scheduling and a smallish chunk size (binsPerTask)
// Right now, not parallel for file-backed systems.
bool fileBacked = (ws->getBoxController()->getFile() != NULL);
PRAGMA_OMP(parallel for schedule(dynamic, binsPerTask) if (!fileBacked) )
for (int i=0; i < numPoints; i++)
{
PARALLEL_START_INTERUPT_REGION
size_t linear_index = size_t(i);
// nd >= outD in all cases so this is safe.
size_t index[nd];
// Get the index at each dimension for this bin.
Utils::NestedForLoop::GetIndicesFromLinearIndex(outD, linear_index, index_maker, index_max, index);
// Construct the bin and its coordinates
MDBin<MDE,nd> bin;
for (size_t bd=0; bd<outD; bd++)
{
// Index in this binning dimension (i_x, i_y, etc.)
size_t idx = index[bd];
// Dimension in the MDEventWorkspace
size_t d = dimensionToBinFrom[bd];
// Corresponding extents
bin.m_min[d] = binDimensions[bd]->getX(idx);
bin.m_max[d] = binDimensions[bd]->getX(idx+1);
}
bin.m_index = linear_index;
bool dimensionsUsed[nd];
for (size_t d=0; d<nd; d++)
dimensionsUsed[d] = (d<3);
// Check if the bin is in the ImplicitFunction (if any)
bool binContained = true;
if (implicitFunction)
{
binContained = implicitFunction->isPointContained(bin.m_min); //TODO. Correct argument passed to this method?
}
if (binContained)
{
// Array of bools set to true when a dimension is fully contained (binary splitting only)
bool fullyContained[nd];
for (size_t d=0; d<nd; d++)
fullyContained[d] = false;
// This will recursively bin into the sub grids
rootBox->centerpointBin(bin, fullyContained);
// Save the data into the dense histogram
outWS->setSignalAt(linear_index, bin.m_signal);
outWS->setErrorAt(linear_index, bin.m_errorSquared);
}
// Report progress but not too often.
if (((linear_index % 100) == 0) && prog ) prog->report();
PARALLEL_END_INTERUPT_REGION
} // (for each linear index)
PARALLEL_CHECK_INTERUPT_REGION
if (DODEBUG) std::cout << tim << " to run the openmp loop.\n";
delete [] index_max;
delete [] index_maker;
}
void BinToMDHistoWorkspace::binByIterating(typename MDEventWorkspace<MDE, nd>::sptr ws)
{
BoxController_sptr bc = ws->getBoxController();
// Start with signal at 0.0
outWS->setTo(0.0, 0.0);
// Cache some data to speed up accessing them a bit
indexMultiplier = new size_t[outD];
for (size_t d=0; d<outD; d++)
{
if (d > 0)
indexMultiplier[d] = outWS->getIndexMultiplier()[d-1];
else
indexMultiplier[d] = 1;
}
signals = outWS->getSignalArray();
errors = outWS->getErrorSquaredArray();
// The dimension (in the output workspace) along which we chunk for parallel processing
// TODO: Find the smartest dimension to chunk against
size_t chunkDimension = 0;
// How many bins (in that dimension) per chunk.
// Try to split it so each core will get 2 tasks:
int chunkNumBins = int(binDimensions[chunkDimension]->getNBins() / (Mantid::Kernel::ThreadPool::getNumPhysicalCores() * 2));
if (chunkNumBins < 1) chunkNumBins = 1;
// Do we actually do it in parallel?
bool doParallel = getProperty("Parallel");
// Not if file-backed!
if (bc->isFileBacked()) doParallel = false;
if (!doParallel)
chunkNumBins = int(binDimensions[chunkDimension]->getNBins());
// Total number of steps
size_t progNumSteps = 0;
if (prog) prog->setNotifyStep(0.1);
if (prog) prog->resetNumSteps(100, 0.00, 1.0);
// Run the chunks in parallel. There is no overlap in the output workspace so it is
// thread safe to write to it..
PRAGMA_OMP( parallel for schedule(dynamic,1) if (doParallel) )
for(int chunk=0; chunk < int(binDimensions[chunkDimension]->getNBins()); chunk += chunkNumBins)
{
PARALLEL_START_INTERUPT_REGION
// Region of interest for this chunk.
size_t * chunkMin = new size_t[outD];
size_t * chunkMax = new size_t[outD];
for (size_t bd=0; bd<outD; bd++)
{
// Same limits in the other dimensions
chunkMin[bd] = 0;
chunkMax[bd] = binDimensions[bd]->getNBins();
}
// Parcel out a chunk in that single dimension dimension
chunkMin[chunkDimension] = size_t(chunk);
if (size_t(chunk+chunkNumBins) > binDimensions[chunkDimension]->getNBins())
chunkMax[chunkDimension] = binDimensions[chunkDimension]->getNBins();
else
chunkMax[chunkDimension] = size_t(chunk+chunkNumBins);
// Build an implicit function (it needs to be in the space of the MDEventWorkspace)
MDImplicitFunction * function = this->getImplicitFunctionForChunk(chunkMin, chunkMax);
// Use getBoxes() to get an array with a pointer to each box
std::vector<IMDBox<MDE,nd>*> boxes;
// Leaf-only; no depth limit; with the implicit function passed to it.
ws->getBox()->getBoxes(boxes, 1000, true, function);
// Sort boxes by file position IF file backed. This reduces seeking time, hopefully.
if (bc->isFileBacked())
IMDBox<MDE, nd>::sortBoxesByFilePos(boxes);
// For progress reporting, the # of boxes
if (prog)
{
PARALLEL_CRITICAL(BinToMDHistoWorkspace_progress)
{
std::cout << "Chunk " << chunk << ": found " << boxes.size() << " boxes within the implicit function." << std::endl;
progNumSteps += boxes.size();
prog->setNumSteps( progNumSteps );
}
}
// Go through every box for this chunk.
for (size_t i=0; i<boxes.size(); i++)
{
MDBox<MDE,nd> * box = dynamic_cast<MDBox<MDE,nd> *>(boxes[i]);
// Perform the binning in this separate method.
if (box)
this->binMDBox(box, chunkMin, chunkMax);
// Progress reporting
if (prog) prog->report();
}// for each box in the vector
PARALLEL_END_INTERUPT_REGION
} // for each chunk in parallel
PARALLEL_CHECK_INTERUPT_REGION
// Now the implicit function
if (implicitFunction)
{
prog->report("Applying implicit function.");
signal_t nan = std::numeric_limits<signal_t>::quiet_NaN();
outWS->applyImplicitFunction(implicitFunction, nan, nan);
}
}
